From: Florian Westphal <fw@strlen.de>
To: <netfilter-devel@vger.kernel.org>
Cc: Florian Westphal <fw@strlen.de>
Subject: [PATCH nf-next v3 3/6] netfilter: nfnetlink_log: fix up dev_parse_header usage
Date: Wed, 18 Nov 2026 18:39:10 +0100 [thread overview]
Message-ID: <20261118173913.22592-4-fw@strlen.de> (raw)
In-Reply-To: <20261118173913.22592-1-fw@strlen.de>
dev_parse_header() grabs the source hw address from the skb and stores
it in the given buffer.
e.g.: eth_header_parse() is a simple:
int eth_header_parse(const struct sk_buff *skb,
const struct net_device *dev, unsigned char *haddr) {
memcpy(haddr, eth_hdr(skb)->h_source, ETH_ALEN);
return ETH_ALEN;
}
Problem is that the compound structure used by the UABI in nfnetlink_log
and nfnetlink_queue has a fixed 8-byte size.
Enough for ETH_ALEN, but not necessarily enough for other address types.
Move this to a helper and allow sharing it between the copypasted users.
The struct layout of the two container structures have same layout:
add BUILD_BUG_ON checks to assert this.
Due to the nature of the dev_parse_header() helpers:
1. skb MUST have a valid mac header, else we read garbage
2. users MUST NOT use this when ingress device and skb->dev
have diverged (we can't know we would decode something
meaningful anymore).
3. Because container structure only has space for 8 octets,
we have to clamp the result.
Do all of this in the common helper.
The helper also rejects skbs whose network header ended up below the
mac header: encapsulation (ip_encap(), xfrm) can do this and the
unsigned skb_mac_header_len() would then wrap to a huge value,
bypassing the sanity checks. The mac header must also span the
device's full hard_header_len, not just min_header_len:
header_ops->parse() implementations read fixed offsets sized for the
complete header format (ipgre leaves min_header_len at 0 but
ipgre_header_parse() reads offset 12..15 of the mac header).
The header_ops pointer is fetched once with READ_ONCE().
This is mostly cargo culting to pacify AI (sashiko) report pointing at
bond_header_parse() and team_header_parse().
The helper runs in the packet path under RCU only.
team_setup_by_port()/bond_setup_by_slave() allegedly can rewrite
dev->header_ops (possibly to NULL) under RTNL. I don't see how traffic
can observe this: AFAICS those writers run while no port/slave exists
that could feed packets to the upper device (team type change requires
an empty port list; bond dev_close()s itself first).
However, giving logging is slowpath anyway, do the READ_ONCE() thing.
This is an LLM finding, no bug was ever reported about this.
Fixes: 0597f2680d66 ("[NETFILTER]: Add new "nfnetlink_log" userspace packet logging facility")
Fixes: 7af4cc3fa158 ("[NETFILTER]: Add "nfnetlink_queue" netfilter queue handler over nfnetlink")
Signed-off-by: Florian Westphal <fw@strlen.de>
---
v3: add a READ_ONCE, demanded by sashiko, even though dev_parse_header()
did not have that either and I could not find any scenario where packet
path could observe changing header_ops.
sashiko also claimed:
checking mac_len < indev->hard_header_len only on the
top-level indev does not cover bond_header_parse() or team_header_parse(),
which delegate to slave->dev->header_ops->parse() or
port->dev->header_ops->parse() under RCU without checking the lower
device's hard_header_len. If a bond or team transitions from a
zero-length tunnel port (such as ipip, with hard_header_len 0) to ipgre
(with hard_header_len 24), the leaf ipgre_header_parse() callback can
still read past a short MAC header.
But I had no choice but to ignore it.
- No suggestion on how to 'fix' this.
- can dev->hard_header_len *really* drift from dev->header_ops
while other CPU is processing packets?
- If yes, how is one supposed to detect that?
- How would that cause oob read?
I could not find evidence on the type-change claim.
team_dev_type_check_change() checks dev->type == port->type.
Even with LLM assistance I could not find any proof of such
a drift.
Moreover, nfnetlink_put_hwaddr() is already shit, with truckload of
'is-this-NULL-is-that-NULL-could-this-offset-be-valid' crap.
I will not tell an LLM to make this even more unreadable
garbage without an exact sequence of events that will result in a
bogus access and at least one poc-style solution.
