sched: fix mysql+oltp regression
[safe/jmp/linux-2.6] / kernel / power / process.c
index b0edfc6..278946a 100644 (file)
@@ -8,7 +8,6 @@
 
 #undef DEBUG
 
-#include <linux/smp_lock.h>
 #include <linux/interrupt.h>
 #include <linux/suspend.h>
 #include <linux/module.h>
  */
 #define TIMEOUT        (20 * HZ)
 
-
 static inline int freezeable(struct task_struct * p)
 {
-       if ((p == current) || 
+       if ((p == current) ||
            (p->flags & PF_NOFREEZE) ||
-           (p->exit_state == EXIT_ZOMBIE) ||
-           (p->exit_state == EXIT_DEAD) ||
-           (p->state == TASK_STOPPED))
+           (p->exit_state != 0))
                return 0;
        return 1;
 }
 
+/*
+ * freezing is complete, mark current process as frozen
+ */
+static inline void frozen_process(void)
+{
+       if (!unlikely(current->flags & PF_NOFREEZE)) {
+               current->flags |= PF_FROZEN;
+               wmb();
+       }
+       clear_freeze_flag(current);
+}
+
 /* Refrigerator is place where frozen processes are stored :-). */
 void refrigerator(void)
 {
        /* Hmm, should we be allowed to suspend when there are realtime
           processes around? */
        long save;
+
+       task_lock(current);
+       if (freezing(current)) {
+               frozen_process();
+               task_unlock(current);
+       } else {
+               task_unlock(current);
+               return;
+       }
        save = current->state;
        pr_debug("%s entered refrigerator\n", current->comm);
 
-       frozen_process(current);
        spin_lock_irq(&current->sighand->siglock);
        recalc_sigpending(); /* We sent fake signal, clean it up */
        spin_unlock_irq(&current->sighand->siglock);
 
-       while (frozen(current)) {
-               current->state = TASK_UNINTERRUPTIBLE;
+       for (;;) {
+               set_current_state(TASK_UNINTERRUPTIBLE);
+               if (!frozen(current))
+                       break;
                schedule();
        }
        pr_debug("%s left refrigerator\n", current->comm);
-       current->state = save;
+       __set_current_state(save);
 }
 
-static inline void freeze_process(struct task_struct *p)
+static void fake_signal_wake_up(struct task_struct *p)
 {
        unsigned long flags;
 
+       spin_lock_irqsave(&p->sighand->siglock, flags);
+       signal_wake_up(p, 0);
+       spin_unlock_irqrestore(&p->sighand->siglock, flags);
+}
+
+static inline bool should_send_signal(struct task_struct *p)
+{
+       return !(p->flags & PF_FREEZER_NOSIG);
+}
+
+/**
+ *     freeze_task - send a freeze request to given task
+ *     @p: task to send the request to
+ *     @sig_only: if set, the request will only be sent if the task has the
+ *             PF_FREEZER_NOSIG flag unset
+ *     Return value: 'false', if @sig_only is set and the task has
+ *             PF_FREEZER_NOSIG set or the task is frozen, 'true', otherwise
+ *
+ *     The freeze request is sent by setting the tasks's TIF_FREEZE flag and
+ *     either sending a fake signal to it or waking it up, depending on whether
+ *     or not it has PF_FREEZER_NOSIG set.  If @sig_only is set and the task
+ *     has PF_FREEZER_NOSIG set (ie. it is a typical kernel thread), its
+ *     TIF_FREEZE flag will not be set.
+ */
+static bool freeze_task(struct task_struct *p, bool sig_only)
+{
+       /*
+        * We first check if the task is freezing and next if it has already
+        * been frozen to avoid the race with frozen_process() which first marks
+        * the task as frozen and next clears its TIF_FREEZE.
+        */
        if (!freezing(p)) {
-               freeze(p);
-               spin_lock_irqsave(&p->sighand->siglock, flags);
-               signal_wake_up(p, 0);
-               spin_unlock_irqrestore(&p->sighand->siglock, flags);
+               rmb();
+               if (frozen(p))
+                       return false;
+
+               if (!sig_only || should_send_signal(p))
+                       set_freeze_flag(p);
+               else
+                       return false;
+       }
+
+       if (should_send_signal(p)) {
+               if (!signal_pending(p))
+                       fake_signal_wake_up(p);
+       } else if (sig_only) {
+               return false;
+       } else {
+               wake_up_state(p, TASK_INTERRUPTIBLE);
        }
+
+       return true;
 }
 
 static void cancel_freezing(struct task_struct *p)
@@ -72,103 +136,132 @@ static void cancel_freezing(struct task_struct *p)
 
        if (freezing(p)) {
                pr_debug("  clean up: %s\n", p->comm);
-               do_not_freeze(p);
+               clear_freeze_flag(p);
                spin_lock_irqsave(&p->sighand->siglock, flags);
-               recalc_sigpending_tsk(p);
+               recalc_sigpending_and_wake(p);
                spin_unlock_irqrestore(&p->sighand->siglock, flags);
        }
 }
 
-/* 0 = success, else # of processes that we failed to stop */
-int freeze_processes(void)
+static int try_to_freeze_tasks(bool sig_only)
 {
-       int todo, nr_user, user_frozen;
-       unsigned long start_time;
        struct task_struct *g, *p;
+       unsigned long end_time;
+       unsigned int todo;
+       struct timeval start, end;
+       u64 elapsed_csecs64;
+       unsigned int elapsed_csecs;
+
+       do_gettimeofday(&start);
 
