On Wed, Oct 18, 2017 at 11:30 PM, Tobin C. Harding <m...@tobin.cc> wrote:
> +static siphash_key_t ptr_secret __read_mostly;
> +static atomic_t have_key = ATOMIC_INIT(0);
> +
> +static void initialize_ptr_secret(void)
> +{
> + if (atomic_read(&have_key) == 1)
> + return;
> +
> + get_random_bytes(&ptr_secret, sizeof(ptr_secret));
> + atomic_set(&have_key, 1);
> +}
> + case -EALREADY:
> + initialize_ptr_secret();
> + break;
Unfortunately the above is racy, and the spinlock you had before was
actually correct (though using an atomic inside a spinlock wasn't
strictly necessary). The race is that two callers might hit
initialize_ptr_secret at the same time, and have_key will be zero at
the beginning for both. Then they'll both scribble over ptr_secret,
and might wind up using a different value after if one finishes before
the other. I see two ways of correcting this:
1) Go back to the spinlock yourself.
2) Use get_random_bytes_once(&ptr_secret, sizeof(ptr_secret)). I don't
know lib/once.c especially well, but from cursory look, it appears to
be taking a spinlock too, which means you're probably good.
+ if (atomic_read(&have_key) == 0) {
+ random_ready.owner = NULL;
+ random_ready.func = schedule_async_key_init;
You can probably take care of this part in the initialization:
static struct random_ready_callback random_ready = {
.func = schedule_async_key_init
};
Alternatively, you could put the actual call to
add_random_ready_callback in an init function, but maybe how you have
it is easier.
Jason
