On 10/27, Tetsuo Handa wrote: > > On 2018/10/26 23:39, Oleg Nesterov wrote: > > On 10/26, Tetsuo Handa wrote: > >> Suppose p1 == p2->real_parent and p2 == p3->real_parent, and p1 exited > >> when someone tried to attach on p2, p2->real_parent was pointing to already > >> (or about to be) freed p1. > > > > I don't see a difference. > > > > If p1 exits it will re-parent p2, p2->real_parent will be updated. > > > >> So, the puzzle part is why p2->real_parent was still pointing p1 even after > >> p1 was freed... > > > > I don't understand the question. > > > > Once again. TASK->real_parent can point to the freed mem only if a) TASK > > exits, > > and b) _after_ that its parent TASK->real_parent exits too. > > Oh, p2 exited and then p1 also exited when someone tried to attach on p2. > Then, p2->real_parent can point to already (or about to be) freed p1.
Then we must see pid_alive(p2) == F as I already explained you in my yesterday's email. Because if p1 exits _after_ p2, then pid_alive(p2) == F must be already completed, p1 can't exit (I mean, release_task(p1) can't be called) until release_task(p2) does detach_pid() and drops tasklist_lock. > (By the way, if p->real_parent were updated to point to init_task when p > exits, > we could omit pid_alive() check?) Sorry, I don't understand the question... Ignoring the PR_SET_CHILD_SUBREAPER parents, ignoring the exiting sub-threads which reparent to group leader, ->real_parent is alwasy updated to point to init_task. But I don't see why we could omit pid_alive() check. Oleg.
