sysctl extern cleanup: module
[linux-2.6.git] / kernel / cpu.c
index 3619e93..f8cced2 100644 (file)
 #include <linux/module.h>
 #include <linux/kthread.h>
 #include <linux/stop_machine.h>
-#include <asm/semaphore.h>
+#include <linux/mutex.h>
 
-/* This protects CPUs going up and down... */
-DECLARE_MUTEX(cpucontrol);
-EXPORT_SYMBOL_GPL(cpucontrol);
+#ifdef CONFIG_SMP
+/* Serializes the updates to cpu_online_mask, cpu_present_mask */
+static DEFINE_MUTEX(cpu_add_remove_lock);
 
-static struct notifier_block *cpu_chain;
+static __cpuinitdata 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;
+
+static struct {
+       struct task_struct *active_writer;
+       struct mutex lock; /* Synchronizes accesses to refcount, */
+       /*
+        * Also blocks the new readers during
+        * an ongoing cpu hotplug operation.
+        */
+       int refcount;
+} cpu_hotplug = {
+       .active_writer = NULL,
+       .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
+       .refcount = 0,
+};
+
+#ifdef CONFIG_HOTPLUG_CPU
+
+void get_online_cpus(void)
+{
+       might_sleep();
+       if (cpu_hotplug.active_writer == current)
+               return;
+       mutex_lock(&cpu_hotplug.lock);
+       cpu_hotplug.refcount++;
+       mutex_unlock(&cpu_hotplug.lock);
+
+}
+EXPORT_SYMBOL_GPL(get_online_cpus);
+
+void put_online_cpus(void)
+{
+       if (cpu_hotplug.active_writer == current)
+               return;
+       mutex_lock(&cpu_hotplug.lock);
+       if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
+               wake_up_process(cpu_hotplug.active_writer);
+       mutex_unlock(&cpu_hotplug.lock);
+
+}
+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_mask, cpu_present_mask.
+ */
+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.
+ *
+ * Note that during a cpu-hotplug operation, the new readers, if any,
+ * will be blocked by the cpu_hotplug.lock
+ *
+ * Since cpu_hotplug_begin() is always called after invoking
+ * cpu_maps_update_begin(), we can be sure that only one writer is active.
+ *
+ * Note that theoretically, there is a possibility of a livelock:
+ * - Refcount goes to zero, last reader wakes up the sleeping
+ *   writer.
+ * - Last reader unlocks the cpu_hotplug.lock.
+ * - A new reader arrives at this moment, bumps up the refcount.
+ * - The writer acquires the cpu_hotplug.lock finds the refcount
+ *   non zero and goes to sleep again.
+ *
+ * However, this is very difficult to achieve in practice since
+ * get_online_cpus() not an api which is called all that often.
+ *
+ */
+static void cpu_hotplug_begin(void)
+{
+       cpu_hotplug.active_writer = current;
+
+       for (;;) {
+               mutex_lock(&cpu_hotplug.lock);
+               if (likely(!cpu_hotplug.refcount))
+                       break;
+               __set_current_state(TASK_UNINTERRUPTIBLE);
+               mutex_unlock(&cpu_hotplug.lock);
+               schedule();
+       }
+}
+
+static void cpu_hotplug_done(void)
+{
+       cpu_hotplug.active_writer = NULL;
+       mutex_unlock(&cpu_hotplug.lock);
+}
 /* Need to know about CPUs going up/down? */
-int register_cpu_notifier(struct notifier_block *nb)
+int __ref register_cpu_notifier(struct notifier_block *nb)
 {
        int ret;
-
-       if ((ret = down_interruptible(&cpucontrol)) != 0)
-               return ret;
-       ret = notifier_chain_register(&cpu_chain, nb);
-       up(&cpucontrol);
+       cpu_maps_update_begin();
+       ret = raw_notifier_chain_register(&cpu_chain, nb);
+       cpu_maps_update_done();
        return ret;
 }
+
+#ifdef CONFIG_HOTPLUG_CPU
+
 EXPORT_SYMBOL(register_cpu_notifier);
 
-void unregister_cpu_notifier(struct notifier_block *nb)
+void __ref unregister_cpu_notifier(struct notifier_block *nb)
 {
-       down(&cpucontrol);
-       notifier_chain_unregister(&cpu_chain, nb);
-       up(&cpucontrol);
+       cpu_maps_update_begin();
+       raw_notifier_chain_unregister(&cpu_chain, nb);
+       cpu_maps_update_done();
 }
 EXPORT_SYMBOL(unregister_cpu_notifier);
 
-#ifdef CONFIG_HOTPLUG_CPU
 static inline void check_for_tasks(int cpu)
 {
        struct task_struct *p;
 
        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 = %lx) \n",
-                                p->comm, p->pid, 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);
 }
 
