[PATCH] do_coredump() should reset group_stop_count earlier

__group_complete_signal() sets ->group_stop_count in sig_kernel_coredump()
path and marks the target thread as ->group_exit_task.  So any thread
except group_exit_task will go to handle_group_stop()->finish_stop().

However, when group_exit_task actually starts do_coredump(), it sets
SIGNAL_GROUP_EXIT, but does not reset ->group_stop_count while killing
other threads.  If we have not yet stopped threads in the same thread
group, they all will spin in kernel mode until group_exit_task sends them
SIGKILL, because ->group_stop_count > 0 means:

	recalc_sigpending_tsk() never clears TIF_SIGPENDING

	get_signal_to_deliver() goes to handle_group_stop()

	handle_group_stop() returns when SIGNAL_GROUP_EXIT set

	syscall_exit/resume_userspace notice TIF_SIGPENDING,
	call get_signal_to_deliver() again.

So we are wasting cpu cycles, and if one of these threads is rt_task() this
may be a serious problem.

NOTE: do_coredump() holds ->mmap_sem, so not stopped threads can't escape
coredumping after clearing ->group_stop_count.

See also this thread: http://marc.theaimsgroup.com/?t=112739139900002

Signed-off-by: Oleg Nesterov <oleg@tv-sign.ru>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
diff --git a/fs/exec.c b/fs/exec.c
index 2075b67..fd02ea4a 100644
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -1462,6 +1462,7 @@
 	if (!(current->signal->flags & SIGNAL_GROUP_EXIT)) {
 		current->signal->flags = SIGNAL_GROUP_EXIT;
 		current->signal->group_exit_code = exit_code;
+		current->signal->group_stop_count = 0;
 		retval = 0;
 	}
 	spin_unlock_irq(&current->sighand->siglock);
@@ -1477,7 +1478,6 @@
 	 * Clear any false indication of pending signals that might
 	 * be seen by the filesystem code called to write the core file.
 	 */
-	current->signal->group_stop_count = 0;
 	clear_thread_flag(TIF_SIGPENDING);
 
 	if (current->signal->rlim[RLIMIT_CORE].rlim_cur < binfmt->min_coredump)