objtool/arm64: Port klp-build to arm64

In preparation for enabling objtool on arm64, annotate intra-function calls.

Signed-off-by Josh Poimboeuf <[email protected]>

The arm64 boot header is mostly data. Move it to a data section to prevent objtool and other tools from trying to disassemble it. The final linked result is the same.

Signed-off-by Josh Poimboeuf <[email protected]>

When building with func-fdata-sections, the .init.bss section gets split up into a bunch of .init.bss.<var> sections. Make sure they get linked into .init.data.

Signed-off-by Josh Poimboeuf <[email protected]>

tools/objtool/check.c +11 -0

@@ -3691,9 +3691,30 @@ static void checksum_update_insn(struct objtool_file *file, struct symbol *func,
 				 struct instruction *insn)
 {
 	struct reloc *reloc = insn_reloc(file, insn);
+	struct alternative *alt;
 	unsigned long offset;
 	struct symbol *sym;
 
+	for (alt = insn->alts; alt; alt = alt->next) {
+		struct alt_group *alt_group = alt->insn->alt_group;
+
+		checksum_update(func, insn, &alt->type, sizeof(alt->type));
+
+		if (alt_group && alt_group->orig_group) {
+			struct instruction *alt_insn;
+
+			checksum_update(func, insn, &alt_group->feature, sizeof(alt_group->feature));
+

Hi,

Quoted 23 lines

tools/objtool/check.c +16 -0

@@ -3691,9 +3691,30 @@ static void checksum_update_insn(struct objtool_file *file, struct symbol *func,
 				 struct instruction *insn)
 {
 	struct reloc *reloc = insn_reloc(file, insn);
+	struct alternative *alt;
 	unsigned long offset;
 	struct symbol *sym;
 
+	for (alt = insn->alts; alt; alt = alt->next) {
+		struct alt_group *alt_group = alt->insn->alt_group;
+
+		checksum_update(func, insn, &alt->type, sizeof(alt->type));
+
+		if (alt_group && alt_group->orig_group) {
+			struct instruction *alt_insn;
+
+			checksum_update(func, insn, &alt_group->feature, sizeof(alt_group->feature));
+
+			for (alt_insn = alt->insn; alt_insn; alt_insn = next_insn_same_sec(file, alt_insn)) {
+				checksum_update_insn(file, func, alt_insn);
+				if (alt_insn == alt_group->last_insn)
+					break;
+			}

tools/objtool/check.c static void checksum_update_insn(struct objtool_file *file, struct symbol *func, · line 3717

Quoted 5 lines

tools/objtool/check.c +5 -0

+		} else {
+			checksum_update(func, insn, &alt->insn->offset, sizeof(alt->insn->offset));
+		}
+	}
+

does this hunk belong to the patch? Unless I am missing something, it might be worth a separate one.

Quoted 25 lines · Miroslav Benes

Hi,

Quoted 23 lines

tools/objtool/check.c +16 -0

@@ -3691,9 +3691,30 @@ static void checksum_update_insn(struct objtool_file *file, struct symbol *func,
 				 struct instruction *insn)
 {
 	struct reloc *reloc = insn_reloc(file, insn);
+	struct alternative *alt;
 	unsigned long offset;
 	struct symbol *sym;
 
+	for (alt = insn->alts; alt; alt = alt->next) {
+		struct alt_group *alt_group = alt->insn->alt_group;
+
+		checksum_update(func, insn, &alt->type, sizeof(alt->type));
+
+		if (alt_group && alt_group->orig_group) {
+			struct instruction *alt_insn;
+
+			checksum_update(func, insn, &alt_group->feature, sizeof(alt_group->feature));
+
+			for (alt_insn = alt->insn; alt_insn; alt_insn = next_insn_same_sec(file, alt_insn)) {
+				checksum_update_insn(file, func, alt_insn);
+				if (alt_insn == alt_group->last_insn)
+					break;
+			}

Quoted 8 lines

tools/objtool/check.c static void checksum_update_insn(struct objtool_file *file, struct symbol *func, · line 3717

Quoted 5 lines

tools/objtool/check.c +5 -0

+		} else {
+			checksum_update(func, insn, &alt->insn->offset, sizeof(alt->insn->offset));
+		}
+	}
+

does this hunk belong to the patch? Unless I am missing something, it might be worth a separate one.

It belongs, but I should have clarified that in the patch description.

This hunk wasn't needed before because validate_branch() already iterates all the alternatives, so it was calling checksum_update_insn() for every instruction in the function, including the alternatives.

