From: Willem de Bruijn <willemdebruijn.kernel@gmail.com> Date: 2023-10-24 19:50:05
From: Willem de Bruijn <willemb@google.com>
LLC reads the mac header with eth_hdr without verifying that the skb
has an Ethernet header.
Syzbot was able to enter llc_rcv on a tun device. Tun can insert
packets without mac len and with user configurable skb->protocol
(passing a tun_pi header when not configuring IFF_NO_PI).
BUG: KMSAN: uninit-value in llc_station_ac_send_test_r net/llc/llc_station.c:81 [inline]
BUG: KMSAN: uninit-value in llc_station_rcv+0x6fb/0x1290 net/llc/llc_station.c:111
llc_station_ac_send_test_r net/llc/llc_station.c:81 [inline]
llc_station_rcv+0x6fb/0x1290 net/llc/llc_station.c:111
llc_rcv+0xc5d/0x14a0 net/llc/llc_input.c:218
__netif_receive_skb_one_core net/core/dev.c:5523 [inline]
__netif_receive_skb+0x1a6/0x5a0 net/core/dev.c:5637
netif_receive_skb_internal net/core/dev.c:5723 [inline]
netif_receive_skb+0x58/0x660 net/core/dev.c:5782
tun_rx_batched+0x3ee/0x980 drivers/net/tun.c:1555
tun_get_user+0x54c5/0x69c0 drivers/net/tun.c:2002
Add a mac_len test before all three eth_hdr(skb) calls under net/llc.
There are further uses in include/net/llc_pdu.h. All these are
protected by a test skb->protocol == ETH_P_802_2. Which does not
protect against this tun scenario.
But the mac_len test added in this patch in llc_fixup_skb will
indirectly protect those too. That is called from llc_rcv before any
other LLC code.
It is tempting to just add a blanket mac_len check in llc_rcv, but
not sure whether that could break valid LLC paths that do not assume
an Ethernet header. 802.2 LLC may be used on top of non-802.3
protocols in principle.
Reported-by: syzbot+a8c7be6dee0de1b669cc@syzkaller.appspotmail.com
Signed-off-by: Willem de Bruijn <willemb@google.com>
---
net/llc/llc_input.c | 10 ++++++++--
net/llc/llc_s_ac.c | 3 +++
net/llc/llc_station.c | 3 +++
3 files changed, 14 insertions(+), 2 deletions(-)
@@ -76,6 +76,9 @@ static int llc_station_ac_send_test_r(struct sk_buff *skb)u32data_size;structsk_buff*nskb;+if(skb->mac_len<ETH_HLEN)+gotoout;+/* The test request command is type U (llc_len = 3) */data_size=ntohs(eth_hdr(skb)->h_proto)-3;nskb=llc_alloc_frame(NULL,skb->dev,LLC_PDU_TYPE_U,data_size);
From: Willem de Bruijn <willemdebruijn.kernel@gmail.com> Date: 2023-10-25 03:12:53
On Tue, Oct 24, 2023 at 3:50 PM Willem de Bruijn
[off-list ref] wrote:
From: Willem de Bruijn <willemb@google.com>
LLC reads the mac header with eth_hdr without verifying that the skb
has an Ethernet header.
Syzbot was able to enter llc_rcv on a tun device. Tun can insert
packets without mac len and with user configurable skb->protocol
(passing a tun_pi header when not configuring IFF_NO_PI).
BUG: KMSAN: uninit-value in llc_station_ac_send_test_r net/llc/llc_station.c:81 [inline]
BUG: KMSAN: uninit-value in llc_station_rcv+0x6fb/0x1290 net/llc/llc_station.c:111
llc_station_ac_send_test_r net/llc/llc_station.c:81 [inline]
llc_station_rcv+0x6fb/0x1290 net/llc/llc_station.c:111
llc_rcv+0xc5d/0x14a0 net/llc/llc_input.c:218
__netif_receive_skb_one_core net/core/dev.c:5523 [inline]
__netif_receive_skb+0x1a6/0x5a0 net/core/dev.c:5637
netif_receive_skb_internal net/core/dev.c:5723 [inline]
netif_receive_skb+0x58/0x660 net/core/dev.c:5782
tun_rx_batched+0x3ee/0x980 drivers/net/tun.c:1555
tun_get_user+0x54c5/0x69c0 drivers/net/tun.c:2002
Add a mac_len test before all three eth_hdr(skb) calls under net/llc.
There are further uses in include/net/llc_pdu.h. All these are
protected by a test skb->protocol == ETH_P_802_2. Which does not
protect against this tun scenario.
But the mac_len test added in this patch in llc_fixup_skb will
indirectly protect those too. That is called from llc_rcv before any
other LLC code.
It is tempting to just add a blanket mac_len check in llc_rcv, but
not sure whether that could break valid LLC paths that do not assume
an Ethernet header. 802.2 LLC may be used on top of non-802.3
protocols in principle.
Reported-by: syzbot+a8c7be6dee0de1b669cc@syzkaller.appspotmail.com
Signed-off-by: Willem de Bruijn <willemb@google.com>
I forgot to add:
Fixes: f83f1768f833 ("[LLC]: skb allocation size for responses")
Can respin if necessary.
At least one of the three eth_hdr uses goes back to before the start
of git history. But the one that syzbot exercises is introduced in
this commit.
