Skip to contentSkip to conversation

net: Add DEVTYPE support for Ethernet based devices

PATCH Applied Marcel Holtmann 4 participants 12 messages Sep 1, 2009 – Sep 11, 2009

Subsystems (7)
lore
Share & tools
lore

Marcel Holtmann

author maintainer
#0 lore raw

PATCH 1/1 net: Add DEVTYPE support for Ethernet based devices

The Ethernet framing is used for a lot of devices these days. Most prominent are WiFi and WiMAX based devices. However for userspace application it is important to classify these devices correctly and not only see them as Ethernet devices. The daemons like HAL, DeviceKit or even NetworkManager with udev support tries to do the classification in userspace with a lot trickery and extra system calls. This is not good and actually reaches its limitations. Especially since the kernel does know the type of the Ethernet device it is pretty stupid.

To solve this problem the underlying device type needs to be set and then the value will be exported as DEVTYPE via uevents and available within udev.

  # cat /sys/class/net/wlan0/uevent
  DEVTYPE=wlan
  INTERFACE=wlan0
  IFINDEX=5

This is similar to subsystems like USB and SCSI that distinguish between hosts, devices, disks, partitions etc.

The new SET_NETDEV_DEVTYPE() is a convenience helper to set the actual device type. All device types are free form, but for convenience the same strings as used with RFKILL are choosen.

---
 drivers/net/usb/hso.c           |    5 +++++
 drivers/net/wimax/i2400m/sdio.c |    5 +++++
 drivers/net/wimax/i2400m/usb.c  |    5 +++++
 include/linux/netdevice.h       |    6 ++++++
 net/bluetooth/bnep/core.c       |    5 +++++
 net/bridge/br_if.c              |    6 ++++++
 net/mac80211/iface.c            |    5 +++++
 7 files changed, 37 insertions(+), 0 deletions(-)
drivers/net/usb/hso.c +5 -0
index ffe4106..5d47a73 100644
@@ -2534,6 +2534,10 @@ static void hso_create_rfkill(struct hso_device *hso_dev,
 	}
 }
 
+static struct device_type hso_type = {
+	.name	= "wwan",
+};
+
 /* Creates our network device */
 static struct hso_device *hso_create_net_device(struct usb_interface *interface,
 						int port_spec)
@@ -2574,6 +2578,7 @@ static struct hso_device *hso_create_net_device(struct usb_interface *interface,
 		goto exit;
 	}
 	SET_NETDEV_DEV(net, &interface->dev);
+	SET_NETDEV_DEVTYPE(net, &hso_type);
 
 	/* registering our net device */
 	result = register_netdev(net);
drivers/net/wimax/i2400m/sdio.c +5 -0
index ea7b290..2981e21 100644
@@ -371,6 +371,10 @@ error:
 }
 
 
+static struct device_type i2400ms_type = {
+	.name	= "wimax",
+};
+
 /*
  * Probe a i2400m interface and register it
  *
@@ -412,6 +416,7 @@ int i2400ms_probe(struct sdio_func *func,
 		goto error_alloc_netdev;
 	}
 	SET_NETDEV_DEV(net_dev, dev);
+	SET_NETDEV_DEVTYPE(net_dev, &i2400ms_type);
 	i2400m = net_dev_to_i2400m(net_dev);
 	i2400ms = container_of(i2400m, struct i2400ms, i2400m);
 	i2400m->wimax_dev.net_dev = net_dev;
drivers/net/wimax/i2400m/usb.c +5 -0
index cfdaf69..7eadd11 100644
@@ -351,6 +351,10 @@ error:
 }
 
 
+static struct device_type i2400mu_type = {
+	.name	= "wimax",
+};
+
 /*
  * Probe a i2400m interface and register it
  *
@@ -388,6 +392,7 @@ int i2400mu_probe(struct usb_interface *iface,
 		goto error_alloc_netdev;
 	}
 	SET_NETDEV_DEV(net_dev, dev);
+	SET_NETDEV_DEVTYPE(net_dev, &i2400mu_type);
 	i2400m = net_dev_to_i2400m(net_dev);
 	i2400mu = container_of(i2400m, struct i2400mu, i2400m);
 	i2400m->wimax_dev.net_dev = net_dev;
include/linux/netdevice.h +6 -0
index 60d3aac..8e00e03 100644
@@ -989,6 +989,12 @@ static inline void *netdev_priv(const struct net_device *dev)
  */
 #define SET_NETDEV_DEV(net, pdev)	((net)->dev.parent = (pdev))
 
