Re: [BUG] 2.6.28-git LOCKDEP: Possible recursive rq->lock

[Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>]

* Peter Zijlstra <a.p.zijlstra@chello.nl> [2009-01-07 15:28:57]:

> On Wed, 2009-01-07 at 19:50 +0530, Vaidyanathan Srinivasan wrote:
> > * Peter Zijlstra <a.p.zijlstra@chello.nl> [2009-01-07 14:12:43]:
> > 
> > > On Wed, 2009-01-07 at 17:59 +0530, Vaidyanathan Srinivasan wrote:
> > > 
> > > > =============================================
> > > > [ INFO: possible recursive locking detected ]
> > > > 2.6.28-autotest-tip-sv #1
> > > > ---------------------------------------------
> > > > klogd/5062 is trying to acquire lock:
> > > >  (&rq->lock){++..}, at: [<ffffffff8022aca2>] task_rq_lock+0x45/0x7e
> > > > 
> > > > but task is already holding lock:
> > > >  (&rq->lock){++..}, at: [<ffffffff805f7354>] schedule+0x158/0xa31
> > > > 
> > > > other info that might help us debug this:
> > > > 1 lock held by klogd/5062:
> > > >  #0:  (&rq->lock){++..}, at: [<ffffffff805f7354>] schedule+0x158/0xa31
> > > > 
> > > > stack backtrace:
> > > > Pid: 5062, comm: klogd Not tainted 2.6.28-autotest-tip-sv #1
> > > > Call Trace:
> > > >  [<ffffffff80259ef1>] __lock_acquire+0xeb9/0x16a4
> > > >  [<ffffffff8025a6c0>] ? __lock_acquire+0x1688/0x16a4
> > > >  [<ffffffff8025a761>] lock_acquire+0x85/0xa9
> > > >  [<ffffffff8022aca2>] ? task_rq_lock+0x45/0x7e
> > > >  [<ffffffff805fa4d4>] _spin_lock+0x31/0x66
> > > >  [<ffffffff8022aca2>] ? task_rq_lock+0x45/0x7e
> > > >  [<ffffffff8022aca2>] task_rq_lock+0x45/0x7e
> > > >  [<ffffffff80233363>] try_to_wake_up+0x88/0x27a
> > > >  [<ffffffff80233581>] wake_up_process+0x10/0x12
> > > >  [<ffffffff805f775c>] schedule+0x560/0xa31
> > > 
> > > I'd be most curious to know where in schedule we are.
> > 
> > ok, we are in sched.c:3777
> > 
> >                 double_unlock_balance(this_rq, busiest);
> >                 if (active_balance)
> > >>>>>>>>>>>              wake_up_process(busiest->migration_thread);
> > 
> >         } else  
> > 
> > In active balance in newidle.  This implies sched_mc was 2 at that time.
> > let me trace this and debug further.
> 
> How about something like this? Strictly speaking we'll not deadlock,
> because ttwu will not be able to place the migration task on our rq, but
> since the code can deal with both rqs getting unlocked, this seems the
> easiest way out.

Hi Peter,

I agree.  Unlocking this_rq is an easy way out.  Thanks for the
suggestion.  I have moved the unlock and lock withing the if
condition.

--Vaidy

        sched: bug fix -- do not call ttwu while holding rq->lock
    
        When sched_mc=2 wake_up_process() is called on busiest_rq
        while holding this_rq lock in load_balance_newidle()
        Though this will not deadlock, this is a lockdep warning
        and the situation is easily solved by releasing the this_rq
        lock at this point in code
    
    Signed-off-by: Vaidyanathan Srinivasan <svaidy@linux.vnet.ibm.com>

diff --git a/kernel/sched.c b/kernel/sched.c
index 71a054f..703a669 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -3773,8 +3773,12 @@ redo:
                }
 
                double_unlock_balance(this_rq, busiest);
-               if (active_balance)
+               if (active_balance) {
+                       /*  Should not call ttwu while holding a rq->lock */
+                       spin_unlock(&this_rq->lock);
                        wake_up_process(busiest->migration_thread);
+                       spin_lock(&this_rq->lock);
+               }
 
        } else
                sd->nr_balance_failed = 0;