futex: Use RCU-based per-CPU reference counting instead of rcuref_t

The use of rcuref_t for reference counting introduces a performance bottleneck
when accessed concurrently by multiple threads during futex operations.

Replace rcuref_t with special crafted per-CPU reference counters. The
lifetime logic remains the same.

The newly allocate private hash starts in FR_PERCPU state. In this state, each
futex operation that requires the private hash uses a per-CPU counter (an
unsigned int) for incrementing or decrementing the reference count.

When the private hash is about to be replaced, the per-CPU counters are
migrated to a atomic_t counter mm_struct::futex_atomic.
The migration process:
- Waiting for one RCU grace period to ensure all users observe the
  current private hash. This can be skipped if a grace period elapsed
  since the private hash was assigned.

- futex_private_hash::state is set to FR_ATOMIC, forcing all users to
  use mm_struct::futex_atomic for reference counting.

- After a RCU grace period, all users are guaranteed to be using the
  atomic counter. The per-CPU counters can now be summed up and added to
  the atomic_t counter. If the resulting count is zero, the hash can be
  safely replaced. Otherwise, active users still hold a valid reference.

- Once the atomic reference count drops to zero, the next futex
  operation will switch to the new private hash.

call_rcu_hurry() is used to speed up transition which otherwise might be
delay with RCU_LAZY. There is nothing wrong with using call_rcu(). The
side effects would be that on auto scaling the new hash is used later
and the SET_SLOTS prctl() will block longer.

[bigeasy: commit description + mm get/ put_async]

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lore.kernel.org/r/20250710110011.384614-3-bigeasy@linutronix.de

3 files changed, 11 insertions, 12 deletions

diff --git a/include/linux/futex.h b/include/linux/futex.h
index b37193653e6b..9e9750f04980 100644
--- a/include/linux/futex.h
+++ b/include/linux/futex.h
@@ -85,18 +85,12 @@ int futex_hash_prctl(unsigned long arg2, unsigned long arg3, unsigned long arg4)
 #ifdef CONFIG_FUTEX_PRIVATE_HASH
 int futex_hash_allocate_default(void);
 void futex_hash_free(struct mm_struct *mm);
-
-static inline void futex_mm_init(struct mm_struct *mm)
-{
-	RCU_INIT_POINTER(mm->futex_phash, NULL);
-	mm->futex_phash_new = NULL;
-	mutex_init(&mm->futex_hash_lock);
-}
+int futex_mm_init(struct mm_struct *mm);
 
 #else /* !CONFIG_FUTEX_PRIVATE_HASH */
 static inline int futex_hash_allocate_default(void) { return 0; }
-static inline void futex_hash_free(struct mm_struct *mm) { }
-static inline void futex_mm_init(struct mm_struct *mm) { }
+static inline int futex_hash_free(struct mm_struct *mm) { return 0; }
+static inline int futex_mm_init(struct mm_struct *mm) { return 0; }
 #endif /* CONFIG_FUTEX_PRIVATE_HASH */
 
 #else /* !CONFIG_FUTEX */
@@ -118,8 +112,8 @@ static inline int futex_hash_allocate_default(void)
 {
 	return 0;
 }
-static inline void futex_hash_free(struct mm_struct *mm) { }
-static inline void futex_mm_init(struct mm_struct *mm) { }
+static inline int futex_hash_free(struct mm_struct *mm) { return 0; }
+static inline int futex_mm_init(struct mm_struct *mm) { return 0; }
 
 #endif
 
diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index d6b91e8a66d6..0f0662157066 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -1070,6 +1070,11 @@ struct mm_struct {
 		struct mutex			futex_hash_lock;
 		struct futex_private_hash	__rcu *futex_phash;
 		struct futex_private_hash	*futex_phash_new;
+		/* futex-ref */
+		unsigned long			futex_batches;
+		struct rcu_head			futex_rcu;
+		atomic_long_t			futex_atomic;
+		unsigned int			__percpu *futex_ref;
 #endif
 
 		unsigned long hiwater_rss; /* High-watermark of RSS usage */
diff --git a/include/linux/sched/mm.h b/include/linux/sched/mm.h
index b13474825130..2201da0afecc 100644
--- a/include/linux/sched/mm.h
+++ b/include/linux/sched/mm.h
@@ -140,7 +140,7 @@ static inline bool mmget_not_zero(struct mm_struct *mm)
 
 /* mmput gets rid of the mappings and all user-space */
 extern void mmput(struct mm_struct *);
-#ifdef CONFIG_MMU
+#if defined(CONFIG_MMU) || defined(CONFIG_FUTEX_PRIVATE_HASH)
 /* same as above but performs the slow path from the async context. Can
  * be called from the atomic context as well
  */