Skip to contentSkip to conversation

net: sched: calls synchronize_net() only when needed

PATCH Applied Eric Dumazet 4 participants 6 messages Jan 9, 2025 – Jan 14, 2025

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 net: sched: calls synchronize_net() only when needed

dev_deactivate_many() role is to remove the qdiscs of a network device.

When/if a qdisc is dismantled, an rcu grace period is needed to make sure all outstanding qdisc enqueue are done before we proceed with a qdisc reset.

Most virtual devices do not have a qdisc (if we exclude noqueue ones).

We can call the expensive synchronize_net() only if needed.

Note that dev_deactivate_many() does not have to deal with qdisc-less dev_queue_xmit, as an old comment was claiming.

---
 net/sched/sch_generic.c | 22 +++++++++++-----------
 1 file changed, 11 insertions(+), 11 deletions(-)
net/sched/sch_generic.c +11 -11
index bb7dd351bd651d4aa16945816ee26df3c4a48645..14ab2f4c190a1e201dd1788b413a06e799a829f2 100644
@@ -1277,15 +1277,17 @@ static void qdisc_deactivate(struct Qdisc *qdisc)
 
 static void dev_deactivate_queue(struct net_device *dev,
 				 struct netdev_queue *dev_queue,
-				 void *_qdisc_default)
+				 void *_sync_needed)
 {
-	struct Qdisc *qdisc_default = _qdisc_default;
+	bool *sync_needed = _sync_needed;
 	struct Qdisc *qdisc;
 
 	qdisc = rtnl_dereference(dev_queue->qdisc);
 	if (qdisc) {
+		if (qdisc->enqueue)
+			*sync_needed = true;
 		qdisc_deactivate(qdisc);
-		rcu_assign_pointer(dev_queue->qdisc, qdisc_default);
+		rcu_assign_pointer(dev_queue->qdisc, &noop_qdisc);
 	}
 }
 
@@ -1352,24 +1354,22 @@ static bool some_qdisc_is_busy(struct net_device *dev)
  */
 void dev_deactivate_many(struct list_head *head)
 {
+	bool sync_needed = false;
 	struct net_device *dev;
 
 	list_for_each_entry(dev, head, close_list) {
 		netdev_for_each_tx_queue(dev, dev_deactivate_queue,
-					 &noop_qdisc);
+					 &sync_needed);
 		if (dev_ingress_queue(dev))
 			dev_deactivate_queue(dev, dev_ingress_queue(dev),
-					     &noop_qdisc);
+					     &sync_needed);
 
 		netdev_watchdog_down(dev);
 	}
 
-	/* Wait for outstanding qdisc-less dev_queue_xmit calls or
-	 * outstanding qdisc enqueuing calls.
-	 * This is avoided if all devices are in dismantle phase :
-	 * Caller will call synchronize_net() for us
-	 */
-	synchronize_net();
+	/* Wait for outstanding qdisc enqueuing calls. */
+	if (sync_needed)
+		synchronize_net();
 
 	list_for_each_entry(dev, head, close_list) {
 		netdev_for_each_tx_queue(dev, dev_reset_queue, NULL);
Signature · 1 line
2.47.1.613.gc27f4b7a9f-goog
#1 lore raw
Quoted 9 lines · Eric Dumazet

dev_deactivate_many() role is to remove the qdiscs of a network device.

When/if a qdisc is dismantled, an rcu grace period is needed to make sure all outstanding qdisc enqueue are done before we proceed with a qdisc reset.

Most virtual devices do not have a qdisc (if we exclude noqueue ones).

Such as? To me, most virtual devices use noqueue:

