Merge branch 'srcu-next.2025.08.21a' into HEAD

SRCU updates:
* Create srcu_read_{,un}lock_fast_notrace()
* Add srcu_read_lock_fast_notrace() and srcu_read_unlock_fast_notrace()
* Add guards for notrace variants of SRCU-fast readers
* Document srcu_read_{,un}lock_fast() use of implicit RCU readers
* Document srcu_flip() memory-barrier D relation to SRCU-fast
* Remove preempt_disable/enable() in Tiny SRCU srcu_gp_start_if_needed()

2 files changed, 11 insertions, 3 deletions

diff --git a/kernel/rcu/srcutiny.c b/kernel/rcu/srcutiny.c
index 6e9fe2ce1075..e3b64a5e0ec7 100644
--- a/kernel/rcu/srcutiny.c
+++ b/kernel/rcu/srcutiny.c
@@ -176,10 +176,9 @@ static void srcu_gp_start_if_needed(struct srcu_struct *ssp)
 {
 	unsigned long cookie;
 
-	preempt_disable();  // Needed for PREEMPT_LAZY
+	lockdep_assert_preemption_disabled(); // Needed for PREEMPT_LAZY
 	cookie = get_state_synchronize_srcu(ssp);
 	if (ULONG_CMP_GE(READ_ONCE(ssp->srcu_idx_max), cookie)) {
-		preempt_enable();
 		return;
 	}
 	WRITE_ONCE(ssp->srcu_idx_max, cookie);
@@ -189,7 +188,6 @@ static void srcu_gp_start_if_needed(struct srcu_struct *ssp)
 		else if (list_empty(&ssp->srcu_work.entry))
 			list_add(&ssp->srcu_work.entry, &srcu_boot_list);
 	}
-	preempt_enable();
 }
 
 /*
diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index c5e8ebc493d5..1ff94b76d91f 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -1168,6 +1168,16 @@ static void srcu_flip(struct srcu_struct *ssp)
 	 * counter update.  Note that both this memory barrier and the
 	 * one in srcu_readers_active_idx_check() provide the guarantee
 	 * for __srcu_read_lock().
+	 *
+	 * Note that this is a performance optimization, in which we spend
+	 * an otherwise unnecessary smp_mb() in order to reduce the number
+	 * of full per-CPU-variable scans in srcu_readers_lock_idx() and
+	 * srcu_readers_unlock_idx().  But this performance optimization
+	 * is not so optimal for SRCU-fast, where we would be spending
+	 * not smp_mb(), but rather synchronize_rcu().  At the same time,
+	 * the overhead of the smp_mb() is in the noise, so there is no
+	 * point in omitting it in the SRCU-fast case.  So the same code
+	 * is executed either way.
 	 */
 	smp_mb(); /* D */  /* Pairs with C. */
 }