TRIM support

[dragonfly.git] / sys / kern / vfs_bio.c

/*
 * Copyright (c) 1994,1997 John S. Dyson
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 1. Redistributions of source code must retain the above copyright
 *    notice immediately at the beginning of the file, without modification,
 *    this list of conditions, and the following disclaimer.
 * 2. Absolutely no warranty of function or purpose is made by the author
 *		John S. Dyson.
 *
 * $FreeBSD: src/sys/kern/vfs_bio.c,v 1.242.2.20 2003/05/28 18:38:10 alc Exp $
 * $DragonFly: src/sys/kern/vfs_bio.c,v 1.115 2008/08/13 11:02:31 swildner Exp $
 */

/*
 * this file contains a new buffer I/O scheme implementing a coherent
 * VM object and buffer cache scheme.  Pains have been taken to make
 * sure that the performance degradation associated with schemes such
 * as this is not realized.
 *
 * Author:  John S. Dyson
 * Significant help during the development and debugging phases
 * had been provided by David Greenman, also of the FreeBSD core team.
 *
 * see man buf(9) for more info.
 */

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/buf.h>
#include <sys/conf.h>
#include <sys/devicestat.h>
#include <sys/eventhandler.h>
#include <sys/lock.h>
#include <sys/malloc.h>
#include <sys/mount.h>
#include <sys/kernel.h>
#include <sys/kthread.h>
#include <sys/proc.h>
#include <sys/reboot.h>
#include <sys/resourcevar.h>
#include <sys/sysctl.h>
#include <sys/vmmeter.h>
#include <sys/vnode.h>
#include <sys/dsched.h>
#include <sys/proc.h>
#include <vm/vm.h>
#include <vm/vm_param.h>
#include <vm/vm_kern.h>
#include <vm/vm_pageout.h>
#include <vm/vm_page.h>
#include <vm/vm_object.h>
#include <vm/vm_extern.h>
#include <vm/vm_map.h>
#include <vm/vm_pager.h>
#include <vm/swap_pager.h>

#include <sys/buf2.h>
#include <sys/thread2.h>
#include <sys/spinlock2.h>
#include <sys/mplock2.h>
#include <vm/vm_page2.h>

#include "opt_ddb.h"
#ifdef DDB
#include <ddb/ddb.h>
#endif

/*
 * Buffer queues.
 */
enum bufq_type {
	BQUEUE_NONE,    	/* not on any queue */
	BQUEUE_LOCKED,  	/* locked buffers */
	BQUEUE_CLEAN,   	/* non-B_DELWRI buffers */
	BQUEUE_DIRTY,   	/* B_DELWRI buffers */
	BQUEUE_DIRTY_HW,   	/* B_DELWRI buffers - heavy weight */
	BQUEUE_EMPTYKVA, 	/* empty buffer headers with KVA assignment */
	BQUEUE_EMPTY,    	/* empty buffer headers */

	BUFFER_QUEUES		/* number of buffer queues */
};

typedef enum bufq_type bufq_type_t;

#define BD_WAKE_SIZE	16384
#define BD_WAKE_MASK	(BD_WAKE_SIZE - 1)

TAILQ_HEAD(bqueues, buf) bufqueues[BUFFER_QUEUES];
static struct spinlock bufqspin = SPINLOCK_INITIALIZER(&bufqspin);
static struct spinlock bufcspin = SPINLOCK_INITIALIZER(&bufcspin);

static MALLOC_DEFINE(M_BIOBUF, "BIO buffer", "BIO buffer");

struct buf *buf;		/* buffer header pool */

static void vfs_clean_pages(struct buf *bp);
static void vfs_clean_one_page(struct buf *bp, int pageno, vm_page_t m);
static void vfs_dirty_one_page(struct buf *bp, int pageno, vm_page_t m);
static void vfs_vmio_release(struct buf *bp);
static int flushbufqueues(bufq_type_t q);
static vm_page_t bio_page_alloc(vm_object_t obj, vm_pindex_t pg, int deficit);

static void bd_signal(int totalspace);
static void buf_daemon(void);
static void buf_daemon_hw(void);

/*
 * bogus page -- for I/O to/from partially complete buffers
 * this is a temporary solution to the problem, but it is not
 * really that bad.  it would be better to split the buffer
 * for input in the case of buffers partially already in memory,
 * but the code is intricate enough already.
 */
vm_page_t bogus_page;

/*
 * These are all static, but make the ones we export globals so we do
 * not need to use compiler magic.
 */
int bufspace;			/* locked by buffer_map */
int maxbufspace;
static int bufmallocspace;	/* atomic ops */
int maxbufmallocspace, lobufspace, hibufspace;
static int bufreusecnt, bufdefragcnt, buffreekvacnt;
static int lorunningspace;
static int hirunningspace;
static int runningbufreq;		/* locked by bufcspin */
static int dirtybufspace;		/* locked by bufcspin */
static int dirtybufcount;		/* locked by bufcspin */
static int dirtybufspacehw;		/* locked by bufcspin */
static int dirtybufcounthw;		/* locked by bufcspin */
static int runningbufspace;		/* locked by bufcspin */
static int runningbufcount;		/* locked by bufcspin */
int lodirtybufspace;
int hidirtybufspace;
static int getnewbufcalls;
static int getnewbufrestarts;
static int recoverbufcalls;
static int needsbuffer;		/* locked by bufcspin */
static int bd_request;		/* locked by bufcspin */
static int bd_request_hw;	/* locked by bufcspin */
static u_int bd_wake_ary[BD_WAKE_SIZE];
static u_int bd_wake_index;
static u_int vm_cycle_point = 40; /* 23-36 will migrate more act->inact */
static int debug_commit;

static struct thread *bufdaemon_td;
static struct thread *bufdaemonhw_td;
static u_int lowmempgallocs;
static u_int lowmempgfails;

/*
 * Sysctls for operational control of the buffer cache.
 */
SYSCTL_INT(_vfs, OID_AUTO, lodirtybufspace, CTLFLAG_RW, &lodirtybufspace, 0,
	"Number of dirty buffers to flush before bufdaemon becomes inactive");
SYSCTL_INT(_vfs, OID_AUTO, hidirtybufspace, CTLFLAG_RW, &hidirtybufspace, 0,
	"High watermark used to trigger explicit flushing of dirty buffers");
SYSCTL_INT(_vfs, OID_AUTO, lorunningspace, CTLFLAG_RW, &lorunningspace, 0,
	"Minimum amount of buffer space required for active I/O");
SYSCTL_INT(_vfs, OID_AUTO, hirunningspace, CTLFLAG_RW, &hirunningspace, 0,
	"Maximum amount of buffer space to usable for active I/O");
SYSCTL_UINT(_vfs, OID_AUTO, lowmempgallocs, CTLFLAG_RW, &lowmempgallocs, 0,
	"Page allocations done during periods of very low free memory");
SYSCTL_UINT(_vfs, OID_AUTO, lowmempgfails, CTLFLAG_RW, &lowmempgfails, 0,
	"Page allocations which failed during periods of very low free memory");
SYSCTL_UINT(_vfs, OID_AUTO, vm_cycle_point, CTLFLAG_RW, &vm_cycle_point, 0,
	"Recycle pages to active or inactive queue transition pt 0-64");
/*
 * Sysctls determining current state of the buffer cache.
 */
SYSCTL_INT(_vfs, OID_AUTO, nbuf, CTLFLAG_RD, &nbuf, 0,
	"Total number of buffers in buffer cache");
SYSCTL_INT(_vfs, OID_AUTO, dirtybufspace, CTLFLAG_RD, &dirtybufspace, 0,
	"Pending bytes of dirty buffers (all)");
SYSCTL_INT(_vfs, OID_AUTO, dirtybufspacehw, CTLFLAG_RD, &dirtybufspacehw, 0,
	"Pending bytes of dirty buffers (heavy weight)");
SYSCTL_INT(_vfs, OID_AUTO, dirtybufcount, CTLFLAG_RD, &dirtybufcount, 0,
	"Pending number of dirty buffers");
SYSCTL_INT(_vfs, OID_AUTO, dirtybufcounthw, CTLFLAG_RD, &dirtybufcounthw, 0,
	"Pending number of dirty buffers (heavy weight)");
SYSCTL_INT(_vfs, OID_AUTO, runningbufspace, CTLFLAG_RD, &runningbufspace, 0,
	"I/O bytes currently in progress due to asynchronous writes");
SYSCTL_INT(_vfs, OID_AUTO, runningbufcount, CTLFLAG_RD, &runningbufcount, 0,
	"I/O buffers currently in progress due to asynchronous writes");
SYSCTL_INT(_vfs, OID_AUTO, maxbufspace, CTLFLAG_RD, &maxbufspace, 0,
	"Hard limit on maximum amount of memory usable for buffer space");
SYSCTL_INT(_vfs, OID_AUTO, hibufspace, CTLFLAG_RD, &hibufspace, 0,
	"Soft limit on maximum amount of memory usable for buffer space");
SYSCTL_INT(_vfs, OID_AUTO, lobufspace, CTLFLAG_RD, &lobufspace, 0,
	"Minimum amount of memory to reserve for system buffer space");
SYSCTL_INT(_vfs, OID_AUTO, bufspace, CTLFLAG_RD, &bufspace, 0,
	"Amount of memory available for buffers");
SYSCTL_INT(_vfs, OID_AUTO, maxmallocbufspace, CTLFLAG_RD, &maxbufmallocspace,
	0, "Maximum amount of memory reserved for buffers using malloc");
SYSCTL_INT(_vfs, OID_AUTO, bufmallocspace, CTLFLAG_RD, &bufmallocspace, 0,
	"Amount of memory left for buffers using malloc-scheme");
SYSCTL_INT(_vfs, OID_AUTO, getnewbufcalls, CTLFLAG_RD, &getnewbufcalls, 0,
	"New buffer header acquisition requests");
SYSCTL_INT(_vfs, OID_AUTO, getnewbufrestarts, CTLFLAG_RD, &getnewbufrestarts,
	0, "New buffer header acquisition restarts");
SYSCTL_INT(_vfs, OID_AUTO, recoverbufcalls, CTLFLAG_RD, &recoverbufcalls, 0,
	"Recover VM space in an emergency");
SYSCTL_INT(_vfs, OID_AUTO, bufdefragcnt, CTLFLAG_RD, &bufdefragcnt, 0,
	"Buffer acquisition restarts due to fragmented buffer map");
SYSCTL_INT(_vfs, OID_AUTO, buffreekvacnt, CTLFLAG_RD, &buffreekvacnt, 0,
	"Amount of time KVA space was deallocated in an arbitrary buffer");
SYSCTL_INT(_vfs, OID_AUTO, bufreusecnt, CTLFLAG_RD, &bufreusecnt, 0,
	"Amount of time buffer re-use operations were successful");
SYSCTL_INT(_vfs, OID_AUTO, debug_commit, CTLFLAG_RW, &debug_commit, 0, "");
SYSCTL_INT(_debug_sizeof, OID_AUTO, buf, CTLFLAG_RD, 0, sizeof(struct buf),
	"sizeof(struct buf)");

