Re: [PATCH net-next,2/2] net: mana: Add support of XDP_REDIRECT action

[Paolo Abeni <pabeni@redhat.com>]

On Thu, 2022-06-09 at 14:57 -0700, Haiyang Zhang wrote:
> Support XDP_REDIRECT action
> 
> Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>

You really should expand the changelog a little bit...

> ---
>  drivers/net/ethernet/microsoft/mana/mana.h    |  6 ++
>  .../net/ethernet/microsoft/mana/mana_bpf.c    | 64 +++++++++++++++++++
>  drivers/net/ethernet/microsoft/mana/mana_en.c | 13 +++-
>  .../ethernet/microsoft/mana/mana_ethtool.c    | 12 +++-
>  4 files changed, 93 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/net/ethernet/microsoft/mana/mana.h b/drivers/net/ethernet/microsoft/mana/mana.h
> index f198b34c232f..d58be64374c8 100644
> --- a/drivers/net/ethernet/microsoft/mana/mana.h
> +++ b/drivers/net/ethernet/microsoft/mana/mana.h
> @@ -53,12 +53,14 @@ struct mana_stats_rx {
>  	u64 bytes;
>  	u64 xdp_drop;
>  	u64 xdp_tx;
> +	u64 xdp_redirect;
>  	struct u64_stats_sync syncp;
>  };
>  
>  struct mana_stats_tx {
>  	u64 packets;
>  	u64 bytes;
> +	u64 xdp_xmit;
>  	struct u64_stats_sync syncp;
>  };
>  
> @@ -311,6 +313,8 @@ struct mana_rxq {
>  	struct bpf_prog __rcu *bpf_prog;
>  	struct xdp_rxq_info xdp_rxq;
>  	struct page *xdp_save_page;
> +	bool xdp_flush;
> +	int xdp_rc; /* XDP redirect return code */
>  
>  	/* MUST BE THE LAST MEMBER:
>  	 * Each receive buffer has an associated mana_recv_buf_oob.
> @@ -396,6 +400,8 @@ int mana_probe(struct gdma_dev *gd, bool resuming);
>  void mana_remove(struct gdma_dev *gd, bool suspending);
>  
>  void mana_xdp_tx(struct sk_buff *skb, struct net_device *ndev);
> +int mana_xdp_xmit(struct net_device *ndev, int n, struct xdp_frame **frames,
> +		  u32 flags);
>  u32 mana_run_xdp(struct net_device *ndev, struct mana_rxq *rxq,
>  		 struct xdp_buff *xdp, void *buf_va, uint pkt_len);
>  struct bpf_prog *mana_xdp_get(struct mana_port_context *apc);
> diff --git a/drivers/net/ethernet/microsoft/mana/mana_bpf.c b/drivers/net/ethernet/microsoft/mana/mana_bpf.c
> index 1d2f948b5c00..421fd39ff3a8 100644
> --- a/drivers/net/ethernet/microsoft/mana/mana_bpf.c
> +++ b/drivers/net/ethernet/microsoft/mana/mana_bpf.c
> @@ -32,9 +32,55 @@ void mana_xdp_tx(struct sk_buff *skb, struct net_device *ndev)
>  	ndev->stats.tx_dropped++;
>  }
>  
> +static int mana_xdp_xmit_fm(struct net_device *ndev, struct xdp_frame *frame,
> +			    u16 q_idx)
> +{
> +	struct sk_buff *skb;
> +
> +	skb = xdp_build_skb_from_frame(frame, ndev);
> +	if (unlikely(!skb))
> +		return -ENOMEM;

... especially considering this implementation choice: converting the
xdp frame to an skb in very bad for performances.

You could implement a mana xmit helper working on top of the xdp_frame
struct, and use it here.

Additionally you could consider revisiting the XDP_TX path: currently
it builds a skb from the xdp_buff to xmit it locally, while it could
resort to a much cheaper xdp_buff to xdp_frame conversion.

The traditional way to handle all the above is keep all the
XDP_TX/XDP_REDIRECT bits in the device-specific _run_xdp helper, that
will additional avoid several conditionals in mana_rx_skb(). 

The above refactoring would probably require a bit of work, but it will
pay-off for sure and will become more costily with time. Your choice ;)

But at the very least we need a better changelog here.

Cheers,

Paolo