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

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?

jirka