PM: Add user-space wake lock api.
[linux-2.6.git] / kernel / cpu.c
index e202a68..12b7458 100644 (file)
 #include <linux/kthread.h>
 #include <linux/stop_machine.h>
 #include <linux/mutex.h>
+#include <linux/gfp.h>
 
-/*
- * Represents all cpu's present in the system
- * In systems capable of hotplug, this map could dynamically grow
- * as new cpu's are detected in the system via any platform specific
- * method, such as ACPI for e.g.
- */
-cpumask_t cpu_present_map __read_mostly;
-EXPORT_SYMBOL(cpu_present_map);
-
-#ifndef CONFIG_SMP
+#ifdef CONFIG_SMP
+/* Serializes the updates to cpu_online_mask, cpu_present_mask */
+static DEFINE_MUTEX(cpu_add_remove_lock);
 
 /*
- * Represents all cpu's that are currently online.
+ * The following two API's must be used when attempting
+ * to serialize the updates to cpu_online_mask, cpu_present_mask.
  */
-cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
-EXPORT_SYMBOL(cpu_online_map);
-
-cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
-EXPORT_SYMBOL(cpu_possible_map);
-
-#else /* CONFIG_SMP */
+void cpu_maps_update_begin(void)
+{
+       mutex_lock(&cpu_add_remove_lock);
+}
 
-/* Serializes the updates to cpu_online_map, cpu_present_map */
-static DEFINE_MUTEX(cpu_add_remove_lock);
+void cpu_maps_update_done(void)
+{
+       mutex_unlock(&cpu_add_remove_lock);
+}
 
-static __cpuinitdata RAW_NOTIFIER_HEAD(cpu_chain);
+static RAW_NOTIFIER_HEAD(cpu_chain);
 
 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
  * Should always be manipulated under cpu_add_remove_lock
  */
 static int cpu_hotplug_disabled;
 
+#ifdef CONFIG_HOTPLUG_CPU
+
 static struct {
        struct task_struct *active_writer;
        struct mutex lock; /* Synchronizes accesses to refcount, */
@@ -55,18 +51,11 @@ static struct {
         * an ongoing cpu hotplug operation.
         */
        int refcount;
-} cpu_hotplug;
-
-void __init cpu_hotplug_init(void)
-{
-       cpu_hotplug.active_writer = NULL;
-       mutex_init(&cpu_hotplug.lock);
-       cpu_hotplug.refcount = 0;
-}
-
-cpumask_t cpu_active_map;
-
-#ifdef CONFIG_HOTPLUG_CPU
+} cpu_hotplug = {
+       .active_writer = NULL,
+       .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
+       .refcount = 0,
+};
 
 void get_online_cpus(void)
 {
@@ -92,22 +81,6 @@ void put_online_cpus(void)
 }
 EXPORT_SYMBOL_GPL(put_online_cpus);
 
-#endif /* CONFIG_HOTPLUG_CPU */
-
-/*
- * The following two API's must be used when attempting
- * to serialize the updates to cpu_online_map, cpu_present_map.
- */
-void cpu_maps_update_begin(void)
-{
-       mutex_lock(&cpu_add_remove_lock);
-}
-
-void cpu_maps_update_done(void)
-{
-       mutex_unlock(&cpu_add_remove_lock);
-}
-
 /*
  * This ensures that the hotplug operation can begin only when the
  * refcount goes to zero.
@@ -149,6 +122,12 @@ static void cpu_hotplug_done(void)
        cpu_hotplug.active_writer = NULL;
        mutex_unlock(&cpu_hotplug.lock);
 }
+
+#else /* #if CONFIG_HOTPLUG_CPU */
+static void cpu_hotplug_begin(void) {}
+static void cpu_hotplug_done(void) {}
+#endif /* #else #if CONFIG_HOTPLUG_CPU */
+
 /* Need to know about CPUs going up/down? */
 int __ref register_cpu_notifier(struct notifier_block *nb)
 {
@@ -159,8 +138,28 @@ int __ref register_cpu_notifier(struct notifier_block *nb)
        return ret;
 }
 
+static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
+                       int *nr_calls)
+{
+       int ret;
+
+       ret = __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
+                                       nr_calls);
+
+       return notifier_to_errno(ret);
+}
+
+static int cpu_notify(unsigned long val, void *v)
+{
+       return __cpu_notify(val, v, -1, NULL);
+}
+
 #ifdef CONFIG_HOTPLUG_CPU
 
