Re: One-off rcu_nocb_rdp_deoffload bug

[Frederic Weisbecker <frederic@kernel.org>]

Le Thu, Sep 05, 2024 at 08:32:16PM +0200, Frederic Weisbecker a écrit :
> Le Wed, Sep 04, 2024 at 06:52:36AM -0700, Paul E. McKenney a écrit :
> > > Yes, I'm preparing an update for the offending patch (which has one more
> > > embarassing issue while I'm going through it again).
> > 
> > Very good, thank you!
> 
> So my proposal for a replacement patch is this (to replace the patch
> of the same name in Neeraj tree):

FYI, the diffstat against the previous version of the same patch is as follows.
The rationale being:

1) rdp->nocb_cb_kthread doesn't need to be protected by nocb_gp_kthread_mutex

2) Once rcuoc is parked, we really _must_ observe the callback list counter decremented
   after the barrier's completion.

3) This fixes another issue: rcuoc must be parked _before_
   rcu_nocb_queue_toggle_rdp() is called, otherwise a nocb locked sequence
   within rcuoc would race with rcuog clearing SEGCBLIST_OFFLOADED concurrently,
   leaving the nocb locked forever.


diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h
index 755ada098035..97b99cd06923 100644
--- a/kernel/rcu/tree_nocb.h
+++ b/kernel/rcu/tree_nocb.h
@@ -1056,6 +1056,13 @@ static int rcu_nocb_rdp_deoffload(struct rcu_data *rdp)
 	/* Flush all callbacks from segcblist and bypass */
 	rcu_barrier();
 
+	/*
+	 * Make sure the rcuoc kthread isn't in the middle of a nocb locked
+	 * sequence while offloading is deactivated, along with nocb locking.
+	 */
+	if (rdp->nocb_cb_kthread)
+		kthread_park(rdp->nocb_cb_kthread);
+
 	rcu_nocb_lock_irqsave(rdp, flags);
 	WARN_ON_ONCE(rcu_cblist_n_cbs(&rdp->nocb_bypass));
 	WARN_ON_ONCE(rcu_segcblist_n_cbs(&rdp->cblist));
@@ -1064,13 +1071,11 @@ static int rcu_nocb_rdp_deoffload(struct rcu_data *rdp)
 	wake_gp = rcu_nocb_queue_toggle_rdp(rdp);
 
 	mutex_lock(&rdp_gp->nocb_gp_kthread_mutex);
+
 	if (rdp_gp->nocb_gp_kthread) {
 		if (wake_gp)
 			wake_up_process(rdp_gp->nocb_gp_kthread);
 
-		if (rdp->nocb_cb_kthread)
-			kthread_park(rdp->nocb_cb_kthread);
-
 		swait_event_exclusive(rdp->nocb_state_wq,
 				      rcu_nocb_rdp_deoffload_wait_cond(rdp));
 	} else {