Re: [PATCH net-next v3] net: pppoe: implement GRO/GSO support

[Richard Gobert <richardbgobert@gmail.com>]

Felix Fietkau wrote:
> On 14.08.25 16:30, Richard Gobert wrote:
>> Felix Fietkau wrote:
>>> Only handles packets where the pppoe header length field matches the exact
>>> packet length. Significantly improves rx throughput.
>>>
>>> When running NAT traffic through a MediaTek MT7621 devices from a host
>>> behind PPPoE to a host directly connected via ethernet, the TCP throughput
>>> that the device is able to handle improves from ~130 Mbit/s to ~630 Mbit/s,
>>> using fraglist GRO.
>>>
>>> Signed-off-by: Felix Fietkau <nbd@nbd.name>
>>> ---
>>> v2: fix compile error
>>> v3:
>>>   - increase priority value
>>>   - implement GSO support
>>>   - use INDIRECT_CALL_INET
>>>   - update pppoe length field
>>>   - remove unnecessary network_offsets update
>>>
>>>  drivers/net/ppp/pppoe.c | 160 +++++++++++++++++++++++++++++++++++++++-
>>>  net/ipv4/af_inet.c      |   2 +
>>>  net/ipv6/ip6_offload.c  |   2 +
>>>  3 files changed, 163 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/net/ppp/pppoe.c b/drivers/net/ppp/pppoe.c
>>> index 410effa42ade..a8d8eb870bce 100644
>>> --- a/drivers/net/ppp/pppoe.c
>>> +++ b/drivers/net/ppp/pppoe.c
>>> +compare_pppoe_header(struct pppoe_hdr *phdr, struct pppoe_hdr *phdr2)
>>> +{
>>> +    return (__force __u16)((phdr->sid ^ phdr2->sid) |
>>> +                   (phdr->tag[0].tag_type ^ phdr2->tag[0].tag_type));
>>> +}
>>> +
>>> +static __be16 pppoe_hdr_proto(struct pppoe_hdr *phdr)
>>> +{
>>> +    switch (phdr->tag[0].tag_type) {
>>> +    case cpu_to_be16(PPP_IP):
>>> +        return cpu_to_be16(ETH_P_IP);
>>> +    case cpu_to_be16(PPP_IPV6):
>>> +        return cpu_to_be16(ETH_P_IPV6);
>>> +    default:
>>> +        return 0;
>>> +    }
>>> +
>>> +}
>>> +
>>> +static struct sk_buff *pppoe_gro_receive(struct list_head *head,
>>> +                     struct sk_buff *skb)
>>> +{
>>> +    const struct packet_offload *ptype;
>>> +    unsigned int hlen, off_pppoe;
>>> +    struct sk_buff *pp = NULL;
>>> +    struct pppoe_hdr *phdr;
>>> +    struct sk_buff *p;
>>> +    int flush = 1;
>>> +    __be16 type;
>>> +
>>> +    off_pppoe = skb_gro_offset(skb);
>>> +    hlen = off_pppoe + sizeof(*phdr);
>>> +    phdr = skb_gro_header(skb, hlen + 2, off_pppoe);
>>> +    if (unlikely(!phdr))
>>> +        goto out;
>>> +
>>> +    /* ignore packets with padding or invalid length */
>>> +    if (skb_gro_len(skb) != be16_to_cpu(phdr->length) + hlen)
>>> +        goto out;
>>> +
>>> +    type = pppoe_hdr_proto(phdr);
>>> +    if (!type)
>>> +        goto out;
>>> +
>>> +    ptype = gro_find_receive_by_type(type);
>>> +    if (!ptype)
>>> +        goto out;
>>> +
>>> +    flush = 0;
>>> +
>>> +    list_for_each_entry(p, head, list) {
>>> +        struct pppoe_hdr *phdr2;
>>> +
>>> +        if (!NAPI_GRO_CB(p)->same_flow)
>>> +            continue;
>>> +
>>> +        phdr2 = (struct pppoe_hdr *)(p->data + off_pppoe);
>>> +        if (compare_pppoe_header(phdr, phdr2))
>>> +            NAPI_GRO_CB(p)->same_flow = 0;
>>> +    }
>>> +
>>> +    skb_gro_pull(skb, sizeof(*phdr) + 2);
>>> +    skb_gro_postpull_rcsum(skb, phdr, sizeof(*phdr) + 2);
>>> +
>>> +    pp = indirect_call_gro_receive_inet(ptype->callbacks.gro_receive,
>>> +                        ipv6_gro_receive, inet_gro_receive,
>>> +                        head, skb);
>>> +
>>> +out:
>>> +    skb_gro_flush_final(skb, pp, flush);
>>> +
>>> +    return pp;
>>> +}
>>> +
>>> +static int pppoe_gro_complete(struct sk_buff *skb, int nhoff)
>>> +{
>>> +    struct pppoe_hdr *phdr = (struct pppoe_hdr *)(skb->data + nhoff);
>>> +    __be16 type = pppoe_hdr_proto(phdr);
>>> +    struct packet_offload *ptype;
>>> +    int len, err;
>>> +
>>> +    ptype = gro_find_complete_by_type(type);
>>> +    if (!ptype)
>>> +        return -ENOENT;
>>> +
>>> +    err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete,
>>> +                 ipv6_gro_complete, inet_gro_complete,
>>> +                 skb, nhoff + sizeof(*phdr) + 2);
>>> +    if (err)
>>> +        return err;
>>> +
>>> +    len = skb->len - (nhoff + sizeof(*phdr));
>>> +    phdr->length = cpu_to_be16(len);
>>> +
>>> +    return 0;
>>> +}
>>> +
>>> +static struct sk_buff *pppoe_gso_segment(struct sk_buff *skb,
>>> +                     netdev_features_t features)
>>> +{
>>
>> I don't think this will be called for PPPoE over GRE packets,
>> since gre_gso_segment skips everything up to the network header.
> 
> What's a good solution to this issue? Use the outer network header instead of the inner one when the protocol is PPPoE?
> 
> - Felix

I don't really have a good solution for this. You could explicitly check
if the protocol is PPPoE in gre_gso_segment, but that wouldn't be very
elegant or future-proof.
 
I think setting skb->inner_network_header in pppoe_gro_complete
(while not resetting it in inet_gro_complete) wouldn't work since other
functions assume that skb->inner_network_header is an IP header.