x86, bts, hw-branch-tracer: add _noirq variants to the debug store interface

The hw-branch-tracer uses debug store functions from an on_each_cpu()
context, which is simply wrong since the functions may sleep.

Add _noirq variants for most functions, which  may be called with
interrupts disabled.

Separate per-cpu and per-task tracing and allow per-cpu tracing to be
controlled from any cpu.

Make the hw-branch-tracer use the new debug store interface, synchronize
with hotplug cpu event using get/put_online_cpus(), and remove the
unnecessary spinlock.

Make the ptrace bts and the ds selftest code use the new interface.

Defer the ds selftest.
Signed-off-by: Markus Metzger <markus.t.metzger@intel.com>
Cc: roland@redhat.com
Cc: eranian@googlemail.com
Cc: oleg@redhat.com
Cc: juan.villacis@intel.com
Cc: ak@linux.jf.intel.com
LKML-Reference: <20090403144555.658136000@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

arch/x86/include/asm/ds.h

@@ -15,8 +15,8 @@
  * - buffer allocation (memory accounting)
  *
  *
- * Copyright (C) 2007-2008 Intel Corporation.
- * Markus Metzger <markus.t.metzger@intel.com>, 2007-2008
+ * Copyright (C) 2007-2009 Intel Corporation.
+ * Markus Metzger <markus.t.metzger@intel.com>, 2007-2009
  */
 
 #ifndef _ASM_X86_DS_H
@@ -83,8 +83,10 @@ enum ds_feature {
  * The interrupt threshold is independent from the overflow callback
  * to allow users to use their own overflow interrupt handling mechanism.
  *
- * task: the task to request recording for;
- *       NULL for per-cpu recording on the current cpu
+ * The function might sleep.
+ *
+ * task: the task to request recording for
+ * cpu:  the cpu to request recording for
  * base: the base pointer for the (non-pageable) buffer;
  * size: the size of the provided buffer in bytes
  * ovfl: pointer to a function to be called on buffer overflow;
@@ -93,19 +95,28 @@ enum ds_feature {
  *     -1 if no interrupt threshold is requested.
  * flags: a bit-mask of the above flags
  */
-extern struct bts_tracer *ds_request_bts(struct task_struct *task,
-					 void *base, size_t size,
-					 bts_ovfl_callback_t ovfl,
-					 size_t th, unsigned int flags);
-extern struct pebs_tracer *ds_request_pebs(struct task_struct *task,
-					   void *base, size_t size,
-					   pebs_ovfl_callback_t ovfl,
-					   size_t th, unsigned int flags);
+extern struct bts_tracer *ds_request_bts_task(struct task_struct *task,
+					      void *base, size_t size,
+					      bts_ovfl_callback_t ovfl,
+					      size_t th, unsigned int flags);
+extern struct bts_tracer *ds_request_bts_cpu(int cpu, void *base, size_t size,
+					     bts_ovfl_callback_t ovfl,
+					     size_t th, unsigned int flags);
+extern struct pebs_tracer *ds_request_pebs_task(struct task_struct *task,
+						void *base, size_t size,
+						pebs_ovfl_callback_t ovfl,
+						size_t th, unsigned int flags);
+extern struct pebs_tracer *ds_request_pebs_cpu(int cpu,
+					       void *base, size_t size,
+					       pebs_ovfl_callback_t ovfl,
+					       size_t th, unsigned int flags);
 
 /*
  * Release BTS or PEBS resources
  * Suspend and resume BTS or PEBS tracing
  *
+ * Must be called with irq's enabled.
+ *
  * tracer: the tracer handle returned from ds_request_~()
  */
 extern void ds_release_bts(struct bts_tracer *tracer);
@@ -115,6 +126,28 @@ extern void ds_release_pebs(struct pebs_tracer *tracer);
 extern void ds_suspend_pebs(struct pebs_tracer *tracer);
 extern void ds_resume_pebs(struct pebs_tracer *tracer);
 
