Re: [PATCH bpf-next] selftests/bpf: Wait until the .data map of test_mmap is freed

[Alexei Starovoitov]

On Thu, Oct 08, 2026 at 06:49 PM Qiliang Yuan <odys.yuan@xxxxxxxxx> wrote:
> The map goes away only when the program that uses it is freed, which
> happens in a work item after an RCU grace period. With other tests
> running in parallel, that can take longer than the polling, and the
> test still gets an FD for the map ID.

Pls include the output of the failed run, the number of workers
and whether it's a debug kernel.
It's not one GP. sys_enter is a faultable tracepoint.
bpf_link_free() does call_rcu_tasks_trace() and only then
bpf_link_dealloc() drops the prog.
Then call_rcu() in __bpf_prog_put_noref() and then the work.
Also test_mmap has no .data map. It's data_map.

> Poll until the map ID is gone. The watchdog of test_progs kills the
> test if the map is never freed.

It doesn't.
watchdog_timer_func() does pthread_kill(SIGSEGV) once.
The timer is not re-armed after that.
crash_handler() prints the backtrace and returns.
Nothing faulted, so the thread goes back to where it was.
That helps when a test is stuck in a syscall. It gets EINTR.
Here usleep() returns and the loop keeps going.
If the kernel leaks the map the test spins forever instead of
failing with get_map_by_id_after.

> -	for (i = 0; i < 10000; i++) {
> +	while (true) {

The get_map_by_id_after check below is dead code after this change.
Keep the loop bounded.
See wait_for_map_free() in tramp_prog_detach.c. It waits up to 10 sec.

pw-bot: cr