arm: tegra: cardhu/enterprise: Remove pinmux conflicts
[linux-2.6.git] / kernel / exit.c
index 019a284..4b4042f 100644 (file)
@@ -50,6 +50,8 @@
 #include <linux/perf_event.h>
 #include <trace/events/sched.h>
 #include <linux/hw_breakpoint.h>
+#include <linux/oom.h>
+#include <linux/writeback.h>
 
 #include <asm/uaccess.h>
 #include <asm/unistd.h>
 
 static void exit_mm(struct task_struct * tsk);
 
-static void __unhash_process(struct task_struct *p)
+static void __unhash_process(struct task_struct *p, bool group_dead)
 {
        nr_threads--;
        detach_pid(p, PIDTYPE_PID);
-       if (thread_group_leader(p)) {
+       if (group_dead) {
                detach_pid(p, PIDTYPE_PGID);
                detach_pid(p, PIDTYPE_SID);
 
                list_del_rcu(&p->tasks);
                list_del_init(&p->sibling);
-               __get_cpu_var(process_counts)--;
+               __this_cpu_dec(process_counts);
        }
        list_del_rcu(&p->thread_group);
 }
@@ -79,25 +81,33 @@ static void __unhash_process(struct task_struct *p)
 static void __exit_signal(struct task_struct *tsk)
 {
        struct signal_struct *sig = tsk->signal;
+       bool group_dead = thread_group_leader(tsk);
        struct sighand_struct *sighand;
-
-       BUG_ON(!sig);
-       BUG_ON(!atomic_read(&sig->count));
+       struct tty_struct *uninitialized_var(tty);
 
        sighand = rcu_dereference_check(tsk->sighand,
-                                       rcu_read_lock_held() ||
                                        lockdep_tasklist_lock_is_held());
        spin_lock(&sighand->siglock);
 
        posix_cpu_timers_exit(tsk);
-       if (atomic_dec_and_test(&sig->count))
+       if (group_dead) {
                posix_cpu_timers_exit_group(tsk);
-       else {
+               tty = sig->tty;
+               sig->tty = NULL;
+       } else {
+               /*
+                * This can only happen if the caller is de_thread().
+                * FIXME: this is the temporary hack, we should teach
+                * posix-cpu-timers to handle this case correctly.
+                */
+               if (unlikely(has_group_leader_pid(tsk)))
+                       posix_cpu_timers_exit_group(tsk);
+
                /*
                 * If there is any task waiting for the group exit
                 * then notify it:
                 */
-               if (sig->group_exit_task && atomic_read(&sig->count) == sig->notify_count)
+               if (sig->notify_count > 0 && !--sig->notify_count)
                        wake_up_process(sig->group_exit_task);
 
                if (tsk == sig->curr_target)
@@ -112,9 +122,9 @@ static void __exit_signal(struct task_struct *tsk)
                 * We won't ever get here for the group leader, since it
                 * will have been the last reference on the signal_struct.
                 */
-               sig->utime = cputime_add(sig->utime, tsk->utime);
-               sig->stime = cputime_add(sig->stime, tsk->stime);
-               sig->gtime = cputime_add(sig->gtime, tsk->gtime);
+               sig->utime += tsk->utime;
+               sig->stime += tsk->stime;
+               sig->gtime += tsk->gtime;
                sig->min_flt += tsk->min_flt;
                sig->maj_flt += tsk->maj_flt;
                sig->nvcsw += tsk->nvcsw;
@@ -123,32 +133,24 @@ static void __exit_signal(struct task_struct *tsk)
                sig->oublock += task_io_get_oublock(tsk);
                task_io_accounting_add(&sig->ioac, &tsk->ioac);
                sig->sum_sched_runtime += tsk->se.sum_exec_runtime;
-               sig = NULL; /* Marker for below. */
        }
 
-       __unhash_process(tsk);
+       sig->nr_threads--;
+       __unhash_process(tsk, group_dead);
 