+/*
+ * Release BTS or PEBS resources
+ * Suspend and resume BTS or PEBS tracing
+ *
+ * Cpu tracers must call this on the traced cpu.
+ * Task tracers must call ds_release_~_noirq() for themselves.
+ *
+ * May be called with irq's disabled.
+ *
+ * Returns 0 if successful;
+ * -EPERM if the cpu tracer does not trace the current cpu.
+ * -EPERM if the task tracer does not trace itself.
+ *
+ * tracer: the tracer handle returned from ds_request_~()
+ */
+extern int ds_release_bts_noirq(struct bts_tracer *tracer);
+extern int ds_suspend_bts_noirq(struct bts_tracer *tracer);
+extern int ds_resume_bts_noirq(struct bts_tracer *tracer);
+extern int ds_release_pebs_noirq(struct pebs_tracer *tracer);
+extern int ds_suspend_pebs_noirq(struct pebs_tracer *tracer);
+extern int ds_resume_pebs_noirq(struct pebs_tracer *tracer);
+
 
 /*
  * The raw DS buffer state as it is used for BTS and PEBS recording.

arch/x86/kernel/ds_selftest.c

@@ -10,11 +10,12 @@
 
 #include <linux/kernel.h>
 #include <linux/string.h>
+#include <linux/smp.h>
 
 #include <asm/ds.h>
 
 
-#define DS_SELFTEST_BUFFER_SIZE 1021 /* Intentionally chose an odd size. */
+#define BUFFER_SIZE 1021 /* Intentionally chose an odd size. */
 
 
 static int ds_selftest_bts_consistency(const struct bts_trace *trace)
@@ -125,12 +126,12 @@ int ds_selftest_bts(void)
 	struct bts_tracer *tracer;
 	int error = 0;
 	void *top;
-	unsigned char buffer[DS_SELFTEST_BUFFER_SIZE];
+	unsigned char buffer[BUFFER_SIZE];
 
 	printk(KERN_INFO "[ds] bts selftest...");
 
-	tracer = ds_request_bts(NULL, buffer, DS_SELFTEST_BUFFER_SIZE,
-				NULL, (size_t)-1, BTS_KERNEL);
+	tracer = ds_request_bts_cpu(smp_processor_id(), buffer, BUFFER_SIZE,
+				    NULL, (size_t)-1, BTS_KERNEL);
 	if (IS_ERR(tracer)) {
 		error = PTR_ERR(tracer);
 		tracer = NULL;

arch/x86/kernel/ptrace.c

@@ -800,8 +800,9 @@ static int ptrace_bts_config(struct task_struct *child,
 	if (cfg.flags & PTRACE_BTS_O_SCHED)
 		flags |= BTS_TIMESTAMPS;
 
-	context->tracer = ds_request_bts(child, context->buffer, context->size,
-					 NULL, (size_t)-1, flags);
+	context->tracer =
+		ds_request_bts_task(child, context->buffer, context->size,
+				    NULL, (size_t)-1, flags);
 	if (unlikely(IS_ERR(context->tracer))) {
 		int error = PTR_ERR(context->tracer);
 

kernel/trace/trace_hw_branches.c

@@ -4,7 +4,6 @@
  * Copyright (C) 2008-2009 Intel Corporation.
  * Markus Metzger <markus.t.metzger@gmail.com>, 2008-2009
  */
-#include <linux/spinlock.h>
 #include <linux/kallsyms.h>
 #include <linux/debugfs.h>
 #include <linux/ftrace.h>
@@ -21,168 +20,113 @@
 
 #define BTS_BUFFER_SIZE (1 << 13)
 
-/*
- * The tracer lock protects the below per-cpu tracer array.
- * It needs to be held to:
- * - start tracing on all cpus
- * - stop tracing on all cpus
- * - start tracing on a single hotplug cpu
- * - stop tracing on a single hotplug cpu
- * - read the trace from all cpus
- * - read the trace from a single cpu
- */
-static DEFINE_SPINLOCK(bts_tracer_lock);
 static DEFINE_PER_CPU(struct bts_tracer *, tracer);
 static DEFINE_PER_CPU(unsigned char[BTS_BUFFER_SIZE], buffer);
 
 #define this_tracer per_cpu(tracer, smp_processor_id())
-#define this_buffer per_cpu(buffer, smp_processor_id())
 
 static int trace_hw_branches_enabled __read_mostly;
 static int trace_hw_branches_suspended __read_mostly;
 static struct trace_array *hw_branch_trace __read_mostly;
 
 
-/*
- * Initialize the tracer for the current cpu.
- * The argument is ignored.
- *
- * pre: bts_tracer_lock must be locked.
- */
-static void bts_trace_init_cpu(void *arg)
+static void bts_trace_init_cpu(int cpu)
 {
-	if (this_tracer)
-		ds_release_bts(this_tracer);
+	per_cpu(tracer, cpu) =
+		ds_request_bts_cpu(cpu, per_cpu(buffer, cpu), BTS_BUFFER_SIZE,
+				   NULL, (size_t)-1, BTS_KERNEL);
 
