net: bridge: fix skb leak of reinjected link-local frames
Subsystems (3)
PATCH 1/1 net: bridge: fix skb leak of reinjected link-local frames
br_handle_frame() passes the link-local frames it neither forwards nor delivers (01:80:c2:00:00:02 to 0d and 0f, unless set in the group forward mask) through NF_HOOK() at NF_BR_LOCAL_IN. Since commit 3b2e2904deb3 ("net: bridge: fix per-port af_packet sockets") the okfn only learns the source and returns 1, which becomes RX_HANDLER_PASS.
That commit sent the nf_queue case to RX_HANDLER_CONSUMED, which is right for the synchronous return, but the frame comes back: on reinjection nf_reinject() calls the okfn and ignores its return value, because every other okfn consumes the skb. This one no longer does, so nothing frees the frame. NFQA_CFG_F_FAIL_OPEN leaks it the same way with no verdict at all, when the queue or the listener's socket is full.
Each frame leaks an skb, and one sent from a local socket keeps that socket charged, so neither it nor the network namespace holding it is freed. Bridge, rule and queue need only CAP_NET_ADMIN in a netns, so with user namespaces an unprivileged user can exhaust kernel memory.
nf_reinject() cannot free the skb on a 1 from the okfn, because other okfns return NET_RX_DROP, which is 1, after freeing it.
Call nf_hook() instead of NF_HOOK() and do the learning and the RX_HANDLER_PASS return in br_handle_frame() on 1, as the 0x00 and 0x0e cases already do and as vrf_rcv_nfhook() does. The okfn then runs only for a reinjected frame, which br_handle_frame() has already reported consumed and cannot pass on the port, so it learns from the frame and frees it. These frames were not delivered before this change either.
Also set *pskb before the new RX_HANDLER_PASS, as the other such returns do; skb_share_check() may have replaced the skb, and the commit above added that return without the assignment.
Discovered by an AI code review agent and reproduced on unmodified mainline under KASAN: an unprivileged user in a new user and network namespace bridged a veth, queued link-local frames to an NFQUEUE listener that accepted every one, and leaked an skb per frame. With this patch the reinjected frames are freed and nothing leaks.
---
Tested in a KVM guest on mainline b3dfbde62cf1 with KASAN and PROVE_LOCKING, running as nobody under "unshare -Urn". No kernel modification, debug patch or injected delay was needed to reproduce.
Before: 192000 bytes charged on the sending socket, unchanged 2 s later. After: 0 bytes, no KASAN report or lockdep warning, and br_input.o builds clean at W=1.
Not tested: NF_STOP, NFQA_CFG_F_FAIL_OPEN, real EAPOL traffic, the netfilter and bridge selftests, CONFIG_NETFILTER=n.
Applies to v7.2-rc5, v6.12.117, next-20261109 and 7e1392a42ea2.
The Fixes commit was checked against torvalds/linux: it adds both the "return 1" okfn and the RX_HANDLER_PASS return without the *pskb assignment.
