MCS Lock: Restructure the MCS lock defines and locking code into its own file

We will need the MCS lock code for doing optimistic spinning for rwsem and queue rwlock. Extracting the MCS code from mutex.c and put into its own file allow us to reuse this code easily.

Signed-off-by Tim Chen <[email protected]>

Signed-off-by Davidlohr Bueso <[email protected]>

---
 include/linux/mcs_spinlock.h |   64 ++++++++++++++++++++++++++++++++++++++++++
 include/linux/mutex.h        |    5 ++-
 kernel/locking/mutex.c       |   60 ++++----------------------------------
 3 files changed, 74 insertions(+), 55 deletions(-)
 create mode 100644 include/linux/mcs_spinlock.h

include/linux/mcs_spinlock.h new file +23 -0

new file mode 100644
index 0000000..b5de3b0
@@ -0,0 +1,64 @@
+/*
+ * MCS lock defines
+ *
+ * This file contains the main data structure and API definitions of MCS lock.
+ *
+ * The MCS lock (proposed by Mellor-Crummey and Scott) is a simple spin-lock
+ * with the desirable properties of being fair, and with each cpu trying
+ * to acquire the lock spinning on a local variable.
+ * It avoids expensive cache bouncings that common test-and-set spin-lock
+ * implementations incur.
+ */
+#ifndef __LINUX_MCS_SPINLOCK_H
+#define __LINUX_MCS_SPINLOCK_H
+
+struct mcs_spinlock {
+	struct mcs_spinlock *next;
+	int locked; /* 1 if lock acquired */
+};
+
+/*
+ * We don't inline mcs_spin_lock() so that perf can correctly account for the
+ * time spent in this lock function.
+ */

include/linux/mcs_spinlock.h new file +6 -0

+}
+
+static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
+{
+	struct mcs_spinlock *next = ACCESS_ONCE(node->next);
+

include/linux/mcs_spinlock.h new file +3 -0

+}
+
+#endif /* __LINUX_MCS_SPINLOCK_H */

include/linux/mutex.h +3 -2

index bab49da..32a32e6 100644
@@ -46,6 +46,7 @@
  * - detects multi-task circular deadlocks and prints out all affected
  *   locks and tasks (and only those tasks)
  */
+struct mcs_spinlock;
 struct mutex {
 	/* 1: unlocked, 0: locked, negative: locked, possible waiters */
 	atomic_t		count;

@@ -55,7 +56,7 @@ struct mutex {
 	struct task_struct	*owner;
 #endif
 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
-	void			*spin_mlock;	/* Spinner MCS lock */
+	struct mcs_spinlock	*mcs_lock;	/* Spinner MCS lock */
 #endif
 #ifdef CONFIG_DEBUG_MUTEXES
 	const char 		*name;

@@ -179,4 +180,4 @@ extern int atomic_dec_and_mutex_lock(atomic_t *cnt, struct mutex *lock);
 # define arch_mutex_cpu_relax() cpu_relax()
 #endif
 
-#endif
+#endif /* __LINUX_MUTEX_H */

kernel/locking/mutex.c +7 -53

index d24105b..e08b183 100644
@@ -25,6 +25,7 @@
 #include <linux/spinlock.h>
 #include <linux/interrupt.h>
 #include <linux/debug_locks.h>
+#include <linux/mcs_spinlock.h>
 
 /*
  * In the DEBUG case we are using the "NULL fastpath" for mutexes,

@@ -52,7 +53,7 @@ __mutex_init(struct mutex *lock, const char *name, struct lock_class_key *key)
 	INIT_LIST_HEAD(&lock->wait_list);
 	mutex_clear_owner(lock);
 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
-	lock->spin_mlock = NULL;
+	lock->mcs_lock = NULL;
 #endif
 
 	debug_mutex_init(lock, name, key);

@@ -111,54 +112,7 @@ EXPORT_SYMBOL(mutex_lock);
  * more or less simultaneously, the spinners need to acquire a MCS lock
  * first before spinning on the owner field.
  *
- * We don't inline mspin_lock() so that perf can correctly account for the
- * time spent in this lock function.
  */
-struct mspin_node {
-	struct mspin_node *next ;
-	int		  locked;	/* 1 if lock acquired */
-};
-#define	MLOCK(mutex)	((struct mspin_node **)&((mutex)->spin_mlock))
-
-static noinline
-void mspin_lock(struct mspin_node **lock, struct mspin_node *node)
-{
-	struct mspin_node *prev;
-
-	/* Init node */
-	node->locked = 0;
-	node->next   = NULL;
-
-	prev = xchg(lock, node);
-	if (likely(prev == NULL)) {
-		/* Lock acquired */
-		node->locked = 1;
-		return;
-	}
-	ACCESS_ONCE(prev->next) = node;
-	smp_wmb();
-	/* Wait until the lock holder passes the lock down */
-	while (!ACCESS_ONCE(node->locked))
-		arch_mutex_cpu_relax();
-}
-
-static void mspin_unlock(struct mspin_node **lock, struct mspin_node *node)
-{
-	struct mspin_node *next = ACCESS_ONCE(node->next);
-
-	if (likely(!next)) {
-		/*
-		 * Release the lock by setting it to NULL
-		 */
-		if (cmpxchg(lock, node, NULL) == node)
-			return;
-		/* Wait until the next pointer is set */
-		while (!(next = ACCESS_ONCE(node->next)))
-			arch_mutex_cpu_relax();
-	}
-	ACCESS_ONCE(next->locked) = 1;
-	smp_wmb();
-}
 
 /*
  * Mutex spinning code migrated from kernel/sched/core.c

@@ -448,7 +402,7 @@ __mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
 
 	for (;;) {
 		struct task_struct *owner;
-		struct mspin_node  node;
+		struct mcs_spinlock  node;
 
 		if (use_ww_ctx && ww_ctx->acquired > 0) {
 			struct ww_mutex *ww;

@@ -470,10 +424,10 @@ __mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
 		 * If there's an owner, wait for it to either
 		 * release the lock or go to sleep.
 		 */
-		mspin_lock(MLOCK(lock), &node);
+		mcs_spin_lock(&lock->mcs_lock, &node);
 		owner = ACCESS_ONCE(lock->owner);
 		if (owner && !mutex_spin_on_owner(lock, owner)) {
-			mspin_unlock(MLOCK(lock), &node);
+			mcs_spin_unlock(&lock->mcs_lock, &node);
 			goto slowpath;
 		}
 

@@ -488,11 +442,11 @@ __mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
 			}
 
 			mutex_set_owner(lock);
-			mspin_unlock(MLOCK(lock), &node);
+			mcs_spin_unlock(&lock->mcs_lock, &node);
 			preempt_enable();
 			return 0;
 		}
-		mspin_unlock(MLOCK(lock), &node);
+		mcs_spin_unlock(&lock->mcs_lock, &node);
 
 		/*
 		 * When there's no owner, we might have preempted between the

Signature · 1 line

1.7.4.4

Quoted 6 lines · Tim Chen

We will need the MCS lock code for doing optimistic spinning for rwsem and queue rwlock. Extracting the MCS code from mutex.c and put into its own file allow us to reuse this code easily.

Signed-off-by Tim Chen <[email protected]>

Signed-off-by Davidlohr Bueso <[email protected]>

Please see comments below.

Thanx, Paul

Quoted 36 lines

---
 include/linux/mcs_spinlock.h |   64 ++++++++++++++++++++++++++++++++++++++++++
 include/linux/mutex.h        |    5 ++-
 kernel/locking/mutex.c       |   60 ++++----------------------------------
 3 files changed, 74 insertions(+), 55 deletions(-)
 create mode 100644 include/linux/mcs_spinlock.h

include/linux/mcs_spinlock.h new file +23 -0

new file mode 100644
index 0000000..b5de3b0
@@ -0,0 +1,64 @@
+/*
+ * MCS lock defines
+ *
+ * This file contains the main data structure and API definitions of MCS lock.
+ *
+ * The MCS lock (proposed by Mellor-Crummey and Scott) is a simple spin-lock
+ * with the desirable properties of being fair, and with each cpu trying
+ * to acquire the lock spinning on a local variable.
+ * It avoids expensive cache bouncings that common test-and-set spin-lock
+ * implementations incur.
+ */
+#ifndef __LINUX_MCS_SPINLOCK_H
+#define __LINUX_MCS_SPINLOCK_H
+
+struct mcs_spinlock {
+	struct mcs_spinlock *next;
+	int locked; /* 1 if lock acquired */
+};
+
+/*
+ * We don't inline mcs_spin_lock() so that perf can correctly account for the
+ * time spent in this lock function.
+ */

include/linux/mcs_spinlock.h line 33

Quoted 10 lines

include/linux/mcs_spinlock.h new file +10 -0

+static noinline
+void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
+{
+	struct mcs_spinlock *prev;
+
+	/* Init node */
+	node->locked = 0;
+	node->next   = NULL;
+
+	prev = xchg(lock, node);

OK, the full memory barriers implied by xchg() ensure that *node will be initialized before the "ACCESS_ONCE(prev->next) = node" below puts the node into the list. This rules out the misordering scenario that Tim Chen called out in message-id <1380322005.3467.186.camel@schen9-DESK> on September 27th.

Assuming of course a corresponding barrier on the lock handoff side.

include/linux/mcs_spinlock.h line 40

Quoted 7 lines

include/linux/mcs_spinlock.h new file +7 -0

+	if (likely(prev == NULL)) {
+		/* Lock acquired */
+		node->locked = 1;
+		return;
+	}
+	ACCESS_ONCE(prev->next) = node;
+	smp_wmb();

I don't see what the above memory barrier does. Here are some things that it cannot be doing:

o Ordering the insertion into the list above with the polling
below. First, smp_wmb() does not order prior writes against later reads, and second misordering is harmless. If we start polling before the insertion is complete, all that happens is that the first few polls have no chance of seeing a lock grant.

o Ordering the polling against the initialization -- the above
xchg() is already doing that for us.

