Pointless overhead to use a work queue to reset the static key
for a DO_ONCE_SLEEPABLE() invocation.

Reported-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
---
 lib/once.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/lib/once.c b/lib/once.c
index 2c306f0e891e..d83bdde78ced 100644
--- a/lib/once.c
+++ b/lib/once.c
@@ -93,6 +93,7 @@ void __do_once_sleepable_done(bool *done, struct static_key_true *once_key,
 {
 	*done = true;
 	mutex_unlock(&once_mutex);
-	once_disable_jump(once_key, mod);
+	BUG_ON(!static_key_enabled(once_key));
+	static_branch_disable(once_key);
 }
 EXPORT_SYMBOL(__do_once_sleepable_done);
-- 
2.52.0

On Thu,  8 Jan 2026 15:27:17 -0800 Tony Luck <tony.luck@intel.com> wrote:

> Pointless overhead to use a work queue to reset the static key
> for a DO_ONCE_SLEEPABLE() invocation.
> 
> ...
>
> --- a/lib/once.c
> +++ b/lib/once.c
> @@ -93,6 +93,7 @@ void __do_once_sleepable_done(bool *done, struct static_key_true *once_key,
>  {
>  	*done = true;
>  	mutex_unlock(&once_mutex);
> -	once_disable_jump(once_key, mod);
> +	BUG_ON(!static_key_enabled(once_key));

Sauron's eye is upon us.  Can we either justify nuking the kernel or
use a WARN_ON?

> +	static_branch_disable(once_key);
>  }
>  EXPORT_SYMBOL(__do_once_sleepable_done);
> -- 
> 2.52.0

Hi Tony,

On 1/8/26 3:27 PM, Tony Luck wrote:
> Pointless overhead to use a work queue to reset the static key
> for a DO_ONCE_SLEEPABLE() invocation.
> 
> Reported-by: Reinette Chatre <reinette.chatre@intel.com>
> Signed-off-by: Tony Luck <tony.luck@intel.com>
> ---
>  lib/once.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/lib/once.c b/lib/once.c
> index 2c306f0e891e..d83bdde78ced 100644
> --- a/lib/once.c
> +++ b/lib/once.c
> @@ -93,6 +93,7 @@ void __do_once_sleepable_done(bool *done, struct static_key_true *once_key,
>  {
>  	*done = true;
>  	mutex_unlock(&once_mutex);
> -	once_disable_jump(once_key, mod);
> +	BUG_ON(!static_key_enabled(once_key));
> +	static_branch_disable(once_key);

Can this be simplified more by disabling once_key with once_mutex held to eliminate the need for the
done boolean? Unless static_branch_disable() is considered to be so slow that it should also be
avoided in sleepable context, deferring to a mutex protected boolean until it completes ...
If "done" boolean is removed this may need an additional check of once_key within
__do_once_sleepable_start() just in case a second caller was blocked on once_mutex.

Reinette

Pointless overhead to use a work queue to reset the static key
for a DO_ONCE_SLEEPABLE() invocation.

Note that the previous code path included a BUG_ON() if the static key
was already disabled. Dropped that as part of this change because:
1) Use of BUG_ON() is highly discouraged.
2) There is a WARN_ON() in the static_branch_disable() code path
   that would provide adequate breadcrumbs to debug any issue.

Reported-by: Reinette Chatre <reinette.chatre@intel.com>
Signed-off-by: Tony Luck <tony.luck@intel.com>
---
Changes since v1:
Andrew Morton pointed out that adding a new BUG_ON() was ill-advised.
Drop it (but explain why it isn't needed in the commit message).

 lib/once.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/lib/once.c b/lib/once.c
index 2c306f0e891e..8557eb489f34 100644
--- a/lib/once.c
+++ b/lib/once.c
@@ -93,6 +93,6 @@ void __do_once_sleepable_done(bool *done, struct static_key_true *once_key,
 {
 	*done = true;
 	mutex_unlock(&once_mutex);
-	once_disable_jump(once_key, mod);
+	static_branch_disable(once_key);
 }
 EXPORT_SYMBOL(__do_once_sleepable_done);
-- 
2.52.0