Re: Build failure in -next due to 'kbuild: allow archs to select link dead code/data elimination'

[Guenter Roeck]

On Wed, Sep 14, 2016 at 12:30:04PM +1000, Nicholas Piggin wrote:
> On Tue, 13 Sep 2016 13:25:22 -0700
> Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
> 
> > On 09/12/2016 08:51 PM, Nicholas Piggin wrote:
> > > On Mon, 12 Sep 2016 20:17:30 -0700
> > > Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
> > >  
> > >> Hi Nicholas,
> > >>
> > >> On 09/12/2016 07:00 PM, Nicholas Piggin wrote:  
> > >>> On Mon, 12 Sep 2016 15:24:43 -0700
> > >>> Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
> > >>>  
> > >>>> Hi,
> > >>>>
> > >>>> your commit 'kbuild: allow archs to select link dead code/data elimination'
> > >>>> is causing the following build failure in -next when building score:defconfig.
> > >>>>
> > >>>> arch/score/kernel/built-in.o: In function `work_resched':
> > >>>> arch/score/kernel/entry.o:(.text+0xe84):
> > >>>> 	relocation truncated to fit: R_SCORE_PC19 against `schedule'
> > >>>>
> > >>>> Reverting the commit fixes the problem.
> > >>>>
> > >>>> Please let me know if I can help tracking down the problem.
> > >>>> In case you need a score toochain, you can find the one I use at
> > >>>> http://server.roeck-us.net/toolchains/score.tgz.
> > >>>>
> > >>>> Thanks,
> > >>>> Guenter  
> > >>>
> > >>> It's not supposed to have any real effect unless the
> > >>> option is selected, but there are a few changes to linker
> > >>> script which must be causing it.
> > >>>
> > >>> There are two changes to vmlinux.lds.h. One is to KEEP a
> > >>> few input sections, that *should* be kept anyway. The other
> > >>> is to bring in additional sections into their correct output
> > >>> section.
> > >>>
> > >>> Could you try reverting those lines of vmlinux.lds.h that
> > >>> change the latter, i.e.:
> > >>>
> > >>> -		*(.text.hot .text .text.fixup .text.unlikely)		\
> > >>> +		*(.text.hot .text .text.fixup .text.unlikely .text.*)	\  
> > >>
> > >> This is the culprit. After removing " .text.*" it builds fine.  
> > >
> > > Thanks for testing it. Some architectures have .text.x sections
> > > not included here, I should have seen that. We should possibly
> > > just revert that line and require implementing archs to do the
> > > right thing.
> > >  
> > I would call the build failure a regression, so it should either be
> > reverted, or we'll need some other solution to fix the build failure.
> 
> Yes it's definitely a regression and that part should be reverted.
> To confirm, the following patch fixes your build?
> 
> Thanks,
> Nick
> 
> commit 0ae28be83b4d6cd03ef5b481487d042f2b91954e
> Author: Nicholas Piggin <npiggin@xxxxxxxxx>
> Date:   Wed Sep 14 12:24:03 2016 +1000
> 
>     kbuild: -ffunction-sections fix for archs with conflicting sections
>     
>     Enabling -ffunction-sections modified the generic linker script to
>     pull .text.* sections into regular TEXT_TEXT section, conflicting
>     with some architectures. Revert that change and require archs that
>     enable the option to ensure they have no conflicting section names,
>     and do the appropriate merging.
> 
>     Signed-off-by: Nicholas Piggin <npiggin@xxxxxxxxx>
> 
Tested-by: Guenter Roeck <linux@xxxxxxxxxxxx>

> diff --git a/arch/Kconfig b/arch/Kconfig
> index 6d5b631..8248037 100644
> --- a/arch/Kconfig
> +++ b/arch/Kconfig
> @@ -487,7 +487,9 @@ config LD_DEAD_CODE_DATA_ELIMINATION
>  	  This requires that the arch annotates or otherwise protects
>  	  its external entry points from being discarded. Linker scripts
>  	  must also merge .text.*, .data.*, and .bss.* correctly into
> -	  output sections.
> +	  output sections. Care must be taken not to pull in unrelated
> +	  sections (e.g., '.text.init'). Typically '.' in section names
> +	  is used to distinguish them from label names / C identifiers.
>  
>  config HAVE_CONTEXT_TRACKING
>  	bool
> diff --git a/include/asm-generic/vmlinux.lds.h b/include/asm-generic/vmlinux.lds.h
> index ad9d8f9..48dd44f 100644
> --- a/include/asm-generic/vmlinux.lds.h
> +++ b/include/asm-generic/vmlinux.lds.h
> @@ -198,9 +198,9 @@
>  
>  /*
>   * .data section
> - * -fdata-sections generates .data.identifier which needs to be pulled in
> - * with .data, but don't want to pull in .data..stuff which has its own
> - * requirements. Same for bss.
> + * LD_DEAD_CODE_DATA_ELIMINATION option enables -fdata-sections generates
> + * .data.identifier which needs to be pulled in with .data, but don't want to
> + * pull in .data..stuff which has its own requirements. Same for bss.
>   */
>  #define DATA_DATA							\
>  	*(.data .data.[0-9a-zA-Z_]*)					\
> @@ -434,10 +434,15 @@
>  	}
>  
>  /* .text section. Map to function alignment to avoid address changes
> - * during second ld run in second ld pass when generating System.map */
> + * during second ld run in second ld pass when generating System.map
> + * LD_DEAD_CODE_DATA_ELIMINATION option enables -ffunction-sections generates
> + * .text.identifier which needs to be pulled in with .text , but some
> + * architectures define .text.foo which is not intended to be pulled in here.
> + * Those enabling LD_DEAD_CODE_DATA_ELIMINATION must ensure they don't have
> + * conflicting section names, and must pull in .text.[0-9a-zA-Z_]* */
>  #define TEXT_TEXT							\
>  		ALIGN_FUNCTION();					\
> -		*(.text.hot .text .text.fixup .text.unlikely .text.*)	\
> +		*(.text.hot .text .text.fixup .text.unlikely)		\
>  		*(.ref.text)						\
>  	MEM_KEEP(init.text)						\
>  	MEM_KEEP(exit.text)						\