So what is its purpose?

include/linux/mcs_spinlock.h line 43

Quoted 4 lines

include/linux/mcs_spinlock.h new file +3 -0

+	/* Wait until the lock holder passes the lock down */
+	while (!ACCESS_ONCE(node->locked))
+		arch_mutex_cpu_relax();

On the other hand, I don't see how we get away without a barrier here. As written, what prevents the caller's load from ->owner from being reordered with the above load from ->locked? (Perhaps you can argue that such reordering is only a performance problem, but if so we need that argument recorded in comments.)

Of course, if anyone ever tries to use mcs_spin_lock() as a full lock, they will need a memory barrier here to prevent the critical section from leaking out.

Quoted 7 lines

include/linux/mcs_spinlock.h new file +6 -0

+}
+
+static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
+{
+	struct mcs_spinlock *next = ACCESS_ONCE(node->next);
+

include/linux/mcs_spinlock.h line 59

Quoted 10 lines

include/linux/mcs_spinlock.h new file +10 -0

+	if (likely(!next)) {
+		/*
+		 * Release the lock by setting it to NULL
+		 */
+		if (cmpxchg(lock, node, NULL) == node)
+			return;
+		/* Wait until the next pointer is set */
+		while (!(next = ACCESS_ONCE(node->next)))
+			arch_mutex_cpu_relax();
+	}

We need a memory barrier somewhere before here in this function, otherwise the critical section can leak out. I do not believe that we can rely on the prohibition against speculative stores that Peter Zijlstra and I have been discussing because that does not provide the transitivity required by locking primitives. I believe that we -could- make the access below be an smp_store_release(), though.

Placing the barrier here (or at least not preceding the initial fetch from node->next) has the advantage of allowing it to pair with the xchg() in mcs_spin_lock(), though given the dependency only an smp_read_barrier_depends() is required for that purpose.

include/linux/mcs_spinlock.h line 61

Quoted 2 lines

include/linux/mcs_spinlock.h new file +2 -0

+	ACCESS_ONCE(next->locked) = 1;
+	smp_wmb();

I don't see what this barrier does for us. It is ordering the unlock store with what, exactly?

If it really is doing something, we need a big fat comment stating what that is, and checkpatch.pl will be happy to inform you. ;-)

Quoted 140 lines

include/linux/mcs_spinlock.h new file +3 -0

+}
+
+#endif /* __LINUX_MCS_SPINLOCK_H */

include/linux/mutex.h +3 -2

index bab49da..32a32e6 100644
@@ -46,6 +46,7 @@
  * - detects multi-task circular deadlocks and prints out all affected
  *   locks and tasks (and only those tasks)
  */
+struct mcs_spinlock;
 struct mutex {
 	/* 1: unlocked, 0: locked, negative: locked, possible waiters */
 	atomic_t		count;

@@ -55,7 +56,7 @@ struct mutex {
 	struct task_struct	*owner;
 #endif
 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
-	void			*spin_mlock;	/* Spinner MCS lock */
+	struct mcs_spinlock	*mcs_lock;	/* Spinner MCS lock */
 #endif
 #ifdef CONFIG_DEBUG_MUTEXES
 	const char 		*name;

@@ -179,4 +180,4 @@ extern int atomic_dec_and_mutex_lock(atomic_t *cnt, struct mutex *lock);
 # define arch_mutex_cpu_relax() cpu_relax()
 #endif
 
-#endif
+#endif /* __LINUX_MUTEX_H */

kernel/locking/mutex.c +7 -53

index d24105b..e08b183 100644
@@ -25,6 +25,7 @@
 #include <linux/spinlock.h>
 #include <linux/interrupt.h>
 #include <linux/debug_locks.h>
+#include <linux/mcs_spinlock.h>
 
 /*
  * In the DEBUG case we are using the "NULL fastpath" for mutexes,

@@ -52,7 +53,7 @@ __mutex_init(struct mutex *lock, const char *name, struct lock_class_key *key)
 	INIT_LIST_HEAD(&lock->wait_list);
 	mutex_clear_owner(lock);
 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
-	lock->spin_mlock = NULL;
+	lock->mcs_lock = NULL;
 #endif
 
 	debug_mutex_init(lock, name, key);

@@ -111,54 +112,7 @@ EXPORT_SYMBOL(mutex_lock);
  * more or less simultaneously, the spinners need to acquire a MCS lock
  * first before spinning on the owner field.
  *
- * We don't inline mspin_lock() so that perf can correctly account for the
- * time spent in this lock function.
  */
-struct mspin_node {
-	struct mspin_node *next ;
-	int		  locked;	/* 1 if lock acquired */
-};
-#define	MLOCK(mutex)	((struct mspin_node **)&((mutex)->spin_mlock))
-
-static noinline
-void mspin_lock(struct mspin_node **lock, struct mspin_node *node)
-{
-	struct mspin_node *prev;
-
-	/* Init node */
-	node->locked = 0;
-	node->next   = NULL;
-
-	prev = xchg(lock, node);
-	if (likely(prev == NULL)) {
-		/* Lock acquired */
-		node->locked = 1;
-		return;
-	}
-	ACCESS_ONCE(prev->next) = node;
-	smp_wmb();
-	/* Wait until the lock holder passes the lock down */
-	while (!ACCESS_ONCE(node->locked))
-		arch_mutex_cpu_relax();
-}
-
-static void mspin_unlock(struct mspin_node **lock, struct mspin_node *node)
-{
-	struct mspin_node *next = ACCESS_ONCE(node->next);
-
-	if (likely(!next)) {
-		/*
-		 * Release the lock by setting it to NULL
-		 */
-		if (cmpxchg(lock, node, NULL) == node)
-			return;
-		/* Wait until the next pointer is set */
-		while (!(next = ACCESS_ONCE(node->next)))
-			arch_mutex_cpu_relax();
-	}
-	ACCESS_ONCE(next->locked) = 1;
-	smp_wmb();
-}
 
 /*
  * Mutex spinning code migrated from kernel/sched/core.c

@@ -448,7 +402,7 @@ __mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
 
 	for (;;) {
 		struct task_struct *owner;
-		struct mspin_node  node;
+		struct mcs_spinlock  node;
 
 		if (use_ww_ctx && ww_ctx->acquired > 0) {
 			struct ww_mutex *ww;

@@ -470,10 +424,10 @@ __mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
 		 * If there's an owner, wait for it to either
 		 * release the lock or go to sleep.
 		 */
-		mspin_lock(MLOCK(lock), &node);
+		mcs_spin_lock(&lock->mcs_lock, &node);
 		owner = ACCESS_ONCE(lock->owner);
 		if (owner && !mutex_spin_on_owner(lock, owner)) {
-			mspin_unlock(MLOCK(lock), &node);
+			mcs_spin_unlock(&lock->mcs_lock, &node);
 			goto slowpath;
 		}
 

@@ -488,11 +442,11 @@ __mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
 			}
 
 			mutex_set_owner(lock);
-			mspin_unlock(MLOCK(lock), &node);
+			mcs_spin_unlock(&lock->mcs_lock, &node);
 			preempt_enable();
 			return 0;
 		}
-		mspin_unlock(MLOCK(lock), &node);
+		mcs_spin_unlock(&lock->mcs_lock, &node);
 

Quoted 7 lines

kernel/locking/mutex.c +0 -0

 		/*
 		 * When there's no owner, we might have preempted between the

Signature · 1 line

1.7.4.4

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Quoted 10 lines · Paul E. McKenney

Quoted 6 lines · Tim Chen

We will need the MCS lock code for doing optimistic spinning for rwsem and queue rwlock. Extracting the MCS code from mutex.c and put into its own file allow us to reuse this code easily.

Signed-off-by Tim Chen <[email protected]>

Signed-off-by Davidlohr Bueso <[email protected]>

Please see comments below.

Thanks for reviewing the code.

Quoted 75 lines

Thanx, Paul

Quoted 36 lines

---
 include/linux/mcs_spinlock.h |   64 ++++++++++++++++++++++++++++++++++++++++++
 include/linux/mutex.h        |    5 ++-
 kernel/locking/mutex.c       |   60 ++++----------------------------------
 3 files changed, 74 insertions(+), 55 deletions(-)
 create mode 100644 include/linux/mcs_spinlock.h

include/linux/mcs_spinlock.h new file +23 -0

new file mode 100644
index 0000000..b5de3b0
@@ -0,0 +1,64 @@
+/*
+ * MCS lock defines
+ *
+ * This file contains the main data structure and API definitions of MCS lock.
+ *
+ * The MCS lock (proposed by Mellor-Crummey and Scott) is a simple spin-lock
+ * with the desirable properties of being fair, and with each cpu trying
+ * to acquire the lock spinning on a local variable.
+ * It avoids expensive cache bouncings that common test-and-set spin-lock
+ * implementations incur.
+ */
+#ifndef __LINUX_MCS_SPINLOCK_H
+#define __LINUX_MCS_SPINLOCK_H
+
+struct mcs_spinlock {
+	struct mcs_spinlock *next;
+	int locked; /* 1 if lock acquired */
+};
+
+/*
+ * We don't inline mcs_spin_lock() so that perf can correctly account for the
+ * time spent in this lock function.
+ */

include/linux/mcs_spinlock.h line 33

Quoted 10 lines

include/linux/mcs_spinlock.h new file +10 -0

+static noinline
+void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
+{
+	struct mcs_spinlock *prev;
+
+	/* Init node */
+	node->locked = 0;
+	node->next   = NULL;
+
+	prev = xchg(lock, node);

OK, the full memory barriers implied by xchg() ensure that *node will be initialized before the "ACCESS_ONCE(prev->next) = node" below puts the node into the list. This rules out the misordering scenario that Tim Chen called out in message-id <1380322005.3467.186.camel@schen9-DESK> on September 27th.