+struct take_cpu_down_param {
+       unsigned long mod;
+       void *hcpu;
+};
+
 /* Take this CPU down. */
-static int take_cpu_down(void *unused)
+static int __ref take_cpu_down(void *_param)
 {
+       struct take_cpu_down_param *param = _param;
        int err;
 
        /* Ensure this CPU doesn't handle any more interrupts. */
@@ -69,59 +178,67 @@ static int take_cpu_down(void *unused)
        if (err < 0)
                return err;
 
+       raw_notifier_call_chain(&cpu_chain, CPU_DYING | param->mod,
+                               param->hcpu);
+
        /* Force idle task to run as soon as we yield: it should
           immediately notice cpu is offline and die quickly. */
        sched_idle_next();
        return 0;
 }
 
-int cpu_down(unsigned int cpu)
+/* Requires cpu_add_remove_lock to be held */
+static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
 {
-       int err;
-       struct task_struct *p;
-       cpumask_t old_allowed, tmp;
+       int err, nr_calls = 0;
+       cpumask_var_t old_allowed;
+       void *hcpu = (void *)(long)cpu;
+       unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
+       struct take_cpu_down_param tcd_param = {
+               .mod = mod,
+               .hcpu = hcpu,
+       };
 
-       if ((err = lock_cpu_hotplug_interruptible()) != 0)
-               return err;
+       if (num_online_cpus() == 1)
+               return -EBUSY;
 
-       if (num_online_cpus() == 1) {
-               err = -EBUSY;
-               goto out;
-       }
+       if (!cpu_online(cpu))
+               return -EINVAL;
 
-       if (!cpu_online(cpu)) {
-               err = -EINVAL;
-               goto out;
-       }
+       if (!alloc_cpumask_var(&old_allowed, GFP_KERNEL))
+               return -ENOMEM;
 
-       err = notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE,
-                                               (void *)(long)cpu);
+       cpu_hotplug_begin();
+       set_cpu_active(cpu, false);
+       err = __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_PREPARE | mod,
+                                       hcpu, -1, &nr_calls);
        if (err == NOTIFY_BAD) {
+               set_cpu_active(cpu, true);
+
+               nr_calls--;
+               __raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
+                                         hcpu, nr_calls, NULL);
                printk("%s: attempt to take down CPU %u failed\n",
-                               __FUNCTION__, cpu);
+                               __func__, cpu);
                err = -EINVAL;
-               goto out;
+               goto out_release;
        }
 
        /* Ensure that we are not runnable on dying cpu */
-       old_allowed = current->cpus_allowed;
-       tmp = CPU_MASK_ALL;
-       cpu_clear(cpu, tmp);
-       set_cpus_allowed(current, tmp);
+       cpumask_copy(old_allowed, &current->cpus_allowed);
+       set_cpus_allowed_ptr(current, cpu_active_mask);
 
-       p = __stop_machine_run(take_cpu_down, NULL, cpu);
-       if (IS_ERR(p)) {
+       err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
+       if (err) {
+               set_cpu_active(cpu, true);
                /* CPU didn't die: tell everyone.  Can't complain. */
-               if (notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED,
-                               (void *)(long)cpu) == NOTIFY_BAD)
+               if (raw_notifier_call_chain(&cpu_chain, CPU_DOWN_FAILED | mod,
+                                           hcpu) == NOTIFY_BAD)
                        BUG();
 
-               err = PTR_ERR(p);
                goto out_allowed;
        }
-
-       if (cpu_online(cpu))
-               goto out_thread;
+       BUG_ON(cpu_online(cpu));
 
        /* Wait for it to sleep (leaving idle task). */
        while (!idle_cpu(cpu))
@@ -130,43 +247,67 @@ int cpu_down(unsigned int cpu)
        /* This actually kills the CPU. */
        __cpu_die(cpu);
 
-       /* Move it here so it can run. */
-       kthread_bind(p, get_cpu());
-       put_cpu();
-
        /* CPU is completely dead: tell everyone.  Too late to complain. */
-       if (notifier_call_chain(&cpu_chain, CPU_DEAD, (void *)(long)cpu)
-           == NOTIFY_BAD)
+       if (raw_notifier_call_chain(&cpu_chain, CPU_DEAD | mod,
+                                   hcpu) == NOTIFY_BAD)
                BUG();
 
        check_for_tasks(cpu);
 
-out_thread:
-       err = kthread_stop(p);
 out_allowed:
-       set_cpus_allowed(current, old_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();
+       }
+       free_cpumask_var(old_allowed);
+       return err;
+}
+
+int __ref cpu_down(unsigned int cpu)
+{
+       int err;
+
+       err = stop_machine_create();
+       if (err)
+               return err;
+       cpu_maps_update_begin();
+
+       if (cpu_hotplug_disabled) {
+               err = -EBUSY;
+               goto out;
+       }
+
+       err = _cpu_down(cpu, 0);
+
 out:
-       unlock_cpu_hotplug();
+       cpu_maps_update_done();
+       stop_machine_destroy();
        return err;
 }
+EXPORT_SYMBOL(cpu_down);
 #endif /*CONFIG_HOTPLUG_CPU*/
 