Existing code did not do anything special and its
been in use for 20+ years.
net/netfilter/nfnetlink_log.c | 18 +---
net/netfilter/nfnetlink_log_common.h | 121 +++++++++++++++++++++++++++
net/netfilter/nfnetlink_queue.c | 18 +---
3 files changed, 129 insertions(+), 28 deletions(-)
create mode 100644 net/netfilter/nfnetlink_log_common.h
diff --git a/net/netfilter/nfnetlink_log.c b/net/netfilter/nfnetlink_log.c
index 90e87c34df85..37f3e7b6b538 100644
--- a/net/netfilter/nfnetlink_log.c
+++ b/net/netfilter/nfnetlink_log.c
@@ -47,6 +47,8 @@
#include <net/netfilter/nf_conntrack.h>
#endif
+#include "nfnetlink_log_common.h"
+
#define NFULNL_COPY_DISABLED 0xff
#define NFULNL_NLBUFSIZ_DEFAULT NLMSG_GOODSIZE
#define NFULNL_TIMEOUT_DEFAULT 100 /* every second */
@@ -587,20 +589,8 @@ __build_packet_message(struct nfnl_log_net *log,
nla_put_be32(inst->skb, NFULA_MARK, htonl(skb->mark)))
goto nla_put_failure;
- if (indev && skb->dev &&
- skb_mac_header_was_set(skb) &&
- skb_mac_header_len(skb) != 0) {
- struct nfulnl_msg_packet_hw phw;
- int len;
-
- memset(&phw, 0, sizeof(phw));
- len = dev_parse_header(skb, phw.hw_addr);
- if (len > 0) {
- phw.hw_addrlen = htons(len);
- if (nla_put(inst->skb, NFULA_HWADDR, sizeof(phw), &phw))
- goto nla_put_failure;
- }
- }
+ if (nfnetlink_put_hwaddr(inst->skb, skb, indev, NFULA_HWADDR))
+ goto nla_put_failure;
if (indev && skb_mac_header_was_set(skb)) {
if (nla_put_be16(inst->skb, NFULA_HWTYPE, htons(skb->dev->type)) ||
diff --git a/net/netfilter/nfnetlink_log_common.h b/net/netfilter/nfnetlink_log_common.h
new file mode 100644
index 000000000000..c6dec09d32a9
--- /dev/null
+++ b/net/netfilter/nfnetlink_log_common.h
@@ -0,0 +1,121 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _NF_NETLINK_LOG_COMMON_H
+#define _NF_NETLINK_LOG_COMMON_H
+
+#include <linux/netdevice.h>
+#include <linux/skbuff.h>
+#include <net/netlink.h>
+#include <linux/netfilter/nfnetlink_queue.h>
+#include <linux/netfilter/nfnetlink_log.h>
+
+/* nfnetlink_log and nfnetlink_queue suffer from copypaste; avoid adding
+ * more and place shared code here.
+ */
+static __always_inline void assert_nfqnl_nfunl_hdrs(void)
+{
+ struct nfulnl_msg_packet_hw phwu;
+ struct nfqnl_msg_packet_hw phwq;
+
+ BUILD_BUG_ON(sizeof(phwu) != sizeof(phwq));
+
+#define X(memb) offsetof(struct nfulnl_msg_packet_hw, memb) != \
+ offsetof(struct nfqnl_msg_packet_hw, memb)
+
+ BUILD_BUG_ON(X(hw_addrlen));
+ BUILD_BUG_ON(X(hw_addr));
+#undef X
+#define X(memb) sizeof(phwu.memb) != sizeof(phwq.memb)
+ BUILD_BUG_ON(X(hw_addrlen));
+#undef X
+}
+
+/**
+ * nfnetlink_put_hwaddr - Serialize L2 hardware address into netlink attr
+ *
+ * @nlskb: netlink message to append the attribute to.
+ * @skb: the packet whose hardware address is being serialized.
+ * @indev: the input device that received @skb.
+ * @attr: netlink attribute type identifier.
+ *
+ * Extracts the MAC header from @skb using @indev's header_ops and appends
+ * it as a netlink attribute of type @attr to @nlskb.
+ *
+ * UAPI NOTE: The attribute is serialized in a legacy container structure
+ * (nfulnl_msg_packet_hw / nfqnl_msg_packet_hw) rather than a plain
+ * nla_put(hwaddr, len).
+ * The hw_addrlen member in that structure is redundant -- userspace knows
+ * the size from the netlink attribute length -- but is retained for ABI
+ * compatibility. nfnetlink_queue duplicates this layout; both structs are
+ * layout-identical, so this one helper serves both.
+ *
+ * The maximum address length is 8 octets, even though netlink itself
+ * supports longer addresses. This is a limitation of the
+ * compound-structure ABI (hw_addr[8] member).
+ *
+ * This helper checks:
+ * - @indev is non-NULL and implements header_ops->parse.
+ * - network_header is not below mac_header (encapsulation), so
+ * skb_mac_header_len() does not wrap.
+ * - L2 header is present and at least @indev's hard_header_len, so
+ * header_ops->parse does not read past a short or missing header.
+ *
+ * @indev, not @skb->dev, is used: @skb->dev may have been changed since
+ * the packet was received, so it is not necessarily the device whose
+ * header_ops can parse the L2 header.
