Hi all,
Various hotplug packages have had trouble dealing with network
interface being renamed. I've decided to tackle this issue from two
angles :
o export ifindex to those apps, as ifindex is persistent.
o expose interface renaming as a hotplug event.
Those two changes are complementary, even an application that
would track interface by ifindex would sometime needs to know about
ifname change. My assumption is that most apps would still use ifname
for a long while to be backward compatible with older kernels.
The kobject framework is well designed, so adding these
features is trivial change and won't run the risk of breaking anything
(famous last words). Obviously, hotplug apps are free to ignore those
additional features.
Patch for 2.6.20 is attached. The patch was tested on a system
running the hotplug scripts, and on another system running udev.
Have fun...
Jean
Signed-off-by: Jean Tourrilhes <[email protected]>
---------------------------------------------------------
diff -u -p linux/include/linux/kobject.j1.h linux/include/linux/kobject.h
--- linux/include/linux/kobject.j1.h 2007-02-26 18:37:55.000000000 -0800
+++ linux/include/linux/kobject.h 2007-02-26 18:38:42.000000000 -0800
@@ -48,6 +48,7 @@ enum kobject_action {
KOBJ_OFFLINE = (__force kobject_action_t) 0x06, /* device offline */
KOBJ_ONLINE = (__force kobject_action_t) 0x07, /* device online */
KOBJ_MOVE = (__force kobject_action_t) 0x08, /* device move */
+ KOBJ_RENAME = (__force kobject_action_t) 0x09, /* device renamed */
};
struct kobject {
diff -u -p linux/net/core/net-sysfs.j1.c linux/net/core/net-sysfs.c
--- linux/net/core/net-sysfs.j1.c 2007-02-27 15:01:08.000000000 -0800
+++ linux/net/core/net-sysfs.c 2007-02-27 15:06:49.000000000 -0800
@@ -412,6 +412,17 @@ static int netdev_uevent(struct class_de
if ((size <= 0) || (i >= num_envp))
return -ENOMEM;
+ /* pass ifindex to uevent.
+ * ifindex is useful as it won't change (interface name may change)
+ * and is what RtNetlink uses natively. */
+ envp[i++] = buf;
+ n = snprintf(buf, size, "IFINDEX=%d", dev->ifindex) + 1;
+ buf += n;
+ size -= n;
+
+ if ((size <= 0) || (i >= num_envp))
+ return -ENOMEM;
+
envp[i] = NULL;
return 0;
}
diff -u -p linux/lib/kobject_uevent.j1.c linux/lib/kobject_uevent.c
--- linux/lib/kobject_uevent.j1.c 2007-02-26 18:38:02.000000000 -0800
+++ linux/lib/kobject_uevent.c 2007-02-26 18:38:42.000000000 -0800
@@ -52,6 +52,8 @@ static char *action_to_string(enum kobje
return "online";
case KOBJ_MOVE:
return "move";
+ case KOBJ_RENAME:
+ return "rename";
default:
return NULL;
}
diff -u -p linux/drivers/base/class.j1.c linux/drivers/base/class.c
--- linux/drivers/base/class.j1.c 2007-02-26 18:38:10.000000000 -0800
+++ linux/drivers/base/class.c 2007-02-27 15:52:37.000000000 -0800
@@ -841,6 +841,8 @@ int class_device_rename(struct class_dev
{
int error = 0;
char *old_class_name = NULL, *new_class_name = NULL;
+ char *devname_string = NULL;
+ char *envp[2];
class_dev = class_device_get(class_dev);
if (!class_dev)
@@ -849,6 +851,15 @@ int class_device_rename(struct class_dev
pr_debug("CLASS: renaming '%s' to '%s'\n", class_dev->class_id,
new_name);
+ devname_string = kmalloc(strlen(class_dev->class_id) + 15, GFP_KERNEL);
+ if (!devname_string) {
+ class_device_put(class_dev);
+ return -ENOMEM;
+ }
+ sprintf(devname_string, "INTERFACE_OLD=%s", class_dev->class_id);
+ envp[0] = devname_string;
+ envp[1] = NULL;
+
#ifdef CONFIG_SYSFS_DEPRECATED
if (class_dev->dev)
old_class_name = make_class_name(class_dev->class->name,
@@ -868,8 +879,16 @@ int class_device_rename(struct class_dev
sysfs_remove_link(&class_dev->dev->kobj, old_class_name);
}
#endif
+
+ /* This function is only used for network interface.
