Skip to contentSkip to conversation

kbuild: Fix lto build issue for aarch64

PATCH Masahiro Yamada 2 participants 2 messages Jan 31, 2025 1 message not in the archive Kbuild

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

#1 lore raw
Quoted 8 lines · Yonghong Song

Commit bede169618c6 ("kbuild: enable objtool for *.mod.o and additional kernel objects") added objtool support for various files like *.mod.o, vmlinux.export.o, etc.
But unfortunately it also breaks lto build for aarch64 where a common source file, e.g., kernel/bpf/syscall.c, is compiled with command:
clang ... -flto=thin ... -c -o kernel/bpf/syscall.o /home/yhs/work/bpf-next/kernel/bpf/syscall.c ;

  ld.lld -EL  -maarch64elf -z norelro -mllvm -import-instr-limit=5 -z noexecstack   \
      -r -o kernel/bpf/.tmp_syscall.o kernel/bpf/syscall.o;
Quoted 7 lines

mv kernel/bpf/.tmp_syscall.o kernel/bpf/syscall.o

The above command line completely destroyed thin-lto since file kernel/bpf/syscall.o is expected to be a llvm bitcode file and later ld.lld can do cross-file inlining.

The issue has been reported in [1] and Masahiro actually provided a suggested fix. But unfortunately that suggested fix does not work.

Thanks for the report, but your patch does not fix the real issue.

I think my original [1] worked, although it was a bit redundant (extra $(LD) is executed where it is unnecessary).

Quoted 2 lines

This patch made improvement on top of [2] and the issue for aarch64 lto is fixed. The x86_64 is also tested without any issue.

You are not fixing the issue.
You are just disabling objtool (i.e. reverting bede169618c6).

Quoted 4 lines

In [1], commit 1b466b29a3bf ("kbuild: re-enable KCSAN for autogenerated *.mod.c intermediaries") enabled kcsan for *.mod.c files. This patch actually disabled that.
Should we revert commit 1b466b29a3bf as well?

No.
I do not want to revert 1b466b29a3bf.

If 1b466b29a3bf were reverted, there would be no point in applying bede169618c6.

Quoted 29 lines

[1] https://lore.kernel.org/lkml/[email protected]/
[2] https://lore.kernel.org/lkml/CAK7LNATpu5zYwx7kmaknsPGLXt8n8uCXyFpdi5vZeFZiBxYkGw@mail.gmail.com/

Fixes bede169618c6 ("kbuild: enable objtool for *.mod.o and additional kernel objects")
---
 scripts/Makefile.build | 2 ++
 scripts/Makefile.lib   | 4 +---
 2 files changed, 3 insertions(+), 3 deletions(-)
scripts/Makefile.build +2 -0
index c16e4cf54d77..e9ae38b71a08 100644
@@ -129,6 +129,8 @@ $(obj)/%.ll: $(obj)/%.c FORCE
 
 is-single-obj-m = $(and $(part-of-module),$(filter $@, $(obj-m)),y)
 
+need-ld-single = $(is-single-obj-m)
+
 ifdef CONFIG_MODVERSIONS
 # When module versioning is enabled the following steps are executed:
 # o compile a <file>.o from <file>.c
scripts/Makefile.lib +0 -0
index 7395200538da..f04672bfe068 100644

scripts/Makefile.lib endef · line 306

Quoted 5 lines
scripts/Makefile.lib +0 -2
@@ -302,11 +302,9 @@ endef
 # ===========================================================================
 # These are shared by some Makefile.* files.
 
-objtool-enabled := y
-

This is wrong.

You are simply disabling objtool for *.mod.o

I am pretty sure the problem reported by 54babdc
will come back.

I submitted a proper fix.

Quoted 10 lines
scripts/Makefile.lib +1 -2
 ifdef CONFIG_LTO_CLANG
 # objtool cannot process LLVM IR. Make $(LD) covert LLVM IR to ELF here.
-cmd_ld_single = $(if $(objtool-enabled), ; $(LD) $(ld_flags) -r -o $(tmp-target) $@; mv $(tmp-target) $@)
+cmd_ld_single = $(if $(need-ld-single), ; $(LD) $(ld_flags) -r -o $(tmp-target) $@; mv $(tmp-target) $@)
 endif
 
