Re: WARN_ON_ONCE(!new_owner) within wake_futex_pi() triggered

[Thomas Gleixner]

On Wed, 30 Jan 2019, Thomas Gleixner wrote:
> On Wed, 30 Jan 2019, Thomas Gleixner wrote:
> The last entries with that uaddr are:
> 
>           <...>-56956 [005] ....   658.923608: sys_futex(uaddr: 3ff9e880140, op: 7, val: 3ff00000007, utime: 3ff9b078910, uaddr2: 3ff9b078910, val3: 3ffea67e3f7)
> 
> UNLOCK
> 
>            <...>-56945 [006] ....   658.923612: sys_futex(uaddr: 3ff9e880140, op: 6, val: 1, utime: 1003ff0, uaddr2: 3ff9e87f910, val3: 3ff0000de71)
> 
> LOCK
> 
>            <...>-56956 [005] ....   658.923612: sys_futex(uaddr: 3ff9e880140, op: 7, val: 3ff00000007, utime: 3ff9b078910, uaddr2: 3ff9b078910, val3: 3ffea67e3f7)
> 
> UNLOCK
> 
>            <...>-56945 [006] ....   658.923830: sys_futex(uaddr: 3ff9e880140, op: 7, val: 3ff00000007, utime: 3ff9e87f910, uaddr2: 3ff9e87f910, val3: 3ffea67e3f7)
> 
> UNLOCK
> 
>            <...>-56496 [001] ....   658.932404: sys_futex(uaddr: 3ff9e880140, op: 6, val: 1, utime: 0, uaddr2: 5, val3: 0)
> 
> LOCK which fails.
> 
> This does not make any sense. The last kernel visible operation of 56956 on
> that uaddr is the UNLOCK above.
> 
> I need to think some more about what might happen.

TBH, no clue. Below are some more traceprintks which hopefully shed some
light on that mystery. See kernel/futex.c line 30  ...

Thanks,

	tglx
8<--------------
--- a/kernel/futex.c
+++ b/kernel/futex.c
@@ -1502,6 +1502,8 @@ static int wake_futex_pi(u32 __user *uad
 	 * died bit, because we are the owner.
 	 */
 	newval = FUTEX_WAITERS | task_pid_vnr(new_owner);
+	trace_printk("uaddr: %lx cur: %x new: %x\n",
+		     (unsigned long) uaddr, uval, newval);
 
 	if (unlikely(should_fail_futex(true)))
 		ret = -EFAULT;
@@ -2431,6 +2433,8 @@ static int fixup_pi_state_owner(u32 __us
 	for (;;) {
 		newval = (uval & FUTEX_OWNER_DIED) | newtid;
 
+		trace_printk("uaddr: %lx cur: %x new: %x\n",
+			     (unsigned long) uaddr, uval, newval);
 		if (cmpxchg_futex_value_locked(&curval, uaddr, uval, newval))
 			goto handle_fault;
 		if (curval == uval)
@@ -2438,6 +2442,8 @@ static int fixup_pi_state_owner(u32 __us
 		uval = curval;
 	}
 
+	trace_printk("uaddr: %lx cur: %x new: %x\n",
+		     (unsigned long) uaddr, uval, newval);
 	/*
 	 * We fixed up user space. Now we need to fix the pi_state
 	 * itself.
@@ -3028,6 +3034,9 @@ static int futex_unlock_pi(u32 __user *u
 		/* drops pi_state->pi_mutex.wait_lock */
 		ret = wake_futex_pi(uaddr, uval, pi_state);
 
+		trace_printk("uaddr: %lx wake: %d\n",
+			     (unsigned long) uaddr, ret);
+
 		put_pi_state(pi_state);
 
 		/*
@@ -3056,6 +3065,8 @@ static int futex_unlock_pi(u32 __user *u
 		goto out_putkey;
 	}
 
+	trace_printk("uaddr: %lx cur: %x new: %x\n",
+		     (unsigned long) uaddr, uval, 0);
 	/*
 	 * We have no kernel internal state, i.e. no waiters in the
 	 * kernel. Waiters which are about to queue themselves are stuck