+ * Some hotplug package track interfaces by their name and
+ * therefore want to know when the name is changed by the user. */
+ if(!error)
+ kobject_uevent_env(&class_dev->kobj, KOBJ_RENAME, envp);
+
class_device_put(class_dev);
+ kfree(devname_string);
kfree(old_class_name);
kfree(new_class_name);
Hi,
> Patch for 2.6.20 is attached.
... and in the meantime netdevices aren't class_device any more :) IOW,
your patch isn't going to work any more. Also, I think wireless could
benefit from this as well.
> The kobject framework is well designed, so adding these
> features is trivial change and won't run the risk of breaking anything
> (famous last words). Obviously, hotplug apps are free to ignore those
> additional features.
Why not just add this to base kobject_rename instead? That way,
userspace is notified for all renames in sysfs.
The patch then collapses down to the change in net's sysfs code to add
the ifindex to the environment, and another change in kobject to invoke
a new event when a name changes and show the old name.
johannes
On 28-02-2007 02:27, Jean Tourrilhes wrote:
> Hi all,
...
> Patch for 2.6.20 is attached. The patch was tested on a system
> running the hotplug scripts, and on another system running udev.
>
> Have fun...
>
> Jean
>
> Signed-off-by: Jean Tourrilhes <[email protected]>
>
> ---------------------------------------------------------
...
> diff -u -p linux/net/core/net-sysfs.j1.c linux/net/core/net-sysfs.c
> --- linux/net/core/net-sysfs.j1.c 2007-02-27 15:01:08.000000000 -0800
> +++ linux/net/core/net-sysfs.c 2007-02-27 15:06:49.000000000 -0800
> @@ -412,6 +412,17 @@ static int netdev_uevent(struct class_de
> if ((size <= 0) || (i >= num_envp))
> return -ENOMEM;
>
> + /* pass ifindex to uevent.
> + * ifindex is useful as it won't change (interface name may change)
> + * and is what RtNetlink uses natively. */
> + envp[i++] = buf;
> + n = snprintf(buf, size, "IFINDEX=%d", dev->ifindex) + 1;
> + buf += n;
> + size -= n;
> +
> + if ((size <= 0) || (i >= num_envp))
Btw.:
1. if size == 10 and snprintf returns 9 (without NULL)
then n == 10 (with NULL), so isn't it enough (here and above):
if ((size < 0) || (i >= num_envp))
2. shouldn't there be (here and above):
envp[--i] = NULL;
> + return -ENOMEM;
> +
> envp[i] = NULL;
> return 0;
> }
...
> diff -u -p linux/drivers/base/class.j1.c linux/drivers/base/class.c
> --- linux/drivers/base/class.j1.c 2007-02-26 18:38:10.000000000 -0800
> +++ linux/drivers/base/class.c 2007-02-27 15:52:37.000000000 -0800
> @@ -841,6 +841,8 @@ int class_device_rename(struct class_dev
> {
> int error = 0;
> char *old_class_name = NULL, *new_class_name = NULL;
> + char *devname_string = NULL;
> + char *envp[2];
>
> class_dev = class_device_get(class_dev);
> if (!class_dev)
> @@ -849,6 +851,15 @@ int class_device_rename(struct class_dev
> pr_debug("CLASS: renaming '%s' to '%s'\n", class_dev->class_id,
> new_name);
>
> + devname_string = kmalloc(strlen(class_dev->class_id) + 15, GFP_KERNEL);
> + if (!devname_string) {
> + class_device_put(class_dev);
> + return -ENOMEM;
> + }
> + sprintf(devname_string, "INTERFACE_OLD=%s", class_dev->class_id);
> + envp[0] = devname_string;
> + envp[1] = NULL;
> +
> #ifdef CONFIG_SYSFS_DEPRECATED
> if (class_dev->dev)
> old_class_name = make_class_name(class_dev->class->name,
> @@ -868,8 +879,16 @@ int class_device_rename(struct class_dev
> sysfs_remove_link(&class_dev->dev->kobj, old_class_name);
> }
> #endif
> +
> + /* This function is only used for network interface.
> + * Some hotplug package track interfaces by their name and
> + * therefore want to know when the name is changed by the user. */
> + if(!error)
> + kobject_uevent_env(&class_dev->kobj, KOBJ_RENAME, envp);
> +
> class_device_put(class_dev);
>
> + kfree(devname_string);
Maybe I miss something, but it seems kobject_uevent_env copies
pointers from envp instead of buffers' contents.
Regards,
Jarek P.
On Wed, Feb 28, 2007 at 10:34:37AM +0100, Jarek Poplawski wrote:
> On 28-02-2007 02:27, Jean Tourrilhes wrote:
...