+static void cpu_notify_nofail(unsigned long val, void *v)
+{
+       BUG_ON(cpu_notify(val, v));
+}
 EXPORT_SYMBOL(register_cpu_notifier);
 
 void __ref unregister_cpu_notifier(struct notifier_block *nb)
@@ -177,13 +176,13 @@ static inline void check_for_tasks(int cpu)
 
        write_lock_irq(&tasklist_lock);
        for_each_process(p) {
-               if (task_cpu(p) == cpu &&
+               if (task_cpu(p) == cpu && p->state == TASK_RUNNING &&
                    (!cputime_eq(p->utime, cputime_zero) ||
                     !cputime_eq(p->stime, cputime_zero)))
-                       printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d\
-                               (state = %ld, flags = %x) \n",
-                                p->comm, task_pid_nr(p), cpu,
-                                p->state, p->flags);
+                       printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d "
+                               "(state = %ld, flags = %x)\n",
+                               p->comm, task_pid_nr(p), cpu,
+                               p->state, p->flags);
        }
        write_unlock_irq(&tasklist_lock);
 }
@@ -199,16 +198,12 @@ static int __ref take_cpu_down(void *_param)
        struct take_cpu_down_param *param = _param;
        int err;
 
-       raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
-                               param->hcpu);
        /* Ensure this CPU doesn't handle any more interrupts. */
        err = __cpu_disable();
        if (err < 0)
                return err;
 
-       /* Force idle task to run as soon as we yield: it should
-          immediately notice cpu is offline and die quickly. */
-       sched_idle_next();
+       cpu_notify(CPU_DYING | param->mod, param->hcpu);
        return 0;
 }
 
@@ -216,7 +211,6 @@ static int __ref take_cpu_down(void *_param)
 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
 {
        int err, nr_calls = 0;
-       cpumask_t old_allowed, tmp;
        void *hcpu = (void *)(long)cpu;
        unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
        struct take_cpu_down_param tcd_param = {
@@ -231,65 +225,53 @@ static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
                return -EINVAL;
 
        cpu_hotplug_begin();
-       err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
-                                       hcpu, -1, &nr_calls);
-       if (err == NOTIFY_BAD) {
+
+       err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
+       if (err) {
                nr_calls--;
-               __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
-                                         hcpu, nr_calls, NULL);
+               __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
                printk("%s: attempt to take down CPU %u failed\n",
                                __func__, cpu);
-               err = -EINVAL;
                goto out_release;
        }
 
-       /* Ensure that we are not runnable on dying cpu */
-       old_allowed = current->cpus_allowed;
-       cpus_setall(tmp);
-       cpu_clear(cpu, tmp);
-       set_cpus_allowed_ptr(current, &tmp);
-       tmp = cpumask_of_cpu(cpu);
-
-       err = __stop_machine(take_cpu_down, &tcd_param, &tmp);
+       err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
        if (err) {
                /* CPU didn't die: tell everyone.  Can't complain. */
-               if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
-                                           hcpu) == NOTIFY_BAD)
-                       BUG();
+               cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
 
-               goto out_allowed;
+               goto out_release;
        }
        BUG_ON(cpu_online(cpu));
 
-       /* Wait for it to sleep (leaving idle task). */
+       /*
+        * The migration_call() CPU_DYING callback will have removed all
+        * runnable tasks from the cpu, there's only the idle task left now
+        * that the migration thread is done doing the stop_machine thing.
+        *
+        * Wait for the stop thread to go away.
+        */
        while (!idle_cpu(cpu))
-               yield();
+               cpu_relax();
 
        /* This actually kills the CPU. */
        __cpu_die(cpu);
 
        /* CPU is completely dead: tell everyone.  Too late to complain. */
-       if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
-                                   hcpu) == NOTIFY_BAD)
-               BUG();
+       cpu_notify_nofail(CPU_DEAD | mod, hcpu);
 
        check_for_tasks(cpu);
 
-out_allowed:
-       set_cpus_allowed_ptr(current, &old_allowed);
 out_release:
        cpu_hotplug_done();
-       if (!err) {
-               if (raw_notifier_call_chain(&cpu_chain, CPU_POST_DEAD | mod,
-                                           hcpu) == NOTIFY_BAD)
-                       BUG();
-       }
+       if (!err)
+               cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
        return err;
 }
 
 int __ref cpu_down(unsigned int cpu)
 {
-       int err = 0;
+       int err;
 
        cpu_maps_update_begin();
 
@@ -298,23 +280,8 @@ int __ref cpu_down(unsigned int cpu)
                goto out;
        }
 
-       cpu_clear(cpu, cpu_active_map);
-
-       /*
-        * Make sure the all cpus did the reschedule and are not
-        * using stale version of the cpu_active_map.
-        * This is not strictly necessary becuase stop_machine()
-        * that we run down the line already provides the required
-        * synchronization. But it's really a side effect and we do not
-        * want to depend on the innards of the stop_machine here.
-        */
-       synchronize_sched();
-
        err = _cpu_down(cpu, 0);
 
