[PATCH] rt_mutex: correctly initialize lockdep in rt_mutex_init_proxy_locked

[PATCH] rt_mutex: correctly initialize lockdep in rt_mutex_init_proxy_locked

[Levin, Alexander (Sasha Levin)]

lockdep can't deal with NULL name or key, and doesn't do anything
with the lock when that happens.

Make rt_mutex_init_proxy_locked pass a name and a key for the lock.

Fixes: f5694788ad8d ("rt_mutex: Add lockdep annotations")
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Sasha Levin <alexander.levin@verizon.com>
---
 kernel/locking/rtmutex.c        |  6 ++++--
 kernel/locking/rtmutex_common.h | 12 ++++++++++--
 2 files changed, 14 insertions(+), 4 deletions(-)

diff --git a/kernel/locking/rtmutex.c b/kernel/locking/rtmutex.c
index 43123533e9b1..f540961cec30 100644
--- a/kernel/locking/rtmutex.c
+++ b/kernel/locking/rtmutex.c
@@ -1679,10 +1679,12 @@ EXPORT_SYMBOL_GPL(__rt_mutex_init);
  * possible at this point because the pi_state which contains the rtmutex
  * is not yet visible to other tasks.
  */
-void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
+void __rt_mutex_init_proxy_locked(struct rt_mutex *lock,
+				const char *name,
+				struct lock_class_key *key,
 				struct task_struct *proxy_owner)
 {
-	__rt_mutex_init(lock, NULL, NULL);
+	__rt_mutex_init(lock, name, key);
 	debug_rt_mutex_proxy_lock(lock, proxy_owner);
 	rt_mutex_set_owner(lock, proxy_owner);
 }
diff --git a/kernel/locking/rtmutex_common.h b/kernel/locking/rtmutex_common.h
index 72ad45a9a794..110dc1ed1e89 100644
--- a/kernel/locking/rtmutex_common.h
+++ b/kernel/locking/rtmutex_common.h
@@ -100,8 +100,16 @@ enum rtmutex_chainwalk {
  * PI-futex support (proxy locking functions, etc.):
  */
 extern struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock);
-extern void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
-				       struct task_struct *proxy_owner);
+extern void __rt_mutex_init_proxy_locked(struct rt_mutex *lock,
+					 const char *name,
+					 struct lock_class_key *key,
+					 struct task_struct *proxy_owner);
+#define rt_mutex_init_proxy_locked(lock, proxy_owner) \
+do { \
+	static struct lock_class_key __key; \
+	__rt_mutex_init_proxy_locked(lock, #lock, &__key, proxy_owner); \
+} while (0)
+
 extern void rt_mutex_proxy_unlock(struct rt_mutex *lock,
 				  struct task_struct *proxy_owner);
 extern void rt_mutex_init_waiter(struct rt_mutex_waiter *waiter);
-- 
2.11.0

Re: [PATCH] rt_mutex: correctly initialize lockdep in rt_mutex_init_proxy_locked

[Peter Zijlstra]

On Sat, Jun 10, 2017 at 02:48:04AM +0000, Levin, Alexander (Sasha Levin) wrote:
> lockdep can't deal with NULL name or key, and doesn't do anything
> with the lock when that happens.

Not doing anything is 'right', the proxy stuff won't be lockdep tracked
anyway. But yeah, the first thing is a wee bit of a problem, for it will
trigger DEBUG_LOCKS_WARN_ON() and fully kill lockdep.

> Make rt_mutex_init_proxy_locked pass a name and a key for the lock.
> 
> Fixes: f5694788ad8d ("rt_mutex: Add lockdep annotations")
> Cc: Linus Torvalds <torvalds@linux-foundation.org>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Thomas Gleixner <tglx@linutronix.de>
> Cc: Ingo Molnar <mingo@kernel.org>
> Cc: linux-kernel@vger.kernel.org
> Signed-off-by: Sasha Levin <alexander.levin@verizon.com>
> ---
>  kernel/locking/rtmutex.c        |  6 ++++--
>  kernel/locking/rtmutex_common.h | 12 ++++++++++--
>  2 files changed, 14 insertions(+), 4 deletions(-)
> 
> diff --git a/kernel/locking/rtmutex.c b/kernel/locking/rtmutex.c
> index 43123533e9b1..f540961cec30 100644
> --- a/kernel/locking/rtmutex.c
> +++ b/kernel/locking/rtmutex.c
> @@ -1679,10 +1679,12 @@ EXPORT_SYMBOL_GPL(__rt_mutex_init);
>   * possible at this point because the pi_state which contains the rtmutex
>   * is not yet visible to other tasks.
>   */
> -void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
> +void __rt_mutex_init_proxy_locked(struct rt_mutex *lock,
> +				const char *name,
> +				struct lock_class_key *key,
>  				struct task_struct *proxy_owner)
>  {
> -	__rt_mutex_init(lock, NULL, NULL);
> +	__rt_mutex_init(lock, name, key);
>  	debug_rt_mutex_proxy_lock(lock, proxy_owner);
>  	rt_mutex_set_owner(lock, proxy_owner);
>  }

Yeah, no need to do that; all we really need here is something like:

