- Remove the old smp cpu topology specification with a new, more flexible

   tree structure that encodes the level of cache sharing and other
   properties.
 - Provide several convenience functions for creating one and two level
   cpu trees as well as a default flat topology.  The system now always
   has some topology.
 - On i386 and amd64 create a seperate level in the hierarchy for HTT
   and multi-core cpus.  This will allow the scheduler to intelligently
   load balance non-uniform cores.  Presently we don't detect what level
   of the cache hierarchy is shared at each level in the topology.
 - Add a mechanism for testing common topologies that have more information
   than the MD code is able to provide via the kern.smp.topology tunable.
   This should be considered a debugging tool only and not a stable api.

Sponsored by:	Nokia

2 files changed, 189 insertions, 78 deletions

diff --git a/sys/kern/sched_ule.c b/sys/kern/sched_ule.c
index 4445b05..d17cd58 100644
--- a/sys/kern/sched_ule.c
+++ b/sys/kern/sched_ule.c
@@ -243,7 +243,6 @@ static int tryself = 1;
 static int steal_htt = 1;
 static int steal_idle = 1;
 static int steal_thresh = 2;
-static int topology = 0;
 
 /*
  * One thread queue per processor.
@@ -1212,43 +1211,6 @@ tdg_add(struct tdq_group *tdg, struct tdq *tdq)
 }
 
 static void
-sched_setup_topology(void)
-{
-	struct tdq_group *tdg;
-	struct cpu_group *cg;
-	int balance_groups;
-	struct tdq *tdq;
-	int i;
-	int j;
-
-	topology = 1;
-	balance_groups = 0;
-	for (i = 0; i < smp_topology->ct_count; i++) {
-		cg = &smp_topology->ct_group[i];
-		tdg = &tdq_groups[i];
-		/*
-		 * Initialize the group.
-		 */
-		tdg_setup(tdg);
-		/*
-		 * Find all of the group members and add them.
-		 */
-		for (j = 0; j < MAXCPU; j++) { 
-			if ((cg->cg_mask & (1 << j)) != 0) {
-				tdq = TDQ_CPU(j);
-				tdq_setup(tdq);
-				tdg_add(tdg, tdq);
-			}
-		}
-		if (tdg->tdg_cpus > 1)
-			balance_groups = 1;
-	}
-	tdg_maxid = smp_topology->ct_count - 1;
-	if (balance_groups)
-		sched_balance_groups();
-}
-
-static void
 sched_setup_smp(void)
 {
 	struct tdq_group *tdg;
@@ -1271,25 +1233,6 @@ sched_setup_smp(void)
 	}
 	tdg_maxid = cpus - 1;
 }
-
-/*
- * Fake a topology with one group containing all CPUs.
- */
-static void
-sched_fake_topo(void)
-{
-#ifdef SCHED_FAKE_TOPOLOGY
-	static struct cpu_top top;
-	static struct cpu_group group;
-
-	top.ct_count = 1;
-	top.ct_group = &group;
-	group.cg_mask = all_cpus;
-	group.cg_count = mp_ncpus;
-	group.cg_children = 0;
-	smp_topology = &top;
-#endif
-}
 #endif
 
 /*
@@ -1303,15 +1246,11 @@ sched_setup(void *dummy)
 
 	tdq = TDQ_SELF();
 #ifdef SMP
-	sched_fake_topo();
 	/*
 	 * Setup tdqs based on a topology configuration or vanilla SMP based
 	 * on mp_maxid.
 	 */
-	if (smp_topology == NULL)
-		sched_setup_smp();
-	else 
-		sched_setup_topology();
+	sched_setup_smp();
 	balance_tdq = tdq;
 	sched_balance();
 #else
@@ -2692,8 +2631,6 @@ SYSCTL_INT(_kern_sched, OID_AUTO, steal_idle, CTLFLAG_RW, &steal_idle, 0,
     "Attempts to steal work from other cores before idling");
 SYSCTL_INT(_kern_sched, OID_AUTO, steal_thresh, CTLFLAG_RW, &steal_thresh, 0,
     "Minimum load on remote cpu before we'll steal");
-SYSCTL_INT(_kern_sched, OID_AUTO, topology, CTLFLAG_RD, &topology, 0,
-    "True when a topology has been specified by the MD code.");
 #endif
 
