On 08/25, Yonghong Song wrote: > > On 8/25/23 10:04 AM, Oleg Nesterov wrote: > >Forgot to mention in the changelog... > > > >In any case this doesn't look right. ->group_leader can exit before other > >threads, call exit_files(), and in this case task_group_seq_get_next() will > >check task->files == NULL. > > It is okay. This won't be affecting correctness. We will end with > calling bpf program for 'next_task'. Well, I didn't mean it is necessarily wrong, I simply do not know. But let's suppose that we have a thread group with the main thread M + 1000 sub-threads. In the likely case they all have the same ->files, CLONE_THREAD without CLONE_FILES is not that common. Let's assume the BPF_TASK_ITER_TGID case for simplicity. Now lets look at task_file_seq_get_next() which passes skip_if_dup_files == 1 to task_seq_get_next() and thus to task_group_seq_get_next(). Now, in this case task_seq_get_next() will return non-NULL only once (OK, unless task_file_seq_ops.stop() was called), it will return the group leader M first, then after task_file_seq_get_next() "reports" all the fd's of M and increments info->tid, the next task_seq_get_next(&info->tid, true) should return NULL because of the skip_if_dup_files check in task_group_seq_get_next(). Right? But. if the group leader M exits then M->files == NULL. And in this case task_seq_get_next() will need to "inspect" all the sub-threads even if they all have the same ->files pointer. No? Again, I am not saying this is a bug and quite possibly I misread this code, but in any case the skip_if_dup_files logic looks sub-optimal and confusing to me. Nevermind, please forget. This is minor even if I am right. Thanks for rewiev! Oleg.
