On 2/22/19 10:21 AM, wenxu wrote:
> On 2019/2/22 下午11:03, David Ahern wrote:
>> On 2/21/19 11:14 PM, wenxu wrote:
>>> build_state in the rcu_read_lock and disable the preempt
>>>
>>> rcu_read_lock();
>>> ops = rcu_dereference(lwtun_encaps[encap_type]);
>>> if (likely(ops && ops->build_state && try_module_get(ops->owner))) {
>>> found = true;
>>> ret = ops->build_state(encap, family, cfg, lws, extack);
>>> if (ret)
>>> module_put(ops->owner);
>>> }
>>> rcu_read_unlock();
>>>
>> Missed that.
>>
>> Once a reference is taken the rcu_read_lock can be dropped before
>> calling build_state allowing the allocations to be GFP_KERNEL.
>>
> I don't think the rcu_read_lock can be dropped. The whole operation of the
> reference should be protect
>
> in the rcu_read_lock that ensure the ops will not be destroy (if it will be)
>
If 'try_module_get(ops->owner)' succeeds then a reference is held so the
ops will not go away.
