marvell/octeontx2: XDP vs skb headroom reservation

marvell/octeontx2: XDP vs skb headroom reservation

[Jakub Sitnicki]

Hi,

I'm auditing calls to skb_metadata_set() in drivers and something seems
off in oceontx2. I may be completely misreading the code. I'm a total
stranger to that driver.

It would seem that there is a difference between the amount of headroom
that the driver reserves for the XDP buffer vs the skb that is built
from on that buffer, while the two should match.

In the receive handler we have:

otx2_rcv_pkt_handler
  otx2_xdp_rcv_pkt_handler if pfvf->xdp_prog
    xdp_prepare_buff(..., OTX2_HEAD_ROOM, ...)
      where OTX2_HEAD_ROOM = OTX2_ALIGN = 128
  skb = napi_get_frags
    napi_alloc_skb
      skb_reserve(NET_SKB_PAD + NET_IP_ALIGN)
        where NET_SKB_PAD = 64, NET_IP_ALIGN = 0 on x86
  napi_gro_frags

There are no other calls to skb_reserve there, so we seem to have a
mismatch. 128B of headroom in XDP buff, and only 64B of headroom in skb
head.

Again, I could be totally wrong about this, I'm not familiar with the
driver, but I thought you might want to double check this.

Thanks,
-jkbs

RE: [EXTERNAL] marvell/octeontx2: XDP vs skb headroom reservation

[Geethasowjanya Akula]

Hi,

>-----Original Message-----
>From: Jakub Sitnicki <jakub@cloudflare.com>
>Sent: Thursday, October 2, 2025 5:16 PM
>To: Sunil Kovvuri Goutham <sgoutham@marvell.com>; Geethasowjanya Akula
><gakula@marvell.com>; Subbaraya Sundeep Bhatta <sbhatta@marvell.com>;
>Hariprasad Kelam <hkelam@marvell.com>; Bharat Bhushan
><bbhushan2@marvell.com>
>Cc: netdev@vger.kernel.org
>Subject: [EXTERNAL] marvell/octeontx2: XDP vs skb headroom reservation
>
>Hi,
>
>I'm auditing calls to skb_metadata_set() in drivers and something seems off in
>oceontx2. I may be completely misreading the code. I'm a total stranger to that
>driver.
>
>It would seem that there is a difference between the amount of headroom that
>the driver reserves for the XDP buffer vs the skb that is built from on that
>buffer, while the two should match.
>
>In the receive handler we have:
>
>otx2_rcv_pkt_handler
>  otx2_xdp_rcv_pkt_handler if pfvf->xdp_prog
>    xdp_prepare_buff(..., OTX2_HEAD_ROOM, ...)
>      where OTX2_HEAD_ROOM = OTX2_ALIGN = 128
>  skb = napi_get_frags
>    napi_alloc_skb
>      skb_reserve(NET_SKB_PAD + NET_IP_ALIGN)
>        where NET_SKB_PAD = 64, NET_IP_ALIGN = 0 on x86
>  napi_gro_frags
>
>There are no other calls to skb_reserve there, so we seem to have a mismatch.
>128B of headroom in XDP buff, and only 64B of headroom in skb head.
>
>Again, I could be totally wrong about this, I'm not familiar with the driver, but I
>thought you might want to double check this.
>

The otx2 netdev driver calculates the receive buffer size to accommodate the full network
packet as follows:
hw.buf_len = mtu + OTX2_ETH_HLEN + OTX2_HW_TIMESTAMP_LEN;
In addition, it reserves OTX2_HEAD_ROOM (128 bytes). The hardware actually copies
the packet starting at (u64*)bufptr + OTX2_HEAD_ROOM. This means the SKB has 
128 bytes of headroom available, which is especially useful in the XDP case for adding any extra headers.

static int otx2_get_rbuf_size(struct otx2_nic *pf, int mtu)
{
        int frame_size;
        int total_size;
        int rbuf_size;

        if (pf->hw.rbuf_len)
                return ALIGN(pf->hw.rbuf_len, OTX2_ALIGN) + OTX2_HEAD_ROOM;
Thanks,
Geetha.
>Thanks,
>-jkbs