Re: [PATCH 2/3] x86/topology: Fix AMD core count

[Huang Rui]

On Fri, Mar 18, 2016 at 05:41:01PM +0100, Borislav Petkov wrote:
> On Fri, Mar 18, 2016 at 04:03:47PM +0100, Peter Zijlstra wrote:
> > It turns out AMD gets x86_max_cores wrong when there are compute
> > units.
> > 
> > The issue is that Linux assumes:
> > 
> > 	nr_logical_cpus = nr_cores * nr_siblings
> > 
> > But AMD reports its CU unit as 2 cores, but then sets num_smp_siblings
> > to 2 as well.
> > 
> > Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> > Cc: Borislav Petkov <bp@xxxxxxxxx>
> > Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > Cc: Andreas Herrmann <aherrmann@xxxxxxxx>
> > Reported-by: Xiong Zhou <jencce.kernel@xxxxxxxxx>
> > Fixes: 1f12e32f4cd5 ("x86/topology: Create logical package id")
> > Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
> > Link: http://lkml.kernel.org/r/20160317095220.GO6344@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
> > ---
> >  arch/x86/kernel/cpu/amd.c |    8 ++++----
> >  arch/x86/kernel/smpboot.c |   11 ++++++-----
> >  2 files changed, 10 insertions(+), 9 deletions(-)
> > 
> > --- a/arch/x86/kernel/cpu/amd.c
> > +++ b/arch/x86/kernel/cpu/amd.c
> > @@ -313,9 +313,9 @@ static void amd_get_topology(struct cpui
> >  		node_id = ecx & 7;
> >  
> >  		/* get compute unit information */
> > -		smp_num_siblings = ((ebx >> 8) & 3) + 1;
> > +		cores_per_cu = smp_num_siblings = ((ebx >> 8) & 3) + 1;
> > +		c->x86_max_cores /= smp_num_siblings;
> >  		c->compute_unit_id = ebx & 0xff;
> > -		cores_per_cu += ((ebx >> 8) & 3);
> >  	} else if (cpu_has(c, X86_FEATURE_NODEID_MSR)) {
> >  		u64 value;
> >  
> > @@ -331,8 +331,8 @@ static void amd_get_topology(struct cpui
> >  		u32 cus_per_node;
> >  
> >  		set_cpu_cap(c, X86_FEATURE_AMD_DCM);
> > -		cores_per_node = c->x86_max_cores / nodes_per_socket;
> > -		cus_per_node = cores_per_node / cores_per_cu;
> > +		cus_per_node = c->x86_max_cores / nodes_per_socket;
> > +		cores_per_node = cus_per_node * cores_per_cu;
> >  
> >  		/* store NodeID, use llc_shared_map to store sibling info */
> >  		per_cpu(cpu_llc_id, cpu) = node_id;
> 
> Looks ok to me, however it probably would be prudent if AMD tested it on
> a bunch of machines just to make sure we don't break anything else. I'm
> thinking F15h and F16h, something big...
> 
> Rui, can you find some time to run this one please?
> 
> Look at before/after info in /proc/cpuinfo, topology in sysfs and dmesg
> before and after might be useful too.
> 

OK, we will find some fam15h, fam16h platforms to verify it. Please
wait for my feedback.

But I am confused with c->x86_max_cores /= smp_num_siblings, what is
the real meaning of c->x86_max_cores here for AMD, the whole compute
unit numbers per socket?

+ Sherry, for her awareness.

Thanks,
Rui