Skip to contentSkip to conversation

ipv6: make icmp6_send() robust against null skb->dev

PATCH In review Eric Dumazet 4 participants 14 messages Jan 4, 2019 – Jan 8, 2019

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ipv6: make icmp6_send() robust against null skb->dev

syzbot was able to crash one host with the following stack trace :

kasan: GPF could be caused by NULL-ptr deref or user memory access

general protection fault: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 8625 Comm: syz-executor4 Not tainted 4.20.0+ #8
RIP: 0010:dev_net include/linux/netdevice.h:2169 [inline]
RIP: 0010:icmp6_send+0x116/0x2d30 net/ipv6/icmp.c:426
 icmpv6_send
 smack_socket_sock_rcv_skb
 security_sock_rcv_skb
 sk_filter_trim_cap
 __sk_receive_skb
 dccp_v6_do_rcv
 release_sock

This is because a RX packet found socket owned by user and was stored into socket backlog. Before leaving RCU protected section, skb->dev was cleared in __sk_receive_skb(). When socket backlog was finally handled at release_sock() time, skb was fed to smack_socket_sock_rcv_skb() then icmp6_send()

We could fix the bug in smack_socket_sock_rcv_skb(), or simply make icmp6_send() more robust against such possibility.

In the future we might provide to icmp6_send() the net pointer instead of infering it.

Fixes d66a8acbda92 ("Smack: Inform peer that IPv6 traffic has been blocked")
Cc Piotr Sawicki <[email protected]>
Cc Casey Schaufler <[email protected]>
---
 net/ipv6/icmp.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
net/ipv6/icmp.c +6 -2
index 5d7aa2c2770ca2b4981d2dd211c3cf0a79a6f9e2..bbcdfd2996926a78c3ea0b274adfa9b5f297efbc 100644
@@ -423,10 +423,10 @@ static int icmp6_iif(const struct sk_buff *skb)
 static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
 		       const struct in6_addr *force_saddr)
 {
-	struct net *net = dev_net(skb->dev);
 	struct inet6_dev *idev = NULL;
 	struct ipv6hdr *hdr = ipv6_hdr(skb);
 	struct sock *sk;
+	struct net *net;
 	struct ipv6_pinfo *np;
 	const struct in6_addr *saddr = NULL;
 	struct dst_entry *dst;
@@ -437,12 +437,16 @@ static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
 	int iif = 0;
 	int addr_type = 0;
 	int len;
-	u32 mark = IP6_REPLY_MARK(net, skb->mark);
+	u32 mark;
 
 	if ((u8 *)hdr < skb->head ||
 	    (skb_network_header(skb) + sizeof(*hdr)) > skb_tail_pointer(skb))
 		return;
 
+	if (!skb->dev)
+		return;
+	net = dev_net(skb->dev);
+	mark = IP6_REPLY_MARK(net, skb->mark);
 	/*
 	 *	Make sure we respect the rules
 	 *	i.e. RFC 1885 2.4(e)
Signature · 1 line
2.20.1.97.g81188d93c3-goog
#1 lore raw
Quoted 15 lines · Eric Dumazet

syzbot was able to crash one host with the following stack trace :

kasan: GPF could be caused by NULL-ptr deref or user memory access

general protection fault: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 8625 Comm: syz-executor4 Not tainted 4.20.0+ #8
RIP: 0010:dev_net include/linux/netdevice.h:2169 [inline]
RIP: 0010:icmp6_send+0x116/0x2d30 net/ipv6/icmp.c:426
 icmpv6_send
 smack_socket_sock_rcv_skb
 security_sock_rcv_skb
 sk_filter_trim_cap
 __sk_receive_skb
 dccp_v6_do_rcv
 release_sock
Quoted 8 lines

This is because a RX packet found socket owned by user and was stored into socket backlog. Before leaving RCU protected section, skb->dev was cleared in __sk_receive_skb(). When socket backlog was finally handled at release_sock() time, skb was fed to smack_socket_sock_rcv_skb() then icmp6_send()

We could fix the bug in smack_socket_sock_rcv_skb(), or simply make icmp6_send() more robust against such possibility.

The Smack patch would be a trivial check for skb->dev == NULL, in which case it wouldn't call icmp6_send(). Unless there's a timing issue, of course. If there are no known timing issues I would be happy to create a Smack patch to address this problem.

Or, I'm happy with the patch below if you like it.

Quoted 47 lines

In the future we might provide to icmp6_send() the net pointer instead of infering it.

Fixes d66a8acbda92 ("Smack: Inform peer that IPv6 traffic has been blocked")
Cc Piotr Sawicki <[email protected]>
Cc Casey Schaufler <[email protected]>
---
 net/ipv6/icmp.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
net/ipv6/icmp.c +6 -2
index 5d7aa2c2770ca2b4981d2dd211c3cf0a79a6f9e2..bbcdfd2996926a78c3ea0b274adfa9b5f297efbc 100644
@@ -423,10 +423,10 @@ static int icmp6_iif(const struct sk_buff *skb)
 static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
 		       const struct in6_addr *force_saddr)
 {
-	struct net *net = dev_net(skb->dev);
 	struct inet6_dev *idev = NULL;
 	struct ipv6hdr *hdr = ipv6_hdr(skb);
 	struct sock *sk;
+	struct net *net;
 	struct ipv6_pinfo *np;
 	const struct in6_addr *saddr = NULL;
 	struct dst_entry *dst;
@@ -437,12 +437,16 @@ static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
 	int iif = 0;
 	int addr_type = 0;
 	int len;
-	u32 mark = IP6_REPLY_MARK(net, skb->mark);
+	u32 mark;
 
 	if ((u8 *)hdr < skb->head ||
 	    (skb_network_header(skb) + sizeof(*hdr)) > skb_tail_pointer(skb))
 		return;
 
