netxen: enable tso6, intr coalescing.
[pandora-kernel.git] / kernel / exit.c
index ceb2587..93d2711 100644 (file)
@@ -71,7 +71,7 @@ static void __unhash_process(struct task_struct *p)
                __get_cpu_var(process_counts)--;
        }
        list_del_rcu(&p->thread_group);
-       remove_parent(p);
+       list_del_init(&p->sibling);
 }
 
 /*
@@ -152,6 +152,18 @@ static void delayed_put_task_struct(struct rcu_head *rhp)
        put_task_struct(container_of(rhp, struct task_struct, rcu));
 }
 
+/*
+ * Do final ptrace-related cleanup of a zombie being reaped.
+ *
+ * Called with write_lock(&tasklist_lock) held.
+ */
+static void ptrace_release_task(struct task_struct *p)
+{
+       BUG_ON(!list_empty(&p->ptraced));
+       ptrace_unlink(p);
+       BUG_ON(!list_empty(&p->ptrace_entry));
+}
+
 void release_task(struct task_struct * p)
 {
        struct task_struct *leader;
@@ -160,8 +172,7 @@ repeat:
        atomic_dec(&p->user->processes);
        proc_flush_task(p);
        write_lock_irq(&tasklist_lock);
-       ptrace_unlink(p);
-       BUG_ON(!list_empty(&p->ptrace_list) || !list_empty(&p->ptrace_children));
+       ptrace_release_task(p);
        __exit_signal(p);
 
        /*
@@ -315,9 +326,8 @@ static void reparent_to_kthreadd(void)
 
        ptrace_unlink(current);
        /* Reparent to init */
-       remove_parent(current);
        current->real_parent = current->parent = kthreadd_task;
-       add_parent(current);
+       list_move_tail(&current->sibling, &current->real_parent->children);
 
        /* Set the exit signal to SIGCHLD so we signal init on exit */
        current->exit_signal = SIGCHLD;
@@ -692,37 +702,97 @@ static void exit_mm(struct task_struct * tsk)
        mmput(mm);
 }
 
-static void
-reparent_thread(struct task_struct *p, struct task_struct *father, int traced)
+/*
+ * Return nonzero if @parent's children should reap themselves.
+ *
+ * Called with write_lock_irq(&tasklist_lock) held.
+ */
+static int ignoring_children(struct task_struct *parent)
 {
-       if (p->pdeath_signal)
-               /* We already hold the tasklist_lock here.  */
-               group_send_sig_info(p->pdeath_signal, SEND_SIG_NOINFO, p);
+       int ret;
+       struct sighand_struct *psig = parent->sighand;
+       unsigned long flags;
+       spin_lock_irqsave(&psig->siglock, flags);
+       ret = (psig->action[SIGCHLD-1].sa.sa_handler == SIG_IGN ||
+              (psig->action[SIGCHLD-1].sa.sa_flags & SA_NOCLDWAIT));
+       spin_unlock_irqrestore(&psig->siglock, flags);
+       return ret;
+}
 
