Summary
The tc_edt selftest fails intermittently on x86_64 test_progs ASAN runs because its 2% rate error tolerance is insufficient when AddressSanitizer instrumentation is active. ASAN adds overhead to every memory access, introducing timing jitter that pushes the measured throughput error above the 2% threshold.
Failure Details
- Test / Component: tc_edt (test_progs ASAN on x86_64)
- Frequency: occasional — observed in "Simplify min()/max()" PR (run 30778181064), independent of the patch content
- Failure mode: flaky assertion failure
- Affected architectures: x86_64 (ASAN only); already denylisted on s390x for the same timing sensitivity
- CI runs observed:
Root Cause Analysis
The test (tools/testing/selftests/bpf/prog_tests/test_tc_edt.c) works by:
- Setting up two network namespaces connected by a veth pair
- Attaching a BPF TC program that shapes traffic to 5 Mbps (via EDT timestamps)
- Sending 1 MB of data and measuring elapsed time
- Computing
rate_error = |measured_rate - 5.0| / 5.0 * 100
- Asserting
rate_error <= 2.0 (line 126)
Under ASAN, malloc/free/memcpy instrumentation adds ~2-5x overhead to userspace operations including the send()/recv() loop in send_recv_data(). While the BPF shaping runs in kernel space (unaffected by ASAN), the userspace timestamp measurements (get_time_ns() at lines 116/118) include ASAN-slowed syscall entry/exit paths. This can push the measured rate slightly away from the true shaped rate, causing marginal failures like the observed "actual 2 > expected 2".
The test is already denylisted on s390x (DENYLIST.s390x) for the same fundamental reason — timing-sensitive rate measurements are unreliable in environments with non-trivial overhead.
Proposed Fix
Add tc_edt to DENYLIST.asan in the CI configuration. This is the same approach used for s390x. The test remains fully exercised in non-ASAN x86_64/aarch64 runs where timing is reliable.
See attached patch: 0001-selftests-bpf-Add-tc_edt-to-CI-ASAN-DENYLIST.patch
Impact
Without the fix, tc_edt occasionally causes false CI failures on ASAN runs, adding noise and forcing re-runs of PRs that don't touch networking code.
References
tools/testing/selftests/bpf/prog_tests/test_tc_edt.c — test implementationgithub/kernel-patches/vmtest/ci/vmtest/configs/DENYLIST.s390x — already denylists tc_edtgithub/kernel-patches/vmtest/ci/vmtest/configs/DENYLIST.asan — target for this fix- Commit b0f82e7ab6fb ("selftests/bpf: integrate test_tc_edt into test_progs") — test introduction
Summary
The
tc_edtselftest fails intermittently on x86_64 test_progs ASAN runs because its 2% rate error tolerance is insufficient when AddressSanitizer instrumentation is active. ASAN adds overhead to every memory access, introducing timing jitter that pushes the measured throughput error above the 2% threshold.Failure Details
Root Cause Analysis
The test (
tools/testing/selftests/bpf/prog_tests/test_tc_edt.c) works by:rate_error = |measured_rate - 5.0| / 5.0 * 100rate_error <= 2.0(line 126)Under ASAN,
malloc/free/memcpy instrumentation adds ~2-5x overhead to userspace operations including thesend()/recv()loop insend_recv_data(). While the BPF shaping runs in kernel space (unaffected by ASAN), the userspace timestamp measurements (get_time_ns()at lines 116/118) include ASAN-slowed syscall entry/exit paths. This can push the measured rate slightly away from the true shaped rate, causing marginal failures like the observed "actual 2 > expected 2".The test is already denylisted on s390x (
DENYLIST.s390x) for the same fundamental reason — timing-sensitive rate measurements are unreliable in environments with non-trivial overhead.Proposed Fix
Add
tc_edttoDENYLIST.asanin the CI configuration. This is the same approach used for s390x. The test remains fully exercised in non-ASAN x86_64/aarch64 runs where timing is reliable.See attached patch:
0001-selftests-bpf-Add-tc_edt-to-CI-ASAN-DENYLIST.patchImpact
Without the fix,
tc_edtoccasionally causes false CI failures on ASAN runs, adding noise and forcing re-runs of PRs that don't touch networking code.References
tools/testing/selftests/bpf/prog_tests/test_tc_edt.c— test implementationgithub/kernel-patches/vmtest/ci/vmtest/configs/DENYLIST.s390x— already denylists tc_edtgithub/kernel-patches/vmtest/ci/vmtest/configs/DENYLIST.asan— target for this fix