 /* ps compat.  All cpu percentages from ULE are weighted. */
diff --git a/sys/kern/subr_smp.c b/sys/kern/subr_smp.c
index fa20e62..9c30a4a 100644
--- a/sys/kern/subr_smp.c
+++ b/sys/kern/subr_smp.c
@@ -68,7 +68,6 @@ int mp_ncpus;
 /* export this for libkvm consumers. */
 int mp_maxcpus = MAXCPU;
 
-struct cpu_top *smp_topology;
 volatile int smp_started;
 u_int mp_maxid;
 
@@ -90,6 +89,11 @@ int smp_cpus = 1;	/* how many cpu's running */
 SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD, &smp_cpus, 0,
     "Number of CPUs online");
 
+int smp_topology = 0;	/* Which topology we're using. */
+SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RD, &smp_topology, 0,
+    "Topology override setting; 0 is default provided by hardware.");
+TUNABLE_INT("kern.smp.topology", &smp_topology);
+
 #ifdef SMP
 /* Enable forwarding of a signal to a process running on a different CPU */
 static int forward_signal_enabled = 1;
@@ -385,22 +389,177 @@ smp_rendezvous(void (* setup_func)(void *),
 	/* release lock */
 	mtx_unlock_spin(&smp_ipi_mtx);
 }
-#else /* !SMP */
 
