Skip to contentSkip to conversation

modpost: Ignore Clang LTO suffixes in symbol matching

PATCH Superseded [email protected] 3 participants 5 messages Jun 15 – Jun 17 Kbuild

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 modpost: Ignore Clang LTO suffixes in symbol matching

From: Rong Xu <[email protected]>

When building the kernel with Clang ThinLTO enabled, the compiler can mangle static variable names by appending suffixes such as ".llvm.<hash>" to prevent naming collisions across translation units.

This name mangling breaks the section mismatch whitelisting in modpost. modpost relies on glob patterns (e.g., "*_ops" or "*_probe") to identify safe references between permanent data and initialization code. Because the LTO suffix modifies the end of the symbol name, legitimately whitelisted structures fail the match, resulting in false positive warnings.

For example, a static pernet_operations struct triggers the following:

  WARNING: modpost: vmlinux: section mismatch in reference: \
  ping_v4_net_ops.llvm.5641696707737373282 (section: .data) -> \
  ping_v4_proc_init_net (section: .init.text)

Fix this by stripping ".llvm." suffixes from the symbol name in match().

---
 scripts/mod/modpost.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
scripts/mod/modpost.c +12 -0
index abbcd3fc1394..1f5a64eeb048 100644
@@ -727,6 +727,18 @@ static const char *sym_name(struct elf_info *elf, Elf_Sym *sym)
 static bool match(const char *string, const char *const patterns[])
 {
 	const char *pattern;
+	char string_stripped[512];
+	const char *ext = strstr(string, ".llvm.");
+
+	/*
+	 * Clang LTO can append .llvm.<hash> to a variable. Safely strip
+	 * the suffix so glob whitelists (like *_ops) work.
+	 */
+	if (ext && (ext - string) < sizeof(string_stripped)) {
+		strncpy(string_stripped, string, ext - string);
+		string_stripped[ext - string] = '\0';
+		string = string_stripped;
+	}
 
 	while ((pattern = *patterns++)) {
 		if (!fnmatch(pattern, string, 0))

base-commit: 2b414a95b8f7307d42173ba9e580d6d3e2bcbfce

Signature · 1 line
2.54.0.1136.gdb2ca164c4-goog
author
#1 lore raw

A more direct fix for the warnings in
https://lore.kernel.org/oe-kbuild-all/[email protected]/
looks like the following. But I believe the solution provided in the patch is more comprehensive.

-Rong

======

scripts/mod/modpost.c +0 -0
index abbcd3fc1394..1f0a379b43c8 100644
@@ -967,7 +967,7 @@ static int secref_whitelist(const char *fromsec,

const char *fromsym,

        /* symbols in data sections that may refer to any init/exit sections */
        if (match(fromsec, PATTERNS(DATA_SECTIONS)) &&
            match(tosec, PATTERNS(ALL_INIT_SECTIONS, ALL_EXIT_SECTIONS)) &&
-           match(fromsym, PATTERNS("*_ops", "*_console")))
+           match(fromsym, PATTERNS("*_ops", "*_ops.llvm.*", "*_console")))
                return 0;

        /* Check for pattern 3 */
Quoted 57 lines · On Mon, Jun 15, 2026 at 3:20 PM <[email protected]> wrote:

From: Rong Xu <[email protected]>

When building the kernel with Clang ThinLTO enabled, the compiler can mangle static variable names by appending suffixes such as ".llvm.<hash>" to prevent naming collisions across translation units.

This name mangling breaks the section mismatch whitelisting in modpost. modpost relies on glob patterns (e.g., "*_ops" or "*_probe") to identify safe references between permanent data and initialization code. Because the LTO suffix modifies the end of the symbol name, legitimately whitelisted structures fail the match, resulting in false positive warnings.

For example, a static pernet_operations struct triggers the following:

  WARNING: modpost: vmlinux: section mismatch in reference: \
  ping_v4_net_ops.llvm.5641696707737373282 (section: .data) -> \
  ping_v4_proc_init_net (section: .init.text)

Fix this by stripping ".llvm." suffixes from the symbol name in match().