Assuming of course a corresponding barrier on the lock handoff side.

include/linux/mcs_spinlock.h line 40

Quoted 7 lines

include/linux/mcs_spinlock.h new file +7 -0

+	if (likely(prev == NULL)) {
+		/* Lock acquired */
+		node->locked = 1;
+		return;
+	}
+	ACCESS_ONCE(prev->next) = node;
+	smp_wmb();

I don't see what the above memory barrier does. Here are some things that it cannot be doing:

o Ordering the insertion into the list above with the polling
below. First, smp_wmb() does not order prior writes against later reads, and second misordering is harmless. If we start polling before the insertion is complete, all that happens is that the first few polls have no chance of seeing a lock grant.

Quoted 4 lines

o Ordering the polling against the initialization -- the above
xchg() is already doing that for us.

So what is its purpose?

Agree that the smp_wmb is not needed. It is in the existing mcs code residing in mutex.c and we're re-factoring the code only here and hasn't corrected the memory barrier.

The particular smp_wmb() is removed in Patch 4/4 that corrects the memory barriers.

Quoted 11 lines

include/linux/mcs_spinlock.h line 43

Quoted 4 lines

include/linux/mcs_spinlock.h new file +3 -0

+	/* Wait until the lock holder passes the lock down */
+	while (!ACCESS_ONCE(node->locked))
+		arch_mutex_cpu_relax();

On the other hand, I don't see how we get away without a barrier here. As written, what prevents the caller's load from ->owner from being reordered with the above load from ->locked? (Perhaps you can argue that such reordering is only a performance problem, but if so we need that argument recorded in comments.)

Quoted 3 lines

Of course, if anyone ever tries to use mcs_spin_lock() as a full lock, they will need a memory barrier here to prevent the critical section from leaking out.

Agree too. The appropriate memory barrier is added in Patch 4/4.

Quoted 192 lines

Quoted 7 lines

include/linux/mcs_spinlock.h new file +6 -0

+}
+
+static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
+{
+	struct mcs_spinlock *next = ACCESS_ONCE(node->next);
+

include/linux/mcs_spinlock.h line 59

Quoted 10 lines

include/linux/mcs_spinlock.h new file +10 -0

+	if (likely(!next)) {
+		/*
+		 * Release the lock by setting it to NULL
+		 */
+		if (cmpxchg(lock, node, NULL) == node)
+			return;
+		/* Wait until the next pointer is set */
+		while (!(next = ACCESS_ONCE(node->next)))
+			arch_mutex_cpu_relax();
+	}

We need a memory barrier somewhere before here in this function, otherwise the critical section can leak out. I do not believe that we can rely on the prohibition against speculative stores that Peter Zijlstra and I have been discussing because that does not provide the transitivity required by locking primitives. I believe that we -could- make the access below be an smp_store_release(), though.

Placing the barrier here (or at least not preceding the initial fetch from node->next) has the advantage of allowing it to pair with the xchg() in mcs_spin_lock(), though given the dependency only an smp_read_barrier_depends() is required for that purpose.

include/linux/mcs_spinlock.h line 61

Quoted 2 lines

include/linux/mcs_spinlock.h new file +2 -0

+	ACCESS_ONCE(next->locked) = 1;
+	smp_wmb();

I don't see what this barrier does for us. It is ordering the unlock store with what, exactly?

If it really is doing something, we need a big fat comment stating what that is, and checkpatch.pl will be happy to inform you. ;-)

Quoted 147 lines

include/linux/mcs_spinlock.h new file +3 -0

+}
+
+#endif /* __LINUX_MCS_SPINLOCK_H */

include/linux/mutex.h +3 -2

index bab49da..32a32e6 100644
@@ -46,6 +46,7 @@
  * - detects multi-task circular deadlocks and prints out all affected
  *   locks and tasks (and only those tasks)
  */
+struct mcs_spinlock;
 struct mutex {
 	/* 1: unlocked, 0: locked, negative: locked, possible waiters */
 	atomic_t		count;

@@ -55,7 +56,7 @@ struct mutex {
 	struct task_struct	*owner;
 #endif
 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
-	void			*spin_mlock;	/* Spinner MCS lock */
+	struct mcs_spinlock	*mcs_lock;	/* Spinner MCS lock */
 #endif
 #ifdef CONFIG_DEBUG_MUTEXES
 	const char 		*name;

@@ -179,4 +180,4 @@ extern int atomic_dec_and_mutex_lock(atomic_t *cnt, struct mutex *lock);
 # define arch_mutex_cpu_relax() cpu_relax()
 #endif
 
-#endif
+#endif /* __LINUX_MUTEX_H */

kernel/locking/mutex.c +7 -53

index d24105b..e08b183 100644
@@ -25,6 +25,7 @@
 #include <linux/spinlock.h>
 #include <linux/interrupt.h>
 #include <linux/debug_locks.h>
+#include <linux/mcs_spinlock.h>
 
 /*
  * In the DEBUG case we are using the "NULL fastpath" for mutexes,

@@ -52,7 +53,7 @@ __mutex_init(struct mutex *lock, const char *name, struct lock_class_key *key)
 	INIT_LIST_HEAD(&lock->wait_list);
 	mutex_clear_owner(lock);
 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
-	lock->spin_mlock = NULL;
+	lock->mcs_lock = NULL;
 #endif
 
 	debug_mutex_init(lock, name, key);

@@ -111,54 +112,7 @@ EXPORT_SYMBOL(mutex_lock);
  * more or less simultaneously, the spinners need to acquire a MCS lock
  * first before spinning on the owner field.
  *
- * We don't inline mspin_lock() so that perf can correctly account for the
- * time spent in this lock function.
  */
-struct mspin_node {
-	struct mspin_node *next ;
-	int		  locked;	/* 1 if lock acquired */
-};
-#define	MLOCK(mutex)	((struct mspin_node **)&((mutex)->spin_mlock))
-
-static noinline
-void mspin_lock(struct mspin_node **lock, struct mspin_node *node)
-{
-	struct mspin_node *prev;
-
-	/* Init node */
-	node->locked = 0;
-	node->next   = NULL;
-
-	prev = xchg(lock, node);
-	if (likely(prev == NULL)) {
-		/* Lock acquired */
-		node->locked = 1;
-		return;
-	}
-	ACCESS_ONCE(prev->next) = node;
-	smp_wmb();
-	/* Wait until the lock holder passes the lock down */
-	while (!ACCESS_ONCE(node->locked))
-		arch_mutex_cpu_relax();
-}
-
-static void mspin_unlock(struct mspin_node **lock, struct mspin_node *node)
-{
-	struct mspin_node *next = ACCESS_ONCE(node->next);
-
-	if (likely(!next)) {
-		/*
-		 * Release the lock by setting it to NULL
-		 */
-		if (cmpxchg(lock, node, NULL) == node)
-			return;
-		/* Wait until the next pointer is set */
-		while (!(next = ACCESS_ONCE(node->next)))
-			arch_mutex_cpu_relax();
-	}
-	ACCESS_ONCE(next->locked) = 1;
-	smp_wmb();
-}
 
 /*
  * Mutex spinning code migrated from kernel/sched/core.c

@@ -448,7 +402,7 @@ __mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
 
 	for (;;) {
 		struct task_struct *owner;
-		struct mspin_node  node;
+		struct mcs_spinlock  node;
 
 		if (use_ww_ctx && ww_ctx->acquired > 0) {
 			struct ww_mutex *ww;

@@ -470,10 +424,10 @@ __mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
 		 * If there's an owner, wait for it to either
 		 * release the lock or go to sleep.
 		 */
-		mspin_lock(MLOCK(lock), &node);
+		mcs_spin_lock(&lock->mcs_lock, &node);
 		owner = ACCESS_ONCE(lock->owner);
 		if (owner && !mutex_spin_on_owner(lock, owner)) {
-			mspin_unlock(MLOCK(lock), &node);
+			mcs_spin_unlock(&lock->mcs_lock, &node);
 			goto slowpath;
 		}
 

@@ -488,11 +442,11 @@ __mutex_lock_common(struct mutex *lock, long state, unsigned int subclass,
 			}
 
 			mutex_set_owner(lock);
-			mspin_unlock(MLOCK(lock), &node);
+			mcs_spin_unlock(&lock->mcs_lock, &node);
 			preempt_enable();
 			return 0;
 		}
-		mspin_unlock(MLOCK(lock), &node);
+		mcs_spin_unlock(&lock->mcs_lock, &node);
 
 		/*
 		 * When there's no owner, we might have preempted between the

Signature · 1 line

1.7.4.4

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

From: Jason Low <[email protected]>

Remove unnecessary operation and make the cmpxchg(lock, node, NULL) == node check in mcs_spin_unlock() likely() as it is likely that a race did not occur most of the time.

Also add in more comments describing how the local node is used in MCS locks.

Reviewed-by Tim Chen <[email protected]>

Signed-off-by Jason Low <[email protected]>

---
 include/linux/mcs_spinlock.h |   13 +++++++++++--
 1 files changed, 11 insertions(+), 2 deletions(-)

include/linux/mcs_spinlock.h +6 -0

index b5de3b0..96f14299 100644
@@ -18,6 +18,12 @@ struct mcs_spinlock {
 };
 
 /*
+ * In order to acquire the lock, the caller should declare a local node and
+ * pass a reference of the node to this function in addition to the lock.
+ * If the lock has already been acquired, then this will proceed to spin
+ * on this node->locked until the previous lock holder sets the node->locked
+ * in mcs_spin_unlock().
+ *
  * We don't inline mcs_spin_lock() so that perf can correctly account for the
  * time spent in this lock function.
  */

include/linux/mcs_spinlock.h +4 -0

 		return;
 	}
 	ACCESS_ONCE(prev->next) = node;

@@ -43,6 +48,10 @@ void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 		arch_mutex_cpu_relax();
 }
 
+/*
+ * Releases the lock. The caller should pass in the corresponding node that
+ * was used to acquire the lock.
+ */
 static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 {
 	struct mcs_spinlock *next = ACCESS_ONCE(node->next);

include/linux/mcs_spinlock.h +0 -0

 			return;
 		/* Wait until the next pointer is set */
 		while (!(next = ACCESS_ONCE(node->next)))

Signature · 1 line

1.7.4.4

Quoted 32 lines · Tim Chen

From: Jason Low <[email protected]>

Remove unnecessary operation and make the cmpxchg(lock, node, NULL) == node check in mcs_spin_unlock() likely() as it is likely that a race did not occur most of the time.

