Re: Build regressions/improvements in v4.17-rc1

On Wed, Aug 8, 2018 at 12:34 PM Michael Ellerman [off-list ref] wrote:

Andrew Morton [off-list ref] writes:

quoted

On Mon, 6 Aug 2018 12:39:21 +0200 Geert Uytterhoeven [off-list ref] wrote:

quoted

CC Dan, Michael, AKPM, powerpc

On Mon, Apr 16, 2018 at 3:10 PM Geert Uytterhoeven [off-list ref] wrote:

quoted

Below is the list of build error/warning regressions/improvements in
v4.17-rc1[1] compared to v4.16[2].

I'd like to point your attention to:

quoted

  + warning: vmlinux.o(.text+0x376518): Section mismatch in reference from the function .devm_memremap_pages() to the function .meminit.text:.arch_add_memory():  => N/A
  + warning: vmlinux.o(.text+0x376d64): Section mismatch in reference from the function .devm_memremap_pages_release() to the function .meminit.text:.arch_remove_memory():  => N/A

hm.  Dan isn't around at present so we're on our own with this one.

x86 doesn't put arch_add_memory and arch_remove_memory into __meminit.
x86 does

#ifdef CONFIG_MEMORY_HOTPLUG
int arch_add_memory(int nid, u64 start, u64 size, struct vmem_altmap *altmap,
              bool want_memblock)
{
      ...


So I guess powerpc should do that as well?

But we only recently added it to fix a section mismatch warning:

  WARNING: vmlinux.o(.text+0x6da88): Section mismatch in reference from the function .arch_add_memory() to the function .meminit.text:.create_section_mapping()
  The function .arch_add_memory() references
  the function __meminit .create_section_mapping().
  This is often because .arch_add_memory lacks a __meminit
  annotation or the annotation of .create_section_mapping is wrong.


I think the problem is that the section mismatch logic isn't able to
cope with __meminit's changing semantics.

When CONFIG_MEMORY_HOTPLUG=y references from .text to .meminit.text
should be allowed, because they're just folded in together in the linker
script.

When CONFIG_MEMORY_HOTPLUG=n references from .text to .meminit.text
should NOT be allowed, because .meminit.text becomes .init.text and will
be freed.

I don't see anything in the section mismatch logic to cope with that
difference.

It looks like __meminit is saving us about 1K on powerpc, so I'm
strongly inclined to just remove it entirely from arch/powerpc.

Also I haven't been seeing this in my local builds because I have:

CONFIG_LD_DEAD_CODE_DATA_ELIMINATION=y

So I guess we need to work out why that's interfering with section
mismatch analysis.

...

Well that's a good question actually. Section mismatch
analysis is done on the throwaway vmlinux.o which is not linked
with --gc-sections (and is not a final link), so the via_pmu_driver
symbol should exist and be picked up.

I wonder if something about the  -ffunction-sections is breaking
the reference detection.
...



ref:
https://www.mail-archive.com/linuxppc-dev@lists.ozlabs.org/msg135431.html

cheers