[PATCH] tvec_bases too large for per-cpu data

With internal Xen-enabled kernels we see the kernel's static per-cpu data
area exceed the limit of 32k on x86-64, and even native x86-64 kernels get
fairly close to that limit.  I generally question whether it is reasonable
to have data structures several kb in size allocated as per-cpu data when
the space there is rather limited.

The biggest arch-independent consumer is tvec_bases (over 4k on 32-bit
archs, over 8k on 64-bit ones), which now gets converted to use dynamically
allocated memory instead.

Signed-off-by: Jan Beulich <jbeulich@novell.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

1 files changed, 34 insertions, 11 deletions

diff --git a/kernel/timer.c b/kernel/timer.c
index 2410c18..4427e72 100644
--- a/kernel/timer.c
+++ b/kernel/timer.c
@@ -86,7 +86,8 @@ struct tvec_t_base_s {
 } ____cacheline_aligned_in_smp;
 
 typedef struct tvec_t_base_s tvec_base_t;
-static DEFINE_PER_CPU(tvec_base_t, tvec_bases);
+static DEFINE_PER_CPU(tvec_base_t *, tvec_bases);
+static tvec_base_t boot_tvec_bases;
 
 static inline void set_running_timer(tvec_base_t *base,
 					struct timer_list *timer)
@@ -157,7 +158,7 @@ EXPORT_SYMBOL(__init_timer_base);
 void fastcall init_timer(struct timer_list *timer)
 {
 	timer->entry.next = NULL;
-	timer->base = &per_cpu(tvec_bases, raw_smp_processor_id()).t_base;
+	timer->base = &per_cpu(tvec_bases, raw_smp_processor_id())->t_base;
 }
 EXPORT_SYMBOL(init_timer);
 
@@ -218,7 +219,7 @@ int __mod_timer(struct timer_list *timer, unsigned long expires)
 		ret = 1;
 	}
 
-	new_base = &__get_cpu_var(tvec_bases);
+	new_base = __get_cpu_var(tvec_bases);
 
 	if (base != &new_base->t_base) {
 		/*
@@ -258,7 +259,7 @@ EXPORT_SYMBOL(__mod_timer);
  */
 void add_timer_on(struct timer_list *timer, int cpu)
 {
-	tvec_base_t *base = &per_cpu(tvec_bases, cpu);
+	tvec_base_t *base = per_cpu(tvec_bases, cpu);
   	unsigned long flags;
 
   	BUG_ON(timer_pending(timer) || !timer->function);
@@ -504,7 +505,7 @@ unsigned long next_timer_interrupt(void)
 	}
 	hr_expires += jiffies;
 
-	base = &__get_cpu_var(tvec_bases);
+	base = __get_cpu_var(tvec_bases);
 	spin_lock(&base->t_base.lock);
 	expires = base->timer_jiffies + (LONG_MAX >> 1);
 	list = NULL;
@@ -901,7 +902,7 @@ EXPORT_SYMBOL(xtime_lock);
  */
 static void run_timer_softirq(struct softirq_action *h)
 {
-	tvec_base_t *base = &__get_cpu_var(tvec_bases);
+	tvec_base_t *base = __get_cpu_var(tvec_bases);
 
  	hrtimer_run_queues();
 	if (time_after_eq(jiffies, base->timer_jiffies))
@@ -1256,12 +1257,32 @@ asmlinkage long sys_sysinfo(struct sysinfo __user *info)
 	return 0;
 }
 
-static void __devinit init_timers_cpu(int cpu)
+static int __devinit init_timers_cpu(int cpu)
 {
 	int j;
 	tvec_base_t *base;
 
-	base = &per_cpu(tvec_bases, cpu);
+	base = per_cpu(tvec_bases, cpu);
+	if (!base) {
+		static char boot_done;
+
+		/*
+		 * Cannot do allocation in init_timers as that runs before the
+		 * allocator initializes (and would waste memory if there are
+		 * more possible CPUs than will ever be installed/brought up).
+		 */
+		if (boot_done) {
+			base = kmalloc_node(sizeof(*base), GFP_KERNEL,
+						cpu_to_node(cpu));
+			if (!base)
+				return -ENOMEM;
+			memset(base, 0, sizeof(*base));
+		} else {
+			base = &boot_tvec_bases;
+			boot_done = 1;
+		}
+		per_cpu(tvec_bases, cpu) = base;
+	}
 	spin_lock_init(&base->t_base.lock);
 	for (j = 0; j < TVN_SIZE; j++) {
 		INIT_LIST_HEAD(base->tv5.vec + j);
@@ -1273,6 +1294,7 @@ static void __devinit init_timers_cpu(int cpu)
 		INIT_LIST_HEAD(base->tv1.vec + j);
 
 	base->timer_jiffies = jiffies;
+	return 0;
 }
 
 #ifdef CONFIG_HOTPLUG_CPU
@@ -1295,8 +1317,8 @@ static void __devinit migrate_timers(int cpu)
 	int i;
 
 	BUG_ON(cpu_online(cpu));
-	old_base = &per_cpu(tvec_bases, cpu);
-	new_base = &get_cpu_var(tvec_bases);
+	old_base = per_cpu(tvec_bases, cpu);
+	new_base = get_cpu_var(tvec_bases);
 
 	local_irq_disable();
 	spin_lock(&new_base->t_base.lock);
@@ -1326,7 +1348,8 @@ static int __devinit timer_cpu_notify(struct notifier_block *self,
 	long cpu = (long)hcpu;
 	switch(action) {
 	case CPU_UP_PREPARE:
-		init_timers_cpu(cpu);
+		if (init_timers_cpu(cpu) < 0)
+			return NOTIFY_BAD;
 		break;
 #ifdef CONFIG_HOTPLUG_CPU
 	case CPU_DEAD: