tracing: Allocate the ftrace event profile buffer dynamically

Currently the trace event profile buffer is allocated in the stack. But
this may be too much for the stack, as the events can have large
statically defined field size and can also grow with dynamic arrays.

Allocate two per cpu buffer for all profiled events. The first cpu
buffer is used to host every non-nmi context traces. It is protected
by disabling the interrupts while writing and committing the trace.

The second buffer is reserved for nmi. So that there is no race between
them and the first buffer.

The whole write/commit section is rcu protected because we release
these buffers while deactivating the last profiling trace event.

v2: Move the buffers from trace_event to be global, as pointed by
    Steven Rostedt.

v3: Fix the syscall events to handle the profiling buffer races
    by disabling interrupts, now that the buffers are globals.

Suggested-by: Steven Rostedt <rostedt@goodmis.org>
Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Jason Baron <jbaron@redhat.com>
Cc: Masami Hiramatsu <mhiramat@redhat.com>
Cc: Ingo Molnar <mingo@elte.hu>

include/linux/ftrace_event.h

@@ -4,6 +4,7 @@
 #include <linux/ring_buffer.h>
 #include <linux/trace_seq.h>
 #include <linux/percpu.h>
+#include <linux/hardirq.h>
 
 struct trace_array;
 struct tracer;
@@ -134,6 +135,11 @@ struct ftrace_event_call {
 	void			(*profile_disable)(void);
 };
 
+#define FTRACE_MAX_PROFILE_SIZE	2048
+
+extern char			*trace_profile_buf;
+extern char			*trace_profile_buf_nmi;
+
 #define MAX_FILTER_PRED		32
 #define MAX_FILTER_STR_VAL	256	/* Should handle KSYM_SYMBOL_LEN */
 

include/trace/ftrace.h

@@ -648,11 +648,12 @@ __attribute__((section("_ftrace_events"))) event_##call = {		\
  *	struct ftrace_raw_##call *entry;
  *	u64 __addr = 0, __count = 1;
  *	unsigned long irq_flags;
+ *	struct trace_entry *ent;
  *	int __entry_size;
  *	int __data_size;
+ *	int __cpu
  *	int pc;
  *
- *	local_save_flags(irq_flags);
  *	pc = preempt_count();
  *
  *	__data_size = ftrace_get_offsets_<call>(&__data_offsets, args);
@@ -663,25 +664,34 @@ __attribute__((section("_ftrace_events"))) event_##call = {		\
  *			     sizeof(u64));
  *	__entry_size -= sizeof(u32);
  *
- *	do {
- *		char raw_data[__entry_size]; <- allocate our sample in the stack
- *		struct trace_entry *ent;
+ *	// Protect the non nmi buffer
+ *	// This also protects the rcu read side
+ *	local_irq_save(irq_flags);
+ *	__cpu = smp_processor_id();
+ *
+ *	if (in_nmi())
+ *		raw_data = rcu_dereference(trace_profile_buf_nmi);
+ *	else
+ *		raw_data = rcu_dereference(trace_profile_buf);
+ *
+ *	if (!raw_data)
+ *		goto end;
  *
- *		zero dead bytes from alignment to avoid stack leak to userspace:
+ *	raw_data = per_cpu_ptr(raw_data, __cpu);
  *
- *		*(u64 *)(&raw_data[__entry_size - sizeof(u64)]) = 0ULL;
- *		entry = (struct ftrace_raw_<call> *)raw_data;
- *		ent = &entry->ent;
- *		tracing_generic_entry_update(ent, irq_flags, pc);
- *		ent->type = event_call->id;
+ *	//zero dead bytes from alignment to avoid stack leak to userspace:
+ *	*(u64 *)(&raw_data[__entry_size - sizeof(u64)]) = 0ULL;
+ *	entry = (struct ftrace_raw_<call> *)raw_data;
+ *	ent = &entry->ent;
+ *	tracing_generic_entry_update(ent, irq_flags, pc);
+ *	ent->type = event_call->id;
  *
- *		<tstruct> <- do some jobs with dynamic arrays
+ *	<tstruct> <- do some jobs with dynamic arrays
  *
- *		<assign>  <- affect our values
+ *	<assign>  <- affect our values
  *
- *		perf_tpcounter_event(event_call->id, __addr, __count, entry,
- *			     __entry_size);  <- submit them to perf counter
- *	} while (0);
+ *	perf_tpcounter_event(event_call->id, __addr, __count, entry,
+ *		     __entry_size);  <- submit them to perf counter
  *
  * }
  */
@@ -704,11 +714,13 @@ static void ftrace_profile_##call(proto)				\
 	struct ftrace_raw_##call *entry;				\
 	u64 __addr = 0, __count = 1;					\
 	unsigned long irq_flags;					\
+	struct trace_entry *ent;					\
 	int __entry_size;						\
 	int __data_size;						\
+	char *raw_data;							\
+	int __cpu;							\
 	int pc;								\
 									\
-	local_save_flags(irq_flags);					\
 	pc = preempt_count();						\
 									\
 	__data_size = ftrace_get_offsets_##call(&__data_offsets, args); \
@@ -716,23 +728,38 @@ static void ftrace_profile_##call(proto)				\
 			     sizeof(u64));				\
 	__entry_size -= sizeof(u32);					\
 									\
-	do {								\
-		char raw_data[__entry_size];				\
-		struct trace_entry *ent;				\
+	if (WARN_ONCE(__entry_size > FTRACE_MAX_PROFILE_SIZE,		\
+		      "profile buffer not large enough"))		\
+		return;							\
+									\
+	local_irq_save(irq_flags);					\
+	__cpu = smp_processor_id();					\
 									\
-		*(u64 *)(&raw_data[__entry_size - sizeof(u64)]) = 0ULL;	\
-		entry = (struct ftrace_raw_##call *)raw_data;		\
-		ent = &entry->ent;					\
-		tracing_generic_entry_update(ent, irq_flags, pc);	\
-		ent->type = event_call->id;				\
+	if (in_nmi())							\
+		raw_data = rcu_dereference(trace_profile_buf_nmi);		\
+	else								\
+		raw_data = rcu_dereference(trace_profile_buf);		\
 									\
-		tstruct							\
+	if (!raw_data)							\
+		goto end;						\
 									\
-		{ assign; }						\
+	raw_data = per_cpu_ptr(raw_data, __cpu);			\
 									\
-		perf_tpcounter_event(event_call->id, __addr, __count, entry,\
+	*(u64 *)(&raw_data[__entry_size - sizeof(u64)]) = 0ULL;		\
+	entry = (struct ftrace_raw_##call *)raw_data;			\
+	ent = &entry->ent;						\
+	tracing_generic_entry_update(ent, irq_flags, pc);		\
+	ent->type = event_call->id;					\
+									\
+	tstruct								\
+									\
+	{ assign; }							\
+									\
+	perf_tpcounter_event(event_call->id, __addr, __count, entry,	\
 			     __entry_size);				\
-	} while (0);							\
+									\
+end:									\
+	local_irq_restore(irq_flags);					\
 									\
 }
 