char *buf_wmesg = BUF_WMESG;

#define VFS_BIO_NEED_ANY	0x01	/* any freeable buffer */
#define VFS_BIO_NEED_UNUSED02	0x02
#define VFS_BIO_NEED_UNUSED04	0x04
#define VFS_BIO_NEED_BUFSPACE	0x08	/* wait for buf space, lo hysteresis */

/*
 * bufspacewakeup:
 *
 *	Called when buffer space is potentially available for recovery.
 *	getnewbuf() will block on this flag when it is unable to free 
 *	sufficient buffer space.  Buffer space becomes recoverable when 
 *	bp's get placed back in the queues.
 */
static __inline void
bufspacewakeup(void)
{
	/*
	 * If someone is waiting for BUF space, wake them up.  Even
	 * though we haven't freed the kva space yet, the waiting
	 * process will be able to now.
	 */
	spin_lock(&bufcspin);
	if (needsbuffer & VFS_BIO_NEED_BUFSPACE) {
		needsbuffer &= ~VFS_BIO_NEED_BUFSPACE;
		spin_unlock(&bufcspin);
		wakeup(&needsbuffer);
	} else {
		spin_unlock(&bufcspin);
	}
}

/*
 * runningbufwakeup:
 *
 *	Accounting for I/O in progress.
 *
 */
static __inline void
runningbufwakeup(struct buf *bp)
{
	int totalspace;
	int limit;

	if ((totalspace = bp->b_runningbufspace) != 0) {
		spin_lock(&bufcspin);
		runningbufspace -= totalspace;
		--runningbufcount;
		bp->b_runningbufspace = 0;

		/*
		 * see waitrunningbufspace() for limit test.
		 */
		limit = hirunningspace * 4 / 6;
		if (runningbufreq && runningbufspace <= limit) {
			runningbufreq = 0;
			spin_unlock(&bufcspin);
			wakeup(&runningbufreq);
		} else {
			spin_unlock(&bufcspin);
		}
		bd_signal(totalspace);
	}
}

/*
 * bufcountwakeup:
 *
 *	Called when a buffer has been added to one of the free queues to
 *	account for the buffer and to wakeup anyone waiting for free buffers.
 *	This typically occurs when large amounts of metadata are being handled
 *	by the buffer cache ( else buffer space runs out first, usually ).
 *
 * MPSAFE
 */
static __inline void
bufcountwakeup(void) 
{
	spin_lock(&bufcspin);
	if (needsbuffer) {
		needsbuffer &= ~VFS_BIO_NEED_ANY;
		spin_unlock(&bufcspin);
		wakeup(&needsbuffer);
	} else {
		spin_unlock(&bufcspin);
	}
}

/*
 * waitrunningbufspace()
 *
 * Wait for the amount of running I/O to drop to hirunningspace * 4 / 6.
 * This is the point where write bursting stops so we don't want to wait
 * for the running amount to drop below it (at least if we still want bioq
 * to burst writes).
 *
 * The caller may be using this function to block in a tight loop, we
 * must block while runningbufspace is greater then or equal to
 * hirunningspace * 4 / 6.
 *
 * And even with that it may not be enough, due to the presence of
 * B_LOCKED dirty buffers, so also wait for at least one running buffer
 * to complete.
 */
void
waitrunningbufspace(void)
{
	int limit = hirunningspace * 4 / 6;
	int dummy;

	spin_lock(&bufcspin);
	if (runningbufspace > limit) {
		while (runningbufspace > limit) {
			++runningbufreq;
			ssleep(&runningbufreq, &bufcspin, 0, "wdrn1", 0);
		}
		spin_unlock(&bufcspin);
	} else if (runningbufspace > limit / 2) {
		++runningbufreq;
		spin_unlock(&bufcspin);
		tsleep(&dummy, 0, "wdrn2", 1);
	} else {
		spin_unlock(&bufcspin);
	}
}

/*
 * buf_dirty_count_severe:
 *
 *	Return true if we have too many dirty buffers.
 */
int
buf_dirty_count_severe(void)
{
	return (runningbufspace + dirtybufspace >= hidirtybufspace ||
	        dirtybufcount >= nbuf / 2);
}

/*
 * Return true if the amount of running I/O is severe and BIOQ should
 * start bursting.
 */
int
buf_runningbufspace_severe(void)
{
	return (runningbufspace >= hirunningspace * 4 / 6);
}

/*
 * vfs_buf_test_cache:
 *
 * Called when a buffer is extended.  This function clears the B_CACHE
 * bit if the newly extended portion of the buffer does not contain
 * valid data.
 *
 * NOTE! Dirty VM pages are not processed into dirty (B_DELWRI) buffer
 * cache buffers.  The VM pages remain dirty, as someone had mmap()'d
 * them while a clean buffer was present.
 */
static __inline__
void
vfs_buf_test_cache(struct buf *bp,
		  vm_ooffset_t foff, vm_offset_t off, vm_offset_t size,
		  vm_page_t m)
{
	if (bp->b_flags & B_CACHE) {
		int base = (foff + off) & PAGE_MASK;
		if (vm_page_is_valid(m, base, size) == 0)
			bp->b_flags &= ~B_CACHE;
	}
}

/*
 * bd_speedup()
 *
 * Spank the buf_daemon[_hw] if the total dirty buffer space exceeds the
 * low water mark.
 *
 * MPSAFE
 */
static __inline__
void
bd_speedup(void)
{
	if (dirtybufspace < lodirtybufspace && dirtybufcount < nbuf / 2)
		return;

	if (bd_request == 0 &&
	    (dirtybufspace - dirtybufspacehw > lodirtybufspace / 2 ||
	     dirtybufcount - dirtybufcounthw >= nbuf / 2)) {
		spin_lock(&bufcspin);
		bd_request = 1;
		spin_unlock(&bufcspin);
		wakeup(&bd_request);
	}
	if (bd_request_hw == 0 &&
	    (dirtybufspacehw > lodirtybufspace / 2 ||
	     dirtybufcounthw >= nbuf / 2)) {
		spin_lock(&bufcspin);
		bd_request_hw = 1;
		spin_unlock(&bufcspin);
		wakeup(&bd_request_hw);
	}
}

/*
 * bd_heatup()
 *
 *	Get the buf_daemon heated up when the number of running and dirty
 *	buffers exceeds the mid-point.
 *
 *	Return the total number of dirty bytes past the second mid point
 *	as a measure of how much excess dirty data there is in the system.
 *
 * MPSAFE
 */
int
bd_heatup(void)
{
	int mid1;
	int mid2;
	int totalspace;

	mid1 = lodirtybufspace + (hidirtybufspace - lodirtybufspace) / 2;

	totalspace = runningbufspace + dirtybufspace;
	if (totalspace >= mid1 || dirtybufcount >= nbuf / 2) {
		bd_speedup();
		mid2 = mid1 + (hidirtybufspace - mid1) / 2;
		if (totalspace >= mid2)
			return(totalspace - mid2);
	}
	return(0);
}

