Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:1 as permitted sender) client-ip=2620:137:e000::3:1;
From:   Valentin Schneider <vschneid@redhat.com>
To:     linux-rt-users@vger.kernel.org, linux-kernel@vger.kernel.org
Cc:     Sebastian Andrzej Siewior <bigeasy@linutronix.de>,
        Thomas Gleixner <tglx@linutronix.de>,
        Juri Lelli <juri.lelli@redhat.com>,
        Clark Williams <williams@redhat.com>,
        "Luis Claudio R. Goncalves" <lgoncalv@redhat.com>
Subject: [RT BUG] Stall caused by eventpoll, rwlocks and CFS bandwidth
 controller
Date:   Thu, 12 Oct 2023 17:07:02 +0200
Message-ID: <xhsmhttqvnall.mognet@vschneid.remote.csb>
MIME-Version: 1.0
Content-Type: text/plain
Precedence: bulk

Hi folks,

We've had reports of stalls happening on our v6.0-ish frankenkernels, and while
we haven't been able to come out with a reproducer (yet), I don't see anything
upstream that would prevent them from happening. 

The setup involves eventpoll, CFS bandwidth controller and timer
expiry, and the sequence looks as follows (time-ordered):

p_read (on CPUn, CFS with bandwidth controller active)
======

ep_poll_callback()
  read_lock_irqsave()
  ...
  try_to_wake_up() <- enqueue causes an update_curr() + sets need_resched
                      due to having no more runtime
    preempt_enable()
      preempt_schedule() <- switch out due to p_read being now throttled

p_write
=======

ep_poll()
  write_lock_irq() <- blocks due to having active readers (p_read)

ktimers/n
=========

timerfd_tmrproc()
`\
  ep_poll_callback()
  `\
    read_lock_irqsave() <- blocks due to having active writer (p_write)


From this point we have a circular dependency:

  p_read -> ktimers/n (to replenish runtime of p_read)
  ktimers/n -> p_write (to let ktimers/n acquire the readlock)
  p_write -> p_read (to let p_write acquire the writelock)

IIUC reverting
  286deb7ec03d ("locking/rwbase: Mitigate indefinite writer starvation")
should unblock this as the ktimers/n thread wouldn't block, but then we're back
to having the indefinite starvation so I wouldn't necessarily call this a win.

Two options I'm seeing:
- Prevent p_read from being preempted when it's doing the wakeups under the
  readlock (icky)
- Prevent ktimers / ksoftirqd (*) from running the wakeups that have
  ep_poll_callback() as a wait_queue_entry callback. Punting that to e.g. a
  kworker /should/ do.

(*) It's not just timerfd, I've also seen it via net::sock_def_readable -
it should be anything that's pollable.

I'm still scratching my head on this, so any suggestions/comments welcome!

Cheers,
Valentin