cpumask: reduce stack usage in SD_x_INIT initializers
authorMike Travis <travis@sgi.com>
Sat, 5 Apr 2008 01:11:11 +0000 (18:11 -0700)
committerIngo Molnar <mingo@elte.hu>
Sat, 19 Apr 2008 17:44:59 +0000 (19:44 +0200)
  * Remove empty cpumask_t (and all non-zero/non-null) variables
    in SD_*_INIT macros.  Use memset(0) to clear.  Also, don't
    inline the initializer functions to save on stack space in
    build_sched_domains().

  * Merge change to include/linux/topology.h that uses the new
    node_to_cpumask_ptr function in the nr_cpus_node macro into
    this patch.

Depends on:
[mm-patch]: asm-generic-add-node_to_cpumask_ptr-macro.patch
[sched-devel]: sched: add new set_cpus_allowed_ptr function

Cc: H. Peter Anvin <hpa@zytor.com>
Signed-off-by: Mike Travis <travis@sgi.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
include/asm-x86/topology.h
include/linux/topology.h
kernel/sched.c

index b167ca9..9ef74c5 100644 (file)
@@ -154,10 +154,6 @@ extern unsigned long node_remap_size[];
 
 /* sched_domains SD_NODE_INIT for NUMAQ machines */
 #define SD_NODE_INIT (struct sched_domain) {           \
-       .span                   = CPU_MASK_NONE,        \
-       .parent                 = NULL,                 \
-       .child                  = NULL,                 \
-       .groups                 = NULL,                 \
        .min_interval           = 8,                    \
        .max_interval           = 32,                   \
        .busy_factor            = 32,                   \
@@ -175,7 +171,6 @@ extern unsigned long node_remap_size[];
                                | SD_WAKE_BALANCE,      \
        .last_balance           = jiffies,              \
        .balance_interval       = 1,                    \
-       .nr_balance_failed      = 0,                    \
 }
 
 #ifdef CONFIG_X86_64_ACPI_NUMA
index bd14f8b..4bb7074 100644 (file)
 #endif
 
 #ifndef nr_cpus_node
-#define nr_cpus_node(node)                                                     \
-       ({                                                                      \
-               cpumask_t __tmp__;                                              \
-               __tmp__ = node_to_cpumask(node);                                \
-               cpus_weight(__tmp__);                                           \
+#define nr_cpus_node(node)                             \
+       ({                                              \
+               node_to_cpumask_ptr(__tmp__, node);     \
+               cpus_weight(*__tmp__);                  \
        })
 #endif
 
-#define for_each_node_with_cpus(node)                                          \
-       for_each_online_node(node)                                              \
+#define for_each_node_with_cpus(node)                  \
+       for_each_online_node(node)                      \
                if (nr_cpus_node(node))
 
 void arch_update_cpu_topology(void);
@@ -80,7 +79,9 @@ void arch_update_cpu_topology(void);
  * by defining their own arch-specific initializer in include/asm/topology.h.
  * A definition there will automagically override these default initializers
  * and allow arch-specific performance tuning of sched_domains.
+ * (Only non-zero and non-null fields need be specified.)
  */
+
 #ifdef CONFIG_SCHED_SMT
 /* MCD - Do we really need this?  It is always on if CONFIG_SCHED_SMT is,
  * so can't we drop this in favor of CONFIG_SCHED_SMT?
@@ -89,20 +90,10 @@ void arch_update_cpu_topology(void);
 /* Common values for SMT siblings */
 #ifndef SD_SIBLING_INIT
 #define SD_SIBLING_INIT (struct sched_domain) {                \
-       .span                   = CPU_MASK_NONE,        \
-       .parent                 = NULL,                 \
-       .child                  = NULL,                 \
-       .groups                 = NULL,                 \
        .min_interval           = 1,                    \
        .max_interval           = 2,                    \
        .busy_factor            = 64,                   \
        .imbalance_pct          = 110,                  \
-       .cache_nice_tries       = 0,                    \
-       .busy_idx               = 0,                    \
-       .idle_idx               = 0,                    \
-       .newidle_idx            = 0,                    \
-       .wake_idx               = 0,                    \
-       .forkexec_idx           = 0,                    \
        .flags                  = SD_LOAD_BALANCE       \
                                | SD_BALANCE_NEWIDLE    \
                                | SD_BALANCE_FORK       \
@@ -112,7 +103,6 @@ void arch_update_cpu_topology(void);
                                | SD_SHARE_CPUPOWER,    \
        .last_balance           = jiffies,              \
        .balance_interval       = 1,                    \
-       .nr_balance_failed      = 0,                    \
 }
 #endif
 #endif /* CONFIG_SCHED_SMT */
@@ -121,18 +111,12 @@ void arch_update_cpu_topology(void);
 /* Common values for MC siblings. for now mostly derived from SD_CPU_INIT */
 #ifndef SD_MC_INIT
 #define SD_MC_INIT (struct sched_domain) {             \
-       .span                   = CPU_MASK_NONE,        \
-       .parent                 = NULL,                 \
-       .child                  = NULL,                 \
-       .groups                 = NULL,                 \
        .min_interval           = 1,                    \
        .max_interval           = 4,                    \
        .busy_factor            = 64,                   \
        .imbalance_pct          = 125,                  \
        .cache_nice_tries       = 1,                    \
        .busy_idx               = 2,                    \
-       .idle_idx               = 0,                    \
-       .newidle_idx            = 0,                    \
        .wake_idx               = 1,                    \
        .forkexec_idx           = 1,                    \
        .flags                  = SD_LOAD_BALANCE       \
@@ -144,7 +128,6 @@ void arch_update_cpu_topology(void);
                                | BALANCE_FOR_MC_POWER, \
        .last_balance           = jiffies,              \
        .balance_interval       = 1,                    \
-       .nr_balance_failed      = 0,                    \
 }
 #endif
 #endif /* CONFIG_SCHED_MC */
@@ -152,10 +135,6 @@ void arch_update_cpu_topology(void);
 /* Common values for CPUs */
 #ifndef SD_CPU_INIT
 #define SD_CPU_INIT (struct sched_domain) {            \
-       .span                   = CPU_MASK_NONE,        \
-       .parent                 = NULL,                 \
-       .child                  = NULL,                 \
-       .groups                 = NULL,                 \
        .min_interval           = 1,                    \
        .max_interval           = 4,                    \
        .busy_factor            = 64,                   \
@@ -174,16 +153,11 @@ void arch_update_cpu_topology(void);
                                | BALANCE_FOR_PKG_POWER,\
        .last_balance           = jiffies,              \
        .balance_interval       = 1,                    \
-       .nr_balance_failed      = 0,                    \
 }
 #endif
 
