[net] net/netfilter: bpf: avoid leakage of skb

  From: "D. Wythe" <alibuda@linux.alibaba.com>

A malicious eBPF program can interrupt the subsequent processing of
a skb by returning an exceptional retval, and no one will be responsible
for releasing the very skb.

Moreover, normal programs can also have the demand to return NF_STOLEN,
usually, the hook needs to take responsibility for releasing this skb
itself, but currently, there is no such helper function to achieve that.
Ignoring NF_STOLEN will also lead to skb leakage.

Fixes: fd9c663b9ad6 ("bpf: minimal support for programs hooked into netfilter framework")
Signed-off-by: D. Wythe <alibuda@linux.alibaba.com>
---
 net/netfilter/nf_bpf_link.c | 19 ++++++++++++++++++-
 1 file changed, 18 insertions(+), 1 deletion(-)
  

D. Wythe <alibuda@linux.alibaba.com> wrote:
> From: "D. Wythe" <alibuda@linux.alibaba.com>
> 
> A malicious eBPF program can interrupt the subsequent processing of
> a skb by returning an exceptional retval, and no one will be responsible
> for releasing the very skb.

How?  The bpf verifier is supposed to reject nf bpf programs that
return a value other than accept or drop.

If this is a real bug, please also figure out why
006c0e44ed92 ("selftests/bpf: add missing netfilter return value and ctx access tests")
failed to catch it.

> Moreover, normal programs can also have the demand to return NF_STOLEN,

No, this should be disallowed already.

>  net/netfilter/nf_bpf_link.c | 19 ++++++++++++++++++-
>  1 file changed, 18 insertions(+), 1 deletion(-)
> 
> diff --git a/net/netfilter/nf_bpf_link.c b/net/netfilter/nf_bpf_link.c
> index e502ec0..03c47d6 100644
> --- a/net/netfilter/nf_bpf_link.c
> +++ b/net/netfilter/nf_bpf_link.c
> @@ -12,12 +12,29 @@ static unsigned int nf_hook_run_bpf(void *bpf_prog, struct sk_buff *skb,
>  				    const struct nf_hook_state *s)
>  {
>  	const struct bpf_prog *prog = bpf_prog;
> +	unsigned int verdict;
>  	struct bpf_nf_ctx ctx = {
>  		.state = s,
>  		.skb = skb,
>  	};
>  
> -	return bpf_prog_run(prog, &ctx);
> +	verdict = bpf_prog_run(prog, &ctx);
> +	switch (verdict) {
> +	case NF_STOLEN:
> +		consume_skb(skb);
> +		fallthrough;

This can't be right.  STOLEN really means STOLEN (free'd,
redirected, etc, "skb" MUST be "leaked".

Which is also why the bpf program is not allowed to return it.
  

On 11/29/23 9:18 PM, Florian Westphal wrote:
> D. Wythe <alibuda@linux.alibaba.com> wrote:
>> From: "D. Wythe" <alibuda@linux.alibaba.com>
>>
>> A malicious eBPF program can interrupt the subsequent processing of
>> a skb by returning an exceptional retval, and no one will be responsible
>> for releasing the very skb.
> How?  The bpf verifier is supposed to reject nf bpf programs that
> return a value other than accept or drop.
>
> If this is a real bug, please also figure out why
> 006c0e44ed92 ("selftests/bpf: add missing netfilter return value and ctx access tests")
> failed to catch it.

Hi Florian,

You are right, i make a mistake.. , it's not a bug..

And my origin intention was to allow ebpf progs to return NF_STOLEN, we 
are trying to modify some netfilter modules via ebpf,
and some scenarios require the use of NF_STOLEN, but from your 
description, it seems that at least currently,
you do not want to return NF_STOLEN, until there is a helper for 
sonsume_skb(), right ?

Again, very sorry to bother you.

Best wishes,
D. Wythe.

>> Moreover, normal programs can also have the demand to return NF_STOLEN,
> No, this should be disallowed already.
>
>>   net/netfilter/nf_bpf_link.c | 19 ++++++++++++++++++-
>>   1 file changed, 18 insertions(+), 1 deletion(-)
>>
>> diff --git a/net/netfilter/nf_bpf_link.c b/net/netfilter/nf_bpf_link.c
>> index e502ec0..03c47d6 100644
>> --- a/net/netfilter/nf_bpf_link.c
>> +++ b/net/netfilter/nf_bpf_link.c
>> @@ -12,12 +12,29 @@ static unsigned int nf_hook_run_bpf(void *bpf_prog, struct sk_buff *skb,
>>   				    const struct nf_hook_state *s)
>>   {
>>   	const struct bpf_prog *prog = bpf_prog;
>> +	unsigned int verdict;
>>   	struct bpf_nf_ctx ctx = {
>>   		.state = s,
>>   		.skb = skb,
>>   	};
>>   
>> -	return bpf_prog_run(prog, &ctx);
>> +	verdict = bpf_prog_run(prog, &ctx);
>> +	switch (verdict) {
>> +	case NF_STOLEN:
>> +		consume_skb(skb);
>> +		fallthrough;
> This can't be right.  STOLEN really means STOLEN (free'd,
> redirected, etc, "skb" MUST be "leaked".
>
> Which is also why the bpf program is not allowed to return it.
  

D. Wythe <alibuda@linux.alibaba.com> wrote:
> And my origin intention was to allow ebpf progs to return NF_STOLEN, we are
> trying to modify some netfilter modules via ebpf,
> and some scenarios require the use of NF_STOLEN, but from your description,

NF_STOLEN can only be supported via a trusted helper, as least as far as
I understand.

Otherwise verifier would have to guarantee that any branch that returns
NF_STOLEN has released the skb, or passed it to a function that will
release the skb in the near future.
  

On 11/29/23 10:47 PM, Florian Westphal wrote:
> D. Wythe <alibuda@linux.alibaba.com> wrote:
>> And my origin intention was to allow ebpf progs to return NF_STOLEN, we are
>> trying to modify some netfilter modules via ebpf,
>> and some scenarios require the use of NF_STOLEN, but from your description,
> NF_STOLEN can only be supported via a trusted helper, as least as far as
> I understand.
>
> Otherwise verifier would have to guarantee that any branch that returns
> NF_STOLEN has released the skb, or passed it to a function that will
> release the skb in the near future.

Thank you very much for your help. I now understand the difficulty here.
The verifier cannot determine whether the consume_skb() was executed or not,
when the return value  goes to NF_STOLEN.

We may use NF_DROP at first, it won't be make much difference for us now.

Also, do you have any plans to support this helper?

Best wishes,
D. Wythe