6fc3e18912
This change fixes intermittent failures of the `core.smp_bind_affinity` utest across SMP CI targets. The original failure was a flaky assertion on unbound threads (`thread_inc[x] != thread_tic[x]` in thread_entry), not incorrect bind_cpu behavior. CI logs already showed T0 binding worked (thread_inc[0] == thread_tic[0] == 100) while the not_equal check failed. The root issue is that unbound-thread CPU placement is platform-dependent and outside the bind_cpu contract. On dual-core RISC-V/ARMv7 CI, an unbound thread may legitimately run all iterations on core 0 while T0 delays on the same core, so any assertion requiring cross-core migration (thread_inc != thread_tic, or core_mask with multiple bits) remains flaky. This patch narrows the test to what bind_affinity actually verifies: T0 bound to core 0 must always execute on core 0. Unbound threads only add scheduling pressure; their core_mask is printed for observation but not asserted. Changes: - Sample `rt_hw_cpu_id()` and update counters under `rt_sched_lock` to avoid migration skew between counting and CPU sampling. - Track `thread_core_mask` for diagnostics; assert only T0 binding via `thread_inc[0] == thread_tic[0]` and `thread_core_mask[0] == 1`. - Remove flaky unbound-thread assertions (`thread_inc != thread_tic` and multi-core core_mask checks). - Move T0 assertions to the main test thread after all workers finish. - Use atomic `finsh_flag`; reset counters and `threads[]` in `utest_tc_init`. - Spin in worker loops after `run_num` until cleanup deletes them (do not return from thread_entry or use `RT_WAITING_FOREVER` with `rt_thread_delay`). - Guard `rt_thread_delete` in cleanup with non-NULL checks. Signed-off-by: GuEe-GUI <2991707448@qq.com> |
||
|---|---|---|
| .. | ||
| dfs | ||
| drivers | ||
| fal | ||
| finsh | ||
| legacy | ||
| libc | ||
| lwp | ||
| mm | ||
| mprotect | ||
| net | ||
| rust | ||
| utilities | ||
| Kconfig | ||
| SConscript | ||