Skip to contentSkip to conversation

libbpf: fix cross-endian build with newer libbpf

PATCH Superseded Dominique Martinet 4 participants 8 messages Nov 15 – Nov 16

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 libbpf: fix cross-endian build with newer libbpf

commit 087f3964f4b3 ("libbpf: Support layout section handling in BTF") added extra data to the BTF header.

This extra data is ignored in the native case, but when building with different endianness the bswap code in old libbpf would check for strict size equality, leading to build failures such as this when BTF data is generated with a newer version of libbpf (through pahole)
BTFIDS vmlinux.unstripped
libbpf: Can't load BTF with non-native endianness due to unsupported header length 32
FAILED: load BTF from vmlinux.unstripped: Operation not supported make[2]: *** [../scripts/Makefile.vmlinux:72: vmlinux.unstripped] Error 255

Newer libbpf versions also properly ignore trailing data in the bswap case, but the patch is too intrusive to backport, just fix the header length comparison.

---
Hey,

we've got a cross build failure report for ppc64 in nixpkgs (link in commit message), that turned out as described above. (build broke for all kernels before 7.1)

In this case I believe it's easier to just fix the check than backport the full layout section handling support, is that ok?
I've confirmed the build succeeds with this, I assume the reporter will eventually also boot check but from my understanding of the code this should be fine.

Feel free to throw this away if there's a better way, just let me know if I can help.

Thanks

 tools/lib/bpf/btf.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/lib/bpf/btf.c +1 -1
index 9f141395c074..ae344cf58e85 100644
@@ -236,7 +236,7 @@ static int btf_parse_hdr(struct btf *btf)
 
 	if (hdr->magic == bswap_16(BTF_MAGIC)) {
 		btf->swapped_endian = true;
-		if (bswap_32(hdr->hdr_len) != sizeof(struct btf_header)) {
+		if (bswap_32(hdr->hdr_len) < sizeof(struct btf_header)) {
 			pr_warn("Can't load BTF with non-native endianness due to unsupported header length %u\n",
 				bswap_32(hdr->hdr_len));
 			return -ENOTSUP;
Signature · 1 line
2.56.1
#1 lore raw
Thank you for your contribution! Sashiko AI review found 2 potential issues to consider: - [Low] libbpf: missing Fixes trailer for btf_parse_hdr() endianness check - [Low] libbpf: out-of-bounds…

Thank you for your contribution! Sashiko AI review found 2 potential issues to consider:
- [Low] libbpf: missing Fixes trailer for btf_parse_hdr() endianness check
- [Low] libbpf: out-of-bounds access in btf_parse_hdr() when hdr_len > sizeof(*hdr)
--

commit c3e5e38661056ceb32071ab50ba595824eefbdbf
Author: Dominique Martinet <[email protected]>

Quoted 14 lines

libbpf: fix cross-endian build with newer libbpf

commit 087f3964f4b3 ("libbpf: Support layout section handling in BTF") added extra data to the BTF header.

This extra data is ignored in the native case, but when building with different endianness the bswap code in old libbpf would check for strict size equality, leading to build failures such as this when BTF data is generated with a newer version of libbpf (through pahole)
BTFIDS vmlinux.unstripped
libbpf: Can't load BTF with non-native endianness due to unsupported header length 32
FAILED: load BTF from vmlinux.unstripped: Operation not supported make[2]: *** [../scripts/Makefile.vmlinux:72: vmlinux.unstripped] Error 255

Quoted 6 lines

Newer libbpf versions also properly ignore trailing data in the bswap case, but the patch is too intrusive to backport, just fix the header length comparison.

