[PATCH] net: arp, ipv6: handle special case of tap device

[PATCH] net: arp, ipv6: handle special case of tap device

[Vladis Dronov]

dev->type can be set to any value for tun/tap devices by TUNSETLINK ioctl.
Nevertheless, all other dev's fields related to the link level like addr_len
and broadcast remain the same as configured previously for ARPHRD_ETHER type,
making dev inconsistent. This can lead to read from uninitialized memory.
Fix this by checking for the case of tun/tap device and act according to the
real link level type of dev while mapping a multicast IP onto multicast MAC

Reported-by: syzbot+d3402c47f680ff24b29c@syzkaller.appspotmail.com
Signed-off-by: Vladis Dronov <vdronov@redhat.com>
---
 net/ipv4/arp.c   | 9 ++++++++-
 net/ipv6/ndisc.c | 9 ++++++++-
 2 files changed, 16 insertions(+), 2 deletions(-)

diff --git a/net/ipv4/arp.c b/net/ipv4/arp.c
index e90c89ef8c08..9ce472cf98a3 100644
--- a/net/ipv4/arp.c
+++ b/net/ipv4/arp.c
@@ -187,7 +187,14 @@ EXPORT_SYMBOL(arp_tbl);
 
 int arp_mc_map(__be32 addr, u8 *haddr, struct net_device *dev, int dir)
 {
-	switch (dev->type) {
+	unsigned short type = dev->type;
+
+#if IS_ENABLED(CONFIG_TAP)
+	if (dev->rtnl_link_ops && !strcmp(dev->rtnl_link_ops->kind, "tun"))
+		type = ARPHRD_ETHER;
+#endif /* CONFIG_TAP */
+
+	switch (type) {
 	case ARPHRD_ETHER:
 	case ARPHRD_FDDI:
 	case ARPHRD_IEEE802:
diff --git a/net/ipv6/ndisc.c b/net/ipv6/ndisc.c
index 0ec273997d1d..9371ff4454f7 100644
--- a/net/ipv6/ndisc.c
+++ b/net/ipv6/ndisc.c
@@ -283,7 +283,14 @@ struct ndisc_options *ndisc_parse_options(const struct net_device *dev,
 
 int ndisc_mc_map(const struct in6_addr *addr, char *buf, struct net_device *dev, int dir)
 {
-	switch (dev->type) {
+	unsigned short type = dev->type;
+
+#if IS_ENABLED(CONFIG_TAP)
+	if (dev->rtnl_link_ops && !strcmp(dev->rtnl_link_ops->kind, "tun"))
+		type = ARPHRD_ETHER;
+#endif /* CONFIG_TAP */
+
+	switch (type) {
 	case ARPHRD_ETHER:
 	case ARPHRD_IEEE802:	/* Not sure. Check it later. --ANK */
 	case ARPHRD_FDDI:
-- 
2.19.0

Re: [PATCH] net: arp, ipv6: handle special case of tap device

[David Miller]

From: Vladis Dronov <vdronov@redhat.com>
Date: Wed, 26 Sep 2018 11:30:18 +0200

> @@ -187,7 +187,14 @@ EXPORT_SYMBOL(arp_tbl);
>  
>  int arp_mc_map(__be32 addr, u8 *haddr, struct net_device *dev, int dir)
>  {
> -	switch (dev->type) {
> +	unsigned short type = dev->type;
> +
> +#if IS_ENABLED(CONFIG_TAP)
> +	if (dev->rtnl_link_ops && !strcmp(dev->rtnl_link_ops->kind, "tun"))
> +		type = ARPHRD_ETHER;
> +#endif /* CONFIG_TAP */

This is insanely ugly.

dev->type determines the link layer header layout and size.

You can fix the kernel, but userspace AF_PACKET applications are still
going to be broken by this behavior.  And that is just the tip of the
iceberg.

I'm not applying this, sorry.  I think tun/tap should be prevented from
allowing the dev->type to be changed, unless it will make those changes
adjust the link layer headers properly as well.

Sorry.

Re: [PATCH] net: arp, ipv6: handle special case of tap device

[Vladis Dronov]

Hello, David, all,

> This is insanely ugly.
> I'm not applying this, sorry. 

It is ugly, I agree.

> dev->type determines the link layer header layout and size.