Also add in more comments describing how the local node is used in MCS locks.

Reviewed-by Tim Chen <[email protected]>

Signed-off-by Jason Low <[email protected]>

---
 include/linux/mcs_spinlock.h |   13 +++++++++++--
 1 files changed, 11 insertions(+), 2 deletions(-)

include/linux/mcs_spinlock.h +6 -0

index b5de3b0..96f14299 100644
@@ -18,6 +18,12 @@ struct mcs_spinlock {
 };
 
 /*
+ * In order to acquire the lock, the caller should declare a local node and
+ * pass a reference of the node to this function in addition to the lock.
+ * If the lock has already been acquired, then this will proceed to spin
+ * on this node->locked until the previous lock holder sets the node->locked
+ * in mcs_spin_unlock().
+ *
  * We don't inline mcs_spin_lock() so that perf can correctly account for the
  * time spent in this lock function.
  */

include/linux/mcs_spinlock.h void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node) · line 36

Quoted 4 lines

include/linux/mcs_spinlock.h +0 -1

@@ -33,7 +39,6 @@ void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 	prev = xchg(lock, node);
 	if (likely(prev == NULL)) {
 		/* Lock acquired */
-		node->locked = 1;

Agreed, no one looks at this field in this case, so no need to initialize it, unless for debug purposes.

Quoted 15 lines

include/linux/mcs_spinlock.h +4 -0

 		return;
 	}
 	ACCESS_ONCE(prev->next) = node;

@@ -43,6 +48,10 @@ void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 		arch_mutex_cpu_relax();
 }
 
+/*
+ * Releases the lock. The caller should pass in the corresponding node that
+ * was used to acquire the lock.
+ */
 static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 {
 	struct mcs_spinlock *next = ACCESS_ONCE(node->next);

include/linux/mcs_spinlock.h static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *nod · line 63

Quoted 5 lines

include/linux/mcs_spinlock.h +1 -1

@@ -51,7 +60,7 @@ static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *nod
 		/*
 		 * Release the lock by setting it to NULL
 		 */
-		if (cmpxchg(lock, node, NULL) == node)
+		if (likely(cmpxchg(lock, node, NULL) == node))

Agreed here as well. Takes a narrow race to hit this.

So, did your testing exercise this path? If the answer is "yes", and if the issues that I called out in patch #1 are resolved:

Reviewed-by Paul E. McKenney <[email protected]>

Quoted 8 lines

include/linux/mcs_spinlock.h +0 -0

 			return;
 		/* Wait until the next pointer is set */
 		while (!(next = ACCESS_ONCE(node->next)))

Signature · 1 line

1.7.4.4

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Quoted 9 lines · Tim Chen

include/linux/mcs_spinlock.h +4 -0

+/*
+ * Releases the lock. The caller should pass in the corresponding node that
+ * was used to acquire the lock.
+ */
 static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 {
 	struct mcs_spinlock *next = ACCESS_ONCE(node->next);

include/linux/mcs_spinlock.h static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *nod · line 63

Quoted 5 lines

include/linux/mcs_spinlock.h +1 -1

@@ -51,7 +60,7 @@ static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *nod
 		/*
 		 * Release the lock by setting it to NULL
 		 */
-		if (cmpxchg(lock, node, NULL) == node)
+		if (likely(cmpxchg(lock, node, NULL) == node))

Agreed here as well. Takes a narrow race to hit this.

So, did your testing exercise this path? If the answer is "yes",

and if the issues that I called out in patch #1 are resolved:

Reviewed-by Paul E. McKenney <[email protected]>

Quoted 4 lines

include/linux/mcs_spinlock.h +0 -0

 			return;
 		/* Wait until the next pointer is set */
 		while (!(next = ACCESS_ONCE(node->next)))

Quoted 18 lines · Paul E. McKenney

Quoted 9 lines · Tim Chen

include/linux/mcs_spinlock.h +4 -0

+/*
+ * Releases the lock. The caller should pass in the corresponding node that
+ * was used to acquire the lock.
+ */
 static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 {
 	struct mcs_spinlock *next = ACCESS_ONCE(node->next);

include/linux/mcs_spinlock.h static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *nod · line 63

Quoted 5 lines

include/linux/mcs_spinlock.h +1 -1

@@ -51,7 +60,7 @@ static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *nod
 		/*
 		 * Release the lock by setting it to NULL
 		 */
-		if (cmpxchg(lock, node, NULL) == node)
+		if (likely(cmpxchg(lock, node, NULL) == node))

Agreed here as well. Takes a narrow race to hit this.

Quoted 1 line

So, did your testing exercise this path? If the answer is "yes",

Paul,

I did some instrumentation and confirmed that the path in question has been exercised. So this patch should be okay.

From: Waiman Long <[email protected]>

The following changes are made:

1) Create a new mcs_spinlock.c file to contain the mcs_spin_lock() and mcs_spin_unlock() function.
2) Include a number of prerequisite header files and define arch_mutex_cpu_relax(), if not previously defined so the mcs functions can be compiled for multiple architecture without causing problems.

Signed-off-by Waiman Long <[email protected]>

Signed-off-by Tim Chen <[email protected]>

---
 include/linux/mcs_spinlock.h                       |   56 ++------------------
 kernel/locking/Makefile                            |    6 +-
 .../locking/mcs_spinlock.c                         |   31 ++++++-----
 3 files changed, 23 insertions(+), 70 deletions(-)
 copy include/linux/mcs_spinlock.h => kernel/locking/mcs_spinlock.c (78%)

include/linux/mcs_spinlock.h +4 -52

index 96f14299..d54bb23 100644
@@ -17,57 +17,9 @@ struct mcs_spinlock {
 	int locked; /* 1 if lock acquired */
 };
 
-/*
- * In order to acquire the lock, the caller should declare a local node and
- * pass a reference of the node to this function in addition to the lock.
- * If the lock has already been acquired, then this will proceed to spin
- * on this node->locked until the previous lock holder sets the node->locked
- * in mcs_spin_unlock().
- *
- * We don't inline mcs_spin_lock() so that perf can correctly account for the
- * time spent in this lock function.
- */
-static noinline
-void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
-{
-	struct mcs_spinlock *prev;
-
-	/* Init node */
-	node->locked = 0;
-	node->next   = NULL;
-
-	prev = xchg(lock, node);
-	if (likely(prev == NULL)) {
-		/* Lock acquired */
-		return;
-	}
-	ACCESS_ONCE(prev->next) = node;
-	smp_wmb();
-	/* Wait until the lock holder passes the lock down */
-	while (!ACCESS_ONCE(node->locked))
-		arch_mutex_cpu_relax();
-}
-
-/*
- * Releases the lock. The caller should pass in the corresponding node that
- * was used to acquire the lock.
- */
-static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
-{
-	struct mcs_spinlock *next = ACCESS_ONCE(node->next);
-
-	if (likely(!next)) {
-		/*
-		 * Release the lock by setting it to NULL
-		 */
-		if (likely(cmpxchg(lock, node, NULL) == node))
-			return;
-		/* Wait until the next pointer is set */
-		while (!(next = ACCESS_ONCE(node->next)))
-			arch_mutex_cpu_relax();
-	}
-	ACCESS_ONCE(next->locked) = 1;
-	smp_wmb();
-}
+extern
+void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node);
+extern
+void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *node);
 
 #endif /* __LINUX_MCS_SPINLOCK_H */

kernel/locking/Makefile +3 -3

index baab8e5..20d9d5c 100644
@@ -13,12 +13,12 @@ obj-$(CONFIG_LOCKDEP) += lockdep.o
 ifeq ($(CONFIG_PROC_FS),y)
 obj-$(CONFIG_LOCKDEP) += lockdep_proc.o
 endif
-obj-$(CONFIG_SMP) += spinlock.o
-obj-$(CONFIG_PROVE_LOCKING) += spinlock.o
+obj-$(CONFIG_SMP) += spinlock.o mcs_spinlock.o
+obj-$(CONFIG_PROVE_LOCKING) += spinlock.o mcs_spinlock.o
 obj-$(CONFIG_RT_MUTEXES) += rtmutex.o
 obj-$(CONFIG_DEBUG_RT_MUTEXES) += rtmutex-debug.o
 obj-$(CONFIG_RT_MUTEX_TESTER) += rtmutex-tester.o
-obj-$(CONFIG_DEBUG_SPINLOCK) += spinlock.o
+obj-$(CONFIG_DEBUG_SPINLOCK) += spinlock.o mcs_spinlock.o
 obj-$(CONFIG_DEBUG_SPINLOCK) += spinlock_debug.o
 obj-$(CONFIG_RWSEM_GENERIC_SPINLOCK) += rwsem-spinlock.o
 obj-$(CONFIG_RWSEM_XCHGADD_ALGORITHM) += rwsem-xadd.o

kernel/locking/mcs_spinlock.c renamed from include/linux/mcs_spinlock.h +16 -15

similarity index 78%
copy from include/linux/mcs_spinlock.h
copy to kernel/locking/mcs_spinlock.c
index 96f14299..b6f27f8 100644
@@ -1,7 +1,5 @@
 /*
- * MCS lock defines
- *
- * This file contains the main data structure and API definitions of MCS lock.
+ * MCS lock
  *
  * The MCS lock (proposed by Mellor-Crummey and Scott) is a simple spin-lock
  * with the desirable properties of being fair, and with each cpu trying