[Severity: Low]
Should this include a Fixes tag for the commit that introduced the strict swapped-endian header size check:

Fixes 3289959b97ca ("libbpf: Support BTF loading and raw data output in both endianness")

tools/lib/bpf/btf.c static int btf_parse_hdr(struct btf *btf) · line 204

Quoted 13 lines
tools/lib/bpf/btf.c +1 -1
index 231b07203e3d2..8f3f942eda704 100644
@@ -198,7 +198,7 @@ static int btf_parse_hdr(struct btf *btf)
 
 	if (hdr->magic == bswap_16(BTF_MAGIC)) {
 		btf->swapped_endian = true;
-		if (bswap_32(hdr->hdr_len) != sizeof(struct btf_header)) {
+		if (bswap_32(hdr->hdr_len) < sizeof(struct btf_header)) {
 			pr_warn("Can't load BTF with non-native endianness due to unsupported header length %u\n",
 				bswap_32(hdr->hdr_len));
 			return -ENOTSUP;

[Severity: Low]
Can relaxing this check cause out-of-bounds reads and writes on btf->raw_data when parsing a malformed cross-endian BTF buffer with hdr->hdr_len > sizeof(*hdr)?

Previously, requiring hdr->hdr_len == sizeof(struct btf_header) together with btf->raw_size >= sizeof(struct btf_header) guaranteed hdr->hdr_len fit within btf->raw_size. btf_parse_hdr() still computes meta_left from sizeof(*hdr) instead of hdr->hdr_len:

tools/lib/bpf/btf.c:btf_parse_hdr() {
	...
	meta_left = btf->raw_size - sizeof(*hdr);
	...
}

However, btf__new() offsets the section pointers by btf->hdr->hdr_len:

tools/lib/bpf/btf.c:btf__new() {
	...
	btf->strs_data = btf->raw_data + btf->hdr->hdr_len + btf->hdr->str_off;
	btf->types_data = btf->raw_data + btf->hdr->hdr_len + btf->hdr->type_off;
	...
}

If hdr->hdr_len exceeds sizeof(*hdr) (or exceeds btf->raw_size), btf->strs_data and btf->types_data can point past btf->raw_data + btf->raw_size, causing btf_parse_str_sec() to read out of bounds and btf_parse_type_sec() to byte-swap memory in place past the end of the heap buffer. Checking that btf->raw_size >= hdr->hdr_len and subtracting hdr->hdr_len when computing meta_left would avoid this.

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#2 lore raw

[email protected] wrote on Sun, Nov 15, 2026 at 08:23:16PM +0000:

Quoted 8 lines
Quoted 2 lines

[Severity: Low]
Should this include a Fixes tag for the commit that introduced the strict swapped-endian header size check:

Fixes 3289959b97ca ("libbpf: Support BTF loading and raw data output in both endianness")

Err, yeah, I guess that bug couldn't happen before support for endianness was added, but it didn't work before either so meh :) Feel free to add when applying if you think it helps.

Quoted 24 lines

tools/lib/bpf/btf.c static int btf_parse_hdr(struct btf *btf) · line 204

Quoted 13 lines
tools/lib/bpf/btf.c +1 -1
index 231b07203e3d2..8f3f942eda704 100644
@@ -198,7 +198,7 @@ static int btf_parse_hdr(struct btf *btf)
 
 	if (hdr->magic == bswap_16(BTF_MAGIC)) {
 		btf->swapped_endian = true;
-		if (bswap_32(hdr->hdr_len) != sizeof(struct btf_header)) {
+		if (bswap_32(hdr->hdr_len) < sizeof(struct btf_header)) {
 			pr_warn("Can't load BTF with non-native endianness due to unsupported header length %u\n",
 				bswap_32(hdr->hdr_len));
 			return -ENOTSUP;

[Severity: Low]
Can relaxing this check cause out-of-bounds reads and writes on btf->raw_data when parsing a malformed cross-endian BTF buffer with hdr->hdr_len > sizeof(*hdr)?

Previously, requiring hdr->hdr_len == sizeof(struct btf_header) together with btf->raw_size >= sizeof(struct btf_header) guaranteed hdr->hdr_len fit within btf->raw_size. btf_parse_hdr() still computes meta_left from sizeof(*hdr) instead of hdr->hdr_len:

Quoted 4 lines
tools/lib/bpf/btf.c:btf_parse_hdr() {
	...
	meta_left = btf->raw_size - sizeof(*hdr);

Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1

sashiko apparently picked the v5.10 tag, but this bad sizeof was fixed in commit c825f5fee19c ("libbpf: Fix BTF header parsing checks") and isn't a problem as far as I can tell (that commit was backported to all stable kernels, and header is properly used, after endianness swap)

Thanks,

Signature · 1 line
Dominique Martinet | Asmadeus
#3 lore raw
In reply to Dominique Martinet
Quoted 44 lines · Dominique Martinet

commit 087f3964f4b3 ("libbpf: Support layout section handling in BTF") added extra data to the BTF header.