/*
 * bd_wait()
 *
 *	Wait for the buffer cache to flush (totalspace) bytes worth of
 *	buffers, then return.
 *
 *	Regardless this function blocks while the number of dirty buffers
 *	exceeds hidirtybufspace.
 *
 * MPSAFE
 */
void
bd_wait(int totalspace)
{
	u_int i;
	int count;

	if (curthread == bufdaemonhw_td || curthread == bufdaemon_td)
		return;

	while (totalspace > 0) {
		bd_heatup();
		if (totalspace > runningbufspace + dirtybufspace)
			totalspace = runningbufspace + dirtybufspace;
		count = totalspace / BKVASIZE;
		if (count >= BD_WAKE_SIZE)
			count = BD_WAKE_SIZE - 1;

		spin_lock(&bufcspin);
		i = (bd_wake_index + count) & BD_WAKE_MASK;
		++bd_wake_ary[i];

		/*
		 * This is not a strict interlock, so we play a bit loose
		 * with locking access to dirtybufspace*
		 */
		tsleep_interlock(&bd_wake_ary[i], 0);
		spin_unlock(&bufcspin);
		tsleep(&bd_wake_ary[i], PINTERLOCKED, "flstik", hz);

		totalspace = runningbufspace + dirtybufspace - hidirtybufspace;
	}
}

/*
 * bd_signal()
 * 
 *	This function is called whenever runningbufspace or dirtybufspace
 *	is reduced.  Track threads waiting for run+dirty buffer I/O
 *	complete.
 *
 * MPSAFE
 */
static void
bd_signal(int totalspace)
{
	u_int i;

	if (totalspace > 0) {
		if (totalspace > BKVASIZE * BD_WAKE_SIZE)
			totalspace = BKVASIZE * BD_WAKE_SIZE;
		spin_lock(&bufcspin);
		while (totalspace > 0) {
			i = bd_wake_index++;
			i &= BD_WAKE_MASK;
			if (bd_wake_ary[i]) {
				bd_wake_ary[i] = 0;
				spin_unlock(&bufcspin);
				wakeup(&bd_wake_ary[i]);
				spin_lock(&bufcspin);
			}
			totalspace -= BKVASIZE;
		}
		spin_unlock(&bufcspin);
	}
}

/*
 * BIO tracking support routines.
 *
 * Release a ref on a bio_track.  Wakeup requests are atomically released
 * along with the last reference so bk_active will never wind up set to
 * only 0x80000000.
 *
 * MPSAFE
 */
static
void
bio_track_rel(struct bio_track *track)
{
	int	active;
	int	desired;

	/*
	 * Shortcut
	 */
	active = track->bk_active;
	if (active == 1 && atomic_cmpset_int(&track->bk_active, 1, 0))
		return;

	/*
	 * Full-on.  Note that the wait flag is only atomically released on
	 * the 1->0 count transition.
	 *
	 * We check for a negative count transition using bit 30 since bit 31
	 * has a different meaning.
	 */
	for (;;) {
		desired = (active & 0x7FFFFFFF) - 1;
		if (desired)
			desired |= active & 0x80000000;
		if (atomic_cmpset_int(&track->bk_active, active, desired)) {
			if (desired & 0x40000000)
				panic("bio_track_rel: bad count: %p\n", track);
			if (active & 0x80000000)
				wakeup(track);
			break;
		}
		active = track->bk_active;
	}
}

/*
 * Wait for the tracking count to reach 0.
 *
 * Use atomic ops such that the wait flag is only set atomically when
 * bk_active is non-zero.
 *
 * MPSAFE
 */
int
bio_track_wait(struct bio_track *track, int slp_flags, int slp_timo)
{
	int	active;
	int	desired;
	int	error;

	/*
	 * Shortcut
	 */
	if (track->bk_active == 0)
		return(0);

	/*
	 * Full-on.  Note that the wait flag may only be atomically set if
	 * the active count is non-zero.
	 *
	 * NOTE: We cannot optimize active == desired since a wakeup could
	 *	 clear active prior to our tsleep_interlock().
	 */
	error = 0;
	while ((active = track->bk_active) != 0) {
		cpu_ccfence();
		desired = active | 0x80000000;
		tsleep_interlock(track, slp_flags);
		if (atomic_cmpset_int(&track->bk_active, active, desired)) {
			error = tsleep(track, slp_flags | PINTERLOCKED,
				       "trwait", slp_timo);
			if (error)
				break;
		}
	}
	return (error);
}

/*
 * bufinit:
 *
 *	Load time initialisation of the buffer cache, called from machine
 *	dependant initialization code. 
 */
void
bufinit(void)
{
	struct buf *bp;
	vm_offset_t bogus_offset;
	int i;

	/* next, make a null set of free lists */
	for (i = 0; i < BUFFER_QUEUES; i++)
		TAILQ_INIT(&bufqueues[i]);

	/* finally, initialize each buffer header and stick on empty q */
	for (i = 0; i < nbuf; i++) {
		bp = &buf[i];
		bzero(bp, sizeof *bp);
		bp->b_flags = B_INVAL;	/* we're just an empty header */
		bp->b_cmd = BUF_CMD_DONE;
		bp->b_qindex = BQUEUE_EMPTY;
		initbufbio(bp);
		xio_init(&bp->b_xio);
		buf_dep_init(bp);
		TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_EMPTY], bp, b_freelist);
	}

	/*
	 * maxbufspace is the absolute maximum amount of buffer space we are 
	 * allowed to reserve in KVM and in real terms.  The absolute maximum
	 * is nominally used by buf_daemon.  hibufspace is the nominal maximum
	 * used by most other processes.  The differential is required to 
	 * ensure that buf_daemon is able to run when other processes might 
	 * be blocked waiting for buffer space.
	 *
	 * maxbufspace is based on BKVASIZE.  Allocating buffers larger then
	 * this may result in KVM fragmentation which is not handled optimally
	 * by the system.
	 */
	maxbufspace = nbuf * BKVASIZE;
	hibufspace = imax(3 * maxbufspace / 4, maxbufspace - MAXBSIZE * 10);
	lobufspace = hibufspace - MAXBSIZE;

	lorunningspace = 512 * 1024;
	/* hirunningspace -- see below */

	/*
	 * Limit the amount of malloc memory since it is wired permanently
	 * into the kernel space.  Even though this is accounted for in
	 * the buffer allocation, we don't want the malloced region to grow
	 * uncontrolled.  The malloc scheme improves memory utilization
	 * significantly on average (small) directories.
	 */
	maxbufmallocspace = hibufspace / 20;

	/*
	 * Reduce the chance of a deadlock occuring by limiting the number
	 * of delayed-write dirty buffers we allow to stack up.
	 *
	 * We don't want too much actually queued to the device at once
	 * (XXX this needs to be per-mount!), because the buffers will
	 * wind up locked for a very long period of time while the I/O
	 * drains.
	 */
	hidirtybufspace = hibufspace / 2;	/* dirty + running */
	hirunningspace = hibufspace / 16;	/* locked & queued to device */
	if (hirunningspace < 1024 * 1024)
		hirunningspace = 1024 * 1024;

	dirtybufspace = 0;
	dirtybufspacehw = 0;

	lodirtybufspace = hidirtybufspace / 2;

	/*
	 * Maximum number of async ops initiated per buf_daemon loop.  This is
	 * somewhat of a hack at the moment, we really need to limit ourselves
	 * based on the number of bytes of I/O in-transit that were initiated
	 * from buf_daemon.
	 */

	bogus_offset = kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
	bogus_page = vm_page_alloc(&kernel_object,
				   (bogus_offset >> PAGE_SHIFT),
				   VM_ALLOC_NORMAL);
	vmstats.v_wire_count++;

}

/*
 * Initialize the embedded bio structures, typically used by
 * deprecated code which tries to allocate its own struct bufs.
 */
void
initbufbio(struct buf *bp)
{
	bp->b_bio1.bio_buf = bp;
	bp->b_bio1.bio_prev = NULL;
	bp->b_bio1.bio_offset = NOOFFSET;
	bp->b_bio1.bio_next = &bp->b_bio2;
	bp->b_bio1.bio_done = NULL;
	bp->b_bio1.bio_flags = 0;

	bp->b_bio2.bio_buf = bp;
	bp->b_bio2.bio_prev = &bp->b_bio1;
	bp->b_bio2.bio_offset = NOOFFSET;
	bp->b_bio2.bio_next = NULL;
	bp->b_bio2.bio_done = NULL;
	bp->b_bio2.bio_flags = 0;

	BUF_LOCKINIT(bp);
}

/*
 * Reinitialize the embedded bio structures as well as any additional
 * translation cache layers.
 */
void
reinitbufbio(struct buf *bp)
{
	struct bio *bio;

	for (bio = &bp->b_bio1; bio; bio = bio->bio_next) {
		bio->bio_done = NULL;
		bio->bio_offset = NOOFFSET;
	}
}

/*
 * Undo the effects of an initbufbio().
 */
void
uninitbufbio(struct buf *bp)
{
	dsched_exit_buf(bp);
	BUF_LOCKFREE(bp);
}