+ *
+ * Return: 0 on success or if the attribute is skipped (no parser, no or
+ * short L2 header, or empty address), -EMSGSIZE if the tailroom of the
+ * skb is insufficient to store the attribute header and payload.
+ */
+static inline int
+nfnetlink_put_hwaddr(struct sk_buff *nlskb,
+ const struct sk_buff *skb,
+ const struct net_device *indev,
+ u16 attr)
+{
+ const struct header_ops *header_ops;
+ char hw_addr[MAX_ADDR_LEN];
+ unsigned int mac_len;
+ int len;
+
+ if (!indev || !skb_mac_header_was_set(skb))
+ return 0;
+
+ header_ops = READ_ONCE(indev->header_ops);
+ if (!header_ops || !header_ops->parse)
+ return 0;
+
+ /* encap (ip_encap, xfrm) can set network_header < mac_header:
+ * skb_mac_header_len() would wrap and bypass the checks below.
+ */
+ if (skb->network_header < skb->mac_header)
+ return 0;
+
+ mac_len = skb_mac_header_len(skb);
+ /* parse() reads fixed offsets sized for the full header
+ * (ipgre: min_header_len 0, reads offset 12..15).
+ */
+ if (mac_len == 0 || mac_len < indev->hard_header_len)
+ return 0;
+
+ len = header_ops->parse(skb, indev, hw_addr);
+ if (len > 0) {
+ struct nfulnl_msg_packet_hw phw;
+ const int max_hwlen = (int)sizeof(phw.hw_addr);
+
+ assert_nfqnl_nfunl_hdrs();
+
+ /* UAPI limitation, must clamp. */
+ if (len > max_hwlen)
+ len = max_hwlen;
+
+ memset(&phw, 0, sizeof(phw));
+ memcpy(phw.hw_addr, hw_addr, len);
+ phw.hw_addrlen = htons(len);
+
+ return nla_put(nlskb, attr, sizeof(phw), &phw);
+ }
+
+ return 0;
+}
+#endif
diff --git a/net/netfilter/nfnetlink_queue.c b/net/netfilter/nfnetlink_queue.c
index 862d2a762106..37f8953b7c49 100644
--- a/net/netfilter/nfnetlink_queue.c
+++ b/net/netfilter/nfnetlink_queue.c
@@ -57,6 +57,8 @@
#include <net/netfilter/nf_conntrack.h>
#endif
+#include "nfnetlink_log_common.h"
+
#define NFQNL_QMAX_DEFAULT 1024
#define NFQNL_HASH_MIN 8
#define NFQNL_HASH_MAX 32768
@@ -917,20 +919,8 @@ nfqnl_build_packet_message(struct net *net, struct nfqnl_instance *queue,
nla_put_be32(skb, NFQA_PRIORITY, htonl(entskb->priority)))
goto nla_put_failure;
- if (indev && entskb->dev &&
- skb_mac_header_was_set(entskb) &&
- skb_mac_header_len(entskb) != 0) {
- struct nfqnl_msg_packet_hw phw;
- int len;
-
- memset(&phw, 0, sizeof(phw));
- len = dev_parse_header(entskb, phw.hw_addr);
- if (len) {
- phw.hw_addrlen = htons(len);
- if (nla_put(skb, NFQA_HWADDR, sizeof(phw), &phw))
- goto nla_put_failure;
- }
- }
+ if (nfnetlink_put_hwaddr(skb, entskb, indev, NFQA_HWADDR))
+ goto nla_put_failure;
if (nfqnl_put_bridge(entry, skb) < 0)
goto nla_put_failure;
--
2.56.1
next prev parent reply other threads:[~2026-11-18 17:39 UTC|newest]
Thread overview: 7+ messages / expand[flat|nested] mbox.gz Atom feed top
2026-11-18 17:39 [PATCH nf-next v3 0/6] netfilter: logging cleanups Florian Westphal
2026-11-18 17:39 ` [PATCH nf-next v3 1/6] netfilter: nf_log: remove dead ARPHRD_SIT handling from MAC/HW header logging Florian Westphal
2026-11-18 17:39 ` [PATCH nf-next v3 2/6] netfilter: nf_log_syslog: fix MAC header dump to use real header length Florian Westphal
2026-11-18 17:39 ` Florian Westphal [this message]
2026-11-18 17:39 ` [PATCH nf-next v3 4/6] netfilter: nfnetlink_log: use real mac header length for hwheader Florian Westphal
2026-11-18 17:39 ` [PATCH nf-next v3 5/6] netfilter: nfnetlink: don't log hwaddr/hwheader on output hooks Florian Westphal
2026-11-18 17:39 ` [PATCH nf-next v3 6/6] netfilter: nfnetlink_log: only add timestamp on the input path Florian Westphal
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20261118173913.22592-4-fw@strlen.de \
--to=fw@strlen.de \
--cc=netfilter-devel@vger.kernel.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.