Re: [PATCH net-next v2 1/3] drop_monitor: Implement namespace filtering/reporting for software drops

[nb <nikolay.borisov@virtuozzo.com>]

On 23.11.22 г. 20:10 ч., Ido Schimmel wrote:
> On Wed, Nov 23, 2022 at 05:21:23PM +0200, nb wrote:
>>
>>
>> On 23.11.22 г. 17:16 ч., Ido Schimmel wrote:
>>> On Wed, Nov 23, 2022 at 04:28:15PM +0200, Nikolay Borisov wrote:
>>>>    static void trace_drop_common(struct sk_buff *skb, void *location)
>>>>    {
>>>>    	struct net_dm_alert_msg *msg;
>>>> @@ -219,7 +233,11 @@ static void trace_drop_common(struct sk_buff *skb, void *location)
>>>>    	int i;
>>>>    	struct sk_buff *dskb;
>>>>    	struct per_cpu_dm_data *data;
>>>> -	unsigned long flags;
>>>> +	unsigned long flags, ns_id = 0;
>>>> +
>>>> +	if (skb->dev && net_dm_ns &&
>>>> +	    dev_net(skb->dev)->ns.inum != net_dm_ns)
>>>
>>> I don't think this is going to work, unfortunately. 'skb->dev' is in a
>>> union with 'dev_scratch' so 'skb->dev' does not necessarily point to a
>>> valid netdev at all times. It can explode when dev_net() tries to
>>> dereference it.
>>>
>>> __skb_flow_dissect() is doing something similar, but I believe there the
>>> code paths were audited to make sure it is safe.
>>>
>>> Did you consider achieving this functionality with a BPF program
>>> attached to skb::kfree_skb tracepoint? I believe BPF programs are run
>>> with page faults disabled, so it should be safe to attempt this there.
>>
>> How would that be different than the trace_drop_common which is called as
>> part of the trace_kfree_skb, as it's really passed as trace point probe via:
> 
> Consider this call path:
> 
> __udp_queue_rcv_skb()
>      __udp_enqueue_schedule_skb()
>          udp_set_dev_scratch() // skb->dev is not NULL, but not a pointer to a netdev either
> 	// error is returned
>      kfree_skb_reason() // probe is called
> 
> dev_net(skb->dev) in the probe will try to dereference skb->dev and
> crash.

This can easily be rectified by using is_kernel() .

> 
> On the other hand, a BPF program that is registered as another probe on
> the tracepoint will access the memory via bpf_probe_read_kernel(), which
> will try to safely read the memory and return an error if it can't. You
> can do that today without any kernel changes.

I did a PoC for this and indeed it works, however I'd still like to 
pursue this code provided there is upstream interest.