 /* sched_domains SD_ALLNODES_INIT for NUMA machines */
 #define SD_ALLNODES_INIT (struct sched_domain) {       \
-       .span                   = CPU_MASK_NONE,        \
-       .parent                 = NULL,                 \
-       .child                  = NULL,                 \
-       .groups                 = NULL,                 \
        .min_interval           = 64,                   \
        .max_interval           = 64*num_online_cpus(), \
        .busy_factor            = 128,                  \
@@ -191,14 +165,10 @@ void arch_update_cpu_topology(void);
        .cache_nice_tries       = 1,                    \
        .busy_idx               = 3,                    \
        .idle_idx               = 3,                    \
-       .newidle_idx            = 0, /* unused */       \
-       .wake_idx               = 0, /* unused */       \
-       .forkexec_idx           = 0, /* unused */       \
        .flags                  = SD_LOAD_BALANCE       \
                                | SD_SERIALIZE, \
        .last_balance           = jiffies,              \
        .balance_interval       = 64,                   \
-       .nr_balance_failed      = 0,                    \
 }
 
 #ifdef CONFIG_NUMA
index 9f7980f..6809178 100644 (file)
@@ -1869,17 +1869,17 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
  * find_idlest_cpu - find the idlest cpu among the cpus in group.
  */
 static int
-find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
+find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
+               cpumask_t *tmp)
 {
-       cpumask_t tmp;
        unsigned long load, min_load = ULONG_MAX;
        int idlest = -1;
        int i;
 
        /* Traverse only the allowed CPUs */
-       cpus_and(tmp, group->cpumask, p->cpus_allowed);
+       cpus_and(*tmp, group->cpumask, p->cpus_allowed);
 
-       for_each_cpu_mask(i, tmp) {
+       for_each_cpu_mask(i, *tmp) {
                load = weighted_cpuload(i);
 
                if (load < min_load || (load == min_load && i == this_cpu)) {
@@ -1918,7 +1918,7 @@ static int sched_balance_self(int cpu, int flag)
        }
 
        while (sd) {
-               cpumask_t span;
+               cpumask_t span, tmpmask;
                struct sched_group *group;
                int new_cpu, weight;
 
@@ -1934,7 +1934,7 @@ static int sched_balance_self(int cpu, int flag)
                        continue;
                }
 