+/* Set the sysfs device type for the network logical device to allow
+ * fin grained indentification of different network device types. For
+ * example Ethernet, Wirelss LAN, Bluetooth, WiMAX etc.
+ */
+#define SET_NETDEV_DEVTYPE(net, devtype)	((net)->dev.type = (devtype))
+
 /**
  *	netif_napi_add - initialize a napi context
  *	@dev:  network device
net/bluetooth/bnep/core.c +5 -0
index 52a6ce0..cafe9f5 100644
@@ -533,6 +533,10 @@ static struct device *bnep_get_device(struct bnep_session *session)
 	return conn ? &conn->dev : NULL;
 }
 
+static struct device_type bnep_type = {
+	.name	= "bluetooth",
+};
+
 int bnep_add_connection(struct bnep_connadd_req *req, struct socket *sock)
 {
 	struct net_device *dev;
@@ -586,6 +590,7 @@ int bnep_add_connection(struct bnep_connadd_req *req, struct socket *sock)
 #endif
 
 	SET_NETDEV_DEV(dev, bnep_get_device(s));
+	SET_NETDEV_DEVTYPE(dev, &bnep_type);
 
 	err = register_netdev(dev);
 	if (err) {
net/bridge/br_if.c +6 -0
index e486f1f..142ebac 100644
@@ -264,6 +264,10 @@ static struct net_bridge_port *new_nbp(struct net_bridge *br,
 	return p;
 }
 
+static struct device_type br_type = {
+	.name	= "bridge",
+};
+
 int br_add_bridge(struct net *net, const char *name)
 {
 	struct net_device *dev;
@@ -280,6 +284,8 @@ int br_add_bridge(struct net *net, const char *name)
 			goto out_free;
 	}
 
+	SET_NETDEV_DEVTYPE(dev, &br_type);
+
 	ret = register_netdevice(dev);
 	if (ret)
 		goto out_free;
net/mac80211/iface.c +5 -0
index f6005ad..b8295cb 100644
@@ -754,6 +754,10 @@ int ieee80211_if_change_type(struct ieee80211_sub_if_data *sdata,
 	return 0;
 }
 
+static struct device_type wiphy_type = {
+	.name	= "wlan",
+};
+
 int ieee80211_if_add(struct ieee80211_local *local, const char *name,
 		     struct net_device **new_dev, enum nl80211_iftype type,
 		     struct vif_params *params)
@@ -785,6 +789,7 @@ int ieee80211_if_add(struct ieee80211_local *local, const char *name,
 
 	memcpy(ndev->dev_addr, local->hw.wiphy->perm_addr, ETH_ALEN);
 	SET_NETDEV_DEV(ndev, wiphy_dev(local->hw.wiphy));
+	SET_NETDEV_DEVTYPE(ndev, &wiphy_type);
 
 	/* don't use IEEE80211_DEV_TO_SUB_IF because it checks too much */
 	sdata = netdev_priv(ndev);
Signature · 1 line
1.6.2.5
maintainer
#1 lore raw
Quoted 18 lines · Marcel Holtmann

The new SET_NETDEV_DEVTYPE() is a convenience helper to set the actual device type. All device types are free form, but for convenience the same strings as used with RFKILL are choosen.

net/mac80211/iface.c +4 -0
@@ -754,6 +754,10 @@ int ieee80211_if_change_type(struct ieee80211_sub_if_data *sdata,
 	return 0;
 }
 
+static struct device_type wiphy_type = {
+	.name	= "wlan",
+};
+
 int ieee80211_if_add(struct ieee80211_local *local, const char *name,
 		     struct net_device **new_dev, enum nl80211_iftype type,
 		     struct vif_params *params)

net/mac80211/iface.c int ieee80211_if_add(struct ieee80211_local *local, const char *name, · line 792

Quoted 4 lines
net/mac80211/iface.c +1 -0
@@ -785,6 +789,7 @@ int ieee80211_if_add(struct ieee80211_local *local, const char *name,
 
 	memcpy(ndev->dev_addr, local->hw.wiphy->perm_addr, ETH_ALEN);
 	SET_NETDEV_DEV(ndev, wiphy_dev(local->hw.wiphy));
+	SET_NETDEV_DEVTYPE(ndev, &wiphy_type);

Can't you move that into the NETDEV_REGISTER hook in cfg80211? That way orinoco, rndis and iwm don't need updating now.

johannes

author maintainer
#2 lore raw

Hi Johannes,

Quoted 19 lines
Quoted 19 lines

The new SET_NETDEV_DEVTYPE() is a convenience helper to set the actual device type. All device types are free form, but for convenience the same strings as used with RFKILL are choosen.

net/mac80211/iface.c +4 -0
@@ -754,6 +754,10 @@ int ieee80211_if_change_type(struct ieee80211_sub_if_data *sdata,
 	return 0;
 }
 
