Re: [PATCH 13/14] lockdep: Prepare for NMI IRQ state tracking

From: Peter Zijlstra
Date: Tue Jun 02 2020 - 16:00:56 EST


On Sat, May 30, 2020 at 12:25:05AM +0200, Peter Zijlstra wrote:
> On Fri, May 29, 2020 at 06:14:01PM -0400, Steven Rostedt wrote:

> > Why remove the check for debug_locks? Isn't that there to disable
> > everything at once to prevent more warnings to be printed?
>
> Yeah, maybe. I was thinking we could keep IRQ state running. But you're
> right, if we mess up the IRQ state itself this might generate a wee
> mess.

How's this then?

---
Subject: lockdep: Prepare for NMI IRQ state tracking
From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Date: Wed May 27 15:00:57 CEST 2020

There is no reason not to always, accurately, track IRQ state.

This change also makes IRQ state tracking ignore lockdep_off().

Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
---
kernel/locking/lockdep.c | 44 +++++++++++++++++++++++++++++++++++++++++---
1 file changed, 41 insertions(+), 3 deletions(-)

--- a/kernel/locking/lockdep.c
+++ b/kernel/locking/lockdep.c
@@ -3646,7 +3646,16 @@ static void __trace_hardirqs_on_caller(v
*/
void lockdep_hardirqs_on_prepare(unsigned long ip)
{
- if (unlikely(!debug_locks || current->lockdep_recursion))
+ if (unlikely(!debug_locks))
+ return;
+
+ /*
+ * NMIs do not (and cannot) track lock dependencies, nothing to do.
+ */
+ if (unlikely(in_nmi()))
+ return;
+
+ if (unlikely(current->lockdep_recursion & LOCKDEP_RECURSION_MASK))
return;

if (unlikely(current->hardirqs_enabled)) {
@@ -3692,7 +3701,27 @@ void noinstr lockdep_hardirqs_on(unsigne
{
struct task_struct *curr = current;

- if (unlikely(!debug_locks || curr->lockdep_recursion))
+ if (unlikely(!debug_locks))
+ return;
+
+ /*
+ * NMIs can happen in the middle of local_irq_{en,dis}able() where the
+ * tracking state and hardware state are out of sync.
+ *
+ * NMIs must save lockdep_hardirqs_enabled() to restore IRQ state from,
+ * and not rely on hardware state like normal interrupts.
+ */
+ if (unlikely(in_nmi())) {
+ /*
+ * Skip:
+ * - recursion check, because NMI can hit lockdep;
+ * - hardware state check, because above;
+ * - chain_key check, see lockdep_hardirqs_on_prepare().
+ */
+ goto skip_checks;
+ }
+
+ if (unlikely(current->lockdep_recursion & LOCKDEP_RECURSION_MASK))
return;

if (curr->hardirqs_enabled) {
@@ -3720,6 +3749,7 @@ void noinstr lockdep_hardirqs_on(unsigne
DEBUG_LOCKS_WARN_ON(current->hardirq_chain_key !=
current->curr_chain_key);

+skip_checks:
/* we'll do an OFF -> ON transition: */
curr->hardirqs_enabled = 1;
curr->hardirq_enable_ip = ip;
@@ -3735,7 +3765,15 @@ void noinstr lockdep_hardirqs_off(unsign
{
struct task_struct *curr = current;

- if (unlikely(!debug_locks || curr->lockdep_recursion))
+ if (unlikely(!debug_locks))
+ return;
+
+ /*
+ * Matching lockdep_hardirqs_on(), allow NMIs in the middle of lockdep;
+ * they will restore the software state. This ensures the software
+ * state is consistent inside NMIs as well.
+ */
+ if (unlikely(!in_nmi() && (current->lockdep_recursion & LOCKDEP_RECURSION_MASK)))
return;

/*