On Wed, Oct 19, 2022 at 12:57 PM <sdf@xxxxxxxxxx> wrote: > > On 10/19, Delyan Kratunov wrote: > > BPF CI has revealed flakiness in the task_local_storage/exit_creds test. > > The failure point in CI [1] is that null_ptr_count is equal to 0, > > which indicates that the program hasn't run yet. This points to the > > kern_sync_rcu (sys_membarrier -> synchronize_rcu underneath) not > > waiting sufficiently. > > > Indeed, synchronize_rcu only waits for read-side sections that started > > before the call. If the program execution starts *during* the > > synchronize_rcu invocation (due to, say, preemption), the test won't > > wait long enough. > > > As a speculative fix, make the synchornize_rcu calls in a loop until > > an explicit run counter has gone up. > > > [1]: > > https://github.com/kernel-patches/bpf/actions/runs/3268263235/jobs/5374940791 > > > Signed-off-by: Delyan Kratunov <delyank@xxxxxx> > > --- > > v1 -> v2: > > Explicit loop counter and MAX_SYNC_RCU_CALLS guard. > > > .../bpf/prog_tests/task_local_storage.c | 18 +++++++++++++++--- > > .../bpf/progs/task_local_storage_exit_creds.c | 3 +++ > > 2 files changed, 18 insertions(+), 3 deletions(-) > > > diff --git a/tools/testing/selftests/bpf/prog_tests/task_local_storage.c > > b/tools/testing/selftests/bpf/prog_tests/task_local_storage.c > > index 035c263aab1b..99a42a2b6e14 100644 > > --- a/tools/testing/selftests/bpf/prog_tests/task_local_storage.c > > +++ b/tools/testing/selftests/bpf/prog_tests/task_local_storage.c > > @@ -39,7 +39,8 @@ static void test_sys_enter_exit(void) > > static void test_exit_creds(void) > > { > > struct task_local_storage_exit_creds *skel; > > - int err; > > + int err, run_count, sync_rcu_calls = 0; > > + const int MAX_SYNC_RCU_CALLS = 1000; > > > skel = task_local_storage_exit_creds__open_and_load(); > > if (!ASSERT_OK_PTR(skel, "skel_open_and_load")) > > @@ -53,8 +54,19 @@ static void test_exit_creds(void) > > if (CHECK_FAIL(system("ls > /dev/null"))) > > goto out; > > > - /* sync rcu to make sure exit_creds() is called for "ls" */ > > - kern_sync_rcu(); > > + /* kern_sync_rcu is not enough on its own as the read section we want > > + * to wait for may start after we enter synchronize_rcu, so our call > > + * won't wait for the section to finish. Loop on the run counter > > + * as well to ensure the program has run. > > + */ > > + do { > > + kern_sync_rcu(); > > + run_count = __atomic_load_n(&skel->bss->run_count, __ATOMIC_SEQ_CST); > > + } while (run_count == 0 && ++sync_rcu_calls < MAX_SYNC_RCU_CALLS); > > Acked-by: Stanislav Fomichev <sdf@xxxxxxxxxx> > > Might have been easier to do the following instead? > > int sync_rcu_calls = 1000; > do { > } while (run_count == 0 && --sync_rcu_calls); I think it's a preference of the author. Both are fine. imo. I was about to apply, but then noticed Delyan's author line and SOB are different. @meta vs @fb :( Delyan, please fix. Fixing SOB is not something maintainers can do while applying.