This extra data is ignored in the native case, but when building with different endianness the bswap code in old libbpf would check for strict size equality, leading to build failures such as this when BTF data is generated with a newer version of libbpf (through pahole)
BTFIDS vmlinux.unstripped
libbpf: Can't load BTF with non-native endianness due to unsupported header length 32
FAILED: load BTF from vmlinux.unstripped: Operation not supported make[2]: *** [../scripts/Makefile.vmlinux:72: vmlinux.unstripped] Error 255

Newer libbpf versions also properly ignore trailing data in the bswap case, but the patch is too intrusive to backport, just fix the header length comparison.

---
Hey,

we've got a cross build failure report for ppc64 in nixpkgs (link in commit message), that turned out as described above. (build broke for all kernels before 7.1)

In this case I believe it's easier to just fix the check than backport the full layout section handling support, is that ok?
I've confirmed the build succeeds with this, I assume the reporter will eventually also boot check but from my understanding of the code this should be fine.

Feel free to throw this away if there's a better way, just let me know if I can help.

Thanks

 tools/lib/bpf/btf.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/lib/bpf/btf.c +0 -0
index 9f141395c074..ae344cf58e85 100644

tools/lib/bpf/btf.c static int btf_parse_hdr(struct btf *btf) · line 239

Quoted 5 lines
tools/lib/bpf/btf.c +1 -1
@@ -236,7 +236,7 @@ static int btf_parse_hdr(struct btf *btf)
 
 	if (hdr->magic == bswap_16(BTF_MAGIC)) {
 		btf->swapped_endian = true;
-		if (bswap_32(hdr->hdr_len) != sizeof(struct btf_header)) {
+		if (bswap_32(hdr->hdr_len) < sizeof(struct btf_header)) {

Thanks for the fix! A small suggestion; what about matching the upstream test instead; i.e.

		if (bswap_32(hdr->hdr_len) < offsetofend(struct btf_header, str_len)) {

That approach is a bit more portable as the str_len field was the last field in the original BTF header, and it will work regardless of the size of struct btf_header.

Quoted 3 lines
tools/lib/bpf/btf.c +0 -0
 			pr_warn("Can't load BTF with non-native endianness due to unsupported header length %u\n",
 				bswap_32(hdr->hdr_len));
 			return -ENOTSUP;
#4 lore raw

Alan Maguire wrote on Mon, Nov 16, 2026 at 08:51:24AM +0000:

Quoted 11 lines

Review

Quoted 4 lines
(quoted diff) +1 -1
 	if (hdr->magic == bswap_16(BTF_MAGIC)) {
 		btf->swapped_endian = true;
-		if (bswap_32(hdr->hdr_len) != sizeof(struct btf_header)) {
+		if (bswap_32(hdr->hdr_len) < sizeof(struct btf_header)) {

Thanks for the fix! A small suggestion; what about matching the upstream test instead; i.e.

		if (bswap_32(hdr->hdr_len) < offsetofend(struct btf_header, str_len)) {

That approach is a bit more portable as the str_len field was the last field in the original BTF header, and it will work regardless of the size of struct btf_header.

Thanks for the look!

I wouldn't expect the struct to change in stable trees and favored a minimal patch; my opinion being that if someone does take the time to backport layout section handling or another btf_header extension in old trees then they can also fix the check properly...

