On 31/07/2017 15:33, Peter Zijlstra wrote:
> On Mon, Jul 31, 2017 at 03:04:06PM +0200, Paolo Bonzini wrote:
>> - the smp_mb (though it could be a smp_wmb or atomic_set_release)
>> initially triggered my paranoia indeed. But even then, I can't see why
>> you would need it because there's nothing it pairs with.
>
> So this one would pair with the mb implied by the cmpxchg loop for
> inc-if-positive. The ordering being that if we see a positive v, we must
> then also see all the text modifications.
>
> So if jump_label_update() were to not fully serialize things, it would
> be possible for the v=1 store to appear before the last text changes.
> And in that case we'd allow the fast path to complete
> static_key_slow_inc() before it was in fact done with changing all text.
>
> Now, I suspect (but did not audit) that anything that changes text must
> in fact serialize world, but I wasn't sure.
I see. Then yes, I agree a smp_wmb would be nicer here.
>> Rather, it's *any use of key->enabled outside jump_label_lock*
>> (meaning: any use of static_key_enabled and static_key_count outside
>> the core jump_label.c code) that should be handled with care.
>
>> - net/ipv4/udp.c and net/ipv6/udp.c want to implement a "once-only"
>> increment of the static key. It's racy and maybe we should provide a
>> new API static_key_enable_forever:
>>
>> void static_key_enable_forever(struct static_key *key)
>> {
>> STATIC_KEY_CHECK_USE();
>> if (atomic_read(&key->enabled) > 0)
>> return;
>>
>> cpus_read_lock();
>> jump_label_lock();
>> if (atomic_read(&key->enabled) == 0) {
>> atomic_set(&key->enabled, -1);
>> jump_label_update(key);
>> atomic_set(&key->enabled, 1);
>> }
>> jump_label_unlock();
>> cpus_read_unlock();
>> }
>> EXPORT_SYMBOL_GPL(static_key_enable_forever);
>>
>> I can prepare a patch if you agree.
>
> Isn't that what we have static_key_enable() for? Which btw also uses
> static_key_count() outside of the mutex.
Yes, they should be fixed and net/ can then use static_key_enable.
Paolo
