selftests/bpf: Fix csum_partial() dropping trailing byte on odd length
Subsystems (3)
PATCH v3 1/1 selftests/bpf: Fix csum_partial() dropping trailing byte on odd length
csum_partial() computes num_u16 = len >> 1 and only sums that many 16-bit words, so the last byte of an odd-length buffer never gets added to the checksum. RFC 1071 says it should be padded with a zero byte and summed as one more word, not dropped.
This backs build_ip_csum(), build_udp_v4_csum() and build_udp_v6_csum(), used by flow_dissector_classification.c and xdp_metadata.c to hand-build packets. No current caller builds an odd-length payload, so nothing fails today, but a future one would get a silently wrong checksum.
Also bump flow_dissector_classification's TEST_PACKET_LEN from 100 to 99 so this actually gets exercised instead of staying latent.
f4504af68575 wrote the len >> 1 division, but it only ever passed sizeof(iphdr), always even, so it couldn't hit the bug. Tagging bcc00987bc56 instead, since it added the first caller, build_udp_v4_csum()/build_udp_v6_csum(), that can pass an odd length.
Verified with test_progs under vmtest.sh: without the fix, an odd TEST_PACKET_LEN makes the kernel drop the packet over a bad checksum and flow_dissector_classification fails; with the fix, both flow_dissector_classification and xdp_metadata pass.
---
v2 -> v3:
- Move the revision changelog and sample test output below --- so they don't end up in the permanent commit history (Alexis)
- Drop the Link tag; one is added automatically on merge (Alexis)
- Paste a passing run too, not just the failure (Alexis)
- Also run xdp_metadata, since it uses the same helpers (Alexis)
v1 -> v2:
- comment style: opening /* on its own line, per BPF selftests style
- Retarget the Fixes tag to bcc00987bc56. commit f4504af68575 ("selftests/bpf: move ip checksum helper to network helpers") moved the helper, but sizeof(iphdr) is always a multiple of 32 bit words / 4 Bytes (iph->ihl counts in 4-byte words), so the odd-length path was never reachable through build_ip_csum(). csum_partial() first gets called with a length that isn't guaranteed even in bcc00987bc56, via build_udp_v4_csum()/build_udp_v6_csum().
- TEST_PACKET_LEN 100 -> 99 so an existing test catches this instead of the bug staying unexercised
Without the fix and with TEST_PACKET_LEN=99, flow_dissector_classification fails under vmtest.sh:
test_flow_dissector_classification:FAIL:test third port unexpected test third port: actual 0 != expected 10
#137/6 flow_dissector_classification/ipv6:FAIL
#137 flow_dissector_classification:FAIL
Summary: 1/0 PASSED, 0 SKIPPED, 1/6 FAILED
With the fix, ./test_progs -t flow_dissector_classification passes under vmtest.sh:
#137/1 flow_dissector_classification/ipv4:OK
#137/2 flow_dissector_classification/ipv4_continue_dissect:OK
#137/3 flow_dissector_classification/ipip:OK
#137/4 flow_dissector_classification/gre:OK
#137/5 flow_dissector_classification/port_range:OK
#137/6 flow_dissector_classification/ipv6:OK
#137 flow_dissector_classification:OKSummary: 1/6 PASSED, 0 SKIPPED, 0/0 FAILED
TEST_PACKET_LEN is 99 by default now, so no need to force it anymore. Ran alongside xdp_metadata too, since it's the other user of build_udp_v4_csum()/build_udp_v6_csum():
#137/1 flow_dissector_classification/ipv4:OK
#137/2 flow_dissector_classification/ipv4_continue_dissect:OK
#137/3 flow_dissector_classification/ipip:OK
#137/4 flow_dissector_classification/gre:OK
#137/5 flow_dissector_classification/port_range:OK
#137/6 flow_dissector_classification/ipv6:OK
#137 flow_dissector_classification:OK
#755 xdp_metadata:OK
Summary: 2/6 PASSED, 0 SKIPPED, 0/0 FAILED
---
tools/testing/selftests/bpf/network_helpers.h | 15 +++++++++++++--
.../prog_tests/flow_dissector_classification.c | 2 +-
2 files changed, 14 insertions(+), 3 deletions(-)tools/testing/selftests/bpf/network_helpers.h +13 -2
@@ -129,12 +129,23 @@ static __u16 csum_fold(__u32 csum)static __wsum csum_partial (const void *buf,int len,__wsum sum) { -__u16 *p = (__u16 *)buf; +const __u8 *p = buf;int num_u16 = len >>1 ;int i;for (i =0 ; i < num_u16; i++) - sum += p[i]; + sum += ((const __u16 *)p)[i]; + +/* +* RFC 1071: an odd-length buffer's trailing byte is paired with +* a zero pad byte to form the final 16-bit word. +*/ +if (len &1 ) { +__u16 tail =0 ; + +__builtin_memcpy (&tail, p + len -1 ,1 ); + sum += tail; + }return sum; }
tools/testing/selftests/bpf/prog_tests/flow_dissector_classification.c +1 -1
@@ -27,7 +27,7 @@#define TEST_NAME_MAX_LEN (32 + SUBTEST_NAME_MAX_LEN)#define MAX_SOURCE_PORTS 3#define TEST_PACKETS_COUNT 10 -#define TEST_PACKET_LEN 100 +#define TEST_PACKET_LEN 99#define TEST_PACKET_PATTERN 'a'#define TEST_IPV4 "192.168.0.1/32"#define TEST_IPV6 "100::a/128"
Signature · 1 line
2.55.0
[email protected]
AI reviewShould the Fixes tag target f4504af68575 instead of bcc00987bc56?