+static struct device_type wiphy_type = {
+	.name	= "wlan",
+};
+
 int ieee80211_if_add(struct ieee80211_local *local, const char *name,
 		     struct net_device **new_dev, enum nl80211_iftype type,
 		     struct vif_params *params)
@@ -785,6 +789,7 @@ int ieee80211_if_add(struct ieee80211_local *local, const char *name,
 
Quoted 6 lines

net/mac80211/iface.c int ieee80211_if_add(struct ieee80211_local *local, const char *name, · line 792

Quoted 3 lines
net/mac80211/iface.c +1 -0
 	memcpy(ndev->dev_addr, local->hw.wiphy->perm_addr, ETH_ALEN);
 	SET_NETDEV_DEV(ndev, wiphy_dev(local->hw.wiphy));
+	SET_NETDEV_DEVTYPE(ndev, &wiphy_type);

Can't you move that into the NETDEV_REGISTER hook in cfg80211? That way orinoco, rndis and iwm don't need updating now.

I am not sure if that would work, but it is worth testing it. If it works, I clearly would prefer this.

Regards

Marcel

author maintainer
#3 lore raw
In reply to Marcel Holtmann

Hi Dave,

Quoted 18 lines

The Ethernet framing is used for a lot of devices these days. Most prominent are WiFi and WiMAX based devices. However for userspace application it is important to classify these devices correctly and not only see them as Ethernet devices. The daemons like HAL, DeviceKit or even NetworkManager with udev support tries to do the classification in userspace with a lot trickery and extra system calls. This is not good and actually reaches its limitations. Especially since the kernel does know the type of the Ethernet device it is pretty stupid.

To solve this problem the underlying device type needs to be set and then the value will be exported as DEVTYPE via uevents and available within udev.

  # cat /sys/class/net/wlan0/uevent
  DEVTYPE=wlan
  INTERFACE=wlan0
  IFINDEX=5
Quoted 6 lines

This is similar to subsystems like USB and SCSI that distinguish between hosts, devices, disks, partitions etc.

The new SET_NETDEV_DEVTYPE() is a convenience helper to set the actual device type. All device types are free form, but for convenience the same strings as used with RFKILL are choosen.

can you please give me some feedback on this patch.

Regards

Marcel

maintainer
#4 lore raw

From: Marcel Holtmann <[email protected]>
Date: Sat, 05 Sep 2009 02:17:26 +0200

Quoted 1 line

can you please give me some feedback on this patch.

Johannes recommended using a NETDEV_REGISTER hook, I'm waiting for you to try and use that instead of this patch.

Your patch is ugly, so I'd like to avoid it if possible.

author maintainer
#5 lore raw

Hi Dave,

Quoted 5 lines
Quoted 1 line

can you please give me some feedback on this patch.

Johannes recommended using a NETDEV_REGISTER hook, I'm waiting for you to try and use that instead of this patch.

that was for the Wireless drivers part. Total different from what this patch trying to achieve.

Quoted 1 line

Your patch is ugly, so I'd like to avoid it if possible.

What is ugly about it. Do you have any other recommendation on how let userspace know what type of Ethernet device it is?

Regards

Marcel

maintainer
#6 lore raw

From: Marcel Holtmann <[email protected]>
Date: Sat, 05 Sep 2009 06:34:00 +0200

Quoted 2 lines

What is ugly about it. Do you have any other recommendation on how let userspace know what type of Ethernet device it is?

Well, for one thing, you're wasting 48 bytes on 64-bit when all you're really intrested in is one string.

Also, you didn't setup an assignment for plain ethernet devices, you just handle the non-traditional devices that do ethernet framing. Is this intentional?

author maintainer
#7 lore raw

Hi David,

Quoted 5 lines
Quoted 2 lines

What is ugly about it. Do you have any other recommendation on how let userspace know what type of Ethernet device it is?

Well, for one thing, you're wasting 48 bytes on 64-bit when all you're really intrested in is one string.

the integration with struct device has plenty of other advantages. You could use all the sysfs magic from it. For example in the next step the hacking of some sysfs directories can be moved into the subsystem itself and could be moved out of the core. Wireless would be one of them were we would benefit here.

Quoted 3 lines

Also, you didn't setup an assignment for plain ethernet devices, you just handle the non-traditional devices that do ethernet framing. Is this intentional?

For now it is. Mainly since some non plain Ethernet devices are using alloc_etherdev instead of just alloc_netdev. So I would have to do some extra work. I have planned it, but not gotten around it yet.

Regards

Marcel

maintainer
#8 lore raw

From: Marcel Holtmann <[email protected]>
Date: Sat, 05 Sep 2009 07:52:20 +0200

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines

What is ugly about it. Do you have any other recommendation on how let userspace know what type of Ethernet device it is?

