[PATCH] Optimize sys_times for a single thread process

Avoid taking the tasklist_lock in sys_times if the process is single
threaded.  In a NUMA system taking the tasklist_lock may cause a bouncing
cacheline if multiple independent processes continually call sys_times to
measure their performance.
Signed-off-by: Christoph Lameter <christoph@lameter.com>
Signed-off-by: Shai Fultheim <shai@scalex86.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

kernel/exit.c

@@ -72,6 +72,11 @@ void release_task(struct task_struct * p)
 	BUG_ON(!list_empty(&p->ptrace_list) || !list_empty(&p->ptrace_children));
 	__exit_signal(p);
 	__exit_sighand(p);
+	/*
+	 * Note that the fastpath in sys_times depends on __exit_signal having
+	 * updated the counters before a task is removed from the tasklist of
+	 * the process by __unhash_process.
+	 */
 	__unhash_process(p);
 
 	/*

kernel/sys.c

@@ -894,9 +894,43 @@ asmlinkage long sys_times(struct tms __user * tbuf)
 	 */
 	if (tbuf) {
 		struct tms tmp;
+		cputime_t utime, stime, cutime, cstime;
+
+#ifdef CONFIG_SMP
+		if (thread_group_empty(current)) {
+			/*
+			 * Single thread case without the use of any locks.
+			 *
+			 * We may race with release_task if two threads are
+			 * executing. However, release task first adds up the
+			 * counters (__exit_signal) before  removing the task
+			 * from the process tasklist (__unhash_process).
+			 * __exit_signal also acquires and releases the
+			 * siglock which results in the proper memory ordering
+			 * so that the list modifications are always visible
+			 * after the counters have been updated.
+			 *
+			 * If the counters have been updated by the second thread
+			 * but the thread has not yet been removed from the list
+			 * then the other branch will be executing which will
+			 * block on tasklist_lock until the exit handling of the
+			 * other task is finished.
+			 *
+			 * This also implies that the sighand->siglock cannot
+			 * be held by another processor. So we can also
+			 * skip acquiring that lock.
+			 */
+			utime = cputime_add(current->signal->utime, current->utime);
+			stime = cputime_add(current->signal->utime, current->stime);
+			cutime = current->signal->cutime;
+			cstime = current->signal->cstime;
+		} else
+#endif
+		{
+
+			/* Process with multiple threads */
 			struct task_struct *tsk = current;
 			struct task_struct *t;
-		cputime_t utime, stime, cutime, cstime;
 
 			read_lock(&tasklist_lock);
 			utime = tsk->signal->utime;
@@ -922,7 +956,7 @@ asmlinkage long sys_times(struct tms __user * tbuf)
 			cstime = tsk->signal->cstime;
 			spin_unlock_irq(&tsk->sighand->siglock);
 			read_unlock(&tasklist_lock);
-
+		}
 		tmp.tms_utime = cputime_to_clock_t(utime);
 		tmp.tms_stime = cputime_to_clock_t(stime);
 		tmp.tms_cutime = cputime_to_clock_t(cutime);