+	if (!skb->dev)
+		return;
+	net = dev_net(skb->dev);
+	mark = IP6_REPLY_MARK(net, skb->mark);
 	/*
 	 *	Make sure we respect the rules
 	 *	i.e. RFC 1885 2.4(e)
#2 lore raw
Quoted 26 lines · Casey Schaufler
Quoted 15 lines · Eric Dumazet

syzbot was able to crash one host with the following stack trace :

kasan: GPF could be caused by NULL-ptr deref or user memory access

general protection fault: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 8625 Comm: syz-executor4 Not tainted 4.20.0+ #8
RIP: 0010:dev_net include/linux/netdevice.h:2169 [inline]
RIP: 0010:icmp6_send+0x116/0x2d30 net/ipv6/icmp.c:426
 icmpv6_send
 smack_socket_sock_rcv_skb
 security_sock_rcv_skb
 sk_filter_trim_cap
 __sk_receive_skb
 dccp_v6_do_rcv
 release_sock
Quoted 8 lines

This is because a RX packet found socket owned by user and was stored into socket backlog. Before leaving RCU protected section, skb->dev was cleared in __sk_receive_skb(). When socket backlog was finally handled at release_sock() time, skb was fed to smack_socket_sock_rcv_skb() then icmp6_send()

We could fix the bug in smack_socket_sock_rcv_skb(), or simply make icmp6_send() more robust against such possibility.

Quoted 7 lines

The Smack patch would be a trivial check for skb->dev == NULL, in which case it wouldn't call icmp6_send(). Unless there's a timing issue, of course. If there are no known timing issues I would be happy to create a Smack patch to address this problem.

Or, I'm happy with the patch below if you like it.

Well, doing the check in icmp6_send() is more generic, this is the path I took, thanks ;)

Quoted 48 lines
Quoted 47 lines

In the future we might provide to icmp6_send() the net pointer instead of infering it.

Fixes d66a8acbda92 ("Smack: Inform peer that IPv6 traffic has been blocked")
Cc Piotr Sawicki <[email protected]>
Cc Casey Schaufler <[email protected]>
---
 net/ipv6/icmp.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
net/ipv6/icmp.c +6 -2
index 5d7aa2c2770ca2b4981d2dd211c3cf0a79a6f9e2..bbcdfd2996926a78c3ea0b274adfa9b5f297efbc 100644
@@ -423,10 +423,10 @@ static int icmp6_iif(const struct sk_buff *skb)
 static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
                     const struct in6_addr *force_saddr)
 {
-     struct net *net = dev_net(skb->dev);
      struct inet6_dev *idev = NULL;
      struct ipv6hdr *hdr = ipv6_hdr(skb);
      struct sock *sk;
+     struct net *net;
      struct ipv6_pinfo *np;
      const struct in6_addr *saddr = NULL;
      struct dst_entry *dst;
@@ -437,12 +437,16 @@ static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
      int iif = 0;
      int addr_type = 0;
      int len;
-     u32 mark = IP6_REPLY_MARK(net, skb->mark);
+     u32 mark;
 
      if ((u8 *)hdr < skb->head ||
          (skb_network_header(skb) + sizeof(*hdr)) > skb_tail_pointer(skb))
              return;
 
+     if (!skb->dev)
+             return;
+     net = dev_net(skb->dev);
+     mark = IP6_REPLY_MARK(net, skb->mark);
      /*
       *      Make sure we respect the rules
       *      i.e. RFC 1885 2.4(e)
#3 lore raw
Quoted 30 lines · Eric Dumazet
Quoted 29 lines · Casey Schaufler
Quoted 15 lines · Eric Dumazet

syzbot was able to crash one host with the following stack trace :

kasan: GPF could be caused by NULL-ptr deref or user memory access

general protection fault: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 8625 Comm: syz-executor4 Not tainted 4.20.0+ #8
RIP: 0010:dev_net include/linux/netdevice.h:2169 [inline]
RIP: 0010:icmp6_send+0x116/0x2d30 net/ipv6/icmp.c:426
 icmpv6_send
 smack_socket_sock_rcv_skb
 security_sock_rcv_skb
 sk_filter_trim_cap
 __sk_receive_skb
 dccp_v6_do_rcv
 release_sock
Quoted 8 lines

This is because a RX packet found socket owned by user and was stored into socket backlog. Before leaving RCU protected section, skb->dev was cleared in __sk_receive_skb(). When socket backlog was finally handled at release_sock() time, skb was fed to smack_socket_sock_rcv_skb() then icmp6_send()

We could fix the bug in smack_socket_sock_rcv_skb(), or simply make icmp6_send() more robust against such possibility.

The Smack patch would be a trivial check for skb->dev == NULL, in which case it wouldn't call icmp6_send(). Unless there's a timing issue, of course. If there are no known timing issues I would be happy to create a Smack patch to address this problem.

Quoted 13 lines
Quoted 2 lines

Or, I'm happy with the patch below if you like it.

Well, doing the check in icmp6_send() is more generic, this is the path I took, thanks ;)

Quoted 8 lines
Quoted 8 lines

In the future we might provide to icmp6_send() the net pointer instead of infering it.