-       if (cpu_online(cpu))
-               cpu_set(cpu, cpu_active_map);
-
 out:
        cpu_maps_update_done();
        return err;
@@ -333,13 +300,11 @@ static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
                return -EINVAL;
 
        cpu_hotplug_begin();
-       ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
-                                                       -1, &nr_calls);
-       if (ret == NOTIFY_BAD) {
+       ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
+       if (ret) {
                nr_calls--;
-               printk("%s: attempt to bring up CPU %u failed\n",
+               printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
                                __func__, cpu);
-               ret = -EINVAL;
                goto out_notify;
        }
 
@@ -350,12 +315,11 @@ static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
        BUG_ON(!cpu_online(cpu));
 
        /* Now call notifier in preparation. */
-       raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
+       cpu_notify(CPU_ONLINE | mod, hcpu);
 
 out_notify:
        if (ret != 0)
-               __raw_notifier_call_chain(&cpu_chain,
-                               CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
+               __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
        cpu_hotplug_done();
 
        return ret;
@@ -364,16 +328,44 @@ out_notify:
 int __cpuinit cpu_up(unsigned int cpu)
 {
        int err = 0;
-       if (!cpu_isset(cpu, cpu_possible_map)) {
+
+#ifdef CONFIG_MEMORY_HOTPLUG
+       int nid;
+       pg_data_t       *pgdat;
+#endif
+
+       if (!cpu_possible(cpu)) {
                printk(KERN_ERR "can't online cpu %d because it is not "
                        "configured as may-hotadd at boot time\n", cpu);
-#if defined(CONFIG_IA64) || defined(CONFIG_X86_64) || defined(CONFIG_S390)
+#if defined(CONFIG_IA64)
                printk(KERN_ERR "please check additional_cpus= boot "
                                "parameter\n");
 #endif
                return -EINVAL;
        }
 
+#ifdef CONFIG_MEMORY_HOTPLUG
+       nid = cpu_to_node(cpu);
+       if (!node_online(nid)) {
+               err = mem_online_node(nid);
+               if (err)
+                       return err;
+       }
+
+       pgdat = NODE_DATA(nid);
+       if (!pgdat) {
+               printk(KERN_ERR
+                       "Can't online cpu %d due to NULL pgdat\n", cpu);
+               return -ENOMEM;
+       }
+
+       if (pgdat->node_zonelists->_zonerefs->zone == NULL) {
+               mutex_lock(&zonelists_mutex);
+               build_all_zonelists(NULL);
+               mutex_unlock(&zonelists_mutex);
+       }
+#endif
+
        cpu_maps_update_begin();
 
        if (cpu_hotplug_disabled) {
@@ -383,41 +375,51 @@ int __cpuinit cpu_up(unsigned int cpu)
 
        err = _cpu_up(cpu, 0);
 
-       if (cpu_online(cpu))
-               cpu_set(cpu, cpu_active_map);
-
 out:
        cpu_maps_update_done();
        return err;
 }
 
 #ifdef CONFIG_PM_SLEEP_SMP
-static cpumask_t frozen_cpus;
+static cpumask_var_t frozen_cpus;
+
+void __weak arch_disable_nonboot_cpus_begin(void)
+{
+}
+
+void __weak arch_disable_nonboot_cpus_end(void)
+{
+}
 
 int disable_nonboot_cpus(void)
 {
        int cpu, first_cpu, error = 0;
 
        cpu_maps_update_begin();
-       first_cpu = first_cpu(cpu_online_map);
-       /* We take down all of the non-boot CPUs in one shot to avoid races
+       first_cpu = cpumask_first(cpu_online_mask);
+       /*
+        * We take down all of the non-boot CPUs in one shot to avoid races
         * with the userspace trying to use the CPU hotplug at the same time
         */
-       cpus_clear(frozen_cpus);
+       cpumask_clear(frozen_cpus);
+       arch_disable_nonboot_cpus_begin();
+
        printk("Disabling non-boot CPUs ...\n");
        for_each_online_cpu(cpu) {
                if (cpu == first_cpu)
                        continue;
                error = _cpu_down(cpu, 1);
-               if (!error) {
-                       cpu_set(cpu, frozen_cpus);
-                       printk("CPU%d is down\n", cpu);
-               } else {
+               if (!error)
+                       cpumask_set_cpu(cpu, frozen_cpus);
+               else {
                        printk(KERN_ERR "Error taking CPU%d down: %d\n",
                                cpu, error);
                        break;
                }
        }
+
+       arch_disable_nonboot_cpus_end();
+
        if (!error) {
                BUG_ON(num_online_cpus() > 1);
                /* Make sure the CPUs won't be enabled by someone else */
@@ -429,6 +431,14 @@ int disable_nonboot_cpus(void)
        return error;
 }
 
