Re: [PATCH net] gso: fix gso fraglist segmentation after pull from frag_list

[Felix Fietkau <nbd@nbd.name>]

On 26.09.24 09:19, Willem de Bruijn wrote:
> Felix Fietkau wrote:
>> On 25.09.24 22:59, Felix Fietkau wrote:
>> > On 25.09.24 21:09, Willem de Bruijn wrote:
>> >> Felix Fietkau wrote:
>> >>> On 22.09.24 17:03, Willem de Bruijn wrote:
>> >>> > From: Willem de Bruijn <willemb@google.com>
>> >>> > 
>> >>> > Detect gso fraglist skbs with corrupted geometry (see below) and
>> >>> > pass these to skb_segment instead of skb_segment_list, as the first
>> >>> > can segment them correctly.
>> >>> > 
>> >>> > Valid SKB_GSO_FRAGLIST skbs
>> >>> > - consist of two or more segments
>> >>> > - the head_skb holds the protocol headers plus first gso_size
>> >>> > - one or more frag_list skbs hold exactly one segment
>> >>> > - all but the last must be gso_size
>> >>> > 
>> >>> > Optional datapath hooks such as NAT and BPF (bpf_skb_pull_data) can
>> >>> > modify these skbs, breaking these invariants.
>> >>> > 
>> >>> > In extreme cases they pull all data into skb linear. For UDP, this
>> >>> > causes a NULL ptr deref in __udpv4_gso_segment_list_csum at
>> >>> > udp_hdr(seg->next)->dest.
>> >>> > 
>> >>> > Detect invalid geometry due to pull, by checking head_skb size.
>> >>> > Don't just drop, as this may blackhole a destination. Convert to be
>> >>> > able to pass to regular skb_segment.
>> >>> > 
>> >>> > Link: https://lore.kernel.org/netdev/20240428142913.18666-1-shiming.cheng@mediatek.com/
>> >>> > Fixes: 3a1296a38d0c ("net: Support GRO/GSO fraglist chaining.")
>> >>> > Signed-off-by: Willem de Bruijn <willemb@google.com>
>> >>> > Cc: stable@vger.kernel.org
>> >>> > 
>> >>> > ---
>> >>> > diff --git a/net/ipv4/udp_offload.c b/net/ipv4/udp_offload.c
>> >>> > index d842303587af..e457fa9143a6 100644
>> >>> > --- a/net/ipv4/udp_offload.c
>> >>> > +++ b/net/ipv4/udp_offload.c
>> >>> > @@ -296,8 +296,16 @@ struct sk_buff *__udp_gso_segment(struct sk_buff *gso_skb,
>> >>> >   		return NULL;
>> >>> >   	}
>> >>> >   
>> >>> > -	if (skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST)
>> >>> > -		return __udp_gso_segment_list(gso_skb, features, is_ipv6);
>> >>> > +	if (skb_shinfo(gso_skb)->gso_type & SKB_GSO_FRAGLIST) {
>> >>> > +		 /* Detect modified geometry and pass these to skb_segment. */
>> >>> > +		if (skb_pagelen(gso_skb) - sizeof(*uh) == skb_shinfo(gso_skb)->gso_size)
>> >>> > +			return __udp_gso_segment_list(gso_skb, features, is_ipv6);
>> >>> > +
>> >>> > +		 /* Setup csum, as fraglist skips this in udp4_gro_receive. */
>> >>> > +		gso_skb->csum_start = skb_transport_header(gso_skb) - gso_skb->head;
>> >>> > +		gso_skb->csum_offset = offsetof(struct udphdr, check);
>> >>> > +		gso_skb->ip_summed = CHECKSUM_PARTIAL;
>> >>> 
>> >>> I also noticed this uh->check update done by udp4_gro_complete only in 
>> >>> case of non-fraglist GRO:
>> >>> 
>> >>>      if (uh->check)
>> >>>          uh->check = ~udp_v4_check(skb->len - nhoff, iph->saddr,
>> >>>                        iph->daddr, 0);
>> >>> 
>> >>> I didn't see any equivalent in your patch. Is it missing or left out 
>> >>> intentionally?
>> >> 
>> >> Thanks. That was not intentional. I think you're right. Am a bit
>> >> concerned that all this testing did not catch it. Perhaps because
>> >> CHECKSUM_PARTIAL looped to ingress on the same machine is simply
>> >> interpreted as CHECKSUM_UNNECESSARY. Need to look into that.
>> >> 
>> >> If respinning this, I should also change the Fixes to
>> >> 
>> >> Fixes: 9fd1ff5d2ac7 ("udp: Support UDP fraglist GRO/GSO.")
>> >> 
>> >> Analogous to the eventual TCP fix to
>> >> 
>> >> Fixes: bee88cd5bd83 ("net: add support for segmenting TCP fraglist GSO packets")
>> > 
>> > In the mean time, I've been working on the TCP side. I managed to
>> > reproduce the issue on one of my devices by routing traffic from
>> > Ethernet to Wifi using your BPF test program.
>> > 
>> > The following patch makes it work for me for TCP v4. Still need to
>> > test and fix v6.
>> 
>> Actually, here is something even simpler that should work for both v4
>> and v6:
> 
> Makes sense. It does come with higher cost of calling skb_checksum.

But only if there is no checksum offload, right? Because the way I 
implemented it, the lines further below starting with
if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL))
will initialize th->check and set ip_summed to CHECKSUM_PARTIAL
Or am I missing something?

Thanks,

- Felix