Re: [for-next][PATCH 04/15] tracepoint: Add lockdep rcu_is_watching() check to trace_##name##_enabled()

[Steven Rostedt]

On Tue, 30 Jun 2026 19:39:02 +0200
Geert Uytterhoeven <geert@xxxxxxxxxxxxxx> wrote:

> Other Renesas ARM32 platforms I tried (R-Mobile A1, RZ/A1H, RZ/A2M)
> are unafffected, perhaps because they are not SMP?
> All Renesas ARM64 platforms I tried (R-Car Gen3/4) are also unaffected.
> 
> Reverting the commit fixes the issue.
> 
> Do you have a clue?

Yes, it means the code was buggy before the commit. The commit will trigger
warnings in places that have issues. Before the commit, the buggy code was
never caught.

It's like enabling KASAN and finding code that has use-after-free.
Disabling KASAN is not the fix.

Tracepoints are managed by using RCU. There's places that RCU is turned
off, meaning a tracepoint in one of those locations can be triggered when
RCU is not active which may have a use-after-free semantic when the
tracepoint is enabled.

Tracepoints hidden by trace_#tracepoint#_enabled() are not caught when RCU
is disabled and the tracepoint is not active. This commit makes these
locations trigger even when the tracepoint is not active.

One way to find out if this is an existing bug or not, could you enable the
preemptirq tracepoints and run the tests again with the commit reverted?

  echo 1 >  /sys/kernel/tracing/events/preemptirq/enable

This will enable the events that are hidden without the commit. If it
triggers when enabled, it shows the commit found a bug.

If you get the same errors, the bug isn't with the commit in question, it's
with the tracepoints being called during suspend/resume. We will need to
fix that if that's the case.

-- Steve