Now that it's no longer called by validate_branch(), checksum_update_insn() has to manually iterate the alternatives.

Quoted 41 lines · Josh Poimboeuf

Quoted 25 lines · Miroslav Benes

Hi,

Quoted 23 lines

tools/objtool/check.c +16 -0

@@ -3691,9 +3691,30 @@ static void checksum_update_insn(struct objtool_file *file, struct symbol *func,
 				 struct instruction *insn)
 {
 	struct reloc *reloc = insn_reloc(file, insn);
+	struct alternative *alt;
 	unsigned long offset;
 	struct symbol *sym;
 
+	for (alt = insn->alts; alt; alt = alt->next) {
+		struct alt_group *alt_group = alt->insn->alt_group;
+
+		checksum_update(func, insn, &alt->type, sizeof(alt->type));
+
+		if (alt_group && alt_group->orig_group) {
+			struct instruction *alt_insn;
+
+			checksum_update(func, insn, &alt_group->feature, sizeof(alt_group->feature));
+
+			for (alt_insn = alt->insn; alt_insn; alt_insn = next_insn_same_sec(file, alt_insn)) {
+				checksum_update_insn(file, func, alt_insn);
+				if (alt_insn == alt_group->last_insn)
+					break;
+			}

Quoted 8 lines

tools/objtool/check.c static void checksum_update_insn(struct objtool_file *file, struct symbol *func, · line 3717

Quoted 5 lines

tools/objtool/check.c +5 -0

+		} else {
+			checksum_update(func, insn, &alt->insn->offset, sizeof(alt->insn->offset));
+		}
+	}
+

does this hunk belong to the patch? Unless I am missing something, it might be worth a separate one.

It belongs, but I should have clarified that in the patch description.

This hunk wasn't needed before because validate_branch() already iterates all the alternatives, so it was calling checksum_update_insn() for every instruction in the function, including the alternatives.

Quoted 2 lines

Now that it's no longer called by validate_branch(), checksum_update_insn() has to manually iterate the alternatives.

After writing the email I had a suspicion it must have been something like above but failed to find it. Now I see that next_insn_to_validate() called in do_validate_branch() handles exactly that. Thanks for the pointer. The patch looks good to me then (and the rest as well as far as I can judge).

I must admit that objtool has gotten so complex that I have a hard time to track everything in the code :).

Instead of returning an error for --mnop without --mcount, just silently ignore it. This will help simplify kbuild's handling of objtool args.

Signed-off-by Josh Poimboeuf <[email protected]>

Split the objtool args into commands and options, such that if no commands have been enabled, objtool doesn't run.

This is in preparation in enabling objtool and klp-build for arm64.

Signed-off-by Josh Poimboeuf <[email protected]>

---
 arch/x86/boot/startup/Makefile |  2 +-
 scripts/Makefile.build         |  4 +--
 scripts/Makefile.lib           | 46 ++++++++++++++++++----------------
 scripts/Makefile.vmlinux_o     | 15 ++++-------
 4 files changed, 33 insertions(+), 34 deletions(-)

arch/x86/boot/startup/Makefile +1 -1

index 5e499cfb29b5..a08297829fc6 100644
@@ -36,7 +36,7 @@ $(patsubst %.o,$(obj)/%.o,$(lib-y)): OBJECT_FILES_NON_STANDARD := y
 # relocations, even if other objtool actions are being deferred.
 #
 $(pi-objs): objtool-enabled	= 1
-$(pi-objs): objtool-args	= $(if $(delay-objtool),--dry-run,$(objtool-args-y)) --noabs
+$(pi-objs): objtool-args	= $(if $(delay-objtool),--dry-run,$(objtool-cmds-y) $(objtool-opts-y)) --noabs
 
 #
 # Confine the startup code by prefixing all symbols with __pi_ (for position

scripts/Makefile.build +2 -2

index 32e209bc7985..d2d0776df947 100644
@@ -277,7 +277,7 @@ endif # CONFIG_FTRACE_MCOUNT_USE_RECORDMCOUNT
 is-standard-object = $(if $(filter-out y%, $(OBJECT_FILES_NON_STANDARD_$(target-stem).o)$(OBJECT_FILES_NON_STANDARD)n),$(is-kernel-object))
 
 ifdef CONFIG_OBJTOOL
-$(obj)/%.o: private objtool-enabled = $(if $(is-standard-object),$(if $(delay-objtool),$(is-single-obj-m),y))
+$(obj)/%.o: private objtool-enabled = $(if $(is-standard-object),$(if $(objtool-cmds-y),$(if $(delay-objtool),$(is-single-obj-m),y)))
 endif
 
 ifneq ($(findstring 1, $(KBUILD_EXTRA_WARN)),)