/*
 * Push another BIO layer onto an existing BIO and return it.  The new
 * BIO layer may already exist, holding cached translation data.
 */
struct bio *
push_bio(struct bio *bio)
{
	struct bio *nbio;

	if ((nbio = bio->bio_next) == NULL) {
		int index = bio - &bio->bio_buf->b_bio_array[0];
		if (index >= NBUF_BIO - 1) {
			panic("push_bio: too many layers bp %p\n",
				bio->bio_buf);
		}
		nbio = &bio->bio_buf->b_bio_array[index + 1];
		bio->bio_next = nbio;
		nbio->bio_prev = bio;
		nbio->bio_buf = bio->bio_buf;
		nbio->bio_offset = NOOFFSET;
		nbio->bio_done = NULL;
		nbio->bio_next = NULL;
	}
	KKASSERT(nbio->bio_done == NULL);
	return(nbio);
}

/*
 * Pop a BIO translation layer, returning the previous layer.  The
 * must have been previously pushed.
 */
struct bio *
pop_bio(struct bio *bio)
{
	return(bio->bio_prev);
}

void
clearbiocache(struct bio *bio)
{
	while (bio) {
		bio->bio_offset = NOOFFSET;
		bio = bio->bio_next;
	}
}

/*
 * bfreekva:
 *
 *	Free the KVA allocation for buffer 'bp'.
 *
 *	Must be called from a critical section as this is the only locking for
 *	buffer_map.
 *
 *	Since this call frees up buffer space, we call bufspacewakeup().
 *
 * MPALMOSTSAFE
 */
static void
bfreekva(struct buf *bp)
{
	int count;

	if (bp->b_kvasize) {
		++buffreekvacnt;
		count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
		vm_map_lock(&buffer_map);
		bufspace -= bp->b_kvasize;
		vm_map_delete(&buffer_map,
		    (vm_offset_t) bp->b_kvabase,
		    (vm_offset_t) bp->b_kvabase + bp->b_kvasize,
		    &count
		);
		vm_map_unlock(&buffer_map);
		vm_map_entry_release(count);
		bp->b_kvasize = 0;
		bp->b_kvabase = NULL;
		bufspacewakeup();
	}
}

/*
 * bremfree:
 *
 *	Remove the buffer from the appropriate free list.
 */
static __inline void
_bremfree(struct buf *bp)
{
	if (bp->b_qindex != BQUEUE_NONE) {
		KASSERT(BUF_REFCNTNB(bp) == 1, 
				("bremfree: bp %p not locked",bp));
		TAILQ_REMOVE(&bufqueues[bp->b_qindex], bp, b_freelist);
		bp->b_qindex = BQUEUE_NONE;
	} else {
		if (BUF_REFCNTNB(bp) <= 1)
			panic("bremfree: removing a buffer not on a queue");
	}
}

void
bremfree(struct buf *bp)
{
	spin_lock(&bufqspin);
	_bremfree(bp);
	spin_unlock(&bufqspin);
}

static void
bremfree_locked(struct buf *bp)
{
	_bremfree(bp);
}

/*
 * bread:
 *
 *	Get a buffer with the specified data.  Look in the cache first.  We
 *	must clear B_ERROR and B_INVAL prior to initiating I/O.  If B_CACHE
 *	is set, the buffer is valid and we do not have to do anything ( see
 *	getblk() ).
 *
 */
int
bread(struct vnode *vp, off_t loffset, int size, struct buf **bpp)
{
	return (breadn(vp, loffset, size, NULL, NULL, 0, bpp));
}

/*
 * This version of bread issues any required I/O asyncnronously and
 * makes a callback on completion.
 *
 * The callback must check whether BIO_DONE is set in the bio and issue
 * the bpdone(bp, 0) if it isn't.  The callback is responsible for clearing
 * BIO_DONE and disposing of the I/O (bqrelse()ing it).
 */
void
breadcb(struct vnode *vp, off_t loffset, int size,
	void (*func)(struct bio *), void *arg)
{
	struct buf *bp;

	bp = getblk(vp, loffset, size, 0, 0);

	/* if not found in cache, do some I/O */
	if ((bp->b_flags & B_CACHE) == 0) {
		bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
		bp->b_cmd = BUF_CMD_READ;
		bp->b_bio1.bio_done = func;
		bp->b_bio1.bio_caller_info1.ptr = arg;
		vfs_busy_pages(vp, bp);
		BUF_KERNPROC(bp);
		vn_strategy(vp, &bp->b_bio1);
	} else if (func) {
		/*
		 * Since we are issuing the callback synchronously it cannot
		 * race the BIO_DONE, so no need for atomic ops here.
		 */
		/*bp->b_bio1.bio_done = func;*/
		bp->b_bio1.bio_caller_info1.ptr = arg;
		bp->b_bio1.bio_flags |= BIO_DONE;
		func(&bp->b_bio1);
	} else {
		bqrelse(bp);
	}
}

/*
 * breadn:
 *
 *	Operates like bread, but also starts asynchronous I/O on
 *	read-ahead blocks.  We must clear B_ERROR and B_INVAL prior
 *	to initiating I/O . If B_CACHE is set, the buffer is valid 
 *	and we do not have to do anything.
 *
 */
int
breadn(struct vnode *vp, off_t loffset, int size, off_t *raoffset,
	int *rabsize, int cnt, struct buf **bpp)
{
	struct buf *bp, *rabp;
	int i;
	int rv = 0, readwait = 0;

	*bpp = bp = getblk(vp, loffset, size, 0, 0);

	/* if not found in cache, do some I/O */
	if ((bp->b_flags & B_CACHE) == 0) {
		bp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
		bp->b_cmd = BUF_CMD_READ;
		bp->b_bio1.bio_done = biodone_sync;
		bp->b_bio1.bio_flags |= BIO_SYNC;
		vfs_busy_pages(vp, bp);
		vn_strategy(vp, &bp->b_bio1);
		++readwait;
	}

	for (i = 0; i < cnt; i++, raoffset++, rabsize++) {
		if (inmem(vp, *raoffset))
			continue;
		rabp = getblk(vp, *raoffset, *rabsize, 0, 0);

		if ((rabp->b_flags & B_CACHE) == 0) {
			rabp->b_flags &= ~(B_ERROR | B_EINTR | B_INVAL);
			rabp->b_cmd = BUF_CMD_READ;
			vfs_busy_pages(vp, rabp);
			BUF_KERNPROC(rabp);
			vn_strategy(vp, &rabp->b_bio1);
		} else {
			brelse(rabp);
		}
	}
	if (readwait)
		rv = biowait(&bp->b_bio1, "biord");
	return (rv);
}

/*
 * bwrite:
 *
 *	Synchronous write, waits for completion.
 *
 *	Write, release buffer on completion.  (Done by iodone
 *	if async).  Do not bother writing anything if the buffer
 *	is invalid.
 *
 *	Note that we set B_CACHE here, indicating that buffer is
 *	fully valid and thus cacheable.  This is true even of NFS
 *	now so we set it generally.  This could be set either here 
 *	or in biodone() since the I/O is synchronous.  We put it
 *	here.
 */
int
bwrite(struct buf *bp)
{
	int error;

	if (bp->b_flags & B_INVAL) {
		brelse(bp);
		return (0);
	}
	if (BUF_REFCNTNB(bp) == 0)
		panic("bwrite: buffer is not busy???");

	/* Mark the buffer clean */
	bundirty(bp);

	bp->b_flags &= ~(B_ERROR | B_EINTR);
	bp->b_flags |= B_CACHE;
	bp->b_cmd = BUF_CMD_WRITE;
	bp->b_bio1.bio_done = biodone_sync;
	bp->b_bio1.bio_flags |= BIO_SYNC;
	vfs_busy_pages(bp->b_vp, bp);

	/*
	 * Normal bwrites pipeline writes.  NOTE: b_bufsize is only
	 * valid for vnode-backed buffers.
	 */
	bsetrunningbufspace(bp, bp->b_bufsize);
	vn_strategy(bp->b_vp, &bp->b_bio1);
	error = biowait(&bp->b_bio1, "biows");
	brelse(bp);

	return (error);
}

/*
 * bawrite:
 *
 *	Asynchronous write.  Start output on a buffer, but do not wait for
 *	it to complete.  The buffer is released when the output completes.
 *
 *	bwrite() ( or the VOP routine anyway ) is responsible for handling
 *	B_INVAL buffers.  Not us.
 */
void
bawrite(struct buf *bp)
{
	if (bp->b_flags & B_INVAL) {
		brelse(bp);
		return;
	}
	if (BUF_REFCNTNB(bp) == 0)
		panic("bwrite: buffer is not busy???");

	/* Mark the buffer clean */
	bundirty(bp);

	bp->b_flags &= ~(B_ERROR | B_EINTR);
	bp->b_flags |= B_CACHE;
	bp->b_cmd = BUF_CMD_WRITE;
	KKASSERT(bp->b_bio1.bio_done == NULL);
	vfs_busy_pages(bp->b_vp, bp);

	/*
	 * Normal bwrites pipeline writes.  NOTE: b_bufsize is only
	 * valid for vnode-backed buffers.
	 */
	bsetrunningbufspace(bp, bp->b_bufsize);
	BUF_KERNPROC(bp);
	vn_strategy(bp->b_vp, &bp->b_bio1);
}