net/bridge/br_input.c | 18 +++++++++++++----- 1 file changed, 13 insertions(+), 5 deletions(-)
net/bridge/br_input.c +13 -5
@@ -257,13 +257,18 @@ static void __br_handle_local_finish(struct sk_buff *skb)br_fdb_update (p->br, p, vlan,eth_hdr (skb)->h_source,0 ); } -/* note: already called with rcu_read_lock */ +/* okfn of the link-local NF_BR_LOCAL_IN hook in br_handle_frame(). It is +* only called when a queued frame is reinjected, after br_handle_frame() +* returned RX_HANDLER_CONSUMED for it, so the frame can no longer be +* passed up on the port and has to be freed here. +* note: already called with rcu_read_lock +*/ static int br_handle_local_finish (struct net *net,struct sock *sk,struct sk_buff *skb) {__br_handle_local_finish (skb); +kfree_skb (skb); -/* return 1 to signal the okfn() was called so it's ok to use the skb */ -return 1 ; +return 0 ; }static int nf_hook_bridge_pre (struct sk_buff *skb,struct sk_buff **pskb)
@@ -424,11 +429,14 @@ static rx_handler_result_t br_handle_frame(struct sk_buff **pskb)/* The else clause should be hit when nf_hook(): * - returns < 0 (drop/error) * - returns = 0 (stolen/nf_queue) -* Thus return 1 from the okfn() to signal the skb is ok to pass +* On 1 the frame is passed on here. The okfn() only runs if a +* queued frame is reinjected later, and must consume it then. */ -if (NF_HOOK (NFPROTO_BRIDGE ,NF_BR_LOCAL_IN , +if (nf_hook (NFPROTO_BRIDGE ,NF_BR_LOCAL_IN ,dev_net (skb->dev),NULL , skb, skb->dev,NULL , br_handle_local_finish) ==1 ) { + *pskb = skb; +__br_handle_local_finish (skb);return RX_HANDLER_PASS ; }else {return RX_HANDLER_CONSUMED ;
Quoted 5 lines · Daniel Wade
br_handle_frame()passes the link-local frames it neither forwards nor delivers (01:80:c2:00:00:02 to 0d and 0f, unless set in the group forward mask) throughNF_HOOK()atNF_BR_LOCAL_IN. Since commit3b2e2904deb3("net: bridge: fix per-portaf_packetsockets") the okfn only learns the source and returns 1, which becomesRX_HANDLER_PASS.
Before that commit the okfn called br_pass_frame_up() and the frame was consumed on reinjection. The leak starts at 3b2e2904deb3.
The queue paths hold up too. Both fail-open exits in __nfqnl_enqueue_packet() reinject NF_ACCEPT inside the enqueue call, and nf_queue() still reports the frame as queued. nf_hook() returns 0, br_handle_frame() answers RX_HANDLER_CONSUMED, and the single free is the kfree_skb() in the new okfn. QUEUE_BYPASS just resumes the hook walk. If the remaining hooks accept, nf_hook() returns 1 and the frame is passed up from br_handle_frame() itself. The branch now matches the 0x00 and 0x0e cases above it, including the *pskb assignment after skb_share_check().
No other bridge okfn leaves the skb with the caller. Delivered frames reach netif_receive_skb() through br_pass_frame_up() and forwarding goes through br_forward_finish(). The link-local hook was the exception.
Quoted 93 lines · Daniel Wade
br_handle_frame()passes the link-local frames it neither forwards nor delivers (01:80:c2:00:00:02 to 0d and 0f, unless set in the group forward mask) throughNF_HOOK()atNF_BR_LOCAL_IN. Since commit3b2e2904deb3("net: bridge: fix per-portaf_packetsockets") the okfn only learns the source and returns 1, which becomesRX_HANDLER_PASS.That commit sent the
nf_queuecase toRX_HANDLER_CONSUMED, which is right for the synchronous return, but the frame comes back: on reinjectionnf_reinject()calls the okfn and ignores its return value, because every other okfn consumes the skb. This one no longer does, so nothing frees the frame.NFQA_CFG_F_FAIL_OPENleaks it the same way with no verdict at all, when the queue or the listener's socket is full.Each frame leaks an skb, and one sent from a local socket keeps that socket charged, so neither it nor the network namespace holding it is freed. Bridge, rule and queue need only
CAP_NET_ADMINin a netns, so with user namespaces an unprivileged user can exhaust kernel memory.
nf_reinject()cannot free the skb on a 1 from the okfn, because other> okfns returnNET_RX_DROP, which is 1, after freeing it.Call
nf_hook()instead ofNF_HOOK()and do the learning and theRX_HANDLER_PASSreturn inbr_handle_frame()on 1, as the0x00and0x0ecases already do and asvrf_rcv_nfhook()does. The okfn then runs only for a reinjected frame, whichbr_handle_frame()has already reported consumed and cannot pass on the port, so it learns from the frame and frees it. These frames were not delivered before this change either.Also set *pskb before the new
RX_HANDLER_PASS, as the other such returns do;skb_share_check()may have replaced the skb, and the commit above added that return without the assignment.Discovered by an AI code review agent and reproduced on unmodified mainline under KASAN: an unprivileged user in a new user and network namespace bridged a veth, queued link-local frames to an NFQUEUE listener that accepted every one, and leaked an skb per frame. With this patch the reinjected frames are freed and nothing leaks.
