tweak the mclgeti algorithm to behave better under load.

instead of letting hardware rings grow on every interrupt, restrict
it so it can only grow once per softclock tick. we can only punish
the rings on softclock ticks, so it make sense to only grow on
softclock tick boundaries too.

the rings are now punished after >1 lost softclock tick rather than
>2. mclgeti is now more aggressive at detecting livelock.

the rings get punished by an 8th, rather than by half.

we now allow the rings to be punished again even if the system is
already considered in livelock.

without this diff a livelocked system will have its rx ring sizes
scale up and down very rapidly, while holding the rings low for too
long. this affected throughput significantly.

discussed and tested heavily at j2k10. there are still some games
with softnet we can play, but this is a good first step.

"put it in" and ok deraadt@
ok claudio@ krw@ henning@ mcbride@

if we find out that it sucks we can pull it out again later. till then
we'll run with it and see how it goes.

1 files changed, 2 insertions, 2 deletions

diff --git a/usr.bin/systat/mbufs.c b/usr.bin/systat/mbufs.c
index 2bbbbc979a0..dac87bc8430 100644
--- a/usr.bin/systat/mbufs.c
+++ b/usr.bin/systat/mbufs.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: mbufs.c,v 1.28 2010/08/01 05:30:13 blambert Exp $ */
+/*	$OpenBSD: mbufs.c,v 1.29 2010/09/23 10:49:55 dlg Exp $ */
 /*
  * Copyright (c) 2008 Can Erkin Acar <canacar@openbsd.org>
  *
@@ -349,9 +349,9 @@ showmbuf(struct if_info *ifi, int p, int showif)
 #if NOTYET
 	print_fld_uint(FLD_MB_RXDELAY, ifi->data.ifi_rxdelay);
 	print_fld_uint(FLD_MB_TXDELAY, ifi->data.ifi_txdelay);
-#endif
 	if (ifi->data.ifi_livelocks)
 		print_fld_size(FLD_MB_LLOCKS, ifi->data.ifi_livelocks);
+#endif
 
 	if (p >= 0 && p < mclpool_count) {
 		struct mclpool *mp = &ifi->data.ifi_mclpool[p];