-               new_cpu = find_idlest_cpu(group, t, cpu);
+               new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
                if (new_cpu == -1 || new_cpu == cpu) {
                        /* Now try balancing at a lower domain level of cpu */
                        sd = sd->child;
@@ -2818,7 +2818,7 @@ static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
 static struct sched_group *
 find_busiest_group(struct sched_domain *sd, int this_cpu,
                   unsigned long *imbalance, enum cpu_idle_type idle,
-                  int *sd_idle, cpumask_t *cpus, int *balance)
+                  int *sd_idle, const cpumask_t *cpus, int *balance)
 {
        struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
        unsigned long max_load, avg_load, total_load, this_load, total_pwr;
@@ -3119,7 +3119,7 @@ ret:
  */
 static struct rq *
 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
-                  unsigned long imbalance, cpumask_t *cpus)
+                  unsigned long imbalance, const cpumask_t *cpus)
 {
        struct rq *busiest = NULL, *rq;
        unsigned long max_load = 0;
@@ -3158,15 +3158,16 @@ find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
  */
 static int load_balance(int this_cpu, struct rq *this_rq,
                        struct sched_domain *sd, enum cpu_idle_type idle,
-                       int *balance)
+                       int *balance, cpumask_t *cpus)
 {
        int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
        struct sched_group *group;
        unsigned long imbalance;
        struct rq *busiest;
-       cpumask_t cpus = CPU_MASK_ALL;
        unsigned long flags;
 
+       cpus_setall(*cpus);
+
        /*
         * When power savings policy is enabled for the parent domain, idle
         * sibling can pick up load irrespective of busy siblings. In this case,
@@ -3181,7 +3182,7 @@ static int load_balance(int this_cpu, struct rq *this_rq,
 
 redo:
        group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
-                                  &cpus, balance);
+                                  cpus, balance);
 
        if (*balance == 0)
                goto out_balanced;
@@ -3191,7 +3192,7 @@ redo:
                goto out_balanced;
        }
 
-       busiest = find_busiest_queue(group, idle, imbalance, &cpus);
+       busiest = find_busiest_queue(group, idle, imbalance, cpus);
        if (!busiest) {
                schedstat_inc(sd, lb_nobusyq[idle]);
                goto out_balanced;
@@ -3224,8 +3225,8 @@ redo:
 
                /* All tasks on this runqueue were pinned by CPU affinity */
                if (unlikely(all_pinned)) {
-                       cpu_clear(cpu_of(busiest), cpus);
-                       if (!cpus_empty(cpus))
+                       cpu_clear(cpu_of(busiest), *cpus);
+                       if (!cpus_empty(*cpus))
                                goto redo;
                        goto out_balanced;
                }
@@ -3310,7 +3311,8 @@ out_one_pinned:
  * this_rq is locked.
  */
 static int
-load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
+load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
+                       cpumask_t *cpus)
 {
        struct sched_group *group;
        struct rq *busiest = NULL;
@@ -3318,7 +3320,8 @@ load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
        int ld_moved = 0;
        int sd_idle = 0;
        int all_pinned = 0;
-       cpumask_t cpus = CPU_MASK_ALL;
+
+       cpus_setall(*cpus);
 
        /*
         * When power savings policy is enabled for the parent domain, idle
@@ -3333,14 +3336,13 @@ load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
        schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
 redo:
        group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
-                                  &sd_idle, &cpus, NULL);
+                                  &sd_idle, cpus, NULL);
        if (!group) {
                schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
                goto out_balanced;
        }
 
-       busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance,
-                               &cpus);
+       busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
        if (!busiest) {
                schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
                goto out_balanced;
@@ -3362,8 +3364,8 @@ redo:
                spin_unlock(&busiest->lock);
 
                if (unlikely(all_pinned)) {
-                       cpu_clear(cpu_of(busiest), cpus);
-                       if (!cpus_empty(cpus))
+                       cpu_clear(cpu_of(busiest), *cpus);
+                       if (!cpus_empty(*cpus))
                                goto redo;
                }
        }
@@ -3397,6 +3399,7 @@ static void idle_balance(int this_cpu, struct rq *this_rq)
        struct sched_domain *sd;
        int pulled_task = -1;
        unsigned long next_balance = jiffies + HZ;
+       cpumask_t tmpmask;
 
        for_each_domain(this_cpu, sd) {
                unsigned long interval;
@@ -3406,8 +3409,8 @@ static void idle_balance(int this_cpu, struct rq *this_rq)
 
                if (sd->flags & SD_BALANCE_NEWIDLE)
                        /* If we've pulled tasks over stop searching: */
-                       pulled_task = load_balance_newidle(this_cpu,
-                                                               this_rq, sd);
+                       pulled_task = load_balance_newidle(this_cpu, this_rq,
+                                                          sd, &tmpmask);
 
                interval = msecs_to_jiffies(sd->balance_interval);
                if (time_after(next_balance, sd->last_balance + interval))
@@ -3566,6 +3569,7 @@ static void rebalance_domains(int cpu, enum cpu_idle_type idle)
        /* Earliest time when we have to do rebalance again */
        unsigned long next_balance = jiffies + 60*HZ;
        int update_next_balance = 0;
+       cpumask_t tmp;
 
        for_each_domain(cpu, sd) {
                if (!(sd->flags & SD_LOAD_BALANCE))
@@ -3589,7 +3593,7 @@ static void rebalance_domains(int cpu, enum cpu_idle_type idle)
                }
 