---
 kernel/locking/rtmutex-debug.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/kernel/locking/rtmutex-debug.c b/kernel/locking/rtmutex-debug.c
index ac35e648b0e5..8dc647dc4b4b 100644
--- a/kernel/locking/rtmutex-debug.c
+++ b/kernel/locking/rtmutex-debug.c
@@ -175,7 +175,8 @@ void debug_rt_mutex_init(struct rt_mutex *lock, const char *name, struct lock_cl
 	lock->name = name;
 
 #ifdef CONFIG_DEBUG_LOCK_ALLOC
-	lockdep_init_map(&lock->dep_map, name, key, 0);
+	if (name && key)
+		lockdep_init_map(&lock->dep_map, name, key, 0);
 #endif
 }
 

Re: [PATCH] rt_mutex: correctly initialize lockdep in rt_mutex_init_proxy_locked

[Levin, Alexander (Sasha Levin)]

On Sat, Jun 10, 2017 at 04:02:12PM +0200, Peter Zijlstra wrote:
> On Sat, Jun 10, 2017 at 02:48:04AM +0000, Levin, Alexander (Sasha Levin) wrote:
> > lockdep can't deal with NULL name or key, and doesn't do anything
> > with the lock when that happens.
> 
> Not doing anything is 'right', the proxy stuff won't be lockdep tracked
> anyway. But yeah, the first thing is a wee bit of a problem, for it will
> trigger DEBUG_LOCKS_WARN_ON() and fully kill lockdep.

But don't we want pi_state->pi_mutex tracked by lockdep?

> Yeah, no need to do that; all we really need here is something like:
> 
> ---
>  kernel/locking/rtmutex-debug.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/locking/rtmutex-debug.c b/kernel/locking/rtmutex-debug.c
> index ac35e648b0e5..8dc647dc4b4b 100644
> --- a/kernel/locking/rtmutex-debug.c
> +++ b/kernel/locking/rtmutex-debug.c
> @@ -175,7 +175,8 @@ void debug_rt_mutex_init(struct rt_mutex *lock, const char *name, struct lock_cl
>  	lock->name = name;
>  
>  #ifdef CONFIG_DEBUG_LOCK_ALLOC
> -	lockdep_init_map(&lock->dep_map, name, key, 0);
> +	if (name && key)
> +		lockdep_init_map(&lock->dep_map, name, key, 0);
>  #endif
>  }

I didn't want to do that because in later calls on that mutex we
will end up going into lockdep code, and I didn't think that doing
that without calling lockdep_init_map() initially was safe.

-- 

Thanks,
Sasha

Re: [PATCH] rt_mutex: correctly initialize lockdep in rt_mutex_init_proxy_locked

[Peter Zijlstra]

On Sun, Jun 11, 2017 at 02:51:09PM +0000, Levin, Alexander (Sasha Levin) wrote:
> On Sat, Jun 10, 2017 at 04:02:12PM +0200, Peter Zijlstra wrote:
> > On Sat, Jun 10, 2017 at 02:48:04AM +0000, Levin, Alexander (Sasha Levin) wrote:
> > > lockdep can't deal with NULL name or key, and doesn't do anything
> > > with the lock when that happens.
> > 
> > Not doing anything is 'right', the proxy stuff won't be lockdep tracked
> > anyway. But yeah, the first thing is a wee bit of a problem, for it will
> > trigger DEBUG_LOCKS_WARN_ON() and fully kill lockdep.
> 
> But don't we want pi_state->pi_mutex tracked by lockdep?

Nope, we can't. That pi_mutex is owned by userspace and not all
operations upon it are visible to the kernel. That is, a userspace
thread can (conceptually) acquire the lock without the kernel ever
knowing.

We typically only create the pi_state when there's contention, at which
point we create the pi_mutex as owned by someone else (hence proxy).

Also, we 'obviously' hold the thing over the return to userspace, which
is something lockdep very much doesn't like.

> > Yeah, no need to do that; all we really need here is something like:
> > 
> > ---
> >  kernel/locking/rtmutex-debug.c | 3 ++-
> >  1 file changed, 2 insertions(+), 1 deletion(-)
> > 
> > diff --git a/kernel/locking/rtmutex-debug.c b/kernel/locking/rtmutex-debug.c
> > index ac35e648b0e5..8dc647dc4b4b 100644
> > --- a/kernel/locking/rtmutex-debug.c
> > +++ b/kernel/locking/rtmutex-debug.c
> > @@ -175,7 +175,8 @@ void debug_rt_mutex_init(struct rt_mutex *lock, const char *name, struct lock_cl
> >  	lock->name = name;
> >  
> >  #ifdef CONFIG_DEBUG_LOCK_ALLOC
> > -	lockdep_init_map(&lock->dep_map, name, key, 0);
> > +	if (name && key)
> > +		lockdep_init_map(&lock->dep_map, name, key, 0);
> >  #endif
> >  }
> 
> I didn't want to do that because in later calls on that mutex we
> will end up going into lockdep code, and I didn't think that doing
> that without calling lockdep_init_map() initially was safe.

So futex has its own private rt_mutex implementation and interface, none
of which include lockdep hooks. So not initializing the field for that
case should not be a problem.