From: Kristian Evensen <kristian.evensen@gmail.com>
Date: Wed, 11 Mar 2020 13:13:10 +0000 (+0100)
Subject: system-linux: improve handling of device rename
X-Git-Url: https://git.librecmc.org/?a=commitdiff_plain;p=oweals%2Fnetifd.git

system-linux: improve handling of device rename

After an interface has been renamed on a "fast" device (for example
x86_64), the interface is sometimes not handled correctly by netifd.
Looking in the logs, I see the following messages when renaming fails:

Wed Mar 11 08:52:44 2020 kern.info kernel: [68383.522038] igb 0000:03:00.0 nlw_1: renamed from eth2
Wed Mar 11 08:52:44 2020 daemon.err netifd[2739]: __device_add_user(710): Add user for device 'nlw_1', refcount=2
Wed Mar 11 08:52:44 2020 daemon.err netifd[2739]: device_claim(413): Claim Network device nlw_1, new active count: 2
Wed Mar 11 08:52:44 2020 daemon.err netifd[2739]: device_claim(432): claim Network device nlw_1 failed: -1

Instrumenting netifd further reveals that there is a race between the hotplug
"@move" event and ioctl(SIOCGIFINDEX). When the above error happens, the
ioctl-call fails with ENODEV. Looking closer at the kernel code, it seems the
hotplug-event is triggered before the renaming is completed. The easiest way to
trigger the race, is if an interface name with the old name is not handled by
netifd and an interface with the new name is. If only the old name is handled,
or both names, I was not able to provoke the race.

When the renaming is complete, a NEWLINK-message is generated. This patch
modifies the logic surrounding renaming, so that we wait for the
NEWLINK-message before marking an interface as present. The changes made are:

* We only handle move-events for interfaces we know, and we return after
device has been set as not present.
* When we receive a NEWLINK message for an interface managed by netifd,
we call device_set_present. device_set_present is guarded by the same
checks as the add hotplug-event.

After these changes, renaming works properly on both "fast" and "slow"
devices. Removing a device is also handled correctly.

Signed-off-by: Kristian Evensen <kristian.evensen@gmail.com>
Signed-off-by: Hans Dedecker <dedeckeh@gmail.com>
---

diff --git a/system-linux.c b/system-linux.c
index c225175..3b09bbb 100644
--- a/system-linux.c
+++ b/system-linux.c
@@ -592,6 +592,9 @@ static int cb_rtnl_event(struct nl_msg *msg, void *arg)
 	if (!system_get_dev_sysctl("/sys/class/net/%s/carrier", dev->ifname, buf, sizeof(buf)))
 		link_state = strtoul(buf, NULL, 0);
 
+	if (dev->type == &simple_device_type && !system_if_force_external(dev->ifname))
+		device_set_present(dev, true);
+
 	device_set_link(dev, link_state ? true : false);
 
 out:
@@ -654,13 +657,15 @@ handle_hotplug_msg(char *data, int size)
 move:
 	dev = device_find(interface_old);
 	if (!dev)
-		goto found;
+		return;
 
 	if (dev->type != &simple_device_type)
 		goto found;
 
 	device_set_present(dev, false);
 
+	return;
+
 found:
 	dev = device_find(interface);
 	if (!dev)