Modprobe as script breaks initramfs kernel?

Modprobe as script breaks initramfs kernel?

[Deepak Saxena]

We attempted an experiment in cleaning up some modprobe messages during 
initramfs bootup when the modules directory is missing by moving modprobe 
to modprobe-bin and replacing modprobe with the following simple shell script:

#!/bin/sh
# Clean up bootup when modules are not present

if [ -e "/lib/modules/'uname -r'/modules.dep" ] ; then
	/sbin/modprobe-bin $*
else
	exit 0
fi

Running this on any architecture causes a non fatal kernel BUG message
out of the do_pipe() code path everytime modprobe is called.

BUG: unable to handle kernel NULL pointer dereference at virtual address 00000014
c016a5e6
*pde = 00000000
Oops: 0000 [#1]
CPU:    0
EIP:    0060:[<c016a5e6>]    Not tainted VLI
Using defaults from ksymoops -t elf32-i386 -a i386
EFLAGS: 00010286   (2.6.18_pro500-pc_target-x86_586 #2) 
eax: 00000000   ebx: bf8f9ae8   ecx: 00000000   edx: c16a00f0
esi: c16f0c80   edi: 00000011   ebp: c16f0bc0   esp: c16e9f40
ds: 007b   es: 007b   ss: 0068   preempt: 00000001
Stack: 00000000 dff64dc0 c13ff580 00000000 ffffffe9 dfe84a40 00000002 c16e8000 
       c169f41c 0808d581 c166b900 c010f5bd c166b934 c169f41c 0808d581 00000000 
       00000004 c16e9fbc bf8f9ae8 c16e9fa8 00000011 c16e8000 c0107737 c16e9fa8 
Call Trace:
 [<c010f5bd>]  [<c0107737>]  [<c01030d7>] <0>Code: c6 c7 44 24 10 e9 ff ff ff 0f 84 1b 02 00 00 e8 91 2e ff ff 85 c0 89 c5 c7 44 24 10 e9 ff ff ff 0f 84 34 02 00 00 a1 d8 75 53 c0 <8b> 40 14 89 04 24 e8 8f ec 00 00 85 c0 89 c3 0f 84 0a 02 00 00 
EIP: [<c016a5e6>]  SS:ESP 0068:c16e9f40
Warning (Oops_read): Code line not seen, dumping what data is available


>>EIP; c016a5e6 <do_pipe+46/2c0>   <=====

>>ebx; bf8f9ae8 <phys_startup_32+bf7f9ae8/c0000000>
>>edx; c16a00f0 <pg0+f220f0/3f880400>
>>esi; c16f0c80 <pg0+f72c80/3f880400>
>>ebp; c16f0bc0 <pg0+f72bc0/3f880400>
>>esp; c16e9f40 <pg0+f6bf40/3f880400>

Trace; c010f5bd <do_page_fault+15d/75a>
Trace; c0107737 <sys_pipe+17/60>
Trace; c01030d7 <syscall_call+7/b>

>>EIP; c016a5e6 <do_pipe+46/2c0>   <=====

I've done a quick skim of the code and don't quite see the issue there.
Can the kernel not handle modprobe as a script that then executes another
process this early in the boot sequence?

Tnx,
~Deepak

ps: Yes, I know this is "fix" to error messages is ugly. :)

-- 
Deepak Saxena - dsaxena@plexity.net - http://www.plexity.net

in the end, they will not say, "those were dark times,"  they will ask
"why were their poets silent?" - Bertold Brecht

Re: Modprobe as script breaks initramfs kernel?

[Michael Tokarev]

Deepak Saxena wrote:
> We attempted an experiment in cleaning up some modprobe messages during 
> initramfs bootup when the modules directory is missing by moving modprobe 
> to modprobe-bin and replacing modprobe with the following simple shell script:
> 
> #!/bin/sh
> # Clean up bootup when modules are not present
> 
> if [ -e "/lib/modules/'uname -r'/modules.dep" ] ; then
> 	/sbin/modprobe-bin $*
> else
> 	exit 0
> fi
> 
[]
> BUG: unable to handle kernel NULL pointer dereference at virtual address 00000014
[]
> Trace; c0107737 <sys_pipe+17/60>
> Trace; c01030d7 <syscall_call+7/b>

This is the same issue I reported much earlier with /sbin/hotplug being a script
in initrfamfs.  The problem is because pipefs isn't initialized yet at the time
the script gets called, and causes a NULL-pointer deref.  Obviously you're using
pipe above.

/mjt

Re: Modprobe as script breaks initramfs kernel?