/*
 * bowrite:
 *
 *	Ordered write.  Start output on a buffer, and flag it so that the
 *	device will write it in the order it was queued.  The buffer is
 *	released when the output completes.  bwrite() ( or the VOP routine
 *	anyway ) is responsible for handling B_INVAL buffers.
 */
int
bowrite(struct buf *bp)
{
	bp->b_flags |= B_ORDERED;
	bawrite(bp);
	return (0);
}

/*
 * bdwrite:
 *
 *	Delayed write. (Buffer is marked dirty).  Do not bother writing
 *	anything if the buffer is marked invalid.
 *
 *	Note that since the buffer must be completely valid, we can safely
 *	set B_CACHE.  In fact, we have to set B_CACHE here rather then in
 *	biodone() in order to prevent getblk from writing the buffer
 *	out synchronously.
 */
void
bdwrite(struct buf *bp)
{
	if (BUF_REFCNTNB(bp) == 0)
		panic("bdwrite: buffer is not busy");

	if (bp->b_flags & B_INVAL) {
		brelse(bp);
		return;
	}
	bdirty(bp);

	if (dsched_is_clear_buf_priv(bp))
		dsched_new_buf(bp);

	/*
	 * Set B_CACHE, indicating that the buffer is fully valid.  This is
	 * true even of NFS now.
	 */
	bp->b_flags |= B_CACHE;

	/*
	 * This bmap keeps the system from needing to do the bmap later,
	 * perhaps when the system is attempting to do a sync.  Since it
	 * is likely that the indirect block -- or whatever other datastructure
	 * that the filesystem needs is still in memory now, it is a good
	 * thing to do this.  Note also, that if the pageout daemon is
	 * requesting a sync -- there might not be enough memory to do
	 * the bmap then...  So, this is important to do.
	 */
	if (bp->b_bio2.bio_offset == NOOFFSET) {
		VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
			 NULL, NULL, BUF_CMD_WRITE);
	}

	/*
	 * Because the underlying pages may still be mapped and
	 * writable trying to set the dirty buffer (b_dirtyoff/end)
	 * range here will be inaccurate.
	 *
	 * However, we must still clean the pages to satisfy the
	 * vnode_pager and pageout daemon, so theythink the pages
	 * have been "cleaned".  What has really occured is that
	 * they've been earmarked for later writing by the buffer
	 * cache.
	 *
	 * So we get the b_dirtyoff/end update but will not actually
	 * depend on it (NFS that is) until the pages are busied for
	 * writing later on.
	 */
	vfs_clean_pages(bp);
	bqrelse(bp);

	/*
	 * note: we cannot initiate I/O from a bdwrite even if we wanted to,
	 * due to the softdep code.
	 */
}

/*
 * Fake write - return pages to VM system as dirty, leave the buffer clean.
 * This is used by tmpfs.
 *
 * It is important for any VFS using this routine to NOT use it for
 * IO_SYNC or IO_ASYNC operations which occur when the system really
 * wants to flush VM pages to backing store.
 */
void
buwrite(struct buf *bp)
{
	vm_page_t m;
	int i;

	/*
	 * Only works for VMIO buffers.  If the buffer is already
	 * marked for delayed-write we can't avoid the bdwrite().
	 */
	if ((bp->b_flags & B_VMIO) == 0 || (bp->b_flags & B_DELWRI)) {
		bdwrite(bp);
		return;
	}

	/*
	 * Set valid & dirty.
	 *
	 * WARNING! vfs_dirty_one_page() assumes vm_token is held for now.
	 */
	lwkt_gettoken(&vm_token);
	for (i = 0; i < bp->b_xio.xio_npages; i++) {
		m = bp->b_xio.xio_pages[i];
		vfs_dirty_one_page(bp, i, m);
	}
	lwkt_reltoken(&vm_token);
	bqrelse(bp);
}

/*
 * bdirty:
 *
 *	Turn buffer into delayed write request by marking it B_DELWRI.
 *	B_RELBUF and B_NOCACHE must be cleared.
 *
 *	We reassign the buffer to itself to properly update it in the
 *	dirty/clean lists. 
 *
 *	Must be called from a critical section.
 *	The buffer must be on BQUEUE_NONE.
 */
void
bdirty(struct buf *bp)
{
	KASSERT(bp->b_qindex == BQUEUE_NONE, ("bdirty: buffer %p still on queue %d", bp, bp->b_qindex));
	if (bp->b_flags & B_NOCACHE) {
		kprintf("bdirty: clearing B_NOCACHE on buf %p\n", bp);
		bp->b_flags &= ~B_NOCACHE;
	}
	if (bp->b_flags & B_INVAL) {
		kprintf("bdirty: warning, dirtying invalid buffer %p\n", bp);
	}
	bp->b_flags &= ~B_RELBUF;

	if ((bp->b_flags & B_DELWRI) == 0) {
		lwkt_gettoken(&bp->b_vp->v_token);
		bp->b_flags |= B_DELWRI;
		reassignbuf(bp);
		lwkt_reltoken(&bp->b_vp->v_token);

		spin_lock(&bufcspin);
		++dirtybufcount;
		dirtybufspace += bp->b_bufsize;
		if (bp->b_flags & B_HEAVY) {
			++dirtybufcounthw;
			dirtybufspacehw += bp->b_bufsize;
		}
		spin_unlock(&bufcspin);

		bd_heatup();
	}
}

/*
 * Set B_HEAVY, indicating that this is a heavy-weight buffer that
 * needs to be flushed with a different buf_daemon thread to avoid
 * deadlocks.  B_HEAVY also imposes restrictions in getnewbuf().
 */
void
bheavy(struct buf *bp)
{
	if ((bp->b_flags & B_HEAVY) == 0) {
		bp->b_flags |= B_HEAVY;
		if (bp->b_flags & B_DELWRI) {
			spin_lock(&bufcspin);
			++dirtybufcounthw;
			dirtybufspacehw += bp->b_bufsize;
			spin_unlock(&bufcspin);
		}
	}
}

/*
 * bundirty:
 *
 *	Clear B_DELWRI for buffer.
 *
 *	Must be called from a critical section.
 *
 *	The buffer is typically on BQUEUE_NONE but there is one case in 
 *	brelse() that calls this function after placing the buffer on
 *	a different queue.
 *
 * MPSAFE
 */
void
bundirty(struct buf *bp)
{
	if (bp->b_flags & B_DELWRI) {
		lwkt_gettoken(&bp->b_vp->v_token);
		bp->b_flags &= ~B_DELWRI;
		reassignbuf(bp);
		lwkt_reltoken(&bp->b_vp->v_token);

		spin_lock(&bufcspin);
		--dirtybufcount;
		dirtybufspace -= bp->b_bufsize;
		if (bp->b_flags & B_HEAVY) {
			--dirtybufcounthw;
			dirtybufspacehw -= bp->b_bufsize;
		}
		spin_unlock(&bufcspin);

		bd_signal(bp->b_bufsize);
	}
	/*
	 * Since it is now being written, we can clear its deferred write flag.
	 */
	bp->b_flags &= ~B_DEFERRED;
}

/*
 * Set the b_runningbufspace field, used to track how much I/O is
 * in progress at any given moment.
 */
void
bsetrunningbufspace(struct buf *bp, int bytes)
{
	bp->b_runningbufspace = bytes;
	if (bytes) {
		spin_lock(&bufcspin);
		runningbufspace += bytes;
		++runningbufcount;
		spin_unlock(&bufcspin);
	}
}

/*
 * brelse:
 *
 *	Release a busy buffer and, if requested, free its resources.  The
 *	buffer will be stashed in the appropriate bufqueue[] allowing it
 *	to be accessed later as a cache entity or reused for other purposes.
 *
 * MPALMOSTSAFE
 */
void
brelse(struct buf *bp)
{
#ifdef INVARIANTS
	int saved_flags = bp->b_flags;
#endif

	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("brelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));