Fixes d66a8acbda92 ("Smack: Inform peer that IPv6 traffic has been blocked")
Cc Piotr Sawicki <[email protected]>
Cc Casey Schaufler <[email protected]>

OK, you can add my

Quoted 38 lines
Quoted 38 lines
Quoted 38 lines
---
 net/ipv6/icmp.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
net/ipv6/icmp.c +6 -2
index 5d7aa2c2770ca2b4981d2dd211c3cf0a79a6f9e2..bbcdfd2996926a78c3ea0b274adfa9b5f297efbc 100644
@@ -423,10 +423,10 @@ static int icmp6_iif(const struct sk_buff *skb)
 static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
                     const struct in6_addr *force_saddr)
 {
-     struct net *net = dev_net(skb->dev);
      struct inet6_dev *idev = NULL;
      struct ipv6hdr *hdr = ipv6_hdr(skb);
      struct sock *sk;
+     struct net *net;
      struct ipv6_pinfo *np;
      const struct in6_addr *saddr = NULL;
      struct dst_entry *dst;
@@ -437,12 +437,16 @@ static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
      int iif = 0;
      int addr_type = 0;
      int len;
-     u32 mark = IP6_REPLY_MARK(net, skb->mark);
+     u32 mark;
 
      if ((u8 *)hdr < skb->head ||
          (skb_network_header(skb) + sizeof(*hdr)) > skb_tail_pointer(skb))
              return;
 
+     if (!skb->dev)
+             return;
+     net = dev_net(skb->dev);
+     mark = IP6_REPLY_MARK(net, skb->mark);
      /*
       *      Make sure we respect the rules
       *      i.e. RFC 1885 2.4(e)
maintainer
#4 lore raw
In reply to Eric Dumazet

From: Eric Dumazet <[email protected]>
Date: Fri, 4 Jan 2019 11:00:00 -0800

Quoted 1 line

syzbot was able to crash one host with the following stack trace :

...

Quoted 9 lines

This is because a RX packet found socket owned by user and was stored into socket backlog. Before leaving RCU protected section, skb->dev was cleared in __sk_receive_skb(). When socket backlog was finally handled at release_sock() time, skb was fed to smack_socket_sock_rcv_skb() then icmp6_send()

We could fix the bug in smack_socket_sock_rcv_skb(), or simply make icmp6_send() more robust against such possibility.

Quoted 8 lines

In the future we might provide to icmp6_send() the net pointer instead of infering it.

Fixes d66a8acbda92 ("Smack: Inform peer that IPv6 traffic has been blocked")
Cc Piotr Sawicki <[email protected]>
Cc Casey Schaufler <[email protected]>

Applied and queued up for -stable, thanks.

#5 lore raw
In reply to Eric Dumazet

dccp_v6_rcv() calls __sk_receive_skb() which calls sk_filter_trim_cap().

sk_filter_trim_cap() should return a value not equal to 0 and cause the skb to be dropped, since icmpv6_send() is called when smack_socket_sock_rcv_skb() returns -EACCES.

So, the packet shouldn't be put into the backlog queue.

How did it get there?

Regards,

Piotr

Quoted 70 lines · Eric Dumazet

syzbot was able to crash one host with the following stack trace :

kasan: GPF could be caused by NULL-ptr deref or user memory access

general protection fault: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 8625 Comm: syz-executor4 Not tainted 4.20.0+ #8
RIP: 0010:dev_net include/linux/netdevice.h:2169 [inline]
RIP: 0010:icmp6_send+0x116/0x2d30 net/ipv6/icmp.c:426
 icmpv6_send
 smack_socket_sock_rcv_skb
 security_sock_rcv_skb
 sk_filter_trim_cap
 __sk_receive_skb
 dccp_v6_do_rcv
 release_sock

This is because a RX packet found socket owned by user and was stored into socket backlog. Before leaving RCU protected section, skb->dev was cleared in __sk_receive_skb(). When socket backlog was finally handled at release_sock() time, skb was fed to smack_socket_sock_rcv_skb() then icmp6_send()

We could fix the bug in smack_socket_sock_rcv_skb(), or simply make icmp6_send() more robust against such possibility.

In the future we might provide to icmp6_send() the net pointer instead of infering it.

Fixes d66a8acbda92 ("Smack: Inform peer that IPv6 traffic has been blocked")
Cc Piotr Sawicki <[email protected]>
Cc Casey Schaufler <[email protected]>
---
 net/ipv6/icmp.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
net/ipv6/icmp.c +6 -2
index 5d7aa2c2770ca2b4981d2dd211c3cf0a79a6f9e2..bbcdfd2996926a78c3ea0b274adfa9b5f297efbc 100644
@@ -423,10 +423,10 @@ static int icmp6_iif(const struct sk_buff *skb)
 static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
 		       const struct in6_addr *force_saddr)
 {
-	struct net *net = dev_net(skb->dev);
 	struct inet6_dev *idev = NULL;
 	struct ipv6hdr *hdr = ipv6_hdr(skb);
 	struct sock *sk;
+	struct net *net;
 	struct ipv6_pinfo *np;
 	const struct in6_addr *saddr = NULL;
 	struct dst_entry *dst;
@@ -437,12 +437,16 @@ static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
 	int iif = 0;
 	int addr_type = 0;
 	int len;
-	u32 mark = IP6_REPLY_MARK(net, skb->mark);
+	u32 mark;
 
 	if ((u8 *)hdr < skb->head ||
 	    (skb_network_header(skb) + sizeof(*hdr)) > skb_tail_pointer(skb))
 		return;
 
+	if (!skb->dev)
+		return;
+	net = dev_net(skb->dev);
+	mark = IP6_REPLY_MARK(net, skb->mark);
 	/*
 	 *	Make sure we respect the rules
 	 *	i.e. RFC 1885 2.4(e)
#6 lore raw
Quoted 9 lines · Piotr Sawicki

dccp_v6_rcv() calls __sk_receive_skb() which calls sk_filter_trim_cap().

sk_filter_trim_cap() should return a value not equal to 0 and cause the skb to be dropped, since icmpv6_send() is called when smack_socket_sock_rcv_skb() returns -EACCES.

So, the packet shouldn't be put into the backlog queue.

How did it get there?

I do not believe crash involved a BPF filter at all (My changelog said nothing about sk_filter_trim_cap())

After packet is queued to backlog, the packet circulates, reaching the smack_socket_sock_rcv_skb() point.