I actually prepared a v2 with this, but then noticed the raw_size check also ought to be updated if you want to handle struct btf_header changes:
---

tools/lib/bpf/btf.c +2 -2
index 9f141395c074..0ab11be8ebc5 100644
@@ -229,14 +229,14 @@ static int btf_parse_hdr(struct btf *btf)
 	struct btf_header *hdr = btf->hdr;
 	__u32 meta_left;
 
-	if (btf->raw_size < sizeof(struct btf_header)) {
+	if (btf->raw_size < offsetofend(struct btf_header, str_len)) {
 		pr_debug("BTF header not found\n");
 		return -EINVAL;
 	}
 
 	if (hdr->magic == bswap_16(BTF_MAGIC)) {
 		btf->swapped_endian = true;
-		if (bswap_32(hdr->hdr_len) != sizeof(struct btf_header)) {
+		if (bswap_32(hdr->hdr_len) < offsetofend(struct btf_header, str_len)) {
 			pr_warn("Can't load BTF with non-native endianness due to unsupported header length %u\n",
 				bswap_32(hdr->hdr_len));
 			return -ENOTSUP;

---
and at this point this isn't just a cross-endian fix anymore but something that looks more like future-proofing, and with my other downstream stable kernel maintainer hat (yay embedded systems full of vendor patches...), I'm not too fond of these and would rather keep this minimal as it's not like it would reduce further merge conflicts either.

But if you think it's better I can send a v2 with either the above or just the bswap_32 check updated, I've checked it builds on 5.10 and 6.18 so have no fundamental problem with either version.

Cheers,

Signature · 1 line
Dominique Martinet | Asmadeus
#5 lore raw
Thank you for your contribution! Sashiko AI review found 2 potential issues to consider: - [High] libbpf: missing Signed-off-by trailer and raw email reply as commit log - [Low] libbpf: missing…

Thank you for your contribution! Sashiko AI review found 2 potential issues to consider:
- [High] libbpf: missing Signed-off-by trailer and raw email reply as commit log
- [Low] libbpf: missing Fixes trailer for btf_parse_hdr() endianness check
--

commit bf42d3c49f2ac6667938fafec9c2e11053665406
Author: Dominique Martinet <[email protected]>

Quoted 17 lines

libbpf: fix cross-endian build with newer libbpf

Alan Maguire wrote on Mon, Nov 16, 2026 at 08:51:24AM +0000:

Quoted 11 lines

Review

Quoted 4 lines
    if (hdr->magic == bswap_16(BTF_MAGIC)) {
            btf->swapped_endian = true;
(quoted diff) +1 -1
-           if (bswap_32(hdr->hdr_len) != sizeof(struct btf_header)) {
+           if (bswap_32(hdr->hdr_len) < sizeof(struct btf_header)) {

Thanks for the fix! A small suggestion; what about matching the upstream test instead; i.e.

              if (bswap_32(hdr->hdr_len) < offsetofend(struct btf_header, str_len)) {

That approach is a bit more portable as the str_len field was the last field in the original BTF header, and it will work regardless of the size of struct btf_header.

Thanks for the look!

Quoted 7 lines

I wouldn't expect the struct to change in stable trees and favored a minimal patch; my opinion being that if someone does take the time to backport layout section handling or another btf_header extension in old trees then they can also fix the check properly...

I actually prepared a v2 with this, but then noticed the raw_size check also ought to be updated if you want to handle struct btf_header changes:

[Severity: High]
The commit message is an unedited mailing-list reply that ends mid-sentence rather than a standalone changelog description, and it is missing a Signed-off-by trailer.

Quoted 8 lines
tools/lib/bpf/btf.c +0 -0
index 231b07203e3d2..2dddee4870343 100644
@@ -191,14 +191,14 @@ static int btf_parse_hdr(struct btf *btf)
 	struct btf_header *hdr = btf->hdr;
 	__u32 meta_left;
 

tools/lib/bpf/btf.c static int btf_parse_hdr(struct btf *btf) · line 201

Quoted 10 lines
tools/lib/bpf/btf.c +2 -2
-	if (btf->raw_size < sizeof(struct btf_header)) {
+	if (btf->raw_size < offsetofend(struct btf_header, str_len)) {
 		pr_debug("BTF header not found\n");
 		return -EINVAL;
 	}
 
 	if (hdr->magic == bswap_16(BTF_MAGIC)) {
 		btf->swapped_endian = true;
-		if (bswap_32(hdr->hdr_len) != sizeof(struct btf_header)) {
+		if (bswap_32(hdr->hdr_len) < offsetofend(struct btf_header, str_len)) {

