Re: [PATCH bpf-next] selftests/bpf: Drop the on_enter recursion check of task_local_storage
From: Alexei Starovoitov
Date: Tue Oct 20 2026 - 20:02:53 EST
On Thu, Oct 08, 2026 at 06:49 PM Qiliang Yuan <odys.yuan@xxxxxxxxx> wrote:
> other tasks. On a preemptible kernel, a task can be preempted while it
> runs on_enter, and a task entering a syscall on the same CPU then misses
> the program. With other tests running in parallel, on_enter ends up
> with recursion misses that have nothing to do with the test task, and
> the check expecting none fails.
Show the log of the failed run.
How many workers and how often does it fail?
> Drop the check. map_a holding 200, which only on_enter of the test task
> writes, already shows that on_enter ran for it.
The race works the other way too.
When another task is preempted inside on_enter and the test task calls
sys_gettid() on that cpu it's the test task that misses on_enter.
Nothing is created and "lookup map_a" fails with ENOENT.
enter_cnt == 2 in test_sys_enter_exit() has the same problem.
sys_enter is only a trigger here.
Make on_enter SEC("raw_tp") and run it with bpf_prog_test_run_opts()
like you did in enable_stats ?
Then no other task runs it and the check can stay.
> Fixes: 387b532138ee ("selftests/bpf: Tracing prog can still do lookup under busy lock")
Wrong tag. In 2022 the check was fine.
Progs attached to sys_enter were running with preemption disabled until
commit 439ebd5b5708 ("bpf: Add sleepable support for raw tracepoint programs").
Say that in the commit log.
pw-bot: cr