[PATCH net] net: initialize ICMP error queue opt_stats

[PATCH net] net: initialize ICMP error queue opt_stats

[Weiming Shi]

ip_icmp_error() and ipv6_icmp_error() clone a received ICMP skb before
building its error queue entry.  skb_clone() preserves the complete control
buffer, while the IP receive paths initialize only the portion aliased by
sock_exterr_skb::header.  The two functions initialize ee, addr_offset, and
port, but leave the final byte containing opt_stats and unused untouched.
A TCX classifier can leave bpf_skb_data_end::data_end across that byte,
forging opt_stats and making __sock_recv_timestamp() pass skb->len bytes
from skb->data to put_cmsg().  For a non-linear ICMP skb, this reads beyond
its linear head.

Initialize both bitfields in the two clone-based producers.  Do not clear
the whole sock_exterr_skb because the error receive paths consume the saved
IP control-buffer header.

  usercopy: Kernel memory exposure attempt detected from SLUB object
            'skbuff_small_head' (offset 72, size 59964)!
  kernel BUG at mm/usercopy.c:102!
  Oops: invalid opcode: 0000 [#1] SMP KASAN NOPTI
  RIP: usercopy_abort (mm/usercopy.c:102)
  Call Trace:
   __check_heap_object (mm/slub.c:8146)
   __check_object_size (mm/usercopy.c:258)
   put_cmsg (net/core/scm.c:276)
   __sock_recv_timestamp (net/socket.c:1061)
   ip_recv_error (net/ipv4/ip_sockglue.c:550)
   udp_recvmsg (net/ipv4/udp.c:2034)
   inet_recvmsg (net/ipv4/af_inet.c:896)
   sock_recvmsg (net/socket.c:1159)
   ____sys_recvmsg (net/socket.c:2918)
   ___sys_recvmsg (net/socket.c:2961)
   __sys_recvmsg (net/socket.c:2993)
   do_syscall_64 (arch/x86/entry/syscall_64.c:94)
   entry_SYSCALL_64_after_hwframe (arch/x86/entry/entry_64.S:130)
  Kernel panic - not syncing: Fatal exception

With HARDENED_USERCOPY enabled, the reproducer aborted the unpatched kernel
on its first packet.  With runtime usercopy checks disabled and KASAN left
in report mode, the same recvmsg() returned 59,964 control-message data
bytes sourced past a 704-byte skb head.  Beyond offset 1,024, 58,932 bytes
differed from the prefilled userspace marker, including 6,288 nonzero
bytes.  The patched build completed 2,000 packets under both configurations
without a forged SCM_TIMESTAMPING_OPT_STATS message or sanitizer report.
Networking selftests were not run.

Fixes: 4ef1b2869447 ("tcp: mark skbs with SCM_TIMESTAMPING_OPT_STATS")
Cc: stable@vger.kernel.org
Reported-by: <co+beaa1ed2d7aaec16@bugs.sh>
Signed-off-by: Weiming Shi <bestswngs@gmail.com>

---
 net/ipv4/ip_sockglue.c | 2 ++
 net/ipv6/datagram.c    | 2 ++
 2 files changed, 4 insertions(+)

diff --git a/net/ipv4/ip_sockglue.c b/net/ipv4/ip_sockglue.c
index a55ef327ec93..3ab84a4c31e5 100644
--- a/net/ipv4/ip_sockglue.c
+++ b/net/ipv4/ip_sockglue.c
@@ -429,6 +429,8 @@ void ip_icmp_error(struct sock *sk, struct sk_buff *skb, int err,
 	serr->addr_offset = (u8 *)&(((struct iphdr *)(icmp_hdr(skb) + 1))->daddr) -
 				   skb_network_header(skb);
 	serr->port = port;
+	serr->opt_stats = 0;
+	serr->unused = 0;
 
 	if (skb_pull(skb, payload - skb->data)) {
 		if (inet_test_bit(RECVERR_RFC4884, sk))
diff --git a/net/ipv6/datagram.c b/net/ipv6/datagram.c
index ca3605acb433..85f646f8d059 100644
--- a/net/ipv6/datagram.c
+++ b/net/ipv6/datagram.c
@@ -330,6 +330,8 @@ void ipv6_icmp_error(struct sock *sk, struct sk_buff *skb, int err,
 	serr->addr_offset = (u8 *)&(((struct ipv6hdr *)(icmph + 1))->daddr) -
 				  skb_network_header(skb);
 	serr->port = port;
+	serr->opt_stats = 0;
+	serr->unused = 0;
 
 	__skb_pull(skb, payload - skb->data);
 
-- 
2.56.0