malicious devices causing unaligned accesses [v2]

malicious devices causing unaligned accesses [v2]

[Oliver Neukum]

Hi,

going through the USB network drivers looking for ways
a malicious device could do us harm I found drivers taking
the alignment coming from the device for granted.

An example can be seen in qmi_wwan:

while (offset + qmimux_hdr_sz < skb->len) {
    hdr = (struct qmimux_hdr*)(skb->data + offset);
    len = be16_to_cpu(hdr->pkt_len);

As you can see the driver accesses stuff coming from the device with the
expectation
that it keep to natural alignment. On some architectures that is a way a
device could use to do bad things to a host. What is to be done about
that?

    Regards
        Oliver

Re: malicious devices causing unaligned accesses [v2]

[Bjørn Mork]

Oliver Neukum <oneukum@suse.com> writes:

> Hi,
>
> going through the USB network drivers looking for ways
> a malicious device could do us harm I found drivers taking
> the alignment coming from the device for granted.
>
> An example can be seen in qmi_wwan:
>
> while (offset + qmimux_hdr_sz < skb->len) {
>     hdr = (struct qmimux_hdr*)(skb->data + offset);
>     len = be16_to_cpu(hdr->pkt_len);
>
> As you can see the driver accesses stuff coming from the device with the
> expectation
> that it keep to natural alignment. On some architectures that is a way a
> device could use to do bad things to a host. What is to be done about
> that?

We can deal with this the same way we deal with hostile hot-plugged CPUs
or memory modules.

Yes, the aligment should probably be verified.  But there are so many
ways a hostile network adapter can mess with us than I don't buy the
"malicious device" argument...

FWIW, the more recent rmnet demuxing implementation from Qualcomm seems
to suffer from the same problem.


struct sk_buff *rmnet_map_deaggregate(struct sk_buff *skb,
				      struct rmnet_port *port)
{
	struct rmnet_map_header *maph;
	struct sk_buff *skbn;
	u32 packet_len;

	if (skb->len == 0)
		return NULL;

	maph = (struct rmnet_map_header *)skb->data;
	packet_len = ntohs(maph->pkt_len) + sizeof(struct rmnet_map_header);


(this implementation moves skb->data by packet_len instead of doing the
offset calculation, but I don't think that makes any difference?)

I guess there is no alignment guarantee here, whether the device is
malicious or not. So we probably have to deal with unaligned accesses to
maph/hdr->pkt_len?



Bjørn

Re: malicious devices causing unaligned accesses [v2]

[Oliver Neukum]

On 17.02.22 11:27, Bjørn Mork wrote:
> Oliver Neukum <oneukum@suse.com> writes:
>
>> Hi,
>>
>> going through the USB network drivers looking for ways
>> a malicious device could do us harm I found drivers taking
>> the alignment coming from the device for granted.
>>
>> An example can be seen in qmi_wwan:
>>
>> while (offset + qmimux_hdr_sz < skb->len) {
>>     hdr = (struct qmimux_hdr*)(skb->data + offset);
>>     len = be16_to_cpu(hdr->pkt_len);
>>
>> As you can see the driver accesses stuff coming from the device with the
>> expectation
>> that it keep to natural alignment. On some architectures that is a way a
>> device could use to do bad things to a host. What is to be done about
>> that?
> We can deal with this the same way we deal with hostile hot-plugged CPUs
> or memory modules.
Yes. That is a basic decision that needs to be made
> Yes, the aligment should probably be verified.  But there are so many
> ways a hostile network adapter can mess with us than I don't buy the
> "malicious device" argument...
Sure, so what is the level of damage that is acceptable?
>
> FWIW, the more recent rmnet demuxing implementation from Qualcomm seems
> to suffer from the same problem.
>
>
> struct sk_buff *rmnet_map_deaggregate(struct sk_buff *skb,
> 				      struct rmnet_port *port)
> {
> 	struct rmnet_map_header *maph;
> 	struct sk_buff *skbn;
> 	u32 packet_len;
>
> 	if (skb->len == 0)
> 		return NULL;
>
> 	maph = (struct rmnet_map_header *)skb->data;
> 	packet_len = ntohs(maph->pkt_len) + sizeof(struct rmnet_map_header);
>
>
> (this implementation moves skb->data by packet_len instead of doing the
> offset calculation, but I don't think that makes any difference?)
>
> I guess there is no alignment guarantee here, whether the device is
> malicious or not. So we probably have to deal with unaligned accesses to
> maph/hdr->pkt_len?
Yes, as far as I can tell a device is fully in spec if it sends frames as
tightly packed as possible, so this is simply a bug, not a security issue.

    Regards
        Oliver