On Tue, 2018-11-27 at 16:07 -0800, Cong Wang wrote: > On Tue, Nov 27, 2018 at 3:48 PM Eric Dumazet <eric.duma...@gmail.com> > wrote: > > > > > > On 11/27/2018 03:21 PM, Cong Wang wrote: > > > When an ethernet frame is padded to meet the minimum ethernet > > > frame > > > size, the padding octets are not covered by the hardware > > > checksum. > > > Fortunately the padding octets are ususally zero's, which don't > > > affect > > > checksum. However, we have a switch which pads non-zero octets, > > > this > > > causes kernel hardware checksum fault repeatedly. > > > > > > Prior to commit 88078d98d1bb ("net: pskb_trim_rcsum() and > > > CHECKSUM_COMPLETE are friends"), > > > skb checksum is forced to be CHECKSUM_NONE when padding is > > > detected. > > > After it, we have to keep skb->csum updated, like what we do for > > > FCS. > > > > > > The logic is a bit complicated when dealing with both FCS and > > > padding, > > > so I wrap it up in a helper function mlx5e_csum_padding(). > > > > > > I tested this patch with RXFCS on and off, it works fine without > > > any > > > warning in both cases. > > > > > > Fixes: 88078d98d1bb ("net: pskb_trim_rcsum() and > > > CHECKSUM_COMPLETE are friends"), > > > Cc: Saeed Mahameed <sae...@mellanox.com> > > > Cc: Eric Dumazet <eduma...@google.com> > > > Signed-off-by: Cong Wang <xiyou.wangc...@gmail.com> > > > --- > > > .../net/ethernet/mellanox/mlx5/core/en_rx.c | 36 > > > +++++++++++++++++++ > > > 1 file changed, 36 insertions(+) > > > > > > diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c > > > b/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c > > > index 16985ca3248d..93c18647ca74 100644 > > > --- a/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c > > > +++ b/drivers/net/ethernet/mellanox/mlx5/core/en_rx.c > > > @@ -732,6 +732,35 @@ static u8 get_ip_proto(struct sk_buff *skb, > > > __be16 proto) > > > ((struct ipv6hdr > > > *)ip_p)->nexthdr; > > > } > > > > > > +static void mlx5e_csum_padding(struct sk_buff *skb, int > > > network_depth, > > > + __be16 proto, bool has_fcs) > > > +{ > > > + u32 frame_len = has_fcs ? skb->len - ETH_FCS_LEN : skb- > > > >len; > > > + void *ip_p = skb->data + network_depth; > > > + u32 pad_offset, pad_len; > > > + void *pad; > > > + > > > + if (likely(frame_len > ETH_ZLEN)) > > > + return; > > > + > > > > But the padding might be added on normal packets (say 1000 bytes + > > 3 bytes of padding) ? > > I never see other padding cases than ETH_ZLEN. Does ethernet standard > require padding for other cases? I only read the section "3.2.8 Pad > field" in > the standard. >
hmm, so we are going to parse each and every packets ? to find if there is an end padding ? that is costly (packet rate wise)! if the ETH_ZLEN check is good enough for now, let's stick with that. We are currently exploring our options if we can W/A this in HW/FW and avoid the extra driver overhead. > > The bug here is that mlx5 csum only includes the data in IP frame. > > > > I would simply force skb->ip_summed to CHECKSUM_NONE if any padding > > is detected. > > Totally agree, let's just skip csum completes for very small packets ( < ETH_ZLEN ). > > Otherwise, your patch needs more work when multiple frags are used > > (ie num_frags > 1 ) > > You mean using skb_header_pointer()? Yeah if we need to handle cases > larger than ETH_ZLEN. > > Thanks.