coredump: zap_threads: comments && use while_each_thread()
authorOleg Nesterov <oleg@tv-sign.ru>
Fri, 25 Jul 2008 08:47:31 +0000 (01:47 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 25 Jul 2008 17:53:38 +0000 (10:53 -0700)
No changes in fs/exec.o

The for_each_process() loop in zap_threads() is very subtle, it is not
clear why we don't race with fork/exit/exec.  Add the fat comment.

Also, change the code to use while_each_thread().

Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Acked-by: Roland McGrath <roland@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/exec.c

index e41aef0fb351df55f2f4a5c6d6dbfbadc740fdf9..af249af4ccab1f487b4f600ee0a82b2ef2681e97 100644 (file)
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -1517,7 +1517,7 @@ static void zap_process(struct task_struct *start)
                        sigaddset(&t->pending.signal, SIGKILL);
                        signal_wake_up(t, 1);
                }
-       } while ((t = next_thread(t)) != start);
+       } while_each_thread(start, t);
 }
 
 static inline int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
@@ -1539,7 +1539,36 @@ static inline int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
 
        if (atomic_read(&mm->mm_users) == mm->core_waiters + 1)
                goto done;
-
+       /*
+        * We should find and kill all tasks which use this mm, and we should
+        * count them correctly into mm->core_waiters. We don't take tasklist
+        * lock, but this is safe wrt:
+        *
+        * fork:
+        *      None of sub-threads can fork after zap_process(leader). All
+        *      processes which were created before this point should be
+        *      visible to zap_threads() because copy_process() adds the new
+        *      process to the tail of init_task.tasks list, and lock/unlock
+        *      of ->siglock provides a memory barrier.
+        *
+        * do_exit:
+        *      The caller holds mm->mmap_sem. This means that the task which
+        *      uses this mm can't pass exit_mm(), so it can't exit or clear
+        *      its ->mm.
+        *
+        * de_thread:
+        *      It does list_replace_rcu(&leader->tasks, &current->tasks),
+        *      we must see either old or new leader, this does not matter.
+        *      However, it can change p->sighand, so lock_task_sighand(p)
+        *      must be used. Since p->mm != NULL and we hold ->mmap_sem
+        *      it can't fail.
+        *
+        *      Note also that "g" can be the old leader with ->mm == NULL
+        *      and already unhashed and thus removed from ->thread_group.
+        *      This is OK, __unhash_process()->list_del_rcu() does not
+        *      clear the ->next pointer, we will find the new leader via
+        *      next_thread().
+        */
        rcu_read_lock();
        for_each_process(g) {
                if (g == tsk->group_leader)
@@ -1549,17 +1578,13 @@ static inline int zap_threads(struct task_struct *tsk, struct mm_struct *mm,
                do {
                        if (p->mm) {
                                if (p->mm == mm) {
-                                       /*
-                                        * p->sighand can't disappear, but
-                                        * may be changed by de_thread()
-                                        */
                                        lock_task_sighand(p, &flags);
                                        zap_process(p);
                                        unlock_task_sighand(p, &flags);
                                }
                                break;
                        }
-               } while ((p = next_thread(p)) != g);
+               } while_each_thread(g, p);
        }
        rcu_read_unlock();
 done: