[Devel] Re: [PATCH net-2.6.25 1/3] Uninline the __inet_hash function

[Eric Dumazet]

Pavel Emelyanov a écrit :
> Eric Dumazet wrote:
>> Pavel Emelyanov a écrit :
>>> This one is used in quite many places in the networking code and
>>> seems to big to be inline.
>>>
>>> After the patch net/ipv4/build-in.o loses 725 bytes:
>>> add/remove: 1/0 grow/shrink: 0/5 up/down: 374/-1099 (-725)
>>> function                                     old     new   delta
>>> __inet_hash                                    -     374    +374
>>> tcp_sacktag_write_queue                     2255    2254      -1
>>> __inet_lookup_listener                       284     274     -10
>>> tcp_v4_syn_recv_sock                         755     495    -260
>>> tcp_v4_hash                                  389      40    -349
>>> inet_hash_connect                           1165     686    -479
>>>
>>> Exporting this is for dccp module.
>>>
>>> Signed-off-by: Pavel Emelyanov <xemul at openvz.org>
>>>
>>> ---
>>>
>>>  include/net/inet_hashtables.h |   27 ++-------------------------
>>>  net/ipv4/inet_hashtables.c    |   27 +++++++++++++++++++++++++++
>>>  2 files changed, 29 insertions(+), 25 deletions(-)
>>>
>>> diff --git a/include/net/inet_hashtables.h b/include/net/inet_hashtables.h
>>> index 37f6cb1..1a43125 100644
>>> --- a/include/net/inet_hashtables.h
>>> +++ b/include/net/inet_hashtables.h
>>> @@ -264,31 +264,8 @@ static inline void inet_listen_unlock(struct inet_hashinfo *hashinfo)
>>>  		wake_up(&hashinfo->lhash_wait);
>>>  }
>>>  
>>> -static inline void __inet_hash(struct inet_hashinfo *hashinfo,
>>> -			       struct sock *sk, const int listen_possible)
>>> -{
>>> -	struct hlist_head *list;
>>> -	rwlock_t *lock;
>>> -
>>> -	BUG_TRAP(sk_unhashed(sk));
>>> -	if (listen_possible && sk->sk_state == TCP_LISTEN) {
>>> -		list = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
>>> -		lock = &hashinfo->lhash_lock;
>>> -		inet_listen_wlock(hashinfo);
>>> -	} else {
>>> -		struct inet_ehash_bucket *head;
>>> -		sk->sk_hash = inet_sk_ehashfn(sk);
>>> -		head = inet_ehash_bucket(hashinfo, sk->sk_hash);
>>> -		list = &head->chain;
>>> -		lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
>>> -		write_lock(lock);
>>> -	}
>>> -	__sk_add_node(sk, list);
>>> -	sock_prot_inc_use(sk->sk_prot);
>>> -	write_unlock(lock);
>>> -	if (listen_possible && sk->sk_state == TCP_LISTEN)
>>> -		wake_up(&hashinfo->lhash_wait);
>>> -}
>>> +extern void __inet_hash(struct inet_hashinfo *hashinfo, struct sock *sk,
>>> +		const int listen_possible);
>>>  
>>>  static inline void inet_hash(struct inet_hashinfo *hashinfo, struct sock *sk)
>>>  {
>>> diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
>>> index 67704da..46f899b 100644
>>> --- a/net/ipv4/inet_hashtables.c
>>> +++ b/net/ipv4/inet_hashtables.c
>>> @@ -267,6 +267,33 @@ static inline u32 inet_sk_port_offset(const struct sock *sk)
>>>  					  inet->dport);
>>>  }
>>>  
>>> +void __inet_hash(struct inet_hashinfo *hashinfo, struct sock *sk,
>>> +		const int listen_possible)
>>> +{
>>> +	struct hlist_head *list;
>>> +	rwlock_t *lock;
>>> +
>>> +	BUG_TRAP(sk_unhashed(sk));
>>> +	if (listen_possible && sk->sk_state == TCP_LISTEN) {
>>> +		list = &hashinfo->listening_hash[inet_sk_listen_hashfn(sk)];
>>> +		lock = &hashinfo->lhash_lock;
>>> +		inet_listen_wlock(hashinfo);
>>> +	} else {
>>> +		struct inet_ehash_bucket *head;
>>> +		sk->sk_hash = inet_sk_ehashfn(sk);
>>> +		head = inet_ehash_bucket(hashinfo, sk->sk_hash);
>>> +		list = &head->chain;
>>> +		lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
>>> +		write_lock(lock);
>>> +	}
>>> +	__sk_add_node(sk, list);
>>> +	sock_prot_inc_use(sk->sk_prot);
>>> +	write_unlock(lock);
>>> +	if (listen_possible && sk->sk_state == TCP_LISTEN)
>>> +		wake_up(&hashinfo->lhash_wait);
>>> +}
>>> +EXPORT_SYMBOL_GPL(__inet_hash);
>>> +
>>>  /*
>>>   * Bind a port for a connect operation and hash it.
>>>   */
>> If you un-inline this (good idea), I am not sure we still need listen_possible 
>> argument.
>>
>> It was usefull only to help compiler to zap dead code (since it was known at 
>> compile time), now it only adds some extra test and argument passing.
> 
> Hm... I've tried to address this issue and got worse result - minus
> 600 bytes (vs minus 725). So, what would be more preferable - get a 
> smaller code with one extra 'if' or get a bit larger code without it?
> 

Strange... What I meant is always assume listen_possible is true.

The if (sk->sk_state == TCP_LISTEN) will finally see the truth.

I did a test here on x86 gcc-4.2.2 and saved 32 bytes.