$ git grep IFF_NO_QUEUE -- drivers/net/
drivers/net/amt.c:      dev->priv_flags         |= IFF_NO_QUEUE;
drivers/net/bareudp.c:  dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/bonding/bond_main.c:        bond_dev->priv_flags |= IFF_BONDING | IFF_UNICAST_FLT | IFF_NO_QUEUE;
drivers/net/caif/caif_serial.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/dummy.c:    dev->priv_flags |= IFF_LIVE_ADDR_CHANGE | IFF_NO_QUEUE;
drivers/net/ethernet/chelsio/cxgb4/cxgb4_main.c:        dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/ethernet/netronome/nfp/nfp_net_repr.c:      netdev->priv_flags |= IFF_NO_QUEUE | IFF_DISABLE_NETPOLL;
drivers/net/geneve.c:   dev->priv_flags |= IFF_LIVE_ADDR_CHANGE | IFF_NO_QUEUE;
drivers/net/gtp.c:      dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/ipvlan/ipvlan_main.c:       dev->priv_flags |= IFF_UNICAST_FLT | IFF_NO_QUEUE;
drivers/net/ipvlan/ipvtap.c:    dev->priv_flags &= ~IFF_NO_QUEUE;
drivers/net/loopback.c: dev->priv_flags         |= IFF_LIVE_ADDR_CHANGE | IFF_NO_QUEUE;
drivers/net/macsec.c:   dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/macvlan.c:  dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/net_failover.c:     failover_dev->priv_flags |= IFF_UNICAST_FLT | IFF_NO_QUEUE;
drivers/net/netdevsim/netdev.c:                    IFF_NO_QUEUE;
drivers/net/netkit.c:   dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/nlmon.c:    dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/pfcp.c:     dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/team/team_core.c:   dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/veth.c:     dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/vrf.c:      dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/vsockmon.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/vxlan/vxlan_core.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/wan/hdlc_fr.c:      dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/wireguard/device.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/wireless/virtual/mac80211_hwsim.c:  dev->priv_flags |= IFF_NO_QUEUE;

And noqueue_qdisc_ops sets ->enqueue to noop_enqueue():

struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {

        .id             =       "noqueue",
        .priv_size      =       0,
        .init           =       noqueue_init,
        .enqueue        =       noop_enqueue,
        .dequeue        =       noop_dequeue,
        .peek           =       noop_dequeue,
        .owner          =       THIS_MODULE,

};

Thanks.

#2 lore raw
Quoted 47 lines · Cong Wang
Quoted 9 lines · Eric Dumazet

dev_deactivate_many() role is to remove the qdiscs of a network device.

When/if a qdisc is dismantled, an rcu grace period is needed to make sure all outstanding qdisc enqueue are done before we proceed with a qdisc reset.

Most virtual devices do not have a qdisc (if we exclude noqueue ones).

Such as? To me, most virtual devices use noqueue:

$ git grep IFF_NO_QUEUE -- drivers/net/
drivers/net/amt.c:      dev->priv_flags         |= IFF_NO_QUEUE;
drivers/net/bareudp.c:  dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/bonding/bond_main.c:        bond_dev->priv_flags |= IFF_BONDING | IFF_UNICAST_FLT | IFF_NO_QUEUE;
drivers/net/caif/caif_serial.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/dummy.c:    dev->priv_flags |= IFF_LIVE_ADDR_CHANGE | IFF_NO_QUEUE;
drivers/net/ethernet/chelsio/cxgb4/cxgb4_main.c:        dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/ethernet/netronome/nfp/nfp_net_repr.c:      netdev->priv_flags |= IFF_NO_QUEUE | IFF_DISABLE_NETPOLL;
drivers/net/geneve.c:   dev->priv_flags |= IFF_LIVE_ADDR_CHANGE | IFF_NO_QUEUE;
drivers/net/gtp.c:      dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/ipvlan/ipvlan_main.c:       dev->priv_flags |= IFF_UNICAST_FLT | IFF_NO_QUEUE;
drivers/net/ipvlan/ipvtap.c:    dev->priv_flags &= ~IFF_NO_QUEUE;
drivers/net/loopback.c: dev->priv_flags         |= IFF_LIVE_ADDR_CHANGE | IFF_NO_QUEUE;
drivers/net/macsec.c:   dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/macvlan.c:  dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/net_failover.c:     failover_dev->priv_flags |= IFF_UNICAST_FLT | IFF_NO_QUEUE;
drivers/net/netdevsim/netdev.c:                    IFF_NO_QUEUE;
drivers/net/netkit.c:   dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/nlmon.c:    dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/pfcp.c:     dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/team/team_core.c:   dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/veth.c:     dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/vrf.c:      dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/vsockmon.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/vxlan/vxlan_core.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/wan/hdlc_fr.c:      dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/wireguard/device.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/wireless/virtual/mac80211_hwsim.c:  dev->priv_flags |= IFF_NO_QUEUE;