-       /* Move the child from its dying parent to the new one.  */
-       if (unlikely(traced)) {
-               /* Preserve ptrace links if someone else is tracing this child.  */
-               list_del_init(&p->ptrace_list);
-               if (ptrace_reparented(p))
-                       list_add(&p->ptrace_list, &p->real_parent->ptrace_children);
-       } else {
-               /* If this child is being traced, then we're the one tracing it
-                * anyway, so let go of it.
+/*
+ * Detach all tasks we were using ptrace on.
+ * Any that need to be release_task'd are put on the @dead list.
+ *
+ * Called with write_lock(&tasklist_lock) held.
+ */
+static void ptrace_exit(struct task_struct *parent, struct list_head *dead)
+{
+       struct task_struct *p, *n;
+       int ign = -1;
+
+       list_for_each_entry_safe(p, n, &parent->ptraced, ptrace_entry) {
+               __ptrace_unlink(p);
+
+               if (p->exit_state != EXIT_ZOMBIE)
+                       continue;
+
+               /*
+                * If it's a zombie, our attachedness prevented normal
+                * parent notification or self-reaping.  Do notification
+                * now if it would have happened earlier.  If it should
+                * reap itself, add it to the @dead list.  We can't call
+                * release_task() here because we already hold tasklist_lock.
+                *
+                * If it's our own child, there is no notification to do.
+                * But if our normal children self-reap, then this child
+                * was prevented by ptrace and we must reap it now.
                 */
-               p->ptrace = 0;
-               remove_parent(p);
-               p->parent = p->real_parent;
-               add_parent(p);
+               if (!task_detached(p) && thread_group_empty(p)) {
+                       if (!same_thread_group(p->real_parent, parent))
+                               do_notify_parent(p, p->exit_signal);
+                       else {
+                               if (ign < 0)
+                                       ign = ignoring_children(parent);
+                               if (ign)
+                                       p->exit_signal = -1;
+                       }
+               }
 
-               if (task_is_traced(p)) {
+               if (task_detached(p)) {
                        /*
-                        * If it was at a trace stop, turn it into
-                        * a normal stop since it's no longer being
-                        * traced.
+                        * Mark it as in the process of being reaped.
                         */
-                       ptrace_untrace(p);
+                       p->exit_state = EXIT_DEAD;
+                       list_add(&p->ptrace_entry, dead);
                }
        }
+}
+
+/*
+ * Finish up exit-time ptrace cleanup.
+ *
+ * Called without locks.
+ */
+static void ptrace_exit_finish(struct task_struct *parent,
+                              struct list_head *dead)
+{
+       struct task_struct *p, *n;
+
+       BUG_ON(!list_empty(&parent->ptraced));
+
+       list_for_each_entry_safe(p, n, dead, ptrace_entry) {
+               list_del_init(&p->ptrace_entry);
+               release_task(p);
+       }
+}
+
+static void reparent_thread(struct task_struct *p, struct task_struct *father)
+{
+       if (p->pdeath_signal)
+               /* We already hold the tasklist_lock here.  */
+               group_send_sig_info(p->pdeath_signal, SEND_SIG_NOINFO, p);
+
+       list_move_tail(&p->sibling, &p->real_parent->children);
 
        /* If this is a threaded reparent there is no need to
         * notify anyone anything has happened.
@@ -737,7 +807,8 @@ reparent_thread(struct task_struct *p, struct task_struct *father, int traced)
        /* If we'd notified the old parent about this child's death,
         * also notify the new parent.
         */
-       if (!traced && p->exit_state == EXIT_ZOMBIE &&
+       if (!ptrace_reparented(p) &&
+           p->exit_state == EXIT_ZOMBIE &&
            !task_detached(p) && thread_group_empty(p))
                do_notify_parent(p, p->exit_signal);
 
@@ -754,12 +825,15 @@ reparent_thread(struct task_struct *p, struct task_struct *father, int traced)
 static void forget_original_parent(struct task_struct *father)
 {
        struct task_struct *p, *n, *reaper = father;
-       struct list_head ptrace_dead;
-
-       INIT_LIST_HEAD(&ptrace_dead);
+       LIST_HEAD(ptrace_dead);
 
        write_lock_irq(&tasklist_lock);
 
+       /*
+        * First clean up ptrace if we were using it.
+        */
+       ptrace_exit(father, &ptrace_dead);
+
        do {
                reaper = next_thread(reaper);
                if (reaper == father) {
@@ -768,58 +842,19 @@ static void forget_original_parent(struct task_struct *father)
                }
        } while (reaper->flags & PF_EXITING);
 
