2016-10-25 02:10:01

by Andrei Vagin

Subject: [PATCH v2] net: skip genenerating uevents for network namespaces that are exiting

No one can see these events, because a network namespace can not be
destroyed, if it has sockets.

Unlike other devices, uevent-s for network devices are generated
only inside their network namespaces. They are filtered in
kobj_bcast_filter()

My experiments shows that net namespaces are destroyed more 30% faster
with this optimization.

Here is a perf output for destroying network namespaces without this
patch.

- 94.76% 0.02% kworker/u48:1 [kernel.kallsyms] [k] cleanup_net
- 94.74% cleanup_net
- 94.64% ops_exit_list.isra.4
- 41.61% default_device_exit_batch
- 41.47% unregister_netdevice_many
- rollback_registered_many
- 40.36% netdev_unregister_kobject
- 14.55% device_del
+ 13.71% kobject_uevent
- 13.04% netdev_queue_update_kobjects
+ 12.96% kobject_put
- 12.72% net_rx_queue_update_kobjects
kobject_put
- kobject_release
+ 12.69% kobject_uevent
+ 0.80% call_netdevice_notifiers_info
+ 19.57% nfsd_exit_net
+ 11.15% tcp_net_metrics_exit
+ 8.25% rpcsec_gss_exit_net

It's very critical to optimize the exit path for network namespaces,
because they are destroyed under net_mutex and many namespaces can be
destroyed for one iteration.

v2: use dev_set_uevent_suppress()

Cc: Cong Wang <[email protected]>
Cc: "David S. Miller" <[email protected]>
Cc: Eric W. Biederman <[email protected]>
Signed-off-by: Andrei Vagin <[email protected]>
---
net/core/net-sysfs.c | 14 +++++++++++---
1 file changed, 11 insertions(+), 3 deletions(-)

diff --git a/net/core/net-sysfs.c b/net/core/net-sysfs.c
index 6e4f347..d4fe286 100644
--- a/net/core/net-sysfs.c
+++ b/net/core/net-sysfs.c
@@ -950,10 +950,13 @@ net_rx_queue_update_kobjects(struct net_device *dev, int old_num, int new_num)
}

while (--i >= new_num) {
+ struct kobject *kobj = &dev->_rx[i].kobj;
+
+ if (!list_empty(&dev_net(dev)->exit_list))
+ kobj->uevent_suppress = 1;
if (dev->sysfs_rx_queue_group)
- sysfs_remove_group(&dev->_rx[i].kobj,
- dev->sysfs_rx_queue_group);
- kobject_put(&dev->_rx[i].kobj);
+ sysfs_remove_group(kobj, dev->sysfs_rx_queue_group);
+ kobject_put(kobj);
}

return error;
@@ -1340,6 +1343,8 @@ netdev_queue_update_kobjects(struct net_device *dev, int old_num, int new_num)
while (--i >= new_num) {
struct netdev_queue *queue = dev->_tx + i;

+ if (!list_empty(&dev_net(dev)->exit_list))
+ queue->kobj.uevent_suppress = 1;
#ifdef CONFIG_BQL
sysfs_remove_group(&queue->kobj, &dql_group);
#endif
@@ -1525,6 +1530,9 @@ void netdev_unregister_kobject(struct net_device *ndev)
{
struct device *dev = &(ndev->dev);

+ if (!list_empty(&dev_net(ndev)->exit_list))
+ dev_set_uevent_suppress(dev, 1);
+
kobject_get(&dev->kobj);

remove_queue_kobjects(ndev);
--
2.7.4

2016-10-27 21:15:28

by David Miller

Subject: Re: [PATCH v2] net: skip genenerating uevents for network namespaces that are exiting

From: Andrei Vagin <[email protected]>
Date: Mon, 24 Oct 2016 19:09:53 -0700

> No one can see these events, because a network namespace can not be
> destroyed, if it has sockets.
>
> Unlike other devices, uevent-s for network devices are generated
> only inside their network namespaces. They are filtered in
> kobj_bcast_filter()
>
> My experiments shows that net namespaces are destroyed more 30% faster
> with this optimization.
>
> Here is a perf output for destroying network namespaces without this
> patch.
...
> It's very critical to optimize the exit path for network namespaces,
> because they are destroyed under net_mutex and many namespaces can be
> destroyed for one iteration.
>
> v2: use dev_set_uevent_suppress()
>
> Cc: Cong Wang <[email protected]>
> Cc: "David S. Miller" <[email protected]>
> Cc: Eric W. Biederman <[email protected]>
> Signed-off-by: Andrei Vagin <[email protected]>

Applied to net-next, thanks.