[PATCH net-next] pppoe: optional deactivation of PADT packet handling

Kernel space shouldn't handle PADT packets since PADT belongs to
PPPoE's control plane.
With "handle_padt" module option, user space can now decide to avoid
kernel interpretation of PADT packets and be responsible for session
disconnection.

Signed-off-by: Guillaume Nault <redacted>
---
The two open-source PPPoE programs I know of are accel-ppp and pppd
(with the rp-pppoe plugin). Accel-ppp already handles PADT, while
pppd + rp-pppoe doesn't (seems like it doesn't even listen to its
PPPoE Discovery socket once session is established).


 drivers/net/ppp/pppoe.c | 18 ++++++++++++++++--
 1 file changed, 16 insertions(+), 2 deletions(-)

diff --git a/drivers/net/ppp/pppoe.c b/drivers/net/ppp/pppoe.c
index b8da2ea..1846c3a 100644
--- a/drivers/net/ppp/pppoe.c
+++ b/drivers/net/ppp/pppoe.c

@@ -109,6 +109,18 @@ struct pppoe_net {
 	rwlock_t hash_lock;
 };
 
+/* PADT packets belong to PPPoE's discovery stage (i.e. control plane)
+ * and should be handled by user space. Unfortunately, some programs
+ * don't listen out for PADT packets and rely on pppoe module for
+ * closing session after reception of PADT.
+ *
+ * This option can be deactivated when PPPoE discovery is handled by
+ * a PADT aware program.
+ */
+static bool handle_padt __read_mostly = true;
+module_param(handle_padt, bool, 0444);
+MODULE_PARM_DESC(handle_padt, "Let kernel interpret incoming PADT");
+
 /*
  * PPPoE could be in the following stages:
  * 1) Discovery stage (to obtain remote MAC and Session ID)

@@ -1173,7 +1185,8 @@ static int __init pppoe_init(void)
 		goto out_unregister_pppoe_proto;
 
 	dev_add_pack(&pppoes_ptype);
-	dev_add_pack(&pppoed_ptype);
+	if (handle_padt)
+		dev_add_pack(&pppoed_ptype);
 	register_netdevice_notifier(&pppoe_notifier);
 
 	return 0;

@@ -1189,7 +1202,8 @@ out:
 static void __exit pppoe_exit(void)
 {
 	unregister_netdevice_notifier(&pppoe_notifier);
-	dev_remove_pack(&pppoed_ptype);
+	if (handle_padt)
+		dev_remove_pack(&pppoed_ptype);
 	dev_remove_pack(&pppoes_ptype);
 	unregister_pppox_proto(PX_PROTO_OE);
 	proto_unregister(&pppoe_sk_proto);

-- 
2.6.4

On Tue, 2015-12-15 at 17:06 +0100, Guillaume Nault wrote:

Kernel space shouldn't handle PADT packets since PADT belongs to
PPPoE's control plane.
With "handle_padt" module option, user space can now decide to avoid
kernel interpretation of PADT packets and be responsible for session
disconnection.

In general, module options like this kinda suck.  How about a new PPPoE
ioctl so the option can be toggled for a given socket/session instead? 
 Then at least you're not locked into a specific PPPoE implementation
for all sessions on the machine.

Dan

From: Guillaume Nault <redacted>
Date: Tue, 15 Dec 2015 17:06:39 +0100

No module options, please.

On Tue, Dec 15, 2015 at 10:50:05AM -0600, Dan Williams wrote:

On Tue, 2015-12-15 at 17:06 +0100, Guillaume Nault wrote:

quoted

Kernel space shouldn't handle PADT packets since PADT belongs to
PPPoE's control plane.
With "handle_padt" module option, user space can now decide to avoid
kernel interpretation of PADT packets and be responsible for session
disconnection.

In general, module options like this kinda suck.  How about a new PPPoE
ioctl so the option can be toggled for a given socket/session instead? 
 Then at least you're not locked into a specific PPPoE implementation
for all sessions on the machine.

I've considered this approach too, but by using per session option,
we still have to register pppoed_ptype in pppoe_init() and thus call
pppoe_disc_recv() for any ETH_P_PPP_DISC frame, just in case a session
needs kernel space PADT handling. This makes the approach much less
interesting IMO, hence the original choice for module wide behaviour.
On the other hand, I understand the need for minimising module options.

Thinking a bit more about it, we could drop dev_add_pack() from
pppoe_init() and let pppoe_connect() register ETH_P_PPP_DISC when
necessary. We could even do it per device, as pppoe_connect() knows the
lower device.
The problem is to know when unregistering ETH_P_PPP_DISC. We could
probably use a reference counter, but I'd prefer to avoid complicating
pppoe's connection and disconnection code given that recent history
has shown its relative fragility.