Re: [PATCH net v6 07/10] ip6mr: Lock RCU before ip6mr_get_table() call in ip6_mroute_setsockopt()

[Paolo Abeni <pabeni@redhat.com>]

On 10/17/24 20:28, Florian Westphal wrote:
> Stefan Wiehler <stefan.wiehler@nokia.com> wrote:
>>  	case MRT6_ADD_MIF:
>> -		if (optlen < sizeof(vif))
>> -			return -EINVAL;
>> -		if (copy_from_sockptr(&vif, optval, sizeof(vif)))
>> -			return -EFAULT;
>> -		if (vif.mif6c_mifi >= MAXMIFS)
>> -			return -ENFILE;
>> +		if (vif.mif6c_mifi >= MAXMIFS) {
>> +			ret = -ENFILE;
>> +			goto out;
>> +		}
>>  		rtnl_lock();
> 
> Same, sleeping function called in rcu read side section.
> 
> Maybe its time to add refcount_t to struct mr_table?

FTR, I agree using a refcount could be a better approach (and would
avoid keeping the RCU lock held across seq start/stop which sounds
dangerous, too.

@Stefan: in any case before your next submission, please have test run
with a debug build, so that the run-time checker could catch similar issue.

Side minor nit: your  'Fixes' tag should come before your Sob in the tag
area.

Cheers,

Paolo