Re: [RT BUG] Stall caused by eventpoll, rwlocks and CFS bandwidth controller

[K Prateek Nayak]

(+ Aaron)

Hello Jan,

On 4/9/2025 12:11 PM, Jan Kiszka wrote:
> On 12.10.23 17:07, Valentin Schneider wrote:
> > 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!
>
> We are hunting for quite some time sporadic lock-ups or RT systems,
> first only in the field (sigh), now finally also in the lab. Those have
> a fairly high overlap with what was described here. Our baselines so
> far: 6.1-rt, Debian and vanilla. We are currently preparing experiments
> with latest mainline.

Do the backtrace from these lockups show tasks (specifically ktimerd)
waiting on a rwsem? Throttle deferral helps if cfs bandwidth throttling
becomes the reason for long delay / circular dependency. Is cfs bandwidth
throttling being used on these systems that run into these lockups?
Otherwise, your issue might be completely different.

> While this thread remained silent afterwards, we have found [1][2][3] as
> apparently related. But this means we are still with this RT bug, even
> in latest 6.15-rc1?

I'm pretty sure a bunch of locking related stuff has been reworked to
accommodate PREEMPT_RT since v6.1. Many rwsem based locking patterns
have been replaced with alternatives like RCU. Recently introduced
dl_server infrastructure also helps prevent starvation of fair tasks
which can allow progress and prevent lockups. I would recommend
checking if the most recent -rt release can still reproduce your
issue:
https://lore.kernel.org/lkml/20250331095610.ulLtPP2C@xxxxxxxxxxxxx/

Note: Aaron Lu is working on Valentin's approach of deferring cfs
throttling to exit to user mode boundary
https://lore.kernel.org/lkml/20250313072030.1032893-1-ziqianlu@xxxxxxxxxxxxx/

If you still run into the issue of a lockup / long latencies on latest
-rt release and your system is using cfs bandwidth controls, you can
perhaps try running with Valentin's or Aaron's series to check if
throttle deferral helps your scenario.

> Jan
>
> [1] https://lore.kernel.org/lkml/20231030145104.4107573-1-vschneid@xxxxxxxxxx/
> [2] https://lore.kernel.org/lkml/20240202080920.3337862-1-vschneid@xxxxxxxxxx/
> [3] https://lore.kernel.org/lkml/20250220093257.9380-1-kprateek.nayak@xxxxxxx/

I'm mostly testing and reviewing Aaron's series now since per-task
throttling seems to be the way forward based on discussions in the
community.

>

--
Thanks and Regards,
Prateek