Re: [Patch net] ipv6,mcast: always hold idev->lock before mca_lock

[dingtianhong <dingtianhong@huawei.com>]

On 2013/5/12 7:11, David Miller wrote:
> From: Cong Wang <amwang@redhat.com>
> Date: Wed,  8 May 2013 15:41:54 +0800
> 
>> @@ -1369,7 +1370,7 @@ static struct sk_buff *mld_newpack(struct net_device *dev, int size)
>>  
>>  	skb_reserve(skb, hlen);
>>  
>> -	if (ipv6_get_lladdr(dev, &addr_buf, IFA_F_TENTATIVE)) {
>> +	if (__ipv6_get_lladdr(idev, &addr_buf, IFA_F_TENTATIVE)) {
>>  		/* <draft-ietf-magma-mld-source-05.txt>:
>>  		 * use unspecified address as the source address
>>  		 * when a valid link-local address is not available.
> 
> You aren't necessarily going to be holding idev->lock, therefore you can't
> just do a lockless traversal of idev->addr_list here.
> 
> Yes, you can elide the rcu_read_lock() because you have a known reference
> to 'idev' in these paths, but you can't get rid of the address list locking
> altogether.
> 
> 
I think the problem is clear:

mld_send_report(...){
	read_lock_bh(&idev->lock);	
	add_grec（...）				
	read_unlock_bh(&idev->lock);
}
	--->add_grec(...){
		add_grhead(...)
	}
		--->add_grhead(...){
			mld_newpack(...)
		}
			--->mld_newpack(...){
				ipv6_get_lladdr(...)
			}
				--->ipv6_get_lladdr(...){
					read_lock_bh(&idev->lock);
					...
					read_unlock_bh(&idev->lock);
				}

so I think it is no need to lock twice and its unsafe here