And noqueue_qdisc_ops sets ->enqueue to noop_enqueue():

struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {

Quoted 8 lines
        .id             =       "noqueue",
        .priv_size      =       0,
        .init           =       noqueue_init,
        .enqueue        =       noop_enqueue,
        .dequeue        =       noop_dequeue,
        .peek           =       noop_dequeue,
        .owner          =       THIS_MODULE,

};

Sure, but please a look at :

static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt, struct netlink_ext_ack *extack)

{
        /* register_qdisc() assigns a default of noop_enqueue if unset,
        * but __dev_queue_xmit() treats noqueue only as such
        * if this is NULL - so clear it here. */
        qdisc->enqueue = NULL;
        return 0;
}
#3 lore raw
Quoted 62 lines · Eric Dumazet
Quoted 47 lines · Cong Wang
Quoted 9 lines · Eric Dumazet

dev_deactivate_many() role is to remove the qdiscs of a network device.

When/if a qdisc is dismantled, an rcu grace period is needed to make sure all outstanding qdisc enqueue are done before we proceed with a qdisc reset.

Most virtual devices do not have a qdisc (if we exclude noqueue ones).

Such as? To me, most virtual devices use noqueue:

$ git grep IFF_NO_QUEUE -- drivers/net/
drivers/net/amt.c:      dev->priv_flags         |= IFF_NO_QUEUE;
drivers/net/bareudp.c:  dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/bonding/bond_main.c:        bond_dev->priv_flags |= IFF_BONDING | IFF_UNICAST_FLT | IFF_NO_QUEUE;
drivers/net/caif/caif_serial.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/dummy.c:    dev->priv_flags |= IFF_LIVE_ADDR_CHANGE | IFF_NO_QUEUE;
drivers/net/ethernet/chelsio/cxgb4/cxgb4_main.c:        dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/ethernet/netronome/nfp/nfp_net_repr.c:      netdev->priv_flags |= IFF_NO_QUEUE | IFF_DISABLE_NETPOLL;
drivers/net/geneve.c:   dev->priv_flags |= IFF_LIVE_ADDR_CHANGE | IFF_NO_QUEUE;
drivers/net/gtp.c:      dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/ipvlan/ipvlan_main.c:       dev->priv_flags |= IFF_UNICAST_FLT | IFF_NO_QUEUE;
drivers/net/ipvlan/ipvtap.c:    dev->priv_flags &= ~IFF_NO_QUEUE;
drivers/net/loopback.c: dev->priv_flags         |= IFF_LIVE_ADDR_CHANGE | IFF_NO_QUEUE;
drivers/net/macsec.c:   dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/macvlan.c:  dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/net_failover.c:     failover_dev->priv_flags |= IFF_UNICAST_FLT | IFF_NO_QUEUE;
drivers/net/netdevsim/netdev.c:                    IFF_NO_QUEUE;
drivers/net/netkit.c:   dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/nlmon.c:    dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/pfcp.c:     dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/team/team_core.c:   dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/veth.c:     dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/vrf.c:      dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/vsockmon.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/vxlan/vxlan_core.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/wan/hdlc_fr.c:      dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/wireguard/device.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/wireless/virtual/mac80211_hwsim.c:  dev->priv_flags |= IFF_NO_QUEUE;

And noqueue_qdisc_ops sets ->enqueue to noop_enqueue():

struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {

Quoted 8 lines
        .id             =       "noqueue",
        .priv_size      =       0,
        .init           =       noqueue_init,
        .enqueue        =       noop_enqueue,
        .dequeue        =       noop_dequeue,
        .peek           =       noop_dequeue,
        .owner          =       THIS_MODULE,

};

Sure, but please a look at :

static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt, struct netlink_ext_ack *extack)

Quoted 7 lines
{
        /* register_qdisc() assigns a default of noop_enqueue if unset,
        * but __dev_queue_xmit() treats noqueue only as such
        * if this is NULL - so clear it here. */
        qdisc->enqueue = NULL;
        return 0;
}

How can we proceed on this patch ?

I can remove the "(if we exclude noqueue ones)" part if this is confusing.

maintainer
#4 lore raw
Quoted 71 lines · Eric Dumazet
Quoted 62 lines · Eric Dumazet
Quoted 47 lines · Cong Wang
Quoted 9 lines · Eric Dumazet

dev_deactivate_many() role is to remove the qdiscs of a network device.

When/if a qdisc is dismantled, an rcu grace period is needed to make sure all outstanding qdisc enqueue are done before we proceed with a qdisc reset.

