net: xdp: don't assume an Ethernet header in generic XDP
Subsystems (3)
PATCH 1/1 net: xdp: don't assume an Ethernet header in generic XDP
Kenneth Lee reported a bug [1]:
skb_vlan_push got skb with skb->data not at mac header (offset 14)
WARNING: net/core/skbuff.c:6474 at skb_vlan_push+0x5ec/0x8a0 ____bpf_skb_vlan_push net/core/filter.c:3239 [inline] bpf_skb_vlan_push+0x216/0x8b0 net/core/filter.c:3229 bpf_prog_run_data_pointers+0x17c/0x240 include/linux/filter.h:917 tcf_bpf_act+0x31c/0x940 net/sched/act_bpf.c:50 tcf_action_exec+0x2c1/0x720 net/sched/act_api.c:1147 tcf_classify+0x6e4/0x1080 net/sched/cls_api.c:1860 tc_run+0x31c/0x5d0 net/core/dev.c:4411 sch_handle_ingress net/core/dev.c:4486 [inline] __netif_receive_skb_core+0x141b/0x2ec0 net/core/dev.c:6054
bpf_prog_run_generic_xdp() just assumes xdp->data is an Ethernet header:
eth = (struct ethhdr *)xdp->data; orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); orig_eth_type = eth->h_proto;
But an L3 device (ARPHRD_NONE, ARPHRD_TUNNEL, ...) has no L2 header at all - its mac_len is 0 and xdp->data is really the L3 header, so those orig_* values are garbage.
When that garbage comparison says the header changed, we then do:
__skb_push(skb, ETH_HLEN); skb->pkt_type = PACKET_HOST; skb->protocol = eth_type_trans(skb, skb->dev);
eth_type_trans() resets mac_header and pulls ETH_HLEN back, so skb->data does not move at all, but we now have mac_len == 0 while skb_mac_header(skb) == skb->data - ETH_HLEN.
tc ingress relies on mac_len == skb->data - skb_mac_header(skb) to get skb->data onto the mac header, so __skb_push(skb, skb->mac_len) pushes nothing, the program runs with skb->data ETH_HLEN past the mac header, and bpf_skb_vlan_push() hits the WARN_ONCE above. skb->protocol also ends up parsed from uninitialised headroom.
So skip the Ethernet part entirely when the skb has no ETH_HLEN sized L2 header. The program still runs and its action is still returned.
Refusing to attach XDP to L3 devices instead would not help: mac_len is a property of the skb, not of the netdev, so an ARPHRD_ETHER device can see mac_len == 0 too.
[1]: https://lore.kernel.org/bpf/[email protected]/
--- net/core/dev.c | 42 ++++++++++++++++++++++++++++-------------- 1 file changed, 28 insertions(+), 14 deletions(-)
net/core/dev.c +28 -14
@@ -5489,11 +5489,12 @@ u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, {void *orig_data, *orig_data_end, *hard_start;struct netdev_rx_queue *rxqueue; -bool orig_bcast, orig_host; +bool orig_bcast = false, orig_host = false; +__be16 orig_eth_type =0 ;u32 mac_len, frame_sz; -__be16 orig_eth_type;struct ethhdr *eth;u32 metalen, act; +bool has_eth;int off;/* The XDP program wants to see the packet starting at the MAC
@@ -5519,10 +5520,21 @@ u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, orig_data_end = xdp->data_end; orig_data = xdp->data; - eth = (struct ethhdr *)xdp->data; - orig_host =ether_addr_equal_64bits (eth->h_dest, skb->dev->dev_addr); - orig_bcast =is_multicast_ether_addr_64bits (eth->h_dest); - orig_eth_type = eth->h_proto; + +/* xdp->data only points at an Ethernet header if this skb actually +* carries one. Devices with a different link layer (mac_len == 0 +* for ARPHRD_NONE/TUNNEL/RAWIP/PPP/..., IPOIB_ENCAP_LEN for IPoIB) +* have nothing to inspect here, and must not have skb->mac_header +* relocated by the ETH_HLEN fixup below. +*/ + has_eth = mac_len ==ETH_HLEN ; +if (has_eth) { + eth = (struct ethhdr *)xdp->data; + orig_host =ether_addr_equal_64bits (eth->h_dest, + skb->dev->dev_addr); + orig_bcast =is_multicast_ether_addr_64bits (eth->h_dest); + orig_eth_type = eth->h_proto; + } act =bpf_prog_run_xdp (xdp_prog, xdp);