The stack trace shows only the 2nd phase of the packet, when the user process calls release_sock()

Quoted 77 lines

Regards,

Piotr

Quoted 70 lines · Eric Dumazet

syzbot was able to crash one host with the following stack trace :

kasan: GPF could be caused by NULL-ptr deref or user memory access

general protection fault: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 8625 Comm: syz-executor4 Not tainted 4.20.0+ #8
RIP: 0010:dev_net include/linux/netdevice.h:2169 [inline]
RIP: 0010:icmp6_send+0x116/0x2d30 net/ipv6/icmp.c:426
 icmpv6_send
 smack_socket_sock_rcv_skb
 security_sock_rcv_skb
 sk_filter_trim_cap
 __sk_receive_skb
 dccp_v6_do_rcv
 release_sock

This is because a RX packet found socket owned by user and was stored into socket backlog. Before leaving RCU protected section, skb->dev was cleared in __sk_receive_skb(). When socket backlog was finally handled at release_sock() time, skb was fed to smack_socket_sock_rcv_skb() then icmp6_send()

We could fix the bug in smack_socket_sock_rcv_skb(), or simply make icmp6_send() more robust against such possibility.

In the future we might provide to icmp6_send() the net pointer instead of infering it.

Fixes d66a8acbda92 ("Smack: Inform peer that IPv6 traffic has been blocked")
Cc Piotr Sawicki <[email protected]>
Cc Casey Schaufler <[email protected]>
---
 net/ipv6/icmp.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
net/ipv6/icmp.c +6 -2
index 5d7aa2c2770ca2b4981d2dd211c3cf0a79a6f9e2..bbcdfd2996926a78c3ea0b274adfa9b5f297efbc 100644
@@ -423,10 +423,10 @@ static int icmp6_iif(const struct sk_buff *skb)
 static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
                     const struct in6_addr *force_saddr)
 {
-     struct net *net = dev_net(skb->dev);
      struct inet6_dev *idev = NULL;
      struct ipv6hdr *hdr = ipv6_hdr(skb);
      struct sock *sk;
+     struct net *net;
      struct ipv6_pinfo *np;
      const struct in6_addr *saddr = NULL;
      struct dst_entry *dst;
@@ -437,12 +437,16 @@ static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
      int iif = 0;
      int addr_type = 0;
      int len;
-     u32 mark = IP6_REPLY_MARK(net, skb->mark);
+     u32 mark;
 
      if ((u8 *)hdr < skb->head ||
          (skb_network_header(skb) + sizeof(*hdr)) > skb_tail_pointer(skb))
              return;
 
+     if (!skb->dev)
+             return;
+     net = dev_net(skb->dev);
+     mark = IP6_REPLY_MARK(net, skb->mark);
      /*
       *      Make sure we respect the rules
       *      i.e. RFC 1885 2.4(e)
#7 lore raw
Quoted 4 lines · Eric Dumazet
Quoted 2 lines · Piotr Sawicki

dccp_v6_rcv() calls __sk_receive_skb() which calls sk_filter_trim_cap().

Quoted 8 lines
Quoted 6 lines

sk_filter_trim_cap() should return a value not equal to 0 and cause the skb to be dropped, since icmpv6_send() is called when smack_socket_sock_rcv_skb() returns -EACCES.

So, the packet shouldn't be put into the backlog queue.

How did it get there?

I do not believe crash involved a BPF filter at all (My changelog said nothing about sk_filter_trim_cap()

Not only BPF but also the LSM subsystem is involved (in this case Smack).

dccp_v6_rcv()

	__sk_receive_skb()

		sk_filter_trim_cap()
			security_sock_rcv_skb()
				smack_sock_rcv_skb()

So, before putting this skb into the backlog queue,

a network packet is checked against Smack rules. If Smack denies access,

the packet is discarded.

__sk_receive_skb()
...
if (sk_filter_trim_cap <https://elixir.bootlin.com/linux/latest/ident/sk_filter_trim_cap>(sk, skb, trim_cap))
goto discard_and_relse; ...

Quoted 84 lines

After packet is queued to backlog, the packet circulates, reaching the smack_socket_sock_rcv_skb() point.

The stack trace shows only the 2nd phase of the packet, when the user process calls release_sock()

Quoted 76 lines

Regards,

Piotr

Quoted 70 lines · Eric Dumazet

syzbot was able to crash one host with the following stack trace :

kasan: GPF could be caused by NULL-ptr deref or user memory access

general protection fault: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 8625 Comm: syz-executor4 Not tainted 4.20.0+ #8
RIP: 0010:dev_net include/linux/netdevice.h:2169 [inline]
RIP: 0010:icmp6_send+0x116/0x2d30 net/ipv6/icmp.c:426
 icmpv6_send
 smack_socket_sock_rcv_skb
 security_sock_rcv_skb
 sk_filter_trim_cap
 __sk_receive_skb
 dccp_v6_do_rcv
 release_sock

This is because a RX packet found socket owned by user and was stored into socket backlog. Before leaving RCU protected section, skb->dev was cleared in __sk_receive_skb(). When socket backlog was finally handled at release_sock() time, skb was fed to smack_socket_sock_rcv_skb() then icmp6_send()

We could fix the bug in smack_socket_sock_rcv_skb(), or simply make icmp6_send() more robust against such possibility.

In the future we might provide to icmp6_send() the net pointer instead of infering it.