Most virtual devices do not have a qdisc (if we exclude noqueue ones).

Such as? To me, most virtual devices use noqueue:

$ git grep IFF_NO_QUEUE -- drivers/net/
drivers/net/amt.c:      dev->priv_flags         |= IFF_NO_QUEUE;
drivers/net/bareudp.c:  dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/bonding/bond_main.c:        bond_dev->priv_flags |= IFF_BONDING | IFF_UNICAST_FLT | IFF_NO_QUEUE;
drivers/net/caif/caif_serial.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/dummy.c:    dev->priv_flags |= IFF_LIVE_ADDR_CHANGE | IFF_NO_QUEUE;
drivers/net/ethernet/chelsio/cxgb4/cxgb4_main.c:        dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/ethernet/netronome/nfp/nfp_net_repr.c:      netdev->priv_flags |= IFF_NO_QUEUE | IFF_DISABLE_NETPOLL;
drivers/net/geneve.c:   dev->priv_flags |= IFF_LIVE_ADDR_CHANGE | IFF_NO_QUEUE;
drivers/net/gtp.c:      dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/ipvlan/ipvlan_main.c:       dev->priv_flags |= IFF_UNICAST_FLT | IFF_NO_QUEUE;
drivers/net/ipvlan/ipvtap.c:    dev->priv_flags &= ~IFF_NO_QUEUE;
drivers/net/loopback.c: dev->priv_flags         |= IFF_LIVE_ADDR_CHANGE | IFF_NO_QUEUE;
drivers/net/macsec.c:   dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/macvlan.c:  dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/net_failover.c:     failover_dev->priv_flags |= IFF_UNICAST_FLT | IFF_NO_QUEUE;
drivers/net/netdevsim/netdev.c:                    IFF_NO_QUEUE;
drivers/net/netkit.c:   dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/nlmon.c:    dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/pfcp.c:     dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/team/team_core.c:   dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/veth.c:     dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/vrf.c:      dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/vsockmon.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/vxlan/vxlan_core.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/wan/hdlc_fr.c:      dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/wireguard/device.c: dev->priv_flags |= IFF_NO_QUEUE;
drivers/net/wireless/virtual/mac80211_hwsim.c:  dev->priv_flags |= IFF_NO_QUEUE;

And noqueue_qdisc_ops sets ->enqueue to noop_enqueue():

struct Qdisc_ops noqueue_qdisc_ops __read_mostly = {

Quoted 8 lines
        .id             =       "noqueue",
        .priv_size      =       0,
        .init           =       noqueue_init,
        .enqueue        =       noop_enqueue,
        .dequeue        =       noop_dequeue,
        .peek           =       noop_dequeue,
        .owner          =       THIS_MODULE,

};

Sure, but please a look at :

static int noqueue_init(struct Qdisc *qdisc, struct nlattr *opt, struct netlink_ext_ack *extack)

Quoted 7 lines
{
        /* register_qdisc() assigns a default of noop_enqueue if unset,
        * but __dev_queue_xmit() treats noqueue only as such
        * if this is NULL - so clear it here. */
        qdisc->enqueue = NULL;
        return 0;
}
Quoted 3 lines

How can we proceed on this patch ?

I can remove the "(if we exclude noqueue ones)" part if this is confusing.

I personally interpret the lack of reply from Cong to your previous message as agreement/understanding. I'll drop the blamed sentence from the commit message when applying the patch.

BTW, in the long run I think it would be nice to remove the noqueue_init hack - AFAICS no in-tree kernel has NULL enqueue/dequeue/peek CB, so the related checks in register_qdisc() could possibly be dropped.

/P

#5 lore raw
In reply to Eric Dumazet
This patch was applied to netdev/net-next.git (main) by Paolo Abeni <[email protected]>:

Hello:

This patch was applied to netdev/net-next.git (main) by Paolo Abeni <[email protected]>:

Quoted 8 lines · On Thu, 9 Jan 2025 17:18:50 +0000 you wrote:

dev_deactivate_many() role is to remove the qdiscs of a network device.

When/if a qdisc is dismantled, an rcu grace period is needed to make sure all outstanding qdisc enqueue are done before we proceed with a qdisc reset.

[...]

Here is the summary with links:
- [net-next] net: sched: calls synchronize_net() only when needed

    https://git.kernel.org/netdev/net-next/c/88df16f851ad

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html