Yes, unfortunately, TUNSETLINK ioctl handler does not adjust all the needed
fields after the tun->dev->type:

[drivers/net/tun.c]
        case TUNSETLINK:
                ...
                tun->dev->type = (int) arg; //<--- that's all

Let me share what I see and let me hope to get an advise for a better fix.
(also another related thread: https://marc.info/?t=153797194500007&r=1&w=2)

- So setting just the tun->dev->type makes the dev struct inconsistent.

- There are more field to adjust, at least dev->addr_len and dev->broadcast.

- There is no get_addr_len_by_link_type() or a simple way to get link layer
  properties by dev->type. Such settings are scattered in *_setup and
  *_init functions, like ipgre_tunnel_init() { ... dev->addr_len = 4; ...}

> I think tun/tap should be prevented from
> allowing the dev->type to be changed, unless it will make those changes
> adjust the link layer headers properly as well.

Probably, this functionality is already used by some userspace, so I believe,
I cannot just remove TUNSETLINK ioctl. Let me try to suggest a patch that
sets dev->addr_len and dev->broadcast (and probably other link-level-related)
fields according to the dev->type.

Best regards,
Vladis Dronov | Red Hat, Inc. | Product Security Engineer

Re: [PATCH] net: arp, ipv6: handle special case of tap device

[David Miller]

From: Vladis Dronov <vdronov@redhat.com>
Date: Mon, 1 Oct 2018 07:32:32 -0400 (EDT)

> Probably, this functionality is already used by some userspace, so I
> believe, I cannot just remove TUNSETLINK ioctl.

Who uses it, how do they use it, and what do they sanely expect about
the link layer header properties of a tap device which has had ->type
changed?

Re: [PATCH] net: arp, ipv6: handle special case of tap device

[Vladis Dronov]

> From: "David Miller" <davem@davemloft.net>
> Sent: Wednesday, October 3, 2018 7:25:40 AM
>
> > Probably, this functionality is already used by some userspace, so I
> > believe, I cannot just remove TUNSETLINK ioctl.
> 
> Who uses it, how do they use it, and what do they sanely expect about
> the link layer header properties of a tap device which has had ->type
> changed?

Honestly, I have no idea. I was working on a syzkaller bug report and
found that the reason for that bug was TUNSETLINK ioctl.

This ioctl was introduced 13 years ago with ff4cc3ac93e1d and it says:

> For use with
> wireless and other networking types it should be possible to set the
> ARP type via an ioctl.

, that's all. Let me try to prepare a patch that sets dev->addr_len and
dev->broadcast according to dev->type. I would probably need an advise
what other fields must be adjusted too.

Otherwise, we can just to remove the TUNSETLINK ioctl.

Best regards,
Vladis Dronov | Red Hat, Inc. | Product Security Engineer

Re: [PATCH] net: arp, ipv6: handle special case of tap device

[David Miller]

From: Vladis Dronov <vdronov@redhat.com>
Date: Wed, 3 Oct 2018 09:25:12 -0400 (EDT)

> Otherwise, we can just to remove the TUNSETLINK ioctl.

I bet that is the best solution unless we can find a real user.

Re: [PATCH] net: arp, ipv6: handle special case of tap device

[Kalle Valo]

Vladis Dronov <vdronov@redhat.com> writes:

>> From: "David Miller" <davem@davemloft.net>
>> Sent: Wednesday, October 3, 2018 7:25:40 AM
>>
>> > Probably, this functionality is already used by some userspace, so I
>> > believe, I cannot just remove TUNSETLINK ioctl.
>> 
>> Who uses it, how do they use it, and what do they sanely expect about
>> the link layer header properties of a tap device which has had ->type
>> changed?
>
> Honestly, I have no idea. I was working on a syzkaller bug report and
> found that the reason for that bug was TUNSETLINK ioctl.
>
> This ioctl was introduced 13 years ago with ff4cc3ac93e1d and it says:
>
>> For use with
>> wireless and other networking types it should be possible to set the
>> ARP type via an ioctl.
>
> , that's all. Let me try to prepare a patch that sets dev->addr_len and
> dev->broadcast according to dev->type. I would probably need an advise
> what other fields must be adjusted too.
>
> Otherwise, we can just to remove the TUNSETLINK ioctl.

Let's add linux-wireless in case someone there knows about the ioctl.

-- 
Kalle Valo