[Patch net-next] pktgen: use the net namespace of current process instead of init_net

From: Cong Wang <redacted>

Use the net namespace of the current process loading
pktgen module.

Cc: Eric W. Biederman <redacted>
Cc: David S. Miller <davem@davemloft.net>
Signed-off-by: Cong Wang <redacted>

---

diff --git a/net/core/pktgen.c b/net/core/pktgen.c
index b29dacf..db1f444 100644
--- a/net/core/pktgen.c
+++ b/net/core/pktgen.c

@@ -440,6 +440,7 @@ static int pg_count_d __read_mostly = 1000;
 static int pg_delay_d __read_mostly;
 static int pg_clone_skb_d  __read_mostly;
 static int debug  __read_mostly;
+static struct net *pktgen_net __read_mostly;
 
 static DEFINE_MUTEX(pktgen_thread_lock);
 static LIST_HEAD(pktgen_threads);

@@ -1910,7 +1911,7 @@ static int pktgen_device_event(struct notifier_block *unused,
 {
 	struct net_device *dev = ptr;
 
-	if (!net_eq(dev_net(dev), &init_net) || pktgen_exiting)
+	if (!net_eq(dev_net(dev), pktgen_net) || pktgen_exiting)
 		return NOTIFY_DONE;
 
 	/* It is OK that we do not hold the group lock right now,

@@ -1944,7 +1945,7 @@ static struct net_device *pktgen_dev_get_by_name(struct pktgen_dev *pkt_dev,
 	}
 	b[i] = 0;
 
-	return dev_get_by_name(&init_net, b);
+	return dev_get_by_name(pktgen_net, b);
 }
 
 

@@ -2205,7 +2206,7 @@ static void get_ipsec_sa(struct pktgen_dev *pkt_dev, int flow)
 	struct xfrm_state *x = pkt_dev->flows[flow].x;
 	if (!x) {
 		/*slow path: we dont already have xfrm_state*/
-		x = xfrm_stateonly_find(&init_net, DUMMY_MARK,
+		x = xfrm_stateonly_find(pktgen_net, DUMMY_MARK,
 					(xfrm_address_t *)&pkt_dev->cur_daddr,
 					(xfrm_address_t *)&pkt_dev->cur_saddr,
 					AF_INET,

@@ -3600,7 +3601,8 @@ static int __init pg_init(void)
 
 	pr_info("%s", version);
 
-	pg_proc_dir = proc_mkdir(PG_PROC_DIR, init_net.proc_net);
+	pktgen_net = current->nsproxy->net_ns;
+	pg_proc_dir = proc_mkdir(PG_PROC_DIR, pktgen_net->proc_net);
 	if (!pg_proc_dir)
 		return -ENODEV;
 

@@ -3634,7 +3636,7 @@ static int __init pg_init(void)
 	unregister_netdevice_notifier(&pktgen_notifier_block);
 	remove_proc_entry(PGCTRL, pg_proc_dir);
  remove_dir:
-	proc_net_remove(&init_net, PG_PROC_DIR);
+	proc_net_remove(pktgen_net, PG_PROC_DIR);
 	return ret;
 }
 

Cong Wang [off-list ref] writes:

From: Cong Wang <redacted>

Use the net namespace of the current process loading
pktgen module.

Why not add proper support for multiple network namespaces?

It looks like the proc files could be created in each network namespace
and be controlled from there.

Your patch below looks like it could have some unintended effects
if pktgen was loaded automatically or compiled into the kernel.

Eric

On Sun, 2013-01-27 at 18:48 -0800, Eric W. Biederman wrote:

Cong Wang [off-list ref] writes:

quoted

From: Cong Wang <redacted>

Use the net namespace of the current process loading
pktgen module.

Why not add proper support for multiple network namespaces?

It looks like the proc files could be created in each network namespace
and be controlled from there.

Hmm, I didn't think much about this. You are right... we could probably
make these proc files per-net.

Your patch below looks like it could have some unintended effects
if pktgen was loaded automatically or compiled into the kernel.

In this case, 'current' should be the init process whose net namespace
is init_net, no?

Thanks.

On Mon, 2013-01-28 at 10:57 +0800, Cong Wang wrote:

Hmm, I didn't think much about this. You are right... we could probably
make these proc files per-net.

Cooking a patch now.

Cong Wang [off-list ref] writes:

On Sun, 2013-01-27 at 18:48 -0800, Eric W. Biederman wrote:

quoted

Cong Wang [off-list ref] writes:

quoted

From: Cong Wang <redacted>

Use the net namespace of the current process loading
pktgen module.

Why not add proper support for multiple network namespaces?

It looks like the proc files could be created in each network namespace
and be controlled from there.

Hmm, I didn't think much about this. You are right... we could probably
make these proc files per-net.

quoted

Your patch below looks like it could have some unintended effects
if pktgen was loaded automatically or compiled into the kernel.

In this case, 'current' should be the init process whose net namespace
is init_net, no?

If it is a kernel thread that decided to load the module.  If it
something else who knows.  To my knowledge there are no guarantees
and I'm not particularly interested in having to worry about current
making sense in a module.

Eric