                if (time_after_eq(jiffies, sd->last_balance + interval)) {
-                       if (load_balance(cpu, rq, sd, idle, &balance)) {
+                       if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
                                /*
                                 * We've pulled tasks over so either we're no
                                 * longer idle, or one of our SMT siblings is
@@ -4945,7 +4949,7 @@ long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
        cpuset_cpus_allowed(p, &cpus_allowed);
        cpus_and(new_mask, new_mask, cpus_allowed);
  again:
-       retval = set_cpus_allowed(p, new_mask);
+       retval = set_cpus_allowed_ptr(p, &new_mask);
 
        if (!retval) {
                cpuset_cpus_allowed(p, &cpus_allowed);
@@ -5700,7 +5704,7 @@ static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
  */
 static void migrate_nr_uninterruptible(struct rq *rq_src)
 {
-       struct rq *rq_dest = cpu_rq(any_online_cpu(CPU_MASK_ALL));
+       struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
        unsigned long flags;
 
        local_irq_save(flags);
@@ -6118,14 +6122,14 @@ EXPORT_SYMBOL(nr_cpu_ids);
 
 #ifdef CONFIG_SCHED_DEBUG
 
-static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
+static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
+                                 cpumask_t *groupmask)
 {
        struct sched_group *group = sd->groups;
-       cpumask_t groupmask;
        char str[256];
 
        cpulist_scnprintf(str, sizeof(str), sd->span);
-       cpus_clear(groupmask);
+       cpus_clear(*groupmask);
 
        printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
 
@@ -6169,13 +6173,13 @@ static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
                        break;
                }
 
-               if (cpus_intersects(groupmask, group->cpumask)) {
+               if (cpus_intersects(*groupmask, group->cpumask)) {
                        printk(KERN_CONT "\n");
                        printk(KERN_ERR "ERROR: repeated CPUs\n");
                        break;
                }
 
-               cpus_or(groupmask, groupmask, group->cpumask);
+               cpus_or(*groupmask, *groupmask, group->cpumask);
 
                cpulist_scnprintf(str, sizeof(str), group->cpumask);
                printk(KERN_CONT " %s", str);
@@ -6184,10 +6188,10 @@ static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
        } while (group != sd->groups);
        printk(KERN_CONT "\n");
 
-       if (!cpus_equal(sd->span, groupmask))
+       if (!cpus_equal(sd->span, *groupmask))
                printk(KERN_ERR "ERROR: groups don't span domain->span\n");
 
-       if (sd->parent && !cpus_subset(groupmask, sd->parent->span))
+       if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
                printk(KERN_ERR "ERROR: parent span is not a superset "
                        "of domain->span\n");
        return 0;
@@ -6195,6 +6199,7 @@ static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level)
 
 static void sched_domain_debug(struct sched_domain *sd, int cpu)
 {
+       cpumask_t *groupmask;
        int level = 0;
 
        if (!sd) {
@@ -6204,14 +6209,21 @@ static void sched_domain_debug(struct sched_domain *sd, int cpu)
 
        printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
 
+       groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
+       if (!groupmask) {
+               printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
+               return;
+       }
+
        for (;;) {
-               if (sched_domain_debug_one(sd, cpu, level))
+               if (sched_domain_debug_one(sd, cpu, level, groupmask))
                        break;
                level++;
                sd = sd->parent;
                if (!sd)
                        break;
        }
+       kfree(groupmask);
 }
 #else
 # define sched_domain_debug(sd, cpu) do { } while (0)
@@ -6399,30 +6411,33 @@ __setup("isolcpus=", isolated_cpu_setup);
  * and ->cpu_power to 0.
  */
 static void
-init_sched_build_groups(cpumask_t span, const cpumask_t *cpu_map,
+init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
                        int (*group_fn)(int cpu, const cpumask_t *cpu_map,
-                                       struct sched_group **sg))
+                                       struct sched_group **sg,
+                                       cpumask_t *tmpmask),
+                       cpumask_t *covered, cpumask_t *tmpmask)
 {
        struct sched_group *first = NULL, *last = NULL;
-       cpumask_t covered = CPU_MASK_NONE;
        int i;
 
-       for_each_cpu_mask(i, span) {
+       cpus_clear(*covered);
+
+       for_each_cpu_mask(i, *span) {
                struct sched_group *sg;
-               int group = group_fn(i, cpu_map, &sg);
+               int group = group_fn(i, cpu_map, &sg, tmpmask);
                int j;
 
-               if (cpu_isset(i, covered))
+               if (cpu_isset(i, *covered))
                        continue;
 
-               sg->cpumask = CPU_MASK_NONE;
+               cpus_clear(sg->cpumask);
                sg->__cpu_power = 0;
 
-               for_each_cpu_mask(j, span) {
-                       if (group_fn(j, cpu_map, NULL) != group)
+               for_each_cpu_mask(j, *span) {
+                       if (group_fn(j, cpu_map, NULL, tmpmask) != group)
                                continue;
 
