Dynamic buffer cache support - a re-commit of what was backed out

after c2k9

allows buffer cache to be extended and grow/shrink dynamically

tested by many, ok oga@, "why not just commit it" deraadt@

26 files changed, 241 insertions, 174 deletions

diff --git a/sys/arch/amd64/amd64/machdep.c b/sys/arch/amd64/amd64/machdep.c
index 14d257736fb..c1745a0f857 100644
--- a/sys/arch/amd64/amd64/machdep.c
+++ b/sys/arch/amd64/amd64/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.97 2009/07/22 20:33:12 deraadt Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.98 2009/08/02 16:28:39 beck Exp $	*/
 /*	$NetBSD: machdep.c,v 1.3 2003/05/07 22:58:18 fvdl Exp $	*/
 
 /*-
@@ -399,12 +399,6 @@ setup_buffers()
 	 */
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
-
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
 }
 
 /*
diff --git a/sys/arch/aviion/aviion/machdep.c b/sys/arch/aviion/aviion/machdep.c
index ebff8ba9465..688d3e06389 100644
--- a/sys/arch/aviion/aviion/machdep.c
+++ b/sys/arch/aviion/aviion/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.29 2009/06/15 17:01:25 beck Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.30 2009/08/02 16:28:39 beck Exp $	*/
 /*
  * Copyright (c) 2007 Miodrag Vallat.
  *
@@ -293,12 +293,6 @@ cpu_startup()
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 	/*
 	 * Allocate a submap for exec arguments.  This map effectively
 	 * limits the number of processes exec'ing at any time.
diff --git a/sys/arch/hp300/hp300/machdep.c b/sys/arch/hp300/hp300/machdep.c
index c4b391071c6..8661452c4f8 100644
--- a/sys/arch/hp300/hp300/machdep.c
+++ b/sys/arch/hp300/hp300/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.122 2009/06/15 17:01:25 beck Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.123 2009/08/02 16:28:39 beck Exp $	*/
 /*	$NetBSD: machdep.c,v 1.121 1999/03/26 23:41:29 mycroft Exp $	*/
 
 /*
@@ -290,12 +290,6 @@ cpu_startup()
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 	/*
 	 * Allocate a submap for exec arguments.  This map effectively
 	 * limits the number of processes exec'ing at any time.
diff --git a/sys/arch/hppa/hppa/machdep.c b/sys/arch/hppa/hppa/machdep.c
index d7053f805f0..68858a6f1bd 100644
--- a/sys/arch/hppa/hppa/machdep.c
+++ b/sys/arch/hppa/hppa/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.174 2009/07/29 18:31:11 kettenis Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.175 2009/08/02 16:28:39 beck Exp $	*/
 
 /*
  * Copyright (c) 1999-2003 Michael Shalayeff
@@ -656,12 +656,6 @@ cpu_startup(void)
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 	/*
 	 * Allocate a submap for exec arguments.  This map effectively
 	 * limits the number of processes exec'ing at any time.
diff --git a/sys/arch/hppa64/hppa64/machdep.c b/sys/arch/hppa64/hppa64/machdep.c
index 33a408a7a02..028dc504818 100644
--- a/sys/arch/hppa64/hppa64/machdep.c
+++ b/sys/arch/hppa64/hppa64/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.15 2009/06/15 17:01:25 beck Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.16 2009/08/02 16:28:39 beck Exp $	*/
 
 /*
  * Copyright (c) 2005 Michael Shalayeff
@@ -406,12 +406,6 @@ cpu_startup(void)
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 printf("here3\n");
 	/*
 	 * Allocate a submap for exec arguments.  This map effectively
diff --git a/sys/arch/i386/i386/machdep.c b/sys/arch/i386/i386/machdep.c
index ea65c0cd522..98750e512b8 100644
--- a/sys/arch/i386/i386/machdep.c
+++ b/sys/arch/i386/i386/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.454 2009/07/27 11:28:55 dms Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.455 2009/08/02 16:28:39 beck Exp $	*/
 /*	$NetBSD: machdep.c,v 1.214 1996/11/10 03:16:17 thorpej Exp $	*/
 
 /*-
@@ -547,11 +547,6 @@ setup_buffers()
 	if (bufpages == 0)
 		bufpages = atop(avail_end) * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4)
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
 }
 