-       /*
-        * There are only two places where our children can be:
-        *
-        * - in our child list
-        * - in our ptraced child list
-        *
-        * Search them and reparent children.
-        */
        list_for_each_entry_safe(p, n, &father->children, sibling) {
-               int ptrace;
-
-               ptrace = p->ptrace;
-
-               /* if father isn't the real parent, then ptrace must be enabled */
-               BUG_ON(father != p->real_parent && !ptrace);
-
-               if (father == p->real_parent) {
-                       /* reparent with a reaper, real father it's us */
-                       p->real_parent = reaper;
-                       reparent_thread(p, father, 0);
-               } else {
-                       /* reparent ptraced task to its real parent */
-                       __ptrace_unlink (p);
-                       if (p->exit_state == EXIT_ZOMBIE && !task_detached(p) &&
-                           thread_group_empty(p))
-                               do_notify_parent(p, p->exit_signal);
-               }
-
-               /*
-                * if the ptraced child is a detached zombie we must collect
-                * it before we exit, or it will remain zombie forever since
-                * we prevented it from self-reap itself while it was being
-                * traced by us, to be able to see it in wait4.
-                */
-               if (unlikely(ptrace && p->exit_state == EXIT_ZOMBIE && task_detached(p)))
-                       list_add(&p->ptrace_list, &ptrace_dead);
-       }
-
-       list_for_each_entry_safe(p, n, &father->ptrace_children, ptrace_list) {
                p->real_parent = reaper;
-               reparent_thread(p, father, 1);
+               if (p->parent == father) {
+                       BUG_ON(p->ptrace);
+                       p->parent = p->real_parent;
+               }
+               reparent_thread(p, father);
        }
 
        write_unlock_irq(&tasklist_lock);
        BUG_ON(!list_empty(&father->children));
-       BUG_ON(!list_empty(&father->ptrace_children));
-
-       list_for_each_entry_safe(p, n, &ptrace_dead, ptrace_list) {
-               list_del_init(&p->ptrace_list);
-               release_task(p);
-       }
 
+       ptrace_exit_finish(father, &ptrace_dead);
 }
 
 /*
@@ -1180,13 +1215,6 @@ static int eligible_child(enum pid_type type, struct pid *pid, int options,
                        return 0;
        }
 
-       /*
-        * Do not consider detached threads that are
-        * not ptraced:
-        */
-       if (task_detached(p) && !p->ptrace)
-               return 0;
-
        /* Wait for all children (clone and not) if __WALL is set;
         * otherwise, wait for clone children *only* if __WCLONE is
         * set; otherwise, wait for non-clone children *only*.  (Note:
@@ -1197,14 +1225,10 @@ static int eligible_child(enum pid_type type, struct pid *pid, int options,
                return 0;
 
        err = security_task_wait(p);
-       if (likely(!err))
-               return 1;
+       if (err)
+               return err;
 
-       if (type != PIDTYPE_PID)
-               return 0;
-       /* This child was explicitly requested, abort */
-       read_unlock(&tasklist_lock);
-       return err;
+       return 1;
 }
 
 static int wait_noreap_copyout(struct task_struct *p, pid_t pid, uid_t uid,
@@ -1238,7 +1262,7 @@ static int wait_noreap_copyout(struct task_struct *p, pid_t pid, uid_t uid,
  * the lock and this task is uninteresting.  If we return nonzero, we have
  * released the lock and the system call should return.
  */
-static int wait_task_zombie(struct task_struct *p, int noreap,
+static int wait_task_zombie(struct task_struct *p, int options,
                            struct siginfo __user *infop,
                            int __user *stat_addr, struct rusage __user *ru)
 {
@@ -1246,7 +1270,10 @@ static int wait_task_zombie(struct task_struct *p, int noreap,
        int retval, status, traced;
        pid_t pid = task_pid_vnr(p);
 
