[PATCH] ia64 cpuset + build_sched_domains() mangles structures

I've already sent this to the maintainers, and this is now being sent to a
larger community audience.  I have fixed a problem with the ia64 version of
build_sched_domains(), but a similar fix still needs to be made to the
generic build_sched_domains() in kernel/sched.c.

The "dynamic sched domains" functionality has recently been merged into
2.6.13-rcN that sees the dynamic declaration of a cpu-exclusive (a.k.a.
"isolated") cpuset and rebuilds the CPU Scheduler sched domains and sched
groups to separate away the CPUs in this cpu-exclusive cpuset from the
remainder of the non-isolated CPUs.  This allows the non-isolated CPUs to
completely ignore the isolated CPUs when doing load-balancing.

Unfortunately, build_sched_domains() expects that a sched domain will
include all the CPUs of each node in the domain, i.e., that no node will
belong in both an isolated cpuset and a non-isolated cpuset.  Declaring a
cpuset that violates this presumption will produce flawed data structures
and will oops the kernel.

To trigger the problem (on a NUMA system with >1 CPUs per node):
   cd /dev/cpuset
   mkdir newcpuset
   cd newcpuset
   echo 0 >cpus
   echo 0 >mems
   echo 1 >cpu_exclusive

I have fixed this shortcoming for ia64 NUMA (with multiple CPUs per node).
A similar shortcoming exists in the generic build_sched_domains() (in
kernel/sched.c) for NUMA, and that needs to be fixed also.  The fix
involves dynamically allocating sched_group_nodes[] and
sched_group_allnodes[] for each invocation of build_sched_domains(), rather
than using global arrays for these structures.  Care must be taken to
remember kmalloc() addresses so that arch_destroy_sched_domains() can
properly kfree() the new dynamic structures.
Signed-off-by: John Hawkes <hawkes@sgi.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: "Luck, Tony" <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

arch/ia64/kernel/domain.c

@@ -120,10 +120,10 @@ static int cpu_to_phys_group(int cpu)
  * gets dynamically allocated.
  */
 static DEFINE_PER_CPU(struct sched_domain, node_domains);
-static struct sched_group *sched_group_nodes[MAX_NUMNODES];
+static struct sched_group **sched_group_nodes_bycpu[NR_CPUS];
 
 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
-static struct sched_group sched_group_allnodes[MAX_NUMNODES];
+static struct sched_group *sched_group_allnodes_bycpu[NR_CPUS];
 
 static int cpu_to_allnodes_group(int cpu)
 {
@@ -138,6 +138,21 @@ static int cpu_to_allnodes_group(int cpu)
 void build_sched_domains(const cpumask_t *cpu_map)
 {
 	int i;
+#ifdef CONFIG_NUMA
+	struct sched_group **sched_group_nodes = NULL;
+	struct sched_group *sched_group_allnodes = NULL;
+
+	/*
+	 * Allocate the per-node list of sched groups
+	 */
+	sched_group_nodes = kmalloc(sizeof(struct sched_group*)*MAX_NUMNODES,
+					   GFP_ATOMIC);
+	if (!sched_group_nodes) {
+		printk(KERN_WARNING "Can not alloc sched group node list\n");
+		return;
+	}
+	sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
+#endif
 
 	/*
 	 * Set up domains for cpus specified by the cpu_map.
@@ -150,8 +165,21 @@ void build_sched_domains(const cpumask_t *cpu_map)
 		cpus_and(nodemask, nodemask, *cpu_map);
 
 #ifdef CONFIG_NUMA
-		if (num_online_cpus()
+		if (cpus_weight(*cpu_map)
 				> SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
+			if (!sched_group_allnodes) {
+				sched_group_allnodes
+					= kmalloc(sizeof(struct sched_group)
+							* MAX_NUMNODES,
+						  GFP_KERNEL);
+				if (!sched_group_allnodes) {
+					printk(KERN_WARNING
+					"Can not alloc allnodes sched group\n");
+					break;
+				}
+				sched_group_allnodes_bycpu[i]
+						= sched_group_allnodes;
+			}
 			sd = &per_cpu(allnodes_domains, i);
 			*sd = SD_ALLNODES_INIT;
 			sd->span = *cpu_map;
@@ -214,6 +242,7 @@ void build_sched_domains(const cpumask_t *cpu_map)
 	}
 
 #ifdef CONFIG_NUMA
+	if (sched_group_allnodes)
 		init_sched_build_groups(sched_group_allnodes, *cpu_map,
 					&cpu_to_allnodes_group);
 
@@ -226,8 +255,10 @@ void build_sched_domains(const cpumask_t *cpu_map)
 		int j;
 
 		cpus_and(nodemask, nodemask, *cpu_map);
-		if (cpus_empty(nodemask))
+		if (cpus_empty(nodemask)) {
+			sched_group_nodes[i] = NULL;
 			continue;
+		}
 
 		domainspan = sched_domain_node_span(i);
 		cpus_and(domainspan, domainspan, *cpu_map);
@@ -372,6 +403,22 @@ void arch_destroy_sched_domains(const cpumask_t *cpu_map)
 {
 #ifdef CONFIG_NUMA
 	int i;
+	int cpu;
+
+	for_each_cpu_mask(cpu, *cpu_map) {
+		struct sched_group *sched_group_allnodes
+			= sched_group_allnodes_bycpu[cpu];
+		struct sched_group **sched_group_nodes
+			= sched_group_nodes_bycpu[cpu];
+
+		if (sched_group_allnodes) {
+			kfree(sched_group_allnodes);
+			sched_group_allnodes_bycpu[cpu] = NULL;
+		}
+
+		if (!sched_group_nodes)
+			continue;
+
 		for (i = 0; i < MAX_NUMNODES; i++) {
 			cpumask_t nodemask = node_to_cpumask(i);
 			struct sched_group *oldsg, *sg = sched_group_nodes[i];
@@ -389,8 +436,9 @@ void arch_destroy_sched_domains(const cpumask_t *cpu_map)
 			kfree(oldsg);
 			if (oldsg != sched_group_nodes[i])
 				goto next_sg;
-		sched_group_nodes[i] = NULL;
+		}
+		kfree(sched_group_nodes);
+		sched_group_nodes_bycpu[cpu] = NULL;
 	}
 #endif
 }
-