 /*
diff --git a/sys/arch/luna88k/luna88k/machdep.c b/sys/arch/luna88k/luna88k/machdep.c
index 03281cec0cb..ff95fd3e908 100644
--- a/sys/arch/luna88k/luna88k/machdep.c
+++ b/sys/arch/luna88k/luna88k/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.64 2009/06/15 17:01:26 beck Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.65 2009/08/02 16:28:39 beck Exp $	*/
 /*
  * Copyright (c) 1998, 1999, 2000, 2001 Steve Murphree, Jr.
  * Copyright (c) 1996 Nivas Madhur
@@ -470,12 +470,6 @@ cpu_startup()
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 	/*
 	 * Allocate a submap for exec arguments.  This map effectively
 	 * limits the number of processes exec'ing at any time.
diff --git a/sys/arch/mac68k/mac68k/machdep.c b/sys/arch/mac68k/mac68k/machdep.c
index 9f992582dca..1122c08d605 100644
--- a/sys/arch/mac68k/mac68k/machdep.c
+++ b/sys/arch/mac68k/mac68k/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.149 2009/06/15 17:01:26 beck Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.150 2009/08/02 16:28:39 beck Exp $	*/
 /*	$NetBSD: machdep.c,v 1.207 1998/07/08 04:39:34 thorpej Exp $	*/
 
 /*
@@ -419,12 +419,6 @@ cpu_startup(void)
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 	/*
 	 * Allocate a submap for exec arguments.  This map effectively
 	 * limits the number of processes exec'ing at any time.
diff --git a/sys/arch/macppc/macppc/machdep.c b/sys/arch/macppc/macppc/machdep.c
index f4c9ba24af1..e7cc2ed692f 100644
--- a/sys/arch/macppc/macppc/machdep.c
+++ b/sys/arch/macppc/macppc/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.109 2009/07/26 18:48:55 miod Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.110 2009/08/02 16:28:39 beck Exp $	*/
 /*	$NetBSD: machdep.c,v 1.4 1996/10/16 19:33:11 ws Exp $	*/
 
 /*
@@ -509,12 +509,6 @@ cpu_startup()
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 	/*
 	 * Allocate a submap for exec arguments.  This map effectively
 	 * limits the number of processes exec'ing at any time.
diff --git a/sys/arch/mvme68k/mvme68k/machdep.c b/sys/arch/mvme68k/mvme68k/machdep.c
index db6c36898db..a17d82742f5 100644
--- a/sys/arch/mvme68k/mvme68k/machdep.c
+++ b/sys/arch/mvme68k/mvme68k/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.116 2009/06/15 17:01:26 beck Exp $ */
+/*	$OpenBSD: machdep.c,v 1.117 2009/08/02 16:28:39 beck Exp $ */
 
 /*
  * Copyright (c) 1995 Theo de Raadt
@@ -265,12 +265,6 @@ cpu_startup()
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 	/*
 	 * Allocate a submap for exec arguments.  This map effectively
 	 * limits the number of processes exec'ing at any time.
diff --git a/sys/arch/mvme88k/mvme88k/machdep.c b/sys/arch/mvme88k/mvme88k/machdep.c
index 063acd72a66..c9d52962de3 100644
--- a/sys/arch/mvme88k/mvme88k/machdep.c
+++ b/sys/arch/mvme88k/mvme88k/machdep.c
@@ -1,4 +1,4 @@
-/* $OpenBSD: machdep.c,v 1.231 2009/06/15 17:01:26 beck Exp $	*/
+/* $OpenBSD: machdep.c,v 1.232 2009/08/02 16:28:39 beck Exp $	*/
 /*
  * Copyright (c) 1998, 1999, 2000, 2001 Steve Murphree, Jr.
  * Copyright (c) 1996 Nivas Madhur
@@ -404,12 +404,6 @@ cpu_startup()
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 	/*
 	 * Allocate a submap for exec arguments.  This map effectively
 	 * limits the number of processes exec'ing at any time.
diff --git a/sys/arch/mvmeppc/mvmeppc/machdep.c b/sys/arch/mvmeppc/mvmeppc/machdep.c
index a2432cce471..ea3a62855c8 100644
--- a/sys/arch/mvmeppc/mvmeppc/machdep.c
+++ b/sys/arch/mvmeppc/mvmeppc/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.58 2009/06/15 17:01:26 beck Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.59 2009/08/02 16:28:39 beck Exp $	*/
 /*	$NetBSD: machdep.c,v 1.4 1996/10/16 19:33:11 ws Exp $	*/
 
 /*
@@ -415,12 +415,6 @@ cpu_startup()
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 	/*
 	 * Allocate a submap for exec arguments.  This map effectively
 	 * limits the number of processes exec'ing at any time.
diff --git a/sys/arch/sgi/sgi/machdep.c b/sys/arch/sgi/sgi/machdep.c
index 0737cc91a60..af07a9de93b 100644
--- a/sys/arch/sgi/sgi/machdep.c
+++ b/sys/arch/sgi/sgi/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.75 2009/07/22 21:29:05 miod Exp $ */
+/*	$OpenBSD: machdep.c,v 1.76 2009/08/02 16:28:39 beck Exp $ */
 