+void __weak arch_enable_nonboot_cpus_begin(void)
+{
+}
+
+void __weak arch_enable_nonboot_cpus_end(void)
+{
+}
+
 void __ref enable_nonboot_cpus(void)
 {
        int cpu, error;
@@ -436,36 +446,69 @@ void __ref enable_nonboot_cpus(void)
        /* Allow everyone to use the CPU hotplug again */
        cpu_maps_update_begin();
        cpu_hotplug_disabled = 0;
-       if (cpus_empty(frozen_cpus))
+       if (cpumask_empty(frozen_cpus))
                goto out;
 
-       printk("Enabling non-boot CPUs ...\n");
-       for_each_cpu_mask_nr(cpu, frozen_cpus) {
+       printk(KERN_INFO "Enabling non-boot CPUs ...\n");
+
+       arch_enable_nonboot_cpus_begin();
+
+       for_each_cpu(cpu, frozen_cpus) {
                error = _cpu_up(cpu, 1);
                if (!error) {
-                       printk("CPU%d is up\n", cpu);
+                       printk(KERN_INFO "CPU%d is up\n", cpu);
                        continue;
                }
                printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
        }
-       cpus_clear(frozen_cpus);
+
+       arch_enable_nonboot_cpus_end();
+
+       cpumask_clear(frozen_cpus);
 out:
        cpu_maps_update_done();
 }
+
+static int alloc_frozen_cpus(void)
+{
+       if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
+               return -ENOMEM;
+       return 0;
+}
+core_initcall(alloc_frozen_cpus);
 #endif /* CONFIG_PM_SLEEP_SMP */
 
+/**
+ * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
+ * @cpu: cpu that just started
+ *
+ * This function calls the cpu_chain notifiers with CPU_STARTING.
+ * It must be called by the arch code on the new cpu, before the new cpu
+ * enables interrupts and before the "boot" cpu returns from __cpu_up().
+ */
+void __cpuinit notify_cpu_starting(unsigned int cpu)
+{
+       unsigned long val = CPU_STARTING;
+
+#ifdef CONFIG_PM_SLEEP_SMP
+       if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
+               val = CPU_STARTING_FROZEN;
+#endif /* CONFIG_PM_SLEEP_SMP */
+       cpu_notify(val, (void *)(long)cpu);
+}
+
 #endif /* CONFIG_SMP */
 
 /*
  * cpu_bit_bitmap[] is a special, "compressed" data structure that
  * represents all NR_CPUS bits binary values of 1<<nr.
  *
- * It is used by cpumask_of_cpu() to get a constant address to a CPU
+ * It is used by cpumask_of() to get a constant address to a CPU
  * mask value that has a single bit set only.
  */
 
 /* cpu_bit_bitmap[0] is empty - so we can back into it */
-#define MASK_DECLARE_1(x)      [x+1][0] = 1UL << (x)
+#define MASK_DECLARE_1(x)      [x+1][0] = (1UL << (x))
 #define MASK_DECLARE_2(x)      MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
 #define MASK_DECLARE_4(x)      MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
 #define MASK_DECLARE_8(x)      MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
@@ -480,3 +523,74 @@ const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
 #endif
 };
 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
+
+const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
+EXPORT_SYMBOL(cpu_all_bits);
+
+#ifdef CONFIG_INIT_ALL_POSSIBLE
+static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
+       = CPU_BITS_ALL;
+#else
+static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
+#endif
+const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
+EXPORT_SYMBOL(cpu_possible_mask);
+
+static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
+const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
+EXPORT_SYMBOL(cpu_online_mask);
+
+static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
+const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
+EXPORT_SYMBOL(cpu_present_mask);
+
+static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
+const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
+EXPORT_SYMBOL(cpu_active_mask);
+
+void set_cpu_possible(unsigned int cpu, bool possible)
+{
+       if (possible)
+               cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
+       else
+               cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
+}
+
+void set_cpu_present(unsigned int cpu, bool present)
+{
+       if (present)
+               cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
+       else
+               cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
+}
+
+void set_cpu_online(unsigned int cpu, bool online)
+{
+       if (online)
+               cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
+       else
+               cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
+}
+
+void set_cpu_active(unsigned int cpu, bool active)
+{
+       if (active)
+               cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
+       else
+               cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
+}
+
+void init_cpu_present(const struct cpumask *src)
+{
+       cpumask_copy(to_cpumask(cpu_present_bits), src);
+}
+
+void init_cpu_possible(const struct cpumask *src)
+{
+       cpumask_copy(to_cpumask(cpu_possible_bits), src);
+}
+
+void init_cpu_online(const struct cpumask *src)
+{
+       cpumask_copy(to_cpumask(cpu_online_bits), src);
+}