@@ -499,7 +499,7 @@ define rule_ld_multi_m
 	$(call cmd,gen_objtooldep)
 endef
 
-$(multi-obj-m): private objtool-enabled := $(delay-objtool)
+$(multi-obj-m): private objtool-enabled := $(if $(objtool-cmds-y),$(delay-objtool))
 $(multi-obj-m): private part-of-module := y
 $(multi-obj-m): %.o: %.mod FORCE
 	$(call if_changed_rule,ld_multi_m)

scripts/Makefile.lib +25 -21

index 0718e39cedda..40a462581666 100644
@@ -183,27 +183,31 @@ ifdef CONFIG_OBJTOOL
 
 objtool := $(objtree)/tools/objtool/objtool
 
-objtool-args-$(CONFIG_HAVE_JUMP_LABEL_HACK)		+= --hacks=jump_label
-objtool-args-$(CONFIG_HAVE_NOINSTR_HACK)		+= --hacks=noinstr
-objtool-args-$(CONFIG_MITIGATION_CALL_DEPTH_TRACKING)	+= --hacks=skylake
-objtool-args-$(CONFIG_X86_KERNEL_IBT)			+= --ibt
-objtool-args-$(CONFIG_FINEIBT)				+= --cfi
-objtool-args-$(CONFIG_FTRACE_MCOUNT_USE_OBJTOOL)	+= --mcount
-ifdef CONFIG_FTRACE_MCOUNT_USE_OBJTOOL
-objtool-args-$(CONFIG_HAVE_OBJTOOL_NOP_MCOUNT)		+= --mnop
-endif
-objtool-args-$(CONFIG_UNWINDER_ORC)			+= --orc
-objtool-args-$(CONFIG_MITIGATION_RETPOLINE)		+= --retpoline
-objtool-args-$(CONFIG_MITIGATION_RETHUNK)		+= --rethunk
-objtool-args-$(CONFIG_MITIGATION_SLS)			+= --sls
-objtool-args-$(CONFIG_STACK_VALIDATION)			+= --stackval
-objtool-args-$(CONFIG_HAVE_STATIC_CALL_INLINE)		+= --static-call
-objtool-args-$(CONFIG_HAVE_UACCESS_VALIDATION)		+= --uaccess
-objtool-args-$(or $(CONFIG_GCOV_KERNEL),$(CONFIG_KCOV))	+= --no-unreachable
-objtool-args-$(CONFIG_PREFIX_SYMBOLS)			+= --prefix=$(CONFIG_FUNCTION_PADDING_BYTES)
-objtool-args-$(CONFIG_OBJTOOL_WERROR)			+= --werror
+# objtool commands
+objtool-cmds-$(CONFIG_HAVE_JUMP_LABEL_HACK)		+= --hacks=jump_label
+objtool-cmds-$(CONFIG_HAVE_NOINSTR_HACK)		+= --hacks=noinstr
+objtool-cmds-$(CONFIG_MITIGATION_CALL_DEPTH_TRACKING)	+= --hacks=skylake
+objtool-cmds-$(CONFIG_X86_KERNEL_IBT)			+= --ibt
+objtool-cmds-$(CONFIG_FINEIBT)				+= --cfi
+objtool-cmds-$(CONFIG_FTRACE_MCOUNT_USE_OBJTOOL)	+= --mcount
+objtool-cmds-$(CONFIG_UNWINDER_ORC)			+= --orc
+objtool-cmds-$(CONFIG_MITIGATION_RETPOLINE)		+= --retpoline
+objtool-cmds-$(CONFIG_MITIGATION_RETHUNK)		+= --rethunk
+objtool-cmds-$(CONFIG_MITIGATION_SLS)			+= --sls
+objtool-cmds-$(CONFIG_STACK_VALIDATION)			+= --stackval
+objtool-cmds-$(CONFIG_HAVE_STATIC_CALL_INLINE)		+= --static-call
+objtool-cmds-$(CONFIG_HAVE_UACCESS_VALIDATION)		+= --uaccess
+objtool-cmds-$(CONFIG_PREFIX_SYMBOLS)			+= --prefix=$(CONFIG_FUNCTION_PADDING_BYTES)
+objtool-cmds-y						+= $(OBJTOOL_ARGS)
 