-       if (unlikely(noreap)) {
+       if (!likely(options & WEXITED))
+               return 0;
+
+       if (unlikely(options & WNOWAIT)) {
                uid_t uid = p->uid;
                int exit_code = p->exit_code;
                int why, status;
@@ -1396,21 +1423,24 @@ static int wait_task_zombie(struct task_struct *p, int noreap,
  * the lock and this task is uninteresting.  If we return nonzero, we have
  * released the lock and the system call should return.
  */
-static int wait_task_stopped(struct task_struct *p,
-                            int noreap, struct siginfo __user *infop,
+static int wait_task_stopped(int ptrace, struct task_struct *p,
+                            int options, struct siginfo __user *infop,
                             int __user *stat_addr, struct rusage __user *ru)
 {
        int retval, exit_code, why;
        uid_t uid = 0; /* unneeded, required by compiler */
        pid_t pid;
 
+       if (!(options & WUNTRACED))
+               return 0;
+
        exit_code = 0;
        spin_lock_irq(&p->sighand->siglock);
 
        if (unlikely(!task_is_stopped_or_traced(p)))
                goto unlock_sig;
 
-       if (!(p->ptrace & PT_PTRACED) && p->signal->group_stop_count > 0)
+       if (!ptrace && p->signal->group_stop_count > 0)
                /*
                 * A group stop is in progress and this is the group leader.
                 * We won't report until all threads have stopped.
@@ -1421,7 +1451,7 @@ static int wait_task_stopped(struct task_struct *p,
        if (!exit_code)
                goto unlock_sig;
 
-       if (!noreap)
+       if (!unlikely(options & WNOWAIT))
                p->exit_code = 0;
 
        uid = p->uid;
@@ -1439,10 +1469,10 @@ unlock_sig:
         */
        get_task_struct(p);
        pid = task_pid_vnr(p);
-       why = (p->ptrace & PT_PTRACED) ? CLD_TRAPPED : CLD_STOPPED;
+       why = ptrace ? CLD_TRAPPED : CLD_STOPPED;
        read_unlock(&tasklist_lock);
 
-       if (unlikely(noreap))
+       if (unlikely(options & WNOWAIT))
                return wait_noreap_copyout(p, pid, uid,
                                           why, exit_code,
                                           infop, ru);
@@ -1476,7 +1506,7 @@ unlock_sig:
  * the lock and this task is uninteresting.  If we return nonzero, we have
  * released the lock and the system call should return.
  */
-static int wait_task_continued(struct task_struct *p, int noreap,
+static int wait_task_continued(struct task_struct *p, int options,
                               struct siginfo __user *infop,
                               int __user *stat_addr, struct rusage __user *ru)
 {
@@ -1484,6 +1514,9 @@ static int wait_task_continued(struct task_struct *p, int noreap,
        pid_t pid;
        uid_t uid;
 
+       if (!unlikely(options & WCONTINUED))
+               return 0;
+
        if (!(p->signal->flags & SIGNAL_STOP_CONTINUED))
                return 0;
 
@@ -1493,7 +1526,7 @@ static int wait_task_continued(struct task_struct *p, int noreap,
                spin_unlock_irq(&p->sighand->siglock);
                return 0;
        }
-       if (!noreap)
+       if (!unlikely(options & WNOWAIT))
                p->signal->flags &= ~SIGNAL_STOP_CONTINUED;
        spin_unlock_irq(&p->sighand->siglock);
 
@@ -1519,89 +1552,161 @@ static int wait_task_continued(struct task_struct *p, int noreap,
        return retval;
 }
 