 /*
  * Copyright (c) 2003-2004 Opsycon AB  (www.opsycon.se / www.opsycon.com)
@@ -766,12 +766,6 @@ cpu_startup()
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm. */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 	/*
 	 * Allocate a submap for exec arguments. This map effectively
 	 * limits the number of processes exec'ing at any time.
diff --git a/sys/arch/socppc/socppc/machdep.c b/sys/arch/socppc/socppc/machdep.c
index 509a5e6a9bb..b6229ea6ce1 100644
--- a/sys/arch/socppc/socppc/machdep.c
+++ b/sys/arch/socppc/socppc/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.14 2009/07/26 18:48:55 miod Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.15 2009/08/02 16:28:39 beck Exp $	*/
 /*	$NetBSD: machdep.c,v 1.4 1996/10/16 19:33:11 ws Exp $	*/
 
 /*
@@ -772,12 +772,6 @@ cpu_startup()
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 	/*
 	 * Allocate a submap for exec arguments.  This map effectively
 	 * limits the number of processes exec'ing at any time.
diff --git a/sys/arch/solbourne/solbourne/machdep.c b/sys/arch/solbourne/solbourne/machdep.c
index 0c90df97b4d..77d353bb009 100644
--- a/sys/arch/solbourne/solbourne/machdep.c
+++ b/sys/arch/solbourne/solbourne/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.10 2009/06/15 17:01:26 beck Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.11 2009/08/02 16:28:39 beck Exp $	*/
 /*	OpenBSD: machdep.c,v 1.105 2005/04/11 15:13:01 deraadt Exp 	*/
 
 /*
@@ -190,12 +190,6 @@ cpu_startup()
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 	/*
 	 * Allocate a submap for exec arguments.  This map effectively
 	 * limits the number of processes exec'ing at any time.
diff --git a/sys/arch/sparc/sparc/machdep.c b/sys/arch/sparc/sparc/machdep.c
index 0cc40477def..b54c0096f4b 100644
--- a/sys/arch/sparc/sparc/machdep.c
+++ b/sys/arch/sparc/sparc/machdep.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: machdep.c,v 1.119 2009/07/13 19:50:00 kettenis Exp $	*/
+/*	$OpenBSD: machdep.c,v 1.120 2009/08/02 16:28:39 beck Exp $	*/
 /*	$NetBSD: machdep.c,v 1.85 1997/09/12 08:55:02 pk Exp $ */
 
 /*
@@ -204,12 +204,6 @@ cpu_startup()
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 	/*
 	 * Allocate a submap for exec arguments.  This map effectively
 	 * limits the number of processes exec'ing at any time.
diff --git a/sys/arch/vax/vax/machdep.c b/sys/arch/vax/vax/machdep.c
index 7d0ac9ae3b4..c9ba240162b 100644
--- a/sys/arch/vax/vax/machdep.c
+++ b/sys/arch/vax/vax/machdep.c
@@ -1,4 +1,4 @@
-/* $OpenBSD: machdep.c,v 1.102 2009/06/20 21:02:15 miod Exp $ */
+/* $OpenBSD: machdep.c,v 1.103 2009/08/02 16:28:39 beck Exp $ */
 /* $NetBSD: machdep.c,v 1.108 2000/09/13 15:00:23 thorpej Exp $	 */
 
 /*
@@ -212,12 +212,6 @@ cpu_startup()
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
-	/* Restrict to at most 25% filled kvm */
-	if (bufpages >
-	    (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) / PAGE_SIZE / 4) 
-		bufpages = (VM_MAX_KERNEL_ADDRESS-VM_MIN_KERNEL_ADDRESS) /
-		    PAGE_SIZE / 4;
-
 	/*
 	 * Allocate a submap for exec arguments.  This map effectively limits
 	 * the number of processes exec'ing at any time.
diff --git a/sys/kern/kern_sysctl.c b/sys/kern/kern_sysctl.c
index 82b2bd64b2f..41782ae2fba 100644
--- a/sys/kern/kern_sysctl.c
+++ b/sys/kern/kern_sysctl.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: kern_sysctl.c,v 1.177 2009/07/21 14:10:14 millert Exp $	*/
+/*	$OpenBSD: kern_sysctl.c,v 1.178 2009/08/02 16:28:39 beck Exp $	*/
 /*	$NetBSD: kern_sysctl.c,v 1.17 1996/05/20 17:49:05 mrg Exp $	*/
 