        /*
         * Do this under ->siglock, we can race with another thread
         * doing sigqueue_free() if we have SIGQUEUE_PREALLOC signals.
         */
        flush_sigqueue(&tsk->pending);
-
-       tsk->signal = NULL;
        tsk->sighand = NULL;
        spin_unlock(&sighand->siglock);
 
        __cleanup_sighand(sighand);
        clear_tsk_thread_flag(tsk,TIF_SIGPENDING);
-       if (sig) {
+       if (group_dead) {
                flush_sigqueue(&sig->shared_pending);
-               taskstats_tgid_free(sig);
-               /*
-                * Make sure ->signal can't go away under rq->lock,
-                * see account_group_exec_runtime().
-                */
-               task_rq_unlock_wait(tsk);
-               __cleanup_signal(sig);
+               tty_kref_put(tty);
        }
 }
 
@@ -156,9 +158,7 @@ static void delayed_put_task_struct(struct rcu_head *rhp)
 {
        struct task_struct *tsk = container_of(rhp, struct task_struct, rcu);
 
-#ifdef CONFIG_PERF_EVENTS
-       WARN_ON_ONCE(tsk->perf_event_ctxp);
-#endif
+       perf_event_delayed_put(tsk);
        trace_sched_process_free(tsk);
        put_task_struct(tsk);
 }
@@ -169,7 +169,6 @@ void release_task(struct task_struct * p)
        struct task_struct *leader;
        int zap_leader;
 repeat:
-       tracehook_prepare_release_task(p);
        /* don't need to get the RCU readlock here - the process is dead and
         * can't be modifying its own credentials. But shut RCU-lockdep up */
        rcu_read_lock();
@@ -179,7 +178,7 @@ repeat:
        proc_flush_task(p);
 
        write_lock_irq(&tasklist_lock);
-       tracehook_finish_release_task(p);
+       ptrace_release_task(p);
        __exit_signal(p);
 
        /*
@@ -190,22 +189,12 @@ repeat:
        zap_leader = 0;
        leader = p->group_leader;
        if (leader != p && thread_group_empty(leader) && leader->exit_state == EXIT_ZOMBIE) {
-               BUG_ON(task_detached(leader));
-               do_notify_parent(leader, leader->exit_signal);
                /*
                 * If we were the last child thread and the leader has
                 * exited already, and the leader's parent ignores SIGCHLD,
                 * then we are the one who should release the leader.
-                *
-                * do_notify_parent() will have marked it self-reaping in
-                * that case.
-                */
-               zap_leader = task_detached(leader);
-
-               /*
-                * This maintains the invariant that release_task()
-                * only runs on a task in EXIT_DEAD, just for sanity.
                 */
+               zap_leader = do_notify_parent(leader, leader->exit_signal);
                if (zap_leader)
                        leader->exit_state = EXIT_DEAD;
        }
@@ -277,18 +266,16 @@ int is_current_pgrp_orphaned(void)
        return retval;
 }
 
-static int has_stopped_jobs(struct pid *pgrp)
+static bool has_stopped_jobs(struct pid *pgrp)
 {
-       int retval = 0;
        struct task_struct *p;
 
        do_each_pid_task(pgrp, PIDTYPE_PGID, p) {
-               if (!task_is_stopped(p))
-                       continue;
-               retval = 1;
-               break;
+               if (p->signal->flags & SIGNAL_STOP_STOPPED)
+                       return true;
        } while_each_pid_task(pgrp, PIDTYPE_PGID, p);
-       return retval;
+
+       return false;
 }
 
 /*
@@ -561,29 +548,28 @@ void exit_files(struct task_struct *tsk)
 
