Re: kobject_add_internal failed for msi_irqs with -EEXIST

[Veaceslav Falico <vfalico@redhat.com>]

On Fri, Sep 27, 2013 at 09:01:58PM +0200, Zdenek Kabelac wrote:
>Dne 27.9.2013 18:01, Veaceslav Falico napsal(a):
>>On Fri, Sep 27, 2013 at 09:58:28AM -0600, Bjorn Helgaas wrote:
>>>[+cc Veaceslav, linux-pci]
>>>
>>>On Fri, Sep 27, 2013 at 7:34 AM, Zdenek Kabelac <zkabelac@redhat.com> wrote:
>>>>Hi
>>>>
>>>>With recent build of 3.12-rc2 I'm getting this warning report from kernel:
>>>>(hw Lenovo T61, C2D, 4GB Ram)
>>>>(repost since linux-kernel@ rejected my gmail email....)
>>>
>>>This looks related to the MSI/kobject issues Veaceslav is working on.
>>>See
>>>http://lkml.kernel.org/r/1379382464-7920-2-git-send-email-vfalico@redhat.com
>>>and related messages.
>>>
>>>We don't have a resolution yet.  If you have
>>>CONFIG_DEBUG_KOBJECT_RELEASE=y, you could try turning that off.  I
>>>don't know if it would help, and it would only be a temporary
>>>workaround anyway.
>>
>>I've looked at the original post -
>>http://www.mail-archive.com/linux-kernel@vger.kernel.org/msg510798.html
>>hope that's it - and it seems that it's disabling DEBUG_KOBJECT_RELEASE
>>won't help - the warning is about the re-registering, not about freeing it.
>>
>>As a workaround I'd suggest adding some kind of delay between removing and
>>adding the msi - as in - rmmod e1000e; sleep 1; modprobe e1000e; - or
>>something like that, so that there is enough time for the /msqi_irqs/ to go
>>away.
>>
>
>I'm not readding e1000e modules myself - however I've no idea what 
>NetworkManager does. Here are messages prior warning:
>
>NetworkManager[304]: <info> monitoring kernel firmware directory '/lib/firmware'.
>NetworkManager[304]: <info> rfkill1: found WiFi radio killswitch (at /sys/devices/pci0000:00/0000:00:1c.1/0000:03:00.0/ieee80211/phy0/rfkill1) 
>(driver iwl3945)
>NetworkManager[304]: <info> WiFi hardware radio set enabled
>NetworkManager[304]: <info> WiFi enabled by radio killswitch; enabled 
>by state file
>NetworkManager[304]: <info> WWAN enabled by radio killswitch; 
>disabled by state file
>NetworkManager[304]: <info> WiMAX enabled by radio killswitch; 
>enabled by state file
>NetworkManager[304]: <info> Networking is enabled by state file
>NetworkManager[304]: <info> (eth0): carrier is OFF
>NetworkManager[304]: <info> (eth0): new Ethernet device (driver: 
>'e1000e' ifindex: 2)
>NetworkManager[304]: <info> (eth0): exported as 
>/org/freedesktop/NetworkManager/Devices/0
>NetworkManager[304]: <info> (eth0): device state change: unmanaged -> 
>unavailable (reason 'managed') [10 20 2]
>NetworkManager[304]: <info> (eth0): bringing up device.
>rpcbind: cannot create socket for udp6
>rpcbind: cannot create socket for tcp6
>kernel: [    5.727025] WARNING! power/level is deprecated; use 
>power/control instead
>
>
>So it looks like  'bringing up' causes recreation of msi_irqs ?

Yep, seems like that. .ndo_stop() (e1000_close(), which removes msi irqs
and the /msi_irqs/ kset) instantly followed by .ndo_open() (e1000_open(),
which enables the requested msi irqs via pci_enable_msi() - and,
consequently, re-creates the /msi_irqs/ kset) triggers it - because the
kset, in the meanwhile, isn't freed - maybe because of some readers, maybe
because of the delay introduced by DEBUG_KOBJECT_RELEASE.

I'm now slowly fighting this (and adjacent) issues, will put you on CC if
I'll come up with patches.

In the meanwhile - can you please try disabling DEBUG_KOBJECT_RELEASE and
test if you see the warning _constantly_?

Cause if you do - we might have another issue here...

>
>Zdenek
>
>
>
>
>