exit: wait: don't use zombie->real_parent
authorOleg Nesterov <oleg@redhat.com>
Wed, 10 Dec 2014 23:54:48 +0000 (15:54 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 11 Dec 2014 01:41:17 +0000 (17:41 -0800)
1. wait_task_zombie() uses p->real_parent to get psig/siglock. This is
   correct but needs tasklist_lock, ->real_parent can exit.

   We can use "current" instead. This is our natural child, its parent
   must be our sub-thread.

2. Read psig/sig outside of ->siglock, ->signal is no longer protected
   by this lock.

3. Fix the outdated comments about tasklist_lock. We can not race with
   __exit_signal(), the whole thread group is dead, nobody but us can
   call it.

   Also clarify the usage of ->stats_lock and ->siglock.

Note: thread_group_cputime_adjusted() is sub-optimal in this case, we
probably want to export cputime_adjust() to avoid thread_group_cputime().
The comment says "all threads" but there are no other threads.

Signed-off-by: Oleg Nesterov <oleg@redhat.com>
Cc: Aaron Tomlin <atomlin@redhat.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Sterling Alexander <stalexan@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
kernel/exit.c

index 20875d6398ae564d1c62480e5a868e13525022a6..457673d659346ed230c1c99f775afed2fa1ed539 100644 (file)
@@ -1010,8 +1010,8 @@ static int wait_task_zombie(struct wait_opts *wo, struct task_struct *p)
         * Check thread_group_leader() to exclude the traced sub-threads.
         */
        if (state == EXIT_DEAD && thread_group_leader(p)) {
-               struct signal_struct *psig;
-               struct signal_struct *sig;
+               struct signal_struct *sig = p->signal;
+               struct signal_struct *psig = current->signal;
                unsigned long maxrss;
                cputime_t tgutime, tgstime;
 
@@ -1023,21 +1023,20 @@ static int wait_task_zombie(struct wait_opts *wo, struct task_struct *p)
                 * accumulate in the parent's signal_struct c* fields.
                 *
                 * We don't bother to take a lock here to protect these
-                * p->signal fields, because they are only touched by
-                * __exit_signal, which runs with tasklist_lock
-                * write-locked anyway, and so is excluded here.  We do
-                * need to protect the access to parent->signal fields,
-                * as other threads in the parent group can be right
-                * here reaping other children at the same time.
+                * p->signal fields because the whole thread group is dead
+                * and nobody can change them.
+                *
+                * psig->stats_lock also protects us from our sub-theads
+                * which can reap other children at the same time. Until
+                * we change k_getrusage()-like users to rely on this lock
+                * we have to take ->siglock as well.
                 *
                 * We use thread_group_cputime_adjusted() to get times for
                 * the thread group, which consolidates times for all threads
                 * in the group including the group leader.
                 */
                thread_group_cputime_adjusted(p, &tgutime, &tgstime);
-               spin_lock_irq(&p->real_parent->sighand->siglock);
-               psig = p->real_parent->signal;
-               sig = p->signal;
+               spin_lock_irq(&current->sighand->siglock);
                write_seqlock(&psig->stats_lock);
                psig->cutime += tgutime + sig->cutime;
                psig->cstime += tgstime + sig->cstime;
@@ -1062,7 +1061,7 @@ static int wait_task_zombie(struct wait_opts *wo, struct task_struct *p)
                task_io_accounting_add(&psig->ioac, &p->ioac);
                task_io_accounting_add(&psig->ioac, &sig->ioac);
                write_sequnlock(&psig->stats_lock);
-               spin_unlock_irq(&p->real_parent->sighand->siglock);
+               spin_unlock_irq(&current->sighand->siglock);
        }
 
        /*