kernel/trace/trace_event_profile.c

@@ -8,12 +8,52 @@
 #include <linux/module.h>
 #include "trace.h"
 
+/*
+ * We can't use a size but a type in alloc_percpu()
+ * So let's create a dummy type that matches the desired size
+ */
+typedef struct {char buf[FTRACE_MAX_PROFILE_SIZE];} profile_buf_t;
+
+char		*trace_profile_buf;
+char 		*trace_profile_buf_nmi;
+
+/* Count the events in use (per event id, not per instance) */
+static int	total_profile_count;
+
 static int ftrace_profile_enable_event(struct ftrace_event_call *event)
 {
+	char *buf;
+	int ret = -ENOMEM;
+
 	if (atomic_inc_return(&event->profile_count))
 		return 0;
 
-	return event->profile_enable();
+	if (!total_profile_count++) {
+		buf = (char *)alloc_percpu(profile_buf_t);
+		if (!buf)
+			goto fail_buf;
+
+		rcu_assign_pointer(trace_profile_buf, buf);
+
+		buf = (char *)alloc_percpu(profile_buf_t);
+		if (!buf)
+			goto fail_buf_nmi;
+
+		rcu_assign_pointer(trace_profile_buf_nmi, buf);
+	}
+
+	ret = event->profile_enable();
+	if (!ret)
+		return 0;
+
+	kfree(trace_profile_buf_nmi);
+fail_buf_nmi:
+	kfree(trace_profile_buf);
+fail_buf:
+	total_profile_count--;
+	atomic_dec(&event->profile_count);
+
+	return ret;
 }
 
 int ftrace_profile_enable(int event_id)
@@ -36,10 +76,29 @@ int ftrace_profile_enable(int event_id)
 
 static void ftrace_profile_disable_event(struct ftrace_event_call *event)
 {
+	char *buf, *nmi_buf;
+
 	if (!atomic_add_negative(-1, &event->profile_count))
 		return;
 
 	event->profile_disable();
+
+	if (!--total_profile_count) {
+		buf = trace_profile_buf;
+		rcu_assign_pointer(trace_profile_buf, NULL);
+
+		nmi_buf = trace_profile_buf_nmi;
+		rcu_assign_pointer(trace_profile_buf_nmi, NULL);
+
+		/*
+		 * Ensure every events in profiling have finished before
+		 * releasing the buffers
+		 */
+		synchronize_sched();
+
+		free_percpu(buf);
+		free_percpu(nmi_buf);
+	}
 }
 
 void ftrace_profile_disable(int event_id)

kernel/trace/trace_syscalls.c

@@ -384,10 +384,13 @@ static int sys_prof_refcount_exit;
 
 static void prof_syscall_enter(struct pt_regs *regs, long id)
 {
-	struct syscall_trace_enter *rec;
 	struct syscall_metadata *sys_data;
+	struct syscall_trace_enter *rec;
+	unsigned long flags;
+	char *raw_data;
 	int syscall_nr;
 	int size;
+	int cpu;
 
 	syscall_nr = syscall_get_nr(current, regs);
 	if (!test_bit(syscall_nr, enabled_prof_enter_syscalls))
@@ -402,20 +405,38 @@ static void prof_syscall_enter(struct pt_regs *regs, long id)
 	size = ALIGN(size + sizeof(u32), sizeof(u64));
 	size -= sizeof(u32);
 
