3.1-rc1 link failure

[Russell King - ARM Linux]

On Mon, Aug 08, 2011 at 01:49:31PM -0500, Rob Herring wrote:
> Russell,
> 
> This commit is causing some link failures:
> 
>     ARM: vmlinux.lds: move discarded sections to beginning
> 
>     Rather than scattering the discarded sections throughout the linker
>     file, move them to the start.
> 
>     Acked-by: Nicolas Pitre <nicolas.pitre at linaro.org>
>     Tested-by: Stephen Boyd <sboyd at codeaurora.org>
>     Signed-off-by: Russell King <rmk+kernel at arm.linux.org.uk>
> 
> The error is this:
> 
> `.exit.text' referenced in section `.alt.smp.init' of
> drivers/built-in.o: defined in discarded section `.exit.text' of
> drivers/built
> -in.o
> `.exit.text' referenced in section `.alt.smp.init' of net/built-in.o:
> defined in discarded section `.exit.text' of net/built-in.o
> 
> I traced the one in net/ to CONFIG_SMP_ON_UP=y and CONFIG_DCB=y.
> dcbnl_exit calls dcb_flushapp which contains a spinlock. dcb_flushapp is
> getting inlined into dcbnl_exit.

Argh.  This is going to be an absolute _pig_ to fix.

This is the relevent part of the linker script (reformatted to make it
clearer):

| SECTIONS
| {
|  /*
|   * unwind exit sections must be discarded before the rest of the
|   * unwind sections get included.
|   */
|  /DISCARD/ : {
|   *(.ARM.exidx.exit.text)
|   *(.ARM.extab.exit.text)
|  }
| ...
|  .exit.text : {
|   *(.exit.text)
|   *(.memexit.text)
|  }
| ...
|  /DISCARD/ : {
|   *(.exit.text)
|   *(.memexit.text)
|   *(.exit.data)
|   *(.memexit.data)
|   *(.memexit.rodata)
|   *(.exitcall.exit)
|   *(.discard)
|   *(.discard.*)
|   }
| }

Now, this is what the linker manual says about discarded output sections:

|    The special output section name `/DISCARD/' may be used to discard
| input sections.  Any input sections which are assigned to an output
| section named `/DISCARD/' are not included in the output file.

No questions, no exceptions.  It doesn't say "unless they are listed
before the /DISCARD/ section."  Now, this is what asn-generic/vmlinux.lds.S
says:

| /*
|  * Default discarded sections.
|  *
|  * Some archs want to discard exit text/data at runtime rather than
|  * link time due to cross-section references such as alt instructions,
|  * bug table, eh_frame, etc.  DISCARDS must be the last of output
|  * section definitions so that such archs put those in earlier section
|  * definitions.
|  */

And guess what - the list _always_ includes .exit.text etc.

Now, what's actually happening is that the linker is reading the script,
and it finds the first /DISCARD/ output section at the beginning of the
script.  It continues reading the script, and finds the 'DISCARD' macro
at the end, which having been postprocessed results in another
/DISCARD/ output section.  As the linker already contains the earlier
/DISCARD/ output section, it adds it to that existing section, so it
effectively is placed at the start.  This can be seen by using the -M
option to ld:

| Linker script and memory map
| 
|                 0xc037c080                jiffies = jiffies_64
| 
| /DISCARD/
|  *(.ARM.exidx.exit.text)
|  *(.ARM.extab.exit.text)
|  *(.exit.text)
|  *(.memexit.text)
|  *(.exit.data)
|  *(.memexit.data)
|  *(.memexit.rodata)
|  *(.exitcall.exit)
|  *(.discard)
|  *(.discard.*)
| 
|                 0xc0008000                . = 0xc0008000
| 
| .head.text      0xc0008000      0x1d0
|                 0xc0008000                _text = .
|  *(.head.text)
|  .head.text     0xc0008000      0x1d0 arch/arm/kernel/head.o
|                 0xc0008000                stext
| 
| .text           0xc0008200   0x2d78d0
|                 0xc0008200                _stext = .
|                 0xc0008200                __exception_text_start = .
|  *(.exception.text)
|  .exception.text
| ...

As you can see, all the discarded sections are grouped together - and
as a result of it being the first output section, they all appear before
any other section.

The result is that not only is the unwind information discarded (as
intended), but also the .exit.text, despite us wanting to have the
.exit.text preserved.

We can't move the unwind information elsewhere, because it'll then be
included even when we do actually discard the .exit.text (and similar)
sections.

The only solution that I can think of is to stop using this
asm-generic/vmlinux.lds.S and write our own fully conditionalized
linker script, ensuring that no input section is mentioned more than
once in the output sections.

Or someone sorts out the asm-generic/vmlinux.lds.S stuff to actually
conform to the linker manual, and stop relying on implementation defined
behaviour of the linker - again by having it fully conditionalized.