| 1 |
On Sat, May 17, 2014 at 2:39 PM, Grant <emailgrant@×××××.com> wrote: |
| 2 |
>>> I have this: |
| 3 |
>>> |
| 4 |
>>> # dmesg | grep enp |
| 5 |
>>> [ 4.297862] systemd-udevd[659]: renamed network interface eth0 to enp0s20u2u1 |
| 6 |
>>> [ 4.778289] systemd-udevd[660]: renamed network interface eth0 to enp0s20u2u2 |
| 7 |
>>> [ 6.496193] ax88179_178a 3-2.1:1.0 enp0s20u2u1: ax88179 - Link status is: 1 |
| 8 |
>>> [ 7.905393] ax88179_178a 3-2.2:1.0 enp0s20u2u2: ax88179 - Link status is: 1 |
| 9 |
>>> # |
| 10 |
>>> |
| 11 |
>>> That doesn't tell us when the network initscripts tried and failed to |
| 12 |
>>> start but this from /var/log/messages/everything/current shows the |
| 13 |
>>> first time in the boot sequence that a dependent service failed to |
| 14 |
>>> start because of the networking failure so it should be before this: |
| 15 |
>>> |
| 16 |
>>> [kernel] [ 0.787433] serio: i8042 AUX port at 0x60,0x64 irq 12 |
| 17 |
>>> [/etc/init.d/unbound] ERROR: cannot start unbound as net.enp0s20u2u1 |
| 18 |
>>> would not start |
| 19 |
>>> [kernel] [ 0.792081] rtc_cmos 00:04: alarms up to one month, y3k, |
| 20 |
>>> 242 bytes nvram, hpet irqs |
| 21 |
>>> |
| 22 |
>> |
| 23 |
>> Yeah, so I think the kernel is detecting your network card after udev |
| 24 |
>> has already started. |
| 25 |
>> |
| 26 |
>> One interesting experiment would be to delay the boot process to allow |
| 27 |
>> the kernel additional time to detect devices. Adding rootdelay=10 to |
| 28 |
>> your kernel command line should do the trick, unless you are using |
| 29 |
>> some broken initramfs. |
| 30 |
> |
| 31 |
> |
| 32 |
> I tried that and it works great which I think confirms our suspicions |
| 33 |
> that the kernel is detecting my network cards after udev has already |
| 34 |
> started. If I remove rootdelay=10 and I do this: |
| 35 |
> |
| 36 |
> # ln -s /dev/null /etc/udev/rules.d/90-network.rules |
| 37 |
> |
| 38 |
> the network interfaces fail to come up which is the same thing I've |
| 39 |
> experienced with rc_hotplug="net.*". |
| 40 |
> |
| 41 |
|
| 42 |
Yeah, so this is not solvable using service dependencies. You will |
| 43 |
either need to make that boot delay permanent, or rely on the hotplug |
| 44 |
functionality to start the net.en* services. In the latter case, you |
| 45 |
should remove them from the default runlevel. |
| 46 |
|
| 47 |
You may want to define rc_need="!net" to prevent init scripts that |
| 48 |
"need net" from automatically starting the net.* services. For most |
| 49 |
services this is fine, but it will obviously break things like ntpdate |
| 50 |
which actually need a usable network connection. |
Archives