-	do {
-		char raw_data[size];
+	if (WARN_ONCE(size > FTRACE_MAX_PROFILE_SIZE,
+		      "profile buffer not large enough"))
+		return;
+
+	/* Protect the per cpu buffer, begin the rcu read side */
+	local_irq_save(flags);
 
-		/* zero the dead bytes from align to not leak stack to user */
-		*(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
+	cpu = smp_processor_id();
+
+	if (in_nmi())
+		raw_data = rcu_dereference(trace_profile_buf_nmi);
+	else
+		raw_data = rcu_dereference(trace_profile_buf);
+
+	if (!raw_data)
+		goto end;
 
-		rec = (struct syscall_trace_enter *) raw_data;
-		tracing_generic_entry_update(&rec->ent, 0, 0);
-		rec->ent.type = sys_data->enter_id;
-		rec->nr = syscall_nr;
-		syscall_get_arguments(current, regs, 0, sys_data->nb_args,
-				       (unsigned long *)&rec->args);
-		perf_tpcounter_event(sys_data->enter_id, 0, 1, rec, size);
-	} while(0);
+	raw_data = per_cpu_ptr(raw_data, cpu);
+
+	/* zero the dead bytes from align to not leak stack to user */
+	*(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
+
+	rec = (struct syscall_trace_enter *) raw_data;
+	tracing_generic_entry_update(&rec->ent, 0, 0);
+	rec->ent.type = sys_data->enter_id;
+	rec->nr = syscall_nr;
+	syscall_get_arguments(current, regs, 0, sys_data->nb_args,
+			       (unsigned long *)&rec->args);
+	perf_tpcounter_event(sys_data->enter_id, 0, 1, rec, size);
+
+end:
+	local_irq_restore(flags);
 }
 
 int reg_prof_syscall_enter(char *name)
@@ -460,8 +481,12 @@ void unreg_prof_syscall_enter(char *name)
 static void prof_syscall_exit(struct pt_regs *regs, long ret)
 {
 	struct syscall_metadata *sys_data;
-	struct syscall_trace_exit rec;
+	struct syscall_trace_exit *rec;
+	unsigned long flags;
 	int syscall_nr;
+	char *raw_data;
+	int size;
+	int cpu;
 
 	syscall_nr = syscall_get_nr(current, regs);
 	if (!test_bit(syscall_nr, enabled_prof_exit_syscalls))
@@ -471,12 +496,46 @@ static void prof_syscall_exit(struct pt_regs *regs, long ret)
 	if (!sys_data)
 		return;
 
-	tracing_generic_entry_update(&rec.ent, 0, 0);
-	rec.ent.type = sys_data->exit_id;
-	rec.nr = syscall_nr;
-	rec.ret = syscall_get_return_value(current, regs);
+	/* We can probably do that at build time */
+	size = ALIGN(sizeof(*rec) + sizeof(u32), sizeof(u64));
+	size -= sizeof(u32);
 
-	perf_tpcounter_event(sys_data->exit_id, 0, 1, &rec, sizeof(rec));
+	/*
+	 * Impossible, but be paranoid with the future
+	 * How to put this check outside runtime?
+	 */
+	if (WARN_ONCE(size > FTRACE_MAX_PROFILE_SIZE,
+		"exit event has grown above profile buffer size"))
+		return;
+
+	/* Protect the per cpu buffer, begin the rcu read side */
+	local_irq_save(flags);
+	cpu = smp_processor_id();
+
+	if (in_nmi())
+		raw_data = rcu_dereference(trace_profile_buf_nmi);
+	else
+		raw_data = rcu_dereference(trace_profile_buf);
+
+	if (!raw_data)
+		goto end;
+
+	raw_data = per_cpu_ptr(raw_data, cpu);
+
+	/* zero the dead bytes from align to not leak stack to user */
+	*(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
+
+	rec = (struct syscall_trace_exit *)raw_data;
+
+	tracing_generic_entry_update(&rec->ent, 0, 0);
+	rec->ent.type = sys_data->exit_id;
+	rec->nr = syscall_nr;
+	rec->ret = syscall_get_return_value(current, regs);
+
+	perf_tpcounter_event(sys_data->exit_id, 0, 1, rec, size);
+
+end:
+	local_irq_restore(flags);
 }
 
 int reg_prof_syscall_exit(char *name)