+/*
+ * Consider @p for a wait by @parent.
+ *
+ * -ECHILD should be in *@notask_error before the first call.
+ * Returns nonzero for a final return, when we have unlocked tasklist_lock.
+ * Returns zero if the search for a child should continue;
+ * then *@notask_error is 0 if @p is an eligible child,
+ * or another error from security_task_wait(), or still -ECHILD.
+ */
+static int wait_consider_task(struct task_struct *parent, int ptrace,
+                             struct task_struct *p, int *notask_error,
+                             enum pid_type type, struct pid *pid, int options,
+                             struct siginfo __user *infop,
+                             int __user *stat_addr, struct rusage __user *ru)
+{
+       int ret = eligible_child(type, pid, options, p);
+       if (!ret)
+               return ret;
+
+       if (unlikely(ret < 0)) {
+               /*
+                * If we have not yet seen any eligible child,
+                * then let this error code replace -ECHILD.
+                * A permission error will give the user a clue
+                * to look for security policy problems, rather
+                * than for mysterious wait bugs.
+                */
+               if (*notask_error)
+                       *notask_error = ret;
+       }
+
+       if (likely(!ptrace) && unlikely(p->ptrace)) {
+               /*
+                * This child is hidden by ptrace.
+                * We aren't allowed to see it now, but eventually we will.
+                */
+               *notask_error = 0;
+               return 0;
+       }
+
+       if (p->exit_state == EXIT_DEAD)
+               return 0;
+
+       /*
+        * We don't reap group leaders with subthreads.
+        */
+       if (p->exit_state == EXIT_ZOMBIE && !delay_group_leader(p))
+               return wait_task_zombie(p, options, infop, stat_addr, ru);
+
+       /*
+        * It's stopped or running now, so it might
+        * later continue, exit, or stop again.
+        */
+       *notask_error = 0;
+
+       if (task_is_stopped_or_traced(p))
+               return wait_task_stopped(ptrace, p, options,
+                                        infop, stat_addr, ru);
+
+       return wait_task_continued(p, options, infop, stat_addr, ru);
+}
+
+/*
+ * Do the work of do_wait() for one thread in the group, @tsk.
+ *
+ * -ECHILD should be in *@notask_error before the first call.
+ * Returns nonzero for a final return, when we have unlocked tasklist_lock.
+ * Returns zero if the search for a child should continue; then
+ * *@notask_error is 0 if there were any eligible children,
+ * or another error from security_task_wait(), or still -ECHILD.
+ */
+static int do_wait_thread(struct task_struct *tsk, int *notask_error,
+                         enum pid_type type, struct pid *pid, int options,
+                         struct siginfo __user *infop, int __user *stat_addr,
+                         struct rusage __user *ru)
+{
+       struct task_struct *p;
+
+       list_for_each_entry(p, &tsk->children, sibling) {
+               /*
+                * Do not consider detached threads.
+                */
+               if (!task_detached(p)) {
+                       int ret = wait_consider_task(tsk, 0, p, notask_error,
+                                                    type, pid, options,
+                                                    infop, stat_addr, ru);
+                       if (ret)
+                               return ret;
+               }
+       }
+
+       return 0;
+}
+
+static int ptrace_do_wait(struct task_struct *tsk, int *notask_error,
+                         enum pid_type type, struct pid *pid, int options,
+                         struct siginfo __user *infop, int __user *stat_addr,
+                         struct rusage __user *ru)
+{
+       struct task_struct *p;
+
+       /*
+        * Traditionally we see ptrace'd stopped tasks regardless of options.
+        */
+       options |= WUNTRACED;
+
+       list_for_each_entry(p, &tsk->ptraced, ptrace_entry) {
+               int ret = wait_consider_task(tsk, 1, p, notask_error,
+                                            type, pid, options,
+                                            infop, stat_addr, ru);
+               if (ret)
+                       return ret;
+       }
+
+       return 0;
+}
+
 static long do_wait(enum pid_type type, struct pid *pid, int options,
                    struct siginfo __user *infop, int __user *stat_addr,
                    struct rusage __user *ru)
 {
        DECLARE_WAITQUEUE(wait, current);
        struct task_struct *tsk;
-       int flag, retval;
+       int retval;
 
        add_wait_queue(&current->signal->wait_chldexit,&wait);
 repeat:
-       /* If there is nothing that can match our critier just get out */
+       /*
+        * If there is nothing that can match our critiera just get out.
+        * We will clear @retval to zero if we see any child that might later
+        * match our criteria, even if we are not able to reap it yet.
+        */
        retval = -ECHILD;
        if ((type < PIDTYPE_MAX) && (!pid || hlist_empty(&pid->tasks[type])))
                goto end;
 