---
 scripts/mod/modpost.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
scripts/mod/modpost.c +12 -0
index abbcd3fc1394..1f5a64eeb048 100644
@@ -727,6 +727,18 @@ static const char *sym_name(struct elf_info *elf, Elf_Sym *sym)
 static bool match(const char *string, const char *const patterns[])
 {
        const char *pattern;
+       char string_stripped[512];
+       const char *ext = strstr(string, ".llvm.");
+
+       /*
+        * Clang LTO can append .llvm.<hash> to a variable. Safely strip
+        * the suffix so glob whitelists (like *_ops) work.
+        */
+       if (ext && (ext - string) < sizeof(string_stripped)) {
+               strncpy(string_stripped, string, ext - string);
+               string_stripped[ext - string] = '\0';
+               string = string_stripped;
+       }
 
        while ((pattern = *patterns++)) {
                if (!fnmatch(pattern, string, 0))

base-commit: 2b414a95b8f7307d42173ba9e580d6d3e2bcbfce --
2.54.0.1136.gdb2ca164c4-goog

#2 lore raw
Quoted 21 lines · Rong Xu

A more direct fix for the warnings in
https://lore.kernel.org/oe-kbuild-all/[email protected]/
looks like the following. But I believe the solution provided in the patch is more comprehensive.

-Rong

======

scripts/mod/modpost.c +1 -1
index abbcd3fc1394..1f0a379b43c8 100644
@@ -967,7 +967,7 @@ static int secref_whitelist(const char *fromsec,
 const char *fromsym,
        /* symbols in data sections that may refer to any init/exit sections */
        if (match(fromsec, PATTERNS(DATA_SECTIONS)) &&
            match(tosec, PATTERNS(ALL_INIT_SECTIONS, ALL_EXIT_SECTIONS)) &&
-           match(fromsym, PATTERNS("*_ops", "*_console")))
+           match(fromsym, PATTERNS("*_ops", "*_ops.llvm.*", "*_console")))
                return 0;
 
Quoted 2 lines
        /* Check for pattern 3 */

This variant makes more sense to me. The initially proposed patch modifies the match() function, which is a generic function to check whether a string matches any of the input patterns. The modpost utility uses it to match both symbol and section names. Adding symbol-specific knowledge to this function seems odd.

I noticed that LLVM has recently made improvements to reduce the number of these ThinLTO renames [1], which might be worth checking out.

[1] https://github.com/llvm/llvm-project/commit/975dba28633d2f3746a8a370741b17024b0f5f9b

Signature · 2 lines
Thanks,
Petr
Quoted 58 lines · On Mon, Jun 15, 2026 at 3:20 PM <[email protected]> wrote:
Quoted 57 lines

From: Rong Xu <[email protected]>

When building the kernel with Clang ThinLTO enabled, the compiler can mangle static variable names by appending suffixes such as ".llvm.<hash>" to prevent naming collisions across translation units.

This name mangling breaks the section mismatch whitelisting in modpost. modpost relies on glob patterns (e.g., "*_ops" or "*_probe") to identify safe references between permanent data and initialization code. Because the LTO suffix modifies the end of the symbol name, legitimately whitelisted structures fail the match, resulting in false positive warnings.

For example, a static pernet_operations struct triggers the following:

  WARNING: modpost: vmlinux: section mismatch in reference: \
  ping_v4_net_ops.llvm.5641696707737373282 (section: .data) -> \
  ping_v4_proc_init_net (section: .init.text)

Fix this by stripping ".llvm." suffixes from the symbol name in match().

---
 scripts/mod/modpost.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)
scripts/mod/modpost.c +12 -0
index abbcd3fc1394..1f5a64eeb048 100644
@@ -727,6 +727,18 @@ static const char *sym_name(struct elf_info *elf, Elf_Sym *sym)
 static bool match(const char *string, const char *const patterns[])
 {
        const char *pattern;
+       char string_stripped[512];
+       const char *ext = strstr(string, ".llvm.");
+
+       /*
+        * Clang LTO can append .llvm.<hash> to a variable. Safely strip
+        * the suffix so glob whitelists (like *_ops) work.
+        */
+       if (ext && (ext - string) < sizeof(string_stripped)) {
+               strncpy(string_stripped, string, ext - string);
+               string_stripped[ext - string] = '\0';
+               string = string_stripped;
+       }
 
        while ((pattern = *patterns++)) {
                if (!fnmatch(pattern, string, 0))

base-commit: 2b414a95b8f7307d42173ba9e580d6d3e2bcbfce --
2.54.0.1136.gdb2ca164c4-goog

#3 lore raw
Quoted 22 lines · Petr Pavlu
Quoted 21 lines · Rong Xu

A more direct fix for the warnings in
https://lore.kernel.org/oe-kbuild-all/[email protected]/
looks like the following. But I believe the solution provided in the patch is more comprehensive.

