Date:   Tue, 14 Mar 2023 14:08:28 -0700
From:   "Paul E. McKenney" <paulmck@kernel.org>
To:     Steven Rostedt <rostedt@goodmis.org>
Cc:     linux-kernel@vger.kernel.org,
        Masami Hiramatsu <mhiramat@kernel.org>,
        Andrew Morton <akpm@linux-foundation.org>,
        stable@vger.kernel.org, Dave Hansen <dave.hansen@linux.intel.com>,
        Mathieu Desnoyers <mathieu.desnoyers@efficios.com>,
        Joel Fernandes <joel@joelfernandes.org>,
        "Peter Zijlstra (Intel)" <peterz@infradead.org>
Subject: Re: [for-linus][PATCH 5/5] tracing: Make tracepoint lockdep check
 actually test something
Message-ID: <f4f52692-9f6c-467a-8988-113aced754eb@paulmck-laptop>
Reply-To: paulmck@kernel.org
References: <20230314190236.203370742@goodmis.org>
 <20230314190310.486609095@goodmis.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20230314190310.486609095@goodmis.org>
Precedence: bulk

On Tue, Mar 14, 2023 at 03:02:41PM -0400, Steven Rostedt wrote:
> From: "Steven Rostedt (Google)" <rostedt@goodmis.org>
> 
> A while ago where the trace events had the following:
> 
>    rcu_read_lock_sched_notrace();
>    rcu_dereference_sched(...);
>    rcu_read_unlock_sched_notrace();
> 
> If the tracepoint is enabled, it could trigger RCU issues if called in
> the wrong place. And this warning was only triggered if lockdep was
> enabled. If the tracepoint was never enabled with lockdep, the bug would
> not be caught. To handle this, the above sequence was done when lockdep
> was enabled regardless if the tracepoint was enabled or not (although the
> always enabled code really didn't do anything, it would still trigger a
> warning).
> 
> But a lot has changed since that lockdep code was added. One is, that
> sequence no longer triggers any warning. Another is, the tracepoint when
> enabled doesn't even do that sequence anymore.
> 
> The main check we care about today is whether RCU is "watching" or not.
> So if lockdep is enabled, always check if rcu_is_watching() which will
> trigger a warning if it is not (tracepoints require RCU to be watching).
> 
> Note, that old sequence did add a bit of overhead when lockdep was enabled,
> and with the latest kernel updates, would cause the system to slow down
> enough to trigger kernel "stalled" warnings.
> 
> Link: http://lore.kernel.org/lkml/20140806181801.GA4605@redhat.com
> Link: http://lore.kernel.org/lkml/20140807175204.C257CAC5@viggo.jf.intel.com
> Link: https://lore.kernel.org/lkml/20230307184645.521db5c9@gandalf.local.home/
> Link: https://lore.kernel.org/linux-trace-kernel/20230310172856.77406446@gandalf.local.home
> 
> Cc: stable@vger.kernel.org
> Cc: Masami Hiramatsu <mhiramat@kernel.org>
> Cc: Dave Hansen <dave.hansen@linux.intel.com>
> Cc: "Paul E. McKenney" <paulmck@kernel.org>
> Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
> Cc: Joel Fernandes <joel@joelfernandes.org>
> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
> Fixes: e6753f23d961 ("tracepoint: Make rcuidle tracepoint callers use SRCU")
> Signed-off-by: Steven Rostedt (Google) <rostedt@goodmis.org>

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

> ---
>  include/linux/tracepoint.h | 15 ++++++---------
>  1 file changed, 6 insertions(+), 9 deletions(-)
> 
> diff --git a/include/linux/tracepoint.h b/include/linux/tracepoint.h
> index fa1004fcf810..2083f2d2f05b 100644
> --- a/include/linux/tracepoint.h
> +++ b/include/linux/tracepoint.h
> @@ -231,12 +231,11 @@ static inline struct tracepoint *tracepoint_ptr_deref(tracepoint_ptr_t *p)
>   * not add unwanted padding between the beginning of the section and the
>   * structure. Force alignment to the same alignment as the section start.
>   *
> - * When lockdep is enabled, we make sure to always do the RCU portions of
> - * the tracepoint code, regardless of whether tracing is on. However,
> - * don't check if the condition is false, due to interaction with idle
> - * instrumentation. This lets us find RCU issues triggered with tracepoints
> - * even when this tracepoint is off. This code has no purpose other than
> - * poking RCU a bit.
> + * When lockdep is enabled, we make sure to always test if RCU is
> + * "watching" regardless if the tracepoint is enabled or not. Tracepoints
> + * require RCU to be active, and it should always warn at the tracepoint
> + * site if it is not watching, as it will need to be active when the
> + * tracepoint is enabled.
>   */
>  #define __DECLARE_TRACE(name, proto, args, cond, data_proto)		\
>  	extern int __traceiter_##name(data_proto);			\
> @@ -249,9 +248,7 @@ static inline struct tracepoint *tracepoint_ptr_deref(tracepoint_ptr_t *p)
>  				TP_ARGS(args),				\
>  				TP_CONDITION(cond), 0);			\
>  		if (IS_ENABLED(CONFIG_LOCKDEP) && (cond)) {		\
> -			rcu_read_lock_sched_notrace();			\
> -			rcu_dereference_sched(__tracepoint_##name.funcs);\
> -			rcu_read_unlock_sched_notrace();		\
> +			WARN_ON_ONCE(!rcu_is_watching());		\
>  		}							\
>  	}								\
>  	__DECLARE_TRACE_RCU(name, PARAMS(proto), PARAMS(args),		\
> -- 
> 2.39.1