-/*
- * Provide dummy SMP support for UP kernels.  Modules that need to use SMP
- * APIs will still work using this dummy support.
- */
-static void
-mp_setvariables_for_up(void *dummy)
+static struct cpu_group group[MAXCPU];
+
+struct cpu_group *
+smp_topo(void)
 {
-	mp_ncpus = 1;
-	mp_maxid = PCPU_GET(cpuid);
-	all_cpus = PCPU_GET(cpumask);
-	KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
+	struct cpu_group *top;
+
+	/*
+	 * Check for a fake topology request for debugging purposes.
+	 */
+	switch (smp_topology) {
+	case 1:
+		/* Dual core with no sharing.  */
+		top = smp_topo_1level(CG_SHARE_NONE, 2, 0);
+		break;
+	case 3:
+		/* Dual core with shared L2.  */
+		top = smp_topo_1level(CG_SHARE_L2, 2, 0);
+		break;
+	case 4:
+		/* quad core, shared l3 among each package, private l2.  */
+		top = smp_topo_1level(CG_SHARE_L3, 4, 0);
+		break;
+	case 5:
+		/* quad core,  2 dualcore parts on each package share l2.  */
+		top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0);
+		break;
+	case 6:
+		/* Single-core 2xHTT */
+		top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT);
+		break;
+	case 7:
+		/* quad core with a shared l3, 8 threads sharing L2.  */
+		top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8,
+		    CG_FLAG_THREAD);
+		break;
+	default:
+		/* Default, ask the system what it wants. */
+		top = cpu_topo();
+		break;
+	}
+	/*
+	 * Verify the returned topology.
+	 */
+	if (top->cg_count != mp_ncpus)
+		panic("Built bad topology at %p.  CPU count %d != %d",
+		    top, top->cg_count, mp_ncpus);
+	if (top->cg_mask != all_cpus)
+		panic("Built bad topology at %p.  CPU mask 0x%X != 0x%X",
+		    top, top->cg_mask, all_cpus);
+	return (top);
 }
-SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
-    mp_setvariables_for_up, NULL)
+
+struct cpu_group *
+smp_topo_none(void)
+{
+	struct cpu_group *top;
+
+	top = &group[0];
+	top->cg_parent = NULL;
+	top->cg_child = NULL;
+	top->cg_mask = (1 << mp_ncpus) - 1;
+	top->cg_count = mp_ncpus;
+	top->cg_children = 0;
+	top->cg_level = CG_SHARE_NONE;
+	top->cg_flags = 0;
+	
+	return (top);
+}
+
+static int
+smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share,
+    int count, int flags, int start)
+{
+	cpumask_t mask;
+	int i;
+
+	for (mask = 0, i = 0; i < count; i++, start++)
+		mask |= (1 << start);
+	child->cg_parent = parent;
+	child->cg_child = NULL;
+	child->cg_children = 0;
+	child->cg_level = share;
+	child->cg_count = count;
+	child->cg_flags = flags;
+	child->cg_mask = mask;
+	parent->cg_children++;
+	for (; parent != NULL; parent = parent->cg_parent) {
+		if ((parent->cg_mask & child->cg_mask) != 0)
+			panic("Duplicate children in %p.  mask 0x%X child 0x%X",
+			    parent, parent->cg_mask, child->cg_mask);
+		parent->cg_mask |= child->cg_mask;
+		parent->cg_count += child->cg_count;
+	}
+
+	return (start);
+}
+
+struct cpu_group *
+smp_topo_1level(int share, int count, int flags)
+{
+	struct cpu_group *child;
+	struct cpu_group *top;
+	int packages;
+	int cpu;
+	int i;
+
+	cpu = 0;
+	top = &group[0];
+	packages = mp_ncpus / count;
+	top->cg_child = child = &group[1];
+	top->cg_level = CG_SHARE_NONE;
+	for (i = 0; i < packages; i++, child++)
+		cpu = smp_topo_addleaf(top, child, share, count, flags, cpu);
+	return (top);
+}
+
+struct cpu_group *
+smp_topo_2level(int l2share, int l2count, int l1share, int l1count,
+    int l1flags)
+{
+	struct cpu_group *top;
+	struct cpu_group *l1g;
+	struct cpu_group *l2g;
+	int cpu;
+	int i;
+	int j;
+
+	cpu = 0;
+	top = &group[0];
+	l2g = &group[1];
+	top->cg_child = l2g;
+	top->cg_level = CG_SHARE_NONE;
+	top->cg_children = mp_ncpus / (l2count * l1count);
+	l1g = l2g + top->cg_children;
+	for (i = 0; i < top->cg_children; i++, l2g++) {
+		l2g->cg_parent = top;
+		l2g->cg_child = l1g;
+		l2g->cg_level = l2share;
+		for (j = 0; j < l2count; j++, l1g++)
+			cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count,
+			    l1flags, cpu);
+	}
+	return (top);
+}
+
+
+struct cpu_group *
+smp_topo_find(struct cpu_group *top, int cpu)
+{
+	struct cpu_group *cg;
+	cpumask_t mask;
+	int children;
+	int i;
+
+	mask = (1 << cpu);
+	cg = top;
+	for (;;) {
+		if ((cg->cg_mask & mask) == 0)
+			return (NULL);
+		if (cg->cg_children == 0)
+			return (cg);
+		children = cg->cg_children;
+		for (i = 0, cg = cg->cg_child; i < children; cg++, i++)
+			if ((cg->cg_mask & mask) != 0)
+				break;
+	}
+	return (NULL);
+}
+#else /* !SMP */
 
 void
 smp_rendezvous(void (*setup_func)(void *), 
@@ -416,4 +575,19 @@ smp_rendezvous(void (*setup_func)(void *),
 	if (teardown_func != NULL)
 		teardown_func(arg);
 }
+
+/*
+ * Provide dummy SMP support for UP kernels.  Modules that need to use SMP
+ * APIs will still work using this dummy support.
+ */
+static void
+mp_setvariables_for_up(void *dummy)
+{
+	mp_ncpus = 1;
+	mp_maxid = PCPU_GET(cpuid);
+	all_cpus = PCPU_GET(cpumask);
+	KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero"));
+}
+SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST,
+    mp_setvariables_for_up, NULL)
 #endif /* SMP */