Date: Thu, 7 Jun 2007 10:59:40 -0400 From: John Baldwin <jhb@freebsd.org> To: Bruce Evans <brde@optusnet.com.au> Cc: src-committers@freebsd.org, Kip Macy <kip.macy@gmail.com>, cvs-all@freebsd.org, Attilio Rao <attilio@freebsd.org>, cvs-src@freebsd.org, Kostik Belousov <kostikbel@gmail.com>, Jeff Roberson <jroberson@chesapeake.net> Subject: Re: cvs commit: src/sys/kern kern_mutex.c Message-ID: <200706071059.41466.jhb@freebsd.org> In-Reply-To: <20070607180257.P7767@besplex.bde.org> References: <200706051420.l55EKEih018925@repoman.freebsd.org> <20070607163724.M7517@besplex.bde.org> <20070607180257.P7767@besplex.bde.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Thursday 07 June 2007 04:15:13 am Bruce Evans wrote:
> On Thu, 7 Jun 2007, Bruce Evans wrote:
>
> > (Cold client):
> > 834.39 real 1300.21 user 192.19 sys
> > 1323006 voluntary context switches
> > 1526348 involuntary context switches
> > ...
> > This is with 4BSD, no PREEMPTION, and pagezero disabled. With the
> > ...
> > The next run will have pagezero resetting its priority when this priority
> > gets clobbered.
>
> That gave only mainly more voluntary context switches (13.5+ million instead
> of the best observed value of 1.3+ million or the value of 2.9+ million
> without priority resetting. It reduced the pagezero time from 30 seconds to
> 24. It didn't change the real time significantly.
Hmm, one problem with the non-preemption page zero is that it doesn't yield
the lock when it voluntarily yields. I wonder if something like this patch
would help things for the non-preemption case:
Index: vm_zeroidle.c
===================================================================
RCS file: /usr/cvs/src/sys/vm/vm_zeroidle.c,v
retrieving revision 1.45
diff -u -r1.45 vm_zeroidle.c
--- vm_zeroidle.c 18 May 2007 07:10:50 -0000 1.45
+++ vm_zeroidle.c 7 Jun 2007 14:56:02 -0000
@@ -147,8 +147,10 @@
#ifndef PREEMPTION
if (sched_runnable()) {
mtx_lock_spin(&sched_lock);
+ mtx_unlock(&vm_page_queue_free_mtx);
mi_switch(SW_VOL, NULL);
mtx_unlock_spin(&sched_lock);
+ mtx_lock(&vm_page_queue_free_mtx);
}
#endif
} else {
We could simulate this behavior some by using a critical section to control
when preemptions happen so that we wait until we drop the lock perhaps to
allow preemptions. Something like this:
> Index: vm_zeroidle.c
===================================================================
RCS file: /usr/cvs/src/sys/vm/vm_zeroidle.c,v
retrieving revision 1.45
diff -u -r1.45 vm_zeroidle.c
--- vm_zeroidle.c 18 May 2007 07:10:50 -0000 1.45
+++ vm_zeroidle.c 7 Jun 2007 14:58:39 -0000
@@ -110,8 +110,10 @@
if (m != NULL && (m->flags & PG_ZERO) == 0) {
vm_pageq_remove_nowakeup(m);
mtx_unlock(&vm_page_queue_free_mtx);
+ critical_exit();
pmap_zero_page_idle(m);
mtx_lock(&vm_page_queue_free_mtx);
+ critical_enter();
m->flags |= PG_ZERO;
vm_pageq_enqueue(PQ_FREE + m->pc, m);
++vm_page_zero_count;
@@ -141,20 +143,25 @@
idlezero_enable = idlezero_enable_default;
mtx_lock(&vm_page_queue_free_mtx);
+ critical_enter();
for (;;) {
if (vm_page_zero_check()) {
vm_page_zero_idle();
#ifndef PREEMPTION
if (sched_runnable()) {
mtx_lock_spin(&sched_lock);
+ mtx_unlock(&vm_page_queue_free_mtx);
mi_switch(SW_VOL, NULL);
mtx_unlock_spin(&sched_lock);
+ mtx_lock(&vm_page_queue_free_mtx);
}
#endif
} else {
+ critical_exit();
wakeup_needed = TRUE;
msleep(&zero_state, &vm_page_queue_free_mtx, 0,
"pgzero", hz * 300);
+ critical_enter();
}
}
}
--
John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200706071059.41466.jhb>