[PATCH 2/12] L2 network namespace: network devices virtualization

[Eric W. Biederman]

Dmitry Mishin <dim at openvz.org> writes:

> Added ability to have per-namespace network devices.
>
> Signed-off-by: Dmitry Mishin <dim at openvz.org>
>
> ---
>  include/linux/net_namespace.h |    6 ++-
>  include/linux/netdevice.h     |   10 +++++
>  net/core/dev.c | 78 +++++++++++++++++++++++++++++++++++-------
>  net/core/net-sysfs.c          |   23 ++++++++++++
>  net/core/net_namespace.c      |   11 +++++
>  5 files changed, 114 insertions(+), 14 deletions(-)
>
> --- linux-2.6.19-rc6-mm2.orig/include/linux/net_namespace.h
> +++ linux-2.6.19-rc6-mm2/include/linux/net_namespace.h
> @@ -6,8 +6,10 @@
>  #include <linux/errno.h>
>  
>  struct net_namespace {
> -	struct kref	kref;
> -	struct nsproxy	*ns;
> +	struct kref		kref;
> +	struct nsproxy		*ns;
> +	struct net_device	*dev_base_p, **dev_tail_p;
> +	unsigned int		hash;
>  };
>  
>  extern struct net_namespace init_net_ns;
> --- linux-2.6.19-rc6-mm2.orig/include/linux/netdevice.h
> +++ linux-2.6.19-rc6-mm2/include/linux/netdevice.h
> @@ -379,6 +379,9 @@ struct net_device
>  	int			promiscuity;
>  	int			allmulti;
>  
> +#ifdef CONFIG_NET_NS
> +	struct net_namespace	*net_ns;
> +#endif
>  
>  	/* Protocol specific pointers */
>  	
> @@ -557,9 +560,16 @@ struct packet_type {
>  
>  #include <linux/interrupt.h>
>  #include <linux/notifier.h>
> +#include <linux/net_namespace.h>
>  
>  extern struct net_device loopback_dev; /* The loopback */
> +#ifndef CONFIG_NET_NS
>  extern struct net_device *dev_base; /* All devices */
> +#define dev_base_ns(dev)		dev_base
> +#else
> +#define dev_base			(current_net_ns->dev_base_p)
> +#define dev_base_ns(dev)		(dev->net_ns->dev_base_p)
> +#endif
>  extern rwlock_t dev_base_lock; /* Device list lock */
>  
>  extern int netdev_boot_setup_check(struct net_device *dev);
> --- linux-2.6.19-rc6-mm2.orig/net/core/dev.c
> +++ linux-2.6.19-rc6-mm2/net/core/dev.c
> @@ -90,6 +90,7 @@
>  #include <linux/if_ether.h>
>  #include <linux/netdevice.h>
>  #include <linux/etherdevice.h>
> +#include <linux/net_namespace.h>
>  #include <linux/notifier.h>
>  #include <linux/skbuff.h>
>  #include <net/sock.h>
> @@ -174,20 +175,28 @@ static spinlock_t net_dma_event_lock;
>   * unregister_netdevice(), which must be called with the rtnl
>   * semaphore held.
>   */
> +#ifndef CONFIG_NET_NS
>  struct net_device *dev_base;
>  static struct net_device **dev_tail = &dev_base;
> -DEFINE_RWLOCK(dev_base_lock);
> -
> +#define dev_tail_ns(dev)	dev_tail
>  EXPORT_SYMBOL(dev_base);
> +#else
> +#define dev_tail		(current_net_ns->dev_tail_p)
> +#define dev_tail_ns(dev)	(dev->net_ns->dev_tail_p)
> +#endif

No.   Please do not hide a macro expansion (i.e. dev_tail).
That has the possibility of hiding costs, that shouldn't be hidden.

> --- linux-2.6.19-rc6-mm2.orig/net/core/net-sysfs.c
> +++ linux-2.6.19-rc6-mm2/net/core/net-sysfs.c
> @@ -453,6 +453,12 @@ static struct class net_class = {
>  
>  void netdev_unregister_sysfs(struct net_device * net)
>  {
> +#ifdef CONFIG_NET_NS
> +	if (net->net_ns != &init_net_ns)
> +		/* not supported yet: sysfs virtualization is required */
> +		return;
> +#endif
> +
>  	device_del(&(net->dev));
>  }

This is something we need to seriously look at how to handle sysfs.

I believe separate directories per namespace and a magic symlink
is going to be the sane approach.


Eric