Re: [PATCH net-next V2] net: track pfmemalloc drops via SKB_DROP_REASON_PFMEMALLOC

[Paolo Abeni <pabeni@redhat.com>]

On 5/9/25 4:36 PM, Jesper Dangaard Brouer wrote:
> diff --git a/include/linux/filter.h b/include/linux/filter.h
> index f5cf4d35d83e..cb31be77dd7e 100644
> --- a/include/linux/filter.h
> +++ b/include/linux/filter.h
> @@ -1073,12 +1073,21 @@ bpf_jit_binary_lock_ro(struct bpf_binary_header *hdr)
>  	return set_memory_rox((unsigned long)hdr, hdr->size >> PAGE_SHIFT);
>  }
>  
> -int sk_filter_trim_cap(struct sock *sk, struct sk_buff *skb, unsigned int cap);
> +int sk_filter_trim_cap(struct sock *sk, struct sk_buff *skb, unsigned int cap,
> +		       enum skb_drop_reason *reason);
>  static inline int sk_filter(struct sock *sk, struct sk_buff *skb)
>  {
> -	return sk_filter_trim_cap(sk, skb, 1);
> +	enum skb_drop_reason ignore_reason;
> +
> +	return sk_filter_trim_cap(sk, skb, 1, &ignore_reason);
> +}

I'm sorry to nit-pick but checkpatch is not happy about the lack of
black lines here, and I think an empty line would make the code more
readable.

[...]
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 03d20a98f8b7..a1e10a13f7c8 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -5910,7 +5910,11 @@ static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc,
>  			dev_core_stats_rx_dropped_inc(skb->dev);
>  		else
>  			dev_core_stats_rx_nohandler_inc(skb->dev);
> -		kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO);
> +
> +		if (pfmemalloc)
> +			kfree_skb_reason(skb, SKB_DROP_REASON_PFMEMALLOC);
> +		else
> +			kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO);

AFAICS we can reach here even if skb_orphan_frags_rx() fails and that
will be accounted as 'SKB_DROP_REASON_UNHANDLED_PROTO'. Perhaps it would
be better to let the 'goto out' caller set the drop reason? And also set
it to 'SKB_DROP_REASON_UNHANDLED_PROTO' in this block before the 'out:'
label.

[...]
> @@ -2637,6 +2635,7 @@ static int udp_unicast_rcv_skb(struct sock *sk, struct sk_buff *skb,
>  int __udp4_lib_rcv(struct sk_buff *skb, struct udp_table *udptable,
>  		   int proto)
>  {
> +	enum skb_drop_reason drop_reason;
>  	struct sock *sk = NULL;
>  	struct udphdr *uh;
>  	unsigned short ulen;
> @@ -2644,7 +2643,6 @@ int __udp4_lib_rcv(struct sk_buff *skb, struct udp_table *udptable,
>  	__be32 saddr, daddr;
>  	struct net *net = dev_net(skb->dev);
>  	bool refcounted;
> -	int drop_reason;
>  
>  	drop_reason = SKB_DROP_REASON_NOT_SPECIFIED;
>  

The above 2 chunks look unrelated/unneeded. Since the patch is already
pretty big, I think it would be better to not include them.

Thanks,

Paolo