[PATCH v7 17/16] arm64: hibernate: Refuse to hibernate if the boot cpu is offline

[Mark Rutland]

On Thu, Apr 21, 2016 at 05:28:52PM +0100, Lorenzo Pieralisi wrote:
> On Thu, Apr 21, 2016 at 01:33:35PM +0100, Mark Rutland wrote:
> > On Thu, Apr 21, 2016 at 12:44:16PM +0100, Mark Rutland wrote:
> > > On Wed, Apr 13, 2016 at 05:31:50PM +0100, James Morse wrote:
> > > > It is important to hibernate/resume on the same CPU, otherwise we may
> > > > change the cpu order or restore a big cpu's register state on a little
> > > > cpu.
> > > > 
> > > > We know cpu 0 is the cpu the firmware booted us on last time, 
> > > 
> > > This assumes that we only kexec from CPU0 also, which we will have to
> > > enforce. For example, disable_nonboot_cpus() does not enforce this if
> > > CPU0 has been hotplugged out.
> > > 
> > > Otherwise, this kernel's CPU0 is not necessarily the CPU the FW booted
> > > a kernel on.
> > 
> > A better approach might be:
> > 
> > * When going down for hibernate, store the physical CPU ID (e.g.
> >   MPIDR_EL1.Aff*) in the header for the hibernate image.
> > 
> > * When restoring a hibernate image, first switch over to the CPU
> >   described in the header (rather than assuming CPU0). 
> > 
> > I think this is a matter of adding a new disable_non_hibernate_cpus()
> > function (defaulting to disable_nonboot_cpus()), and overriding that in
> > the arch code. Then that can be called in resume_target_kernel.
> 
> Yes, it looks feasible at least by code inspection, given that it
> requires changes in core code I wonder whether it is a restriction
> that we can remove later or we want it in from the beginning (given
> that the issue with kexec you mention above is not present in the
> current kernel, hopefully not for long :)).

So long as we're not tied to a specific ABI for the hibernate image then
we should be fine to change that later; I agree that this can be a
subsequent improvement.

So long as we have a mechanism to detect that the hibernate image format
changed, we should be OK to add stuff to it.

> > Though there are likely caveats I've missed.
> 
> Same here, it *should* just be a matter of rescheduling on the
> cpu whose MPIDR corresponds to the cpu that created the snapshot
> image instead of the first cpu online, that can be done as you
> said with a generalized disable_nonboot_cpus() that looks up
> the MPIDR and get the corresponding cpu index, in the *current*
> logical map that I expect to be identical to the one in the
> hibernated kernel (actually I am not even sure that's a requirement).

That can't be a requirement. The logical ID may be arbitrarily
different.

Imagine a two CPU system (0x000 and 0x100, say):

* Kernel A boots (CPU0: 0x000, CPU1: 0x100);
* Kernel A hotplugs out CPU0 (0x000)
* Kernel A kexecs to kernel B on CPU1 (0x100)
* Kernel B boots (CPU0: 0x100, CPU1: 0x000)
* Kernel B hibernates on CPU0 (0x100)
* Kernel C boots (CPU0: 0x000, CPU1: 0x100)
* Kernel C switches to CPU1 (0x100)
* Kernel C resumes Kernel B, CPU1 becomes CPU0 (0x100)

Thanks,
Mark.