That commit is old enough (2008), that effectively all stable kernels
should receive this. This commit also introduces llc_mac_header_len,
which shows at least one valid L2 header that is not an Ethernet
header, back in the day:
+#ifdef CONFIG_TR
+ case ARPHRD_IEEE802_TR:
+ return sizeof(struct trh_hdr);
+#endif
But that token ring variant was removed in 2012 in commit 211ed865108e
("net: delete all instances of special processing for token ring").
From: Eric Dumazet <edumazet@google.com> Date: 2023-10-25 07:09:24
On Tue, Oct 24, 2023 at 9:50 PM Willem de Bruijn
[off-list ref] wrote:
From: Willem de Bruijn <willemb@google.com>
LLC reads the mac header with eth_hdr without verifying that the skb
has an Ethernet header.
Syzbot was able to enter llc_rcv on a tun device. Tun can insert
packets without mac len and with user configurable skb->protocol
(passing a tun_pi header when not configuring IFF_NO_PI).
BUG: KMSAN: uninit-value in llc_station_ac_send_test_r net/llc/llc_station.c:81 [inline]
BUG: KMSAN: uninit-value in llc_station_rcv+0x6fb/0x1290 net/llc/llc_station.c:111
llc_station_ac_send_test_r net/llc/llc_station.c:81 [inline]
llc_station_rcv+0x6fb/0x1290 net/llc/llc_station.c:111
llc_rcv+0xc5d/0x14a0 net/llc/llc_input.c:218
__netif_receive_skb_one_core net/core/dev.c:5523 [inline]
__netif_receive_skb+0x1a6/0x5a0 net/core/dev.c:5637
netif_receive_skb_internal net/core/dev.c:5723 [inline]
netif_receive_skb+0x58/0x660 net/core/dev.c:5782
tun_rx_batched+0x3ee/0x980 drivers/net/tun.c:1555
tun_get_user+0x54c5/0x69c0 drivers/net/tun.c:2002
Add a mac_len test before all three eth_hdr(skb) calls under net/llc.
There are further uses in include/net/llc_pdu.h. All these are
protected by a test skb->protocol == ETH_P_802_2. Which does not
protect against this tun scenario.
But the mac_len test added in this patch in llc_fixup_skb will
indirectly protect those too. That is called from llc_rcv before any
other LLC code.
It is tempting to just add a blanket mac_len check in llc_rcv, but
not sure whether that could break valid LLC paths that do not assume
an Ethernet header. 802.2 LLC may be used on top of non-802.3
protocols in principle.
Reported-by: syzbot+a8c7be6dee0de1b669cc@syzkaller.appspotmail.com
Note that the syzbot report is still private in syzbot infra.
Reviewed-by: Eric Dumazet <edumazet@google.com>
@@ -76,6 +76,9 @@ static int llc_station_ac_send_test_r(struct sk_buff *skb)u32data_size;structsk_buff*nskb;+if(skb->mac_len<ETH_HLEN)+gotoout;+/* The test request command is type U (llc_len = 3) */data_size=ntohs(eth_hdr(skb)->h_proto)-3;nskb=llc_alloc_frame(NULL,skb->dev,LLC_PDU_TYPE_U,data_size);--
From: Jakub Kicinski <kuba@kernel.org> Date: 2023-10-25 23:00:17
On Tue, 24 Oct 2023 23:12:13 -0400 Willem de Bruijn wrote:
Fixes: f83f1768f833 ("[LLC]: skb allocation size for responses")
Can respin if necessary.
At least one of the three eth_hdr uses goes back to before the start
of git history.
Right, good enough.
But the one that syzbot exercises is introduced in this commit.
That commit is old enough (2008), that effectively all stable kernels
should receive this. This commit also introduces llc_mac_header_len,
which shows at least one valid L2 header that is not an Ethernet
header, back in the day:
+#ifdef CONFIG_TR
+ case ARPHRD_IEEE802_TR:
+ return sizeof(struct trh_hdr);
+#endif
But that token ring variant was removed in 2012 in commit 211ed865108e
("net: delete all instances of special processing for token ring").
From: Jakub Kicinski <kuba@kernel.org> Date: 2023-10-25 23:14:14
On Wed, 25 Oct 2023 19:05:59 -0400 Willem de Bruijn wrote:
quoted
I think this one may want 1 to indicate error, technically. No?
Absolutely, thanks. For both tests.
Both state machine callbacks (functions with names ending with 'test_r'
in the patch), right? Perhaps 'goto out' is what's expected here.
The fixup function seems to have inverse return semantics, because
why not.
From: Willem de Bruijn <willemdebruijn.kernel@gmail.com> Date: 2023-10-25 23:19:37
On Wed, Oct 25, 2023 at 7:14 PM Jakub Kicinski [off-list ref] wrote:
On Wed, 25 Oct 2023 19:05:59 -0400 Willem de Bruijn wrote:
quoted
quoted
I think this one may want 1 to indicate error, technically. No?
Absolutely, thanks. For both tests.
Both state machine callbacks (functions with names ending with 'test_r'
in the patch), right? Perhaps 'goto out' is what's expected here.
The fixup function seems to have inverse return semantics, because
why not.
Exactly :) Extrapolating from that is where I went wrong.