On Mon, 2021-01-11 at 09:02 +0100, Magnus Karlsson wrote:
> On Tue, Jan 5, 2021 at 9:51 PM Saeed Mahameed <sa...@kernel.org>
> wrote:
> > On Mon, 2021-01-04 at 18:59 +0800, Xuan Zhuo wrote:
> > > hi
> > >
> > > In the process of developing xdp socket, we tried to directly use
> > > page to
> > > construct skb directly, to avoid data copy. And the MAC
> > > information
> > > is also in
> > > the page, which caused the linear space of skb to be empty. In
> > > this
> > > case, I
> > > encountered a problem :
> > >
> > > mlx5_core 0000:3b:00.1 eth1: Error cqe on cqn 0x817, ci 0x8, qn
> > > 0x1dbb, opcode 0xd, syndrome 0x1, vendor syndrome 0x68
> > > 00000000: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> > > 00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> > > 00000020: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> > > 00000030: 00 00 00 00 60 10 68 01 0a 00 1d bb 00 0f 9f d2
> > > WQE DUMP: WQ size 1024 WQ cur size 0, WQE index 0xf, len: 64
> > > 00000000: 00 00 0f 0a 00 1d bb 03 00 00 00 08 00 00 00 00
> > > 00000010: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> > > 00000020: 00 00 00 2b 00 08 00 00 00 00 00 05 9e e3 08 00
> > > 00000030: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
> > > mlx5_core 0000:3b:00.1 eth1: ERR CQE on SQ: 0x1dbb
> > >
> > >
> > > And when I try to copy only the mac address into the linear space
> > > of
> > > skb, the
> > > other parts are still placed in the page. When constructing skb
> > > in
> > > this way, I
> > > found that although the data can be sent successfully, the
> > > sending
> > > performance
> > > is relatively poor!!
> > >
> >
> > Hi,
> >
> > This is an expected behavior of ConnectX4-LX, ConnectX4-LX requires
> > the
> > driver to copy at least the L2 headers into the linear part, in
> > some
> > DCB/DSCP configuration it will require L3 headers.
>
> Do I understand this correctly if I say whatever is calling
> ndo_start_xmit has to make sure at least the L2 headers is in the
> linear part of the skb? If Xuan does not do this, the ConnectX4
> driver
> crashes, but if he does, it works. So from an ndo_start_xmit
> interface
> perspective, what is the requirement of an skb that is passed to it?
> Do all users of ndo_start_xmit make sure the L2 header is in the
> linear part, or are there users that do not make sure this is the
> case? Judging from the ConnectX5 code it seems that the latter is
> possible (since it has code to deal with this), but from the
> ConnectX4, it seems like the former is true (since it does not copy
> the L2 headers into the linear part as far as I can see). Sorry for
> my
> confusion, but I think it is important to get some clarity here as it
> will decide if Xuan's patch is a good idea or not in its current
> form.
>
To clarify:
Connectx4Lx, doesn't really require data to be in the linear part, I
was refereing to a HW limitation that requires the driver to copy the
L2/L3 headers (depending on current HW config) to a special area in the
tx descriptor, currently the driver copy the L2/L3 headers only from
the linear part of the SKB, but this can be changed via calling
pskb_may_pull in mlx5 ConnectX4LX tx path to make sure the linear part
has the needed data ..
Something like:
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_tx.c
b/drivers/net/ethernet/mellanox/mlx5/core/en_tx.c
index 61ed671fe741..5939fd8eed2c 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_tx.c
@@ -159,6 +159,7 @@ static inline int mlx5e_skb_l2_header_offset(struct
sk_buff *skb)
{
#define MLX5E_MIN_INLINE (ETH_HLEN + VLAN_HLEN)
+ /* need to check ret val */
+ pskb_may_pull(skb, MLX5E_MIN_INLINE);
return max(skb_network_offset(skb), MLX5E_MIN_INLINE);
}
