tracing, perf_events: Protect the buffer from recursion in perf

While tracing using events with perf, if one enables the
lockdep:lock_acquire event, it will infect every other perf
trace events.

Basically, you can enable whatever set of trace events through
perf but if this event is part of the set, the only result we
can get is a long list of lock_acquire events of rcu read lock,
and only that.

This is because of a recursion inside perf.

1) When a trace event is triggered, it will fill a per cpu
   buffer and submit it to perf.

2) Perf will commit this event but will also protect some data
   using rcu_read_lock

3) A recursion appears: rcu_read_lock triggers a lock_acquire
   event that will fill the per cpu event and then submit the
   buffer to perf.

4) Perf detects a recursion and ignores it

5) Perf continues its work on the previous event, but its buffer
   has been overwritten by the lock_acquire event, it has then
   been turned into a lock_acquire event of rcu read lock

Such scenario also happens with lock_release with
rcu_read_unlock().

We could turn the rcu_read_lock() into __rcu_read_lock() to drop
the lock debugging from perf fast path, but that would make us
lose the rcu debugging and that doesn't prevent from other
possible kind of recursion from perf in the future.

This patch adds a recursion protection based on a counter on the
perf trace per cpu buffers to solve the problem.

-v2: Fixed lost whitespace, added reviewed-by tag

Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
Reviewed-by: Masami Hiramatsu <mhiramat@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Li Zefan <lizf@cn.fujitsu.com>
Cc: Jason Baron <jbaron@redhat.com>
LKML-Reference: <1257477185-7838-1-git-send-email-fweisbec@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

1 files changed, 36 insertions, 8 deletions

diff --git a/kernel/trace/trace_syscalls.c b/kernel/trace/trace_syscalls.c
index 58b8e5370767..51213b0aa81b 100644
--- a/kernel/trace/trace_syscalls.c
+++ b/kernel/trace/trace_syscalls.c
@@ -477,6 +477,7 @@ static int sys_prof_refcount_exit;
 static void prof_syscall_enter(struct pt_regs *regs, long id)
 {
         struct syscall_metadata *sys_data;
+        struct perf_trace_buf *trace_buf;
         struct syscall_trace_enter *rec;
         unsigned long flags;
         char *raw_data;
@@ -507,14 +508,25 @@ static void prof_syscall_enter(struct pt_regs *regs, long id)
         cpu = smp_processor_id();
 
         if (in_nmi())
-                raw_data = rcu_dereference(trace_profile_buf_nmi);
+                trace_buf = rcu_dereference(perf_trace_buf_nmi);
         else
-                raw_data = rcu_dereference(trace_profile_buf);
+                trace_buf = rcu_dereference(perf_trace_buf);
 
-        if (!raw_data)
+        if (!trace_buf)
                 goto end;
 
-        raw_data = per_cpu_ptr(raw_data, cpu);
+        trace_buf = per_cpu_ptr(trace_buf, cpu);
+
+        if (trace_buf->recursion++)
+                goto end_recursion;
+
+        /*
+         * Make recursion update visible before entering perf_tp_event
+         * so that we protect from perf recursions.
+         */
+        barrier();
+
+        raw_data = trace_buf->buf;
 
         /* zero the dead bytes from align to not leak stack to user */
         *(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
@@ -527,6 +539,8 @@ static void prof_syscall_enter(struct pt_regs *regs, long id)
                                (unsigned long *)&rec->args);
         perf_tp_event(sys_data->enter_id, 0, 1, rec, size);
 
+end_recursion:
+        trace_buf->recursion--;
 end:
         local_irq_restore(flags);
 }
@@ -574,6 +588,7 @@ static void prof_syscall_exit(struct pt_regs *regs, long ret)
 {
         struct syscall_metadata *sys_data;
         struct syscall_trace_exit *rec;
+        struct perf_trace_buf *trace_buf;
         unsigned long flags;
         int syscall_nr;
         char *raw_data;
@@ -605,14 +620,25 @@ static void prof_syscall_exit(struct pt_regs *regs, long ret)
         cpu = smp_processor_id();
 
         if (in_nmi())
-                raw_data = rcu_dereference(trace_profile_buf_nmi);
+                trace_buf = rcu_dereference(perf_trace_buf_nmi);
         else
-                raw_data = rcu_dereference(trace_profile_buf);
+                trace_buf = rcu_dereference(perf_trace_buf);
 
-        if (!raw_data)
+        if (!trace_buf)
                 goto end;
 
-        raw_data = per_cpu_ptr(raw_data, cpu);
+        trace_buf = per_cpu_ptr(trace_buf, cpu);
+
+        if (trace_buf->recursion++)
+                goto end_recursion;
+
+        /*
+         * Make recursion update visible before entering perf_tp_event
+         * so that we protect from perf recursions.
+         */
+        barrier();
+
+        raw_data = trace_buf->buf;
 
         /* zero the dead bytes from align to not leak stack to user */
         *(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
@@ -626,6 +652,8 @@ static void prof_syscall_exit(struct pt_regs *regs, long ret)
 
         perf_tp_event(sys_data->exit_id, 0, 1, rec, size);
 
+end_recursion:
+        trace_buf->recursion--;
 end:
         local_irq_restore(flags);
 }