Linux panics when suspend cannot offline the secondary cores

[Rafael J. Wysocki]

On Monday, June 13, 2016 03:50:56 PM Mason wrote:
> On 13/06/2016 15:30, Rafael J. Wysocki wrote:
> 
> > On Monday, June 13, 2016 02:06:14 PM Mason wrote:
> >
> >> On 10/06/2016 23:37, Mason wrote:
> >>
> >>> On 10/06/2016 23:35, Rafael J. Wysocki wrote:
> >>>
> >>>> On Friday, June 10, 2016 05:41:32 PM Mason wrote:
> >>>>
> >>>>> I'm playing with S3 Suspend-to-RAM, and I noticed that Linux is really
> >>>>> unhappy when the suspend framework fails to offline secondary cores.
> >>>>>
> >>>>> Is this expected/by design, or could it fail more gracefully?
> >>>>> (It could also be something missing in my platform's code.)
> >>>>
> >>>> This looks like a CPU offline bug to me which is more general than just
> >>>> system suspend.
> >>>
> >>> You may be right, I will try just off-lining cpu1.
> >>> Suspend may be a red herring.
> >>>
> >>> By the way, I know my implementation of tango_cpu_die
> >>> is incorrect, I was testing the failure mode.
> >>
> >> Hello Rafael,
> >>
> >> Suspend was indeed a red herring. Manually requesting cpu1 off-lining
> >> also makes Linux panic when cpu_die() unexpectedly returns.
> >>
> >> The subject should perhaps have been:
> >>
> >>   Linux panics when secondary core off-lining fails
> >>
> >> Could it be made to fail more gracefully?
> >> Or is this borkage inherent to the failed operation?
> >> Or is it a bug in my platform code?
> >> (A bug other than tango_cpu_die() failing to kill the core.)
> > 
> > Well, smp_ops.cpu_die() is not expected to return AFAICS, so that may be
> > the reason why it fails for you the way it does.
> 
> I am aware that smp_ops.cpu_die() is not expected to return.
> (I was wondering if the framework could handle it gracefully.)
> 
> The actual implementation for cpu_die() asks the firmware to off-line
> the current core. If the operation fails, for whatever reason, firmware
> is not supposed to return control to Linux?

Firmware can do what it wants (although ideally it should just do what it is
asked for).  smp_ops.cpu_die() is not supposed to return to its caller anyway.

> Is panic the only safe thing to do in Linux:
> (If yes, then why doesn't the framework panic immediately?)

I guess all of the existing implementations of smp_ops.cpu_die() don't return
to the caller no matter what, so the caller did not have to consider anything
else.

And quite frankly I don't see why it would have to.  smp_ops.cpu_die() simply
needs to be implemented to never return.

Thanks,
Rafael