Re: [PATCH RFC v2] futex: Fix might_sleep() warning in futex_pivot_pending()

On 8/13/2026 11:15 AM, syzbot wrote:
> A recent change in commit 8e7ff730dd96 ("futex: Fix race in
> futex_pivot_pending() during private hash resize") modified
> futex_pivot_pending() to acquire a mutex to fix a race condition. However,
> futex_pivot_pending() is evaluated as a condition inside wait_var_event()
> in futex_hash_allocate(). Since wait_var_event() sets the task state to
> TASK_UNINTERRUPTIBLE before evaluating the condition, calling a blocking
> operation like mutex_lock() is invalid and triggers a might_sleep()
> warning:
> 
> do not call blocking ops when !TASK_RUNNING; state=2 set at
> [<ffffffff819e8c8d>] prepare_to_wait_event+0x3dd/0x480
> kernel/sched/wait.c:317
> WARNING: kernel/sched/core.c:9124 at __might_sleep+0x92/0xf0
> kernel/sched/core.c:9120
> Call Trace:
>   <TASK>
>   __mutex_lock_common kernel/locking/mutex.c:623 [inline]
>   __mutex_lock+0x118/0x1550 kernel/locking/mutex.c:821
>   class_mutex_constructor include/linux/mutex.h:253 [inline]
>   futex_pivot_pending kernel/futex/core.c:1789 [inline]
>   futex_hash_allocate+0x7fb/0xf00 kernel/futex/core.c:1872
>   __do_sys_prctl kernel/sys.c:2885 [inline]
>   __se_sys_prctl+0x78c/0x1910 kernel/sys.c:2534
> 
> Furthermore, if the mutex is contended, mutex_lock() will block. When it
> acquires the lock and returns, the task state will be reset to
> TASK_RUNNING. This causes the subsequent schedule() in the wait loop to
> return immediately, leading to a busy loop that consumes 100% CPU until the
> condition is met.
> 
> Fix this by reverting futex_pivot_pending() to a lockless implementation
> using RCU and memory barriers, which is the idiomatic way to handle
> conditions in wait_event loops. By reading the hash pointer first,
> executing an smp_rmb() memory barrier, and then reading hash_new, we
> leverage the Message Passing (MP) pattern to guarantee correctness without
> blocking. This pairs with the rcu_assign_pointer() release barrier in
> __futex_pivot_hash(). If the reader sees the new hash, it is guaranteed to
> see the cleared hash_new and correctly return true. If the reader sees the
> old hash, it will check futex_ref_is_dead(old), which will return true if
> the writer has already completed the pivot. The old hash memory is
> guaranteed to remain valid for the duration of the check in
> futex_ref_is_dead() because futex_pivot_pending() executes within an RCU
> read-side critical section and the old hash is freed using kvfree_rcu().
> 
> Fixes: 8e7ff730dd96 ("futex: Fix race in futex_pivot_pending() during private hash resize")
> Assisted-by: Gemini:gemini-3.6-flash Gemini:gemini-3.1-pro-preview syzbot
> Reported-by: [email protected]
> Closes: https://syzkaller.appspot.com/bug?extid=350a93852ac854927f45
> Link: https://syzkaller.appspot.com/ai_job?id=d25bd376-be2e-491a-a0a7-ef39dc13a9a5
> To: <[email protected]>
> To: "Ingo Molnar" <[email protected]>
> To: "Thomas Gleixner" <[email protected]>
> To: "Yao Kai" <[email protected]>
> Cc: =?utf-8?q?Andr=C3=A9_Almeida?= <[email protected]>
> Cc: "Davidlohr Bueso" <[email protected]>
> Cc: "Darren Hart" <[email protected]>
> Cc: "Peter Zijlstra" <[email protected]>
> 
> ---
> v2:
> - Use WRITE_ONCE() for stores to hash_new to complement READ_ONCE() in futex_pivot_pending().
> 
> v1:
> https://lore.kernel.org/all/[email protected]/T/
> ---
> diff --git a/kernel/futex/core.c b/kernel/futex/core.c
> index 128c5752f..e84be5410 100644
> --- a/kernel/futex/core.c
> +++ b/kernel/futex/core.c
> @@ -202,7 +202,7 @@ static bool __futex_pivot_hash(struct mm_struct *mm, struct futex_private_hash *
>   	fph = rcu_dereference_protected(mmph->hash, lockdep_is_held(&mmph->lock));
>   	if (fph) {
>   		if (!futex_ref_is_dead(fph)) {
> -			mmph->hash_new = new;
> +			WRITE_ONCE(mmph->hash_new, new);
>   			return false;
>   		}
>   
> @@ -224,7 +224,7 @@ static void futex_pivot_hash(struct mm_struct *mm)
>   
>   		fph = mm->futex.phash.hash_new;
>   		if (fph) {
> -			mm->futex.phash.hash_new = NULL;
> +			WRITE_ONCE(mm->futex.phash.hash_new, NULL);
>   			__futex_pivot_hash(mm, fph);
>   		}
>   	}
> @@ -1786,12 +1786,18 @@ static bool futex_pivot_pending(struct mm_struct *mm)
>   	struct futex_mm_phash *mmph = &mm->futex.phash;
>   	struct futex_private_hash *fph;
>   
> -	guard(mutex)(&mmph->lock);
> +	guard(rcu)();
>   
> -	if (!mmph->hash_new)
> +	fph = rcu_dereference(mmph->hash);
> +	/*
> +	 * Ensure that if we see the new hash, we will also see the cleared
> +	 * hash_new pointer. Pairs with rcu_assign_pointer() in
> +	 * __futex_pivot_hash().
> +	 */
> +	smp_rmb();
> +	if (!READ_ONCE(mmph->hash_new))
>   		return true;
>   
> -	fph = rcu_dereference_raw(mmph->hash);
>   	return futex_ref_is_dead(fph);
>   }
>   
> @@ -1879,7 +1885,7 @@ static int futex_hash_allocate(unsigned int hash_slots, unsigned int flags)
>   		cur = rcu_dereference_protected(mm->futex.phash.hash,
>   						lockdep_is_held(&mm->futex.phash.lock));
>   		new = mm->futex.phash.hash_new;
> -		mm->futex.phash.hash_new = NULL;
> +		WRITE_ONCE(mm->futex.phash.hash_new, NULL);
>   
>   		if (fph) {
>   			if (cur && !cur->hash_mask) {
> @@ -1889,7 +1895,7 @@ static int futex_hash_allocate(unsigned int hash_slots, unsigned int flags)
>   				 * the second one returns here.
>   				 */
>   				free = fph;
> -				mm->futex.phash.hash_new = new;
> +				WRITE_ONCE(mm->futex.phash.hash_new, new);
>   				return -EBUSY;
>   			}
>   			if (cur && !new) {
> 
> 
> base-commit: db2ddb87143519e20a95aa36c60b36107b736a58

Thanks, the WRITE_ONCE() changes address my comment and the code looks
correct.

One commit-message correction remains: mutex_lock() may restore
TASK_RUNNING when it takes the slow path, causing the immediately
following schedule() not to sleep. However, the next wait-loop iteration
sets TASK_UNINTERRUPTIBLE again, so this does not necessarily result in a
100% CPU busy loop until the condition is met. Please drop or reword that
paragraph.

Please also add:

Cc: [email protected]

since the commit being fixed is marked for stable.

Please send a v3 with these changes.