Date: Mon, 18 Feb 2002 11:12:16 -0800 (PST) From: Matthew Dillon <dillon@apollo.backplane.com> To: current@freebsd.org Subject: Patch to improve mutex collision performance Message-ID: <200202181912.g1IJCGK32122@apollo.backplane.com>
next in thread | raw e-mail | index | archive | help
While testing some Giant removal stuff I noticed that my current
system sometimes got into an extremely non-optimal flip-flop situation
between two processes contesting Giant on an SMP system which halved the
syscall performance in the test.
In my getuid test, for example, with Giant in place I was getting
683Kcalls/sec with one process and 427Kcalls/sec with two. Giant
was being obtained in two places: in userret and in getuid().
When I turned off Giant in getuid() the syscall performance actually
went DOWN, to 250Kcalls/sec with two processes. This was a totally
unexpected result.
It turns out that the two processes got into an extremely non-optimal
contested/sleep/wakeup situation, even though they do not actually have
to sleep on Giant in this situation.
The solution is to allow _mtx_lock_sleep() to spin instead of sleep in
the situation where: (1) there are no runnable processes other then
the ones already running on a cpu, (2) interrupts are enabled, and
(3) the mutex in question is not contested (to avoid starving the thread
contesting the mutex). In this case we can spin.
This will go in tonight if no problems arise.
-Matt
Index: kern/kern_mutex.c
===================================================================
RCS file: /home/ncvs/src/sys/kern/kern_mutex.c,v
retrieving revision 1.80
diff -u -r1.80 kern_mutex.c
--- kern/kern_mutex.c 18 Feb 2002 17:51:47 -0000 1.80
+++ kern/kern_mutex.c 18 Feb 2002 19:11:17 -0000
@@ -287,7 +287,9 @@
_mtx_lock_sleep(struct mtx *m, int opts, const char *file, int line)
{
struct thread *td = curthread;
+#if 0
struct ksegrp *kg = td->td_ksegrp;
+#endif
if ((m->mtx_lock & MTX_FLAGMASK) == (uintptr_t)td) {
m->mtx_recurse++;
@@ -312,6 +314,22 @@
* the sched_lock.
*/
if ((v = m->mtx_lock) == MTX_UNOWNED) {
mtx_unlock_spin(&sched_lock);
continue;
}
+
+ /*
+ * Check to see if there are any runnable processes. If
+ * there aren't and nobody is contesting the mutex (to avoid
+ * starving a contester) and interrupts are enabled, then
+ * we can safely spin.
+ *
+ * This prevents a silly-sleep-flip-flop situation on SMP
+ * systems where two running processes need Giant (or any
+ * other sleep mutex).
+ */
+ if (td->td_critnest == 0 && (v & MTX_CONTESTED) == 0 &&
+ procrunnable() == 0) {
+ mtx_unlock_spin(&sched_lock);
+ continue;
+ }
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-current" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200202181912.g1IJCGK32122>