[RFC PATCH 0/5] ARM: introducing DT topology

Lorenzo Pieralisi lorenzo.pieralisi at arm.com
Thu Jan 19 04:50:28 EST 2012


On Wed, Jan 18, 2012 at 04:24:23PM +0000, Russell King - ARM Linux wrote:
> On Wed, Jan 18, 2012 at 02:36:43PM +0000, Lorenzo Pieralisi wrote:
> > Current code in the kernel, in particular the boot sequence, hinges upon a
> > sequential mapping of MPIDR values for cpus and related interrupt
> > controller CPU interfaces to logical cpu indexing. 
> 
> I don't believe it does.  What it does rely upon is having cpu_logical_map()
> correctly setup for the logical-to-hardware mapping - which, if it's
> different, should be done in smp_init_cpus(), taking note that logical
> CPU 0 is _always_ the boot CPU.  (That's a restriction caused by the
> way suspend/resume unplugs all but the lowest numbered logical online
> CPU.)
> 
> So, with the code today, there's nothing in the code which prevents this
> from happening:
> 
> a) you boot on h/w CPU 1, which becomes logical CPU 0.
> b) you boot h/w CPU 3 as logical CPU 1.
> c) you boot h/w CPU 0 as logical CPU 2.
> d) you boot h/w CPU 2 as logical CPU 3.
> 
> You just need to ensure that cpu_logical_map() contains the array
> {1, 3, 0, 2} instead of the default (because we _have_ to have a
> default) of {1, 0, 2, 3}.

I agree with you Russell, that's 100% valid on a single cluster. But on
a multi-cluster (eg dual-cluster) the MPIDR might be wired like this:

MPIDR[15:8] - cluster id
MPIDR[7:0] - core id

no hyperthreading

* CLUSTER 0 *
                       MPIDR[23:16]   MPIDR[15:8]    MPIDR[7:0]
HWCPU0: MPIDR=0x0          0x0            0x0           0x0
HWCPU1: MPIDR=0x1          0x0            0x0           0x1 

* CLUSTER 1 *

HWCPU2: MPIDR=0x100        0x0            0x1           0x0               
HWCPU3: MPIDR=0x101        0x0            0x1           0x1  

MPIDR is not a sequential index anymore, that's what I am going on about. 

And yes, code like cpu_resume, that relies on MPIDR[7:0] to be unique
needs patching, since that just takes into account the first affinity
level, which can have same values for different CPUs in the system if
they belong to different clusters.

> > This hypothesis is not valid when the concept of cluster is introduced since
> > the MPIDR cannot be represented as a single index and interrupt controller
> > CPU interfaces might be wired with a numbering scheme following per-SoC
> > design parameters which cannot be extrapolated easily through generic functions
> > by the primary CPU.
> 
> So what you're saying is that the GIC CPU index may not be the CPU number
> given by MPIDR?
>

Yes, that's correct. Taking the same example as above:

MPIDR[15:8] - cluster id
MPIDR[7:0] - core id

no hyperthreading

* CLUSTER 0 *
                         MPIDR[15:8]    MPIDR[7:0]  GIC-CPU-ID
		       
HWCPU0: MPIDR=0x0            0x0           0x0          0x0
HWCPU1: MPIDR=0x1            0x0           0x1          0x1 

* CLUSTER 1 *

HWCPU2: MPIDR=0x100          0x1           0x0          0x2     
HWCPU3: MPIDR=0x101          0x1           0x1          0x3

There is just one GIC distributor shared across all clusters.

> > Furthermore, relying on the MPIDR to be wired according to real topology
> > levels might turn out to be an unreliable solution, hence a SW
> > representation is needed to override possibly incorrect MPIDR register
> > values.
> 
> This sounds like you're saying that the contents of MPIDR might be buggy
> sometime in the future?  Do we actually know of any situations where the
> information in there is currently wrong (outside of the development lab)?
> If not, it's not something we should cater for until it's actually happened,
> and then the pain should be felt by those silly enough to allow the chip
> to go out the door.

I share your view Russell. Having said that: MPIDR is IMPLEMENTATION DEFINED.

There are three possibilities:

1- MPIDR is not unique (there are CPUs with duplicated values) 
2- MPIDR is unique but affinity levels do not represent hierarchy
   properly (cluster level, core level)
3- MPIDR is unique and affinity levels are properly set

IMHO:

case 1) We should not care and designers must get their act together
case 2) We might use DT to build the topology properly and rely on
        uniqueness to have a properly running system
case 3) We are home and dry	

Thoughts ?

Thanks a lot for your feedback,
Lorenzo



More information about the devicetree-discuss mailing list