IRQ disable vs NMI

This patch set tries to solve the local_irq_disable() vs NMI problem that SPARC has by providing new arch hooks and instrumenting the existing interface to WARN on conflicting usage.

Provide local_irq_{save,restore}_nmi() which will allow us to help architectures that implement NMIs using IRQ priorities like SPARC64 does.

Sparc uses IRQ prio 15 for NMIs and implements local_irq_disable() as disable <= 14. However if you do that while inside an NMI you re- enable the NMI priority again, causing all kinds of fun.

A more solid implementation would first check the disable level and never lower it, however that is more costly and would slow down the rest of the kernel for no particular reason.

Therefore introduce local_irq_save_nmi() which can implement this slower but more solid scheme and dis-allow local_irq_save() from NMI context.

---
 include/linux/irqflags.h     |   51 ++++++++++++++++++++++++++++++++++++++++---
 kernel/lockdep.c             |    7 +++++
 kernel/trace/trace_irqsoff.c |    8 ++++++
 3 files changed, 63 insertions(+), 3 deletions(-)

Index: linux-2.6/include/linux/irqflags.h
===================================================================
@@ -18,6 +18,7 @@
   extern void trace_softirqs_off(unsigned long ip);
   extern void trace_hardirqs_on(void);
   extern void trace_hardirqs_off(void);
+  extern void trace_hardirqs_off_no_nmi(void);
 # define trace_hardirq_context(p)	((p)->hardirq_context)
 # define trace_softirq_context(p)	((p)->softirq_context)
 # define trace_hardirqs_enabled(p)	((p)->hardirqs_enabled)

@@ -30,6 +31,7 @@
 #else
 # define trace_hardirqs_on()		do { } while (0)
 # define trace_hardirqs_off()		do { } while (0)
+# define trace_hardirqs_off_no_nmi()	do { } while (0)
 # define trace_softirqs_on(ip)		do { } while (0)
 # define trace_softirqs_off(ip)		do { } while (0)
 # define trace_hardirq_context(p)	0

@@ -59,15 +61,15 @@
 #define local_irq_enable() \
 	do { trace_hardirqs_on(); raw_local_irq_enable(); } while (0)
 #define local_irq_disable() \
-	do { raw_local_irq_disable(); trace_hardirqs_off(); } while (0)
+	do { raw_local_irq_disable(); trace_hardirqs_off_no_nmi(); } while (0)
+
 #define local_irq_save(flags)				\
 	do {						\
 		typecheck(unsigned long, flags);	\
 		raw_local_irq_save(flags);		\
-		trace_hardirqs_off();			\
+		trace_hardirqs_off_no_nmi();		\
 	} while (0)
 
-
 #define local_irq_restore(flags)			\
 	do {						\
 		typecheck(unsigned long, flags);	\

@@ -79,6 +81,30 @@
 			raw_local_irq_restore(flags);	\
 		}					\
 	} while (0)
+
+#ifndef local_irq_save_nmi
+# define local_irq_save_nmi(flags) 			\
+	 do {						\
+		 typecheck(unsigned long, flags);	\
+		 raw_local_irq_save(flags);		\
+		 trace_hardirqs_off();			\
+	 } while (0)
+#endif
+
+#ifndef local_irq_restore_nmi
+#define local_irq_restore_nmi(flags)			\
+	do {						\
+		typecheck(unsigned long, flags);	\
+		if (raw_irqs_disabled_flags(flags)) {	\
+			raw_local_irq_restore(flags);	\
+			trace_hardirqs_off();		\
+		} else {				\
+			trace_hardirqs_on();		\
+			raw_local_irq_restore(flags);	\
+		}					\
+	} while (0)
+#endif
+
 #else /* !CONFIG_TRACE_IRQFLAGS_SUPPORT */
 /*
  * The local_irq_*() APIs are equal to the raw_local_irq*()

@@ -86,16 +112,35 @@
  */
 # define raw_local_irq_disable()	local_irq_disable()
 # define raw_local_irq_enable()		local_irq_enable()
+
 # define raw_local_irq_save(flags)			\
 	do {						\
 		typecheck(unsigned long, flags);	\
 		local_irq_save(flags);			\
 	} while (0)
+
 # define raw_local_irq_restore(flags)			\
 	do {						\
 		typecheck(unsigned long, flags);	\
 		local_irq_restore(flags);		\
 	} while (0)
+
+#ifndef local_irq_save_nmi
+# define local_irq_save_nmi(flags)			\
+	 do {						\
+		 typecheck(unsigned long, flags);	\
+		 local_irq_save(flags);			\
+	 } while (0)
+#endif
+
+#ifndef local_irq_restore_nmi
+# define local_irq_restore_nmi(flags)			\
+	 do {						\
+		 typecheck(unsigned long, flags);	\
+		 local_irq_restore(flags);		\
+	 } while (0)
+#endif
+
 #endif /* CONFIG_TRACE_IRQFLAGS_SUPPORT */
 
 #ifdef CONFIG_TRACE_IRQFLAGS_SUPPORT

Index: linux-2.6/kernel/lockdep.c
===================================================================
@@ -2369,6 +2369,13 @@ void trace_hardirqs_off(void)
 }
 EXPORT_SYMBOL(trace_hardirqs_off);
 