@@ -9,13 +7,20 @@
  * It avoids expensive cache bouncings that common test-and-set spin-lock
  * implementations incur.
  */
-#ifndef __LINUX_MCS_SPINLOCK_H
-#define __LINUX_MCS_SPINLOCK_H
+/*
+ * asm/processor.h may define arch_mutex_cpu_relax().
+ * If it is not defined, cpu_relax() will be used.
+ */
+#include <asm/barrier.h>
+#include <asm/cmpxchg.h>
+#include <asm/processor.h>
+#include <linux/compiler.h>
+#include <linux/mcs_spinlock.h>
+#include <linux/export.h>
 
-struct mcs_spinlock {
-	struct mcs_spinlock *next;
-	int locked; /* 1 if lock acquired */
-};
+#ifndef arch_mutex_cpu_relax
+# define arch_mutex_cpu_relax() cpu_relax()
+#endif
 
 /*
  * In order to acquire the lock, the caller should declare a local node and

@@ -23,11 +28,7 @@ struct mcs_spinlock {
  * If the lock has already been acquired, then this will proceed to spin
  * on this node->locked until the previous lock holder sets the node->locked
  * in mcs_spin_unlock().
- *
- * We don't inline mcs_spin_lock() so that perf can correctly account for the
- * time spent in this lock function.
  */
-static noinline
 void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 {
 	struct mcs_spinlock *prev;

@@ -47,6 +48,7 @@ void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 	while (!ACCESS_ONCE(node->locked))
 		arch_mutex_cpu_relax();
 }
+EXPORT_SYMBOL_GPL(mcs_spin_lock);
 
 /*
  * Releases the lock. The caller should pass in the corresponding node that

@@ -69,5 +71,4 @@ static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *nod
 	ACCESS_ONCE(next->locked) = 1;
 	smp_wmb();
 }
-
-#endif /* __LINUX_MCS_SPINLOCK_H */
+EXPORT_SYMBOL_GPL(mcs_spin_unlock);

Signature · 1 line

1.7.4.4

From: Waiman Long <[email protected]>

This patch corrects the way memory barriers are used in the MCS lock with smp_load_acquire and smp_store_release fucnction.
It removes ones that are not needed.

It uses architecture specific load-acquire and store-release primitives for synchronization, if available. Generic implementations are provided in case they are not defined even though they may not be optimal. These generic implementation could be removed later on once changes are made in all the relevant header files.

Suggested-by Michel Lespinasse <[email protected]>

Signed-off-by Waiman Long <[email protected]>

Signed-off-by Jason Low <[email protected]>

Signed-off-by Tim Chen <[email protected]>

---
 kernel/locking/mcs_spinlock.c |   48 +++++++++++++++++++++++++++++++++++------
 1 files changed, 41 insertions(+), 7 deletions(-)

kernel/locking/mcs_spinlock.c +13 -0

index b6f27f8..df5c167 100644
@@ -23,6 +23,31 @@
 #endif
 
 /*
+ * Fall back to use the regular atomic operations and memory barrier if
+ * the acquire/release versions are not defined.
+ */
+#ifndef	xchg_acquire
+# define xchg_acquire(p, v)		xchg(p, v)
+#endif
+
+#ifndef	smp_load_acquire
+# define smp_load_acquire(p)				\
+	({						\
+		typeof(*p) __v = ACCESS_ONCE(*(p));	\
+		smp_mb();				\
+		__v;					\

kernel/locking/mcs_spinlock.c +3 -1

+/*
  * In order to acquire the lock, the caller should declare a local node and
  * pass a reference of the node to this function in addition to the lock.
  * If the lock has already been acquired, then this will proceed to spin

@@ -37,15 +62,19 @@ void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 	node->locked = 0;
 	node->next   = NULL;
 
-	prev = xchg(lock, node);
+	/* xchg() provides a memory barrier */
+	prev = xchg_acquire(lock, node);
 	if (likely(prev == NULL)) {
 		/* Lock acquired */
 		return;
 	}
 	ACCESS_ONCE(prev->next) = node;

Quoted 4 lines

kernel/locking/mcs_spinlock.c +4 -0

+	/*
+	 * Wait until the lock holder passes the lock down.
+	 * Using smp_load_acquire() provides a memory barrier that
+	 * ensures subsequent operations happen after the lock is acquired.

kernel/locking/mcs_spinlock.c void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node) · line 78

Quoted 3 lines

kernel/locking/mcs_spinlock.c +2 -0

+	 */
+	while (!(smp_load_acquire(&node->locked)))
 		arch_mutex_cpu_relax();

Thinking more about that, the real issue here is that arch_mutex_cpu_relax() doesn't have a corresponding hook on the unlock side. On ARM, for example, we can enter a low-power state using the wfe instruction, but that requires the unlocker to wake up the core when the lock is released.

Quoted 4 lines · Will Deacon

Quoted 4 lines

kernel/locking/mcs_spinlock.c +4 -0

+	/*
+	 * Wait until the lock holder passes the lock down.
+	 * Using smp_load_acquire() provides a memory barrier that
+	 * ensures subsequent operations happen after the lock is acquired.

Quoted 8 lines

kernel/locking/mcs_spinlock.c void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node) · line 78

Quoted 3 lines

kernel/locking/mcs_spinlock.c +2 -0

+	 */
+	while (!(smp_load_acquire(&node->locked)))
 		arch_mutex_cpu_relax();

Thinking more about that, the real issue here is that arch_mutex_cpu_relax() doesn't have a corresponding hook on the unlock side. On ARM, for example, we can enter a low-power state using the wfe instruction, but that requires the unlocker to wake up the core when the lock is released.

That said, it would be ever so awesome if we could come to some sort of conclusion on control dependencies here.

I _know_ C/C++ doesn't do them, but at the end of the day the compiler still generates ASM and as long as we're relatively certain there's a loop there (there has to be, right? :-), we could maybe rely on it anyway.

Hello,

Quoted 46 lines · Tim Chen

From: Waiman Long <[email protected]>

This patch corrects the way memory barriers are used in the MCS lock with smp_load_acquire and smp_store_release fucnction.
It removes ones that are not needed.

It uses architecture specific load-acquire and store-release primitives for synchronization, if available. Generic implementations are provided in case they are not defined even though they may not be optimal. These generic implementation could be removed later on once changes are made in all the relevant header files.

Suggested-by Michel Lespinasse <[email protected]>

Signed-off-by Waiman Long <[email protected]>

Signed-off-by Jason Low <[email protected]>

Signed-off-by Tim Chen <[email protected]>

---
 kernel/locking/mcs_spinlock.c |   48 +++++++++++++++++++++++++++++++++++------
 1 files changed, 41 insertions(+), 7 deletions(-)

kernel/locking/mcs_spinlock.c +18 -0

index b6f27f8..df5c167 100644
@@ -23,6 +23,31 @@
 #endif
 
 /*
+ * Fall back to use the regular atomic operations and memory barrier if
+ * the acquire/release versions are not defined.
+ */
+#ifndef	xchg_acquire
+# define xchg_acquire(p, v)		xchg(p, v)
+#endif
+
+#ifndef	smp_load_acquire
+# define smp_load_acquire(p)				\
+	({						\
+		typeof(*p) __v = ACCESS_ONCE(*(p));	\
+		smp_mb();				\
+		__v;					\
+	})
+#endif
+
+#ifndef smp_store_release
+# define smp_store_release(p, v)		\

kernel/locking/mcs_spinlock.c line 48

Quoted 5 lines

kernel/locking/mcs_spinlock.c +5 -0

+	do {					\
+		smp_mb();			\
+		ACCESS_ONCE(*(p)) = v;		\
+	} while (0)
+#endif

PeterZ already has a series implementing acquire/release accessors, so you should probably take a look at that rather than rolling your own here.

You could then augment that with [cmp]xchg_{acquire,release} as appropriate.

Quoted 18 lines

kernel/locking/mcs_spinlock.c +3 -3

+/*
  * In order to acquire the lock, the caller should declare a local node and
  * pass a reference of the node to this function in addition to the lock.
  * If the lock has already been acquired, then this will proceed to spin

@@ -37,15 +62,19 @@ void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 	node->locked = 0;
 	node->next   = NULL;
 
-	prev = xchg(lock, node);
+	/* xchg() provides a memory barrier */
+	prev = xchg_acquire(lock, node);
 	if (likely(prev == NULL)) {
 		/* Lock acquired */
 		return;
 	}
 	ACCESS_ONCE(prev->next) = node;
-	smp_wmb();
-	/* Wait until the lock holder passes the lock down */

kernel/locking/mcs_spinlock.c void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node) · line 78

Quoted 8 lines

kernel/locking/mcs_spinlock.c +6 -1

-	while (!ACCESS_ONCE(node->locked))
+	/*
+	 * Wait until the lock holder passes the lock down.
+	 * Using smp_load_acquire() provides a memory barrier that
+	 * ensures subsequent operations happen after the lock is acquired.
+	 */
+	while (!(smp_load_acquire(&node->locked)))
 		arch_mutex_cpu_relax();

After a chat with some micro-architects, I'm going to have to disagree with Paul here. For architectures where acquire/release are implemented with explicit barriers (similarly for simple microarchitectures), emitting barriers in a loop *is* going to have an affect on overall performance, since those barriers may well result in traffic outside of the core (at least, on ARM).

Thinking more about that, the real issue here is that arch_mutex_cpu_relax() doesn't have a corresponding hook on the unlock side. On ARM, for example, we can enter a low-power state using the wfe instruction, but that requires the unlocker to wake up the core when the lock is released.

