Re: [PATCH] sched_groups are expected to be circular linked list, make it so right after allocation

From: Igor Mammedov
Date: Thu May 10 2012 - 13:02:10 EST


----- Original Message -----
> From: "Peter Zijlstra" <peterz@xxxxxxxxxxxxx>
> To: "Igor Mammedov" <imammedo@xxxxxxxxxx>
> Cc: "Jiang Liu" <liuj97@xxxxxxxxx>, linux-kernel@xxxxxxxxxxxxxxx, mingo@xxxxxxxxxx, pjt@xxxxxxxxxx,
> tglx@xxxxxxxxxxxxx, "seto hidetoshi" <seto.hidetoshi@xxxxxxxxxxxxxx>
> Sent: Thursday, May 10, 2012 3:45:44 PM
> Subject: Re: [PATCH] sched_groups are expected to be circular linked list, make it so right after allocation
>
> On Thu, 2012-05-10 at 15:26 +0200, Igor Mammedov wrote:
> > [ 141.699854] sched: Bonkers domain doesn't include its own cpu: 3
> > 0-1,3
> > [ 141.725038] sched: Bonkers domain doesn't include its own cpu: 3
> > 0-1
>
> Whee!! so cpu_mask (active_mask) does include 3, but the tl->mask()
> doesn't.
>
> > [ 141.775040] sched: Topology is hosed for CPU-3!!
> > [ 141.775596] sched: domain: NODE 0-1
> > [ 141.776004] sched: group: 0-1
> >
> This seems to suggest its the node topology being wrecked.
>
> which with your code-base would be
> cpu_node_mask()->sched_domain_node_span()..
>
> Did you specify any node topology on the qemu command line? If not,
> it
> should all reduce to cpumask_of_node(0).
>
> identify_secondary_cpu()->identify_cpu()->numa_add_cpu() should set
> that
> bit. which is well before the CPU_ONLINE->cpuset_update_active_cpus()
> sched domain rebuild.
>
>
> Most puzzling. Can you dig a little deeper as to why these masks
> might
> be wrong? Also, can you reproduce on actual hardware? The reason I
probably not

> never
> use kvm or other virt for debugging is that I always end up spending
> time chasing virt bugs, and I hate virt..

qemu started with -smp 4,sockets=4,cores=1,threads=1

I can answer it right away. It happens when one of cpus failed
to come online on time and boot cpu decided to abandon attempt and exit
do_boot_cpu with error. While exiting with error it clears some masks:

if (boot_error) {
/* Try to put things back the way they were before ... */
numa_remove_cpu(cpu); /* was set by numa_add_cpu */

/* was set by do_boot_cpu() */
cpumask_clear_cpu(cpu, cpu_callout_mask);

/* was set by cpu_init() */
cpumask_clear_cpu(cpu, cpu_initialized_mask);

set_cpu_present(cpu, false);
per_cpu(x86_cpu_to_apicid, cpu) = BAD_APICID;
}

However cpu that has been blissfully forgotten, continues to boot and sets
cpu_active_mask. So causing all sorts of havoc in places that use this mask.
Then when next cpu is being onlined/offlined it eventually calls
CPU_[ONLINE|DOWN]->cpuset_update_active_cpus()
result is OOPS.

That is why I've posted https://lkml.org/lkml/2012/5/9/125
that prevent all of this from happening by 'gracefully' aborting boot
process on failed cpu with necessary cleanup.

Checked Ingo's git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git:master
races, that patches 1, 2, 4 are fixing, are still there.
If it is preferable for this patches to go into -tip, then I'll rebase them
on top of it and repost. ???

Please review them if you think it has a sense to fix races in mainline,
before cpu-hotplug overhaul of this part is complete.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/