 #ifdef CONFIG_MM_OWNER
 /*
- * Task p is exiting and it owned mm, lets find a new owner for it
+ * A task is exiting.   If it owned this mm, find a new owner for the mm.
  */
-static inline int
-mm_need_new_owner(struct mm_struct *mm, struct task_struct *p)
-{
-       /*
-        * If there are other users of the mm and the owner (us) is exiting
-        * we need to find a new owner to take on the responsibility.
-        */
-       if (atomic_read(&mm->mm_users) <= 1)
-               return 0;
-       if (mm->owner != p)
-               return 0;
-       return 1;
-}
-
 void mm_update_next_owner(struct mm_struct *mm)
 {
        struct task_struct *c, *g, *p = current;
 
 retry:
-       if (!mm_need_new_owner(mm, p))
+       /*
+        * If the exiting or execing task is not the owner, it's
+        * someone else's problem.
+        */
+       if (mm->owner != p)
+               return;
+       /*
+        * The current owner is exiting/execing and there are no other
+        * candidates.  Do not leave the mm pointing to a possibly
+        * freed task structure.
+        */
+       if (atomic_read(&mm->mm_users) <= 1) {
+               mm->owner = NULL;
                return;
+       }
 
        read_lock(&tasklist_lock);
        /*
@@ -694,8 +680,6 @@ static void exit_mm(struct task_struct * tsk)
        tsk->mm = NULL;
        up_read(&mm->mmap_sem);
        enter_lazy_tlb(mm, current);
-       /* We don't want this task to be frozen prematurely */
-       clear_freeze_flag(tsk);
        task_unlock(tsk);
        mm_update_next_owner(mm);
        mmput(mm);
@@ -709,6 +693,8 @@ static void exit_mm(struct task_struct * tsk)
  * space.
  */
 static struct task_struct *find_new_reaper(struct task_struct *father)
+       __releases(&tasklist_lock)
+       __acquires(&tasklist_lock)
 {
        struct pid_namespace *pid_ns = task_active_pid_ns(father);
        struct task_struct *thread;
@@ -748,7 +734,7 @@ static void reparent_leader(struct task_struct *father, struct task_struct *p,
 {
        list_move_tail(&p->sibling, &p->real_parent->children);
 
-       if (task_detached(p))
+       if (p->exit_state == EXIT_DEAD)
                return;
        /*
         * If this is a threaded reparent there is no need to
@@ -761,10 +747,9 @@ static void reparent_leader(struct task_struct *father, struct task_struct *p,
        p->exit_signal = SIGCHLD;
 
        /* If it has exited notify the new parent about this child's death. */
-       if (!task_ptrace(p) &&
+       if (!p->ptrace &&
            p->exit_state == EXIT_ZOMBIE && thread_group_empty(p)) {
-               do_notify_parent(p, p->exit_signal);
-               if (task_detached(p)) {
+               if (do_notify_parent(p, p->exit_signal)) {
                        p->exit_state = EXIT_DEAD;
                        list_move_tail(&p->sibling, dead);
                }
@@ -778,9 +763,12 @@ static void forget_original_parent(struct task_struct *father)
        struct task_struct *p, *n, *reaper;
        LIST_HEAD(dead_children);
 
-       exit_ptrace(father);
-
        write_lock_irq(&tasklist_lock);
+       /*
+        * Note that exit_ptrace() and find_new_reaper() might
+        * drop tasklist_lock and reacquire it.
+        */
+       exit_ptrace(father);
        reaper = find_new_reaper(father);
 
        list_for_each_entry_safe(p, n, &father->children, sibling) {
@@ -788,7 +776,7 @@ static void forget_original_parent(struct task_struct *father)
                do {
                        t->real_parent = reaper;
                        if (t->parent == father) {
-                               BUG_ON(task_ptrace(t));
+                               BUG_ON(t->ptrace);
                                t->parent = t->real_parent;
                        }
                        if (t->pdeath_signal)
@@ -813,8 +801,7 @@ static void forget_original_parent(struct task_struct *father)
  */
 static void exit_notify(struct task_struct *tsk, int group_dead)
 {
-       int signal;
-       void *cookie;
+       bool autoreap;
 