@@ -5554,14 +5566,16 @@ u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, skb->data_len =0 ;/* check if XDP changed eth hdr such SKB needs update */ - eth = (struct ethhdr *)xdp->data; -if ((orig_eth_type != eth->h_proto) || - (orig_host !=ether_addr_equal_64bits (eth->h_dest, - skb->dev->dev_addr)) || - (orig_bcast !=is_multicast_ether_addr_64bits (eth->h_dest))) { -__skb_push (skb,ETH_HLEN ); - skb->pkt_type =PACKET_HOST ; - skb->protocol =eth_type_trans (skb, skb->dev); +if (has_eth) { + eth = (struct ethhdr *)xdp->data; +if ((orig_eth_type != eth->h_proto) || + (orig_host !=ether_addr_equal_64bits (eth->h_dest, + skb->dev->dev_addr)) || + (orig_bcast !=is_multicast_ether_addr_64bits (eth->h_dest))) { +__skb_push (skb,ETH_HLEN ); + skb->pkt_type =PACKET_HOST ; + skb->protocol =eth_type_trans (skb, skb->dev); + } }/* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull
Signature · 1 line
2.43.0
[email protected]
AI reviewThank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:
Pre-existing issues:
- [High] Incorrect SKB length and tail pointer calculation after generic XDP tail adjustment. --
--- Patch [1]: [PATCH net-next] net: xdp: don't assume an Ethernet header in generic XDP --- commit 3332d9a3f6b94e19ff995230c0e36bef6be765df
Author: Jiayuan Chen <[email protected]>
net: xdp: don't assume an Ethernet header in generic XDP
This patch fixes generic XDP for L3 devices (like tun/tap or ARPHRD_NONE) that do not have an Ethernet header. It skips Ethernet-specific checks and header adjustments when the SKB does not carry an ETH_HLEN sized L2 header.
Quoted 16 lines
net/core/dev.c +8
-0@@ -5519,10 +5520,21 @@ u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, orig_data_end = xdp->data_end; orig_data = xdp->data; + +/* xdp->data only points at an Ethernet header if this skb actually +* carries one. Devices with a different link layer (mac_len == 0 +* for ARPHRD_NONE/TUNNEL/RAWIP/PPP/..., IPOIB_ENCAP_LEN for IPoIB) +* have nothing to inspect here, and must not have skb->mac_header +* relocated by the ETH_HLEN fixup below. +*/ + has_eth = mac_len ==ETH_HLEN ;
net/core/dev.c u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, · line 5540
Quoted 10 lines
net/core/dev.c +7
-0+if (has_eth) { + eth = (struct ethhdr *)xdp->data; + orig_host =ether_addr_equal_64bits (eth->h_dest, + skb->dev->dev_addr); + orig_bcast =is_multicast_ether_addr_64bits (eth->h_dest); + orig_eth_type = eth->h_proto; + } act =bpf_prog_run_xdp (xdp_prog, xdp);
[Severity: High]
This isn't a bug introduced by this patch, but looking at how xdp->data and skb->data relate to the MAC header, could there be a tail pointer calculation issue when bpf_xdp_adjust_tail() is called?
Before bpf_prog_run_generic_xdp() executes, the MAC header is stripped via __skb_pull(skb, mac_len), leaving skb->data pointing at the network layer. However, xdp->data is initialized to point at the MAC header.
If the XDP program adjusts the tail, the new tail pointer is updated further down in bpf_prog_run_generic_xdp():
/* check if bpf_xdp_adjust_tail was used */
off = xdp->data_end - orig_data_end;
if (off != 0) {
skb_set_tail_pointer(skb, xdp->data_end - xdp->data);
skb->len += off;
}Because xdp->data_end - xdp->data includes the MAC header length, but skb_set_tail_pointer() applies this offset relative to skb->data (which is already at the network header), doesn't this push the tail pointer mac_len bytes too far past the actual end of the packet?
