Message-ID: <773fd8246a1ec4ef79142d9e31b8ba4163a0d496.camel@gmx.de>
Subject: [rcf/patch] netpoll: Make it RT friendly
From:   Mike Galbraith <efault@gmx.de>
To:     Sebastian Andrzej Siewior <bigeasy@linutronix.de>,
        Thomas Gleixner <tglx@linutronix.de>
Cc:     LKML <linux-kernel@vger.kernel.org>,
        RT <linux-rt-users@vger.kernel.org>
Date:   Fri, 19 Nov 2021 15:41:25 +0100
In-Reply-To: <20211118163405.adczki7ibk22bw7e@linutronix.de>
References: <20211118163405.adczki7ibk22bw7e@linutronix.de>
Content-Type: text/plain; charset="UTF-8"
User-Agent: Evolution 3.42.1 
MIME-Version: 1.0
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Thu, 2021-11-18 at 17:34 +0100, Sebastian Andrzej Siewior wrote:
> Dear RT folks!
>
> I'm pleased to announce the v5.16-rc1-rt2 patch set.
>
> Changes since v5.16-rc1-rt1:
>
> =C2=A0 - Redo the delayed deallocation of the task-stack.
>
> Known issues
> =C2=A0=C2=A0=C2=A0=C2=A0 - netconsole triggers WARN.

The below seems to do the trick for netconsole in master.today-rt.

netpoll: Make it RT friendly

PREEMPT_RT cannot alloc/free memory when not preemptible, making
zap_completion_queue() disabling preemption across kfree() an
issue, as well as disabling IRQs asross __netpoll_send_skb(),
which leads to zap_completion_queue() and our pesky kfree().

We can let rcu_read_lock_bh() provide local exclusion for RT across
__netpoll_send_skb() (via softirq_ctrl.lock) instead of disabling
IRQs, and since zap_completion_queue() replaces sd->completion_queue
with IRQs disabled and makes a private copy, there's no need to
keep preemption disabled during traversal/freeing, so put_cpu_var()
before doing so.  Disable a couple warnings for RT, and we're done.

Signed-off-by: Mike Galbraith <efault@gmx.de>
=2D--
 net/core/netpoll.c |   23 +++++++++++++++--------
 1 file changed, 15 insertions(+), 8 deletions(-)

=2D-- a/net/core/netpoll.c
+++ b/net/core/netpoll.c
@@ -252,6 +252,7 @@ static void zap_completion_queue(void)
 		clist =3D sd->completion_queue;
 		sd->completion_queue =3D NULL;
 		local_irq_restore(flags);
+		put_cpu_var(softnet_data);

 		while (clist !=3D NULL) {
 			struct sk_buff *skb =3D clist;
@@ -263,9 +264,8 @@ static void zap_completion_queue(void)
 				__kfree_skb(skb);
 			}
 		}
-	}
-
-	put_cpu_var(softnet_data);
+	} else
+		put_cpu_var(softnet_data);
 }

 static struct sk_buff *find_skb(struct netpoll *np, int len, int reserve)
@@ -314,7 +314,8 @@ static netdev_tx_t __netpoll_send_skb(st
 	/* It is up to the caller to keep npinfo alive. */
 	struct netpoll_info *npinfo;

-	lockdep_assert_irqs_disabled();
+	if (!IS_ENABLED(CONFIG_PREEMPT_RT))
+		lockdep_assert_irqs_disabled();

 	dev =3D np->dev;
 	npinfo =3D rcu_dereference_bh(dev->npinfo);
@@ -350,7 +351,7 @@ static netdev_tx_t __netpoll_send_skb(st
 			udelay(USEC_PER_POLL);
 		}

-		WARN_ONCE(!irqs_disabled(),
+		WARN_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RT) && !irqs_disabled(),
 			"netpoll_send_skb_on_dev(): %s enabled interrupts in poll (%pS)\n",
 			dev->name, dev->netdev_ops->ndo_start_xmit);

@@ -365,16 +366,22 @@ static netdev_tx_t __netpoll_send_skb(st

 netdev_tx_t netpoll_send_skb(struct netpoll *np, struct sk_buff *skb)
 {
-	unsigned long flags;
+	unsigned long __maybe_unused flags;
 	netdev_tx_t ret;

 	if (unlikely(!np)) {
 		dev_kfree_skb_irq(skb);
 		ret =3D NET_XMIT_DROP;
 	} else {
-		local_irq_save(flags);
+		if (!IS_ENABLED(CONFIG_PREEMPT_RT))
+			local_irq_save(flags);
+		else
+			rcu_read_lock_bh();
 		ret =3D __netpoll_send_skb(np, skb);
-		local_irq_restore(flags);
+		if (!IS_ENABLED(CONFIG_PREEMPT_RT))
+			local_irq_restore(flags);
+		else
+			rcu_read_unlock_bh();
 	}
 	return ret;
 }