        /*
         * This does two things:
@@ -834,7 +821,7 @@ static void exit_notify(struct task_struct *tsk, int group_dead)
        /* Let father know we died
         *
         * Thread signals are configurable, but you aren't going to use
-        * that to send signals to arbitary processes.
+        * that to send signals to arbitrary processes.
         * That stops right now.
         *
         * If the parent exec id doesn't match the exec id we saved
@@ -845,29 +832,33 @@ static void exit_notify(struct task_struct *tsk, int group_dead)
         * we have changed execution domain as these two values started
         * the same after a fork.
         */
-       if (tsk->exit_signal != SIGCHLD && !task_detached(tsk) &&
+       if (thread_group_leader(tsk) && tsk->exit_signal != SIGCHLD &&
            (tsk->parent_exec_id != tsk->real_parent->self_exec_id ||
             tsk->self_exec_id != tsk->parent_exec_id))
                tsk->exit_signal = SIGCHLD;
 
-       signal = tracehook_notify_death(tsk, &cookie, group_dead);
-       if (signal >= 0)
-               signal = do_notify_parent(tsk, signal);
+       if (unlikely(tsk->ptrace)) {
+               int sig = thread_group_leader(tsk) &&
+                               thread_group_empty(tsk) &&
+                               !ptrace_reparented(tsk) ?
+                       tsk->exit_signal : SIGCHLD;
+               autoreap = do_notify_parent(tsk, sig);
+       } else if (thread_group_leader(tsk)) {
+               autoreap = thread_group_empty(tsk) &&
+                       do_notify_parent(tsk, tsk->exit_signal);
+       } else {
+               autoreap = true;
+       }
 
-       tsk->exit_state = signal == DEATH_REAP ? EXIT_DEAD : EXIT_ZOMBIE;
+       tsk->exit_state = autoreap ? EXIT_DEAD : EXIT_ZOMBIE;
 
-       /* mt-exec, de_thread() is waiting for us */
-       if (thread_group_leader(tsk) &&
-           tsk->signal->group_exit_task &&
-           tsk->signal->notify_count < 0)
+       /* mt-exec, de_thread() is waiting for group leader */
+       if (unlikely(tsk->signal->notify_count < 0))
                wake_up_process(tsk->signal->group_exit_task);
-
        write_unlock_irq(&tasklist_lock);
 
-       tracehook_report_death(tsk, signal, cookie, group_dead);
-
        /* If the process is dead, release it - nobody will wait for it */
-       if (signal == DEATH_REAP)
+       if (autoreap)
                release_task(tsk);
 }
 
@@ -896,21 +887,30 @@ static void check_stack_usage(void)
 static inline void check_stack_usage(void) {}
 #endif
 
-NORET_TYPE void do_exit(long code)
+void do_exit(long code)
 {
        struct task_struct *tsk = current;
        int group_dead;
 
        profile_task_exit(tsk);
 
-       WARN_ON(atomic_read(&tsk->fs_excl));
+       WARN_ON(blk_needs_flush_plug(tsk));
 
        if (unlikely(in_interrupt()))
                panic("Aiee, killing interrupt handler!");
        if (unlikely(!tsk->pid))
                panic("Attempted to kill the idle task!");
 
-       tracehook_report_exit(&code);
+       /*
+        * If do_exit is called because this processes oopsed, it's possible
+        * that get_fs() was left as KERNEL_DS, so reset it to USER_DS before
+        * continuing. Amongst other possible reasons, this is to prevent
+        * mm_release()->clear_child_tid() from writing to a user-controlled
+        * kernel address.
+        */
+       set_fs(USER_DS);
+
+       ptrace_event(PTRACE_EVENT_EXIT, code);
 
        validate_creds_for_do_exit(tsk);
 
@@ -964,8 +964,7 @@ NORET_TYPE void do_exit(long code)
        acct_collect(code, group_dead);
        if (group_dead)
                tty_audit_exit();
-       if (unlikely(tsk->audit_context))
-               audit_free(tsk);
+       audit_free(tsk);
 
        tsk->exit_code = code;
        taskstats_exit(tsk, group_dead);
@@ -977,10 +976,20 @@ NORET_TYPE void do_exit(long code)
        trace_sched_process_exit(tsk);
 
        exit_sem(tsk);
+       exit_shm(tsk);
        exit_files(tsk);
        exit_fs(tsk);
        check_stack_usage();
        exit_thread();
+
+       /*
+        * Flush inherited counters to the parent - before the parent
+        * gets woken up by child-exit notifications.
+        *
+        * because of cgroup mode, must be called before cgroup_exit()
+        */
+       perf_event_exit_task(tsk);
+
        cgroup_exit(tsk, 1);
 
