powernv: Include NUMA locality information in the CPU topology

ULE uses this topology to try and preserve locality when migrating
threads between CPUs and when performing work stealing.  Ensure that on
NUMA systems it will at least take the NUMA topology into account.

Reviewed by:	bdragon, jhibbits (previous version)
Tested by:	bdragon
MFC after:	2 weeks
Differential Revision:	https://reviews.freebsd.org/D28580

1 files changed, 58 insertions, 8 deletions

diff --git a/sys/powerpc/powernv/platform_powernv.c b/sys/powerpc/powernv/platform_powernv.c
index d7acc544c2ed..8c7da5c257ee 100644
--- a/sys/powerpc/powernv/platform_powernv.c
+++ b/sys/powerpc/powernv/platform_powernv.c
@@ -481,21 +481,71 @@ powernv_smp_probe_threads(platform_t plat)
 }
 
 static struct cpu_group *
+cpu_group_init(struct cpu_group *group, struct cpu_group *parent,
+    const cpuset_t *cpus, int children, int level, int flags)
+{
+	struct cpu_group *child;
+
+	child = children != 0 ? smp_topo_alloc(children) : NULL;
+
+	group->cg_parent = parent;
+	group->cg_child = child;
+	CPU_COPY(cpus, &group->cg_mask);
+	group->cg_count = CPU_COUNT(cpus);
+	group->cg_children = children;
+	group->cg_level = level;
+	group->cg_flags = flags;
+
+	return (child);
+}
+
+static struct cpu_group *
 powernv_smp_topo(platform_t plat)
 {
+	struct cpu_group *core, *dom, *root;
+	cpuset_t corecpus, domcpus;
+	int cpuid, i, j, k, ncores;
+
 	if (mp_ncpus % smp_threads_per_core != 0) {
-		printf("WARNING: Irregular SMP topology. Performance may be "
-		     "suboptimal (%d threads, %d on first core)\n",
-		     mp_ncpus, smp_threads_per_core);
+		printf("%s: irregular SMP topology (%d threads, %d per core)\n",
+		    __func__, mp_ncpus, smp_threads_per_core);
 		return (smp_topo_none());
 	}
 
-	/* Don't do anything fancier for non-threaded SMP */
-	if (smp_threads_per_core == 1)
-		return (smp_topo_none());
+	root = smp_topo_alloc(1);
+	dom = cpu_group_init(root, NULL, &all_cpus, vm_ndomains, CG_SHARE_NONE,
+	    0);
+
+	/*
+	 * Redundant layers will be collapsed by the caller so we don't need a
+	 * special case for a single domain.
+	 */
+	for (i = 0; i < vm_ndomains; i++, dom++) {
+		CPU_COPY(&cpuset_domain[i], &domcpus);
+		ncores = CPU_COUNT(&domcpus) / smp_threads_per_core;
+		KASSERT(CPU_COUNT(&domcpus) % smp_threads_per_core == 0,
+		    ("%s: domain %d core count not divisible by thread count",
+		    __func__, i));
+
+		core = cpu_group_init(dom, root, &domcpus, ncores, CG_SHARE_L3,
+		    0);
+		for (j = 0; j < ncores; j++, core++) {
+			/*
+			 * Assume that consecutive CPU IDs correspond to sibling
+			 * threads.
+			 */
+			CPU_ZERO(&corecpus);
+			for (k = 0; k < smp_threads_per_core; k++) {
+				cpuid = CPU_FFS(&domcpus) - 1;
+				CPU_CLR(cpuid, &domcpus);
+				CPU_SET(cpuid, &corecpus);
+			}
+			(void)cpu_group_init(core, dom, &corecpus, 0,
+			    CG_SHARE_L1, CG_FLAG_SMT);
+		}
+	}
 
-	return (smp_topo_1level(CG_SHARE_L1, smp_threads_per_core,
-	    CG_FLAG_SMT));
+	return (root);
 }
 
 #endif