-int __devinit cpu_up(unsigned int cpu)
+/* Requires cpu_add_remove_lock to be held */
+static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
 {
-       int ret;
+       int ret, nr_calls = 0;
        void *hcpu = (void *)(long)cpu;
+       unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
 
-       if ((ret = down_interruptible(&cpucontrol)) != 0)
-               return ret;
+       if (cpu_online(cpu) || !cpu_present(cpu))
+               return -EINVAL;
 
-       if (cpu_online(cpu) || !cpu_present(cpu)) {
-               ret = -EINVAL;
-               goto out;
-       }
-       ret = notifier_call_chain(&cpu_chain, CPU_UP_PREPARE, hcpu);
+       cpu_hotplug_begin();
+       ret = __raw_notifier_call_chain(&cpu_chain, CPU_UP_PREPARE | mod, hcpu,
+                                                       -1, &nr_calls);
        if (ret == NOTIFY_BAD) {
+               nr_calls--;
                printk("%s: attempt to bring up CPU %u failed\n",
-                               __FUNCTION__, cpu);
+                               __func__, cpu);
                ret = -EINVAL;
                goto out_notify;
        }
@@ -175,16 +316,253 @@ int __devinit cpu_up(unsigned int cpu)
        ret = __cpu_up(cpu);
        if (ret != 0)
                goto out_notify;
-       if (!cpu_online(cpu))
-               BUG();
+       BUG_ON(!cpu_online(cpu));
+
+       set_cpu_active(cpu, true);
 
        /* Now call notifier in preparation. */
-       notifier_call_chain(&cpu_chain, CPU_ONLINE, hcpu);
+       raw_notifier_call_chain(&cpu_chain, CPU_ONLINE | mod, hcpu);
 
 out_notify:
        if (ret != 0)
-               notifier_call_chain(&cpu_chain, CPU_UP_CANCELED, hcpu);
-out:
-       up(&cpucontrol);
+               __raw_notifier_call_chain(&cpu_chain,
+                               CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
+       cpu_hotplug_done();
+
        return ret;
 }
+
+int __cpuinit cpu_up(unsigned int cpu)
+{
+       int err = 0;
+       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)
+               printk(KERN_ERR "please check additional_cpus= boot "
+                               "parameter\n");
+#endif
+               return -EINVAL;
+       }
+
+       cpu_maps_update_begin();
+
+       if (cpu_hotplug_disabled) {
+               err = -EBUSY;
+               goto out;
+       }
+
+       err = _cpu_up(cpu, 0);
+
+out:
+       cpu_maps_update_done();
+       return err;
+}
+
+#ifdef CONFIG_PM_SLEEP_SMP
+static cpumask_var_t frozen_cpus;
+
+int disable_nonboot_cpus(void)
+{
+       int cpu, first_cpu, error;
+
+       error = stop_machine_create();
+       if (error)
+               return error;
+       cpu_maps_update_begin();
+       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
+        */
+       cpumask_clear(frozen_cpus);
+
+       printk("Disabling non-boot CPUs ...\n");
+       for_each_online_cpu(cpu) {
+               if (cpu == first_cpu)
+                       continue;
+               error = _cpu_down(cpu, 1);
+               if (!error)
+                       cpumask_set_cpu(cpu, frozen_cpus);
+               else {
+                       printk(KERN_ERR "Error taking CPU%d down: %d\n",
+                               cpu, error);
+                       break;
+               }
+       }
+
+       if (!error) {
+               BUG_ON(num_online_cpus() > 1);
+               /* Make sure the CPUs won't be enabled by someone else */
+               cpu_hotplug_disabled = 1;
+       } else {
+               printk(KERN_ERR "Non-boot CPUs are not disabled\n");
+       }
+       cpu_maps_update_done();
+       stop_machine_destroy();
+       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;
+
+       /* Allow everyone to use the CPU hotplug again */
+       cpu_maps_update_begin();
+       cpu_hotplug_disabled = 0;
+       if (cpumask_empty(frozen_cpus))
+               goto out;
+
+       printk("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);
+                       continue;
+               }
+               printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
+       }
+
+       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 */
+       raw_notifier_call_chain(&cpu_chain, 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() 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_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)
+
+const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
+
+       MASK_DECLARE_8(0),      MASK_DECLARE_8(8),
+       MASK_DECLARE_8(16),     MASK_DECLARE_8(24),
+#if BITS_PER_LONG > 32
+       MASK_DECLARE_8(32),     MASK_DECLARE_8(40),
+       MASK_DECLARE_8(48),     MASK_DECLARE_8(56),
+#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);
+}