Re: Add memory barrier when waiting on futex

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Nov 25, 2013 at 01:15:17PM +0000, Ma, Xindong wrote:
> We encountered following panic several times:

> [   74.671982] BUG: unable to handle kernel NULL pointer dereference at 00000008
> [   74.672101] IP: [<c129bb27>] wake_futex+0x47/0x80

> [   74.674144]  [<c129bc29>] futex_wake+0xc9/0x110
> [   74.674195]  [<c129da0b>] do_futex+0xeb/0x950
> [   74.674484]  [<c129e30b>] SyS_futex+0x9b/0x140
> [   74.674582]  [<c195a718>] syscall_call+0x7/0xb
> 
> On smp systems, setting current task to q->task in queue_me() may
> not visible immediately to another cpu, some times this will
> cause panic in wake_futex(). Adding memory barrier to avoid this.
> 
> Signed-off-by: Leon Ma <xindong.ma@intel.com>
> Signed-off-by: xiaobing tu <xiaobing.tu@intel.com>
> ---
>  kernel/futex.c |    1 +
>  1 files changed, 1 insertions(+), 0 deletions(-)
> 
> diff --git a/kernel/futex.c b/kernel/futex.c
> index 80ba086..792cd41 100644
> --- a/kernel/futex.c
> +++ b/kernel/futex.c
> @@ -1529,6 +1529,7 @@ static inline void queue_me(struct futex_q *q, struct futex_hash_bucket *hb)
>  	plist_node_init(&q->list, prio);
>  	plist_add(&q->list, &hb->chain);
>  	q->task = current;
> +	smp_mb();
>  	spin_unlock(&hb->lock);
>  }

This is wrong, because an uncommented barrier is wrong per definition.

This is further wrong because wake_futex() is always called with
hb->lock held, and since queue_me also has hb->lock held, this is in
fact guaranteed.

This is even more wrong for adding stable.