Re: WARNING: at fs/sysfs/group.c:138 sysfs_remove_group

[Andrew Morton <akpm@linux-foundation.org>]

On Fri, 17 Apr 2009 11:13:08 +0200
Zdenek Kabelac <zdenek.kabelac@gmail.com> wrote:

> Hi
> 
> 
> When I've checked  2.6.30-rc2 kernel and  I've got among other errors
> with Wifi message below.
> As I've noticed - some people were reporting this problem already earlier:
> 
> http://www.gossamer-threads.com/lists/linux/kernel/1058276?page=last
> 
> But for me - it started with 2.6.30-rc kernels.
> I've T61, 4GB, C2D, X86_64.
> 
> 
> 
> Zdenek
> 
> 
> platform microcode: firmware: requesting intel-ucode/06-0f-0a
> platform microcode: firmware: requesting intel-ucode/06-0f-0a

Lots of people are reporting this.  One person reported that each
attempt takes 60 seconds, thus delaying boot be several minutes.

> Microcode Update Driver: v2.00 <tigran@aivazian.fsnet.co.uk>, Peter Oruba
> ------------[ cut here ]------------
> WARNING: at fs/sysfs/group.c:138 sysfs_remove_group+0x10b/0x110()
> Hardware name: 6464CTO
> sysfs group ffffffffa038d900 not found for kobject 'cpu0'
> Modules linked in: microcode(-) binfmt_misc dm_mirror dm_region_hash
> dm_log dm_mod kvm_intel kvm i915 dr
> m i2c_algo_bit uinput btusb bluetooth arc4 ecb cryptomgr aead
> pcompress thinkpad_acpi snd_hda_codec_analog crypto_blkcipher
> crypto_hash snd_hda_intel cry
> pto_algapi snd_hda_codec iwl3945 snd_seq_oss iwlcore
> snd_seq_midi_event mac80211 snd_seq rfkill backlight i2c_i801 psmouse
> intel_agp snd_seq_device snd_p
> cm_oss rtc_cmos rtc_core sr_mod cdrom i2c_core sdhci_pci sdhci
> mmc_core serio_raw led_class iTCO_wdt iTCO_vendor_support nvram
> cfg80211 snd_mixer_oss e10
> 00e rtc_lib snd_pcm snd_timer snd evdev soundcore button ac battery
> snd_page_alloc uhci_hcd ohci_hcd ehci_hcd usbcore [last unloaded:
> scsi_wait_scan]
> Pid: 1983, comm: rmmod Not tainted 2.6.30-rc2-00076-g3ee8da8 #51
> Call Trace:
>  [<ffffffff802434bf>] warn_slowpath+0xaf/0xf0
>  [<ffffffff8039c3dd>] ? kobject_release+0x5d/0xa0
>  [<ffffffff80333f7b>] sysfs_remove_group+0x10b/0x110
>  [<ffffffffa038b3f8>] mc_sysdev_remove+0x38/0x50 [microcode]
>  [<ffffffff8043c08f>] sysdev_driver_unregister+0x5f/0xa0
>  [<ffffffffa038bee8>] microcode_exit+0x39/0x65 [microcode]
>  [<ffffffff8026e974>] sys_delete_module+0x1b4/0x280
>  [<ffffffff8039ff18>] ? __up_write+0xd8/0x150
>  [<ffffffff8053e35f>] ? trace_hardirqs_on_thunk+0x3a/0x3c
>  [<ffffffff8020c19b>] system_call_fastpath+0x16/0x1b

ugh, where to start?

sysdev_driver_register() calls all the ->add methods and cheerily drops
their return codes on the floor.  Consequently
sysdev_driver_unregister() will call the ->remove method on objects
which were not successfully add()ed.

mc_sysdev_add() will return an error if microcode_init_cpu() failed but
fails to remove its attr_group in that case.  So if
sysdev_driver_register()/sysdev_driver_unregister() get fixed,
microcode will misbehave.

This patch:

: commit af5c820a3169e81af869c113e18ec7588836cd50
: Author:     Rusty Russell <rusty@rustcorp.com.au>
: AuthorDate: Wed Mar 11 16:32:36 2009 +1030
: Commit:     Ingo Molnar <mingo@elte.hu>
: CommitDate: Wed Mar 18 13:50:47 2009 +0100
: 
:     x86: cpumask: use work_on_cpu in arch/x86/kernel/microcode_core.c
: 

attempted to fix things up a bit:

: @@ -379,8 +390,11 @@ static int mc_sysdev_add(struct sys_device *sys_dev)
:         if (err)
:                 return err;
:  
: -       microcode_init_cpu(cpu);
: -       return 0;
: +       err = microcode_init_cpu(cpu);
: +       if (err)
: +               sysfs_remove_group(&sys_dev->kobj, &mc_attr_group);
: +
: +       return err;
:  }
:  
:  static int mc_sysdev_remove(struct sys_device *sys_dev)

But it seems that the sysfs_remove_group() later got removed again.


Sigh, what a pickle.  Rafael, please add this as a post-2.6.29 regression.