-                       cpu_set(j, covered);
+                       cpu_set(j, *covered);
                        cpu_set(j, sg->cpumask);
                }
                if (!first)
@@ -6520,7 +6535,8 @@ static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
 
 static int
-cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
+cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
+                cpumask_t *unused)
 {
        if (sg)
                *sg = &per_cpu(sched_group_cpus, cpu);
@@ -6538,19 +6554,22 @@ static DEFINE_PER_CPU(struct sched_group, sched_group_core);
 
 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
 static int
-cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
+cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
+                 cpumask_t *mask)
 {
        int group;
-       cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
-       cpus_and(mask, mask, *cpu_map);
-       group = first_cpu(mask);
+
+       *mask = per_cpu(cpu_sibling_map, cpu);
+       cpus_and(*mask, *mask, *cpu_map);
+       group = first_cpu(*mask);
        if (sg)
                *sg = &per_cpu(sched_group_core, group);
        return group;
 }
 #elif defined(CONFIG_SCHED_MC)
 static int
-cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
+cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
+                 cpumask_t *unused)
 {
        if (sg)
                *sg = &per_cpu(sched_group_core, cpu);
@@ -6562,17 +6581,18 @@ static DEFINE_PER_CPU(struct sched_domain, phys_domains);
 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
 
 static int
-cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg)
+cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
+                 cpumask_t *mask)
 {
        int group;
 #ifdef CONFIG_SCHED_MC
-       cpumask_t mask = cpu_coregroup_map(cpu);
-       cpus_and(mask, mask, *cpu_map);
-       group = first_cpu(mask);
+       *mask = cpu_coregroup_map(cpu);
+       cpus_and(*mask, *mask, *cpu_map);
+       group = first_cpu(*mask);
 #elif defined(CONFIG_SCHED_SMT)
-       cpumask_t mask = per_cpu(cpu_sibling_map, cpu);
-       cpus_and(mask, mask, *cpu_map);
-       group = first_cpu(mask);
+       *mask = per_cpu(cpu_sibling_map, cpu);
+       cpus_and(*mask, *mask, *cpu_map);
+       group = first_cpu(*mask);
 #else
        group = cpu;
 #endif
@@ -6594,13 +6614,13 @@ static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
 
 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
-                                struct sched_group **sg)
+                                struct sched_group **sg, cpumask_t *nodemask)
 {
-       cpumask_t nodemask = node_to_cpumask(cpu_to_node(cpu));
        int group;
 
-       cpus_and(nodemask, nodemask, *cpu_map);
-       group = first_cpu(nodemask);
+       *nodemask = node_to_cpumask(cpu_to_node(cpu));
+       cpus_and(*nodemask, *nodemask, *cpu_map);
+       group = first_cpu(*nodemask);
 
        if (sg)
                *sg = &per_cpu(sched_group_allnodes, group);
@@ -6636,7 +6656,7 @@ static void init_numa_sched_groups_power(struct sched_group *group_head)
 
 #ifdef CONFIG_NUMA
 /* Free memory allocated for various sched_group structures */
-static void free_sched_groups(const cpumask_t *cpu_map)
+static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
 {
        int cpu, i;
 
@@ -6648,11 +6668,11 @@ static void free_sched_groups(const cpumask_t *cpu_map)
                        continue;
 
                for (i = 0; i < MAX_NUMNODES; i++) {
-                       cpumask_t nodemask = node_to_cpumask(i);
                        struct sched_group *oldsg, *sg = sched_group_nodes[i];
 
-                       cpus_and(nodemask, nodemask, *cpu_map);
-                       if (cpus_empty(nodemask))
+                       *nodemask = node_to_cpumask(i);
+                       cpus_and(*nodemask, *nodemask, *cpu_map);
+                       if (cpus_empty(*nodemask))
                                continue;
 
                        if (sg == NULL)
@@ -6670,7 +6690,7 @@ next_sg:
        }
 }
 #else
-static void free_sched_groups(const cpumask_t *cpu_map)
+static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
 {
 }
 #endif
@@ -6728,6 +6748,65 @@ static void init_sched_groups_power(int cpu, struct sched_domain *sd)
 }
 
