Re: [PATCH] stop_machine: Avoid potential race behaviour of multi_stop_data::state

[Mark Rutland <mark.rutland@arm.com>]

On Tue, Oct 17, 2023 at 04:50:52PM +0800, Rong Tao wrote:
> From: Rong Tao <rongtao@cestc.cn>
> 
> In commit b1fc58333575 ("stop_machine: Avoid potential race behaviour")
> fix both multi_cpu_stop() and set_state() access multi_stop_data::state,
> We should ensure that multi_stop_data::state is accessed using the rwonce
> method.
> 
> Signed-off-by: Rong Tao <rongtao@cestc.cn>
> ---
>  kernel/stop_machine.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/kernel/stop_machine.c b/kernel/stop_machine.c
> index cedb17ba158a..73de9ab77132 100644
> --- a/kernel/stop_machine.c
> +++ b/kernel/stop_machine.c
> @@ -191,7 +191,7 @@ static void set_state(struct multi_stop_data *msdata,
>  static void ack_state(struct multi_stop_data *msdata)
>  {
>  	if (atomic_dec_and_test(&msdata->thread_ack))
> -		set_state(msdata, msdata->state + 1);
> +		set_state(msdata, READ_ONCE(msdata->state) + 1);

IIUC this is bening, as the state machine only ever has a single writer to
msdata->state (though which thread is the writer may change per iteration of
the loop).

At this point we know that the active thread *is* the writer, and so no other
threads can write to msdata->state, so there is no race and reading that
non-atomically is fine.

I'm not opposed to making this use READ_ONCE(), but I don't think that it's
strictly necessary to do so.

That said, if we really want to avoid the non-atomic read, it's probably better
to have multi_cpu_stop() pass curstate as a paramter to ack_state. That or fold
ack_state() into multi_cpu_stop() and use curstate directly.

Mark.

>  }
>  
>  notrace void __weak stop_machine_yield(const struct cpumask *cpumask)
> -- 
> 2.42.0
>