[Jan Engelhardt]

On Feb 22 2007 14:30, Michael Tokarev wrote:
>Deepak Saxena wrote:
>> We attempted an experiment in cleaning up some modprobe messages during 
>> initramfs bootup when the modules directory is missing by moving modprobe 
>> to modprobe-bin and replacing modprobe with the following simple shell script:
>> 
>> #!/bin/sh
>> # Clean up bootup when modules are not present
>> 
>> if [ -e "/lib/modules/'uname -r'/modules.dep" ] ; then
>> 	/sbin/modprobe-bin $*

The nitpick guide says: `uname -r` and "$@" instead of $*  ;-)

>> else
>> 	exit 0
>> fi
>[]
>> BUG: unable to handle kernel NULL pointer dereference at virtual address 00000014
>[]
>> Trace; c0107737 <sys_pipe+17/60>
>> Trace; c01030d7 <syscall_call+7/b>
>
>This is the same issue I reported much earlier with /sbin/hotplug being a script
>in initrfamfs.  The problem is because pipefs isn't initialized yet at the time
>the script gets called, and causes a NULL-pointer deref.  Obviously you're using
>pipe above.

Btw, has this pipefs issue been adressed (by moving pipefs before initramfs
stage), or something?


Jan
-- 

Re: Modprobe as script breaks initramfs kernel?

[Michael Tokarev]

Jan Engelhardt wrote:
> On Feb 22 2007 14:30, Michael Tokarev wrote:
>> Deepak Saxena wrote:
>>> We attempted an experiment in cleaning up some modprobe messages during 
>>> initramfs bootup when the modules directory is missing by moving modprobe 
>>> to modprobe-bin and replacing modprobe with the following simple shell script:
>>>
>>> #!/bin/sh
>>> # Clean up bootup when modules are not present
>>>
>>> if [ -e "/lib/modules/'uname -r'/modules.dep" ] ; then
>>> 	/sbin/modprobe-bin $*
> 
> The nitpick guide says: `uname -r` and "$@" instead of $*  ;-)

Here, it's possible to workaround the pipefs issue... with another kernel
bugfix.  uname -r is available in /proc/sys/kernel/version, but the problem
is that reading almost any file in /proc with shell, which does read by a
single byte, returns only first byte of of the file being read.  The bug
has been addressed by Andrew recently, but I don't remember if the fix is
in 2.6.20 or 2.6.19.x yet.  I discovered it exactly because of this very
issue - hit pipefs bug and tried to work around it by replacing the above
construct with

   read version < /proc/sys/kernel/version
   ... /lib/modules/$version/modules.dep ...

wich results in $version being set to.. "2" (from "2.6...." :)

But sure, the usage is wrong, albiet ok for some minimal debugging.

>>> BUG: unable to handle kernel NULL pointer dereference at virtual address 00000014
>>> Trace; c0107737 <sys_pipe+17/60>

>> This is the same issue I reported much earlier with /sbin/hotplug being a script
>> in initrfamfs.  The problem is because pipefs isn't initialized yet at the time
>> the script gets called, and causes a NULL-pointer deref.  Obviously you're using
>> pipe above.
> 
> Btw, has this pipefs issue been adressed (by moving pipefs before initramfs
> stage), or something?

Not as I know of.  It has been discussed, but seems to be of low priority.

At least, reading /proc file is fixed... somewhere... :)  So it's now possible
to work around it, at least in some places (because sometimes pipe is really
useful anyway and difficult to substitute with something ;)

/mjt

Re: Modprobe as script breaks initramfs kernel?

[Deepak Saxena]

On Feb 22 2007, at 19:44, Jan Engelhardt was caught saying:
> 
> On Feb 22 2007 14:30, Michael Tokarev wrote:
> >Deepak Saxena wrote:
> >> We attempted an experiment in cleaning up some modprobe messages during 
> >> initramfs bootup when the modules directory is missing by moving modprobe 
> >> to modprobe-bin and replacing modprobe with the following simple shell script:
> >> 
> >> #!/bin/sh
> >> # Clean up bootup when modules are not present
> >> 
> >> if [ -e "/lib/modules/'uname -r'/modules.dep" ] ; then
> >> 	/sbin/modprobe-bin $*
> 
> The nitpick guide says: `uname -r` and "$@" instead of $*  ;-)

Yeah. The quotes were a typo in my email. :)

> >This is the same issue I reported much earlier with /sbin/hotplug being a script
> >in initrfamfs.  The problem is because pipefs isn't initialized yet at the time
> >the script gets called, and causes a NULL-pointer deref.  Obviously you're using
> >pipe above.
> 
> Btw, has this pipefs issue been adressed (by moving pipefs before initramfs
> stage), or something?

If it hasn't, it's now on my todo list.

-- 
Deepak Saxena - dsaxena@plexity.net - http://www.plexity.net

In the end, they will not say, "those were dark times,"  they will ask
"why were their poets silent?" - Bertolt Brecht