-objtool-args = $(objtool-args-y)					\
+# objtool options
+ifdef CONFIG_FTRACE_MCOUNT_USE_OBJTOOL
+objtool-opts-$(CONFIG_HAVE_OBJTOOL_NOP_MCOUNT)		+= --mnop
+endif
+objtool-opts-$(or $(CONFIG_GCOV_KERNEL),$(CONFIG_KCOV))	+= --no-unreachable
+objtool-opts-$(CONFIG_OBJTOOL_WERROR)			+= --werror
+
+objtool-args = $(objtool-cmds-y) $(objtool-opts-y)			\
 	$(if $(delay-objtool), --link)					\
 	$(if $(part-of-module), --module)
 

@@ -212,7 +216,7 @@ delay-objtool := $(or $(CONFIG_LTO_CLANG),$(CONFIG_X86_KERNEL_IBT),$(CONFIG_KLP_
 cmd_objtool = $(if $(objtool-enabled), ; $(objtool) $(objtool-args) $@)
 cmd_gen_objtooldep = $(if $(objtool-enabled), { echo ; echo '$@: $$(wildcard $(objtool))' ; } >> $(dot-target).cmd)
 
-objtool-enabled := y
+objtool-enabled = $(if $(objtool-cmds-y),y)
 
 endif # CONFIG_OBJTOOL
 

scripts/Makefile.vmlinux_o +5 -10

index 527352c222ff..09af33203bd8 100644
@@ -36,18 +36,13 @@ endif
 # For !delay-objtool + CONFIG_NOINSTR_VALIDATION, it runs on both translation
 # units and vmlinux.o, with the latter only used for noinstr/unret validation.
 
-objtool-enabled := $(or $(delay-objtool),$(CONFIG_NOINSTR_VALIDATION))
-
-ifeq ($(delay-objtool),y)
-vmlinux-objtool-args-y					+= $(objtool-args-y)
-else
-vmlinux-objtool-args-$(CONFIG_OBJTOOL_WERROR)		+= --werror
+ifneq ($(delay-objtool),y)
+objtool-cmds-y					 =
+objtool-opts-y					+= --link
 endif
 
-vmlinux-objtool-args-$(CONFIG_NOINSTR_VALIDATION)	+= --noinstr \
-							   $(if $(or $(CONFIG_MITIGATION_UNRET_ENTRY),$(CONFIG_MITIGATION_SRSO)), --unret)
-
-objtool-args = $(vmlinux-objtool-args-y) --link
+objtool-cmds-$(CONFIG_NOINSTR_VALIDATION)	+= --noinstr \
+						   $(if $(or $(CONFIG_MITIGATION_UNRET_ENTRY),$(CONFIG_MITIGATION_SRSO)), --unret)
 
 # Link of vmlinux.o used for section mismatch analysis
 # ---------------------------------------------------------------------------

Signature · 1 line

2.53.0

Sometimes Clang arm64 code jumps to the end of the function for UB. No need to make that an error, arm64 doesn't reverse engineer the CFG anyway.

Signed-off-by Josh Poimboeuf <[email protected]>

arm64 can have empty alternatives, which are effectively no-ops. Ignore them.

Signed-off-by Josh Poimboeuf <[email protected]>

For duplicate strings, elf_add_string() just blindly adds duplicates.

That can be a problem for arm64 which often uses two consecutive instructions (and corresponding relocations) to put an address into a register, like:

  d8:   90000001        adrp    x1, 0 <meminfo_proc_show>       d8: R_AARCH64_ADR_PREL_PG_HI21  .rodata.meminfo_proc_show.str1.8
  dc:   91000021        add     x1, x1, #0x0    dc: R_AARCH64_ADD_ABS_LO12_NC   .rodata.meminfo_proc_show.str1.8

Referencing two different string addresses in the adrp+add pair can result in a corrupt string addresses. Detect such consecutive reuses and force them to use the same string.

Signed-off-by Josh Poimboeuf <[email protected]>

Add support for cross-compilation. The user must export ARCH, and either CROSS_COMPILE or LLVM.

Signed-off-by Josh Poimboeuf <[email protected]>

---
 scripts/livepatch/klp-build | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

scripts/livepatch/klp-build +0 -0

index 809e198a561d..b6c057e2120f 100755

Quoted 17 lines · Josh Poimboeuf

Add support for cross-compilation. The user must export ARCH, and either CROSS_COMPILE or LLVM.

