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

[K Prateek Nayak]

On 4/14/2025 8:20 PM, K Prateek Nayak wrote:
> > BTW, does anyone already have a test case that produces the lockup issue
> > with one or two simple programs and some hectic CFS bandwidth settings?
>
> This is your cue to grab a brown paper bag since what I'm about to paste
> below is probably lifetime without parole in the RT land but I believe
> it gets close to the scenario described by Valentin:
>
> (Based on v6.15-rc1; I haven't yet tested this with Aaron's series yet)

I tried this with Aaron's series [1] and I did not run into any rcu
stalls yet. following are dmesg logs:

    [  122.853909] sched_cfs_period_timer: Started on CPU248
    [  122.853912] sched_cfs_period_timer: Finished on CPU248
    [  123.726232] dumb_ways_to_die: Started on CPU248 with 50000000 loops
    [  123.726574] dumb_ways_to_die: Queuing timer on CPU248
    [  123.726577] dumb_ways_to_die: Waking up RT kthread on CPU248
    [  125.768969] RT kthread: Started on CPU248
    [  125.769050] deadlock_timer: Started on CPU248

    # Fair task runs, drops rwlock, is preempted

    [  126.666709] RT kthread: Finished on CPU248

    # RT kthread finishes

    [  126.666737] deadlock_timer: Finished on CPU248

    # ktimerd function finishes and unblocks replenish

    [  126.666740] sched_cfs_period_timer: Started on CPU248
    [  126.666741] sched_cfs_period_timer: Finished on CPU248

    # cfs task runs prctl() to completion and is throttled

    [  126.666762] dumb_ways_to_die: Finished on CPU248 with 50000000 loops

    # cfs_bandwidth continues to catch up on slack accumulated

    [  126.851820] sched_cfs_period_timer: Started on CPU248
    [  126.851825] sched_cfs_period_timer: Finished on CPU248

[1] https://lore.kernel.org/all/20250409120746.635476-1-ziqianlu@xxxxxxxxxxxxx/

--
Thanks and Regards,
Prateek