Re: linux-next: stall warnings and deadlock on Arm64 (was: [PATCH] kfence: Avoid stalling...)

["Paul E. McKenney" <paulmck@kernel.org>]

On Thu, Nov 19, 2020 at 08:38:19PM +0100, Marco Elver wrote:
> On Thu, Nov 19, 2020 at 10:48AM -0800, Paul E. McKenney wrote:
> > On Thu, Nov 19, 2020 at 06:02:59PM +0100, Marco Elver wrote:

[ . . . ]

> > > I can try bisection again, or reverting some commits that might be
> > > suspicious? But we'd need some selection of suspicious commits.
> > 
> > The report claims that one of the rcu_node ->lock fields is held
> > with interrupts enabled, which would indeed be bad.  Except that all
> > of the stack traces that it shows have these locks held within the
> > scheduling-clock interrupt handler.  Now with the "rcu: Don't invoke
> > try_invoke_on_locked_down_task() with irqs disabled" but without the
> > "sched/core: Allow try_invoke_on_locked_down_task() with irqs disabled"
> > commit, I understand why.  With both, I don't see how this happens.
> 
> I'm at a loss, but happy to keep bisecting and trying patches. I'm also
> considering:
> 
> 	Is it the compiler? Probably not, I tried 2 versions of GCC.
> 
> 	Can we trust lockdep to precisely know IRQ state? I know there's
> 	been some recent work around this, but hopefully we're not
> 	affected here?
> 
> 	Is QEMU buggy?
> 
> > At this point, I am reduced to adding lockdep_assert_irqs_disabled()
> > calls at various points in that code, as shown in the patch below.
> > 
> > At this point, I would guess that your first priority would be the
> > initial bug rather than this following issue, but you never know, this
> > might well help diagnose the initial bug.
> 
> I don't mind either way. I'm worried deadlocking the whole system might
> be worse.

Here is another set of lockdep_assert_irqs_disabled() calls on the
off-chance that they actually find something.

							Thanx, Paul

------------------------------------------------------------------------

commit bcca5277df3f24db15e15ccc8b05ecf346d05169
Author: Paul E. McKenney <paulmck@kernel.org>
Date:   Thu Nov 19 13:30:33 2020 -0800

    rcu: Add lockdep_assert_irqs_disabled() to raw_spin_unlock_rcu_node() macros
    
    This commit adds a lockdep_assert_irqs_disabled() call to the
    helper macros that release the rcu_node structure's ->lock, namely
    to raw_spin_unlock_rcu_node(), raw_spin_unlock_irq_rcu_node() and
    raw_spin_unlock_irqrestore_rcu_node().  The point of this is to help track
    down a situation where lockdep appears to be insisting that interrupts
    are enabled while holding an rcu_node structure's ->lock.
    
    Link: https://lore.kernel.org/lkml/20201111133813.GA81547@elver.google.com/
    Signed-off-by: Paul E. McKenney <paulmck@kernel.org>

diff --git a/kernel/rcu/rcu.h b/kernel/rcu/rcu.h
index 59ef1ae..bf0827d 100644
--- a/kernel/rcu/rcu.h
+++ b/kernel/rcu/rcu.h
@@ -378,7 +378,11 @@ do {									\
 	smp_mb__after_unlock_lock();					\
 } while (0)
 
-#define raw_spin_unlock_rcu_node(p) raw_spin_unlock(&ACCESS_PRIVATE(p, lock))
+#define raw_spin_unlock_rcu_node(p)					\
+do {									\
+	lockdep_assert_irqs_disabled();					\
+	raw_spin_unlock(&ACCESS_PRIVATE(p, lock));			\
+} while (0)
 
 #define raw_spin_lock_irq_rcu_node(p)					\
 do {									\
@@ -387,7 +391,10 @@ do {									\
 } while (0)
 
 #define raw_spin_unlock_irq_rcu_node(p)					\
-	raw_spin_unlock_irq(&ACCESS_PRIVATE(p, lock))
+do {									\
+	lockdep_assert_irqs_disabled();					\
+	raw_spin_unlock_irq(&ACCESS_PRIVATE(p, lock));			\
+} while (0)
 
 #define raw_spin_lock_irqsave_rcu_node(p, flags)			\
 do {									\
@@ -396,7 +403,10 @@ do {									\
 } while (0)
 
 #define raw_spin_unlock_irqrestore_rcu_node(p, flags)			\
-	raw_spin_unlock_irqrestore(&ACCESS_PRIVATE(p, lock), flags)
+do {									\
+	lockdep_assert_irqs_disabled();					\
+	raw_spin_unlock_irqrestore(&ACCESS_PRIVATE(p, lock), flags);	\
+} while (0)
 
 #define raw_spin_trylock_rcu_node(p)					\
 ({									\