On Tue, May 17, 2022 at 12:13:45PM +0200, Jiri Olsa wrote:

On Mon, May 16, 2022 at 01:49:22PM +0200, Frederic Weisbecker wrote:

quoted

On Sun, May 15, 2022 at 09:25:35PM -0700, Paul E. McKenney wrote:

quoted

On Sun, May 15, 2022 at 10:36:52PM +0200, Jiri Olsa wrote:

quoted

Making arch_cpu_idle and rcu_idle_exit noinstr. Both functions run
in rcu 'not watching' context and if there's tracer attached to
them, which uses rcu (e.g. kprobe multi interface) it will hit RCU
warning like:

  [    3.017540] WARNING: suspicious RCU usage
  ...
  [    3.018363]  kprobe_multi_link_handler+0x68/0x1c0
  [    3.018364]  ? kprobe_multi_link_handler+0x3e/0x1c0
  [    3.018366]  ? arch_cpu_idle_dead+0x10/0x10
  [    3.018367]  ? arch_cpu_idle_dead+0x10/0x10
  [    3.018371]  fprobe_handler.part.0+0xab/0x150
  [    3.018374]  0xffffffffa00080c8
  [    3.018393]  ? arch_cpu_idle+0x5/0x10
  [    3.018398]  arch_cpu_idle+0x5/0x10
  [    3.018399]  default_idle_call+0x59/0x90
  [    3.018401]  do_idle+0x1c3/0x1d0

The call path is following:

default_idle_call
  rcu_idle_enter
  arch_cpu_idle
  rcu_idle_exit

The arch_cpu_idle and rcu_idle_exit are the only ones from above
path that are traceble and cause this problem on my setup.

Signed-off-by: Jiri Olsa <jolsa@kernel.org>

From an RCU viewpoint:

Reviewed-by: Paul E. McKenney <paulmck@kernel.org>

[ I considered asking for an instrumentation_on() in rcu_idle_exit(),
but there is no point given that local_irq_restore() isn't something
you instrument anyway. ]

So local_irq_save() in the beginning of rcu_idle_exit() is unsafe because
it is instrumentable by the function (graph)  tracers and the irqsoff tracer.

Also it calls into lockdep that might make use of RCU.

That's why rcu_idle_exit() is not noinstr yet. See this patch:

https://lore.kernel.org/lkml/20220503100051.2799723-4-frederic@kernel.org/ (local)

I see, could we mark it at least with notrace meanwhile?

For the RCU part, how about as follows?

If this approach is reasonable, my guess would be that Frederic will pull
it into his context-tracking series, perhaps using a revert of this patch
to maintain sanity in the near term.

If this approach is unreasonable, well, that is Murphy for you!

For the x86 idle part, my feeling is still that the rcu_idle_enter()
and rcu_idle_exit() need to be pushed deeper into the code.  Perhaps
an ongoing process as the idle loop continues to be dug deeper?

							Thanx, Paul

------------------------------------------------------------------------

commit cd338be719a0a692e0d50e1a8438e1f6c7165d9c
Author: Paul E. McKenney [off-list ref]
Date:   Tue May 17 21:00:04 2022 -0700

    rcu: Apply noinstr to rcu_idle_enter() and rcu_idle_exit()
    
    This commit applies the "noinstr" tag to the rcu_idle_enter() and
    rcu_idle_exit() functions, which are invoked from portions of the idle
    loop that cannot be instrumented.  These tags require reworking the
    rcu_eqs_enter() and rcu_eqs_exit() functions that these two functions
    invoke in order to cause them to use normal assertions rather than
    lockdep.  In addition, within rcu_idle_exit(), the raw versions of
    local_irq_save() and local_irq_restore() are used, again to avoid issues
    with lockdep in uninstrumented code.
    
    This patch is based in part on an earlier patch by Jiri Olsa, discussions
    with Peter Zijlstra and Frederic Weisbecker, earlier changes by Thomas
    Gleixner, and off-list discussions with Yonghong Song.
    
    Link: https://lore.kernel.org/lkml/20220515203653.4039075-1-jolsa@kernel.org/ (local)
    Reported-by: Jiri Olsa [off-list ref]
    Reported-by: Alexei Starovoitov [off-list ref]
    Reported-by: Andrii Nakryiko [off-list ref]
    Signed-off-by: Paul E. McKenney [off-list ref]
    Reviewed-by: Yonghong Song [off-list ref]

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index 222d59299a2af..02233b17cce0e 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c

@@ -635,8 +635,8 @@ static noinstr void rcu_eqs_enter(bool user)
 		return;
 	}
 
-	lockdep_assert_irqs_disabled();
 	instrumentation_begin();
+	lockdep_assert_irqs_disabled();
 	trace_rcu_dyntick(TPS("Start"), rdp->dynticks_nesting, 0, atomic_read(&rdp->dynticks));
 	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
 	rcu_preempt_deferred_qs(current);

@@ -663,9 +663,9 @@ static noinstr void rcu_eqs_enter(bool user)
  * If you add or remove a call to rcu_idle_enter(), be sure to test with
  * CONFIG_RCU_EQS_DEBUG=y.
  */
-void rcu_idle_enter(void)
+void noinstr rcu_idle_enter(void)
 {
-	lockdep_assert_irqs_disabled();
+	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !raw_irqs_disabled());
 	rcu_eqs_enter(false);
 }
 EXPORT_SYMBOL_GPL(rcu_idle_enter);

@@ -865,7 +865,7 @@ static void noinstr rcu_eqs_exit(bool user)
 	struct rcu_data *rdp;
 	long oldval;
 
-	lockdep_assert_irqs_disabled();
+	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !raw_irqs_disabled());
 	rdp = this_cpu_ptr(&rcu_data);
 	oldval = rdp->dynticks_nesting;
 	WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);

@@ -900,13 +900,13 @@ static void noinstr rcu_eqs_exit(bool user)
  * If you add or remove a call to rcu_idle_exit(), be sure to test with
  * CONFIG_RCU_EQS_DEBUG=y.
  */
-void rcu_idle_exit(void)
+void noinstr rcu_idle_exit(void)
 {
 	unsigned long flags;
 
-	local_irq_save(flags);
+	raw_local_irq_save(flags);
 	rcu_eqs_exit(false);
-	local_irq_restore(flags);
+	raw_local_irq_restore(flags);
 }
 EXPORT_SYMBOL_GPL(rcu_idle_exit);