> > + /* This function is only used for network interface.
> > + * Some hotplug package track interfaces by their name and
> > + * therefore want to know when the name is changed by the user. */
> > + if(!error)
> > + kobject_uevent_env(&class_dev->kobj, KOBJ_RENAME, envp);
> > +
> > class_device_put(class_dev);
> >
> > + kfree(devname_string);
>
> Maybe I miss something, but it seems kobject_uevent_env copies
> pointers from envp instead of buffers' contents.
And it's enough - sorry.
Jarek P.
On Tue, Feb 27, 2007 at 05:27:41PM -0800, Jean Tourrilhes wrote:
> diff -u -p linux/drivers/base/class.j1.c linux/drivers/base/class.c
> --- linux/drivers/base/class.j1.c 2007-02-26 18:38:10.000000000 -0800
> +++ linux/drivers/base/class.c 2007-02-27 15:52:37.000000000 -0800
> @@ -841,6 +841,8 @@ int class_device_rename(struct class_dev
This function is not in the 2.6.21-rc2 kernel, so you might want to
rework this patch a bit :)
Also, it's userspace that causes the rename to happen, so it knows it
did it, why should the kernel have to emit a message to tell userspace
again what just happened?
thanks,
greg k-h
On Wed, Feb 28, 2007 at 07:36:17AM -0800, Greg KH wrote:
> On Tue, Feb 27, 2007 at 05:27:41PM -0800, Jean Tourrilhes wrote:
> > diff -u -p linux/drivers/base/class.j1.c linux/drivers/base/class.c
> > --- linux/drivers/base/class.j1.c 2007-02-26 18:38:10.000000000 -0800
> > +++ linux/drivers/base/class.c 2007-02-27 15:52:37.000000000 -0800
> > @@ -841,6 +841,8 @@ int class_device_rename(struct class_dev
>
> This function is not in the 2.6.21-rc2 kernel, so you might want to
> rework this patch a bit :)
It was a trial balloon to gather feedback. I will do.
> Also, it's userspace that causes the rename to happen, so it knows it
> did it, why should the kernel have to emit a message to tell userspace
> again what just happened?
Username is not one big program, but a collection of program,
and one program does not know what another program do.
In particular, udev does not know when people are using
iproute2 to rename interface and loose its marbles. We don't really
want to ban iproute2 or udev ;-)
> thanks,
>
> greg k-h
Have fun...
Jean
On Wed, Feb 28, 2007 at 10:34:37AM +0100, Jarek Poplawski wrote:
> On 28-02-2007 02:27, Jean Tourrilhes wrote:
> > Hi all,
> ...
> > Patch for 2.6.20 is attached. The patch was tested on a system
> > running the hotplug scripts, and on another system running udev.
> >
> > Have fun...
> >
> > Jean
> >
> > Signed-off-by: Jean Tourrilhes <[email protected]>
> >
> > ---------------------------------------------------------
> ...
> > diff -u -p linux/net/core/net-sysfs.j1.c linux/net/core/net-sysfs.c
> > --- linux/net/core/net-sysfs.j1.c 2007-02-27 15:01:08.000000000 -0800
> > +++ linux/net/core/net-sysfs.c 2007-02-27 15:06:49.000000000 -0800
> > @@ -412,6 +412,17 @@ static int netdev_uevent(struct class_de
> > if ((size <= 0) || (i >= num_envp))
> > return -ENOMEM;
> >
> > + /* pass ifindex to uevent.
> > + * ifindex is useful as it won't change (interface name may change)
> > + * and is what RtNetlink uses natively. */
> > + envp[i++] = buf;
> > + n = snprintf(buf, size, "IFINDEX=%d", dev->ifindex) + 1;
> > + buf += n;
> > + size -= n;
> > +
> > + if ((size <= 0) || (i >= num_envp))
>
> Btw.:
> 1. if size == 10 and snprintf returns 9 (without NULL)
> then n == 10 (with NULL), so isn't it enough (here and above):
>
> if ((size < 0) || (i >= num_envp))
I just cut'n'pasted the code a few line above. If the original
code is incorrect, it need fixing. And it will need fixing in probably
a lot of places.
> 2. shouldn't there be (here and above):
>
> envp[--i] = NULL;
>
No, envp is local, so who cares.
Thanks.
Jean
On Wed, Feb 28, 2007 at 10:16:05AM +0100, Johannes Berg wrote:
> Hi,
>
> > Patch for 2.6.20 is attached.