-Rong

======

scripts/mod/modpost.c +1 -1
index abbcd3fc1394..1f0a379b43c8 100644
@@ -967,7 +967,7 @@ static int secref_whitelist(const char *fromsec,
 const char *fromsym,
        /* symbols in data sections that may refer to any init/exit sections */
        if (match(fromsec, PATTERNS(DATA_SECTIONS)) &&
            match(tosec, PATTERNS(ALL_INIT_SECTIONS, ALL_EXIT_SECTIONS)) &&
-           match(fromsym, PATTERNS("*_ops", "*_console")))
+           match(fromsym, PATTERNS("*_ops", "*_ops.llvm.*", "*_console")))
                return 0;
 
Quoted 8 lines
Quoted 2 lines
        /* Check for pattern 3 */

This variant makes more sense to me. The initially proposed patch modifies the match() function, which is a generic function to check whether a string matches any of the input patterns. The modpost utility uses it to match both symbol and section names. Adding symbol-specific knowledge to this function seems odd.

Yeah, I had the feeling that it was a bit of a layering violation as well. I think I would prefer this targeted fix for now since I will take it as a fix for 7.2 in my planned second Kbuild pull request. We can revisit the more comprehensive fix if it is needed in the future (while seeing if we can address it outside of match() regardless).

Quoted 4 lines

I noticed that LLVM has recently made improvements to reduce the number of these ThinLTO renames [1], which might be worth checking out.

[1] https://github.com/llvm/llvm-project/commit/975dba28633d2f3746a8a370741b17024b0f5f9b

Indeed, we have started using that for regular ThinLTO when it is available as of commit dc3b90751d6f ("kbuild: Reduce the number of compiler-generated suffixes for clang thin-lto build"). Distributed ThinLTO support will follow when LLVM 23.1.0 is released.

Signature · 2 lines
Cheers,
Nathan
author
#4 lore raw

Petr and Nathan, thanks for the review and suggestions.

I will send v2 shortly to implement the more targeted approach.

Best regards,

Rong

Quoted 52 lines · Nathan Chancellor
Quoted 22 lines · Petr Pavlu
Quoted 21 lines · Rong Xu

A more direct fix for the warnings in
https://lore.kernel.org/oe-kbuild-all/[email protected]/
looks like the following. But I believe the solution provided in the patch is more comprehensive.

-Rong

======

scripts/mod/modpost.c +1 -1
index abbcd3fc1394..1f0a379b43c8 100644
@@ -967,7 +967,7 @@ static int secref_whitelist(const char *fromsec,
 const char *fromsym,
        /* symbols in data sections that may refer to any init/exit sections */
        if (match(fromsec, PATTERNS(DATA_SECTIONS)) &&
            match(tosec, PATTERNS(ALL_INIT_SECTIONS, ALL_EXIT_SECTIONS)) &&
-           match(fromsym, PATTERNS("*_ops", "*_console")))
+           match(fromsym, PATTERNS("*_ops", "*_ops.llvm.*", "*_console")))
                return 0;
 
Quoted 8 lines
Quoted 2 lines
        /* Check for pattern 3 */

This variant makes more sense to me. The initially proposed patch modifies the match() function, which is a generic function to check whether a string matches any of the input patterns. The modpost utility uses it to match both symbol and section names. Adding symbol-specific knowledge to this function seems odd.

Yeah, I had the feeling that it was a bit of a layering violation as well. I think I would prefer this targeted fix for now since I will take it as a fix for 7.2 in my planned second Kbuild pull request. We can revisit the more comprehensive fix if it is needed in the future (while seeing if we can address it outside of match() regardless).

Quoted 4 lines

I noticed that LLVM has recently made improvements to reduce the number of these ThinLTO renames [1], which might be worth checking out.

[1] https://github.com/llvm/llvm-project/commit/975dba28633d2f3746a8a370741b17024b0f5f9b

Indeed, we have started using that for regular ThinLTO when it is available as of commit dc3b90751d6f ("kbuild: Reduce the number of compiler-generated suffixes for clang thin-lto build"). Distributed ThinLTO support will follow when LLVM 23.1.0 is released.

--
Cheers,
Nathan