	/*
	 * If B_NOCACHE is set we are being asked to destroy the buffer and
	 * its backing store.  Clear B_DELWRI.
	 *
	 * B_NOCACHE is set in two cases: (1) when the caller really wants
	 * to destroy the buffer and backing store and (2) when the caller
	 * wants to destroy the buffer and backing store after a write 
	 * completes.
	 */
	if ((bp->b_flags & (B_NOCACHE|B_DELWRI)) == (B_NOCACHE|B_DELWRI)) {
		bundirty(bp);
	}

	if ((bp->b_flags & (B_INVAL | B_DELWRI)) == B_DELWRI) {
		/*
		 * A re-dirtied buffer is only subject to destruction
		 * by B_INVAL.  B_ERROR and B_NOCACHE are ignored.
		 */
		/* leave buffer intact */
	} else if ((bp->b_flags & (B_NOCACHE | B_INVAL | B_ERROR)) ||
		   (bp->b_bufsize <= 0)) {
		/*
		 * Either a failed read or we were asked to free or not
		 * cache the buffer.  This path is reached with B_DELWRI
		 * set only if B_INVAL is already set.  B_NOCACHE governs
		 * backing store destruction.
		 *
		 * NOTE: HAMMER will set B_LOCKED in buf_deallocate if the
		 * buffer cannot be immediately freed.
		 */
		bp->b_flags |= B_INVAL;
		if (LIST_FIRST(&bp->b_dep) != NULL)
			buf_deallocate(bp);
		if (bp->b_flags & B_DELWRI) {
			spin_lock(&bufcspin);
			--dirtybufcount;
			dirtybufspace -= bp->b_bufsize;
			if (bp->b_flags & B_HEAVY) {
				--dirtybufcounthw;
				dirtybufspacehw -= bp->b_bufsize;
			}
			spin_unlock(&bufcspin);

			bd_signal(bp->b_bufsize);
		}
		bp->b_flags &= ~(B_DELWRI | B_CACHE);
	}

	/*
	 * We must clear B_RELBUF if B_DELWRI or B_LOCKED is set,
	 * or if b_refs is non-zero.
	 *
	 * If vfs_vmio_release() is called with either bit set, the
	 * underlying pages may wind up getting freed causing a previous
	 * write (bdwrite()) to get 'lost' because pages associated with
	 * a B_DELWRI bp are marked clean.  Pages associated with a
	 * B_LOCKED buffer may be mapped by the filesystem.
	 *
	 * If we want to release the buffer ourselves (rather then the
	 * originator asking us to release it), give the originator a
	 * chance to countermand the release by setting B_LOCKED.
	 * 
	 * We still allow the B_INVAL case to call vfs_vmio_release(), even
	 * if B_DELWRI is set.
	 *
	 * If B_DELWRI is not set we may have to set B_RELBUF if we are low
	 * on pages to return pages to the VM page queues.
	 */
	if ((bp->b_flags & (B_DELWRI | B_LOCKED)) || bp->b_refs) {
		bp->b_flags &= ~B_RELBUF;
	} else if (vm_page_count_severe()) {
		if (LIST_FIRST(&bp->b_dep) != NULL)
			buf_deallocate(bp);		/* can set B_LOCKED */
		if (bp->b_flags & (B_DELWRI | B_LOCKED))
			bp->b_flags &= ~B_RELBUF;
		else
			bp->b_flags |= B_RELBUF;
	}

	/*
	 * Make sure b_cmd is clear.  It may have already been cleared by
	 * biodone().
	 *
	 * At this point destroying the buffer is governed by the B_INVAL 
	 * or B_RELBUF flags.
	 */
	bp->b_cmd = BUF_CMD_DONE;
	dsched_exit_buf(bp);

	/*
	 * VMIO buffer rundown.  Make sure the VM page array is restored
	 * after an I/O may have replaces some of the pages with bogus pages
	 * in order to not destroy dirty pages in a fill-in read.
	 *
	 * Note that due to the code above, if a buffer is marked B_DELWRI
	 * then the B_RELBUF and B_NOCACHE bits will always be clear.
	 * B_INVAL may still be set, however.
	 *
	 * For clean buffers, B_INVAL or B_RELBUF will destroy the buffer
	 * but not the backing store.   B_NOCACHE will destroy the backing
	 * store.
	 *
	 * Note that dirty NFS buffers contain byte-granular write ranges
	 * and should not be destroyed w/ B_INVAL even if the backing store
	 * is left intact.
	 */
	if (bp->b_flags & B_VMIO) {
		/*
		 * Rundown for VMIO buffers which are not dirty NFS buffers.
		 */
		int i, j, resid;
		vm_page_t m;
		off_t foff;
		vm_pindex_t poff;
		vm_object_t obj;
		struct vnode *vp;

		vp = bp->b_vp;

		/*
		 * Get the base offset and length of the buffer.  Note that 
		 * in the VMIO case if the buffer block size is not
		 * page-aligned then b_data pointer may not be page-aligned.
		 * But our b_xio.xio_pages array *IS* page aligned.
		 *
		 * block sizes less then DEV_BSIZE (usually 512) are not 
		 * supported due to the page granularity bits (m->valid,
		 * m->dirty, etc...). 
		 *
		 * See man buf(9) for more information
		 */

		resid = bp->b_bufsize;
		foff = bp->b_loffset;

		lwkt_gettoken(&vm_token);
		for (i = 0; i < bp->b_xio.xio_npages; i++) {
			m = bp->b_xio.xio_pages[i];
			vm_page_flag_clear(m, PG_ZERO);
			/*
			 * If we hit a bogus page, fixup *all* of them
			 * now.  Note that we left these pages wired
			 * when we removed them so they had better exist,
			 * and they cannot be ripped out from under us so
			 * no critical section protection is necessary.
			 */
			if (m == bogus_page) {
				obj = vp->v_object;
				poff = OFF_TO_IDX(bp->b_loffset);

				for (j = i; j < bp->b_xio.xio_npages; j++) {
					vm_page_t mtmp;

					mtmp = bp->b_xio.xio_pages[j];
					if (mtmp == bogus_page) {
						mtmp = vm_page_lookup(obj, poff + j);
						if (!mtmp) {
							panic("brelse: page missing");
						}
						bp->b_xio.xio_pages[j] = mtmp;
					}
				}
				bp->b_flags &= ~B_HASBOGUS;

				if ((bp->b_flags & B_INVAL) == 0) {
					pmap_qenter(trunc_page((vm_offset_t)bp->b_data),
						bp->b_xio.xio_pages, bp->b_xio.xio_npages);
				}
				m = bp->b_xio.xio_pages[i];
			}

			/*
			 * Invalidate the backing store if B_NOCACHE is set
			 * (e.g. used with vinvalbuf()).  If this is NFS
			 * we impose a requirement that the block size be
			 * a multiple of PAGE_SIZE and create a temporary
			 * hack to basically invalidate the whole page.  The
			 * problem is that NFS uses really odd buffer sizes
			 * especially when tracking piecemeal writes and
			 * it also vinvalbuf()'s a lot, which would result
			 * in only partial page validation and invalidation
			 * here.  If the file page is mmap()'d, however,
			 * all the valid bits get set so after we invalidate
			 * here we would end up with weird m->valid values
			 * like 0xfc.  nfs_getpages() can't handle this so
			 * we clear all the valid bits for the NFS case
			 * instead of just some of them.
			 *
			 * The real bug is the VM system having to set m->valid
			 * to VM_PAGE_BITS_ALL for faulted-in pages, which
			 * itself is an artifact of the whole 512-byte
			 * granular mess that exists to support odd block 
			 * sizes and UFS meta-data block sizes (e.g. 6144).
			 * A complete rewrite is required.
			 *
			 * XXX
			 */
			if (bp->b_flags & (B_NOCACHE|B_ERROR)) {
				int poffset = foff & PAGE_MASK;
				int presid;

				presid = PAGE_SIZE - poffset;
				if (bp->b_vp->v_tag == VT_NFS &&
				    bp->b_vp->v_type == VREG) {
					; /* entire page */
				} else if (presid > resid) {
					presid = resid;
				}
				KASSERT(presid >= 0, ("brelse: extra page"));
				vm_page_set_invalid(m, poffset, presid);

				/*
				 * Also make sure any swap cache is removed
				 * as it is now stale (HAMMER in particular
				 * uses B_NOCACHE to deal with buffer
				 * aliasing).
				 */
				swap_pager_unswapped(m);
			}
			resid -= PAGE_SIZE - (foff & PAGE_MASK);
			foff = (foff + PAGE_SIZE) & ~(off_t)PAGE_MASK;
		}
		if (bp->b_flags & (B_INVAL | B_RELBUF))
			vfs_vmio_release(bp);
		lwkt_reltoken(&vm_token);
	} else {
		/*
		 * Rundown for non-VMIO buffers.
		 */
		if (bp->b_flags & (B_INVAL | B_RELBUF)) {
			if (bp->b_bufsize)
				allocbuf(bp, 0);
			KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
			if (bp->b_vp)
				brelvp(bp);
		}
	}
			
	if (bp->b_qindex != BQUEUE_NONE)
		panic("brelse: free buffer onto another queue???");
	if (BUF_REFCNTNB(bp) > 1) {
		/* Temporary panic to verify exclusive locking */
		/* This panic goes away when we allow shared refs */
		panic("brelse: multiple refs");
		/* NOT REACHED */
		return;
	}

