signals: re-assign CLD_CONTINUED notification from the sender to reciever
authorOleg Nesterov <oleg@tv-sign.ru>
Wed, 30 Apr 2008 07:52:44 +0000 (00:52 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 30 Apr 2008 15:29:34 +0000 (08:29 -0700)
Based on discussion with Jiri and Roland.

In short: currently handle_stop_signal(SIGCONT, p) sends the notification to
p->parent, with this patch p itself notifies its parent when it becomes
running.

handle_stop_signal(SIGCONT) has to drop ->siglock temporary in order to notify
the parent with do_notify_parent_cldstop().  This leads to multiple problems:

- as Jiri Kosina pointed out, the stopped task can resume without
  actually seeing SIGCONT which may have a handler.

- we race with another sig_kernel_stop() signal which may come in
  that window.

- we race with sig_fatal() signals which may set SIGNAL_GROUP_EXIT
  in that window.

- we can't avoid taking tasklist_lock() while sending SIGCONT.

With this patch handle_stop_signal() just sets the new SIGNAL_CLD_CONTINUED
flag in p->signal->flags and returns.  The notification is sent by the first
task which returns from finish_stop() (there should be at least one) or any
other signalled thread from get_signal_to_deliver().

This is a user-visible change.  Say, currently kill(SIGCONT, stopped_child)
can't return without seeing SIGCHLD, with this patch SIGCHLD can be delayed
unpredictably.  Another difference is that if the child is ptraced by another
process, CLD_CONTINUED may be delivered to ->real_parent after ptrace_detach()
while currently it always goes to the tracer which doesn't actually need this
notification.  Hopefully not a problem.

The patch asks for the futher obvious cleanups, I'll send them separately.

Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Cc: Roland McGrath <roland@redhat.com>
Cc: Jiri Kosina <jkosina@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/sched.h
kernel/signal.c

index 1d02bab..ef56152 100644 (file)
@@ -554,6 +554,12 @@ struct signal_struct {
 #define SIGNAL_STOP_DEQUEUED   0x00000002 /* stop signal dequeued */
 #define SIGNAL_STOP_CONTINUED  0x00000004 /* SIGCONT since WCONTINUED reap */
 #define SIGNAL_GROUP_EXIT      0x00000008 /* group exit in progress */
+/*
+ * Pending notifications to parent.
+ */
+#define SIGNAL_CLD_STOPPED     0x00000010
+#define SIGNAL_CLD_CONTINUED   0x00000020
+#define SIGNAL_CLD_MASK                (SIGNAL_CLD_STOPPED|SIGNAL_CLD_CONTINUED)
 
 /* If true, all threads except ->group_exit_task have pending SIGKILL */
 static inline int signal_group_exit(const struct signal_struct *sig)
index 91d57f8..115c04f 100644 (file)
@@ -603,10 +603,8 @@ static void handle_stop_signal(int sig, struct task_struct *p)
                         * the SIGCHLD was pending on entry to this kill.
                         */
                        p->signal->group_stop_count = 0;
-                       p->signal->flags = SIGNAL_STOP_CONTINUED;
-                       spin_unlock(&p->sighand->siglock);
-                       do_notify_parent_cldstop(p, CLD_STOPPED);
-                       spin_lock(&p->sighand->siglock);
+                       p->signal->flags = SIGNAL_STOP_CONTINUED |
+                                               SIGNAL_CLD_STOPPED;
                }
                rm_from_queue(SIG_KERNEL_STOP_MASK, &p->signal->shared_pending);
                t = p;
@@ -643,25 +641,23 @@ static void handle_stop_signal(int sig, struct task_struct *p)
                         * We were in fact stopped, and are now continued.
                         * Notify the parent with CLD_CONTINUED.
                         */
-                       p->signal->flags = SIGNAL_STOP_CONTINUED;
+                       p->signal->flags = SIGNAL_STOP_CONTINUED |
+                                               SIGNAL_CLD_CONTINUED;
                        p->signal->group_exit_code = 0;
-                       spin_unlock(&p->sighand->siglock);
-                       do_notify_parent_cldstop(p, CLD_CONTINUED);
-                       spin_lock(&p->sighand->siglock);
                } else {
                        /*
                         * We are not stopped, but there could be a stop
                         * signal in the middle of being processed after
                         * being removed from the queue.  Clear that too.
                         */
-                       p->signal->flags = 0;
+                       p->signal->flags &= ~SIGNAL_STOP_DEQUEUED;
                }
        } else if (sig == SIGKILL) {
                /*
                 * Make sure that any pending stop signal already dequeued
                 * is undone by the wakeup for SIGKILL.
                 */
-               p->signal->flags = 0;
+               p->signal->flags &= ~SIGNAL_STOP_DEQUEUED;
        }
 }
 
@@ -1784,6 +1780,19 @@ relock:
        try_to_freeze();
 
        spin_lock_irq(&current->sighand->siglock);
+
+       if (unlikely(current->signal->flags & SIGNAL_CLD_MASK)) {
+               int why = (current->signal->flags & SIGNAL_STOP_CONTINUED)
+                               ? CLD_CONTINUED : CLD_STOPPED;
+               current->signal->flags &= ~SIGNAL_CLD_MASK;
+               spin_unlock_irq(&current->sighand->siglock);
+
+               read_lock(&tasklist_lock);
+               do_notify_parent_cldstop(current->group_leader, why);
+               read_unlock(&tasklist_lock);
+               goto relock;
+       }
+
        for (;;) {
                struct k_sigaction *ka;