Bug 855107

Created attachment 610441 [details]
eth0 config

Forgot to mention hardware specs:

System Information
        Manufacturer: HP
        Product Name: ProLiant BL460c G1
BIOS Information
        Vendor: HP
        Version: I15
        Release Date: 10/25/2010

[root@localhost ~]# ethtool -i eth0
driver: bnx2
version: 2.1.11
firmware-version: bc 4.4.1
bus-info: 0000:03:00.0
[root@localhost ~]# modinfo bnx2
filename:       /lib/modules/2.6.32-220.el6.x86_64/kernel/drivers/net/bnx2.ko
firmware:       bnx2/bnx2-rv2p-09ax-6.0.17.fw
firmware:       bnx2/bnx2-rv2p-09-6.0.17.fw
firmware:       bnx2/bnx2-mips-09-6.2.1a.fw
firmware:       bnx2/bnx2-rv2p-06-6.0.15.fw
firmware:       bnx2/bnx2-mips-06-6.2.1.fw
version:        2.1.11
license:        GPL
description:    Broadcom NetXtreme II BCM5706/5708/5709/5716 Driver
author:         Michael Chan <mchan>
srcversion:     61BD2699C6587068253C2BB
alias:          pci:v000014E4d0000163Csv*sd*bc*sc*i*
alias:          pci:v000014E4d0000163Bsv*sd*bc*sc*i*
alias:          pci:v000014E4d0000163Asv*sd*bc*sc*i*
alias:          pci:v000014E4d00001639sv*sd*bc*sc*i*
alias:          pci:v000014E4d000016ACsv*sd*bc*sc*i*
alias:          pci:v000014E4d000016AAsv*sd*bc*sc*i*
alias:          pci:v000014E4d000016AAsv0000103Csd00003102bc*sc*i*
alias:          pci:v000014E4d0000164Csv*sd*bc*sc*i*
alias:          pci:v000014E4d0000164Asv*sd*bc*sc*i*
alias:          pci:v000014E4d0000164Asv0000103Csd00003106bc*sc*i*
alias:          pci:v000014E4d0000164Asv0000103Csd00003101bc*sc*i*
depends:
vermagic:       2.6.32-220.el6.x86_64 SMP mod_unload modversions
parm:           disable_msi:Disable Message Signaled Interrupt (MSI) (int)

Not sure if it's important but I'm including it just in case. If you have more questions don't hesitate to ask.

Thanks for the report, I was able to reproduce this.
The main issue here is that if we just call ifup eth.280, ifup is started twice
ifup.280 -> kernel event -> udev reaction -> net.hotplug -> ifup.280 (which is definitely bad behavior) and same thing happens with ifdown.

I don't think that some locking would help, so we have two options
1) Ignore hotplug's calls of ifup and ifdown for vlans (but I am not sure if this will not break something)
2) Reassign this to kernel or maybe udev and they might be able solve this better on their level.

I've tried option 1) and it did not work:

# Ethernet 802.1Q VLAN support
-if [ "${VLAN}" = "yes" ] && [ "$ISALIAS" = "no" ]; then
+if [ "${VLAN}" = "yes" ] && [ "$ISALIAS" = "no" ] && [ -z "$IN_HOTPLUG" ]; then

not only the vlan ended in down state, I still could see one unnecessary kernel/udev events but they do not loop forever. Also as you've said it may actually break even more things...

[root@localhost ~]# ifdown eth0.280; ifup eth0.280
KERNEL[1347284812.857702] remove   /devices/virtual/net/eth0.280/queues/rx-0 (queues)
KERNEL[1347284812.857771] remove   /devices/virtual/net/eth0.280/queues/tx-0 (queues)
KERNEL[1347284812.857884] remove   /devices/virtual/net/eth0.280 (net)
UDEV  [1347284812.857980] remove   /devices/virtual/net/eth0.280/queues/rx-0 (queues)
UDEV  [1347284812.858008] remove   /devices/virtual/net/eth0.280/queues/tx-0 (queues)
KERNEL[1347284812.885215] add      /devices/virtual/net/eth0.280 (net)
KERNEL[1347284812.885254] add      /devices/virtual/net/eth0.280/queues/rx-0 (queues)
KERNEL[1347284812.885270] add      /devices/virtual/net/eth0.280/queues/tx-0 (queues)
KERNEL[1347284813.000878] remove   /devices/virtual/net/eth0.280/queues/rx-0 (queues)
KERNEL[1347284813.000930] remove   /devices/virtual/net/eth0.280/queues/tx-0 (queues)
KERNEL[1347284813.001009] remove   /devices/virtual/net/eth0.280 (net)
UDEV  [1347284814.016270] remove   /devices/virtual/net/eth0.280 (net)
UDEV  [1347284814.100786] add      /devices/virtual/net/eth0.280 (net)
UDEV  [1347284814.100987] add      /devices/virtual/net/eth0.280/queues/rx-0 (queues)
UDEV  [1347284814.101012] add      /devices/virtual/net/eth0.280/queues/tx-0 (queues)
UDEV  [1347284814.101143] remove   /devices/virtual/net/eth0.280/queues/tx-0 (queues)
UDEV  [1347284814.101160] remove   /devices/virtual/net/eth0.280/queues/rx-0 (queues)
UDEV  [1347284814.202761] remove   /devices/virtual/net/eth0.280 (net)


Having said that I'm fine with reassigning to kernel/udev if you think they are the ones that should be fixing it. However I personally think that initscripts are still responsible. Kernel/udev may have limited ways of knowing if the ifup/ifdown was the trigger for the event they received. In the end it's your call.

Created attachment 663210 [details]
patch proposal

Wait for udev to process all current events before exiting. This fixes race condition we've been having with vlan interfaces. All comments are appreciated.

Comment on attachment 663210 [details]
patch proposal

swaped filenames in diff

Created attachment 663211 [details]
patch

Hi, any update? Did anyone had time to look at the proposed patch?

This, patch looks quite sane. We will consider to include it in next release.

Great. Thanks.

udevadm settle --timeout=5

so, timeout=5 hardcoded? I don't think, that is a good idea.

Fair enough. I don't like that hardcoded value either but could not come up with better solution. What do you suggest?

And I also think, that ifup/ifdown should somehow file lock (see flock(1) for shell).

Concurrently operating on routing tables, interface settings, etc. does not seem to be a way to get consistent settings.

I can confirm this is happening on Red Hat Enterprise Linux Server release 6.4 (Santiago), x84_64 arch.

Sleeping for 1 second between ifdown and ifup works around the issue.

*** Bug 952538 has been marked as a duplicate of this bug. ***

According to Lukas, he thinks this BZ has been already fixed:
https://github.com/fedora-sysv/initscripts/commit/0c78d0c

The locking mechanism for ifup/ifdown is nice to have feature, but it would still not work correctly if someone would call different networking subscripts manually, or from some other non-RHEL scripts.

Therefore, I'm closing this BZ as WONTFIX. In case anyone still faces this issue, please use the workaround:
> HOTPLUG=no
as mentioned in comment #0.

Best regards,

David

*** Bug 1398326 has been marked as a duplicate of this bug. ***