Oleg Nesterov <oleg@xxxxxxxxxx> writes:
> On 05/24, Oleg Nesterov wrote:
>>
>> On 05/24, Oleg Nesterov wrote:
>> >
>> > I fail to understand this patch...
>> >
>> > On 05/18, Eric W. Biederman wrote:
>> > >
>> > > Today if a process is ptraced only the ptracer will ever be woken up in
>> > > wait
>> >
>> > and why is this wrong?
>> >
>> > > Fixes: 75b95953a569 ("job control: Add @for_ptrace to do_notify_parent_cldstop()")
>> >
>> > how does this change fix 75b95953a569?
>>
>> OK, I guess you mean the 2nd do_notify_parent_cldstop() in ptrace_stop(),
>> the problematic case is current->ptrace == T. Right?
>>
>> I dislike this patch anyway, but let me think more about it.
>
> OK, now that I understand the problem, the patch doesn't look bad to me,
> although I'd ask to make the changelog more clear.
I will see what I can do.
> After this change __wake_up_parent() can't accept any "parent" from
> p->parent thread group, but all callers look fine except
> ptrace_detach().
Having looked at it a little more I think the change was too
restrictive. For the !ptrace_reparented case there are possibly
two threads of the parent process that wait_consider_task will
allow to wait even with __WNOTHREAD specified. It is desirable
to wake them both up.
Which if I have had enough sleep reduces this patch to just:
diff --git a/kernel/exit.c b/kernel/exit.c
index f072959fcab7..c8156366b722 100644
--- a/kernel/exit.c
+++ b/kernel/exit.c
@@ -1431,8 +1431,10 @@ static int child_wait_callback(wait_queue_entry_t *wait, unsigned mode,
if (!eligible_pid(wo, p))
return 0;
- if ((wo->wo_flags & __WNOTHREAD) && wait->private != p->parent)
- return 0;
+ if ((wo->wo_flags & __WNOTHREAD) &&
+ (wait->private != p->parent) &&
+ (wait->private != p->real_parent))
+ return 0;
return default_wake_function(wait, mode, sync, key);
}
I think that solves the issue without missing wake-ups without adding
any more.
For the same set of reasons it looks like the __wake_up_parent in
__ptrace_detach is just simply dead code. I don't think there is a case
where when !ptrace_reparented the thread that is the real_parent can
sleep in do_wait when the thread that was calling ptrace could not.
That needs a very close look to confirm.
Eric
|