So, although I'm completely in favour of introducing acquire/release accessors, I really think the mcs locking routines would benefit from some arch-specific backend code, even if it's optional (although I would imagine most architectures implementing something to improve power and/or performance).

Will

Quoted 21 lines

You could then augment that with [cmp]xchg_{acquire,release} as appropriate.

Quoted 18 lines

kernel/locking/mcs_spinlock.c +3 -3

+/*
   * In order to acquire the lock, the caller should declare a local node and
   * pass a reference of the node to this function in addition to the lock.
   * If the lock has already been acquired, then this will proceed to spin

@@ -37,15 +62,19 @@ void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
  	node->locked = 0;
  	node->next   = NULL;
 
-	prev = xchg(lock, node);
+	/* xchg() provides a memory barrier */
+	prev = xchg_acquire(lock, node);
  	if (likely(prev == NULL)) {
  		/* Lock acquired */
  		return;
  	}
  	ACCESS_ONCE(prev->next) = node;
-	smp_wmb();
-	/* Wait until the lock holder passes the lock down */

Quoted 8 lines

kernel/locking/mcs_spinlock.c void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node) · line 78

Quoted 8 lines

kernel/locking/mcs_spinlock.c +6 -1

-	while (!ACCESS_ONCE(node->locked))
+	/*
+	 * Wait until the lock holder passes the lock down.
+	 * Using smp_load_acquire() provides a memory barrier that
+	 * ensures subsequent operations happen after the lock is acquired.
+	 */
+	while (!(smp_load_acquire(&node->locked)))
  		arch_mutex_cpu_relax();

An alternate implementation is

	while (!ACCESS_ONCE(node->locked))
		arch_mutex_cpu_relax();
	smp_load_acquire(&node->locked);

Leaving the smp_load_acquire at the end to provide appropriate barrier. Will that be acceptable?

Tim

Quoted 37 lines · Tim Chen

Quoted 21 lines

You could then augment that with [cmp]xchg_{acquire,release} as appropriate.

Quoted 18 lines

kernel/locking/mcs_spinlock.c +3 -3

+/*
  * In order to acquire the lock, the caller should declare a local node and
  * pass a reference of the node to this function in addition to the lock.
  * If the lock has already been acquired, then this will proceed to spin

@@ -37,15 +62,19 @@ void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 	node->locked = 0;
 	node->next   = NULL;
 
-	prev = xchg(lock, node);
+	/* xchg() provides a memory barrier */
+	prev = xchg_acquire(lock, node);
 	if (likely(prev == NULL)) {
 		/* Lock acquired */
 		return;
 	}
 	ACCESS_ONCE(prev->next) = node;
-	smp_wmb();
-	/* Wait until the lock holder passes the lock down */

Quoted 8 lines

kernel/locking/mcs_spinlock.c void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node) · line 78

Quoted 8 lines

kernel/locking/mcs_spinlock.c +6 -1

-	while (!ACCESS_ONCE(node->locked))
+	/*
+	 * Wait until the lock holder passes the lock down.
+	 * Using smp_load_acquire() provides a memory barrier that
+	 * ensures subsequent operations happen after the lock is acquired.
+	 */
+	while (!(smp_load_acquire(&node->locked)))
 		arch_mutex_cpu_relax();

An alternate implementation is

	while (!ACCESS_ONCE(node->locked))
		arch_mutex_cpu_relax();
	smp_load_acquire(&node->locked);

Leaving the smp_load_acquire at the end to provide appropriate barrier. Will that be acceptable?

Quoted 1 line

Tim

I second Tim's opinion. It will be help to have a smp_mb_load_acquire() function that provide a memory barrier with load-acquire semantic. I don't think we need one for store-release as that will not be in a loop.

Peter, what do you think about adding that to your patch?

-Longman

Quoted 38 lines · Waiman Long

Quoted 37 lines · Tim Chen

Quoted 21 lines

You could then augment that with [cmp]xchg_{acquire,release} as appropriate.

Quoted 18 lines

kernel/locking/mcs_spinlock.c +3 -3

+/*
  * In order to acquire the lock, the caller should declare a local node and
  * pass a reference of the node to this function in addition to the lock.
  * If the lock has already been acquired, then this will proceed to spin

@@ -37,15 +62,19 @@ void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 	node->locked = 0;
 	node->next   = NULL;
 
-	prev = xchg(lock, node);
+	/* xchg() provides a memory barrier */
+	prev = xchg_acquire(lock, node);
 	if (likely(prev == NULL)) {
 		/* Lock acquired */
 		return;
 	}
 	ACCESS_ONCE(prev->next) = node;
-	smp_wmb();
-	/* Wait until the lock holder passes the lock down */

Quoted 8 lines

kernel/locking/mcs_spinlock.c void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node) · line 78

Quoted 8 lines

kernel/locking/mcs_spinlock.c +6 -1

-	while (!ACCESS_ONCE(node->locked))
+	/*
+	 * Wait until the lock holder passes the lock down.
+	 * Using smp_load_acquire() provides a memory barrier that
+	 * ensures subsequent operations happen after the lock is acquired.
+	 */
+	while (!(smp_load_acquire(&node->locked)))
 		arch_mutex_cpu_relax();

An alternate implementation is

	while (!ACCESS_ONCE(node->locked))
		arch_mutex_cpu_relax();
	smp_load_acquire(&node->locked);

Leaving the smp_load_acquire at the end to provide appropriate barrier. Will that be acceptable?

Quoted 6 lines

Quoted 1 line

Tim

I second Tim's opinion. It will be help to have a smp_mb_load_acquire() function that provide a memory barrier with load-acquire semantic. I don't think we need one for store-release as that will not be in a loop.

Hmmm... I guess the ACCESS_ONCE() in the smp_load_acquire() should prevent it from being optimized away. But yes, you then end up with an extra load on the critical lock hand-off patch. And something like an smp_mb_acquire() could then be useful, although I believe that on all current hardware smp_mb_acquire() emits the same code as would an smp_mb_release():

o barrier() on TSO systems such as x86 and s390.

o lwsync instruction on powerpc. (Really old systems would
want a different instruction for smp_mb_acquire(), but let's not optimize for really old systems.)

o dmb instruction on ARM.

o mf instruction on ia64.

So how about an smp_mb_acquire_release() to cover both use cases?

This could be used to further optimize circular buffers, for example.

Thanx, Paul

Quoted 4 lines

Peter, what do you think about adding that to your patch?

-Longman

Quoted 63 lines · Paul E. McKenney

Quoted 38 lines · Waiman Long

Quoted 37 lines · Tim Chen

Quoted 21 lines

You could then augment that with [cmp]xchg_{acquire,release} as appropriate.

Quoted 18 lines

kernel/locking/mcs_spinlock.c +3 -3

+/*
  * In order to acquire the lock, the caller should declare a local node and
  * pass a reference of the node to this function in addition to the lock.
  * If the lock has already been acquired, then this will proceed to spin

@@ -37,15 +62,19 @@ void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 	node->locked = 0;
 	node->next   = NULL;
 
-	prev = xchg(lock, node);
+	/* xchg() provides a memory barrier */
+	prev = xchg_acquire(lock, node);
 	if (likely(prev == NULL)) {
 		/* Lock acquired */
 		return;
 	}
 	ACCESS_ONCE(prev->next) = node;
-	smp_wmb();
-	/* Wait until the lock holder passes the lock down */

Quoted 8 lines

kernel/locking/mcs_spinlock.c void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node) · line 78

Quoted 8 lines

kernel/locking/mcs_spinlock.c +6 -1

-	while (!ACCESS_ONCE(node->locked))
+	/*
+	 * Wait until the lock holder passes the lock down.
+	 * Using smp_load_acquire() provides a memory barrier that
+	 * ensures subsequent operations happen after the lock is acquired.
+	 */
+	while (!(smp_load_acquire(&node->locked)))
 		arch_mutex_cpu_relax();

An alternate implementation is

	while (!ACCESS_ONCE(node->locked))
		arch_mutex_cpu_relax();
	smp_load_acquire(&node->locked);

Leaving the smp_load_acquire at the end to provide appropriate barrier. Will that be acceptable?

Quoted 6 lines

Quoted 1 line

Tim

I second Tim's opinion. It will be help to have a smp_mb_load_acquire() function that provide a memory barrier with load-acquire semantic. I don't think we need one for store-release as that will not be in a loop.

Hmmm... I guess the ACCESS_ONCE() in the smp_load_acquire() should prevent it from being optimized away. But yes, you then end up with an extra load on the critical lock hand-off patch. And something like an smp_mb_acquire() could then be useful, although I believe that on all current hardware smp_mb_acquire() emits the same code as would an smp_mb_release():

o barrier() on TSO systems such as x86 and s390.

o lwsync instruction on powerpc. (Really old systems would
want a different instruction for smp_mb_acquire(), but let's not optimize for really old systems.)

o dmb instruction on ARM.

o mf instruction on ia64.

Quoted 1 line

So how about an smp_mb_acquire_release() to cover both use cases?

I guess we haven't addressed Will's preference to use wfe for ARM instead of doing a spin loop. I'll like some suggestions on how to proceed here. Should we do arch_mcs_lock and arch_mcs_unlock, which defaults to the existing mcs_lock and mcs_unlock code, but allow architecture specific implementation?

Quoted 21 lines

You could then augment that with [cmp]xchg_{acquire,release} as appropriate.