Fixes d66a8acbda92 ("Smack: Inform peer that IPv6 traffic has been blocked")
Cc Piotr Sawicki <[email protected]>
Cc Casey Schaufler <[email protected]>
---
 net/ipv6/icmp.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
net/ipv6/icmp.c +6 -2
index 5d7aa2c2770ca2b4981d2dd211c3cf0a79a6f9e2..bbcdfd2996926a78c3ea0b274adfa9b5f297efbc 100644
@@ -423,10 +423,10 @@ static int icmp6_iif(const struct sk_buff *skb)
 static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
                     const struct in6_addr *force_saddr)
 {
-     struct net *net = dev_net(skb->dev);
      struct inet6_dev *idev = NULL;
      struct ipv6hdr *hdr = ipv6_hdr(skb);
      struct sock *sk;
+     struct net *net;
      struct ipv6_pinfo *np;
      const struct in6_addr *saddr = NULL;
      struct dst_entry *dst;
@@ -437,12 +437,16 @@ static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
      int iif = 0;
      int addr_type = 0;
      int len;
-     u32 mark = IP6_REPLY_MARK(net, skb->mark);
+     u32 mark;
 
      if ((u8 *)hdr < skb->head ||
          (skb_network_header(skb) + sizeof(*hdr)) > skb_tail_pointer(skb))
              return;
 
+     if (!skb->dev)
+             return;
+     net = dev_net(skb->dev);
+     mark = IP6_REPLY_MARK(net, skb->mark);
      /*
       *      Make sure we respect the rules
       *      i.e. RFC 1885 2.4(e)
#8 lore raw
Quoted 27 lines · Piotr Sawicki
Quoted 4 lines · Eric Dumazet
Quoted 2 lines · Piotr Sawicki

dccp_v6_rcv() calls __sk_receive_skb() which calls sk_filter_trim_cap().

Quoted 8 lines
Quoted 6 lines

sk_filter_trim_cap() should return a value not equal to 0 and cause the skb to be dropped, since icmpv6_send() is called when smack_socket_sock_rcv_skb() returns -EACCES.

So, the packet shouldn't be put into the backlog queue.

How did it get there?

I do not believe crash involved a BPF filter at all (My changelog said nothing about sk_filter_trim_cap()

Not only BPF but also the LSM subsystem is involved (in this case Smack).

dccp_v6_rcv()

        __sk_receive_skb()

                sk_filter_trim_cap()
                        security_sock_rcv_skb()
                                smack_sock_rcv_skb()

So, before putting this skb into the backlog queue,

Quoted 9 lines

a network packet is checked against Smack rules. If Smack denies access,

the packet is discarded.

__sk_receive_skb()
...
if (sk_filter_trim_cap <https://elixir.bootlin.com/linux/latest/ident/sk_filter_trim_cap>(sk, skb, trim_cap))
goto discard_and_relse; ...

Crash did not involve sk_filter_trim_cap() here...

Not sure what you are trying to say.

Quoted 84 lines
Quoted 84 lines

After packet is queued to backlog, the packet circulates, reaching the smack_socket_sock_rcv_skb() point.

The stack trace shows only the 2nd phase of the packet, when the user process calls release_sock()

Quoted 76 lines

Regards,

Piotr

Quoted 70 lines · Eric Dumazet

syzbot was able to crash one host with the following stack trace :

kasan: GPF could be caused by NULL-ptr deref or user memory access

general protection fault: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 8625 Comm: syz-executor4 Not tainted 4.20.0+ #8
RIP: 0010:dev_net include/linux/netdevice.h:2169 [inline]
RIP: 0010:icmp6_send+0x116/0x2d30 net/ipv6/icmp.c:426
 icmpv6_send
 smack_socket_sock_rcv_skb
 security_sock_rcv_skb
 sk_filter_trim_cap
 __sk_receive_skb
 dccp_v6_do_rcv
 release_sock

This is because a RX packet found socket owned by user and was stored into socket backlog. Before leaving RCU protected section, skb->dev was cleared in __sk_receive_skb(). When socket backlog was finally handled at release_sock() time, skb was fed to smack_socket_sock_rcv_skb() then icmp6_send()

We could fix the bug in smack_socket_sock_rcv_skb(), or simply make icmp6_send() more robust against such possibility.

In the future we might provide to icmp6_send() the net pointer instead of infering it.

Fixes d66a8acbda92 ("Smack: Inform peer that IPv6 traffic has been blocked")
Cc Piotr Sawicki <[email protected]>
Cc Casey Schaufler <[email protected]>
---
 net/ipv6/icmp.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
net/ipv6/icmp.c +6 -2
index 5d7aa2c2770ca2b4981d2dd211c3cf0a79a6f9e2..bbcdfd2996926a78c3ea0b274adfa9b5f297efbc 100644
@@ -423,10 +423,10 @@ static int icmp6_iif(const struct sk_buff *skb)
 static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
                     const struct in6_addr *force_saddr)
 {
-     struct net *net = dev_net(skb->dev);
      struct inet6_dev *idev = NULL;
      struct ipv6hdr *hdr = ipv6_hdr(skb);
      struct sock *sk;
+     struct net *net;
      struct ipv6_pinfo *np;
      const struct in6_addr *saddr = NULL;
      struct dst_entry *dst;
@@ -437,12 +437,16 @@ static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
      int iif = 0;
      int addr_type = 0;
      int len;
-     u32 mark = IP6_REPLY_MARK(net, skb->mark);
+     u32 mark;
 
      if ((u8 *)hdr < skb->head ||
          (skb_network_header(skb) + sizeof(*hdr)) > skb_tail_pointer(skb))
              return;
 
+     if (!skb->dev)
+             return;
+     net = dev_net(skb->dev);
+     mark = IP6_REPLY_MARK(net, skb->mark);
      /*
       *      Make sure we respect the rules
       *      i.e. RFC 1885 2.4(e)
#9 lore raw
Quoted 31 lines · Eric Dumazet
Quoted 29 lines · Piotr Sawicki
Quoted 4 lines · Eric Dumazet
Quoted 2 lines · Piotr Sawicki

dccp_v6_rcv() calls __sk_receive_skb() which calls sk_filter_trim_cap().

Quoted 8 lines
Quoted 6 lines

sk_filter_trim_cap() should return a value not equal to 0 and cause the skb to be dropped, since icmpv6_send() is called when smack_socket_sock_rcv_skb() returns -EACCES.

So, the packet shouldn't be put into the backlog queue.

