Re: [PATCH 2/2] perf/x86/intel/uncore: With > 8 nodes, get pci bus die id from NUMA info

[Steve Wahl <steve.wahl@hpe.com>]

On Mon, Jan 11, 2021 at 02:00:33PM +0100, Peter Zijlstra wrote:
> On Fri, Jan 08, 2021 at 09:35:49AM -0600, Steve Wahl wrote:
> 
> 
> > +		/*
> > +		 * The nodeid and idmap registers only contain enough
> > +		 * information to handle 8 nodes.  On systems with more
> > +		 * than 8 nodes, we need to rely on NUMA information,
> > +		 * filled in from BIOS supplied information, to determine
> > +		 * the topology.
> > +		 */
> 
> Egads.. do we realy have to trust BIOS data? BIOS crud tends to be
> bonghits qualitee :/

I work too close to BIOS people (virtually, at least for the moment)
to safely make disparaging remarks. :-) While the origin is the BIOS,
I'm using pieces that were already being pulled from the BIOS tables
for NUMA purposes.  

> > +		if (nr_node_ids <= 8) {
> > +			/* get the Node ID of the local register */
> > +			err = pci_read_config_dword(ubox_dev, nodeid_loc, &config);
> > +			if (err)
> > +				break;
> > +			nodeid = config & NODE_ID_MASK;
> > +			/* get the Node ID mapping */
> > +			err = pci_read_config_dword(ubox_dev, idmap_loc, &config);
> > +			if (err)
> > +				break;
> >  
> > +			segment = pci_domain_nr(ubox_dev->bus);
> > +			raw_spin_lock(&pci2phy_map_lock);
> > +			map = __find_pci2phy_map(segment);
> > +			if (!map) {
> > +				raw_spin_unlock(&pci2phy_map_lock);
> > +				err = -ENOMEM;
> > +				break;
> > +			}
> > +
> > +			/*
> > +			 * every three bits in the Node ID mapping register maps
> > +			 * to a particular node.
> > +			 */
> > +			for (i = 0; i < 8; i++) {
> > +				if (nodeid == ((config >> (3 * i)) & 0x7)) {
> > +					if (topology_max_die_per_package() > 1)
> > +						die_id = i;
> > +					else
> > +						die_id = topology_phys_to_logical_pkg(i);
> > +					map->pbus_to_dieid[bus] = die_id;
> > +					break;
> > +				}
> > +			}
> >  			raw_spin_unlock(&pci2phy_map_lock);
> > +		} else {
> > +			int node = pcibus_to_node(ubox_dev->bus);
> > +			int cpu;
> > +
> > +			segment = pci_domain_nr(ubox_dev->bus);
> > +			raw_spin_lock(&pci2phy_map_lock);
> > +			map = __find_pci2phy_map(segment);
> > +			if (!map) {
> > +				raw_spin_unlock(&pci2phy_map_lock);
> > +				err = -ENOMEM;
> > +				break;
> > +			}
> > +			die_id = -1;
> > +			for_each_cpu(cpu, cpumask_of_pcibus(ubox_dev->bus)) {
> > +				struct cpuinfo_x86 *c = &cpu_data(cpu);
> > +
> > +				if (c->initialized && cpu_to_node(cpu) == node) {
> > +					map->pbus_to_dieid[bus] = die_id = c->logical_die_id;
> > +					break;
> > +				}
> > +			}
> > +			raw_spin_unlock(&pci2phy_map_lock);
> > +
> > +			if (WARN_ON_ONCE(die_id == -1)) {
> > +				err = -EINVAL;
> >  				break;
> >  			}
> 
> This seems to assume a single die per node; is that fundemantally
> correct?

It should work for one or more nodes per die; i.e. sub-NUMA clustering
should work.  If there are any systems with fewer nodes than dies
(more than one die in a NUMA node) it will likely fail.  It's not
clear to me whether nodes < dies is a possibility or not; however,
note that this situation would be broken with or without my changes.

> Did you consider malicious BIOS data? I think we're good, but I didn't
> look too hard.

I did not consider malicious BIOS data.  With quick thought toward it,
I believe the worst that could happen is the counters get associated
with the wrong die, and only under circumstances where the previous
code would have aborted mapping the counters to dies completely (which
it does when there are more than 8 dies).

Thank you for taking the time to look at this!

--> Steve

-- 
Steve Wahl, Hewlett Packard Enterprise