Re: [PATCH net-next] net: use cmpxchg() in lock_sock_nested()

[David Laight <david.laight.linux@gmail.com>]

On Wed, 25 Feb 2026 03:32:16 +0000
Eric Dumazet <edumazet@google.com> wrote:

> Add a fast path in lock_sock_nested(), to avoid acquiring
> the socket spinlock only to set @owned to one:
> 
>         spin_lock_bh(&sk->sk_lock.slock);
>         if (unlikely(sock_owned_by_user_nocheck(sk)))
>                 __lock_sock(sk);
>         sk->sk_lock.owned = 1;
>         spin_unlock_bh(&sk->sk_lock.slock);

There has to be a better commit message that just quoting the old code.
The whole thing is quite subtle.
Not helped by the 'helper' sock_owned_by_user_nocheck() obfuscating things
and that __lock_sock() is 'wait for socket to be unlocked' and doesn't
actually lock it at all.

	David

> 
> On x86_64 compiler generates something quite efficient:
> 
> ffffffff81e96100 <lock_sock_nested>:
> ffffffff81e96100:       f3 0f 1e fa             endbr64
> ffffffff81e96104:       0f 1f 44 00 00          nopl   0x0(%rax,%rax,1)
> ffffffff81e96109:       b9 01 00 00 00          mov    $0x1,%ecx
> ffffffff81e9610e:       31 c0                   xor    %eax,%eax
> ffffffff81e96110:       f0 48 0f b1 8f 48 01    lock cmpxchg %rcx,0x148(%rdi)
> ffffffff81e96117:       00 00
> ffffffff81e96119:       48 85 c0                test   %rax,%rax
> ffffffff81e9611c:       74 33                   je     ffffffff81e96151 <lock_sock_nested+0x51>
> ...// slow path
> ffffffff81e96151:       2e e9 19 c4 33 00       cs jmp ffffffff821d2570 <__pi___x86_return_thunk>
> 
> Signed-off-by: Eric Dumazet <edumazet@google.com>
> ---
>  include/net/sock.h |  9 +++++++--
>  net/core/sock.c    | 10 ++++++++++
>  2 files changed, 17 insertions(+), 2 deletions(-)
> 
> diff --git a/include/net/sock.h b/include/net/sock.h
> index 55b61e4b0d8318887d527e919fc1103d78ac6d14..84c21fb38a28406e387dc33e0fb5decd18893950 100644
> --- a/include/net/sock.h
> +++ b/include/net/sock.h
> @@ -81,8 +81,13 @@
>   * mini-semaphore synchronizes multiple users amongst themselves.
>   */
>  typedef struct {
> -	spinlock_t		slock;
> -	int			owned;
> +	union {
> +		struct slock_owned {
> +			int		owned;
> +			spinlock_t	slock;
> +		};
> +		long	combined;
> +	};
>  	wait_queue_head_t	wq;
>  	/*
>  	 * We express the mutex-alike socket_lock semantics
> diff --git a/net/core/sock.c b/net/core/sock.c
> index cfb2a6209946089669882cdbd5d1b36c53838989..2a03d1432111f2dcea9d700f80efb7b0bb4a47d8 100644
> --- a/net/core/sock.c
> +++ b/net/core/sock.c
> @@ -3780,6 +3780,16 @@ void noinline lock_sock_nested(struct sock *sk, int subclass)
>  	mutex_acquire(&sk->sk_lock.dep_map, subclass, 0, _RET_IP_);
>  
>  	might_sleep();
> +#ifdef CONFIG_64BIT
> +	if (sizeof(struct slock_owned) == sizeof(long)) {
> +		socket_lock_t tmp;
> +
> +		tmp.slock = __SPIN_LOCK_UNLOCKED(tmp.slock);
> +		tmp.owned = 1;
> +		if (likely(!cmpxchg(&sk->sk_lock.combined, 0, tmp.combined)))
> +			return;
> +	}
> +#endif
>  	spin_lock_bh(&sk->sk_lock.slock);
>  	if (unlikely(sock_owned_by_user_nocheck(sk)))
>  		__lock_sock(sk);