How did it get there?

I do not believe crash involved a BPF filter at all (My changelog said nothing about sk_filter_trim_cap()

Not only BPF but also the LSM subsystem is involved (in this case Smack).

dccp_v6_rcv()

        __sk_receive_skb()

                sk_filter_trim_cap()
                        security_sock_rcv_skb()
                                smack_sock_rcv_skb()

So, before putting this skb into the backlog queue,

a network packet is checked against Smack rules. If Smack denies access,

the packet is discarded.

Quoted 8 lines
Quoted 5 lines

__sk_receive_skb()
...
if (sk_filter_trim_cap <https://elixir.bootlin.com/linux/latest/ident/sk_filter_trim_cap>(sk, skb, trim_cap))
goto discard_and_relse; ...

Crash did not involve sk_filter_trim_cap() here...

Not sure what you are trying to say.

Are you sure that sk_filter_trim_cap() is not on the stack trace?

Maybe I'm missing something. How should I read the below dump?

kasan: GPF could be caused by NULL-ptr deref or user memory access

general protection fault: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 8625 Comm: syz-executor4 Not tainted 4.20.0+ #8
RIP: 0010:dev_net include/linux/netdevice.h:2169 [inline]
RIP: 0010:icmp6_send+0x116/0x2d30 net/ipv6/icmp.c:426
 icmpv6_send
 smack_socket_sock_rcv_skb
 security_sock_rcv_skb
 sk_filter_trim_cap <---- here
 __sk_receive_skb
 dccp_v6_do_rcv
 release_sock
Quoted 84 lines
Quoted 83 lines
Quoted 83 lines

After packet is queued to backlog, the packet circulates, reaching the smack_socket_sock_rcv_skb() point.

The stack trace shows only the 2nd phase of the packet, when the user process calls release_sock()

Quoted 76 lines

Regards,

Piotr

Quoted 70 lines · Eric Dumazet

syzbot was able to crash one host with the following stack trace :

kasan: GPF could be caused by NULL-ptr deref or user memory access

general protection fault: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 8625 Comm: syz-executor4 Not tainted 4.20.0+ #8
RIP: 0010:dev_net include/linux/netdevice.h:2169 [inline]
RIP: 0010:icmp6_send+0x116/0x2d30 net/ipv6/icmp.c:426
 icmpv6_send
 smack_socket_sock_rcv_skb
 security_sock_rcv_skb
 sk_filter_trim_cap
 __sk_receive_skb
 dccp_v6_do_rcv
 release_sock

This is because a RX packet found socket owned by user and was stored into socket backlog. Before leaving RCU protected section, skb->dev was cleared in __sk_receive_skb(). When socket backlog was finally handled at release_sock() time, skb was fed to smack_socket_sock_rcv_skb() then icmp6_send()

We could fix the bug in smack_socket_sock_rcv_skb(), or simply make icmp6_send() more robust against such possibility.

In the future we might provide to icmp6_send() the net pointer instead of infering it.

Fixes d66a8acbda92 ("Smack: Inform peer that IPv6 traffic has been blocked")
Cc Piotr Sawicki <[email protected]>
Cc Casey Schaufler <[email protected]>
---
 net/ipv6/icmp.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
net/ipv6/icmp.c +6 -2
index 5d7aa2c2770ca2b4981d2dd211c3cf0a79a6f9e2..bbcdfd2996926a78c3ea0b274adfa9b5f297efbc 100644
@@ -423,10 +423,10 @@ static int icmp6_iif(const struct sk_buff *skb)
 static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
                     const struct in6_addr *force_saddr)
 {
-     struct net *net = dev_net(skb->dev);
      struct inet6_dev *idev = NULL;
      struct ipv6hdr *hdr = ipv6_hdr(skb);
      struct sock *sk;
+     struct net *net;
      struct ipv6_pinfo *np;
      const struct in6_addr *saddr = NULL;
      struct dst_entry *dst;
@@ -437,12 +437,16 @@ static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
      int iif = 0;
      int addr_type = 0;
      int len;
-     u32 mark = IP6_REPLY_MARK(net, skb->mark);
+     u32 mark;
 
      if ((u8 *)hdr < skb->head ||
          (skb_network_header(skb) + sizeof(*hdr)) > skb_tail_pointer(skb))
              return;
 
+     if (!skb->dev)
+             return;
+     net = dev_net(skb->dev);
+     mark = IP6_REPLY_MARK(net, skb->mark);
      /*
       *      Make sure we respect the rules
       *      i.e. RFC 1885 2.4(e)
#10 lore raw
Quoted 39 lines · Piotr Sawicki
Quoted 36 lines · Eric Dumazet
Quoted 34 lines · Piotr Sawicki
Quoted 4 lines · Eric Dumazet
Quoted 2 lines · Piotr Sawicki

dccp_v6_rcv() calls __sk_receive_skb() which calls sk_filter_trim_cap().

Quoted 8 lines
Quoted 6 lines

sk_filter_trim_cap() should return a value not equal to 0 and cause the skb to be dropped, since icmpv6_send() is called when smack_socket_sock_rcv_skb() returns -EACCES.

So, the packet shouldn't be put into the backlog queue.

How did it get there?

I do not believe crash involved a BPF filter at all (My changelog said nothing about sk_filter_trim_cap()

Not only BPF but also the LSM subsystem is involved (in this case Smack).

dccp_v6_rcv()

        __sk_receive_skb()

                sk_filter_trim_cap()
                        security_sock_rcv_skb()
                                smack_sock_rcv_skb()

So, before putting this skb into the backlog queue,

a network packet is checked against Smack rules. If Smack denies access,

the packet is discarded.

__sk_receive_skb()
...
if (sk_filter_trim_cap <https://elixir.bootlin.com/linux/latest/ident/sk_filter_trim_cap>(sk, skb, trim_cap))
goto discard_and_relse; ...

