Re: [PATCH] random32: Use rcuidle variant for tracepoint

[Marco Elver]

On Fri, 21 Aug 2020 at 08:30, Marco Elver <elver@xxxxxxxxxx> wrote:
> With KCSAN enabled, prandom_u32() may be called from any context,
> including idle CPUs.
>
> Therefore, switch to using trace_prandom_u32_rcuidle(), to avoid various
> issues due to recursion and lockdep warnings when KCSAN and tracing is
> enabled.
>
> Fixes: 94c7eb54c4b8 ("random32: add a tracepoint for prandom_u32()")
> Link: https://lkml.kernel.org/r/20200820155923.3d5c4873@xxxxxxxxxxxxxxxx
> Suggested-by: Steven Rostedt <rostedt@xxxxxxxxxxx>
> Signed-off-by: Marco Elver <elver@xxxxxxxxxx>
> Cc: Eric Dumazet <edumazet@xxxxxxxxxx>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> ---
>  lib/random32.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/lib/random32.c b/lib/random32.c
> index 932345323af0..1c5607a411d4 100644
> --- a/lib/random32.c
> +++ b/lib/random32.c
> @@ -83,7 +83,7 @@ u32 prandom_u32(void)
>         u32 res;
>
>         res = prandom_u32_state(state);
> -       trace_prandom_u32(res);
> +       trace_prandom_u32_rcuidle(res);
>         put_cpu_var(net_rand_state);
>

Note that, for KCSAN's use, there is still a small chance this will
mess things up. So, as a short-term fix, I'd appreciate if this patch
will be applied to 5.9, assuming it doesn't mess with the original
usecase.

Longer-term, I think I need to switch KCSAN to use either
prandom_u32_state() or open-code its own prandom function, as the
trace-rcuidle code calls into srcu, which might have other weird
effects if KCSAN instruments that. Still working on the longer-term
fix.

Thanks,
-- Marco