Re: [PATCH] locking/static_key: Fix false positive warnings on concurrent dec/inc

[Jakub Kicinski]

Thanks for looking at the patch!

On Tue, 19 Mar 2019 13:18:56 +0100, Peter Zijlstra wrote:
> On Mon, Mar 18, 2019 at 02:58:14PM -0700, Jakub Kicinski wrote:
> > Even though the atomic_dec_and_mutex_lock() in
> > __static_key_slow_dec_cpuslocked() can never see a negative
> > value in key->enabled the subsequent sanity check is re-reading
> > key->enabled, which may have been set to -1 in the meantime by
> > static_key_slow_inc_cpuslocked().  
> 
> A little extra detail might not hurt, or a diagram or something.

Like this:

                CPU  A                               CPU B

 __static_key_slow_dec_cpuslocked():
 static_key_slow_inc_cpuslocked():
                               # enabled = 1
        atomic_dec_and_mutex_lock() 
                               # enabled = 0
                                              atomic_read() == 0
                                              atomic_set(-1)
                               # enabled = -1
        val = atomic_read() 
        # Oops - val == -1!

?

The test case is TCP's clean_acked_data_enable() /
clean_acked_data_disable() as tickled by ktls (net/ktls).
Which should probably use the delayed version in the first 
place, hopefully I can get to adding delayed version of 
static branches and converting at some point..

> > Instead of using -1 as a "enable in progress" constant use
> > -0xffff, this way we can still treat smaller negative values
> > as errors.  
> 
> Those offset games always hurt my brain, but see below.
> 
> > Fixes: 4c5ea0a9cd02 ("locking/static_key: Fix concurrent static_key_slow_inc()")
> > Signed-off-by: Jakub Kicinski <jakub.kicinski@xxxxxxxxxxxxx>
> > ---
> >  kernel/jump_label.c | 21 ++++++++++-----------
> >  1 file changed, 10 insertions(+), 11 deletions(-)
> > 
> > diff --git a/kernel/jump_label.c b/kernel/jump_label.c
> > index bad96b476eb6..4a227e70a8f3 100644
> > --- a/kernel/jump_label.c
> > +++ b/kernel/jump_label.c
> > @@ -89,7 +89,7 @@ static void jump_label_update(struct static_key *key);
> >  int static_key_count(struct static_key *key)
> >  {
> >  	/*
> > -	 * -1 means the first static_key_slow_inc() is in progress.
> > +	 * -0xffff means the first static_key_slow_inc() is in progress.
> >  	 *  static_key_enabled() must return true, so return 1 here.
> >  	 */
> >  	int n = atomic_read(&key->enabled);
> > @@ -125,7 +125,10 @@ void static_key_slow_inc_cpuslocked(struct static_key *key)
> >  
> >  	jump_label_lock();
> >  	if (atomic_read(&key->enabled) == 0) {
> > -		atomic_set(&key->enabled, -1);
> > +		/* Use a large enough negative number so we can still
> > +		 * catch underflow bugs in static_key_slow_dec().
> > +		 */  
> 
> Broken comment style.

Ah, sorry, netdev.

> > +		atomic_set(&key->enabled, -0xffff);
> >  		jump_label_update(key);
> >  		/*
> >  		 * Ensure that if the above cmpxchg loop observes our positive
> > @@ -158,7 +161,7 @@ void static_key_enable_cpuslocked(struct static_key *key)
> >  
> >  	jump_label_lock();
> >  	if (atomic_read(&key->enabled) == 0) {
> > -		atomic_set(&key->enabled, -1);
> > +		atomic_set(&key->enabled, -0xffff);
> >  		jump_label_update(key);
> >  		/*
> >  		 * See static_key_slow_inc().
> > @@ -208,15 +211,11 @@ static void __static_key_slow_dec_cpuslocked(struct static_key *key,
> >  {
> >  	lockdep_assert_cpus_held();
> >  
> > -	/*
> > -	 * The negative count check is valid even when a negative
> > -	 * key->enabled is in use by static_key_slow_inc(); a
> > -	 * __static_key_slow_dec() before the first static_key_slow_inc()
> > -	 * returns is unbalanced, because all other static_key_slow_inc()
> > -	 * instances block while the update is in progress.
> > -	 */
> >  	if (!atomic_dec_and_mutex_lock(&key->enabled, &jump_label_mutex)) {
> > -		WARN(atomic_read(&key->enabled) < 0,
> > +		int v;
> > +
> > +		v = atomic_read(&key->enabled);
> > +		WARN(v < 0 && v != -0xffff,
> >  		     "jump label: negative count!\n");
> >  		return;
> >  	}  
> 
> > Alternatively we could implement atomic_dec_and_mutex_lock_return().  
> 
> I think I like that better, something like:

That indeed looks far cleanest, thanks!

Tested-by: Jakub Kicinski <jakub.kicinski@xxxxxxxxxxxxx>

>  kernel/jump_label.c | 21 +++++++++++++--------
>  1 file changed, 13 insertions(+), 8 deletions(-)
> 
> diff --git a/kernel/jump_label.c b/kernel/jump_label.c
> index bad96b476eb6..a799b1ac6b2f 100644
> --- a/kernel/jump_label.c
> +++ b/kernel/jump_label.c
> @@ -206,6 +206,8 @@ static void __static_key_slow_dec_cpuslocked(struct static_key *key,
>  					   unsigned long rate_limit,
>  					   struct delayed_work *work)
>  {
> +	int val;
> +
>  	lockdep_assert_cpus_held();
>  
>  	/*
> @@ -215,17 +217,20 @@ static void __static_key_slow_dec_cpuslocked(struct static_key *key,
>  	 * returns is unbalanced, because all other static_key_slow_inc()
>  	 * instances block while the update is in progress.
>  	 */
> -	if (!atomic_dec_and_mutex_lock(&key->enabled, &jump_label_mutex)) {
> -		WARN(atomic_read(&key->enabled) < 0,
> -		     "jump label: negative count!\n");
> +	val = atomic_fetch_add_unless(&key->enabled, -1, 1);
> +	if (val != 1) {
> +		WARN(val < 0, "jump label: negative count!\n");
>  		return;
>  	}
>  
> -	if (rate_limit) {
> -		atomic_inc(&key->enabled);
> -		schedule_delayed_work(work, rate_limit);
> -	} else {
> -		jump_label_update(key);
> +	jump_label_lock();
> +	if (atomic_dec_and_test(&key->enabled)) {
> +		if (rate_limit) {
> +			atomic_inc(&key->enabled);
> +			schedule_delayed_work(work, rate_limit);
> +		} else {
> +			jump_label_update(key);
> +		}
>  	}
>  	jump_label_unlock();
>  }