        if (group_dead)
@@ -993,12 +1002,7 @@ NORET_TYPE void do_exit(long code)
        /*
         * FIXME: do that only when needed, using sched_exit tracepoint
         */
-       flush_ptrace_hw_breakpoint(tsk);
-       /*
-        * Flush inherited counters to the parent - before the parent
-        * gets woken up by child-exit notifications.
-        */
-       perf_event_exit_task(tsk);
+       ptrace_put_breakpoints(tsk);
 
        exit_notify(tsk, group_dead);
 #ifdef CONFIG_NUMA
@@ -1031,9 +1035,28 @@ NORET_TYPE void do_exit(long code)
        validate_creds_for_do_exit(tsk);
 
        preempt_disable();
+       if (tsk->nr_dirtied)
+               __this_cpu_add(dirty_throttle_leaks, tsk->nr_dirtied);
        exit_rcu();
+
+       /*
+        * The setting of TASK_RUNNING by try_to_wake_up() may be delayed
+        * when the following two conditions become true.
+        *   - There is race condition of mmap_sem (It is acquired by
+        *     exit_mm()), and
+        *   - SMI occurs before setting TASK_RUNINNG.
+        *     (or hypervisor of virtual machine switches to other guest)
+        *  As a result, we may become TASK_RUNNING after becoming TASK_DEAD
+        *
+        * To avoid it, we have to wait for releasing tsk->pi_lock which
+        * is held by try_to_wake_up()
+        */
+       smp_mb();
+       raw_spin_unlock_wait(&tsk->pi_lock);
+
        /* causes final put_task_struct in finish_task_switch(). */
        tsk->state = TASK_DEAD;
+       tsk->flags |= PF_NOFREEZE;      /* tell freezer to ignore us */
        schedule();
        BUG();
        /* Avoid "noreturn function does return".  */
@@ -1043,7 +1066,7 @@ NORET_TYPE void do_exit(long code)
 
 EXPORT_SYMBOL_GPL(do_exit);
 
-NORET_TYPE void complete_and_exit(struct completion *comp, long code)
+void complete_and_exit(struct completion *comp, long code)
 {
        if (comp)
                complete(comp);
@@ -1062,7 +1085,7 @@ SYSCALL_DEFINE1(exit, int, error_code)
  * Take down every thread in the group.  This is called by fatal signals
  * as well as by sys_exit_group (below).
  */
-NORET_TYPE void
+void
 do_group_exit(int exit_code)
 {
        struct signal_struct *sig = current->signal;
@@ -1218,9 +1241,9 @@ static int wait_task_zombie(struct wait_opts *wo, struct task_struct *p)
        traced = ptrace_reparented(p);
        /*
         * It can be ptraced but not reparented, check
-        * !task_detached() to filter out sub-threads.
+        * thread_group_leader() to filter out sub-threads.
         */
-       if (likely(!traced) && likely(!task_detached(p))) {
+       if (likely(!traced) && thread_group_leader(p)) {
                struct signal_struct *psig;
                struct signal_struct *sig;
                unsigned long maxrss;
@@ -1249,19 +1272,9 @@ static int wait_task_zombie(struct wait_opts *wo, struct task_struct *p)
                spin_lock_irq(&p->real_parent->sighand->siglock);
                psig = p->real_parent->signal;
                sig = p->signal;
-               psig->cutime =
-                       cputime_add(psig->cutime,
-                       cputime_add(tgutime,
-                                   sig->cutime));
-               psig->cstime =
-                       cputime_add(psig->cstime,
-                       cputime_add(tgstime,
-                                   sig->cstime));
-               psig->cgtime =
-                       cputime_add(psig->cgtime,
-                       cputime_add(p->gtime,
-                       cputime_add(sig->gtime,
-                                   sig->cgtime)));
+               psig->cutime += tgutime + sig->cutime;
+               psig->cstime += tgstime + sig->cstime;
+               psig->cgtime += p->gtime + sig->gtime + sig->cgtime;
                psig->cmin_flt +=
                        p->min_flt + sig->min_flt + sig->cmin_flt;
                psig->cmaj_flt +=
@@ -1328,16 +1341,13 @@ static int wait_task_zombie(struct wait_opts *wo, struct task_struct *p)
                /* We dropped tasklist, ptracer could die and untrace */
                ptrace_unlink(p);
                /*
-                * If this is not a detached task, notify the parent.
-                * If it's still not detached after that, don't release
-                * it now.
+                * If this is not a sub-thread, notify the parent.
+                * If parent wants a zombie, don't release it now.
                 */
-               if (!task_detached(p)) {
-                       do_notify_parent(p, p->exit_signal);
-                       if (!task_detached(p)) {
-                               p->exit_state = EXIT_ZOMBIE;
-                               p = NULL;
-                       }
+               if (thread_group_leader(p) &&
+                   !do_notify_parent(p, p->exit_signal)) {
+                       p->exit_state = EXIT_ZOMBIE;
+                       p = NULL;
                }
                write_unlock_irq(&tasklist_lock);
        }
@@ -1350,7 +1360,8 @@ static int wait_task_zombie(struct wait_opts *wo, struct task_struct *p)
 static int *task_stopped_code(struct task_struct *p, bool ptrace)
 {
        if (ptrace) {
-               if (task_is_stopped_or_traced(p))
+               if (task_is_stopped_or_traced(p) &&
+                   !(p->jobctl & JOBCTL_LISTENING))
                        return &p->exit_code;
        } else {
                if (p->signal->flags & SIGNAL_STOP_STOPPED)
@@ -1359,11 +1370,23 @@ static int *task_stopped_code(struct task_struct *p, bool ptrace)
        return NULL;
 }
 
