[v4,net] r8169: work around RTL8125 UDP hw bug

Commit Message

Heiner Kallweit Jan. 28, 2021, 10:01 p.m. UTC

It was reported that on RTL8125 network breaks under heavy UDP load,
e.g. torrent traffic ([0], from comment 27). Realtek confirmed a hw bug
and provided me with a test version of the r8125 driver including a
workaround. Tests confirmed that the workaround fixes the issue.
I modified the original version of the workaround to meet mainline
code style.

[0] https://bugzilla.kernel.org/show_bug.cgi?id=209839

v2:
- rebased to net
v3:
- make rtl_skb_is_udp() more robust and use skb_header_pointer()
  to access the ip(v6) header
v4:
- remove dependency on ptp_classify.h
- replace magic number with offsetof(struct udphdr, len)

Fixes: f1bce4ad2f1c ("r8169: add support for RTL8125")
Tested-by: xplo <[e-mail]>
Signed-off-by: Heiner Kallweit <[e-mail]>
---
 drivers/net/ethernet/realtek/r8169_main.c | 71 +++++++++++++++++++++--
 1 file changed, 65 insertions(+), 6 deletions(-)

Comments

Willem de Bruijn Jan. 28, 2021, 10:16 p.m. UTC | #1

On Thu, Jan 28, 2021 at 5:02 PM Heiner Kallweit <[e-mail]> wrote:
>
> It was reported that on RTL8125 network breaks under heavy UDP load,
> e.g. torrent traffic ([0], from comment 27). Realtek confirmed a hw bug
> and provided me with a test version of the r8125 driver including a
> workaround. Tests confirmed that the workaround fixes the issue.
> I modified the original version of the workaround to meet mainline
> code style.
>
> [0] https://bugzilla.kernel.org/show_bug.cgi?id=209839
>
> v2:
> - rebased to net
> v3:
> - make rtl_skb_is_udp() more robust and use skb_header_pointer()
>   to access the ip(v6) header
> v4:
> - remove dependency on ptp_classify.h
> - replace magic number with offsetof(struct udphdr, len)
>
> Fixes: f1bce4ad2f1c ("r8169: add support for RTL8125")
> Tested-by: xplo <[e-mail]>
> Signed-off-by: Heiner Kallweit <[e-mail]>

Acked-by: Willem de Bruijn <[e-mail]>

Thanks for the quick iterations. I was afraid that the switch with
implicit fall-through to the default label could trigger
Wimplicit-fallthrough, but it seems that warning is suppressed if the
case only has a break or return. It's not strictly needed in that
case, but no need to respin just for that, either.

[e-mail] Jan. 30, 2021, 5:30 a.m. UTC | #2

Hello:

This patch was applied to netdev/net.git (refs/heads/master):

On Thu, 28 Jan 2021 23:01:54 +0100 you wrote:
> It was reported that on RTL8125 network breaks under heavy UDP load,
> e.g. torrent traffic ([0], from comment 27). Realtek confirmed a hw bug
> and provided me with a test version of the r8125 driver including a
> workaround. Tests confirmed that the workaround fixes the issue.
> I modified the original version of the workaround to meet mainline
> code style.
> 
> [...]

Here is the summary with links:
  - [v4,net] r8169: work around RTL8125 UDP hw bug
    https://git.kernel.org/netdev/net/c/8d520b4de3ed

You are awesome, thank you!
--
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html

Patch

diff --git a/drivers/net/ethernet/realtek/r8169_main.c b/drivers/net/ethernet/realtek/r8169_main.c
index a569abe7f..f2269c9f5 100644
--- a/drivers/net/ethernet/realtek/r8169_main.c
+++ b/drivers/net/ethernet/realtek/r8169_main.c
@@ -4046,17 +4046,72 @@  static int rtl8169_xmit_frags(struct rtl8169_private *tp, struct sk_buff *skb,
 	return -EIO;
 }
 
