In the Linux kernel, the following vulnerability has been resolved:
flowdissector: check device type before reading ETHADDRS
__skbflowdissect() unconditionally reads 12 bytes from ethhdr(skb) when FLOWDISSECTORKEYETHADDRS is requested. This assumes the skb has a valid Ethernet header at macheader, which is not always the case.
The problem can be triggered by: 1. Creating a TUN device in L3 mode (IFFTUN, hardheaderlen=0) 2. Attaching a multiq qdisc with a flower filter matching on ethsrc 3. Sending a packet through AF_PACKET
Since TUN in L3 mode has no link-layer header, macheader points to the L3 data area. The flow dissector reads 12 bytes of uninitialized skb memory, which then propagates through flsetmaskedkey() and is used as a rhashtable lookup key in _fllookup(), as reported by KMSAN.
Rejecting the filter in the control path (at tc filter add time) is not feasible because TC filter blocks can be shared between arbitrary devices -- a filter installed on an Ethernet device may later classify packets on a headerless device through a shared block. The device association is not fixed at filter creation time.
Fix this by gating the memcpy on dev->type == ARPHRDETHER, which ensures only true Ethernet-framed packets have their addresses read. This is more precise than the previous hardheaderlen >= 12 check, which would incorrectly pass for non-Ethernet link types like IPoIB (ARPHRDINFINIBAND, hardheaderlen=24) and FDDI (hardheaderlen=21) whose L2 headers are not in Ethernet format. Additionally check skbmacheaderwasset() to guard against the pathological case where macheader is the unset sentinel (~0U), which would cause ethhdr() to return a wild pointer.
For the act_mirred redirect case (Ethernet packet redirected to a non-Ethernet device sharing a TC block), zeroing the key is the correct behavior: the packet is now being classified on the target device, where Ethernet address matching is not semantically meaningful.
Note: on non-Ethernet devices, the zeroed key will match a filter configured with all-zero MAC addresses. This is an improvement over the previous behavior where uninitialized memory could randomly match any filter.
{
"cna_assigner": "Linux",
"osv_generated_from": "https://github.com/CVEProject/cvelistV5/tree/main/cves/2026/72xxx/CVE-2026-72444.json"
}