 /*-
@@ -555,12 +555,21 @@ kern_sysctl(int *name, u_int namelen, void *oldp, size_t *oldlenp, void *newp,
 		return (sysctl_cptime2(name + 1, namelen -1, oldp, oldlenp,
 		    newp, newlen));
 	case KERN_CACHEPCT: {
-		int opct = 0;
-
+		int opct, pgs;
+		opct = bufcachepercent;
 		error = sysctl_int(oldp, oldlenp, newp, newlen,
-		    &opct);
+		    &bufcachepercent);
 		if (error)
 			return(error);
+		if (bufcachepercent > 90 || bufcachepercent < 5) {
+			bufcachepercent = opct;
+			return (EINVAL);
+		}
+		if (bufcachepercent != opct) {
+			pgs = bufcachepercent * physmem / 100;
+			bufadjust(pgs); /* adjust bufpages */
+			bufhighpages = bufpages; /* set high water mark */
+		}
 		return(0);
 	}
 	default:
diff --git a/sys/kern/vfs_bio.c b/sys/kern/vfs_bio.c
index 1322bb93feb..045bd969ee3 100644
--- a/sys/kern/vfs_bio.c
+++ b/sys/kern/vfs_bio.c
@@ -1,7 +1,7 @@
-/*	$OpenBSD: vfs_bio.c,v 1.118 2009/06/25 15:49:26 thib Exp $	*/
+/*	$OpenBSD: vfs_bio.c,v 1.119 2009/08/02 16:28:40 beck Exp $	*/
 /*	$NetBSD: vfs_bio.c,v 1.44 1996/06/11 11:15:36 pk Exp $	*/
 
-/*-
+/*
  * Copyright (c) 1994 Christopher G. Demetriou
  * Copyright (c) 1982, 1986, 1989, 1993
  *	The Regents of the University of California.  All rights reserved.
@@ -62,20 +62,6 @@
 #include <miscfs/specfs/specdev.h>
 
 /*
- * Definitions for the buffer hash lists.
- */
-#define	BUFHASH(dvp, lbn)	\
-	(&bufhashtbl[((long)(dvp) / sizeof(*(dvp)) + (int)(lbn)) & bufhash])
-LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash;
-u_long	bufhash;
-
-/*
- * Insq/Remq for the buffer hash lists.
- */
-#define	binshash(bp, dp)	LIST_INSERT_HEAD(dp, bp, b_hash)
-#define	bremhash(bp)		LIST_REMOVE(bp, b_hash)
-
-/*
  * Definitions for the buffer free lists.
  */
 #define	BQUEUES		2		/* number of free buffer queues */
@@ -123,6 +109,9 @@ long hidirtypages;
 long locleanpages;
 long hicleanpages;
 long maxcleanpages;
+long backoffpages;	/* backoff counter for page allocations */
+long buflowpages;	/* bufpages low water mark */
+long bufhighpages; 	/* bufpages high water mark */
 
 /* XXX - should be defined here. */
 extern int bufcachepercent;
@@ -182,9 +171,13 @@ buf_put(struct buf *bp)
 		panic("buf_put: b_dep is not empty");
 #endif
 
-	bremhash(bp);
 	LIST_REMOVE(bp, b_list);
 	bcstats.numbufs--;
+	if (backoffpages) {
+		backoffpages -= atop(bp->b_bufsize);
+		if (backoffpages < 0)
+			backoffpages = 0;
+	}
 
 	if (buf_dealloc_mem(bp) != 0)
 		return;
@@ -200,7 +193,7 @@ bufinit(void)
 	struct bqueues *dp;
 
 	/* XXX - for now */
-	bufpages = bufcachepercent = bufkvm = 0;
+	bufhighpages = buflowpages = bufpages = bufcachepercent = bufkvm = 0;
 
 	/*
 	 * If MD code doesn't say otherwise, use 10% of kvm for mappings and
@@ -211,6 +204,16 @@ bufinit(void)
 	if (bufpages == 0)
 		bufpages = physmem * bufcachepercent / 100;
 
+	bufhighpages = bufpages;
+
+	/*
+	 * set the base backoff level for the buffer cache to bufpages.
+	 * we will not allow uvm to steal back more than this number of
+	 * pages
+	 */
+	buflowpages = physmem * 10 / 100;
+
+
 	if (bufkvm == 0)
 		bufkvm = (VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS) / 10;
 
@@ -237,7 +240,6 @@ bufinit(void)
  	 */
 	buf_mem_init(bufkvm);
 