Well, for one thing, you're wasting 48 bytes on 64-bit when all you're really intrested in is one string.

the integration with struct device has plenty of other advantages

Ok, this I agree with.

#9 lore raw
In reply to Marcel Holtmann
Quoted 9 lines · On Tue, 1 Sep 2009 00:08:19 -0700 Marcel Holtmann <[email protected]> wrote:

The Ethernet framing is used for a lot of devices these days. Most prominent are WiFi and WiMAX based devices. However for userspace application it is important to classify these devices correctly and not only see them as Ethernet devices. The daemons like HAL, DeviceKit or even NetworkManager with udev support tries to do the classification in userspace with a lot trickery and extra system calls. This is not good and actually reaches its limitations. Especially since the kernel does know the type of the Ethernet device it is pretty stupid.

Quoted 9 lines

To solve this problem the underlying device type needs to be set and then the value will be exported as DEVTYPE via uevents and available within udev.

  # cat /sys/class/net/wlan0/uevent
  DEVTYPE=wlan
  INTERFACE=wlan0
  IFINDEX=5

The problem with your idea is that there is only a nebulous definition of what is a Wifi device, and what is a WiMAX device etc. What userspace should be looking for is "does device XXX support API yyy?" and if it supports API
yyy then I it can be configured that way.

There already is some information in sysfs like /sys/class/net/XXX/type which gives the hardware type (ARPHRD_ETHER, etc). Why not a better version of something like this that provides "can do FOO" interface?

Doing a several system calls (open/read/close) per device is not a big issue. Even an android phone can do open/read/close in less than a millisecond I bet.

author maintainer
#10 lore raw

Hi Stephen,

Quoted 19 lines
Quoted 9 lines

The Ethernet framing is used for a lot of devices these days. Most prominent are WiFi and WiMAX based devices. However for userspace application it is important to classify these devices correctly and not only see them as Ethernet devices. The daemons like HAL, DeviceKit or even NetworkManager with udev support tries to do the classification in userspace with a lot trickery and extra system calls. This is not good and actually reaches its limitations. Especially since the kernel does know the type of the Ethernet device it is pretty stupid.

Quoted 9 lines

To solve this problem the underlying device type needs to be set and then the value will be exported as DEVTYPE via uevents and available within udev.

  # cat /sys/class/net/wlan0/uevent
  DEVTYPE=wlan
  INTERFACE=wlan0
  IFINDEX=5
Quoted 4 lines

The problem with your idea is that there is only a nebulous definition of what is a Wifi device, and what is a WiMAX device etc. What userspace should be looking for is "does device XXX support API yyy?" and if it supports API
yyy then I it can be configured that way.

We don't need that. We just need to know what technology is behind that Ethernet device. Otherwise we have no real starting point.

Quoted 3 lines

There already is some information in sysfs like /sys/class/net/XXX/type which gives the hardware type (ARPHRD_ETHER, etc). Why not a better version of something like this that provides "can do FOO" interface?

that was my initial idea, but I couldn't come up with something good. And the DEVTYPE from struct device is a standard way of exposing such information. Actually subsystem like SCSI, USB, Bluetooth etc. use that already. So why not do the same for Ethernet based devices.

Quoted 3 lines

Doing a several system calls (open/read/close) per device is not a big issue. Even an android phone can do open/read/close in less than a millisecond I bet.

It is not only the system calls that userspace has to do. There is no central place for this stuff and why should be do it all if the kernel already has all the details.

Regards

Marcel

maintainer Applied
#11 lore raw
In reply to Marcel Holtmann

From: Marcel Holtmann <[email protected]>
Date: Tue, 1 Sep 2009 00:08:19 -0700

Quoted 18 lines

The Ethernet framing is used for a lot of devices these days. Most prominent are WiFi and WiMAX based devices. However for userspace application it is important to classify these devices correctly and not only see them as Ethernet devices. The daemons like HAL, DeviceKit or even NetworkManager with udev support tries to do the classification in userspace with a lot trickery and extra system calls. This is not good and actually reaches its limitations. Especially since the kernel does know the type of the Ethernet device it is pretty stupid.

To solve this problem the underlying device type needs to be set and then the value will be exported as DEVTYPE via uevents and available within udev.

  # cat /sys/class/net/wlan0/uevent
  DEVTYPE=wlan
  INTERFACE=wlan0
  IFINDEX=5
Quoted 8 lines

This is similar to subsystems like USB and SCSI that distinguish between hosts, devices, disks, partitions etc.

The new SET_NETDEV_DEVTYPE() is a convenience helper to set the actual device type. All device types are free form, but for convenience the same strings as used with RFKILL are choosen.

Applied to net-next-2.6, thanks.