[PATCH AUTOSEL 6.17-6.12] selftests/bpf: Fix flaky bpf_cookie selftest

[Sasha Levin <sashal@kernel.org>]

From: Mykyta Yatsenko <yatsenko@meta.com>

[ Upstream commit 105eb5dc74109a9f53c2f26c9a918d9347a73595 ]

bpf_cookie can fail on perf_event_open(), when it runs after the task_work
selftest. The task_work test causes perf to lower
sysctl_perf_event_sample_rate, and bpf_cookie uses sample_freq,
which is validated against that sysctl. As a result,
perf_event_open() rejects the attr if the (now tighter) limit is
exceeded.

>From perf_event_open():
if (attr.freq) {
	if (attr.sample_freq > sysctl_perf_event_sample_rate)
		return -EINVAL;
} else {
	if (attr.sample_period & (1ULL << 63))
		return -EINVAL;
}

Switch bpf_cookie to use sample_period, which is not checked against
sysctl_perf_event_sample_rate.

Signed-off-by: Mykyta Yatsenko <yatsenko@meta.com>
Signed-off-by: Andrii Nakryiko <andrii@kernel.org>
Link: https://lore.kernel.org/bpf/20250925215230.265501-1-mykyta.yatsenko5@gmail.com
Signed-off-by: Sasha Levin <sashal@kernel.org>
---

LLM Generated explanations, may be completely bogus:

YES

- What it fixes and why it matters
  - The perf-event subtest in bpf_cookie can fail with -EINVAL when run
    after tests that cause perf to throttle sampling (e.g., task_work
    stress), because the test uses frequency mode and sets
    `attr.sample_freq` above the current global limit. The kernel
    validates frequency mode against `sysctl_perf_event_sample_rate` and
    rejects it when exceeded (kernel/events/core.c:13403). In contrast,
    period mode only rejects if the high bit is set
    (kernel/events/core.c:13406), so it avoids this global-throttling
    pitfall.
  - Perf can dynamically lower `sysctl_perf_event_sample_rate` under
    high overhead (see assignment in kernel/events/core.c:654), so this
    flakiness can affect real test runs on slower systems or after heavy
    tests.

- Specific code change
  - In `tools/testing/selftests/bpf/prog_tests/bpf_cookie.c:453-454`,
    the test currently sets:
    - `attr.freq = 1;`
    - `attr.sample_freq = 10000;`
  - The commit switches to period mode by replacing those with:
    - `attr.sample_period = 100000;`
  - This removes reliance on `sysctl_perf_event_sample_rate` entirely
    for this test, eliminating the spurious -EINVAL from
    `perf_event_open()` and making the selftest deterministic.

- Scope, risk, and stable criteria
  - Selftests-only change; no kernel runtime code touched.
  - Minimal and contained (1 insertion, 2 deletions in a single file).
  - No API or architectural changes; uses long-supported perf_event_attr
    fields.
  - Purpose is purely to fix test flakiness, not to add features.
  - Low regression risk: switching from frequency to period mode is
    semantically equivalent for this test’s goal (ensuring perf samples
    fire to trigger the attached BPF program during `burn_cpu()`), while
    avoiding global sysctl dependency.
  - The issue exists in this stable tree: the local file still uses
    `attr.freq`/`attr.sample_freq` at
    `tools/testing/selftests/bpf/prog_tests/bpf_cookie.c:453-454`.

- Additional context
  - The upstream kernel already contains this exact fix (commit
    105eb5dc74109 “selftests/bpf: Fix flaky bpf_cookie selftest”).
  - Earlier attempts at hardening tests by lowering frequency (e.g., to
    1000) still risk hitting the dynamic throttle; period mode is the
    robust approach.

Given this is a small, targeted selftest flakiness fix with negligible
risk and clear benefit to stable testing reliability, it is suitable for
backporting.

 tools/testing/selftests/bpf/prog_tests/bpf_cookie.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/tools/testing/selftests/bpf/prog_tests/bpf_cookie.c b/tools/testing/selftests/bpf/prog_tests/bpf_cookie.c
index 4a0670c056bad..75f4dff7d0422 100644
--- a/tools/testing/selftests/bpf/prog_tests/bpf_cookie.c
+++ b/tools/testing/selftests/bpf/prog_tests/bpf_cookie.c
@@ -450,8 +450,7 @@ static void pe_subtest(struct test_bpf_cookie *skel)
 	attr.size = sizeof(attr);
 	attr.type = PERF_TYPE_SOFTWARE;
 	attr.config = PERF_COUNT_SW_CPU_CLOCK;
-	attr.freq = 1;
-	attr.sample_freq = 10000;
+	attr.sample_period = 100000;
 	pfd = syscall(__NR_perf_event_open, &attr, -1, 0, -1, PERF_FLAG_FD_CLOEXEC);
 	if (!ASSERT_GE(pfd, 0, "perf_fd"))
 		goto cleanup;
-- 
2.51.0