+void trace_hardirqs_off_no_nmi(void)
+{
+	WARN_ON_ONCE(in_nmi());
+	trace_hardirqs_off_caller(CALLER_ADDR0);
+}
+EXPORT_SYMBOL(trace_hardirqs_off_no_nmi);
+
 /*
  * Softirqs will be enabled:
  */

Index: linux-2.6/kernel/trace/trace_irqsoff.c
===================================================================
@@ -316,6 +316,14 @@ void trace_hardirqs_off(void)
 }
 EXPORT_SYMBOL(trace_hardirqs_off);
 
+void trace_hardirqs_off_no_nmi(void)
+{
+	WARN_ON_ONCE(in_nmi());
+	if (!preempt_trace() && irq_trace())
+		start_critical_timing(CALLER_ADDR0, CALLER_ADDR1);
+}
+EXPORT_SYMBOL(trace_hardirqs_off_no_nmi);
+
 void trace_hardirqs_on_caller(unsigned long caller_addr)
 {
 	if (!preempt_trace() && irq_trace())

From: Steven Rostedt <[email protected]>
Date: Tue, 06 Apr 2010 21:13:10 -0400

That's the whole point, so that the problem is more easily noticed and it gets fixed long before I end up accidently testing the code on my machines :-)

To be honest, the fix is so trivial, you just need to add '_nmi' to the local_irq_{save,restore}() calls that warn like this.

I'm even willing to have you forward all of those reports to me and I'll be responsible for fixing them.

How's that? :-)

Sure.

/me sets up his procmailrc to search for the WARN_ON line in trace_irqsoff.c and have it forward to davem.

-- Steve

Patch 8bb39f9 (perf: Fix 'perf sched record' deadlock) introduced a local_irq_save() in NMI context, convert that to local_irq_save_nmi() and move the IRQ disable into perf_output_lock/unlock().

The former is needed because we now disallow local_irq_disable() from NMI context due to some arch limitations.

The second is because its really about IRQ lock inversion with that funny output lock, and perf_event_task_output() is only one site that could trigger it.

---
 include/linux/perf_event.h |    1 +
 kernel/perf_event.c        |   17 ++++++-----------
 2 files changed, 7 insertions(+), 11 deletions(-)

Index: linux-2.6/include/linux/perf_event.h =================================--- linux-2.6.orig/include/linux/perf_event.h +++ linux-2.6/include/linux/perf_event.h
@@ -758,6 +758,7 @@ struct perf_output_handle {
struct perf_mmap_data *data;
unsigned long head;
unsigned long offset;
+ unsigned long flags;
int nmi;
int sample;
int locked;
Index: linux-2.6/kernel/perf_event.c =================================--- linux-2.6.orig/kernel/perf_event.c +++ linux-2.6/kernel/perf_event.c
@@ -2848,6 +2848,10 @@ static void perf_output_lock(struct perf
struct perf_mmap_data *data = handle->data;
int cur, cpu = get_cpu();

+ /*
+ * Since this is a lock we need to be IRQ-safe
+ */
+ local_irq_save_nmi(handle->flags);
handle->locked = 0;

for (;;) {
@@ -2906,6 +2910,7 @@ again:
if (atomic_xchg(&data->wakeup, 0))
perf_output_wakeup(handle);
out:
+ local_irq_restore_nmi(handle->flags);
put_cpu();
}

@@ -3385,19 +3390,10 @@ static void perf_event_task_output(struc
unsigned long flags;
int size, ret;

- /*
- * If this CPU attempts to acquire an rq lock held by a CPU spinning
- * in perf_output_lock() from interrupt context, it's game over.
- */
- local_irq_save(flags);
-
size = task_event->event_id.header.size; ret = perf_output_begin(&handle, event, size, 0, 0);
-
- if (ret) {
- local_irq_restore(flags);
+ if (ret)
return;
- }

task_event->event_id.pid = perf_event_pid(event, task); task_event->event_id.ppid = perf_event_pid(event, current);
@@ -3408,7 +3404,6 @@ static void perf_event_task_output(struc
perf_output_put(&handle, task_event->event_id);

perf_output_end(&handle);
- local_irq_restore(flags);
}

static int perf_event_task_match(struct perf_event *event)

Since we can call cpu_clock() from NMI context fix up the IRQ disabling to conform to the new rules.

---
 kernel/sched_clock.c |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

Index: linux-2.6/kernel/sched_clock.c =================================--- linux-2.6.orig/kernel/sched_clock.c +++ linux-2.6/kernel/sched_clock.c
@@ -241,9 +241,9 @@ unsigned long long cpu_clock(int cpu)
unsigned long long clock;
unsigned long flags;

- local_irq_save(flags);
+ local_irq_save_nmi(flags);
clock = sched_clock_cpu(cpu);
- local_irq_restore(flags);
+ local_irq_restore_nmi(flags);

return clock;
}

That is already so, see 1/3.

Ah you're right.

From: Peter Zijlstra <[email protected]>
Date: Tue, 06 Apr 2010 15:28:07 +0200

Thanks Peter, looks good at first sight.

I'll toss together the sparc bits and give this a go.

From: Peter Zijlstra <[email protected]>
Date: Tue, 06 Apr 2010 15:28:07 +0200

Ok, two patches coming.

One which adds the sparc64 irqflags.h methods.

And one which annotates the ftrace functions, as needed.

With this I have the function tracer working.

Thanks!