Virtualization - Jan 2023 - [PATCH 3/6] ftrace/x86: Warn and ignore graph tracing when RCU is disabled

On Mon, 23 Jan 2023 16:53:04 -0500
Steven Rostedt <rostedt at goodmis.org> wrote:
> On Mon, 23 Jan 2023 21:50:12 +0100
> Peter Zijlstra <peterz at infradead.org> wrote:
> 
> > All RCU disabled code should be noinstr and hence we should never get
> > here -- when we do, WARN about it and make sure to not actually do
> > tracing.
> > 
> > Signed-off-by: Peter Zijlstra (Intel) <peterz at infradead.org>
> > ---
> >  arch/x86/kernel/ftrace.c |    3 +++
> >  1 file changed, 3 insertions(+)
> > 
> > --- a/arch/x86/kernel/ftrace.c
> > +++ b/arch/x86/kernel/ftrace.c
> > @@ -646,6 +646,9 @@ void prepare_ftrace_return(unsigned long
> >  	if (unlikely(atomic_read(&current->tracing_graph_pause)))
> >  		return;
> >  
> > +	if (WARN_ONCE(!rcu_is_watching(), "RCU not on for: %pS\n",
(void *)ip))
> > +		return;
> > +  
> 
> Please add this to after recursion trylock below. Although WARN_ONCE()
> should not not have recursion issues, as function tracing can do weird
> things, I rather be safe than sorry, and not have the system triple boot
> due to some path that might get added in the future.
> 
> If rcu_is_watching() is false, it will still get by the below recursion
> check and warn. That is, the below check should be done before this
> function calls any other function.
> 
> >  	bit = ftrace_test_recursion_trylock(ip, *parent);
> >  	if (bit < 0)
> >  		return;
> >   
> 
Actually, perhaps we can just add this, and all you need to do is create
and set CONFIG_NO_RCU_TRACING (or some other name).

This should cover all ftrace locations. (Uncompiled).

-- Steve

diff --git a/include/linux/trace_recursion.h b/include/linux/trace_recursion.h
index c303f7a114e9..10ee3fbb9113 100644
--- a/include/linux/trace_recursion.h
+++ b/include/linux/trace_recursion.h
@@ -135,6 +135,22 @@ extern void ftrace_record_recursion(unsigned long ip,
unsigned long parent_ip);
 # define do_ftrace_record_recursion(ip, pip)	do { } while (0)
 #endif
 
+#ifdef CONFIG_NO_RCU_TRACING
+# define trace_warn_on_no_rcu(ip)					\
+	({								\
+		bool __ret = false;					\
+		if (!trace_recursion_test(TRACE_RECORD_RECURSION_BIT)) { \
+			trace_recursion_set(TRACE_RECORD_RECURSION_BIT); \
+			__ret = WARN_ONCE(!rcu_is_watching(),		\
+					  "RCU not on for: %pS\n", (void *)ip); \
+			trace_recursion_clear(TRACE_RECORD_RECURSION_BIT); \
+		}							\
+		__ret;							\
+	})
+#else
+# define trace_warn_on_no_rcu(ip)	false
+#endif
+
 /*
  * Preemption is promised to be disabled when return bit >= 0.
  */
@@ -144,6 +160,9 @@ static __always_inline int
trace_test_and_set_recursion(unsigned long ip, unsign
 	unsigned int val = READ_ONCE(current->trace_recursion);
 	int bit;
 
+	if (trace_warn_on_no_rcu(ip))
+		return -1;
+
 	bit = trace_get_context_bit() + start;
 	if (unlikely(val & (1 << bit))) {
 		/*

On Mon, Jan 23, 2023 at 05:07:53PM -0500, Steven Rostedt wrote:
> Actually, perhaps we can just add this, and all you need to do is create
> and set CONFIG_NO_RCU_TRACING (or some other name).
Elsewhere I've used CONFIG_ARCH_WANTS_NO_INSTR for this.

Anyway, I took it for a spin and it .... doesn't seems to do the job.

With my patch the first splat is

  "RCU not on for: cpuidle_poll_time+0x0/0x70"

While with yours I seems to get the endless:

  "WARNING: suspicious RCU usage"

thing. Let me see if I can figure out where it goes side-ways.