Re: [tip:locking/urgent] locking/static_key: Fix concurrent static_key_slow_inc()

[Paolo Bonzini]

On 31/07/2017 11:36, Peter Zijlstra wrote:
> On Fri, Jun 24, 2016 at 01:59:06AM -0700, tip-bot for Paolo Bonzini wrote:
>> +++ b/kernel/jump_label.c
>> @@ -58,13 +58,36 @@ static void jump_label_update(struct static_key *key);
>>  
>>  void static_key_slow_inc(struct static_key *key)
>>  {
>> +	int v, v1;
>> +
>>  	STATIC_KEY_CHECK_USE();
>> -	if (atomic_inc_not_zero(&key->enabled))
>> -		return;
>> +
>> +	/*
>> +	 * Careful if we get concurrent static_key_slow_inc() calls;
>> +	 * later calls must wait for the first one to _finish_ the
>> +	 * jump_label_update() process.  At the same time, however,
>> +	 * the jump_label_update() call below wants to see
>> +	 * static_key_enabled(&key) for jumps to be updated properly.
>> +	 *
>> +	 * So give a special meaning to negative key->enabled: it sends
>> +	 * static_key_slow_inc() down the slow path, and it is non-zero
>> +	 * so it counts as "enabled" in jump_label_update().  Note that
>> +	 * atomic_inc_unless_negative() checks >= 0, so roll our own.
>> +	 */
>> +	for (v = atomic_read(&key->enabled); v > 0; v = v1) {
>> +		v1 = atomic_cmpxchg(&key->enabled, v, v + 1);
>> +		if (likely(v1 == v))
>> +			return;
>> +	}
>>  
>>  	jump_label_lock();
>> -	if (atomic_inc_return(&key->enabled) == 1)
>> +	if (atomic_read(&key->enabled) == 0) {
>> +		atomic_set(&key->enabled, -1);
>>  		jump_label_update(key);
>> +		atomic_set(&key->enabled, 1);
>> +	} else {
>> +		atomic_inc(&key->enabled);
>> +	}
>>  	jump_label_unlock();
>>  }
> 
> 
> So I was recently looking at this again and got all paranoid. Do we want
> something like so?

Though I agree with the paranoia sentiment, no:

- key->enabled cannot go from 0 to nonzero outside jump_label_mutex.
For this reason the atomic_try_cmpxchg is unnecessary.

- the (implied) smp_mb before jump_label_update is not interesting, but
I don't think it is useful because: 1) during the jump_label_update
there is no correspondence between what static_key_enabled returns and
what the text look like; 2) what would it even be pairing with?

- 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.  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.

And indeed, while there aren't many, I think two of them are wrong (and
not fixed by your patch):

- include/linux/cpuset.h defines nr_cpusets which uses static_key_count.
 It makes no sense to call it outside cpuset_mutex, and indeed that's
how kernel/cgroup/cpuset.c uses it (nr_cpusets <- generate_sched_domains
<- rebuild_sched_domains_locked).  The function should be moved inside
kernel/cgroup/cpuset.c since the mutex is static.

- kernel/cgroup/cgroup.c only enables/disables at init, so using
static_key_enabled should be fine.

- kernel/tracepoint.c only manipulates tracepoint static keys under
tracepoints_mutex, and uses static_key_enabled under the same mutex, so
it's fine.

- 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.

Paolo

> ---
>  kernel/jump_label.c | 24 +++++++++++++++++-------
>  1 file changed, 17 insertions(+), 7 deletions(-)
> 
> diff --git a/kernel/jump_label.c b/kernel/jump_label.c
> index d11c506a6ac3..69d07e78e48b 100644
> --- a/kernel/jump_label.c
> +++ b/kernel/jump_label.c
> @@ -103,7 +103,7 @@ EXPORT_SYMBOL_GPL(static_key_disable);
>  
>  void static_key_slow_inc(struct static_key *key)
>  {
> -	int v, v1;
> +	int v;
>  
>  	STATIC_KEY_CHECK_USE();
>  
> @@ -119,18 +119,28 @@ void static_key_slow_inc(struct static_key *key)
>  	 * so it counts as "enabled" in jump_label_update().  Note that
>  	 * atomic_inc_unless_negative() checks >= 0, so roll our own.
>  	 */
> -	for (v = atomic_read(&key->enabled); v > 0; v = v1) {
> -		v1 = atomic_cmpxchg(&key->enabled, v, v + 1);
> -		if (likely(v1 == v))
> +	for (v = atomic_read(&key->enabled); v > 0;) {
> +		if (atomic_try_cmpxchg(&key->enabled, &v, v+1))
>  			return;
>  	}
>  
>  	cpus_read_lock();
>  	jump_label_lock();
> -	if (atomic_read(&key->enabled) == 0) {
> -		atomic_set(&key->enabled, -1);
> +
> +	if (atomic_try_cmpxchg(&key->enabled, 0, -1)) {
> +		/*
> +		 * smp_mb implied, must have -1 before proceeding to change
> +		 * text.
> +		 */
>  		jump_label_update(key);
> -		atomic_set(&key->enabled, 1);
> +
> +		/*
> +		 * smp_mb, such that we finish modifying text before enabling
> +		 * the fast path. Probably not needed because modifying text is
> +		 * likely to serialize everything. Be paranoid.
> +		 */
> +		smp_mb__before_atomic();
> +		atomic_add(2, &key->enabled); /* -1 -> 1 */
>  	} else {
>  		atomic_inc(&key->enabled);
>  	}
>