Re: [PATCH 1/1] Use RCU for the UDP hash lock

[Corey Minyard <minyard@acm.org>]

Stephen Hemminger wrote:
> On Thu, 25 Sep 2008 08:29:36 -0700
> "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> wrote:
>
>   
>> On Wed, Sep 24, 2008 at 03:46:20PM -0500, Corey Minyard wrote:
>>     
>>> Stephen Hemminger wrote:
>>>       
>>>>   
>>>>         
>>>>>   static inline void udp_lib_unhash(struct sock *sk)
>>>>>  {
>>>>> -	write_lock_bh(&udp_hash_lock);
>>>>> -	if (sk_del_node_init(sk)) {
>>>>> +	spin_lock_bh(&udp_hash_wlock);
>>>>> +	if (sk_del_node_rcu(sk)) {
>>>>>  		inet_sk(sk)->num = 0;
>>>>>  		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
>>>>>  	}
>>>>> -	write_unlock_bh(&udp_hash_lock);
>>>>> +	spin_unlock_bh(&udp_hash_wlock);
>>>>> +	synchronize_sched();
>>>>>     
>>>>>           
>>>> Could this be synchronize_rcu? You are using rcu_read_lock() protected 
>>>> sections.
>>>>   
>>>>         
>>> I meant to comment on that.  I wasn't sure which to use, so I chose the 
>>> more conservative approach.  synchronize_rcu() might be appropriate.
>>>       
>> You do indeed need to match the update-side and read-side primitives:
>>
>> 	Update-side				Read-side
>>
>> 	synchronize_rcu()			rcu_read_lock()
>> 	call_rcu()				rcu_read_unlock()
>>
>> 	call_rcu_bh()				rcu_read_lock_bh()
>> 						rcu_read_unlock_bh()
>>
>> 	synchronize_sched()			preempt_disable()
>> 						preempt_enable()
>> 						[and anything else
>> 						 that disables either
>> 						 preemption or irqs]
>>
>> 	synchronize_srcu()			srcu_read_lock()
>> 						srcu_read_unlock()
>>
>>
>> Mixing RCU or RCU-SCHED with RCU-BH will fail in Classic RCU systems,
>> while mixing RCU or RCU-BH with RCU-SCHED will fail in preemptable RCU
>> systems.  Mixing SRCU with any of the other flavors of RCU will fail
>> on any system.
>>
>> So please match them up correctly!
>>     
Ok, will do.  I read more on this, and I think I understand the issues 
better.

>>     
>
> Also, for consistency with other parts of networking code, don't introduce
> the synchronize_sched() or synchronize_srcu() pattern to network protocols
> unless there is a no other way to achieve the desired result.
>   
Do you mean synchronize_rcu(), too?  It seems to be used in the net 
code.  To avoid that I'd need to add a struct rcu_head to struct sock.  
Would that be preferable?

Thanks,

-corey