>
> ... and in the meantime netdevices aren't class_device any more :) IOW,
> your patch isn't going to work any more.
That's why I always specify the kernel version. I'll look into
that, I'm sure it's not the end of the world ;-)
> Also, I think wireless could benefit from this as well.
In which sense ? Wireless interface are regular netdevices.
> > The kobject framework is well designed, so adding these
> > features is trivial change and won't run the risk of breaking anything
> > (famous last words). Obviously, hotplug apps are free to ignore those
> > additional features.
>
> Why not just add this to base kobject_rename instead? That way,
> userspace is notified for all renames in sysfs.
> The patch then collapses down to the change in net's sysfs code to add
> the ifindex to the environment, and another change in kobject to invoke
> a new event when a name changes and show the old name.
I'm just trying to follow the established pattern. Both
class_device_add() and class_device_del() are generating the
event. Also, I'm not sure if other subsystem would benefit from it, I
don't want to generate too many useless events.
> johannes
Thanks !
Jean
On Wed, 2007-02-28 at 10:51 -0800, Jean Tourrilhes wrote:
> That's why I always specify the kernel version. I'll look into
> that, I'm sure it's not the end of the world ;-)
Sure, just wanted to point it out.
> In which sense ? Wireless interface are regular netdevices.
Yeah but in mac80211 we have the wiphy concept since multiple virtual
interfaces can be associated to one hardware, and that is where QoS is
done, not the netdevs. Of course, those interested can just listen to
nl80211 events to figure out if someone renamed a 802.11 phy, but things
like hal would probably not want to and still know about the name
change.
> I'm just trying to follow the established pattern. Both
> class_device_add() and class_device_del() are generating the
> event. Also, I'm not sure if other subsystem would benefit from it, I
> don't want to generate too many useless events.
I don't think many other subsystems (can) rename things ;)
johannes
On Wed, Feb 28, 2007 at 07:36:17AM -0800, Greg KH wrote:
> On Tue, Feb 27, 2007 at 05:27:41PM -0800, Jean Tourrilhes wrote:
> > diff -u -p linux/drivers/base/class.j1.c linux/drivers/base/class.c
> > --- linux/drivers/base/class.j1.c 2007-02-26 18:38:10.000000000 -0800
> > +++ linux/drivers/base/class.c 2007-02-27 15:52:37.000000000 -0800
> > @@ -841,6 +841,8 @@ int class_device_rename(struct class_dev
>
> This function is not in the 2.6.21-rc2 kernel, so you might want to
> rework this patch a bit :)
Thanks for all you good comments. I ported my patch to
2.6.21-rc2, and tested it both on a hotplug and a udev system. Patch
is attached, I would be glad if you could push that through the usual
channels.
Also, I realised that I forgot to say in my original e-mail
that migrating udev to use ifindex instead of ifname would fix the
remove/add race condition for network devices. But that's not going to
happen overnight...
Have fun...
Jean
Signed-off-by: Jean Tourrilhes <[email protected]>
---------------------------------------------------------
diff -u -p linux/include/linux/kobject.j1.h linux/include/linux/kobject.h
--- linux/include/linux/kobject.j1.h 2007-02-28 14:26:29.000000000 -0800
+++ linux/include/linux/kobject.h 2007-02-28 14:27:54.000000000 -0800
@@ -48,6 +48,7 @@ enum kobject_action {
KOBJ_OFFLINE = (__force kobject_action_t) 0x06, /* device offline */
KOBJ_ONLINE = (__force kobject_action_t) 0x07, /* device online */
KOBJ_MOVE = (__force kobject_action_t) 0x08, /* device move */
+ KOBJ_RENAME = (__force kobject_action_t) 0x09, /* device renamed */
};
struct kobject {
diff -u -p linux/net/core/net-sysfs.j1.c linux/net/core/net-sysfs.c
--- linux/net/core/net-sysfs.j1.c 2007-02-28 14:26:45.000000000 -0800
+++ linux/net/core/net-sysfs.c 2007-02-28 14:27:54.000000000 -0800
@@ -424,6 +424,17 @@ static int netdev_uevent(struct device *
if ((size <= 0) || (i >= num_envp))
return -ENOMEM;
+ /* pass ifindex to uevent.
