Re: [PATCH net-next] xfrm: replay: Fix ESN wrap around for GSO

[Leon Romanovsky <leon@kernel.org>]

On Tue, Sep 27, 2022 at 11:28:08AM +0200, Christian Langrock wrote:
> When using GSO it can happen that the wrong seq_hi is used for the last
> packets before the wrap around. To avoid this, we should serialize this
> last GSO packet.
> 
> Fixes: d7dbefc45cf55 ("xfrm: Add xfrm_replay_overflow functions for
> offloading")
> 

Please remove extra line between Fixes and SOB.

> Signed-off-by: Christian Langrock <christian.langrock@secunet.com>
> ---
>  include/net/xfrm.h     |  1 +
>  net/xfrm/xfrm_output.c |  2 +-
>  net/xfrm/xfrm_replay.c | 36 ++++++++++++++++++++++++++++++++++++
>  3 files changed, 38 insertions(+), 1 deletion(-)
> 
> diff --git a/include/net/xfrm.h b/include/net/xfrm.h
> index 6e8fa98f786f..49d6d974f493 100644
> --- a/include/net/xfrm.h
> +++ b/include/net/xfrm.h
> @@ -1749,6 +1749,7 @@ void xfrm_replay_advance(struct xfrm_state *x,
> __be32 net_seq);
>  int xfrm_replay_check(struct xfrm_state *x, struct sk_buff *skb, __be32
> net_seq);
>  void xfrm_replay_notify(struct xfrm_state *x, int event);
>  int xfrm_replay_overflow(struct xfrm_state *x, struct sk_buff *skb);
> +int xfrm_replay_overflow_check(struct xfrm_state *x, struct sk_buff *skb);
>  int xfrm_replay_recheck(struct xfrm_state *x, struct sk_buff *skb,
> __be32 net_seq);
> 
>  static inline int xfrm_aevent_is_on(struct net *net)
> diff --git a/net/xfrm/xfrm_output.c b/net/xfrm/xfrm_output.c
> index 9a5e79a38c67..c470a68d9c88 100644
> --- a/net/xfrm/xfrm_output.c
> +++ b/net/xfrm/xfrm_output.c
> @@ -738,7 +738,7 @@ int xfrm_output(struct sock *sk, struct sk_buff *skb)
>  		skb->encapsulation = 1;
> 
>  		if (skb_is_gso(skb)) {
> -			if (skb->inner_protocol)
> +			if (skb->inner_protocol || xfrm_replay_overflow_check(x, skb))

Maybe it is perfectly fine to call xfrm_output_gso(), but your commit
message doesn't explain what is wrong with standard flow.

>  				return xfrm_output_gso(net, sk, skb);
> 
>  			skb_shinfo(skb)->gso_type |= SKB_GSO_ESP;
> diff --git a/net/xfrm/xfrm_replay.c b/net/xfrm/xfrm_replay.c
> index 9277d81b344c..6c696b6c0a22 100644
> --- a/net/xfrm/xfrm_replay.c
> +++ b/net/xfrm/xfrm_replay.c
> @@ -750,6 +750,37 @@ int xfrm_replay_overflow(struct xfrm_state *x,
> struct sk_buff *skb)
> 
>  	return xfrm_replay_overflow_offload(x, skb);
>  }
> +
> +static int xfrm_replay_overflow_check_offload_esn(struct xfrm_state *x,
> struct sk_buff *skb)
> +{

The function returns true or false and better to have "static bool ..."
as a prototype.

> +	int ret = 0;
> +	struct xfrm_offload *xo = xfrm_offload(skb);
> +	struct xfrm_replay_state_esn *replay_esn = x->replay_esn;
> +	__u32 oseq = replay_esn->oseq;

Reversed Christmas tree.

> +
> +	if (xo && x->type->flags & XFRM_TYPE_REPLAY_PROT) {
> +		if (skb_is_gso(skb)) {

You already checked this. Maybe it is more future proof to write like
this, but it is not optimal from performance POV as you perform same
checks in datapath.

> +			oseq = oseq + 1 + skb_shinfo(skb)->gso_segs;
> +			if (unlikely(oseq < replay_esn->oseq)) {
> +				ret = 1;
> +			}
> +		}
> +	}
> +
> +	return ret;
> +}
> +
> +int xfrm_replay_overflow_check(struct xfrm_state *x, struct sk_buff *skb)

This function doesn't do much except call to another function.

> +{
> +	switch (x->repl_mode) {
> +	case XFRM_REPLAY_MODE_ESN:
> +		return xfrm_replay_overflow_check_offload_esn(x, skb);
> +	}
> +
> +	return 0;
> +
> +}
> +
>  #else
>  int xfrm_replay_overflow(struct xfrm_state *x, struct sk_buff *skb)
>  {
> @@ -764,6 +795,11 @@ int xfrm_replay_overflow(struct xfrm_state *x,
> struct sk_buff *skb)
> 
>  	return __xfrm_replay_overflow(x, skb);
>  }
> +
> +int xfrm_replay_overflow_check(struct xfrm_state *x, struct sk_buff *skb)
> +{
> +	return 0;
> +}
>  #endif
> 
>  int xfrm_init_replay(struct xfrm_state *x)
> -- 
> 2.37.1.223.g6a475b71f8