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

[Thomas Gleixner]

On Wed, 30 Jan 2019, Heiko Carstens wrote:
> On Tue, Jan 29, 2019 at 06:16:53PM +0100, Sebastian Sewior wrote:
> >  	if (unlikely(p->flags & PF_KTHREAD)) {
> >  		put_task_struct(p);
> 
> Last lines of the trace with your additional patch (full log attached):
> 
>            <...>-50539 [003] ....  2376.398223: sys_futex -> 0x0
>            <...>-50539 [003] ....  2376.398223: sys_futex(uaddr: 3ffb7700208, op: 6, val: 1, utime: 0, uaddr2: 3, val3: 0)
>            <...>-50539 [003] ....  2376.398225: attach_to_pi_owner: Missing pid 50734
>            <...>-50539 [003] ....  2376.398226: handle_exit_race: uval2 vs uval 8000c62e vs 8000c62e (-1)

So the user space value is: 8000c62e. FUTEX_WAITER bit is set and the owner
of the futex is PID 50734, which exited long time ago:

           <...>-50734 [000] ....  2376.394936: sched_process_exit: comm=ld64.so.1 pid=50734 prio=120

But at least from the kernel view 50734 has released it last:

           <...>-50734 [000] ....  2376.394930: sys_futex(uaddr: 3ffb7700208, op: 7, val: 3ff00000007, utime: 3ffb3ef8910, uaddr2: 3ffb3ef8910, val3: 3ffc0afe987)
           <...>-50539 [003] ....  2376.398223: sys_futex(uaddr: 3ffb7700208, op: 6, val: 1, utime: 0, uaddr2: 3, val3: 0)

Now, if it would have acquired it in userspace again before exiting, then
the robust list exit code should have set the OWNER_DIED bit as well, but
that's not set....

debug patch for the robust list exit handling below.

Thanks,

	tglx

8<-----------------
--- a/kernel/futex.c
+++ b/kernel/futex.c
@@ -3440,10 +3440,13 @@ static int handle_futex_death(u32 __user
 {
 	u32 uval, uninitialized_var(nval), mval;
 
+	trace_printk("uaddr: %lx pi: %d\n", (unsigned long) uaddr, pi);
 retry:
 	if (get_user(uval, uaddr))
 		return -1;
 
+	trace_printk("uaddr: %lx uval: %08x\n", (unsigned long) uaddr, uval);
+
 	if ((uval & FUTEX_TID_MASK) == task_pid_vnr(curr)) {
 		/*
 		 * Ok, this dying thread is truly holding a futex
@@ -3480,6 +3483,7 @@ static int handle_futex_death(u32 __user
 		if (!pi && (uval & FUTEX_WAITERS))
 			futex_wake(uaddr, 1, 1, FUTEX_BITSET_MATCH_ANY);
 	}
+	trace_printk("uaddr: %lx success\n", (unsigned long) uaddr);
 	return 0;
 }