[PATCH] revert "Optimize sys_times for a single thread process"
Oleg Nesterov [Wed, 29 Mar 2006 00:11:19 +0000 (16:11 -0800)]
This patch reverts 'CONFIG_SMP && thread_group_empty()' optimization in
sys_times().  The reason is that the next patch breaks memory ordering which
is needed for that optimization.

tasklist_lock in sys_times() will be eliminated completely by further patch.

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>

kernel/exit.c
kernel/sys.c

index 3823ec8..6b2e4cf 100644 (file)
@@ -139,11 +139,7 @@ repeat:
        ptrace_unlink(p);
        BUG_ON(!list_empty(&p->ptrace_list) || !list_empty(&p->ptrace_children));
        __exit_signal(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);
 
        /*
index c93d37f..84371fd 100644 (file)
@@ -1202,69 +1202,35 @@ asmlinkage long sys_times(struct tms __user * tbuf)
         */
        if (tbuf) {
                struct tms tmp;
+               struct task_struct *tsk = current;
+               struct task_struct *t;
                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;
-
-                       read_lock(&tasklist_lock);
-                       utime = tsk->signal->utime;
-                       stime = tsk->signal->stime;
-                       t = tsk;
-                       do {
-                               utime = cputime_add(utime, t->utime);
-                               stime = cputime_add(stime, t->stime);
-                               t = next_thread(t);
-                       } while (t != tsk);
+               read_lock(&tasklist_lock);
+               utime = tsk->signal->utime;
+               stime = tsk->signal->stime;
+               t = tsk;
+               do {
+                       utime = cputime_add(utime, t->utime);
+                       stime = cputime_add(stime, t->stime);
+                       t = next_thread(t);
+               } while (t != tsk);
+
+               /*
+                * While we have tasklist_lock read-locked, no dying thread
+                * can be updating current->signal->[us]time.  Instead,
+                * we got their counts included in the live thread loop.
+                * However, another thread can come in right now and
+                * do a wait call that updates current->signal->c[us]time.
+                * To make sure we always see that pair updated atomically,
+                * we take the siglock around fetching them.
+                */
+               spin_lock_irq(&tsk->sighand->siglock);
+               cutime = tsk->signal->cutime;
+               cstime = tsk->signal->cstime;
+               spin_unlock_irq(&tsk->sighand->siglock);
+               read_unlock(&tasklist_lock);
 
-                       /*
-                        * While we have tasklist_lock read-locked, no dying thread
-                        * can be updating current->signal->[us]time.  Instead,
-                        * we got their counts included in the live thread loop.
-                        * However, another thread can come in right now and
-                        * do a wait call that updates current->signal->c[us]time.
-                        * To make sure we always see that pair updated atomically,
-                        * we take the siglock around fetching them.
-                        */
-                       spin_lock_irq(&tsk->sighand->siglock);
-                       cutime = tsk->signal->cutime;
-                       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);