-static bool rtl_test_hw_pad_bug(struct rtl8169_private *tp)
+static bool rtl_skb_is_udp(struct sk_buff *skb)
+{
+	int no = skb_network_offset(skb);
+	struct ipv6hdr *i6h, _i6h;
+	struct iphdr *ih, _ih;
+
+	switch (vlan_get_protocol(skb)) {
+	case htons(ETH_P_IP):
+		ih = skb_header_pointer(skb, no, sizeof(_ih), &_ih);
+		return ih && ih->protocol == IPPROTO_UDP;
+	case htons(ETH_P_IPV6):
+		i6h = skb_header_pointer(skb, no, sizeof(_i6h), &_i6h);
+		return i6h && i6h->nexthdr == IPPROTO_UDP;
+	default:
+		return false;
+	}
+}
+
+#define RTL_MIN_PATCH_LEN	47
+
+/* see rtl8125_get_patch_pad_len() in r8125 vendor driver */
+static unsigned int rtl8125_quirk_udp_padto(struct rtl8169_private *tp,
+					    struct sk_buff *skb)
 {
+	unsigned int padto = 0, len = skb->len;
+
+	if (rtl_is_8125(tp) && len < 128 + RTL_MIN_PATCH_LEN &&
+	    rtl_skb_is_udp(skb) && skb_transport_header_was_set(skb)) {
+		unsigned int trans_data_len = skb_tail_pointer(skb) -
+					      skb_transport_header(skb);
+
+		if (trans_data_len >= offsetof(struct udphdr, len) &&
+		    trans_data_len < RTL_MIN_PATCH_LEN) {
+			u16 dest = ntohs(udp_hdr(skb)->dest);
+
+			/* dest is a standard PTP port */
+			if (dest == 319 || dest == 320)
+				padto = len + RTL_MIN_PATCH_LEN - trans_data_len;
+		}
+
+		if (trans_data_len < sizeof(struct udphdr))
+			padto = max_t(unsigned int, padto,
+				      len + sizeof(struct udphdr) - trans_data_len);
+	}
+
+	return padto;
+}
+
+static unsigned int rtl_quirk_packet_padto(struct rtl8169_private *tp,
+					   struct sk_buff *skb)
+{
+	unsigned int padto;
+
+	padto = rtl8125_quirk_udp_padto(tp, skb);
+
 	switch (tp->mac_version) {
 	case RTL_GIGA_MAC_VER_34:
 	case RTL_GIGA_MAC_VER_60:
 	case RTL_GIGA_MAC_VER_61:
 	case RTL_GIGA_MAC_VER_63:
-		return true;
+		padto = max_t(unsigned int, padto, ETH_ZLEN);
 	default:
-		return false;
+		break;
 	}
+
+	return padto;
 }
 
 static void rtl8169_tso_csum_v1(struct sk_buff *skb, u32 *opts)
@@ -4128,9 +4183,10 @@  static bool rtl8169_tso_csum_v2(struct rtl8169_private *tp,
 
 		opts[1] |= transport_offset << TCPHO_SHIFT;
 	} else {
-		if (unlikely(skb->len < ETH_ZLEN && rtl_test_hw_pad_bug(tp)))
-			/* eth_skb_pad would free the skb on error */
-			return !__skb_put_padto(skb, ETH_ZLEN, false);
+		unsigned int padto = rtl_quirk_packet_padto(tp, skb);
+
+		/* skb_padto would free the skb on error */
+		return !__skb_put_padto(skb, padto, false);
 	}
 
 	return true;
@@ -4307,6 +4363,9 @@  static netdev_features_t rtl8169_features_check(struct sk_buff *skb,
 		if (skb->len < ETH_ZLEN)
 			features &= ~NETIF_F_CSUM_MASK;
 
+		if (rtl_quirk_packet_padto(tp, skb))
+			features &= ~NETIF_F_CSUM_MASK;
+
 		if (transport_offset > TCPHO_MAX &&
 		    rtl_chip_supports_csum_v2(tp))
 			features &= ~NETIF_F_CSUM_MASK;