Quoted 19 lines
csum_partial()computesnum_u16= len >> 1 and only sums that many 16-bit words, so the last byte of an odd-length buffer never gets added to the checksum. RFC 1071 says it should be padded with a zero byte and summed as one more word, not dropped.This backs
build_ip_csum(),build_udp_v4_csum()andbuild_udp_v6_csum(), used byflow_dissector_classification.candxdp_metadata.cto hand-build packets. No current caller builds an odd-length payload, so nothing fails today, but a future one would get a silently wrong checksum.Also bump flow_dissector_classification's
TEST_PACKET_LENfrom 100 to 99 so this actually gets exercised instead of staying latent.
f4504af68575wrote the len >> 1 division, but it only ever passed sizeof(iphdr), always even, so it couldn't hit the bug. Taggingbcc00987bc56instead, since it added the first caller, build_udp_v4_csum()/build_udp_v6_csum(), that can pass an odd length.
Quoted 7 lines
Verified with
test_progsundervmtest.sh: without the fix, an oddTEST_PACKET_LENmakes the kernel drop the packet over a bad checksum andflow_dissector_classificationfails; with the fix, bothflow_dissector_classificationandxdp_metadatapass.Fixes bcc00987bc56 ("selftests/bpf: add network helpers to generate udp checksums")Signed-off-by Madhav Khosla <[email protected]>
Should the Fixes tag target f4504af68575 instead of bcc00987bc56?
The tag currently points to bcc00987bc56, but that commit never touched csum_partial(). Looking at the commit:
git show --stat bcc00987bc56
shows it adds 42 lines to network_helpers.h consisting only of csum_tcpudp_magic(), csum_ipv6_magic(), build_udp_v4_csum() and build_udp_v6_csum().
The buggy code being fixed here was introduced by f4504af68575 ("selftests/bpf: move ip checksum helper to network helpers"), whose diff adds:
+static __wsum csum_partial(const void *buf, int len, __wsum sum)
+{
+ __u16 *p = (__u16 *)buf;
+ int num_u16 = len >> 1;The commit message's rationale for retargeting the tag is also self-undermining. It says bcc00987bc56 "added the first caller ... that *can* pass an odd length", while the paragraph above states "No current caller builds an odd-length payload, so nothing fails today".
As of every released kernel, no caller of build_udp_v4_csum() or build_udp_v6_csum() passes an odd length. Before this patch TEST_PACKET_LEN was 100, so udph->len was 108, and xdp_metadata.c only ever passes iph->ihl << 2 == 20. The odd length is introduced by this very patch.
So either the tag belongs on f4504af68575, which actually wrote the len >> 1 truncation, or the bug is latent-only and there should be no Fixes tag at all. The tag as written points at a commit that neither contains the defect nor exposes it.
---
AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md
CI run summary: https://github.com/kernel-patches/bpf/actions/runs/35501508732
This patch was applied to bpf/bpf-next.git (master) by Alexei Starovoitov <[email protected]>:
Hello:
This patch was applied to bpf/bpf-next.git (master) by Alexei Starovoitov <[email protected]>:
Quoted 5 lines · On Sun, 20 Sep 2026 14:25:49 +0530 you wrote:
csum_partial()computesnum_u16= len >> 1 and only sums that many 16-bit words, so the last byte of an odd-length buffer never gets added to the checksum. RFC 1071 says it should be padded with a zero byte and summed as one more word, not dropped.
Quoted 7 lines
This backs
build_ip_csum(),build_udp_v4_csum()andbuild_udp_v6_csum(), used byflow_dissector_classification.candxdp_metadata.cto hand-build packets. No current caller builds an odd-length payload, so nothing fails today, but a future one would get a silently wrong checksum.[...]
Here is the summary with links:
- [bpf-next,v3] selftests/bpf: Fix csum_partial() dropping trailing byte on odd length
https://git.kernel.org/bpf/bpf-next/c/ddaa827dcea4
You are awesome, thank you!
Signature · 2 lines
Deet-doot-dot, I am a bot. https://korg.docs.kernel.org/patchwork/pwbot.html