Re: [PATCH 4/6] vlan: Optimize multiple unregistration

[Eric Dumazet <eric.dumazet@gmail.com>]

Patrick McHardy a écrit :
> Eric Dumazet wrote:
>> Use unregister_netdevice_many() to speedup master device unregister.
>>
>> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
>> ---
>>  include/linux/if_vlan.h |    1 
>>  net/8021q/vlan.c        |   49 +++++++++++++++++++++++++-------------
>>  net/core/dev.c          |    1 
>>  3 files changed, 35 insertions(+), 16 deletions(-)
>>
>> diff --git a/include/linux/if_vlan.h b/include/linux/if_vlan.h
>> index 8898cbe..71a4870 100644
>> --- a/include/linux/if_vlan.h
>> +++ b/include/linux/if_vlan.h
>> @@ -85,6 +85,7 @@ struct vlan_group {
>>  					    * the vlan is attached to.
>>  					    */
>>  	unsigned int		nr_vlans;
>> +	int			killall;
>>  	struct hlist_node	hlist;	/* linked list */
>>  	struct net_device **vlan_devices_arrays[VLAN_GROUP_ARRAY_SPLIT_PARTS];
>>  	struct rcu_head		rcu;
>> diff --git a/net/8021q/vlan.c b/net/8021q/vlan.c
>> index 6b5c9dd..511afe7 100644
>> --- a/net/8021q/vlan.c
>> +++ b/net/8021q/vlan.c
>> @@ -159,11 +159,12 @@ void unregister_vlan_dev(struct net_device *dev, struct list_head *head)
>>  	if (real_dev->features & NETIF_F_HW_VLAN_FILTER)
>>  		ops->ndo_vlan_rx_kill_vid(real_dev, vlan_id);
>>  
>> -	vlan_group_set_device(grp, vlan_id, NULL);
>>  	grp->nr_vlans--;
>>  
>> -	synchronize_net();
>> -
>> +	if (!grp->killall) {
>> +		vlan_group_set_device(grp, vlan_id, NULL);
>> +		synchronize_net();
>> +	}
>>  	unregister_netdevice_queue(dev, head);
>>  
>>  	/* If the group is now empty, kill off the group. */
>> @@ -183,6 +184,34 @@ void unregister_vlan_dev(struct net_device *dev, struct list_head *head)
>>  	dev_put(real_dev);
>>  }
>>  
>> +void unregister_vlan_dev_alls(struct vlan_group *grp)
> 
> This could be static.
> 
>> +{
>> +	LIST_HEAD(list);
>> +	int i;
>> +	struct net_device *vlandev;
>> +	struct vlan_group save;
>> +
>> +	memcpy(&save, grp, sizeof(save));
>> +	memset(&grp->vlan_devices_arrays, 0, sizeof(grp->vlan_devices_arrays));
> 
> This shouldn't be necessary since the lower device is already in the
> process of being unregistered. If it was necessary, it could cause
> crashes since the individual pointers are not set to zero atomically.
> Or maybe I'm misunderstanding the purpose entirely :)

Very good point indeed, even if in practice memset() use long word transferts

I'll make a cleanup patch, or do you want to do it ?