Re: [PATCH 0/2] introduce __next_thread(), change next_thread()

[Oleg Nesterov]

On 08/24, Linus Torvalds wrote:
>
> On Thu, 24 Aug 2023 at 07:32, Oleg Nesterov <oleg@xxxxxxxxxx> wrote:
> >
> > After document-while_each_thread-change-first_tid-to-use-for_each_thread.patch
> > in mm tree + this series
>
> Looking at your patch 2/2, I started looking at users ("Maybe we
> *want* NULL for the end case, and make next_thread() and __next_thread
> be the same?").

Yes, but see below.

> One of the main users is while_each_thread(), which certainly wants
> that NULL case, both for an easier loop condition,

No. Please note that, say,

	do {
		do_something(t);
	} while_each_thread(current, t);

differs from for_each_thread() in that it loops starting from current,
not current->parent. I guess in most cases the order doesn't matter,
and I am going to audit the users and change them to use
for_each_thread() when possible.

Or,
	while_each_thread(current, t)
		do_something(t);

means do_something for every thread except current. And this have a
couple of valid users (say, zap_other_threads), but perhaps we can
change them too.

> but also because
> the only user that uses the 't' pointer after the loop is
> fs/proc/base.c, which wants it to be NULL.

Do you mean first_tid() ? Not only it is the only user that uses
the 't' pointer after the loop, it is the only user of lockless
while_each_thread() which (in general) is NOT rcu-safe.

But I have already changed it to use for_each_thread(), see
https://lore.kernel.org/all/20230823170806.GA11724@xxxxxxxxxx/

This is
document-while_each_thread-change-first_tid-to-use-for_each_thread.patch
in mm tree.

> And kernel/bpf/task_iter.c seems to *expect* NULL at the end?

Yes! I think the same and I even documented this in 1/2.
To me this code looks simply wrong, but so far I don't understand
it enough. Currently I am trying to push the initial cleanups into
this code. See the

	https://lore.kernel.org/all/20230821150909.GA2431@xxxxxxxxxx/

thread.

> End result: if you're changing next_thread() anyway, please just
> change it to be a completely new thing that returns NULL at the end,

See above.

I'd prefer to audit/change the current users of while_each_thread()
and next_thread(), then (perhaps) kill while_each_thread() and/or
next_thread().

Oleg.