linux-kernel - Re: [PATCH 13/14] sched: Add {DE,EN}QUEUE

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20250912141904.GA3289052@noisy.programming.kicks-ass.net>
Date: Fri, 12 Sep 2025 16:19:04 +0200
From: Peter Zijlstra <peterz@...radead.org>
To: Tejun Heo <tj@...nel.org>
Cc: linux-kernel@...r.kernel.org, mingo@...hat.com, juri.lelli@...hat.com,
	vincent.guittot@...aro.org, dietmar.eggemann@....com,
	rostedt@...dmis.org, bsegall@...gle.com, mgorman@...e.de,
	vschneid@...hat.com, longman@...hat.com, hannes@...xchg.org,
	mkoutny@...e.com, void@...ifault.com, arighi@...dia.com,
	changwoo@...lia.com, cgroups@...r.kernel.org,
	sched-ext@...ts.linux.dev, liuwenfang@...or.com, tglx@...utronix.de
Subject: Re: [PATCH 13/14] sched: Add {DE,EN}QUEUE_LOCKED

On Thu, Sep 11, 2025 at 10:40:06AM -1000, Tejun Heo wrote:
> Hello,
> 
> On Thu, Sep 11, 2025 at 11:42:40AM +0200, Peter Zijlstra wrote:
> ...
> > I didn't immediately see how to do that. Doesn't that
> > list_for_each_entry_safe_reverse() rely on rq->lock to retain integrity?
> 
> Ah, sorry, I was thinking it was iterating scx_tasks list. Yes, as
> implemented, it needs to hold rq lock throughout.
> 
> > Moreover, since the goal is to allow:
> > 
> >  __schedule()
> >    lock(rq->lock);
> >    next = pick_task() := pick_task_scx()
> >      lock(dsq->lock);
> >      p = some_dsq_task(dsq);
> >      task_unlink_from_dsq(p, dsq);
> >      set_task_cpu(p, cpu_of(rq));
> >      move_task_to_local_dsq(p, ...);
> >      return p;
> > 
> > without dropping rq->lock, by relying on dsq->lock to serialize things,
> > I don't see how we can retain the runnable list at all.
> >
> > And at this point, I'm not sure I understand ext well enough to know
> > what this bypass stuff does at all, let alone suggest means to
> > re architect this.
> 
> Bypass mode is enabled when the kernel side can't trust the BPF scheduling
> anymore and wants to fall back to dumb FIFO scheduling to guarantee forward
> progress (e.g. so that we can switch back to fair).
> 
> It comes down to flipping scx_rq_bypassing() on, which makes scheduling
> paths bypass most BPF parts and fall back to FIFO behavior, and then making
> sure every thread is on FIFO behavior. The latter part is what the loop is
> doing. It scans all currently runnable tasks and dequeues and re-enqueues
> them. As scx_rq_bypass() is true at this point, if a task were queued on the
> BPF side, the cycling takes it out of the BPF side and puts it on the
> fallback FIFO queue.
> 
> If we want to get rid of the locking requirement:
> 
> - Walk scx_tasks list which is iterated with a cursor and allows dropping
>   locks while iterating. However, on some hardware, there are cases where
>   CPUs are extremely slowed down from BPF scheduler making bad decisions and
>   causing a lot of sync cacheline pingponging across e.g. NUMA nodes. As
>   scx_bypass() is what's supposed to extricate the system from this state,
>   walking all tasks while going through each's locking probably isn't going
>   to be great.
> 
> - We can update ->runnable_list iteration to allow dropping rq lock e.g.
>   with a cursor based iteration. Maybe some code can be shared with
>   scx_tasks iteration. Cycling through locks still isn't going to be great
>   but here it's likely a lot fewer of them at least.
> 
> Neither option is great. Leave it as-is for now?

Ah, but I think we *have* to change it :/ The thing is that with the new
pick you can change 'rq' without holding the source rq->lock. So we
can't maintain this list.

Could something like so work?

	scoped_guard (rcu) for_each_process_thread(g, p) {
		if (p->flags & PF_EXITING || p->sched_class != ext_sched_class)
			continue;

		guard(task_rq_lock)(p);
		scoped_guard (sched_change, p) {
			/* no-op */
		}
	}