Signed-off-by Josh Poimboeuf <[email protected]>

---
 scripts/livepatch/klp-build | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

scripts/livepatch/klp-build +0 -0

index 809e198a561d..b6c057e2120f 100755
@@ -404,6 +404,14 @@ validate_patches() {
        revert_patches
 }
 

scripts/livepatch/klp-build validate_patches() { · line 413

Quoted 7 lines

scripts/livepatch/klp-build +7 -0

+cross_compile_init() {
+       if [[ -v LLVM ]]; then
+               OBJCOPY=llvm-objcopy
+       else
+               OBJCOPY="${CROSS_COMPILE:-}objcopy"
+       fi
+}

Shall we show a specific warning if
- ARCH is set; and
- ARCH is not the same as (uname -m); and
- neither LLVM nor CROSS_COMPILE is set.

Quoted 18 lines · Song Liu

Quoted 17 lines · Josh Poimboeuf

Add support for cross-compilation. The user must export ARCH, and either CROSS_COMPILE or LLVM.

Signed-off-by Josh Poimboeuf <[email protected]>

---
 scripts/livepatch/klp-build | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

scripts/livepatch/klp-build +0 -0

index 809e198a561d..b6c057e2120f 100755
@@ -404,6 +404,14 @@ validate_patches() {
        revert_patches
 }
 

Quoted 12 lines

scripts/livepatch/klp-build validate_patches() { · line 413

Quoted 7 lines

scripts/livepatch/klp-build +7 -0

+cross_compile_init() {
+       if [[ -v LLVM ]]; then
+               OBJCOPY=llvm-objcopy
+       else
+               OBJCOPY="${CROSS_COMPILE:-}objcopy"
+       fi
+}

Shall we show a specific warning if
- ARCH is set; and
- ARCH is not the same as (uname -m); and
- neither LLVM nor CROSS_COMPILE is set.

Yeah, I think that would be a good idea. Will do that for v2.

Quoted 2 lines

Yeah, I think that would be a good idea. Will do that for v2.

So, in general ARCH is complicated. For example:

- ARCH=arm64 would never match "uname -m" (aarch64)
- ARCH=i386 would use the same gcc binary (no cross-compiler needed)
- I'm sure there are many other edge cases...

Instead of a manual error, it may be simpler to just let the build fail naturally if the user doesn't set the right ARCH.

Though, I think the check can be improved slightly, as ARCH is a reasonably good indicator that cross-compiling is happening. So I can at least add an ARCH check at the beginning like so?

cross_compile_init() {
        if [[ ! -v ARCH ]]; then
                OBJCOPY=objcopy
                return 0
        fi

        if [[ -v LLVM ]]; then
                OBJCOPY=llvm-objcopy
        else
                OBJCOPY="${CROSS_COMPILE:-}objcopy"
        fi
}

On Tue, Mar 17, 2026 at 10:52 AM Josh Poimboeuf <[email protected]> wrote:
[...]

Quoted 8 lines

Quoted 2 lines

Yeah, I think that would be a good idea. Will do that for v2.

So, in general ARCH is complicated. For example:

- ARCH=arm64 would never match "uname -m" (aarch64)
- ARCH=i386 would use the same gcc binary (no cross-compiler needed)
- I'm sure there are many other edge cases...

Agreed. I haven't worked with i386 for a long time, but I did notice the arm64 vs. aarch64 difference.

Quoted 7 lines

Instead of a manual error, it may be simpler to just let the build fail naturally if the user doesn't set the right ARCH.

Though, I think the check can be improved slightly, as ARCH is a reasonably good indicator that cross-compiling is happening. So I can at least add an ARCH check at the beginning like so?

Quoted 12 lines

cross_compile_init() {
        if [[ ! -v ARCH ]]; then
                OBJCOPY=objcopy
                return 0
        fi

        if [[ -v LLVM ]]; then
                OBJCOPY=llvm-objcopy
        else
                OBJCOPY="${CROSS_COMPILE:-}objcopy"
        fi
}

Do we need ARCH when CROSS_COMPILE is set? I was under the impression that CROSS_COMPILE doesn't require ARCH.

Quoted 4 lines

Do we need ARCH when CROSS_COMPILE is set? I was under the impression that CROSS_COMPILE doesn't require ARCH.

If CROSS_COMPILE is used without ARCH, it will just try to use the host
arch. I'm not sure if that's considered cross-compiling? I suppose it
should use the CROSS_COMPILE version of objcopy in that case? Though in practice it probably doesn't matter.

I guess the original version of the function is probably fine and we don't need to complicate matters.

Port objtool and the klp-build tooling (for building livepatch modules) to arm64.

Note this doesn't bring all the objtool bells and whistles to arm64, nor any of the CFG reverse engineering. This only adds the bare minimum needed for 'objtool --checksum'.

And note that objtool still doesn't get enabled at all for normal arm64 kernel builds, so this doesn't affect any users other than those running klp-build directly.