Commit 71a2224d authored by Christoph Lameter's avatar Christoph Lameter Committed by Linus Torvalds

[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: default avatarChristoph Lameter <christoph@lameter.com>
Signed-off-by: default avatarShai Fultheim <shai@scalex86.org>
Signed-off-by: default avatarAndrew Morton <akpm@osdl.org>
Signed-off-by: default avatarLinus Torvalds <torvalds@osdl.org>
parent b030a4dd
......@@ -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);
/*
......
......@@ -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);
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment