exec: rework the group exit and fix the race with kill

As Roland pointed out, we have the very old problem with exec.  de_thread()
sets SIGNAL_GROUP_EXIT, kills other threads, changes ->group_leader and then
clears signal->flags.  All signals (even fatal ones) sent in this window
(which is not too small) will be lost.

With this patch exec doesn't abuse SIGNAL_GROUP_EXIT.  signal_group_exit(),
the new helper, should be used to detect exit_group() or exec() in progress.
It can have more users, but this patch does only strictly necessary changes.

Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Cc: Davide Libenzi <davidel@xmailserver.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Robin Holt <holt@sgi.com>
Cc: Roland McGrath <roland@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 7 insertions, 0 deletions

diff --git a/include/linux/sched.h b/include/linux/sched.h
index 680bb03a4b90..483ea4e1accf 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -555,6 +555,13 @@ struct signal_struct {
 #define SIGNAL_STOP_CONTINUED   0x00000004 /* SIGCONT since WCONTINUED reap */
 #define SIGNAL_GROUP_EXIT       0x00000008 /* group exit in progress */
 
+/* If true, all threads except ->group_exit_task have pending SIGKILL */
+static inline int signal_group_exit(const struct signal_struct *sig)
+{
+        return  (sig->flags & SIGNAL_GROUP_EXIT) ||
+                (sig->group_exit_task != NULL);
+}
+
 /*
  * Some day this will be a full-fledged user tracking system..
  */