Re: [PATCH 1/2] tracing: fix uptime overflow problem

[Steven Rostedt <rostedt@goodmis.org>]

On Sat, 28 Jun 2014 19:10:00 +0800
Xie XiuQi <xiexiuqi@huawei.com> wrote:

 
>  /*
> - * trace_jiffy_clock(): Simply use jiffies as a clock counter.
> + * trace_clock_uptime(): Use lockless version __current_kernel_time,
> + * so it's safe in NMI context.
>   */
> -u64 notrace trace_clock_jiffies(void)
> +u64 notrace trace_clock_uptime(void)
>  {
> -	u64 jiffy = jiffies - INITIAL_JIFFIES;
> +	struct timespec uptime, now, boottime;
> +
> +	/* Does not take xtime_lock, so it's safe in NMI context. */
> +	now = __current_kernel_time();

The problem with this patch is that you have a race:

        ts.tv_sec = tk->xtime_sec;
        ts.tv_nsec = (long)(tk->xtime_nsec >> tk->shift);

There's a chance that gets done between the update of xtime_sec an
xtime_nsec, and make time go backwards.

This call is lockless but is not reliable.

-- Steve


> +	getboottime(&boottime);
> +	uptime = timespec_sub(now, boottime);
>  
>  	/* Return nsecs */
> -	return (u64)jiffies_to_usecs(jiffy) * 1000ULL;
> +	return timespec_to_ns(&uptime);
>  }
>  
>  /*