Link breakage on all architectures which implement their own show_mem

Link breakage on all architectures which implement their own show_mem

[James Bottomley]

This is the problem a lot of architectures will see:

Specifically this:
  LD      vmlinux.o
lib/lib.a(show_mem.o): In function `show_mem':
(.text.show_mem+0x0): multiple definition of `show_mem'
arch/parisc/mm/built-in.o:(.text.show_mem+0x0): first defined here
make: *** [vmlinux.o] Error 1

Caused by this patch:

commit ddd588b5dd55f14320379961e47683db4e4c1d90
Author: David Rientjes <rientjes@google.com>
Date:   Tue Mar 22 16:30:46 2011 -0700

    oom: suppress nodes that are not allowed from meminfo on oom kill

Is the cause.  What it does is introduce a new __show_mem() which is
required by files in mm/, so the object containing it: show_mem.o gets
pulled in all the time in the link and that gives every architecture
that implements their own show_mem() a link failure because of the
double definition.  Library linking works at the file level, not at the
function level.  To work, you have to put these functions in separate
files.

James

Re: Link breakage on all architectures which implement their own show_mem

[David Rientjes]

On Wed, 23 Mar 2011, James Bottomley wrote:

> This is the problem a lot of architectures will see:
> 
> Specifically this:
>   LD      vmlinux.o
> lib/lib.a(show_mem.o): In function `show_mem':
> (.text.show_mem+0x0): multiple definition of `show_mem'
> arch/parisc/mm/built-in.o:(.text.show_mem+0x0): first defined here
> make: *** [vmlinux.o] Error 1
> 
> Caused by this patch:
> 
> commit ddd588b5dd55f14320379961e47683db4e4c1d90
> Author: David Rientjes <rientjes@google.com>
> Date:   Tue Mar 22 16:30:46 2011 -0700
> 
>     oom: suppress nodes that are not allowed from meminfo on oom kill
> 
> Is the cause.  What it does is introduce a new __show_mem() which is
> required by files in mm/, so the object containing it: show_mem.o gets
> pulled in all the time in the link and that gives every architecture
> that implements their own show_mem() a link failure because of the
> double definition.  Library linking works at the file level, not at the
> function level.  To work, you have to put these functions in separate
> files.
> 

Thanks for the report, James.

This was also reported by Stephen on LKML when he merged the latest Linus 
tree into linux-next.  Andrew suggested that we remove __show_mem() 
entirely and simply add its argument to show_mem() in all the 
architectures.  I have a patch at:

	http://marc.info/?l=linux-kernel&m=130100511521954

I don't have a parisc cross compiler, so if you could try it out and add 
your Tested-by that would be great

Thanks again!

Re: Link breakage on all architectures which implement their own show_mem

[Benjamin Herrenschmidt]

On Wed, 2011-03-23 at 11:41 -0500, James Bottomley wrote:
>     oom: suppress nodes that are not allowed from meminfo on oom kill
> 
> Is the cause.  What it does is introduce a new __show_mem() which is
> required by files in mm/, so the object containing it: show_mem.o gets
> pulled in all the time in the link and that gives every architecture
> that implements their own show_mem() a link failure because of the
> double definition.  Library linking works at the file level, not at
> the
> function level.  To work, you have to put these functions in separate 

Can't this be fixed by making show_mem() weak ? Or am I missing
something ?

Cheers,
Ben.

Re: Link breakage on all architectures which implement their own show_mem

[James Bottomley]

On Fri, 2011-03-25 at 21:55 +1100, Benjamin Herrenschmidt wrote:
> On Wed, 2011-03-23 at 11:41 -0500, James Bottomley wrote:
> >     oom: suppress nodes that are not allowed from meminfo on oom kill
> > 
> > Is the cause.  What it does is introduce a new __show_mem() which is
> > required by files in mm/, so the object containing it: show_mem.o gets
> > pulled in all the time in the link and that gives every architecture
> > that implements their own show_mem() a link failure because of the
> > double definition.  Library linking works at the file level, not at
> > the
> > function level.  To work, you have to put these functions in separate 
> 
> Can't this be fixed by making show_mem() weak ? Or am I missing
> something ?

Yes, I think so; there are multiple ways of solving this ... part of the
point I was making was that there's little point having something in
lib.a if it always gets linked (which this will because of the multiple
callsites for __show_mem())

James

Re: Link breakage on all architectures which implement their own show_mem

[James Bottomley]

On Thu, 2011-03-24 at 15:21 -0700, David Rientjes wrote:
> I don't have a parisc cross compiler, so if you could try it out and add 
> your Tested-by that would be great

I'm afraid this hit at exactly the wrong time (we're relocating the
parisc dev cluster in cupertino), so all my parisc build systems are
currently trekking across country on fedex mules.

James

Re: Link breakage on all architectures which implement their own show_mem

[Josh Boyer]

On Thu, Mar 24, 2011 at 6:21 PM, David Rientjes <rientjes@google.com> wrote:
> On Wed, 23 Mar 2011, James Bottomley wrote:
>
>> This is the problem a lot of architectures will see:
>>
>> Specifically this:
>>   LD      vmlinux.o
>> lib/lib.a(show_mem.o): In function `show_mem':
>> (.text.show_mem+0x0): multiple definition of `show_mem'
>> arch/parisc/mm/built-in.o:(.text.show_mem+0x0): first defined here
>> make: *** [vmlinux.o] Error 1
>>
>> Caused by this patch:
>>
>> commit ddd588b5dd55f14320379961e47683db4e4c1d90
>> Author: David Rientjes <rientjes@google.com>
>> Date:   Tue Mar 22 16:30:46 2011 -0700
>>
>>     oom: suppress nodes that are not allowed from meminfo on oom kill
>>
>> Is the cause.  What it does is introduce a new __show_mem() which is
>> required by files in mm/, so the object containing it: show_mem.o gets
>> pulled in all the time in the link and that gives every architecture
>> that implements their own show_mem() a link failure because of the
>> double definition.  Library linking works at the file level, not at the
>> function level.  To work, you have to put these functions in separate
>> files.
>>
>
> Thanks for the report, James.
>
> This was also reported by Stephen on LKML when he merged the latest Linus
> tree into linux-next.  Andrew suggested that we remove __show_mem()
> entirely and simply add its argument to show_mem() in all the
> architectures.  I have a patch at:
>
>        http://marc.info/?l=linux-kernel&m=130100511521954
>
> I don't have a parisc cross compiler, so if you could try it out and add
> your Tested-by that would be great

You can find basic cross compilers for a variety of architectures
already built here:

http://www.kernel.org/pub/tools/crosstool/

josh

Re: Link breakage on all architectures which implement their own show_mem

[David Rientjes]

On Fri, 25 Mar 2011, James Bottomley wrote:

> > I don't have a parisc cross compiler, so if you could try it out and add 
> > your Tested-by that would be great
> 
> I'm afraid this hit at exactly the wrong time (we're relocating the
> parisc dev cluster in cupertino), so all my parisc build systems are
> currently trekking across country on fedex mules.
> 

Haha, ok.  Linus merged the patch into his tree so you should hear from 
parisc people if there is additional breakage, which there shouldn't be.  
Thanks again for the report.