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>

2 files changed, 138 insertions, 20 deletions

diff --git a/kernel/trace/trace_event_profile.c b/kernel/trace/trace_event_profile.c
index df4a74efd50..3aaa77c3309 100644
--- a/kernel/trace/trace_event_profile.c
+++ b/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)
diff --git a/kernel/trace/trace_syscalls.c b/kernel/trace/trace_syscalls.c
index 8712ce3c6a0..7a3550cf259 100644
--- a/kernel/trace/trace_syscalls.c
+++ b/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)