rcu: Reset stall detection after ftrace dump

[PATCH] rcu: Reset stall detection after ftrace dump

Posted by Joel Fernandes 2 days ago

During some recent expedited RCU testing with rcutorture, I noticed RCU
stalls when using ftrace_dump() to analyze traces. The issue that since
ftrace_dump() can take a long time, any accumulated issue can show up
just after the stall is unsupressed.  Due to this a stall can be
detected and reported.

The problem is something like this:

  1. rcu_ftrace_dump_stall_suppress() sets suppression value.

  2. ftrace_dump() runs for a long time.

  3. rcu_ftrace_dump_stall_unsuppress() clears suppression.

  4. Immediately later, stall is detected.

Fix this by calling rcu_cpu_stall_reset() when clearing suppression.
This resets jiffies_stall to ULONG_MAX, preventing the issue.

Tested with gp_exp=true enabled with rcutorture and ftrace_dump() and no
stall is reported, with only clean traces in the output.

Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>
---
This is a non-urgent fix and could go in either this or the next merge
window (Boqun's call).

 kernel/rcu/rcu.h        | 5 ++++-
 kernel/rcu/tree_stall.h | 1 +
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/kernel/rcu/rcu.h b/kernel/rcu/rcu.h
index 9cf01832a6c3..87fcf528b9ac 100644
--- a/kernel/rcu/rcu.h
+++ b/kernel/rcu/rcu.h
@@ -288,6 +288,7 @@ extern int rcu_cpu_stall_cputime;
 extern bool rcu_exp_stall_task_details __read_mostly;
 int rcu_jiffies_till_stall_check(void);
 int rcu_exp_jiffies_till_stall_check(void);
+void rcu_cpu_stall_reset(void);
 
 static inline bool rcu_stall_is_suppressed(void)
 {
@@ -302,8 +303,10 @@ do { \
 
 #define rcu_ftrace_dump_stall_unsuppress() \
 do { \
-	if (rcu_cpu_stall_suppress == 3) \
+	if (rcu_cpu_stall_suppress == 3) { \
 		rcu_cpu_stall_suppress = 0; \
+		rcu_cpu_stall_reset(); \
+	} \
 } while (0)
 
 #else /* #endif #ifdef CONFIG_RCU_STALL_COMMON */
diff --git a/kernel/rcu/tree_stall.h b/kernel/rcu/tree_stall.h
index b67532cb8770..9a036fc27fed 100644
--- a/kernel/rcu/tree_stall.h
+++ b/kernel/rcu/tree_stall.h
@@ -192,6 +192,7 @@ void rcu_cpu_stall_reset(void)
 	WRITE_ONCE(rcu_state.nr_fqs_jiffies_stall, 3);
 	WRITE_ONCE(rcu_state.jiffies_stall, ULONG_MAX);
 }
+EXPORT_SYMBOL_GPL(rcu_cpu_stall_reset);
 
 //////////////////////////////////////////////////////////////////////////////
 //
-- 
2.34.1

Re: [PATCH] rcu: Reset stall detection after ftrace dump

Posted by Paul E. McKenney 1 day, 16 hours ago

On Sun, Jan 04, 2026 at 03:36:42PM -0500, Joel Fernandes wrote:
> During some recent expedited RCU testing with rcutorture, I noticed RCU
> stalls when using ftrace_dump() to analyze traces. The issue that since
> ftrace_dump() can take a long time, any accumulated issue can show up
> just after the stall is unsupressed.  Due to this a stall can be
> detected and reported.
> 
> The problem is something like this:
> 
>   1. rcu_ftrace_dump_stall_suppress() sets suppression value.
> 
>   2. ftrace_dump() runs for a long time.
> 
>   3. rcu_ftrace_dump_stall_unsuppress() clears suppression.
> 
>   4. Immediately later, stall is detected.
> 
> Fix this by calling rcu_cpu_stall_reset() when clearing suppression.
> This resets jiffies_stall to ULONG_MAX, preventing the issue.
> 
> Tested with gp_exp=true enabled with rcutorture and ftrace_dump() and no
> stall is reported, with only clean traces in the output.
> 
> Signed-off-by: Joel Fernandes <joelagnelf@nvidia.com>

Good catch!

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

> ---
> This is a non-urgent fix and could go in either this or the next merge
> window (Boqun's call).
> 
>  kernel/rcu/rcu.h        | 5 ++++-
>  kernel/rcu/tree_stall.h | 1 +
>  2 files changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/rcu/rcu.h b/kernel/rcu/rcu.h
> index 9cf01832a6c3..87fcf528b9ac 100644
> --- a/kernel/rcu/rcu.h
> +++ b/kernel/rcu/rcu.h
> @@ -288,6 +288,7 @@ extern int rcu_cpu_stall_cputime;
>  extern bool rcu_exp_stall_task_details __read_mostly;
>  int rcu_jiffies_till_stall_check(void);
>  int rcu_exp_jiffies_till_stall_check(void);
> +void rcu_cpu_stall_reset(void);
>  
>  static inline bool rcu_stall_is_suppressed(void)
>  {
> @@ -302,8 +303,10 @@ do { \
>  
>  #define rcu_ftrace_dump_stall_unsuppress() \
>  do { \
> -	if (rcu_cpu_stall_suppress == 3) \
> +	if (rcu_cpu_stall_suppress == 3) { \
>  		rcu_cpu_stall_suppress = 0; \
> +		rcu_cpu_stall_reset(); \
> +	} \
>  } while (0)
>  
>  #else /* #endif #ifdef CONFIG_RCU_STALL_COMMON */
> diff --git a/kernel/rcu/tree_stall.h b/kernel/rcu/tree_stall.h
> index b67532cb8770..9a036fc27fed 100644
> --- a/kernel/rcu/tree_stall.h
> +++ b/kernel/rcu/tree_stall.h
> @@ -192,6 +192,7 @@ void rcu_cpu_stall_reset(void)
>  	WRITE_ONCE(rcu_state.nr_fqs_jiffies_stall, 3);
>  	WRITE_ONCE(rcu_state.jiffies_stall, ULONG_MAX);
>  }
> +EXPORT_SYMBOL_GPL(rcu_cpu_stall_reset);
>  
>  //////////////////////////////////////////////////////////////////////////////
>  //
> -- 
> 2.34.1
>