Re: potential linking bug in recursive directory descent

[Randy Dunlap <randy.dunlap@oracle.com>]

Fred Isaman wrote:
> On Tue, Feb 24, 2009 at 4:03 PM, Sam Ravnborg <sam@ravnborg.org> wrote:
>> On Tue, Feb 24, 2009 at 02:05:25PM -0500, Fred Isaman wrote:
>>> The pnfs tree base of of 2.6.29-rc5 uses the new find_last_bit
>>> function (defined in
>>> lib/find_last_bit.c) in a file in the fs/nfs directory.  Myself and
>>> another developer (though no one else so far) get the following error
>>> during compile:
>>>
>>>    Kernel: arch/x86/boot/bzImage is ready  (#5)
>>>      Building modules, stage 2.
>>>      MODPOST 515 modules
>>>    ERROR: "find_last_bit" [fs/nfs/nfs.ko] undefined!
>>>    make[2]: *** [__modpost] Error 1
>>>    make[1]: *** [modules] Error 2
>>>    make: *** [sub-make] Error 2
>>>
>>>
>>> Note that find_last_bit() is not used in any file in the fs directory.
>>> If I add it to any function that is EXPORT_SYMBOL'ed
>>> from the fs directory, suddenly the compile errors go away.  (See the
>>> below patch for a more concrete example.)
>>>
>>> I am not sure what is going on, and why it only affects some
>>> developers, but it looks a lot like the kbuild system
>>> is deciding that the library does not need to be included at the fs
>>> directory level, so isn't including it in fs/nfs where
>>> it is needed.
>> kbuild only links in libaries if there is any users of siad library.
>> And your normal config has no users of lib/find_last_bit.c so
>> it is not linked into the kernel.
>>
> 
> I don't understand your comment.  There *is* a user of the library in the
> (modified) fs/nfs/nfs4proc.c.  The problem is that it doesn't compile unless
> there is also a user in some fs/*.c.

If you build nfs (or pnfs) into the kernel image instead of as a
loadable module, does the build succeed?  I expect that it does
(or would).

>> The general rule is that if any in tree module uses a library
>> then the library is always linked in.
>> And to always link in a library you need to patch the Makefile
>> as Randy suggest in his mail.
>>
>>        Sam
>>
> 
> Again, I suspect I am misunderstanding.  It seems you are suggesting
> that to compile the nfs code I have to alter lib/Makefile.  While that is a
> useful workaround, that is certainly not a viable solution.

Why not?  We do that same change for other library options.


-- 
~Randy