-	bufhashtbl = hashinit(bufpages / 4, M_CACHE, M_WAITOK, &bufhash);
 	hidirtypages = (bufpages / 4) * 3;
 	lodirtypages = bufpages / 2;
 
@@ -251,6 +253,104 @@ bufinit(void)
 	maxcleanpages = locleanpages;
 }
 
+/*
+ * Change cachepct
+ */
+void
+bufadjust(int newbufpages)
+{
+	/*
+	 * XXX - note, bufkvm was allocated once, based on 10% of physmem
+	 * see above.
+	 */
+	struct buf *bp;
+	int s;
+
+	s = splbio();
+	bufpages = newbufpages;
+
+	hidirtypages = (bufpages / 4) * 3;
+	lodirtypages = bufpages / 2;
+
+	/*
+	 * When we hit 95% of pages being clean, we bring them down to
+	 * 90% to have some slack.
+	 */
+	hicleanpages = bufpages - (bufpages / 20);
+	locleanpages = bufpages - (bufpages / 10);
+
+	maxcleanpages = locleanpages;
+
+	/*
+	 * If we we have more buffers allocated than bufpages,
+	 * free them up to get back down. this may possibly consume
+	 * all our clean pages...
+	 */
+	while ((bp = TAILQ_FIRST(&bufqueues[BQ_CLEAN])) &&
+	    (bcstats.numbufpages > bufpages)) {
+		bremfree(bp);
+		if (bp->b_vp) {
+			RB_REMOVE(buf_rb_bufs,
+			    &bp->b_vp->v_bufs_tree, bp);
+			brelvp(bp);
+		}
+		buf_put(bp);
+	}
+
+	/*
+	 * Wake up cleaner if we're getting low on pages. We might
+	 * now have too much dirty, or have fallen below our low
+	 * water mark on clean pages so we need to free more stuff
+	 * up.
+	 */
+	if (bcstats.numdirtypages >= hidirtypages ||
+	    bcstats.numcleanpages <= locleanpages)
+		wakeup(&bd_req);
+
+	/*
+	 * if immediate action has not freed up enough goo for us
+	 * to proceed - we tsleep and wait for the cleaner above
+	 * to do it's work and get us reduced down to sanity.
+	 */
+	while (bcstats.numbufpages > bufpages) {
+		tsleep(&needbuffer, PRIBIO, "needbuffer", 0);
+	}
+	splx(s);
+}
+
+/*
+ * Make the buffer cache back off from cachepct.
+ */
+int
+bufbackoff()
+{
+	/*
+	 * Back off the amount of buffer cache pages. Called by the page
+	 * daemon to consume buffer cache pages rather than swapping.
+	 *
+	 * On success, it frees N pages from the buffer cache, and sets
+	 * a flag so that the next N allocations from buf_get will recycle
+	 * a buffer rather than allocate a new one. It then returns 0 to the
+	 * caller. 
+	 *
+	 * on failure, it could free no pages from the buffer cache, does
+	 * nothing and returns -1 to the caller. 
+	 */
+	long d;
+
+	if (bufpages <= buflowpages) 
+		return(-1);
+
+	if (bufpages - BACKPAGES >= buflowpages)
+		d = BACKPAGES;
+	else
+		d = bufpages - buflowpages;
+	backoffpages = BACKPAGES;
+	bufadjust(bufpages - d);
+	backoffpages = BACKPAGES;
+	return(0);
+}
+
 struct buf *
 bio_doread(struct vnode *vp, daddr64_t blkno, int size, int async)
 {
@@ -676,10 +776,12 @@ brelse(struct buf *bp)
 			CLR(bp->b_flags, B_DELWRI);
 		}
 
-		if (bp->b_vp)
+		if (bp->b_vp) {
+			RB_REMOVE(buf_rb_bufs, &bp->b_vp->v_bufs_tree,
+			    bp);
 			brelvp(bp);
-		bremhash(bp);
-		binshash(bp, &invalhash);
+		}
+		bp->b_vp = NULL;
 
 		/*
 		 * If the buffer has no associated data, place it back in the
@@ -697,6 +799,9 @@ brelse(struct buf *bp)
 				CLR(bp->b_flags, B_WANTED);
 				wakeup(bp);
 			}
+			if (bp->b_vp != NULL)
+				RB_REMOVE(buf_rb_bufs,
+				    &bp->b_vp->v_bufs_tree, bp);
 			buf_put(bp);
 			splx(s);
 			return;
@@ -758,15 +863,14 @@ struct buf *
 incore(struct vnode *vp, daddr64_t blkno)
 {
 	struct buf *bp;
-
-	/* Search hash chain */
-	LIST_FOREACH(bp, BUFHASH(vp, blkno), b_hash) {
-		if (bp->b_lblkno == blkno && bp->b_vp == vp &&
-		    !ISSET(bp->b_flags, B_INVAL))
-			return (bp);
-	}
-
-	return (NULL);
+	struct buf b;
+
+	/* Search buf lookup tree */
+	b.b_lblkno = blkno;
+	bp = RB_FIND(buf_rb_bufs, &vp->v_bufs_tree, &b);
+	if (bp && !ISSET(bp->b_flags, B_INVAL))
+		return(bp);
+	return(NULL);
 }
 