	/*
	 * Figure out the correct queue to place the cleaned up buffer on.
	 * Buffers placed in the EMPTY or EMPTYKVA had better already be
	 * disassociated from their vnode.
	 */
	spin_lock(&bufqspin);
	if (bp->b_flags & B_LOCKED) {
		/*
		 * Buffers that are locked are placed in the locked queue
		 * immediately, regardless of their state.
		 */
		bp->b_qindex = BQUEUE_LOCKED;
		TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
	} else if (bp->b_bufsize == 0) {
		/*
		 * Buffers with no memory.  Due to conditionals near the top
		 * of brelse() such buffers should probably already be
		 * marked B_INVAL and disassociated from their vnode.
		 */
		bp->b_flags |= B_INVAL;
		KASSERT(bp->b_vp == NULL, ("bp1 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
		KKASSERT((bp->b_flags & B_HASHED) == 0);
		if (bp->b_kvasize) {
			bp->b_qindex = BQUEUE_EMPTYKVA;
		} else {
			bp->b_qindex = BQUEUE_EMPTY;
		}
		TAILQ_INSERT_HEAD(&bufqueues[bp->b_qindex], bp, b_freelist);
	} else if (bp->b_flags & (B_INVAL | B_NOCACHE | B_RELBUF)) {
		/*
		 * Buffers with junk contents.   Again these buffers had better
		 * already be disassociated from their vnode.
		 */
		KASSERT(bp->b_vp == NULL, ("bp2 %p flags %08x/%08x vnode %p unexpectededly still associated!", bp, saved_flags, bp->b_flags, bp->b_vp));
		KKASSERT((bp->b_flags & B_HASHED) == 0);
		bp->b_flags |= B_INVAL;
		bp->b_qindex = BQUEUE_CLEAN;
		TAILQ_INSERT_HEAD(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
	} else {
		/*
		 * Remaining buffers.  These buffers are still associated with
		 * their vnode.
		 */
		switch(bp->b_flags & (B_DELWRI|B_HEAVY)) {
		case B_DELWRI:
		    bp->b_qindex = BQUEUE_DIRTY;
		    TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY], bp, b_freelist);
		    break;
		case B_DELWRI | B_HEAVY:
		    bp->b_qindex = BQUEUE_DIRTY_HW;
		    TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_DIRTY_HW], bp,
				      b_freelist);
		    break;
		default:
		    /*
		     * NOTE: Buffers are always placed at the end of the
		     * queue.  If B_AGE is not set the buffer will cycle
		     * through the queue twice.
		     */
		    bp->b_qindex = BQUEUE_CLEAN;
		    TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
		    break;
		}
	}
	spin_unlock(&bufqspin);

	/*
	 * If B_INVAL, clear B_DELWRI.  We've already placed the buffer
	 * on the correct queue.
	 */
	if ((bp->b_flags & (B_INVAL|B_DELWRI)) == (B_INVAL|B_DELWRI))
		bundirty(bp);

	/*
	 * The bp is on an appropriate queue unless locked.  If it is not
	 * locked or dirty we can wakeup threads waiting for buffer space.
	 *
	 * We've already handled the B_INVAL case ( B_DELWRI will be clear
	 * if B_INVAL is set ).
	 */
	if ((bp->b_flags & (B_LOCKED|B_DELWRI)) == 0)
		bufcountwakeup();

	/*
	 * Something we can maybe free or reuse
	 */
	if (bp->b_bufsize || bp->b_kvasize)
		bufspacewakeup();

	/*
	 * Clean up temporary flags and unlock the buffer.
	 */
	bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF | B_DIRECT);
	BUF_UNLOCK(bp);
}

/*
 * bqrelse:
 *
 *	Release a buffer back to the appropriate queue but do not try to free
 *	it.  The buffer is expected to be used again soon.
 *
 *	bqrelse() is used by bdwrite() to requeue a delayed write, and used by
 *	biodone() to requeue an async I/O on completion.  It is also used when
 *	known good buffers need to be requeued but we think we may need the data
 *	again soon.
 *
 *	XXX we should be able to leave the B_RELBUF hint set on completion.
 *
 * MPSAFE
 */
void
bqrelse(struct buf *bp)
{
	KASSERT(!(bp->b_flags & (B_CLUSTER|B_PAGING)), ("bqrelse: inappropriate B_PAGING or B_CLUSTER bp %p", bp));

	if (bp->b_qindex != BQUEUE_NONE)
		panic("bqrelse: free buffer onto another queue???");
	if (BUF_REFCNTNB(bp) > 1) {
		/* do not release to free list */
		panic("bqrelse: multiple refs");
		return;
	}

	buf_act_advance(bp);

	spin_lock(&bufqspin);
	if (bp->b_flags & B_LOCKED) {
		/*
		 * Locked buffers are released to the locked queue.  However,
		 * if the buffer is dirty it will first go into the dirty
		 * queue and later on after the I/O completes successfully it
		 * will be released to the locked queue.
		 */
		bp->b_qindex = BQUEUE_LOCKED;
		TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_LOCKED], bp, b_freelist);
	} else if (bp->b_flags & B_DELWRI) {
		bp->b_qindex = (bp->b_flags & B_HEAVY) ?
			       BQUEUE_DIRTY_HW : BQUEUE_DIRTY;
		TAILQ_INSERT_TAIL(&bufqueues[bp->b_qindex], bp, b_freelist);
	} else if (vm_page_count_severe()) {
		/*
		 * We are too low on memory, we have to try to free the
		 * buffer (most importantly: the wired pages making up its
		 * backing store) *now*.
		 */
		spin_unlock(&bufqspin);
		brelse(bp);
		return;
	} else {
		bp->b_qindex = BQUEUE_CLEAN;
		TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
	}
	spin_unlock(&bufqspin);

	if ((bp->b_flags & B_LOCKED) == 0 &&
	    ((bp->b_flags & B_INVAL) || (bp->b_flags & B_DELWRI) == 0)) {
		bufcountwakeup();
	}

	/*
	 * Something we can maybe free or reuse.
	 */
	if (bp->b_bufsize && !(bp->b_flags & B_DELWRI))
		bufspacewakeup();

	/*
	 * Final cleanup and unlock.  Clear bits that are only used while a
	 * buffer is actively locked.
	 */
	bp->b_flags &= ~(B_ORDERED | B_NOCACHE | B_RELBUF);
	dsched_exit_buf(bp);
	BUF_UNLOCK(bp);
}

/*
 * Hold a buffer, preventing it from being reused.  This will prevent
 * normal B_RELBUF operations on the buffer but will not prevent B_INVAL
 * operations.  If a B_INVAL operation occurs the buffer will remain held
 * but the underlying pages may get ripped out.
 *
 * These functions are typically used in VOP_READ/VOP_WRITE functions
 * to hold a buffer during a copyin or copyout, preventing deadlocks
 * or recursive lock panics when read()/write() is used over mmap()'d
 * space.
 *
 * NOTE: bqhold() requires that the buffer be locked at the time of the
 *	 hold.  bqdrop() has no requirements other than the buffer having
 *	 previously been held.
 */
void
bqhold(struct buf *bp)
{
	atomic_add_int(&bp->b_refs, 1);
}

void
bqdrop(struct buf *bp)
{
	KKASSERT(bp->b_refs > 0);
	atomic_add_int(&bp->b_refs, -1);
}

/*
 * vfs_vmio_release:
 *
 *	Return backing pages held by the buffer 'bp' back to the VM system
 *	if possible.  The pages are freed if they are no longer valid or
 *	attempt to free if it was used for direct I/O otherwise they are
 *	sent to the page cache.
 *
 *	Pages that were marked busy are left alone and skipped.
 *
 *	The KVA mapping (b_data) for the underlying pages is removed by
 *	this function.
 */
static void
vfs_vmio_release(struct buf *bp)
{
	int i;
	vm_page_t m;

	lwkt_gettoken(&vm_token);
	for (i = 0; i < bp->b_xio.xio_npages; i++) {
		m = bp->b_xio.xio_pages[i];
		bp->b_xio.xio_pages[i] = NULL;

		/*
		 * The VFS is telling us this is not a meta-data buffer
		 * even if it is backed by a block device.
		 */
		if (bp->b_flags & B_NOTMETA)
			vm_page_flag_set(m, PG_NOTMETA);

		/*
		 * This is a very important bit of code.  We try to track
		 * VM page use whether the pages are wired into the buffer
		 * cache or not.  While wired into the buffer cache the
		 * bp tracks the act_count.
		 *
		 * We can choose to place unwired pages on the inactive
		 * queue (0) or active queue (1).  If we place too many
		 * on the active queue the queue will cycle the act_count
		 * on pages we'd like to keep, just from single-use pages
		 * (such as when doing a tar-up or file scan).
		 */
		if (bp->b_act_count < vm_cycle_point)
			vm_page_unwire(m, 0);
		else
			vm_page_unwire(m, 1);

		/*
		 * We don't mess with busy pages, it is the responsibility
		 * of the process that busied the pages to deal with them.
		 *
		 * However, the caller may have marked the page invalid and
		 * we must still make sure the page is no longer mapped.
		 */
		if ((m->flags & PG_BUSY) || (m->busy != 0)) {
			vm_page_protect(m, VM_PROT_NONE);
			continue;
		}
			
		if (m->wire_count == 0) {
			vm_page_flag_clear(m, PG_ZERO);
			/*
			 * Might as well free the page if we can and it has
			 * no valid data.  We also free the page if the
			 * buffer was used for direct I/O.
			 */
#if 0
			if ((bp->b_flags & B_ASYNC) == 0 && !m->valid &&
					m->hold_count == 0) {
				vm_page_busy(m);
				vm_page_protect(m, VM_PROT_NONE);
				vm_page_free(m);
			} else
#endif
			/*
			 * Cache the page if we are really low on free
			 * pages.
			 *
			 * Also bypass the active and inactive queues
			 * if B_NOTMETA is set.  This flag is set by HAMMER
			 * on a regular file buffer when double buffering
			 * is enabled or on a block device buffer representing
			 * file data when double buffering is not enabled.
			 * The flag prevents two copies of the same data from
			 * being cached for long periods of time.
			 */
			if (bp->b_flags & B_DIRECT) {
				vm_page_try_to_free(m);
			} else if ((bp->b_flags & B_NOTMETA) ||
				   vm_page_count_severe()) {
				m->act_count = bp->b_act_count;
				vm_page_try_to_cache(m);
			} else {
				m->act_count = bp->b_act_count;
			}
		}
	}
	lwkt_reltoken(&vm_token);

	pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
		     bp->b_xio.xio_npages);
	if (bp->b_bufsize) {
		bufspacewakeup();
		bp->b_bufsize = 0;
	}
	bp->b_xio.xio_npages = 0;
	bp->b_flags &= ~B_VMIO;
	KKASSERT (LIST_FIRST(&bp->b_dep) == NULL);
	if (bp->b_vp)
		brelvp(bp);
}

