Re: [PATCH 1/5] lib: introduce call_once()

[Nick Piggin <nickpiggin@yahoo.com.au>]

On Tuesday 11 March 2008 01:57, Akinobu Mita wrote:

> +static inline int call_once(struct once_control *once_control,
> +			    int (*init_rouine)(void))
> +{
> +	return likely(once_control->done) ? 0
> +			: call_once_slow(once_control, init_rouine);
> +}
> +
> +#endif /* __LINUX_ONCE_H */
> Index: 2.6-rc/lib/once.c
> ===================================================================
> --- /dev/null
> +++ 2.6-rc/lib/once.c
> @@ -0,0 +1,18 @@
> +#include <linux/module.h>
> +#include <linux/once.h>
> +
> +int call_once_slow(struct once_control *once_control, int
> (*init_rouine)(void)) +{
> +	int err = 0;
> +
> +	mutex_lock(&once_control->lock);
> +	if (!once_control->done) {
> +		err = init_rouine();
> +		if (!err)
> +			once_control->done = 1;
> +	}
> +	mutex_unlock(&once_control->lock);
> +
> +	return err;
> +}
> +EXPORT_SYMBOL_GPL(call_once_slow);

The store "once_control->done = 1" can become visible before
init_routine has finished. The code after calling call_once may
also speculatively load some memory before the load of
once_control->done completes, so you can likewise have a data
race that way too.

To fix this, you need smp_wmb after init_rouine(), and probably
smp_mb() in the fastpath after the check but before returning.

Basically any time you have this situation where you're touching
a shared variable without using locks, then you're vastly
increasing the complexity of the code, and so you must have a
good reason for it.

So acquiring the mutex unconditionally would be the best way to
go, unless you're calling this a lot in fastpaths (in which case
I would say you should probably rework your code)

Thanks,
Nick