oprofile, x86: Reimplement nmi timer mode using perf event

The legacy x86 nmi watchdog code was removed with the implementation
of the perf based nmi watchdog. This broke Oprofile's nmi timer
mode. To run nmi timer mode we relied on a continuous ticking nmi
source which the nmi watchdog provided. The nmi tick was no longer
available and current watchdog can not be used anymore since it runs
with very long periods in the range of seconds. This patch
reimplements the nmi timer mode using a perf counter nmi source.

V2:
* removing pr_info()
* fix undefined reference to `__udivdi3' for 32 bit build
* fix section mismatch of .cpuinit.data:nmi_timer_cpu_nb
* removed nmi timer setup in arch/x86
* implemented function stubs for op_nmi_init/exit()
* made code more readable in oprofile_init()

V3:
* fix architectural initialization in oprofile_init()
* fix CONFIG_OPROFILE_NMI_TIMER dependencies

Acked-by: Peter Zijlstra <[email protected]>
Signed-off-by: Robert Richter <[email protected]>

arch/Kconfig

@@ -30,6 +30,10 @@ config OPROFILE_EVENT_MULTIPLEX
 config HAVE_OPROFILE
 	bool
 
+config OPROFILE_NMI_TIMER
+	def_bool y
+	depends on PERF_EVENTS && HAVE_PERF_EVENTS_NMI
+
 config KPROBES
 	bool "Kprobes"
 	depends on MODULES

arch/x86/oprofile/Makefile

@@ -4,9 +4,8 @@ DRIVER_OBJS = $(addprefix ../../../drivers/oprofile/, \
 		oprof.o cpu_buffer.o buffer_sync.o \
 		event_buffer.o oprofile_files.o \
 		oprofilefs.o oprofile_stats.o  \
-		timer_int.o )
+		timer_int.o nmi_timer_int.o )
 
 oprofile-y				:= $(DRIVER_OBJS) init.o backtrace.o
 oprofile-$(CONFIG_X86_LOCAL_APIC) 	+= nmi_int.o op_model_amd.o \
 					   op_model_ppro.o op_model_p4.o
-oprofile-$(CONFIG_X86_IO_APIC)		+= nmi_timer_int.o

arch/x86/oprofile/init.c

@@ -16,37 +16,23 @@
  * with the NMI mode driver.
  */
 
+#ifdef CONFIG_X86_LOCAL_APIC
 extern int op_nmi_init(struct oprofile_operations *ops);
-extern int op_nmi_timer_init(struct oprofile_operations *ops);
 extern void op_nmi_exit(void);
-extern void x86_backtrace(struct pt_regs * const regs, unsigned int depth);
+#else
+static int op_nmi_init(struct oprofile_operations *ops) { return -ENODEV; }
+static void op_nmi_exit(void) { }
+#endif
 
-static int nmi_timer;
+extern void x86_backtrace(struct pt_regs * const regs, unsigned int depth);
 
 int __init oprofile_arch_init(struct oprofile_operations *ops)
 {
-	int ret;
-
-	ret = -ENODEV;
-
-#ifdef CONFIG_X86_LOCAL_APIC
-	ret = op_nmi_init(ops);
-#endif
-	nmi_timer = (ret != 0);
-#ifdef CONFIG_X86_IO_APIC
-	if (nmi_timer)
-		ret = op_nmi_timer_init(ops);
-#endif
 	ops->backtrace = x86_backtrace;
-
-	return ret;
+	return op_nmi_init(ops);
 }
 
-
 void oprofile_arch_exit(void)
 {
-#ifdef CONFIG_X86_LOCAL_APIC
-	if (!nmi_timer)
-		op_nmi_exit();
-#endif
+	op_nmi_exit();
 }