-       /*
-        * We will set this flag if we see any child that might later
-        * match our criteria, even if we are not able to reap it yet.
-        */
-       flag = retval = 0;
        current->state = TASK_INTERRUPTIBLE;
        read_lock(&tasklist_lock);
        tsk = current;
        do {
-               struct task_struct *p;
-
-               list_for_each_entry(p, &tsk->children, sibling) {
-                       int ret = eligible_child(type, pid, options, p);
-                       if (!ret)
-                               continue;
-
-                       if (unlikely(ret < 0)) {
-                               retval = ret;
-                       } else if (task_is_stopped_or_traced(p)) {
-                               /*
-                                * It's stopped now, so it might later
-                                * continue, exit, or stop again.
-                                */
-                               flag = 1;
-                               if (!(p->ptrace & PT_PTRACED) &&
-                                   !(options & WUNTRACED))
-                                       continue;
-
-                               retval = wait_task_stopped(p,
-                                               (options & WNOWAIT), infop,
-                                               stat_addr, ru);
-                       } else if (p->exit_state == EXIT_ZOMBIE &&
-                                       !delay_group_leader(p)) {
-                               /*
-                                * We don't reap group leaders with subthreads.
-                                */
-                               if (!likely(options & WEXITED))
-                                       continue;
-                               retval = wait_task_zombie(p,
-                                               (options & WNOWAIT), infop,
-                                               stat_addr, ru);
-                       } else if (p->exit_state != EXIT_DEAD) {
-                               /*
-                                * It's running now, so it might later
-                                * exit, stop, or stop and then continue.
-                                */
-                               flag = 1;
-                               if (!unlikely(options & WCONTINUED))
-                                       continue;
-                               retval = wait_task_continued(p,
-                                               (options & WNOWAIT), infop,
-                                               stat_addr, ru);
-                       }
-                       if (retval != 0) /* tasklist_lock released */
-                               goto end;
-               }
-               if (!flag) {
-                       list_for_each_entry(p, &tsk->ptrace_children,
-                                                               ptrace_list) {
-                               flag = eligible_child(type, pid, options, p);
-                               if (!flag)
-                                       continue;
-                               if (likely(flag > 0))
-                                       break;
-                               retval = flag;
-                               goto end;
-                       }
+               int tsk_result = do_wait_thread(tsk, &retval,
+                                               type, pid, options,
+                                               infop, stat_addr, ru);
+               if (!tsk_result)
+                       tsk_result = ptrace_do_wait(tsk, &retval,
+                                                   type, pid, options,
+                                                   infop, stat_addr, ru);
+               if (tsk_result) {
+                       /*
+                        * tasklist_lock is unlocked and we have a final result.
+                        */
+                       retval = tsk_result;
+                       goto end;
                }
+
                if (options & __WNOTHREAD)
                        break;
                tsk = next_thread(tsk);
@@ -1609,16 +1714,14 @@ repeat:
        } while (tsk != current);
        read_unlock(&tasklist_lock);
 
-       if (flag) {
-               if (options & WNOHANG)
-                       goto end;
+       if (!retval && !(options & WNOHANG)) {
                retval = -ERESTARTSYS;
-               if (signal_pending(current))
-                       goto end;
-               schedule();
-               goto repeat;
+               if (!signal_pending(current)) {
+                       schedule();
+                       goto repeat;
+               }
        }
-       retval = -ECHILD;
+
 end:
        current->state = TASK_RUNNING;
        remove_wait_queue(&current->signal->wait_chldexit,&wait);