-	this_tracer = ds_request_bts(NULL, this_buffer, BTS_BUFFER_SIZE,
-				     NULL, (size_t)-1, BTS_KERNEL);
-	if (IS_ERR(this_tracer)) {
-		this_tracer = NULL;
-		return;
-	}
+	if (IS_ERR(per_cpu(tracer, cpu)))
+		per_cpu(tracer, cpu) = NULL;
 }
 
 static int bts_trace_init(struct trace_array *tr)
 {
-	int cpu, avail;
-
-	spin_lock(&bts_tracer_lock);
+	int cpu;
 
 	hw_branch_trace = tr;
+	trace_hw_branches_enabled = 0;
 
-	on_each_cpu(bts_trace_init_cpu, NULL, 1);
-
-	/* Check on how many cpus we could enable tracing */
-	avail = 0;
-	for_each_online_cpu(cpu)
-		if (per_cpu(tracer, cpu))
-			avail++;
+	get_online_cpus();
+	for_each_online_cpu(cpu) {
+		bts_trace_init_cpu(cpu);
 
-	trace_hw_branches_enabled = (avail ? 1 : 0);
+		if (likely(per_cpu(tracer, cpu)))
+			trace_hw_branches_enabled = 1;
+	}
 	trace_hw_branches_suspended = 0;
-
-	spin_unlock(&bts_tracer_lock);
-
+	put_online_cpus();
 
 	/* If we could not enable tracing on a single cpu, we fail. */
-	return avail ? 0 : -EOPNOTSUPP;
-}
-
-/*
- * Release the tracer for the current cpu.
- * The argument is ignored.
- *
- * pre: bts_tracer_lock must be locked.
- */
-static void bts_trace_release_cpu(void *arg)
-{
-	if (this_tracer) {
-		ds_release_bts(this_tracer);
-		this_tracer = NULL;
-	}
+	return trace_hw_branches_enabled ? 0 : -EOPNOTSUPP;
 }
 
 static void bts_trace_reset(struct trace_array *tr)
 {
-	spin_lock(&bts_tracer_lock);
+	int cpu;
 
-	on_each_cpu(bts_trace_release_cpu, NULL, 1);
+	get_online_cpus();
+	for_each_online_cpu(cpu) {
+		if (likely(per_cpu(tracer, cpu))) {
+			ds_release_bts(per_cpu(tracer, cpu));
+			per_cpu(tracer, cpu) = NULL;
+		}
+	}
 	trace_hw_branches_enabled = 0;
 	trace_hw_branches_suspended = 0;
-
-	spin_unlock(&bts_tracer_lock);
-}
-
-/*
- * Resume tracing on the current cpu.
- * The argument is ignored.
- *
- * pre: bts_tracer_lock must be locked.
- */
-static void bts_trace_resume_cpu(void *arg)
-{
-	if (this_tracer)
-		ds_resume_bts(this_tracer);
+	put_online_cpus();
 }
 
 static void bts_trace_start(struct trace_array *tr)
 {
-	spin_lock(&bts_tracer_lock);
+	int cpu;
 
-	on_each_cpu(bts_trace_resume_cpu, NULL, 1);
+	get_online_cpus();
+	for_each_online_cpu(cpu)
+		if (likely(per_cpu(tracer, cpu)))
+			ds_resume_bts(per_cpu(tracer, cpu));
 	trace_hw_branches_suspended = 0;
-
-	spin_unlock(&bts_tracer_lock);
-}
-
-/*
- * Suspend tracing on the current cpu.
- * The argument is ignored.
- *
- * pre: bts_tracer_lock must be locked.
- */
-static void bts_trace_suspend_cpu(void *arg)
-{
-	if (this_tracer)
-		ds_suspend_bts(this_tracer);
+	put_online_cpus();
 }
 
 static void bts_trace_stop(struct trace_array *tr)
 {
-	spin_lock(&bts_tracer_lock);
+	int cpu;
 