 /*
+ * Initializers for schedule domains
+ * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
+ */
+
+#define        SD_INIT(sd, type)       sd_init_##type(sd)
+#define SD_INIT_FUNC(type)     \
+static noinline void sd_init_##type(struct sched_domain *sd)   \
+{                                                              \
+       memset(sd, 0, sizeof(*sd));                             \
+       *sd = SD_##type##_INIT;                                 \
+}
+
+SD_INIT_FUNC(CPU)
+#ifdef CONFIG_NUMA
+ SD_INIT_FUNC(ALLNODES)
+ SD_INIT_FUNC(NODE)
+#endif
+#ifdef CONFIG_SCHED_SMT
+ SD_INIT_FUNC(SIBLING)
+#endif
+#ifdef CONFIG_SCHED_MC
+ SD_INIT_FUNC(MC)
+#endif
+
+/*
+ * To minimize stack usage kmalloc room for cpumasks and share the
+ * space as the usage in build_sched_domains() dictates.  Used only
+ * if the amount of space is significant.
+ */
+struct allmasks {
+       cpumask_t tmpmask;                      /* make this one first */
+       union {
+               cpumask_t nodemask;
+               cpumask_t this_sibling_map;
+               cpumask_t this_core_map;
+       };
+       cpumask_t send_covered;
+
+#ifdef CONFIG_NUMA
+       cpumask_t domainspan;
+       cpumask_t covered;
+       cpumask_t notcovered;
+#endif
+};
+
+#if    NR_CPUS > 128
+#define        SCHED_CPUMASK_ALLOC             1
+#define        SCHED_CPUMASK_FREE(v)           kfree(v)
+#define        SCHED_CPUMASK_DECLARE(v)        struct allmasks *v
+#else
+#define        SCHED_CPUMASK_ALLOC             0
+#define        SCHED_CPUMASK_FREE(v)
+#define        SCHED_CPUMASK_DECLARE(v)        struct allmasks _v, *v = &_v
+#endif
+
+#define        SCHED_CPUMASK_VAR(v, a)         cpumask_t *v = (cpumask_t *) \
+                       ((unsigned long)(a) + offsetof(struct allmasks, v))
+
+/*
  * Build sched domains for a given set of cpus and attach the sched domains
  * to the individual cpus
  */
@@ -6735,6 +6814,8 @@ static int build_sched_domains(const cpumask_t *cpu_map)
 {
        int i;
        struct root_domain *rd;
+       SCHED_CPUMASK_DECLARE(allmasks);
+       cpumask_t *tmpmask;
 #ifdef CONFIG_NUMA
        struct sched_group **sched_group_nodes = NULL;
        int sd_allnodes = 0;
@@ -6748,38 +6829,60 @@ static int build_sched_domains(const cpumask_t *cpu_map)
                printk(KERN_WARNING "Can not alloc sched group node list\n");
                return -ENOMEM;
        }
-       sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
 #endif
 
        rd = alloc_rootdomain();
        if (!rd) {
                printk(KERN_WARNING "Cannot alloc root domain\n");
+#ifdef CONFIG_NUMA
+               kfree(sched_group_nodes);
+#endif
                return -ENOMEM;
        }
 
+#if SCHED_CPUMASK_ALLOC
+       /* get space for all scratch cpumask variables */
+       allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
+       if (!allmasks) {
+               printk(KERN_WARNING "Cannot alloc cpumask array\n");
+               kfree(rd);
+#ifdef CONFIG_NUMA
+               kfree(sched_group_nodes);
+#endif
+               return -ENOMEM;
+       }
+#endif
+       tmpmask = (cpumask_t *)allmasks;
+
+
+#ifdef CONFIG_NUMA
+       sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
+#endif
+
        /*
         * Set up domains for cpus specified by the cpu_map.
         */
        for_each_cpu_mask(i, *cpu_map) {
                struct sched_domain *sd = NULL, *p;
-               cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
+               SCHED_CPUMASK_VAR(nodemask, allmasks);
 
-               cpus_and(nodemask, nodemask, *cpu_map);
+               *nodemask = node_to_cpumask(cpu_to_node(i));
+               cpus_and(*nodemask, *nodemask, *cpu_map);
 
 #ifdef CONFIG_NUMA
                if (cpus_weight(*cpu_map) >
-                               SD_NODES_PER_DOMAIN*cpus_weight(nodemask)) {
+                               SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
                        sd = &per_cpu(allnodes_domains, i);
-                       *sd = SD_ALLNODES_INIT;
+                       SD_INIT(sd, ALLNODES);
                        sd->span = *cpu_map;
-                       cpu_to_allnodes_group(i, cpu_map, &sd->groups);
+                       cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
                        p = sd;
                        sd_allnodes = 1;
                } else
                        p = NULL;
 
                sd = &per_cpu(node_domains, i);
-               *sd = SD_NODE_INIT;
+               SD_INIT(sd, NODE);
                sd->span = sched_domain_node_span(cpu_to_node(i));
                sd->parent = p;
                if (p)
@@ -6789,94 +6892,114 @@ static int build_sched_domains(const cpumask_t *cpu_map)
 
                p = sd;
                sd = &per_cpu(phys_domains, i);
-               *sd = SD_CPU_INIT;
-               sd->span = nodemask;
+               SD_INIT(sd, CPU);
+               sd->span = *nodemask;
                sd->parent = p;
                if (p)
                        p->child = sd;
-               cpu_to_phys_group(i, cpu_map, &sd->groups);
+               cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
 