 /*
@@ -781,6 +885,7 @@ struct buf *
 getblk(struct vnode *vp, daddr64_t blkno, int size, int slpflag, int slptimeo)
 {
 	struct buf *bp;
+	struct buf b;
 	int s, error;
 
 	/*
@@ -794,9 +899,9 @@ getblk(struct vnode *vp, daddr64_t blkno, int size, int slpflag, int slptimeo)
 	 * the block until the write is finished.
 	 */
 start:
-	LIST_FOREACH(bp, BUFHASH(vp, blkno), b_hash) {
-		if (bp->b_lblkno != blkno || bp->b_vp != vp)
-			continue;
+	b.b_lblkno = blkno;
+	bp = RB_FIND(buf_rb_bufs, &vp->v_bufs_tree, &b);
+	if (bp != NULL) {
 
 		s = splbio();
 		if (ISSET(bp->b_flags, B_BUSY)) {
@@ -846,11 +951,24 @@ geteblk(int size)
 struct buf *
 buf_get(struct vnode *vp, daddr64_t blkno, size_t size)
 {
+	static int gcount = 0;
 	struct buf *bp;
 	int poolwait = size == 0 ? PR_NOWAIT : PR_WAITOK;
 	int npages;
 	int s;
 
+	/*
+	 * if we were previously backed off, slowly climb back up
+	 * to the high water mark again.
+	 */
+	if ((backoffpages == 0) && (bufpages < bufhighpages)) {
+		if ( gcount == 0 )  {
+			bufadjust(bufpages + BACKPAGES);
+			gcount += BACKPAGES;
+		} else
+			gcount--;
+	}
+
 	s = splbio();
 	if (size) {
 		/*
@@ -868,8 +986,11 @@ buf_get(struct vnode *vp, daddr64_t blkno, size_t size)
 			while (bcstats.numcleanpages > locleanpages) {
 				bp = TAILQ_FIRST(&bufqueues[BQ_CLEAN]);
 				bremfree(bp);
-				if (bp->b_vp)
+				if (bp->b_vp) {
+					RB_REMOVE(buf_rb_bufs,
+					    &bp->b_vp->v_bufs_tree, bp);
 					brelvp(bp);
+				}
 				buf_put(bp);
 			}
 		}
@@ -879,16 +1000,21 @@ buf_get(struct vnode *vp, daddr64_t blkno, size_t size)
 		/*
 		 * Free some buffers until we have enough space.
 		 */
-		while (bcstats.numbufpages + npages > bufpages) {
+		while ((bcstats.numbufpages + npages > bufpages)
+		    || backoffpages) {
 			int freemax = 5;
 			int i = freemax;
 			while ((bp = TAILQ_FIRST(&bufqueues[BQ_CLEAN])) && i--) {
 				bremfree(bp);
-				if (bp->b_vp)
+				if (bp->b_vp) {
+					RB_REMOVE(buf_rb_bufs,
+					    &bp->b_vp->v_bufs_tree, bp);
 					brelvp(bp);
+				}
 				buf_put(bp);
 			}
-			if (freemax == i) {
+			if (freemax == i &&
+			    (bcstats.numbufpages + npages > bufpages)) {
 				needbuffer++;
 				tsleep(&needbuffer, PRIBIO, "needbuffer", 0);
 				splx(s);
@@ -929,11 +1055,12 @@ buf_get(struct vnode *vp, daddr64_t blkno, size_t size)
 
 		bp->b_blkno = bp->b_lblkno = blkno;
 		bgetvp(vp, bp);
-		binshash(bp, BUFHASH(vp, blkno));
+		if (RB_INSERT(buf_rb_bufs, &vp->v_bufs_tree, bp))
+			panic("buf_get: dup lblk vp %p bp %p", vp, bp);
 	} else {
 		bp->b_vnbufs.le_next = NOLIST;
 		SET(bp->b_flags, B_INVAL);
-		binshash(bp, &invalhash);
+		bp->b_vp = NULL;
 	}
 