 quiet_cmd_cc_o_c = CC $(quiet_modtag)  $@
--

2.43.5

--
Best Regards

Masahiro Yamada

#2 lore raw
Quoted 17 lines · Masahiro Yamada
Quoted 9 lines · Yonghong Song

Commit bede169618c6 ("kbuild: enable objtool for *.mod.o and additional kernel objects") added objtool support for various files like *.mod.o, vmlinux.export.o, etc. But unfortunately it also breaks lto build for aarch64 where a common source file, e.g., kernel/bpf/syscall.c, is compiled with command:

   clang ... -flto=thin ... -c -o kernel/bpf/syscall.o /home/yhs/work/bpf-next/kernel/bpf/syscall.c  ;
   ld.lld -EL  -maarch64elf -z norelro -mllvm -import-instr-limit=5 -z noexecstack   \
       -r -o kernel/bpf/.tmp_syscall.o kernel/bpf/syscall.o;
   mv kernel/bpf/.tmp_syscall.o kernel/bpf/syscall.o
Quoted 5 lines

The above command line completely destroyed thin-lto since file kernel/bpf/syscall.o is expected to be a llvm bitcode file and later ld.lld can do cross-file inlining.

The issue has been reported in [1] and Masahiro actually provided a suggested fix. But unfortunately that suggested fix does not work.

Thanks for the report, but your patch does not fix the real issue.

Quoted 8 lines

I think my original [1] worked, although it was a bit redundant (extra $(LD) is executed where it is unnecessary).

Quoted 2 lines

This patch made improvement on top of [2] and the issue for aarch64 lto is fixed. The x86_64 is also tested without any issue.

You are not fixing the issue.
You are just disabling objtool (i.e. reverting bede169618c6).

You are correct. Essentially my patch just reverted bede169618c6 since I am not familar with Makefile.build and Makefile.lib so won't be able to come up with a proper solution.

Quoted 47 lines
Quoted 5 lines

In [1], commit 1b466b29a3bf ("kbuild: re-enable KCSAN for autogenerated *.mod.c intermediaries") enabled kcsan for *.mod.c files. This patch actually disabled that.
Should we revert commit 1b466b29a3bf as well?

No.
I do not want to revert 1b466b29a3bf.

If 1b466b29a3bf were reverted, there would be no point in applying bede169618c6.

Quoted 28 lines
   [1] https://lore.kernel.org/lkml/[email protected]/
   [2] https://lore.kernel.org/lkml/CAK7LNATpu5zYwx7kmaknsPGLXt8n8uCXyFpdi5vZeFZiBxYkGw@mail.gmail.com/
Fixes bede169618c6 ("kbuild: enable objtool for *.mod.o and additional kernel objects")
---
  scripts/Makefile.build | 2 ++
  scripts/Makefile.lib   | 4 +---
  2 files changed, 3 insertions(+), 3 deletions(-)
scripts/Makefile.build +2 -0
index c16e4cf54d77..e9ae38b71a08 100644
@@ -129,6 +129,8 @@ $(obj)/%.ll: $(obj)/%.c FORCE
 
  is-single-obj-m = $(and $(part-of-module),$(filter $@, $(obj-m)),y)
 
+need-ld-single = $(is-single-obj-m)
+
  ifdef CONFIG_MODVERSIONS
  # When module versioning is enabled the following steps are executed:
  # o compile a <file>.o from <file>.c
scripts/Makefile.lib +0 -0
index 7395200538da..f04672bfe068 100644

scripts/Makefile.lib endef · line 306

Quoted 5 lines
scripts/Makefile.lib +0 -2
@@ -302,11 +302,9 @@ endef
  # ===========================================================================
  # These are shared by some Makefile.* files.
 
-objtool-enabled := y
-

This is wrong.

Quoted 7 lines

You are simply disabling objtool for *.mod.o

I am pretty sure the problem reported by 54babdc
will come back.

I submitted a proper fix.

I tested the patch with arm64/x86_64 + lto/no-lto. Everything works fine. I have replied with my Tested-by tag. Thanks!

Quoted 18 lines
Quoted 11 lines
scripts/Makefile.lib +1 -2
  ifdef CONFIG_LTO_CLANG
  # objtool cannot process LLVM IR. Make $(LD) covert LLVM IR to ELF here.
-cmd_ld_single = $(if $(objtool-enabled), ; $(LD) $(ld_flags) -r -o $(tmp-target) $@; mv $(tmp-target) $@)
+cmd_ld_single = $(if $(need-ld-single), ; $(LD) $(ld_flags) -r -o $(tmp-target) $@; mv $(tmp-target) $@)
  endif
 
  quiet_cmd_cc_o_c = CC $(quiet_modtag)  $@
--
Quoted 2 lines

2.43.5

--
Best Regards

Masahiro Yamada