let me change the subject to avoid the confusion with the already confusing disccussion about task_is_descendant().
On 10/29, Oleg Nesterov wrote: > > I still think we need a single pid_alive() check and I even sent the patch. > Attached again at the end. > > To clarify, let me repeat that ptracer_exception_found() may need some fixes > too, right now I am only talking about task_is_descendant(). so yes, the ptracer_relations code looks very broken to me, but perhaps I misread this code, please correct me. RCU can only protect the ptracer_relations list itself, you can do nothing with (say) relation->tracer. relation->tracer can be already freed when ptracer_exception_found() checks relation->tracee == tracee. Not only pid_alive(parent) can not help in this case, pid_alive(parent) is equally unsafe because, again, this memory can be freed. security_task_free(tsk) is called right before free_task(tsk), there is no a gp pass in between, and of course we can't rely on the ->invalid check. _At first glance_ we can fix this if we simply turn both ->tracer/tracee pointers into "signal_struct *", then we can turn all same_thread_group()'s into walker->signal == parent which doesn't need to dereference the possibly- freed parent. This also allows to remove all thread_group_leader() checks. We need to ensure that false-positive is not possible (if, say, ->tracer was already re-allocated and points to another task->signal), but this doesn't look difficult. Oleg.