Quoted 18 lines

kernel/locking/mcs_spinlock.c +3 -3

+/*
   * In order to acquire the lock, the caller should declare a local node and
   * pass a reference of the node to this function in addition to the lock.
   * If the lock has already been acquired, then this will proceed to spin

@@ -37,15 +62,19 @@ void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
  	node->locked = 0;
  	node->next   = NULL;
 
-	prev = xchg(lock, node);
+	/* xchg() provides a memory barrier */
+	prev = xchg_acquire(lock, node);
  	if (likely(prev == NULL)) {
  		/* Lock acquired */
  		return;
  	}
  	ACCESS_ONCE(prev->next) = node;
-	smp_wmb();
-	/* Wait until the lock holder passes the lock down */

Quoted 8 lines

kernel/locking/mcs_spinlock.c void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node) · line 78

Quoted 8 lines

kernel/locking/mcs_spinlock.c +6 -1

-	while (!ACCESS_ONCE(node->locked))
+	/*
+	 * Wait until the lock holder passes the lock down.
+	 * Using smp_load_acquire() provides a memory barrier that
+	 * ensures subsequent operations happen after the lock is acquired.
+	 */
+	while (!(smp_load_acquire(&node->locked)))
  		arch_mutex_cpu_relax();

An alternate implementation is

	while (!ACCESS_ONCE(node->locked))
		arch_mutex_cpu_relax();
	smp_load_acquire(&node->locked);

Leaving the smp_load_acquire at the end to provide appropriate barrier. Will that be acceptable?

Tim

Quoted 21 lines

You could then augment that with [cmp]xchg_{acquire,release} as appropriate.

Quoted 18 lines

kernel/locking/mcs_spinlock.c +3 -3

+/*
   * In order to acquire the lock, the caller should declare a local node and
   * pass a reference of the node to this function in addition to the lock.
   * If the lock has already been acquired, then this will proceed to spin

@@ -37,15 +62,19 @@ void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
  	node->locked = 0;
  	node->next   = NULL;
 
-	prev = xchg(lock, node);
+	/* xchg() provides a memory barrier */
+	prev = xchg_acquire(lock, node);
  	if (likely(prev == NULL)) {
  		/* Lock acquired */
  		return;
  	}
  	ACCESS_ONCE(prev->next) = node;
-	smp_wmb();
-	/* Wait until the lock holder passes the lock down */

Quoted 8 lines

kernel/locking/mcs_spinlock.c void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node) · line 78

Quoted 8 lines

kernel/locking/mcs_spinlock.c +6 -1

-	while (!ACCESS_ONCE(node->locked))
+	/*
+	 * Wait until the lock holder passes the lock down.
+	 * Using smp_load_acquire() provides a memory barrier that
+	 * ensures subsequent operations happen after the lock is acquired.
+	 */
+	while (!(smp_load_acquire(&node->locked)))
  		arch_mutex_cpu_relax();

An alternate implementation is

	while (!ACCESS_ONCE(node->locked))
		arch_mutex_cpu_relax();
	smp_load_acquire(&node->locked);

Leaving the smp_load_acquire at the end to provide appropriate barrier. Will that be acceptable?

Tim

Quoted 27 lines · Will Deacon

Quoted 26 lines · Tim Chen

kernel/locking/mcs_spinlock.c +18 -0

index b6f27f8..df5c167 100644
@@ -23,6 +23,31 @@
 #endif
 
 /*
+ * Fall back to use the regular atomic operations and memory barrier if
+ * the acquire/release versions are not defined.
+ */
+#ifndef	xchg_acquire
+# define xchg_acquire(p, v)		xchg(p, v)
+#endif
+
+#ifndef	smp_load_acquire
+# define smp_load_acquire(p)				\
+	({						\
+		typeof(*p) __v = ACCESS_ONCE(*(p));	\
+		smp_mb();				\
+		__v;					\
+	})
+#endif
+
+#ifndef smp_store_release
+# define smp_store_release(p, v)		\

Quoted 8 lines

kernel/locking/mcs_spinlock.c line 48

Quoted 5 lines

kernel/locking/mcs_spinlock.c +5 -0

+	do {					\
+		smp_mb();			\
+		ACCESS_ONCE(*(p)) = v;		\
+	} while (0)
+#endif

PeterZ already has a series implementing acquire/release accessors, so you should probably take a look at that rather than rolling your own here.

Yes, we are using Peter Z's implementation here. The above is for anything where smp_load_acquire and smp_store_release are *not* defined. We can remove this once all architectures implement the acquire and release functions as mentioned in the comments of the patch.

Quoted 15 lines

kernel/locking/mcs_spinlock.c void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node) · line 78

Quoted 7 lines

kernel/locking/mcs_spinlock.c +6 -0

+	/*
+	 * Wait until the lock holder passes the lock down.
+	 * Using smp_load_acquire() provides a memory barrier that
+	 * ensures subsequent operations happen after the lock is acquired.
+	 */
+	while (!(smp_load_acquire(&node->locked)))
 		arch_mutex_cpu_relax();

After a chat with some micro-architects, I'm going to have to disagree with Paul here. For architectures where acquire/release are implemented with explicit barriers (similarly for simple microarchitectures), emitting barriers in a loop *is* going to have an affect on overall performance, since those barriers may well result in traffic outside of the core (at least, on ARM).

Quoted 4 lines

Thinking more about that, the real issue here is that arch_mutex_cpu_relax() doesn't have a corresponding hook on the unlock side. On ARM, for example, we can enter a low-power state using the wfe instruction, but that requires the unlocker to wake up the core when the lock is released.

An alternate implementation is

	while (!ACCESS_ONCE(node->locked))
		arch_mutex_cpu_relax();
	smp_load_acquire(&node->locked);

Leaving the smp_load_acquire at the end to provide appropriate barrier. Will that be acceptable?

Quoted 7 lines

An alternate implementation is

	while (!ACCESS_ONCE(node->locked))
		arch_mutex_cpu_relax();
	smp_load_acquire(&node->locked);

Leaving the smp_load_acquire at the end to provide appropriate barrier. Will that be acceptable?

It still doesn't solve my problem though: I want a way to avoid that busy loop by some architecture-specific manner. The arch_mutex_cpu_relax() hook is a start, but there is no corresponding hook on the unlock side to issue a wakeup. Given a sensible relax implementation, I don't have an issue with putting a load-acquire in a loop, since it shouldn't be aggresively spinning anymore.

Quoted 8 lines · Tim Chen

Quoted 7 lines

An alternate implementation is

	while (!ACCESS_ONCE(node->locked))
		arch_mutex_cpu_relax();
	smp_load_acquire(&node->locked);

Leaving the smp_load_acquire at the end to provide appropriate barrier. Will that be acceptable?

Quoted 6 lines · Will Deacon

It still doesn't solve my problem though: I want a way to avoid that busy loop by some architecture-specific manner. The arch_mutex_cpu_relax() hook is a start, but there is no corresponding hook on the unlock side to issue a wakeup. Given a sensible relax implementation, I don't have an issue with putting a load-acquire in a loop, since it shouldn't be aggresively spinning anymore.

So you want something like this?

/*
* This is a spin-wait with acquire semantics. That is, accesses after
* this are not allowed to be reordered before the load that meets
* the specified condition. This requires that it end with either a
* load-acquire or a full smp_mb(). The optimal way to do this is likely
* to be architecture-dependent. E.g. x86 MONITOR/MWAIT instructions. */
#ifndef smp_load_acquire_until

#define smp_load_acquire_until(addr, cond) \
	while (!(smp_load_acquire(addr) cond)) { \
		do { \
			arch_mutex_cpu_relax(); \
		} while (!(ACCESS_ONCE(*(addr)) cond)); \
	}

#endif

	smp_load_acquire_until(&node->locked, != 0);

Alternative implementations:

#define smp_load_acquire_until(addr, cond) { \
	while (!(ACCESS_ONCE(*(addr)) cond)) \
		arch_mutex_cpu_relax(); \
	smp_mb(); }

#define smp_load_acquire_until(addr, cond) \
	if (!(smp_load_acquire(addr) cond)) { \
		do { \
			arch_mutex_cpu_relax(); \
		} while (!(ACCESS_ONCE(*(addr)) cond)); \
		smp_mb(); \
	}

Hello,

Quoted 41 lines · Tim Chen

From: Waiman Long <[email protected]>

This patch corrects the way memory barriers are used in the MCS lock with smp_load_acquire and smp_store_release fucnction.
It removes ones that are not needed.

It uses architecture specific load-acquire and store-release primitives for synchronization, if available. Generic implementations are provided in case they are not defined even though they may not be optimal. These generic implementation could be removed later on once changes are made in all the relevant header files.

Suggested-by Michel Lespinasse <[email protected]>

Signed-off-by Waiman Long <[email protected]>

Signed-off-by Jason Low <[email protected]>

Signed-off-by Tim Chen <[email protected]>

---
 kernel/locking/mcs_spinlock.c |   48 +++++++++++++++++++++++++++++++++++------
 1 files changed, 41 insertions(+), 7 deletions(-)

kernel/locking/mcs_spinlock.c +13 -0

index b6f27f8..df5c167 100644
@@ -23,6 +23,31 @@
 #endif
 
 /*
+ * Fall back to use the regular atomic operations and memory barrier if
+ * the acquire/release versions are not defined.
+ */
+#ifndef	xchg_acquire
+# define xchg_acquire(p, v)		xchg(p, v)
+#endif
+
+#ifndef	smp_load_acquire
+# define smp_load_acquire(p)				\
+	({						\
+		typeof(*p) __v = ACCESS_ONCE(*(p));	\
+		smp_mb();				\
+		__v;					\

kernel/locking/mcs_spinlock.c line 48

Quoted 10 lines

kernel/locking/mcs_spinlock.c +10 -0

+	})
+#endif
+
+#ifndef smp_store_release
+# define smp_store_release(p, v)		\
+	do {					\
+		smp_mb();			\
+		ACCESS_ONCE(*(p)) = v;		\
+	} while (0)
+#endif

PeterZ already has a series implementing acquire/release accessors, so you should probably take a look at that rather than rolling your own here.

You could then augment that with [cmp]xchg_{acquire,release} as appropriate.