+ * ifindex is useful as it won't change (interface name may change)
+ * and is what RtNetlink uses natively. */
+ envp[i++] = buf;
+ n = snprintf(buf, size, "IFINDEX=%d", dev->ifindex) + 1;
+ buf += n;
+ size -= n;
+
+ if ((size <= 0) || (i >= num_envp))
+ return -ENOMEM;
+
envp[i] = NULL;
return 0;
}
diff -u -p linux/lib/kobject_uevent.j1.c linux/lib/kobject_uevent.c
--- linux/lib/kobject_uevent.j1.c 2007-02-28 14:26:58.000000000 -0800
+++ linux/lib/kobject_uevent.c 2007-02-28 14:27:54.000000000 -0800
@@ -52,6 +52,8 @@ static char *action_to_string(enum kobje
return "online";
case KOBJ_MOVE:
return "move";
+ case KOBJ_RENAME:
+ return "rename";
default:
return NULL;
}
diff -u -p linux/drivers/base/core.j1.c linux/drivers/base/core.c
--- linux/drivers/base/core.j1.c 2007-02-28 15:45:45.000000000 -0800
+++ linux/drivers/base/core.c 2007-02-28 15:47:30.000000000 -0800
@@ -1007,6 +1007,8 @@ int device_rename(struct device *dev, ch
char *new_class_name = NULL;
char *old_symlink_name = NULL;
int error;
+ char *devname_string = NULL;
+ char *envp[2];
dev = get_device(dev);
if (!dev)
@@ -1014,6 +1016,15 @@ int device_rename(struct device *dev, ch
pr_debug("DEVICE: renaming '%s' to '%s'\n", dev->bus_id, new_name);
+ devname_string = kmalloc(strlen(dev->bus_id) + 15, GFP_KERNEL);
+ if (!devname_string) {
+ put_device(dev);
+ return -ENOMEM;
+ }
+ sprintf(devname_string, "INTERFACE_OLD=%s", dev->bus_id);
+ envp[0] = devname_string;
+ envp[1] = NULL;
+
#ifdef CONFIG_SYSFS_DEPRECATED
if ((dev->class) && (dev->parent))
old_class_name = make_class_name(dev->class->name, &dev->kobj);
@@ -1049,12 +1060,20 @@ int device_rename(struct device *dev, ch
sysfs_create_link(&dev->class->subsys.kset.kobj, &dev->kobj,
dev->bus_id);
}
+
+ /* This function is only used for network interface.
+ * Some hotplug package track interfaces by their name and
+ * therefore want to know when the name is changed by the user. */
+ if(!error)
+ kobject_uevent_env(&dev->kobj, KOBJ_RENAME, envp);
+
put_device(dev);
kfree(new_class_name);
kfree(old_symlink_name);
out_free_old_class:
kfree(old_class_name);
+ kfree(devname_string);
return error;
}
On Wed, 2007-02-28 at 16:26 -0800, Jean Tourrilhes wrote:
> + /* This function is only used for network interface.
> + * Some hotplug package track interfaces by their name and
> + * therefore want to know when the name is changed by the user. */
Right now, that's true, but wireless is going to start using
device_rename pretty soon as well. Could you rephrase this comment?
johannes
On Thu, Mar 01, 2007 at 01:37:46AM +0100, Johannes Berg wrote:
> On Wed, 2007-02-28 at 16:26 -0800, Jean Tourrilhes wrote:
>
> > + /* This function is only used for network interface.
> > + * Some hotplug package track interfaces by their name and
> > + * therefore want to know when the name is changed by the user. */
>
> Right now, that's true, but wireless is going to start using
> device_rename pretty soon as well. Could you rephrase this comment?
>
> johannes
I would prefer to fix the comment when this change actually
happens. I prefer comments to refer to the current reality, rather
than past/future situation. When you introduce wireless renaming, you
will need to verify the whole chain anyway, so you might as well fix
the comment while merging wireless renaming.
Note also that my comment is technically correct. I did not
say 'netdev' but the more generic term 'network interface', and I
believe your wireless interface is a 'network interface', even if it's
not a netdev ;-)
But if this really bugs you, please feel free to respin my
patch.
Have fun...
Jean
On Wed, 2007-02-28 at 16:51 -0800, Jean Tourrilhes wrote:
> I would prefer to fix the comment when this change actually
> happens. I prefer comments to refer to the current reality, rather
> than past/future situation.
Uh, no. device_rename is perfectly fine, even other people may use it in
the future.
> When you introduce wireless renaming, you
> will need to verify the whole chain anyway, so you might as well fix
> the comment while merging wireless renaming.
No again, device_rename is perfectly fine API, I shouldn't have to look
at it's internals to see if it's broken in my use case. Even if it's
only a broken comment.
I'm not going to respin your patches though, if this doesn't make it in
I don't care.
johannes