Commit 18dbcbfabfff ("perf: Fix the POLL_HUP delivery breakage") added a
direct pmu->stop() call when the refresh limit reaches zero. The change
was based on a test program [1] that reported missing POLL_HUP
notifications.However, the test program used SIGIO, which is a standard signal and can be coalesced. As a result, userspace may miss signal delivery even though the signal was generated by the kernel. This is expected behaviour for standard signals. This series adds a selftest for the PERF_EVENT_IOC_REFRESH limit using a real-time signal, which guarantees queued signal delivery and confirms that POLL_HUP notifications are delivered reliably on arm64. The second patch replaces the direct PMU stop with an explicit pending-disable guard, this can avoid redundant stop for most cases and logic is easier to understand. [1] https://lore.kernel.org/lkml/[email protected]/ Signed-off-by: Leo Yan <[email protected]> --- Changes in v2: - Replaced ASSERT_EQ() with EXPECT_EQ() in test tear down (Sashiko). - Handled a race case for high frequency overflow before disable irq_work (Sashiko). - Link to v1: https://lore.kernel.org/r/[email protected] --- Leo Yan (2): selftests/perf_events: Add test for refresh limit signals perf/core: Ignore overflows while disable is pending kernel/events/core.c | 8 +- tools/testing/selftests/perf_events/.gitignore | 1 + tools/testing/selftests/perf_events/Makefile | 3 +- .../testing/selftests/perf_events/refresh_signal.c | 120 +++++++++++++++++++++ 4 files changed, 130 insertions(+), 2 deletions(-) --- base-commit: e1914add2799225a87502051415fc5c32aeb02ae change-id: 20260429-arm_cs_clean_perf_handle-763cc339c1f5 Best regards, -- Leo Yan <[email protected]>