Quoted 25 lines

kernel/locking/mcs_spinlock.c +9 -4

+/*
  * In order to acquire the lock, the caller should declare a local node and
  * pass a reference of the node to this function in addition to the lock.
  * If the lock has already been acquired, then this will proceed to spin

@@ -37,15 +62,19 @@ void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 	node->locked = 0;
 	node->next   = NULL;
 
-	prev = xchg(lock, node);
+	/* xchg() provides a memory barrier */
+	prev = xchg_acquire(lock, node);
 	if (likely(prev == NULL)) {
 		/* Lock acquired */
 		return;
 	}
 	ACCESS_ONCE(prev->next) = node;
-	smp_wmb();
-	/* Wait until the lock holder passes the lock down */
-	while (!ACCESS_ONCE(node->locked))
+	/*
+	 * Wait until the lock holder passes the lock down.
+	 * Using smp_load_acquire() provides a memory barrier that
+	 * ensures subsequent operations happen after the lock is acquired.
+	 */
+	while (!(smp_load_acquire(&node->locked)))

kernel/locking/mcs_spinlock.c void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node) · line 78

Quoted 1 line

kernel/locking/mcs_spinlock.c +0 -0

 		arch_mutex_cpu_relax();

After a chat with some micro-architects, I'm going to have to disagree with Paul here. For architectures where acquire/release are implemented with explicit barriers (similarly for simple microarchitectures), emitting barriers in a loop *is* going to have an affect on overall performance, since those barriers may well result in traffic outside of the core (at least, on ARM).

Thinking more about that, the real issue here is that arch_mutex_cpu_relax() doesn't have a corresponding hook on the unlock side. On ARM, for example, we can enter a low-power state using the wfe instruction, but that requires the unlocker to wake up the core when the lock is released.

So, although I'm completely in favour of introducing acquire/release accessors, I really think the mcs locking routines would benefit from some arch-specific backend code, even if it's optional (although I would imagine most architectures implementing something to improve power and/or performance).

From: Waiman Long <[email protected]>

This patch corrects the way memory barriers are used in the MCS lock with smp_load_acquire and smp_store_release fucnction.
It removes ones that are not needed.

It uses architecture specific load-acquire and store-release primitives for synchronization, if available. Generic implementations are provided in case they are not defined even though they may not be optimal. These generic implementation could be removed later on once changes are made in all the relevant header files.

Suggested-by Michel Lespinasse <[email protected]>

Signed-off-by Waiman Long <[email protected]>

Signed-off-by Jason Low <[email protected]>

Signed-off-by Tim Chen <[email protected]>

---
 kernel/locking/mcs_spinlock.c |   48 +++++++++++++++++++++++++++++++++++------
 1 files changed, 41 insertions(+), 7 deletions(-)

kernel/locking/mcs_spinlock.c +25 -0

index b6f27f8..df5c167 100644
@@ -23,6 +23,31 @@
 #endif
 
 /*
+ * Fall back to use the regular atomic operations and memory barrier if
+ * the acquire/release versions are not defined.
+ */
+#ifndef	xchg_acquire
+# define xchg_acquire(p, v)		xchg(p, v)
+#endif
+
+#ifndef	smp_load_acquire
+# define smp_load_acquire(p)				\
+	({						\
+		typeof(*p) __v = ACCESS_ONCE(*(p));	\
+		smp_mb();				\
+		__v;					\
+	})
+#endif
+
+#ifndef smp_store_release
+# define smp_store_release(p, v)		\
+	do {					\
+		smp_mb();			\
+		ACCESS_ONCE(*(p)) = v;		\
+	} while (0)
+#endif
+
+/*
  * In order to acquire the lock, the caller should declare a local node and
  * pass a reference of the node to this function in addition to the lock.
  * If the lock has already been acquired, then this will proceed to spin

kernel/locking/mcs_spinlock.c +0 -0

 	if (likely(prev == NULL)) {
 		/* Lock acquired */
 		return;
 	}
 	ACCESS_ONCE(prev->next) = node;

kernel/locking/mcs_spinlock.c +1 -1

 }
 EXPORT_SYMBOL_GPL(mcs_spin_lock);

@@ -54,7 +83,7 @@ EXPORT_SYMBOL_GPL(mcs_spin_lock);
  * Releases the lock. The caller should pass in the corresponding node that
  * was used to acquire the lock.
  */
-static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
+void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 {
 	struct mcs_spinlock *next = ACCESS_ONCE(node->next);
 

@@ -68,7 +97,12 @@ static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *nod
 		while (!(next = ACCESS_ONCE(node->next)))
 			arch_mutex_cpu_relax();

kernel/locking/mcs_spinlock.c +0 -0

 }
 EXPORT_SYMBOL_GPL(mcs_spin_unlock);

Signature · 1 line

1.7.4.4

Quoted 12 lines · Tim Chen

From: Waiman Long <[email protected]>

This patch corrects the way memory barriers are used in the MCS lock with smp_load_acquire and smp_store_release fucnction.
It removes ones that are not needed.

It uses architecture specific load-acquire and store-release primitives for synchronization, if available. Generic implementations are provided in case they are not defined even though they may not be optimal. These generic implementation could be removed later on once changes are made in all the relevant header files.

Quoted 4 lines

Suggested-by Michel Lespinasse <[email protected]>

Signed-off-by Waiman Long <[email protected]>

Signed-off-by Jason Low <[email protected]>

Signed-off-by Tim Chen <[email protected]>

Please see comments below.

Thanx, Paul

Quoted 44 lines

---
 kernel/locking/mcs_spinlock.c |   48 +++++++++++++++++++++++++++++++++++------
 1 files changed, 41 insertions(+), 7 deletions(-)

kernel/locking/mcs_spinlock.c +25 -0

index b6f27f8..df5c167 100644
@@ -23,6 +23,31 @@
 #endif
 
 /*
+ * Fall back to use the regular atomic operations and memory barrier if
+ * the acquire/release versions are not defined.
+ */
+#ifndef	xchg_acquire
+# define xchg_acquire(p, v)		xchg(p, v)
+#endif
+
+#ifndef	smp_load_acquire
+# define smp_load_acquire(p)				\
+	({						\
+		typeof(*p) __v = ACCESS_ONCE(*(p));	\
+		smp_mb();				\
+		__v;					\
+	})
+#endif
+
+#ifndef smp_store_release
+# define smp_store_release(p, v)		\
+	do {					\
+		smp_mb();			\
+		ACCESS_ONCE(*(p)) = v;		\
+	} while (0)
+#endif
+
+/*
  * In order to acquire the lock, the caller should declare a local node and
  * pass a reference of the node to this function in addition to the lock.
  * If the lock has already been acquired, then this will proceed to spin

@@ -37,15 +62,19 @@ void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 	node->locked = 0;
 	node->next   = NULL;
 

kernel/locking/mcs_spinlock.c void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node) · line 66

Quoted 3 lines

kernel/locking/mcs_spinlock.c +2 -1

-	prev = xchg(lock, node);
+	/* xchg() provides a memory barrier */
+	prev = xchg_acquire(lock, node);

But if this is xchg_acquire() with only acquire semantics, it need not ensure that the initializations of node->locked and node->next above will happen before the "ACCESS_ONCE(prev->next) = node" below. This therefore needs to remain xchg(). Or you need an smp_store_release() below instead of an ACCESS_ONCE() assignment.

As currently written, the poor CPU doing the unlock can be fatally disappointed by seeing pre-initialized values of ->locked and ->next. This could, among other things, result in a hang where the handoff happens before the initialization.

Quoted 5 lines

kernel/locking/mcs_spinlock.c +0 -0

 	if (likely(prev == NULL)) {
 		/* Lock acquired */
 		return;
 	}
 	ACCESS_ONCE(prev->next) = node;

kernel/locking/mcs_spinlock.c void mcs_spin_lock(struct mcs_spinlock **lock, struct mcs_spinlock *node) · line 78

Quoted 10 lines

kernel/locking/mcs_spinlock.c +6 -3

-	smp_wmb();
-	/* Wait until the lock holder passes the lock down */
-	while (!ACCESS_ONCE(node->locked))
+	/*
+	 * Wait until the lock holder passes the lock down.
+	 * Using smp_load_acquire() provides a memory barrier that
+	 * ensures subsequent operations happen after the lock is acquired.
+	 */
+	while (!(smp_load_acquire(&node->locked)))
 		arch_mutex_cpu_relax();

OK, this smp_load_acquire() makes sense!

Quoted 14 lines

kernel/locking/mcs_spinlock.c +1 -1

 }
 EXPORT_SYMBOL_GPL(mcs_spin_lock);

@@ -54,7 +83,7 @@ EXPORT_SYMBOL_GPL(mcs_spin_lock);
  * Releases the lock. The caller should pass in the corresponding node that
  * was used to acquire the lock.
  */
-static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
+void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *node)
 {
 	struct mcs_spinlock *next = ACCESS_ONCE(node->next);
 

@@ -68,7 +97,12 @@ static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *nod
 		while (!(next = ACCESS_ONCE(node->next)))
 			arch_mutex_cpu_relax();

kernel/locking/mcs_spinlock.c static void mcs_spin_unlock(struct mcs_spinlock **lock, struct mcs_spinlock *nod · line 106

Quoted 10 lines

kernel/locking/mcs_spinlock.c +7 -2

 	}
-	ACCESS_ONCE(next->locked) = 1;
-	smp_wmb();
+	/*
+	 * Pass lock to next waiter.
+	 * smp_store_release() provides a memory barrier to ensure
+	 * all operations in the critical section has been completed
+	 * before unlocking.
+	 */
+	smp_store_release(&next->locked , 1);

This smp_store_release() makes sense as well!

Could you please get rid of the extraneous space before the comma?

Quoted 7 lines

kernel/locking/mcs_spinlock.c +0 -0

 }
 EXPORT_SYMBOL_GPL(mcs_spin_unlock);

Signature · 1 line

1.7.4.4