Quoted 9 lines
Quoted 3 lines

Crash did not involve sk_filter_trim_cap() here...

Not sure what you are trying to say.

Are you sure that sk_filter_trim_cap() is not on the stack trace?

Maybe I'm missing something. How should I read the below dump?

Crash happens from smack_socket_sock_rcv_skb() calling icmpv6_send()

You missed the fact that this stuff can be called twice per packet.

First invocation might have allowed the packet being queued (into the backlog)

Anything can happen when user owns the socket lock, including changing security parameters/filters.

Quoted 98 lines

kasan: GPF could be caused by NULL-ptr deref or user memory access

general protection fault: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 8625 Comm: syz-executor4 Not tainted 4.20.0+ #8
RIP: 0010:dev_net include/linux/netdevice.h:2169 [inline]
RIP: 0010:icmp6_send+0x116/0x2d30 net/ipv6/icmp.c:426
 icmpv6_send
 smack_socket_sock_rcv_skb
 security_sock_rcv_skb
 sk_filter_trim_cap <---- here
 __sk_receive_skb
 dccp_v6_do_rcv
 release_sock
Quoted 84 lines
Quoted 83 lines
Quoted 83 lines

After packet is queued to backlog, the packet circulates, reaching the smack_socket_sock_rcv_skb() point.

The stack trace shows only the 2nd phase of the packet, when the user process calls release_sock()

Quoted 76 lines

Regards,

Piotr

Quoted 70 lines · Eric Dumazet

syzbot was able to crash one host with the following stack trace :

kasan: GPF could be caused by NULL-ptr deref or user memory access

general protection fault: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 8625 Comm: syz-executor4 Not tainted 4.20.0+ #8
RIP: 0010:dev_net include/linux/netdevice.h:2169 [inline]
RIP: 0010:icmp6_send+0x116/0x2d30 net/ipv6/icmp.c:426
 icmpv6_send
 smack_socket_sock_rcv_skb
 security_sock_rcv_skb
 sk_filter_trim_cap
 __sk_receive_skb
 dccp_v6_do_rcv
 release_sock

This is because a RX packet found socket owned by user and was stored into socket backlog. Before leaving RCU protected section, skb->dev was cleared in __sk_receive_skb(). When socket backlog was finally handled at release_sock() time, skb was fed to smack_socket_sock_rcv_skb() then icmp6_send()

We could fix the bug in smack_socket_sock_rcv_skb(), or simply make icmp6_send() more robust against such possibility.

In the future we might provide to icmp6_send() the net pointer instead of infering it.

Fixes d66a8acbda92 ("Smack: Inform peer that IPv6 traffic has been blocked")
Cc Piotr Sawicki <[email protected]>
Cc Casey Schaufler <[email protected]>
---
 net/ipv6/icmp.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
net/ipv6/icmp.c +6 -2
index 5d7aa2c2770ca2b4981d2dd211c3cf0a79a6f9e2..bbcdfd2996926a78c3ea0b274adfa9b5f297efbc 100644
@@ -423,10 +423,10 @@ static int icmp6_iif(const struct sk_buff *skb)
 static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
                     const struct in6_addr *force_saddr)
 {
-     struct net *net = dev_net(skb->dev);
      struct inet6_dev *idev = NULL;
      struct ipv6hdr *hdr = ipv6_hdr(skb);
      struct sock *sk;
+     struct net *net;
      struct ipv6_pinfo *np;
      const struct in6_addr *saddr = NULL;
      struct dst_entry *dst;
@@ -437,12 +437,16 @@ static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
      int iif = 0;
      int addr_type = 0;
      int len;
-     u32 mark = IP6_REPLY_MARK(net, skb->mark);
+     u32 mark;
 
      if ((u8 *)hdr < skb->head ||
          (skb_network_header(skb) + sizeof(*hdr)) > skb_tail_pointer(skb))
              return;
 
+     if (!skb->dev)
+             return;
+     net = dev_net(skb->dev);
+     mark = IP6_REPLY_MARK(net, skb->mark);
      /*
       *      Make sure we respect the rules
       *      i.e. RFC 1885 2.4(e)
#11 lore raw
Quoted 47 lines · Eric Dumazet
Quoted 37 lines · Piotr Sawicki
Quoted 35 lines · Eric Dumazet
Quoted 33 lines · Piotr Sawicki
Quoted 4 lines · Eric Dumazet
Quoted 2 lines · Piotr Sawicki

dccp_v6_rcv() calls __sk_receive_skb() which calls sk_filter_trim_cap().

Quoted 8 lines
Quoted 6 lines

sk_filter_trim_cap() should return a value not equal to 0 and cause the skb to be dropped, since icmpv6_send() is called when smack_socket_sock_rcv_skb() returns -EACCES.

So, the packet shouldn't be put into the backlog queue.

How did it get there?

I do not believe crash involved a BPF filter at all (My changelog said nothing about sk_filter_trim_cap()

Not only BPF but also the LSM subsystem is involved (in this case Smack).

dccp_v6_rcv()

        __sk_receive_skb()

                sk_filter_trim_cap()
                        security_sock_rcv_skb()
                                smack_sock_rcv_skb()

So, before putting this skb into the backlog queue,

a network packet is checked against Smack rules. If Smack denies access,

the packet is discarded.

__sk_receive_skb()
...
if (sk_filter_trim_cap <https://elixir.bootlin.com/linux/latest/ident/sk_filter_trim_cap>(sk, skb, trim_cap))
goto discard_and_relse; ...

Quoted 8 lines
Quoted 3 lines

Crash did not involve sk_filter_trim_cap() here...

Not sure what you are trying to say.

Are you sure that sk_filter_trim_cap() is not on the stack trace?

Maybe I'm missing something. How should I read the below dump?