-/*
- * Handle sys_wait4 work for one task in state TASK_STOPPED.  We hold
- * read_lock(&tasklist_lock) on entry.  If we return zero, we still hold
- * the lock and this task is uninteresting.  If we return nonzero, we have
- * released the lock and the system call should return.
+/**
+ * wait_task_stopped - Wait for %TASK_STOPPED or %TASK_TRACED
+ * @wo: wait options
+ * @ptrace: is the wait for ptrace
+ * @p: task to wait for
+ *
+ * Handle sys_wait4() work for %p in state %TASK_STOPPED or %TASK_TRACED.
+ *
+ * CONTEXT:
+ * read_lock(&tasklist_lock), which is released if return value is
+ * non-zero.  Also, grabs and releases @p->sighand->siglock.
+ *
+ * RETURNS:
+ * 0 if wait condition didn't exist and search for other wait conditions
+ * should continue.  Non-zero return, -errno on failure and @p's pid on
+ * success, implies that tasklist_lock is released and wait condition
+ * search should terminate.
  */
 static int wait_task_stopped(struct wait_opts *wo,
                                int ptrace, struct task_struct *p)
@@ -1379,6 +1402,9 @@ static int wait_task_stopped(struct wait_opts *wo,
        if (!ptrace && !(wo->wo_flags & WUNTRACED))
                return 0;
 
+       if (!task_stopped_code(p, ptrace))
+               return 0;
+
        exit_code = 0;
        spin_lock_irq(&p->sighand->siglock);
 
@@ -1393,8 +1419,7 @@ static int wait_task_stopped(struct wait_opts *wo,
        if (!unlikely(wo->wo_flags & WNOWAIT))
                *p_code = 0;
 
-       /* don't need the RCU readlock here as we're holding a spinlock */
-       uid = __task_cred(p)->uid;
+       uid = task_uid(p);
 unlock_sig:
        spin_unlock_irq(&p->sighand->siglock);
        if (!exit_code)
@@ -1467,7 +1492,7 @@ static int wait_task_continued(struct wait_opts *wo, struct task_struct *p)
        }
        if (!unlikely(wo->wo_flags & WNOWAIT))
                p->signal->flags &= ~SIGNAL_STOP_CONTINUED;
-       uid = __task_cred(p)->uid;
+       uid = task_uid(p);
        spin_unlock_irq(&p->sighand->siglock);
 
        pid = task_pid_vnr(p);
@@ -1521,33 +1546,90 @@ static int wait_consider_task(struct wait_opts *wo, int ptrace,
                return 0;
        }
 
