On Mon, 2012-10-29 at 14:28 +0100, Frederic Weisbecker wrote:
> Context requirements on irq work claim are not entirely
> clear. But it appears that we can try to claim a work that
> may be already claimed by another CPU.
> 
> If so then the early check on IRQ_WORK_PENDING in
> irq_work_claim() is racy because another CPU may be
> changing the flags concurrently and we have nothing
> to synchronize against that. So the value we deal with
> may be stale for a while already.
> 
> To fix this, start with our best wish as the initial
> value for the work flags and feed cmpxchg with it. But
> only do the check against IRQ_WORK_PENDING flag with the
> cmpxchg result.
> 
> Nonetheless, if the work is not pending but our best wish
> was wrong, restart with the old value returned by cmpxchg.
> 
> Signed-off-by: Frederic Weisbecker <[email protected]>
> Cc: Peter Zijlstra <[email protected]>
> Cc: Ingo Molnar <[email protected]>
> Cc: Thomas Gleixner <[email protected]>
> Cc: Andrew Morton <[email protected]>
> Cc: Steven Rostedt <[email protected]>
> Cc: Paul Gortmaker <[email protected]>
> ---
>  kernel/irq_work.c |   17 ++++++++++++-----
>  1 files changed, 12 insertions(+), 5 deletions(-)
> 
> diff --git a/kernel/irq_work.c b/kernel/irq_work.c
> index 1588e3b..679c13e 100644
> --- a/kernel/irq_work.c
> +++ b/kernel/irq_work.c
> @@ -34,15 +34,22 @@ static DEFINE_PER_CPU(struct llist_head, irq_work_list);
>   */
>  static bool irq_work_claim(struct irq_work *work)
>  {
> -     unsigned long flags, nflags;
> +     unsigned long flags, oflags, nflags;
>  
> +     /*
> +      * Can't check IRQ_WORK_PENDING bit right now because the work
> +      * can be running on another CPU and we are not sync with its
> +      * changes to work flags. Only cmpxchg can reliably check for us.
> +      */
> +     flags = work->flags & ~IRQ_WORK_PENDING;
>       for (;;) {
> -             flags = work->flags;

I wonder if the bug is just a memory barrier missing here? But that also
suggests that the other CPU used a memory barrier too (or cmpxchg()
which implies one).

But this change looks fine too.

-- Steve

> -             if (flags & IRQ_WORK_PENDING)
> -                     return false;
>               nflags = flags | IRQ_WORK_FLAGS;
> -             if (cmpxchg(&work->flags, flags, nflags) == flags)
> +             oflags = cmpxchg(&work->flags, flags, nflags);
> +             if (oflags == flags)
>                       break;
> +             if (oflags & IRQ_WORK_PENDING)
> +                     return false;
> +             flags = oflags;
>               cpu_relax();
>       }
>  


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Reply via email to