RE: [PATCH V2] Driver cpu: update online when cpu_up/down besides sysfs

From: Neil Zhang
Date: Wed Oct 29 2014 - 22:00:26 EST


Rafael,

> -----Original Message-----
> From: Rafael J. Wysocki [mailto:rjw@xxxxxxxxxxxxx]
> Sent: 2014å10æ30æ 5:47
> To: Dan Streetman
> Cc: Rafael J. Wysocki; Neil Zhang; linux-kernel; Greg Kroah-Hartman;
> nfont@xxxxxxxxxxxxxxxxxx
> Subject: Re: [PATCH V2] Driver cpu: update online when cpu_up/down besides
> sysfs
>
> On Monday, October 27, 2014 08:46:08 PM Dan Streetman wrote:
> > On Mon, Oct 27, 2014 at 5:38 PM, Rafael J. Wysocki
> > <rafael.j.wysocki@xxxxxxxxx> wrote:
> > > On 10/27/2014 3:59 AM, Neil Zhang wrote:
> > >>
> > >> The current per-cpu offline info won't be updated when we use any
> > >> other method besides sysfs to call cpu_up/down.
> > >> Thus the cpu/online can't reflect the real online status.
> > >>
> > >> This patch is going to fix the issue introduced by commit
> > >> 0902a9044fa5b7a0456ea4daacec2c2b3189ba8c (Driver core:
> > >> Use generic offline/online for CPU offline/online)
> > >>
> > >> CC: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> > >> Tested-by: Dan Streetman <ddstreet@xxxxxxxx>
> > >> Signed-off-by: Neil Zhang <zhangwm@xxxxxxxxxxx>
> > >
> > >
> > > Oh dear, no.
> > >
> > > Please first tell me what exactly the problem you're seeing is.
> >
> > For some background, here is my last comment on the first email thread on
> this:
> > https://lkml.org/lkml/2014/10/27/595
> >
> > I didn't create this patch, but the problem essentially is that before
> > your commit the individual cpu online nodes
> > (/sys/devices/system/cpu/cpuN/online) stayed in sync during
> > cpu_down/up, because they used the cpu_online_mask; while after the
> > commit, they are tracked by the cpu's generic dev->offline flag, which
> > isn't updated during cpu_down/up.
>
> Which is not triggered from sysfs.
>
> > So now, any place in the kernel
> > that brings a cpu up or down must also update the cpu->dev->offline
> > flag.
>
> Not any place. In particular, system suspend-resume doesn't need to do that,
> because it takes CPUs offline and then brings them back online.
>
> If there's a place in the kernel where CPUs are taken offline and left in that
> state, then it needs to be updated.

Many ARM SoCs will have an in kernel profiler to determine whether we need
to remove or add a cpu into system for power and performance consideration.
Thus we will call cpu_up / down in kernel directly.

>
> > My interest in the patch was coincidental because I was seeing the
> > same problem when using dlpar operations to hotplug cpus, which uses
> > the arch/powerpc/platform/pseries/dlpar.c code; that code brings a cpu
> > offline when it's hot-removed (and the cpu online when it's
> > hot-added), but it hasn't been changed to also update the cpu's
> > dev->offline flag.
>
> It should be modified to do that.
>
> > As I said in the previous email to the first thread, the ppc dlpar
> > operation might be changed in the future to fully unregister a cpu
> > when it's hot-removed, which would remove the entire sysfs cpuN
> > directory. Alternately and/or until then, it could be updated to
> > simply update the cpu'd dev->offline flag (that's what I originally
> > did for my own testing). However, without a central place to update
> > the cpu's dev->offline field, like this, or possibly in
> > set_cpu_online(), or elsewhere during cpu_down/up, each place in the
> > kernel that calls cpu_down() or cpu_up() also needs to update the
> > dev->offline flag. It's possible that the ppc dlpar code is the only
> > place in the kernel that has this problem; I haven't searched.
>
> It is quite likely to be the only place like that.
>
> While I'm not familiar with the code in question, the most straightforward way
> to fix the problem would be to replace cpu_down() in there with
> device_offline(get_cpu_device(cpu)), but that needs to be called under
> device_hotplug_lock.

Great, we can use this way to fix our problems.
Thanks for the remind!

>
> --
> I speak only for myself.
> Rafael J. Wysocki, Intel Open Source Technology Center.

Best Regards,
Neil Zhang