-       if (likely(!ptrace) && unlikely(task_ptrace(p))) {
+       /* dead body doesn't have much to contribute */
+       if (unlikely(p->exit_state == EXIT_DEAD)) {
                /*
-                * This child is hidden by ptrace.
-                * We aren't allowed to see it now, but eventually we will.
+                * But do not ignore this task until the tracer does
+                * wait_task_zombie()->do_notify_parent().
                 */
-               wo->notask_error = 0;
+               if (likely(!ptrace) && unlikely(ptrace_reparented(p)))
+                       wo->notask_error = 0;
                return 0;
        }
 
-       if (p->exit_state == EXIT_DEAD)
-               return 0;
+       /* slay zombie? */
+       if (p->exit_state == EXIT_ZOMBIE) {
+               /*
+                * A zombie ptracee is only visible to its ptracer.
+                * Notification and reaping will be cascaded to the real
+                * parent when the ptracer detaches.
+                */
+               if (likely(!ptrace) && unlikely(p->ptrace)) {
+                       /* it will become visible, clear notask_error */
+                       wo->notask_error = 0;
+                       return 0;
+               }
+
+               /* we don't reap group leaders with subthreads */
+               if (!delay_group_leader(p))
+                       return wait_task_zombie(wo, p);
+
+               /*
+                * Allow access to stopped/continued state via zombie by
+                * falling through.  Clearing of notask_error is complex.
+                *
+                * When !@ptrace:
+                *
+                * If WEXITED is set, notask_error should naturally be
+                * cleared.  If not, subset of WSTOPPED|WCONTINUED is set,
+                * so, if there are live subthreads, there are events to
+                * wait for.  If all subthreads are dead, it's still safe
+                * to clear - this function will be called again in finite
+                * amount time once all the subthreads are released and
+                * will then return without clearing.
+                *
+                * When @ptrace:
+                *
+                * Stopped state is per-task and thus can't change once the
+                * target task dies.  Only continued and exited can happen.
+                * Clear notask_error if WCONTINUED | WEXITED.
+                */
+               if (likely(!ptrace) || (wo->wo_flags & (WCONTINUED | WEXITED)))
+                       wo->notask_error = 0;
+       } else {
+               /*
+                * If @p is ptraced by a task in its real parent's group,
+                * hide group stop/continued state when looking at @p as
+                * the real parent; otherwise, a single stop can be
+                * reported twice as group and ptrace stops.
+                *
+                * If a ptracer wants to distinguish the two events for its
+                * own children, it should create a separate process which
+                * takes the role of real parent.
+                */
+               if (likely(!ptrace) && p->ptrace && !ptrace_reparented(p))
+                       return 0;
+
+               /*
+                * @p is alive and it's gonna stop, continue or exit, so
+                * there always is something to wait for.
+                */
+               wo->notask_error = 0;
+       }
 
        /*
-        * We don't reap group leaders with subthreads.
+        * Wait for stopped.  Depending on @ptrace, different stopped state
+        * is used and the two don't interact with each other.
         */
-       if (p->exit_state == EXIT_ZOMBIE && !delay_group_leader(p))
-               return wait_task_zombie(wo, p);
+       ret = wait_task_stopped(wo, ptrace, p);
+       if (ret)
+               return ret;
 
        /*
-        * It's stopped or running now, so it might
-        * later continue, exit, or stop again.
+        * Wait for continued.  There's only one continued state and the
+        * ptracer can consume it which can confuse the real parent.  Don't
+        * use WCONTINUED from ptracer.  You don't need or want it.
         */
-       wo->notask_error = 0;
-
-       if (task_stopped_code(p, ptrace))
-               return wait_task_stopped(wo, ptrace, p);
-
        return wait_task_continued(wo, p);
 }