[Severity: Low]
This fixes the strict cross-endian header length check in btf_parse_hdr(). Should the commit message include:

Fixes 3289959b97ca ("libbpf: Support BTF loading and raw data output in both endianness")
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
Acked-by
#6 lore raw
In reply to Dominique Martinet
Quoted 51 lines · Dominique Martinet

Alan Maguire wrote on Mon, Nov 16, 2026 at 08:51:24AM +0000:

Quoted 11 lines

tools/lib/bpf/btf.c static int btf_parse_hdr(struct btf *btf)

Quoted 4 lines
tools/lib/bpf/btf.c +1 -1
 	if (hdr->magic == bswap_16(BTF_MAGIC)) {
 		btf->swapped_endian = true;
-		if (bswap_32(hdr->hdr_len) != sizeof(struct btf_header)) {
+		if (bswap_32(hdr->hdr_len) < sizeof(struct btf_header)) {

Thanks for the fix! A small suggestion; what about matching the upstream test instead; i.e.

		if (bswap_32(hdr->hdr_len) < offsetofend(struct btf_header, str_len)) {

That approach is a bit more portable as the str_len field was the last field in the original BTF header, and it will work regardless of the size of struct btf_header.

Thanks for the look!

I wouldn't expect the struct to change in stable trees and favored a minimal patch; my opinion being that if someone does take the time to backport layout section handling or another btf_header extension in old trees then they can also fix the check properly...

I actually prepared a v2 with this, but then noticed the raw_size check also ought to be updated if you want to handle struct btf_header changes:
---

tools/lib/bpf/btf.c +2 -3
index 9f141395c074..0ab11be8ebc5 100644
@@ -229,14 +229,14 @@ static int btf_parse_hdr(struct btf *btf)
 	struct btf_header *hdr = btf->hdr;
 	__u32 meta_left;
 
-	if (btf->raw_size < sizeof(struct btf_header)) {
+	if (btf->raw_size < offsetofend(struct btf_header, str_len)) {
 		pr_debug("BTF header not found\n");
 		return -EINVAL;
 	}
 
 	if (hdr->magic == bswap_16(BTF_MAGIC)) {
 		btf->swapped_endian = true;
-		if (bswap_32(hdr->hdr_len) != sizeof(struct btf_header)) {
+		if (bswap_32(hdr->hdr_len) < offsetofend(struct btf_header, str_len)) {
 			pr_warn("Can't load BTF with non-native endianness due to unsupported header length %u\n",
 				bswap_32(hdr->hdr_len));
 			return -ENOTSUP;
---

and at this point this isn't just a cross-endian fix anymore but something that looks more like future-proofing, and with my other downstream stable kernel maintainer hat (yay embedded systems full of vendor patches...), I'm not too fond of these and would rather keep this minimal as it's not like it would reduce further merge conflicts either.

Quoted 4 lines

But if you think it's better I can send a v2 with either the above or just the bswap_32 check updated, I've checked it builds on 5.10 and 6.18 so have no fundamental problem with either version.

The only risk with relying on struct btf_header size is if the libbpf is compiled against system headers which for some reason have a newer "struct btf"; that shouldn't happen, and these days the standalone libbpf mirror carries its own uAPI headers, but just in case someone somewhere is doing something unusual in building libbpf it might be best to use the v2 approach. So for v2

Thanks again!

Quoted 1 line

Cheers,

maintainer
#7 lore raw
In reply to Dominique Martinet
Quoted 3 lines

Newer libbpf versions also properly ignore trailing data in the bswap case, but the patch is too intrusive to backport, just fix the header length comparison.

I'm holding this until there is either an ack from Alan on v1, or a v2.

Signature · 2 lines
Thanks,
Sasha