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
Jakub Kicinski
maintainerQuoted 3 lines · On Thu, 13 Aug 2026 19:32:24 +0800 Jiayuan Chen wrote:
Refusing to attach XDP to L3 devices instead would not help:
mac_lenis a property of the skb, not of the netdev, so anARPHRD_ETHERdevice can seemac_len== 0 too.
Let's try. There is no native XDP on any non-ether device, making generic xdp work on those is silly. Just attach the BPF in TC.
AI says that one example for mac_len=0 packet on ether device is IP mcast loopback, which should not go via generic XDP in the first place.
Quoted 1 line
Fixes 22b6034323fd ("net, xdp: Update pkt_type if generic XDP changes unicast MAC")
Please don't send fixes to net-next :|
Signature · 1 line
pw-bot: cr
From: Jiayuan Chen <[email protected]>
Date: Thu, 13 Aug 2026 19:32:24 +0800
Quoted 14 lines
Kenneth Lee reported a bug [1]:
skb_vlan_pushgot skb withskb->datanot at mac header (offset 14)WARNING: net/core/skbuff.c:6474 at skb_vlan_push+0x5ec/0x8a0 ____bpf_skb_vlan_push net/core/filter.c:3239 [inline] bpf_skb_vlan_push+0x216/0x8b0 net/core/filter.c:3229 bpf_prog_run_data_pointers+0x17c/0x240 include/linux/filter.h:917 tcf_bpf_act+0x31c/0x940 net/sched/act_bpf.c:50 tcf_action_exec+0x2c1/0x720 net/sched/act_api.c:1147 tcf_classify+0x6e4/0x1080 net/sched/cls_api.c:1860 tc_run+0x31c/0x5d0 net/core/dev.c:4411 sch_handle_ingress net/core/dev.c:4486 [inline] __netif_receive_skb_core+0x141b/0x2ec0 net/core/dev.c:6054
Quoted 1 line
bpf_prog_run_generic_xdp()just assumesxdp->datais an Ethernet header:
Most XDP programs expect Ethernet header at the beginning of a frame. Unless you write a custom one which doesn't.
But then you may face that XDP_TX, XDP_REDIRECT won't work properly -- cpumap Rx, each .ndo_xdp_xmit() implementation -- all expect Ethernet header at the beginning.
Thanks,
Olek
Quoted 17 lines · Alexander Lobakin
From: Jiayuan Chen <[email protected]>
Date: Thu, 13 Aug 2026 19:32:24 +0800Quoted 14 lines
Kenneth Lee reported a bug [1]:
skb_vlan_pushgot skb withskb->datanot at mac header (offset 14)WARNING: net/core/skbuff.c:6474 at skb_vlan_push+0x5ec/0x8a0 ____bpf_skb_vlan_push net/core/filter.c:3239 [inline] bpf_skb_vlan_push+0x216/0x8b0 net/core/filter.c:3229 bpf_prog_run_data_pointers+0x17c/0x240 include/linux/filter.h:917 tcf_bpf_act+0x31c/0x940 net/sched/act_bpf.c:50 tcf_action_exec+0x2c1/0x720 net/sched/act_api.c:1147 tcf_classify+0x6e4/0x1080 net/sched/cls_api.c:1860 tc_run+0x31c/0x5d0 net/core/dev.c:4411 sch_handle_ingress net/core/dev.c:4486 [inline] __netif_receive_skb_core+0x141b/0x2ec0 net/core/dev.c:6054
Quoted 6 lines
Quoted 1 line
bpf_prog_run_generic_xdp()just assumesxdp->datais an Ethernet header:Most XDP programs expect Ethernet header at the beginning of a frame. Unless you write a custom one which doesn't.
But then you may face thatXDP_TX,XDP_REDIRECTwon't work properly -- cpumap Rx, each .ndo_xdp_xmit() implementation -- all expect Ethernet header at the beginning.
Yeah, I don't think we should start messing with the "XDP is Ethernet only" assumption at this point...
-Toke