 	LIST_INSERT_HEAD(&bufhead, bp, b_list);
diff --git a/sys/kern/vfs_subr.c b/sys/kern/vfs_subr.c
index 4b19f59ef2a..7a420752b26 100644
--- a/sys/kern/vfs_subr.c
+++ b/sys/kern/vfs_subr.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: vfs_subr.c,v 1.179 2009/06/25 15:49:26 thib Exp $	*/
+/*	$OpenBSD: vfs_subr.c,v 1.180 2009/08/02 16:28:40 beck Exp $	*/
 /*	$NetBSD: vfs_subr.c,v 1.53 1996/04/22 01:39:13 christos Exp $	*/
 
 /*
@@ -59,6 +59,7 @@
 #include <sys/mbuf.h>
 #include <sys/syscallargs.h>
 #include <sys/pool.h>
+#include <sys/tree.h>
 
 #include <uvm/uvm_extern.h>
 #include <sys/sysctl.h>
@@ -115,6 +116,19 @@ void printlockedvnodes(void);
 
 struct pool vnode_pool;
 
+static int rb_buf_compare(struct buf *b1, struct buf *b2);
+RB_GENERATE(buf_rb_bufs, buf, b_rbbufs, rb_buf_compare);
+
+static int
+rb_buf_compare(struct buf *b1, struct buf *b2)
+{
+	if (b1->b_lblkno < b2->b_lblkno)
+		return(-1);
+	if (b1->b_lblkno > b2->b_lblkno)
+		return(1);
+	return(0);
+}
+
 /*
  * Initialize the vnode management data structures.
  */
@@ -345,6 +359,7 @@ getnewvnode(enum vtagtype tag, struct mount *mp, int (**vops)(void *),
 	    ((TAILQ_FIRST(listhd = &vnode_hold_list) == NULL) || toggle))) {
 		splx(s);
 		vp = pool_get(&vnode_pool, PR_WAITOK | PR_ZERO);
+		RB_INIT(&vp->v_bufs_tree);
 		numvnodes++;
 	} else {
 		for (vp = TAILQ_FIRST(listhd); vp != NULLVP;
diff --git a/sys/sys/buf.h b/sys/sys/buf.h
index ded9e03f319..3abc87d19d9 100644
--- a/sys/sys/buf.h
+++ b/sys/sys/buf.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: buf.h,v 1.66 2009/06/17 01:30:32 thib Exp $	*/
+/*	$OpenBSD: buf.h,v 1.67 2009/08/02 16:28:40 beck Exp $	*/
 /*	$NetBSD: buf.h,v 1.25 1997/04/09 21:12:17 mycroft Exp $	*/
 
 /*
@@ -40,12 +40,16 @@
 #ifndef _SYS_BUF_H_
 #define	_SYS_BUF_H_
 #include <sys/queue.h>
+#include <sys/tree.h>
 
 #define NOLIST ((struct buf *)0x87654321)
 
 struct buf;
 struct vnode;
 
+struct buf_rb_bufs;
+RB_PROTOTYPE(buf_rb_bufs, buf, b_rbbufs, rb_buf_compare);
+
 LIST_HEAD(bufhead, buf);
 
 /*
@@ -72,8 +76,8 @@ extern struct bio_ops {
  * The buffer header describes an I/O operation in the kernel.
  */
 struct buf {
+	RB_ENTRY(buf) b_rbbufs;		/* vnode "hash" tree */
 	LIST_ENTRY(buf) b_list;		/* All allocated buffers. */
-	LIST_ENTRY(buf) b_hash;		/* Hash chain. */
 	LIST_ENTRY(buf) b_vnbufs;	/* Buffer's associated vnode. */
 	TAILQ_ENTRY(buf) b_freelist;	/* Free list position if not active. */
 	time_t	b_synctime;		/* Time this buffer should be flushed */
diff --git a/sys/sys/mount.h b/sys/sys/mount.h
index b856a781827..ac3da0f4959 100644
--- a/sys/sys/mount.h
+++ b/sys/sys/mount.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: mount.h,v 1.95 2009/06/15 17:01:26 beck Exp $	*/
+/*	$OpenBSD: mount.h,v 1.96 2009/08/02 16:28:40 beck Exp $	*/
 /*	$NetBSD: mount.h,v 1.48 1996/02/18 11:55:47 fvdl Exp $	*/
 