drivers/oprofile/nmi_timer_int.c

@@ -0,0 +1,173 @@
+/**
+ * @file nmi_timer_int.c
+ *
+ * @remark Copyright 2011 Advanced Micro Devices, Inc.
+ *
+ * @author Robert Richter <[email protected]>
+ */
+
+#include <linux/init.h>
+#include <linux/smp.h>
+#include <linux/errno.h>
+#include <linux/oprofile.h>
+#include <linux/perf_event.h>
+
+#ifdef CONFIG_OPROFILE_NMI_TIMER
+
+static DEFINE_PER_CPU(struct perf_event *, nmi_timer_events);
+static int ctr_running;
+
+static struct perf_event_attr nmi_timer_attr = {
+	.type           = PERF_TYPE_HARDWARE,
+	.config         = PERF_COUNT_HW_CPU_CYCLES,
+	.size           = sizeof(struct perf_event_attr),
+	.pinned         = 1,
+	.disabled       = 1,
+};
+
+static void nmi_timer_callback(struct perf_event *event,
+			       struct perf_sample_data *data,
+			       struct pt_regs *regs)
+{
+	event->hw.interrupts = 0;       /* don't throttle interrupts */
+	oprofile_add_sample(regs, 0);
+}
+
+static int nmi_timer_start_cpu(int cpu)
+{
+	struct perf_event *event = per_cpu(nmi_timer_events, cpu);
+
+	if (!event) {
+		event = perf_event_create_kernel_counter(&nmi_timer_attr, cpu, NULL,
+							 nmi_timer_callback, NULL);
+		if (IS_ERR(event))
+			return PTR_ERR(event);
+		per_cpu(nmi_timer_events, cpu) = event;
+	}
+
+	if (event && ctr_running)
+		perf_event_enable(event);
+
+	return 0;
+}
+
+static void nmi_timer_stop_cpu(int cpu)
+{
+	struct perf_event *event = per_cpu(nmi_timer_events, cpu);
+
+	if (event && ctr_running)
+		perf_event_disable(event);
+}
+
+static int nmi_timer_cpu_notifier(struct notifier_block *b, unsigned long action,
+				  void *data)
+{
+	int cpu = (unsigned long)data;
+	switch (action) {
+	case CPU_DOWN_FAILED:
+	case CPU_ONLINE:
+		nmi_timer_start_cpu(cpu);
+		break;
+	case CPU_DOWN_PREPARE:
+		nmi_timer_stop_cpu(cpu);
+		break;
+	}
+	return NOTIFY_DONE;
+}
+
+static struct notifier_block nmi_timer_cpu_nb = {
+	.notifier_call = nmi_timer_cpu_notifier
+};
+
+static int nmi_timer_start(void)
+{
+	int cpu;
+
+	get_online_cpus();
+	ctr_running = 1;
+	for_each_online_cpu(cpu)
+		nmi_timer_start_cpu(cpu);
+	put_online_cpus();
+
+	return 0;
+}
+
+static void nmi_timer_stop(void)
+{
+	int cpu;
+
+	get_online_cpus();
+	for_each_online_cpu(cpu)
+		nmi_timer_stop_cpu(cpu);
+	ctr_running = 0;
+	put_online_cpus();
+}
+
+static void nmi_timer_shutdown(void)
+{
+	struct perf_event *event;
+	int cpu;
+
+	get_online_cpus();
+	unregister_cpu_notifier(&nmi_timer_cpu_nb);
+	for_each_possible_cpu(cpu) {
+		event = per_cpu(nmi_timer_events, cpu);
+		if (!event)
+			continue;
+		perf_event_disable(event);
+		per_cpu(nmi_timer_events, cpu) = NULL;
+		perf_event_release_kernel(event);
+	}
+
+	put_online_cpus();
+}
+
+static int nmi_timer_setup(void)
+{
+	int cpu, err;
+	u64 period;
+
+	/* clock cycles per tick: */
+	period = (u64)cpu_khz * 1000;
+	do_div(period, HZ);
+	nmi_timer_attr.sample_period = period;
+
+	get_online_cpus();
+	err = register_cpu_notifier(&nmi_timer_cpu_nb);
+	if (err)
+		goto out;
+	/* can't attach events to offline cpus: */
+	for_each_online_cpu(cpu) {
+		err = nmi_timer_start_cpu(cpu);
+		if (err)
+			break;
+	}
+	if (err)
+		nmi_timer_shutdown();
+out:
+	put_online_cpus();
+	return err;
+}
+
+int __init op_nmi_timer_init(struct oprofile_operations *ops)
+{
+	int err = 0;
+
+	err = nmi_timer_setup();
+	if (err)
+		return err;
+	nmi_timer_shutdown();		/* only check, don't alloc */
+
+	ops->create_files	= NULL;
+	ops->setup		= nmi_timer_setup;
+	ops->shutdown		= nmi_timer_shutdown;
+	ops->start		= nmi_timer_start;
+	ops->stop		= nmi_timer_stop;
+	ops->cpu_type		= "timer";
+
+	printk(KERN_INFO "oprofile: using NMI timer interrupt.\n");
+
+	return 0;
+}
+
+#endif

drivers/oprofile/oprof.c

@@ -246,26 +246,24 @@ static int __init oprofile_init(void)
 	int err;
 
 	/* always init architecture to setup backtrace support */
+	timer_mode = 0;
 	err = oprofile_arch_init(&oprofile_ops);
+	if (!err) {
+		if (!timer && !oprofilefs_register())
+			return 0;
+		oprofile_arch_exit();
+	}
 
-	timer_mode = err || timer;	/* fall back to timer mode on errors */
-	if (timer_mode) {
-		if (!err)
-			oprofile_arch_exit();
+	/* setup timer mode: */
+	timer_mode = 1;
+	/* no nmi timer mode if oprofile.timer is set */
+	if (timer || op_nmi_timer_init(&oprofile_ops)) {
 		err = oprofile_timer_init(&oprofile_ops);
 		if (err)
 			return err;
 	}
 
-	err = oprofilefs_register();
-	if (!err)
-		return 0;
-
-	/* failed */
-	if (!timer_mode)
-		oprofile_arch_exit();
-
-	return err;
+	return oprofilefs_register();
 }
 
 

drivers/oprofile/oprof.h

@@ -36,6 +36,15 @@ struct dentry;
 void oprofile_create_files(struct super_block *sb, struct dentry *root);
 int oprofile_timer_init(struct oprofile_operations *ops);
 void oprofile_timer_exit(void);
+#ifdef CONFIG_OPROFILE_NMI_TIMER
+int op_nmi_timer_init(struct oprofile_operations *ops);
+#else
+static inline int op_nmi_timer_init(struct oprofile_operations *ops)
+{
+	return -ENODEV;
+}
+#endif
+
 
 int oprofile_set_ulong(unsigned long *addr, unsigned long val);
 int oprofile_set_timeout(unsigned long time);

kernel/events/core.c

@@ -1322,6 +1322,7 @@ void perf_event_disable(struct perf_event *event)
 	}
 	raw_spin_unlock_irq(&ctx->lock);
 }
+EXPORT_SYMBOL_GPL(perf_event_disable);
 
 static void perf_set_shadow_time(struct perf_event *event,
 				 struct perf_event_context *ctx,
@@ -1806,6 +1807,7 @@ void perf_event_enable(struct perf_event *event)
 out:
 	raw_spin_unlock_irq(&ctx->lock);
 }
+EXPORT_SYMBOL_GPL(perf_event_enable);
 
 int perf_event_refresh(struct perf_event *event, int refresh)
 {