cpumask: use cpumask_var_t in uv_flush_tlb_others.

Impact: remove cpumask_t, reduce per-cpu size for CONFIG_CPUMASK_OFFSTACK=y
Signed-off-by: NRusty Russell <rusty@rustcorp.com.au>

arch/x86/kernel/tlb_uv.c

@@ -275,6 +275,8 @@ const struct cpumask *uv_flush_send_and_wait(int cpu, int this_blade,
 	return NULL;
 }
 
+static DEFINE_PER_CPU(cpumask_var_t, uv_flush_tlb_mask);
+
 /**
  * uv_flush_tlb_others - globally purge translation cache of a virtual
  * address or all TLB's
@@ -304,8 +306,7 @@ const struct cpumask *uv_flush_tlb_others(const struct cpumask *cpumask,
 					  struct mm_struct *mm,
 					  unsigned long va, unsigned int cpu)
 {
-	static DEFINE_PER_CPU(cpumask_t, flush_tlb_mask);
-	struct cpumask *flush_mask = &__get_cpu_var(flush_tlb_mask);
+	struct cpumask *flush_mask = __get_cpu_var(uv_flush_tlb_mask);
 	int i;
 	int bit;
 	int blade;
@@ -755,6 +756,10 @@ static int __init uv_bau_init(void)
 	if (!is_uv_system())
 		return 0;
 
+	for_each_possible_cpu(cur_cpu)
+		alloc_cpumask_var_node(&per_cpu(uv_flush_tlb_mask, cur_cpu),
+				       GFP_KERNEL, cpu_to_node(cur_cpu));
+
 	uv_bau_retry_limit = 1;
 	uv_nshift = uv_hub_info->n_val;
 	uv_mmask = (1UL << uv_hub_info->n_val) - 1;