 /*
@@ -504,8 +504,14 @@ struct bcachestats {
 	int64_t cachehits;		/* total reads found in cache */
 };
 #ifdef _KERNEL
+#define BACKPAGES 100
 extern struct bcachestats bcstats;
-#define BUFPAGES_DEFICIT (bufpages - bcstats.numbufpages)
+extern long buflowpages, bufhighpages;
+#define BUFPAGES_DEFICIT (((buflowpages - bcstats.numbufpages) < 0) ? 0 \
+    : buflowpages - bcstats.numbufpages)
+extern int bufcachepercent;
+extern void bufadjust(int);
+extern int bufbackoff(void);
 #endif
 
 /*
diff --git a/sys/sys/syscall.h b/sys/sys/syscall.h
index 645d9314448..8dc92f61d30 100644
--- a/sys/sys/syscall.h
+++ b/sys/sys/syscall.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: syscall.h,v 1.106 2009/06/15 17:31:49 deraadt Exp $	*/
+/*	$OpenBSD: syscall.h,v 1.107 2009/08/02 16:28:40 beck Exp $	*/
 
 /*
  * System call numbers.
diff --git a/sys/sys/syscallargs.h b/sys/sys/syscallargs.h
index e77e06394e1..ae8104b81b9 100644
--- a/sys/sys/syscallargs.h
+++ b/sys/sys/syscallargs.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: syscallargs.h,v 1.108 2009/06/15 17:31:49 deraadt Exp $	*/
+/*	$OpenBSD: syscallargs.h,v 1.109 2009/08/02 16:28:40 beck Exp $	*/
 
 /*
  * System call argument lists.
diff --git a/sys/sys/vnode.h b/sys/sys/vnode.h
index ffa752ca279..1608736d418 100644
--- a/sys/sys/vnode.h
+++ b/sys/sys/vnode.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: vnode.h,v 1.101 2009/07/09 22:29:56 thib Exp $	*/
+/*	$OpenBSD: vnode.h,v 1.102 2009/08/02 16:28:40 beck Exp $	*/
 /*	$NetBSD: vnode.h,v 1.38 1996/02/29 20:59:05 cgd Exp $	*/
 
 /*
@@ -32,10 +32,12 @@
  *	@(#)vnode.h	8.11 (Berkeley) 11/21/94
  */
 
+#include <sys/buf.h>
 #include <sys/types.h>
 #include <sys/queue.h>
 #include <sys/lock.h>
 #include <sys/selinfo.h>
+#include <sys/tree.h>
 
 #include <uvm/uvm.h>
 #include <uvm/uvm_vnode.h>
@@ -79,6 +81,8 @@ enum vtagtype	{
  */
 LIST_HEAD(buflists, buf);
 
+RB_HEAD(buf_rb_bufs, buf);
+
 struct vnode {
 	struct uvm_vnode v_uvm;			/* uvm data */
 	int	(**v_op)(void *);		/* vnode operations vector */
@@ -94,6 +98,7 @@ struct vnode {
 	struct	mount *v_mount;			/* ptr to vfs we are in */
 	TAILQ_ENTRY(vnode) v_freelist;		/* vnode freelist */
 	LIST_ENTRY(vnode) v_mntvnodes;		/* vnodes for mount point */
+	struct	buf_rb_bufs v_bufs_tree;	/* lookup of all bufs */
 	struct	buflists v_cleanblkhd;		/* clean blocklist head */
 	struct	buflists v_dirtyblkhd;		/* dirty blocklist head */
 	u_int   v_numoutput;			/* num of writes in progress */
diff --git a/sys/uvm/uvm_pdaemon.c b/sys/uvm/uvm_pdaemon.c
index 046368e697a..000ef14d0a2 100644
--- a/sys/uvm/uvm_pdaemon.c
+++ b/sys/uvm/uvm_pdaemon.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: uvm_pdaemon.c,v 1.52 2009/07/22 21:05:37 oga Exp $	*/
+/*	$OpenBSD: uvm_pdaemon.c,v 1.53 2009/08/02 16:28:40 beck Exp $	*/
 /*	$NetBSD: uvm_pdaemon.c,v 1.23 2000/08/20 10:24:14 bjh21 Exp $	*/
 
 /* 
@@ -239,11 +239,13 @@ uvm_pageout(void *arg)
 		    uvmexp.inactarg);
 
 		/*
-		 * scan if needed
+		 * get pages from the buffer cache, or scan if needed
 		 */
-		if ((uvmexp.free - BUFPAGES_DEFICIT) < uvmexp.freetarg ||
-		    uvmexp.inactive < uvmexp.inactarg) {
+		if (uvmexp.inactive < uvmexp.inactarg)
 			uvmpd_scan();
+		else if ((uvmexp.free - BUFPAGES_DEFICIT) < uvmexp.freetarg) {
+			if (bufbackoff() == -1)
+				uvmpd_scan();
 		}
 
 		/*