[RFC,DO,NOT,APPLY] vmlinux.lds: revert link speed regression

Commit Message

Arnd Bergmann Jan. 20, 2025, 9:21 p.m. UTC

From: Arnd Bergmann <[e-mail]>

I noticed a regression in the time it takes to fully link some randconfig
kernels and bisected this to commit 0043ecea2399 ("vmlinux.lds.h: Adjust
symbol ordering in text output section"), which (among other changes) moves
.text.unlikely ahead of .text.

Partially reverting this makes the final link over six times faster again,
back to what it was in linux-6.12:

		linux-6.12	linux-6.13
ld.lld v20	1.2s		1.2s
ld.bfd v2.36	3.2s		5.2s
ld.bfd v2.39	59s		388s

According to the commit description, that revert is not allowed here
because with CONFIG_LD_DEAD_CODE_DATA_ELIMINATION, the .text.unlikely
section name conflicts with the function-section names. On the other
hand, the excessive link time happens both with and without that
option, so the order could be conditional.

I did not try to bisect the linker beyond trying multiple versions
I had installed already, and it does feel like the behavior of recent
versions (tested 2.39 and 2.42 with identical results) is broken in
some form that earlier versions were not. According to 'perf', most
of the time is spent in elf_link_adjust_relocs() and ext64l_r_offset().

I also did not try to narrow the problem down to specific kernel
configuration options, but from my first impression it does appear
to be rare, and unrelated to the Propeller options added in 6.13.

Cc: [e-mail]
Cc: Han Shen <[e-mail]>
Cc: Masahiro Yamada <[e-mail]>
Cc: Nathan Chancellor <[e-mail]>
Cc: Kees Cook <[e-mail]>
Fixes: 0043ecea2399 ("vmlinux.lds.h: Adjust symbol ordering in text output section")
Link: https://pastebin.com/raw/sWpbkapL (config)
Signed-off-by: Arnd Bergmann <[e-mail]>
---
 include/asm-generic/vmlinux.lds.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Comments

Masahiro Yamada Jan. 21, 2025, 12:19 a.m. UTC | #1

On Tue, Jan 21, 2025 at 6:29 AM Arnd Bergmann <[e-mail]> wrote:
>
> From: Arnd Bergmann <[e-mail]>
>
> I noticed a regression in the time it takes to fully link some randconfig
> kernels and bisected this to commit 0043ecea2399 ("vmlinux.lds.h: Adjust
> symbol ordering in text output section"), which (among other changes) moves
> .text.unlikely ahead of .text.
>
> Partially reverting this makes the final link over six times faster again,
> back to what it was in linux-6.12:
>
>                 linux-6.12      linux-6.13
> ld.lld v20      1.2s            1.2s
> ld.bfd v2.36    3.2s            5.2s
> ld.bfd v2.39    59s             388s
>
> According to the commit description, that revert is not allowed here
> because with CONFIG_LD_DEAD_CODE_DATA_ELIMINATION, the .text.unlikely
> section name conflicts with the function-section names. On the other
> hand, the excessive link time happens both with and without that
> option, so the order could be conditional.
>
> I did not try to bisect the linker beyond trying multiple versions
> I had installed already, and it does feel like the behavior of recent
> versions (tested 2.39 and 2.42 with identical results) is broken in
> some form that earlier versions were not. According to 'perf', most
> of the time is spent in elf_link_adjust_relocs() and ext64l_r_offset().

Is this problem specific to the BFD linker from binutils?

Did you observe a link speed regression with LLVM=1 build?





--
Best Regards
Masahiro Yamada

Patch

diff --git a/include/asm-generic/vmlinux.lds.h b/include/asm-generic/vmlinux.lds.h
index 54504013c749..61fa047023b5 100644
--- a/include/asm-generic/vmlinux.lds.h
+++ b/include/asm-generic/vmlinux.lds.h
@@ -588,10 +588,10 @@  defined(CONFIG_AUTOFDO_CLANG) || defined(CONFIG_PROPELLER_CLANG)
 		*(.text.asan.* .text.tsan.*)				\
 		*(.text.unknown .text.unknown.*)			\
 		TEXT_SPLIT						\
-		TEXT_UNLIKELY						\
 		. = ALIGN(PAGE_SIZE);					\
 		TEXT_HOT						\
 		*(TEXT_MAIN .text.fixup)				\
+		TEXT_UNLIKELY						\
 		NOINSTR_TEXT						\
 		*(.ref.text)