On 28.01.2021 19:36, Willem de Bruijn wrote:
> On Thu, Jan 28, 2021 at 2:44 AM Heiner Kallweit <hkallwe...@gmail.com> 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
>>
>> Fixes: f1bce4ad2f1c ("r8169: add support for RTL8125")
>> Tested-by: xplo <xplo...@gmail.com>
>> Signed-off-by: Heiner Kallweit <hkallwe...@gmail.com>
>> ---
>> drivers/net/ethernet/realtek/r8169_main.c | 70 +++++++++++++++++++++--
>> 1 file changed, 64 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/net/ethernet/realtek/r8169_main.c
>> b/drivers/net/ethernet/realtek/r8169_main.c
>> index a569abe7f..457fa1404 100644
>> --- a/drivers/net/ethernet/realtek/r8169_main.c
>> +++ b/drivers/net/ethernet/realtek/r8169_main.c
>> @@ -28,6 +28,7 @@
>> #include <linux/bitfield.h>
>> #include <linux/prefetch.h>
>> #include <linux/ipv6.h>
>> +#include <linux/ptp_classify.h>
>
> this is only for UDP_HLEN? perhaps use sizeof(struct udphdr) instead
> to remove the dependency.
>
It's also for PTP_EV_PORT. But yes, we can define this locally to
remove the dependency.
>> #include <net/ip6_checksum.h>
>>
>> #include "r8169.h"
>> @@ -4046,17 +4047,70 @@ 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
>> +#define PTP_GEN_PORT 320
>> +
>> +/* 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 > 3 && trans_data_len <
>> RTL_MIN_PATCH_LEN) {
>
> trans_data_len > 3
>
> here probably means
>
> trans_data_len >= offsetof(struct udphdr, len)
>
offsetof(struct udphdr, len) = 5
Presumably the check ensures that udphdr->dest is accessible. I think we should
use skb_header_pointer() here too.
> to safely access dest. Then that is a bit more self documenting
>
>> + u16 dest = ntohs(udp_hdr(skb)->dest);
>> +
>> + if (dest == PTP_EV_PORT || dest == PTP_GEN_PORT)
>> + padto = len + RTL_MIN_PATCH_LEN -
>> trans_data_len;
>> + }
>> +
>> + if (trans_data_len < UDP_HLEN)
>
> nit: else if ?
>
I don't think so. If e.g. trans_data_len == 5 and dest port isn't a PTP port,
then the trans_data_len < UDP_HLEN branch wouldn't be execeuted.
>> + padto = max(padto, len + UDP_HLEN - 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 +4182,10 @@ 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);
>
> should this path still pad to ETH_ZLEN as a minimum when the other
> cases do not hit?
>
For most chip versions that's not needed because hw does the padding
to ETH_ZLEN. Few chip versions have a hw bug and padding needs to be
done in sw, that's handled in rtl_quirk_packet_padto().
>> }
>>
>> return true;
>> @@ -4307,6 +4362,9 @@ 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;
>> --
>> 2.30.0
>>