 #ifdef CONFIG_SCHED_MC
                p = sd;
                sd = &per_cpu(core_domains, i);
-               *sd = SD_MC_INIT;
+               SD_INIT(sd, MC);
                sd->span = cpu_coregroup_map(i);
                cpus_and(sd->span, sd->span, *cpu_map);
                sd->parent = p;
                p->child = sd;
-               cpu_to_core_group(i, cpu_map, &sd->groups);
+               cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
 #endif
 
 #ifdef CONFIG_SCHED_SMT
                p = sd;
                sd = &per_cpu(cpu_domains, i);
-               *sd = SD_SIBLING_INIT;
+               SD_INIT(sd, SIBLING);
                sd->span = per_cpu(cpu_sibling_map, i);
                cpus_and(sd->span, sd->span, *cpu_map);
                sd->parent = p;
                p->child = sd;
-               cpu_to_cpu_group(i, cpu_map, &sd->groups);
+               cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
 #endif
        }
 
 #ifdef CONFIG_SCHED_SMT
        /* Set up CPU (sibling) groups */
        for_each_cpu_mask(i, *cpu_map) {
-               cpumask_t this_sibling_map = per_cpu(cpu_sibling_map, i);
-               cpus_and(this_sibling_map, this_sibling_map, *cpu_map);
-               if (i != first_cpu(this_sibling_map))
+               SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
+               SCHED_CPUMASK_VAR(send_covered, allmasks);
+
+               *this_sibling_map = per_cpu(cpu_sibling_map, i);
+               cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
+               if (i != first_cpu(*this_sibling_map))
                        continue;
 
                init_sched_build_groups(this_sibling_map, cpu_map,
-                                       &cpu_to_cpu_group);
+                                       &cpu_to_cpu_group,
+                                       send_covered, tmpmask);
        }
 #endif
 
 #ifdef CONFIG_SCHED_MC
        /* Set up multi-core groups */
        for_each_cpu_mask(i, *cpu_map) {
-               cpumask_t this_core_map = cpu_coregroup_map(i);
-               cpus_and(this_core_map, this_core_map, *cpu_map);
-               if (i != first_cpu(this_core_map))
+               SCHED_CPUMASK_VAR(this_core_map, allmasks);
+               SCHED_CPUMASK_VAR(send_covered, allmasks);
+
+               *this_core_map = cpu_coregroup_map(i);
+               cpus_and(*this_core_map, *this_core_map, *cpu_map);
+               if (i != first_cpu(*this_core_map))
                        continue;
+
                init_sched_build_groups(this_core_map, cpu_map,
-                                       &cpu_to_core_group);
+                                       &cpu_to_core_group,
+                                       send_covered, tmpmask);
        }
 #endif
 
        /* Set up physical groups */
        for (i = 0; i < MAX_NUMNODES; i++) {
-               cpumask_t nodemask = node_to_cpumask(i);
+               SCHED_CPUMASK_VAR(nodemask, allmasks);
+               SCHED_CPUMASK_VAR(send_covered, allmasks);
 
-               cpus_and(nodemask, nodemask, *cpu_map);
-               if (cpus_empty(nodemask))
+               *nodemask = node_to_cpumask(i);
+               cpus_and(*nodemask, *nodemask, *cpu_map);
+               if (cpus_empty(*nodemask))
                        continue;
 
-               init_sched_build_groups(nodemask, cpu_map, &cpu_to_phys_group);
+               init_sched_build_groups(nodemask, cpu_map,
+                                       &cpu_to_phys_group,
+                                       send_covered, tmpmask);
        }
 
 #ifdef CONFIG_NUMA
        /* Set up node groups */
-       if (sd_allnodes)
-               init_sched_build_groups(*cpu_map, cpu_map,
-                                       &cpu_to_allnodes_group);
+       if (sd_allnodes) {
+               SCHED_CPUMASK_VAR(send_covered, allmasks);
+
+               init_sched_build_groups(cpu_map, cpu_map,
+                                       &cpu_to_allnodes_group,
+                                       send_covered, tmpmask);
+       }
 
        for (i = 0; i < MAX_NUMNODES; i++) {
                /* Set up node groups */
                struct sched_group *sg, *prev;
-               cpumask_t nodemask = node_to_cpumask(i);
-               cpumask_t domainspan;
-               cpumask_t covered = CPU_MASK_NONE;
+               SCHED_CPUMASK_VAR(nodemask, allmasks);
+               SCHED_CPUMASK_VAR(domainspan, allmasks);
+               SCHED_CPUMASK_VAR(covered, allmasks);
                int j;
 
-               cpus_and(nodemask, nodemask, *cpu_map);
-               if (cpus_empty(nodemask)) {
+               *nodemask = node_to_cpumask(i);
+               cpus_clear(*covered);
+
+               cpus_and(*nodemask, *nodemask, *cpu_map);
+               if (cpus_empty(*nodemask)) {
                        sched_group_nodes[i] = NULL;
                        continue;
                }
 
