From: Oleg Nesterov Date: Wed, 30 Apr 2008 07:52:44 +0000 (-0700) Subject: signals: re-assign CLD_CONTINUED notification from the sender to reciever X-Git-Tag: v2.6.26-rc1~302 X-Git-Url: http://git.openpandora.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=e442055193e4584218006e616c9bdce0c5e9ae5c;p=pandora-kernel.git signals: re-assign CLD_CONTINUED notification from the sender to reciever 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 Cc: Roland McGrath Cc: Jiri Kosina Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds --- Reading git-diff-tree failed