Re: [PATCH 6/6] clone4: Introduce new CLONE_FD flag to get task exit notification via fd

On Sat, Mar 14, 2015 at 08:47:21PM +0100, Oleg Nesterov wrote:

On 03/14, Oleg Nesterov wrote:

quoted

On 03/14, Josh Triplett wrote:

quoted

On Sat, Mar 14, 2015 at 11:38:29AM -0700, Thiago Macieira wrote:

quoted

On Saturday 14 March 2015 15:32:35 Oleg Nesterov wrote:

quoted

It is not clear to me what do_wait() should do with ->autoreap child, even
ignoring ptrace.

Just suppose that real_parent has a single "autoreap" child. Should
wait(NULL) hanf then?

It should ignore the child that is set to autoreap. wait(NULL) should return -
ECHILD, indicating there are no children waiting to be reaped.

Right.  And I don't think the current code does this.  I think we need
to change wait_consider_task to early-return for ->autoreap just as it
does for task_state == EXIT_DEAD.

No. This EXIT_DEAD is absolutely different. And this is another indication
that you might use it wrongly ;)

What we actually want is BUG_ON(task_state == EXIT_DEAD) here. We do not
want the EXIT_DEAD tasks in ->children/ptraced lists. These EXIT_DEAD tasks
complicate the exit/wait/reparent paths.

However, currently this is TODO. The main problem is the locking in
wait_task_zombie(), we can set EXIT_DEAD and remove the task from list
under read_lock().

Let me clarify in case I confused you.

The EXIT_DEAD check in do_wait() paths doesn't mean "autoreap". It means
that this thread/process (depending on ptrace) was already reaped. It was
reaped by our sub-thread, or it was reaped because we ignore SIGCHLD, or
other reasons. This doesn't matter.

In short, EXIT_DEAD means: we have to keep this thread on lists until the
task which set this state calls release_task().

That much I already understood from reading through the code, since
exit_notify doesn't set task_state to EXIT_DEAD until the task is
actually completely dead.  When wait_consider_task sees p->task_state ==
EXIT_DEAD, that task isn't eligible for waiting at all.

What I was proposing was that a task that isn't yet dead, but that is
going to be autoreaped, is not eligible for waiting either.  All the
various wait* familiy of system calls should pretend it doesn't exist at
all, because returning an autoreaped task from a wait* call introduces a
race condition if the parent tries to *do* anything with the returned
PID.  If you launch a process with CLONE_FD, you need to manage it
exclusively with that fd, not with the wait* family of system calls.

That also implies that the child-stop and child-continued mechanisms
(do_notify_parent_cldstop, WSTOPPED, WCONTINUED) should ignore the task
too.  In the future there could be a flag to clone4 that lets you get
stop and continue notifications through the file descriptor.

- Josh Triplett

On 03/14, Josh Triplett wrote:

What I was proposing was that a task that isn't yet dead, but that is
going to be autoreaped, is not eligible for waiting either.  All the
various wait* familiy of system calls should pretend it doesn't exist at
all, because returning an autoreaped task from a wait* call introduces a
race condition if the parent tries to *do* anything with the returned
PID.  If you launch a process with CLONE_FD, you need to manage it
exclusively with that fd, not with the wait* family of system calls.

That also implies that the child-stop and child-continued mechanisms
(do_notify_parent_cldstop, WSTOPPED, WCONTINUED) should ignore the task
too.  In the future there could be a flag to clone4 that lets you get
stop and continue notifications through the file descriptor.

So far I strongly disagree, and I think that "autoreap" feature should
not depend on CLONE_FD.

However, as I already said, perhaps something like W_IGNORE_AUTOREAP
for wait* makes sense.

Plus we should also discuss the reparenting. Ok, let me leave this
discussion until I read 0/4 at least.

Oleg.