-               domainspan = sched_domain_node_span(i);
-               cpus_and(domainspan, domainspan, *cpu_map);
+               *domainspan = sched_domain_node_span(i);
+               cpus_and(*domainspan, *domainspan, *cpu_map);
 
                sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
                if (!sg) {
@@ -6885,31 +7008,31 @@ static int build_sched_domains(const cpumask_t *cpu_map)
                        goto error;
                }
                sched_group_nodes[i] = sg;
-               for_each_cpu_mask(j, nodemask) {
+               for_each_cpu_mask(j, *nodemask) {
                        struct sched_domain *sd;
 
                        sd = &per_cpu(node_domains, j);
                        sd->groups = sg;
                }
                sg->__cpu_power = 0;
-               sg->cpumask = nodemask;
+               sg->cpumask = *nodemask;
                sg->next = sg;
-               cpus_or(covered, covered, nodemask);
+               cpus_or(*covered, *covered, *nodemask);
                prev = sg;
 
                for (j = 0; j < MAX_NUMNODES; j++) {
-                       cpumask_t tmp, notcovered;
+                       SCHED_CPUMASK_VAR(notcovered, allmasks);
                        int n = (i + j) % MAX_NUMNODES;
                        node_to_cpumask_ptr(pnodemask, n);
 
-                       cpus_complement(notcovered, covered);
-                       cpus_and(tmp, notcovered, *cpu_map);
-                       cpus_and(tmp, tmp, domainspan);
-                       if (cpus_empty(tmp))
+                       cpus_complement(*notcovered, *covered);
+                       cpus_and(*tmpmask, *notcovered, *cpu_map);
+                       cpus_and(*tmpmask, *tmpmask, *domainspan);
+                       if (cpus_empty(*tmpmask))
                                break;
 
-                       cpus_and(tmp, tmp, *pnodemask);
-                       if (cpus_empty(tmp))
+                       cpus_and(*tmpmask, *tmpmask, *pnodemask);
+                       if (cpus_empty(*tmpmask))
                                continue;
 
                        sg = kmalloc_node(sizeof(struct sched_group),
@@ -6920,9 +7043,9 @@ static int build_sched_domains(const cpumask_t *cpu_map)
                                goto error;
                        }
                        sg->__cpu_power = 0;
-                       sg->cpumask = tmp;
+                       sg->cpumask = *tmpmask;
                        sg->next = prev->next;
-                       cpus_or(covered, covered, tmp);
+                       cpus_or(*covered, *covered, *tmpmask);
                        prev->next = sg;
                        prev = sg;
                }
@@ -6958,7 +7081,8 @@ static int build_sched_domains(const cpumask_t *cpu_map)
        if (sd_allnodes) {
                struct sched_group *sg;
 
-               cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg);
+               cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
+                                                               tmpmask);
                init_numa_sched_groups_power(sg);
        }
 #endif
@@ -6976,11 +7100,13 @@ static int build_sched_domains(const cpumask_t *cpu_map)
                cpu_attach_domain(sd, rd, i);
        }
 
+       SCHED_CPUMASK_FREE((void *)allmasks);
        return 0;
 
 #ifdef CONFIG_NUMA
 error:
-       free_sched_groups(cpu_map);
+       free_sched_groups(cpu_map, tmpmask);
+       SCHED_CPUMASK_FREE((void *)allmasks);
        return -ENOMEM;
 #endif
 }
@@ -7020,9 +7146,10 @@ static int arch_init_sched_domains(const cpumask_t *cpu_map)
        return err;
 }
 
-static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
+static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
+                                      cpumask_t *tmpmask)
 {
-       free_sched_groups(cpu_map);
+       free_sched_groups(cpu_map, tmpmask);
 }
 
 /*
@@ -7031,6 +7158,7 @@ static void arch_destroy_sched_domains(const cpumask_t *cpu_map)
  */
 static void detach_destroy_domains(const cpumask_t *cpu_map)
 {
+       cpumask_t tmpmask;
        int i;
 
        unregister_sched_domain_sysctl();
@@ -7038,7 +7166,7 @@ static void detach_destroy_domains(const cpumask_t *cpu_map)
        for_each_cpu_mask(i, *cpu_map)
                cpu_attach_domain(NULL, &def_root_domain, i);
        synchronize_sched();
-       arch_destroy_sched_domains(cpu_map);
+       arch_destroy_sched_domains(cpu_map, &tmpmask);
 }
 
 /*
@@ -7246,7 +7374,7 @@ void __init sched_init_smp(void)
        hotcpu_notifier(update_sched_domains, 0);
 
        /* Move init over to a non-isolated CPU */
-       if (set_cpus_allowed(current, non_isolated_cpus) < 0)
+       if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
                BUG();
        sched_init_granularity();
 }