kvfree_rcu: Add a bulk-list check when a scheduler is run
authorUladzislau Rezki (Sony) <urezki@gmail.com>
Thu, 15 Apr 2021 17:19:58 +0000 (19:19 +0200)
committerPaul E. McKenney <paulmck@kernel.org>
Mon, 10 May 2021 23:00:48 +0000 (16:00 -0700)
The rcu_scheduler_active flag is set to RCU_SCHEDULER_RUNNING once the
scheduler is up and running.  That signal is used in order to check
and queue a "monitor work" to reclaim freed objects (if there are any)
during early boot.  This flag is used by kvfree_rcu() to determine when
work can safely be queued, at which point memory passed to earlier
invocations of kvfree_rcu() can be processed.

However, only "krcp->head" is checked for objects that need to be
released, and there are now two more, namely, "krcp->bkvhead[0]" and
"krcp->bkvhead[1]".  Therefore, check these two additional channels.

Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com>
Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
kernel/rcu/tree.c

index 676a49a..e86f32d 100644 (file)
@@ -3739,7 +3739,8 @@ void __init kfree_rcu_scheduler_running(void)
                struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
 
                raw_spin_lock_irqsave(&krcp->lock, flags);
-               if (!krcp->head || krcp->monitor_todo) {
+               if ((!krcp->bkvhead[0] && !krcp->bkvhead[1] && !krcp->head) ||
+                               krcp->monitor_todo) {
                        raw_spin_unlock_irqrestore(&krcp->lock, flags);
                        continue;
                }