Josh, I am really sorry for delay. On 03/15, Josh Triplett wrote: > > On Sun, Mar 15, 2015 at 08:55:06PM +0100, Oleg Nesterov wrote: > > > It should be per-process simply because this "autoreap" affects the whole > > process. And the sub-threads are already "autoreap". And these 2 autoreap's > > semantics differ, we should not confuse them. > > Will the approach I suggested, of having clones with CLONE_THREAD > inherit the autoreap value rather than setting it from CLONE_AUTOREAP, > implement the semantics you're looking for? Not sure I understand... CLONE_THREAD should not inherit the autoreap. A sub-thread is always autoreapable. > Also, are you suggesting that CLONE_AUTOREAP with CLONE_THREAD should > produce -EINVAL, or just that it should be ignored? Yes, I think CLONE_AUTOREAP | CLONE_THREAD should return -EINVAL. But this all is minor... The main problem is how/when we should check this "autoreap" without making this code even more ugly. I still think we need a preparation patch. I tried to make it today but failed. Will try again on weekend... Note that we can't solely rely on do_notify_parent() which (with your patch) correctly checks !ptrace && autoreap. Just for example. Please look at __ptrace_detach(). Note that if we add CLONE_AUTOREAP this needs a fix in any case. The tracee can be "autoreap" but zombie, because "autoreap" should be ignored until the tracer detaches. But the "same_thread_group" should not call do_notify_parent() again. So this needs another check. And let me quote our discussion from the previous email: > > EXCEPT: do we really want SIGCHLD from the exiting child? I think we > > do not. I won't really argue though, but this should be discussed and > > documented. IIUC, with your patch it is still sent. > > I think we do, yes. The caller of clone can already specify what signal > they want, including no signal at all. If they specify a signal > (SIGCHLD or otherwise) along with CLONE_AUTOREAP, we can send that > signal. OK. Agreed. Yes, I agree... But the changes in __ptrace_detach() depend on whether we need to send a signal or not. Either way the changle is simple, but looks ugly. It would be nice to cleanup this somehow. Also. I forgot that the kernel always resets ->exit_signal to SIGCHLD on exec or reparenting. Reparenting is probably fine. But what about exec? Should it keep ->exit_signal == 0 if "autoreap" ? I think it should not, to avoid the strange special case. > > > > And there are ptrace/mt issues, > > > > it seems. Just for example, we should avoid EXIT_TRACE if autoreap in > > > > wait_task_zombie() even if we are going to re-notify parent. > > > > > > I don't see how EXIT_TRACE can happen in wait_task_zombie if autoreap is > > > set. wait_task_zombie does a cmpxchg with exit_state and doesn't > > > proceed unless exit_state was EXIT_ZOMBIE, and I don't see how we can > > > ever reach the EXIT_ZOMBIE state if autoreap. > > > > Because you again forgot about ptrace ;) And this too asks for preparation before CLONE_AUTOREAP... So I'll try to think about this all again on weekend. I'll try very much to not disappear again ;) Oleg. -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