-       printk( "Stopping tasks: " );
-       start_time = jiffies;
-       user_frozen = 0;
+       end_time = jiffies + TIMEOUT;
        do {
-               nr_user = todo = 0;
+               todo = 0;
                read_lock(&tasklist_lock);
                do_each_thread(g, p) {
-                       if (!freezeable(p))
+                       if (frozen(p) || !freezeable(p))
                                continue;
-                       if (frozen(p))
-                               continue;
-                       if (p->state == TASK_TRACED && frozen(p->parent)) {
-                               cancel_freezing(p);
+
+                       if (!freeze_task(p, sig_only))
                                continue;
-                       }
-                       if (p->mm && !(p->flags & PF_BORROWED_MM)) {
-                               /* The task is a user-space one.
-                                * Freeze it unless there's a vfork completion
-                                * pending
-                                */
-                               if (!p->vfork_done)
-                                       freeze_process(p);
-                               nr_user++;
-                       } else {
-                               /* Freeze only if the user space is frozen */
-                               if (user_frozen)
-                                       freeze_process(p);
+
+                       /*
+                        * Now that we've done set_freeze_flag, don't
+                        * perturb a task in TASK_STOPPED or TASK_TRACED.
+                        * It is "frozen enough".  If the task does wake
+                        * up, it will immediately call try_to_freeze.
+                        */
+                       if (!task_is_stopped_or_traced(p) &&
+                           !freezer_should_skip(p))
                                todo++;
-                       }
                } while_each_thread(g, p);
                read_unlock(&tasklist_lock);
-               todo += nr_user;
-               if (!user_frozen && !nr_user) {
-                       sys_sync();
-                       start_time = jiffies;
-               }
-               user_frozen = !nr_user;
                yield();                        /* Yield is okay here */
-               if (todo && time_after(jiffies, start_time + TIMEOUT))
+               if (time_after(jiffies, end_time))
                        break;
-       } while(todo);
+       } while (todo);
+
+       do_gettimeofday(&end);
+       elapsed_csecs64 = timeval_to_ns(&end) - timeval_to_ns(&start);
+       do_div(elapsed_csecs64, NSEC_PER_SEC / 100);
+       elapsed_csecs = elapsed_csecs64;
 
-       /* This does not unfreeze processes that are already frozen
-        * (we have slightly ugly calling convention in that respect,
-        * and caller must call thaw_processes() if something fails),
-        * but it cleans up leftover PF_FREEZE requests.
-        */
        if (todo) {
-               printk( "\n" );
-               printk(KERN_ERR " stopping tasks timed out "
-                       "after %d seconds (%d tasks remaining):\n",
-                       TIMEOUT / HZ, todo);
+               /* This does not unfreeze processes that are already frozen
+                * (we have slightly ugly calling convention in that respect,
+                * and caller must call thaw_processes() if something fails),
+                * but it cleans up leftover PF_FREEZE requests.
+                */
+               printk("\n");
+               printk(KERN_ERR "Freezing of tasks failed after %d.%02d seconds "
+                               "(%d tasks refusing to freeze):\n",
+                               elapsed_csecs / 100, elapsed_csecs % 100, todo);
+               show_state();
                read_lock(&tasklist_lock);
                do_each_thread(g, p) {
-                       if (freezeable(p) && !frozen(p))
-                               printk(KERN_ERR "  %s\n", p->comm);
+                       task_lock(p);
+                       if (freezing(p) && !freezer_should_skip(p))
+                               printk(KERN_ERR " %s\n", p->comm);
                        cancel_freezing(p);
+                       task_unlock(p);
                } while_each_thread(g, p);
                read_unlock(&tasklist_lock);
-               return todo;
+       } else {
+               printk("(elapsed %d.%02d seconds) ", elapsed_csecs / 100,
+                       elapsed_csecs % 100);
        }
 
-       printk( "|\n" );
+       return todo ? -EBUSY : 0;
+}
+
+/**
+ *     freeze_processes - tell processes to enter the refrigerator
+ */
+int freeze_processes(void)
+{
+       int error;
+
+       printk("Freezing user space processes ... ");
+       error = try_to_freeze_tasks(true);
+       if (error)
+               goto Exit;
+       printk("done.\n");
+
+       printk("Freezing remaining freezable tasks ... ");
+       error = try_to_freeze_tasks(false);
+       if (error)
+               goto Exit;
+       printk("done.");
+ Exit:
        BUG_ON(in_atomic());
-       return 0;
+       printk("\n");
+       return error;
 }
 
-void thaw_processes(void)
+static void thaw_tasks(bool nosig_only)
 {
        struct task_struct *g, *p;
 
-       printk( "Restarting tasks..." );
        read_lock(&tasklist_lock);
        do_each_thread(g, p) {
                if (!freezeable(p))
                        continue;
-               if (!thaw_process(p))
-                       printk(KERN_INFO " Strange, %s not stopped\n", p->comm );
-       } while_each_thread(g, p);
 
+               if (nosig_only && should_send_signal(p))
+                       continue;
+
+               thaw_process(p);
+       } while_each_thread(g, p);
        read_unlock(&tasklist_lock);
+}
+
+void thaw_processes(void)
+{
+       printk("Restarting tasks ... ");
+       thaw_tasks(true);
+       thaw_tasks(false);
        schedule();
-       printk( " done\n" );
+       printk("done.\n");
 }
 
 EXPORT_SYMBOL(refrigerator);