-	on_each_cpu(bts_trace_suspend_cpu, NULL, 1);
+	get_online_cpus();
+	for_each_online_cpu(cpu)
+		if (likely(per_cpu(tracer, cpu)))
+			ds_suspend_bts(per_cpu(tracer, cpu));
 	trace_hw_branches_suspended = 1;
-
-	spin_unlock(&bts_tracer_lock);
+	put_online_cpus();
 }
 
 static int __cpuinit bts_hotcpu_handler(struct notifier_block *nfb,
 				     unsigned long action, void *hcpu)
 {
-	unsigned int cpu = (unsigned long)hcpu;
-
-	spin_lock(&bts_tracer_lock);
-
-	if (!trace_hw_branches_enabled)
-		goto out;
+	int cpu = (long)hcpu;
 
 	switch (action) {
 	case CPU_ONLINE:
 	case CPU_DOWN_FAILED:
-		smp_call_function_single(cpu, bts_trace_init_cpu, NULL, 1);
-
-		if (trace_hw_branches_suspended)
-			smp_call_function_single(cpu, bts_trace_suspend_cpu,
-						 NULL, 1);
+		/* The notification is sent with interrupts enabled. */
+		if (trace_hw_branches_enabled) {
+			bts_trace_init_cpu(cpu);
+
+			if (trace_hw_branches_suspended &&
+			    likely(per_cpu(tracer, cpu)))
+				ds_suspend_bts(per_cpu(tracer, cpu));
+		}
 		break;
+
 	case CPU_DOWN_PREPARE:
-		smp_call_function_single(cpu, bts_trace_release_cpu, NULL, 1);
-		break;
+		/* The notification is sent with interrupts enabled. */
+		if (likely(per_cpu(tracer, cpu))) {
+			ds_release_bts(per_cpu(tracer, cpu));
+			per_cpu(tracer, cpu) = NULL;
+		}
 	}
 
- out:
-	spin_unlock(&bts_tracer_lock);
 	return NOTIFY_DONE;
 }
 
@@ -274,7 +218,7 @@ static void trace_bts_at(const struct bts_trace *trace, void *at)
 /*
  * Collect the trace on the current cpu and write it into the ftrace buffer.
  *
- * pre: bts_tracer_lock must be locked
+ * pre: tracing must be suspended on the current cpu
  */
 static void trace_bts_cpu(void *arg)
 {
@@ -291,10 +235,9 @@ static void trace_bts_cpu(void *arg)
 	if (unlikely(!this_tracer))
 		return;
 
-	ds_suspend_bts(this_tracer);
 	trace = ds_read_bts(this_tracer);
 	if (!trace)
-		goto out;
+		return;
 
 	for (at = trace->ds.top; (void *)at < trace->ds.end;
 	     at += trace->ds.size)
@@ -303,18 +246,27 @@ static void trace_bts_cpu(void *arg)
 	for (at = trace->ds.begin; (void *)at < trace->ds.top;
 	     at += trace->ds.size)
 		trace_bts_at(trace, at);
-
-out:
-	ds_resume_bts(this_tracer);
 }
 
 static void trace_bts_prepare(struct trace_iterator *iter)
 {
-	spin_lock(&bts_tracer_lock);
+	int cpu;
 
+	get_online_cpus();
+	for_each_online_cpu(cpu)
+		if (likely(per_cpu(tracer, cpu)))
+			ds_suspend_bts(per_cpu(tracer, cpu));
+	/*
+	 * We need to collect the trace on the respective cpu since ftrace
+	 * implicitly adds the record for the current cpu.
+	 * Once that is more flexible, we could collect the data from any cpu.
+	 */
 	on_each_cpu(trace_bts_cpu, iter->tr, 1);
 
-	spin_unlock(&bts_tracer_lock);
+	for_each_online_cpu(cpu)
+		if (likely(per_cpu(tracer, cpu)))
+			ds_resume_bts(per_cpu(tracer, cpu));
+	put_online_cpus();
 }
 
 static void trace_bts_close(struct trace_iterator *iter)
@@ -324,12 +276,11 @@ static void trace_bts_close(struct trace_iterator *iter)
 
 void trace_hw_branch_oops(void)
 {
-	spin_lock(&bts_tracer_lock);
-
-	if (trace_hw_branches_enabled)
+	if (this_tracer) {
+		ds_suspend_bts_noirq(this_tracer);
 		trace_bts_cpu(hw_branch_trace);
-
-	spin_unlock(&bts_tracer_lock);
+		ds_resume_bts_noirq(this_tracer);
+	}
 }
 
 struct tracer bts_tracer __read_mostly =