Quoted 8 lines

Crash happens from smack_socket_sock_rcv_skb() calling icmpv6_send()

You missed the fact that this stuff can be called twice per packet.

First invocation might have allowed the packet being queued (into the backlog)

Anything can happen when user owns the socket lock, including changing security parameters/filters.

Yes I know.  It looks like the Smack's security rule was changed during this process.

Firstly the packet was allowed to be received and it was put into the backlog queue. Then, the

rule was changed, and during the release phase LSM was called again for the same packet.

But this time, Smack denied access and tried to send an ICMPv6 packet to inform a peer.

I want to make sure if it is the root cause of this problem.

Besides, what is the purpose of setting skb->dev to NULL in __sk_receive_skb() ?

Quoted 98 lines
Quoted 97 lines

kasan: GPF could be caused by NULL-ptr deref or user memory access

general protection fault: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 8625 Comm: syz-executor4 Not tainted 4.20.0+ #8
RIP: 0010:dev_net include/linux/netdevice.h:2169 [inline]
RIP: 0010:icmp6_send+0x116/0x2d30 net/ipv6/icmp.c:426
 icmpv6_send
 smack_socket_sock_rcv_skb
 security_sock_rcv_skb
 sk_filter_trim_cap <---- here
 __sk_receive_skb
 dccp_v6_do_rcv
 release_sock
Quoted 83 lines
Quoted 83 lines
Quoted 83 lines

After packet is queued to backlog, the packet circulates, reaching the smack_socket_sock_rcv_skb() point.

The stack trace shows only the 2nd phase of the packet, when the user process calls release_sock()

Quoted 76 lines

Regards,

Piotr

Quoted 70 lines · Eric Dumazet

syzbot was able to crash one host with the following stack trace :

kasan: GPF could be caused by NULL-ptr deref or user memory access

general protection fault: 0000 [#1] PREEMPT SMP KASAN
CPU: 0 PID: 8625 Comm: syz-executor4 Not tainted 4.20.0+ #8
RIP: 0010:dev_net include/linux/netdevice.h:2169 [inline]
RIP: 0010:icmp6_send+0x116/0x2d30 net/ipv6/icmp.c:426
 icmpv6_send
 smack_socket_sock_rcv_skb
 security_sock_rcv_skb
 sk_filter_trim_cap
 __sk_receive_skb
 dccp_v6_do_rcv
 release_sock

This is because a RX packet found socket owned by user and was stored into socket backlog. Before leaving RCU protected section, skb->dev was cleared in __sk_receive_skb(). When socket backlog was finally handled at release_sock() time, skb was fed to smack_socket_sock_rcv_skb() then icmp6_send()

We could fix the bug in smack_socket_sock_rcv_skb(), or simply make icmp6_send() more robust against such possibility.

In the future we might provide to icmp6_send() the net pointer instead of infering it.

Fixes d66a8acbda92 ("Smack: Inform peer that IPv6 traffic has been blocked")
Cc Piotr Sawicki <[email protected]>
Cc Casey Schaufler <[email protected]>
---
 net/ipv6/icmp.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
net/ipv6/icmp.c +6 -2
index 5d7aa2c2770ca2b4981d2dd211c3cf0a79a6f9e2..bbcdfd2996926a78c3ea0b274adfa9b5f297efbc 100644
@@ -423,10 +423,10 @@ static int icmp6_iif(const struct sk_buff *skb)
 static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
                     const struct in6_addr *force_saddr)
 {
-     struct net *net = dev_net(skb->dev);
      struct inet6_dev *idev = NULL;
      struct ipv6hdr *hdr = ipv6_hdr(skb);
      struct sock *sk;
+     struct net *net;
      struct ipv6_pinfo *np;
      const struct in6_addr *saddr = NULL;
      struct dst_entry *dst;
@@ -437,12 +437,16 @@ static void icmp6_send(struct sk_buff *skb, u8 type, u8 code, __u32 info,
      int iif = 0;
      int addr_type = 0;
      int len;
-     u32 mark = IP6_REPLY_MARK(net, skb->mark);
+     u32 mark;
 
      if ((u8 *)hdr < skb->head ||
          (skb_network_header(skb) + sizeof(*hdr)) > skb_tail_pointer(skb))
              return;
 
+     if (!skb->dev)
+             return;
+     net = dev_net(skb->dev);
+     mark = IP6_REPLY_MARK(net, skb->mark);
      /*
       *      Make sure we respect the rules
       *      i.e. RFC 1885 2.4(e)
#12 lore raw
Quoted 4 lines · Piotr Sawicki

Yes I know. It looks like the Smack's security rule was changed during this process.

Firstly the packet was allowed to be received and it was put into the backlog queue. Then, the

Quoted 8 lines

rule was changed, and during the release phase LSM was called again for the same packet.

But this time, Smack denied access and tried to send an ICMPv6 packet to inform a peer.

I want to make sure if it is the root cause of this problem.

Besides, what is the purpose of setting skb->dev to NULL in __sk_receive_skb() ?

We can not keep a pointer to the device, the device might be dismantled/freed before socket backlog can be processed.

Input processing is using RCU, meaning no refcount is taken on the device.

#13 lore raw
Quoted 14 lines · Eric Dumazet
Quoted 11 lines · Piotr Sawicki

Yes I know. It looks like the Smack's security rule was changed during this process.

Firstly the packet was allowed to be received and it was put into the backlog queue. Then, the

rule was changed, and during the release phase LSM was called again for the same packet.

But this time, Smack denied access and tried to send an ICMPv6 packet to inform a peer.

I want to make sure if it is the root cause of this problem.

Quoted 6 lines
Quoted 1 line

Besides, what is the purpose of setting skb->dev to NULL in __sk_receive_skb() ?

We can not keep a pointer to the device, the device might be dismantled/freed before socket backlog can be processed.

Input processing is using RCU, meaning no refcount is taken on the device.

Thanks for explanation.