Re: [PATCH] sched/core: Don't pin the idle task in migrate_disable_switch()

[Peter Zijlstra]

On Thu, Aug 06, 2026 at 11:01:32AM +0200, Krystian Slowik wrote:
> [Resend as plain text, the lists rejected my first attempt for HTML.]
> 
> > What is actually doing migrate_disable() here? Why would the idle thread
> > ever hit this...
> 
> Nothing, as far as I can tell. I went back into the vmcore and the
> counter didn't come from migrate_disable() at all:
> 
>   swapper/3->migration_disabled == 8      /* 0b1000 */
>   cpu3 rq->nr_pinned == 0
>   the other 15 idle tasks: 0 and 0
> 
> That pair can't come from the API. Surrounding bytes are clean, the
> whole anomaly is one set bit:
> 
>   swapper/2 +2416: 00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00
>   swapper/3 +2416: 00 00 00 00 00 00 00 00  08 00 00 00 00 00 00 00
>                    migration_pending        ^^ migration_disabled
> 
> A second dump from the same box two days later (rcu_cblist_dequeue,
> unrelated) has all 16 idle tasks at 0.
> 
> Correction to my patch description while I'm here: "two separate
> machines" was wrong, both idle-loop oopses are the same box.
> 
> Fleet context, since it changes how much you should care: three
> Ryzen 7 7840HS boxes, no ECC, DDR5 at JEDEC 5600, no EXPO. Same
> board, BIOS and microcode, identical apart from DIMM vendor and SSD.
> Taint 0 on all three, no out-of-tree modules ever, BPF is systemd
> cgroup boilerplate only. Two of the three have produced corruption
> oopses over five weeks (css_rstat_flush GPF twice at the same RIP
> plus a slab freelist on one, the rcu_cblist fault plus this bit flip
> on the other), and they were on different kernel builds at the time.
> So I can't separate a scribbler from a platform fault yet. DDR5
> on-die ECC makes marginal cells the less likely half, though it's
> SEC per burst so not excluded. Memtest next, with low expectations
> at 11 events in five weeks.

Oof, memory corruption is a pain.

> Which makes this not a fix but an assertion, and the stable Cc
> should go. If you want it at all:
> 
>     if (likely(!p->migration_disabled))
>         return;
> 
>     if (WARN_ON_ONCE(p == rq->idle))
>         return;
> 
> behind the existing early return, so the hot path doesn't pay for it
> and the state screams instead of oopsing under the rq lock.
> 
> Placement is a guess though. Anything reaching do_set_cpus_allowed()
> with p == rq->idle hits the same missing enqueue_task(), so
> sched_change_begin() might be the better home. And if you'd rather
> not paper over corruption in the scheduler at all, that's a fine
> answer too. The thread already answered the question I actually had.

Well, do_set_cpus_allowed() or any of the other sched_change users. I
think I'm leaning towards no change here, as you found we already get a
'nice' splat if this happens. Additionally, tripping a WARN inside
schedule() is commonly fatal all on its own (printk likes to do a
wakeup, which doesn't really work all that great from inside the
scheduler -- this is being worked on on the prink side).

Thanks!