Fixes 3b2e2904deb3 ("net: bridge: fix per-port af_packet sockets")Signed-off-by Daniel Wade <[email protected]>---
Tested in a KVM guest on mainlineb3dfbde62cf1with KASAN andPROVE_LOCKING, running as nobody under "unshare -Urn". No kernel modification, debug patch or injected delay was needed to reproduce.
Before: 192000 bytes charged on the sending socket, unchanged 2 s later. After: 0 bytes, no KASAN report or lockdep warning, and br_input.o builds clean at W=1.Not tested:
NF_STOP,NFQA_CFG_F_FAIL_OPEN, real EAPOL traffic, the netfilter and bridge selftests,CONFIG_NETFILTER=n.Applies to v7.2-rc5, v6.12.117, next-20261109 and
7e1392a42ea2.The Fixes commit was checked against torvalds/linux: it adds both the "return 1" okfn and the
RX_HANDLER_PASSreturn without the *pskb assignment.net/bridge/br_input.c | 18 +++++++++++++----- 1 file changed, 13 insertions(+), 5 deletions(-)net/bridge/br_input.c +10
-4@@ -257,13 +257,18 @@ static void __br_handle_local_finish(struct sk_buff *skb)br_fdb_update (p->br, p, vlan,eth_hdr (skb)->h_source,0 ); } -/* note: already called with rcu_read_lock */ +/* okfn of the link-local NF_BR_LOCAL_IN hook in br_handle_frame(). It is +* only called when a queued frame is reinjected, after br_handle_frame() +* returned RX_HANDLER_CONSUMED for it, so the frame can no longer be +* passed up on the port and has to be freed here. +* note: already called with rcu_read_lock +*/ static int br_handle_local_finish (struct net *net,struct sock *sk,struct sk_buff *skb) {__br_handle_local_finish (skb); +kfree_skb (skb); -/* return 1 to signal the okfn() was called so it's ok to use the skb */ -return 1 ; +return 0 ; }static int nf_hook_bridge_pre (struct sk_buff *skb,struct sk_buff **pskb)@@ -424,11 +429,14 @@ static rx_handler_result_t br_handle_frame(struct sk_buff **pskb)/* The else clause should be hit when nf_hook(): * - returns < 0 (drop/error) * - returns = 0 (stolen/nf_queue) -* Thus return 1 from the okfn() to signal the skb is ok to pass +* On 1 the frame is passed on here. The okfn() only runs if a +* queued frame is reinjected later, and must consume it then.
net/bridge/br_input.c static rx_handler_result_t br_handle_frame(struct sk_buff **pskb) · line 442
Quoted 10 lines
net/bridge/br_input.c +3
-1*/ -if (NF_HOOK (NFPROTO_BRIDGE ,NF_BR_LOCAL_IN , +if (nf_hook (NFPROTO_BRIDGE ,NF_BR_LOCAL_IN ,dev_net (skb->dev),NULL , skb, skb->dev,NULL , br_handle_local_finish) ==1 ) { + *pskb = skb; +__br_handle_local_finish (skb);return RX_HANDLER_PASS ; }else {return RX_HANDLER_CONSUMED ;
The fix looks ok, the commit message though - next time trim the AI slop, all of that could've been 3-4 sentences.
Actually I think there is a worse bug in there, by the time an skb was stolen and reinjected, its port could've gone away (I'll have to verify it) but we deref it in the helper without any checks, that could lead to a null ptr deref or wrong ptr. I will take care of that.
Anyway let's not delay this patch,