/*
 * vfs_bio_awrite:
 *
 *	Implement clustered async writes for clearing out B_DELWRI buffers.
 *	This is much better then the old way of writing only one buffer at
 *	a time.  Note that we may not be presented with the buffers in the 
 *	correct order, so we search for the cluster in both directions.
 *
 *	The buffer is locked on call.
 */
int
vfs_bio_awrite(struct buf *bp)
{
	int i;
	int j;
	off_t loffset = bp->b_loffset;
	struct vnode *vp = bp->b_vp;
	int nbytes;
	struct buf *bpa;
	int nwritten;
	int size;

	/*
	 * right now we support clustered writing only to regular files.  If
	 * we find a clusterable block we could be in the middle of a cluster
	 * rather then at the beginning.
	 *
	 * NOTE: b_bio1 contains the logical loffset and is aliased
	 * to b_loffset.  b_bio2 contains the translated block number.
	 */
	if ((vp->v_type == VREG) && 
	    (vp->v_mount != 0) && /* Only on nodes that have the size info */
	    (bp->b_flags & (B_CLUSTEROK | B_INVAL)) == B_CLUSTEROK) {

		size = vp->v_mount->mnt_stat.f_iosize;

		for (i = size; i < MAXPHYS; i += size) {
			if ((bpa = findblk(vp, loffset + i, FINDBLK_TEST)) &&
			    BUF_REFCNT(bpa) == 0 &&
			    ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
			    (B_DELWRI | B_CLUSTEROK)) &&
			    (bpa->b_bufsize == size)) {
				if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
				    (bpa->b_bio2.bio_offset !=
				     bp->b_bio2.bio_offset + i))
					break;
			} else {
				break;
			}
		}
		for (j = size; i + j <= MAXPHYS && j <= loffset; j += size) {
			if ((bpa = findblk(vp, loffset - j, FINDBLK_TEST)) &&
			    BUF_REFCNT(bpa) == 0 &&
			    ((bpa->b_flags & (B_DELWRI | B_CLUSTEROK | B_INVAL)) ==
			    (B_DELWRI | B_CLUSTEROK)) &&
			    (bpa->b_bufsize == size)) {
				if ((bpa->b_bio2.bio_offset == NOOFFSET) ||
				    (bpa->b_bio2.bio_offset !=
				     bp->b_bio2.bio_offset - j))
					break;
			} else {
				break;
			}
		}
		j -= size;
		nbytes = (i + j);

		/*
		 * this is a possible cluster write
		 */
		if (nbytes != size) {
			BUF_UNLOCK(bp);
			nwritten = cluster_wbuild(vp, size,
						  loffset - j, nbytes);
			return nwritten;
		}
	}

	/*
	 * default (old) behavior, writing out only one block
	 *
	 * XXX returns b_bufsize instead of b_bcount for nwritten?
	 */
	nwritten = bp->b_bufsize;
	bremfree(bp);
	bawrite(bp);

	return nwritten;
}

/*
 * getnewbuf:
 *
 *	Find and initialize a new buffer header, freeing up existing buffers 
 *	in the bufqueues as necessary.  The new buffer is returned locked.
 *
 *	Important:  B_INVAL is not set.  If the caller wishes to throw the
 *	buffer away, the caller must set B_INVAL prior to calling brelse().
 *
 *	We block if:
 *		We have insufficient buffer headers
 *		We have insufficient buffer space
 *		buffer_map is too fragmented ( space reservation fails )
 *		If we have to flush dirty buffers ( but we try to avoid this )
 *
 *	To avoid VFS layer recursion we do not flush dirty buffers ourselves.
 *	Instead we ask the buf daemon to do it for us.  We attempt to
 *	avoid piecemeal wakeups of the pageout daemon.
 *
 * MPALMOSTSAFE
 */
struct buf *
getnewbuf(int blkflags, int slptimeo, int size, int maxsize)
{
	struct buf *bp;
	struct buf *nbp;
	int defrag = 0;
	int nqindex;
	int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
	static int flushingbufs;

	/*
	 * We can't afford to block since we might be holding a vnode lock,
	 * which may prevent system daemons from running.  We deal with
	 * low-memory situations by proactively returning memory and running
	 * async I/O rather then sync I/O.
	 */
	
	++getnewbufcalls;
	--getnewbufrestarts;
restart:
	++getnewbufrestarts;

	/*
	 * Setup for scan.  If we do not have enough free buffers,
	 * we setup a degenerate case that immediately fails.  Note
	 * that if we are specially marked process, we are allowed to
	 * dip into our reserves.
	 *
	 * The scanning sequence is nominally:  EMPTY->EMPTYKVA->CLEAN
	 *
	 * We start with EMPTYKVA.  If the list is empty we backup to EMPTY.
	 * However, there are a number of cases (defragging, reusing, ...)
	 * where we cannot backup.
	 */
	nqindex = BQUEUE_EMPTYKVA;
	spin_lock(&bufqspin);
	nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA]);

	if (nbp == NULL) {
		/*
		 * If no EMPTYKVA buffers and we are either
		 * defragging or reusing, locate a CLEAN buffer
		 * to free or reuse.  If bufspace useage is low
		 * skip this step so we can allocate a new buffer.
		 */
		if (defrag || bufspace >= lobufspace) {
			nqindex = BQUEUE_CLEAN;
			nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
		}

		/*
		 * If we could not find or were not allowed to reuse a
		 * CLEAN buffer, check to see if it is ok to use an EMPTY
		 * buffer.  We can only use an EMPTY buffer if allocating
		 * its KVA would not otherwise run us out of buffer space.
		 */
		if (nbp == NULL && defrag == 0 &&
		    bufspace + maxsize < hibufspace) {
			nqindex = BQUEUE_EMPTY;
			nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTY]);
		}
	}

	/*
	 * Run scan, possibly freeing data and/or kva mappings on the fly
	 * depending.
	 *
	 * WARNING!  bufqspin is held!
	 */
	while ((bp = nbp) != NULL) {
		int qindex = nqindex;

		nbp = TAILQ_NEXT(bp, b_freelist);

		/*
		 * BQUEUE_CLEAN - B_AGE special case.  If not set the bp
		 * cycles through the queue twice before being selected.
		 */
		if (qindex == BQUEUE_CLEAN && 
		    (bp->b_flags & B_AGE) == 0 && nbp) {
			bp->b_flags |= B_AGE;
			TAILQ_REMOVE(&bufqueues[qindex], bp, b_freelist);
			TAILQ_INSERT_TAIL(&bufqueues[qindex], bp, b_freelist);
			continue;
		}

		/*
		 * Calculate next bp ( we can only use it if we do not block
		 * or do other fancy things ).
		 */
		if (nbp == NULL) {
			switch(qindex) {
			case BQUEUE_EMPTY:
				nqindex = BQUEUE_EMPTYKVA;
				if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_EMPTYKVA])))
					break;
				/* fall through */
			case BQUEUE_EMPTYKVA:
				nqindex = BQUEUE_CLEAN;
				if ((nbp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN])))
					break;
				/* fall through */
			case BQUEUE_CLEAN:
				/*
				 * nbp is NULL. 
				 */
				break;
			}
		}

		/*
		 * Sanity Checks
		 */
		KASSERT(bp->b_qindex == qindex,
			("getnewbuf: inconsistent queue %d bp %p", qindex, bp));

		/*
		 * Note: we no longer distinguish between VMIO and non-VMIO
		 * buffers.
		 */
		KASSERT((bp->b_flags & B_DELWRI) == 0,
			("delwri buffer %p found in queue %d", bp, qindex));

		/*
		 * Do not try to reuse a buffer with a non-zero b_refs.
		 * This is an unsynchronized test.  A synchronized test
		 * is also performed after we lock the buffer.
		 */
		if (bp->b_refs)
			continue;

		/*
		 * If we are defragging then we need a buffer with 
		 * b_kvasize != 0.  XXX this situation should no longer
		 * occur, if defrag is non-zero the buffer's b_kvasize
		 * should also be non-zero at this point.  XXX
		 */
		if (defrag && bp->b_kvasize == 0) {
			kprintf("Warning: defrag empty buffer %p\n", bp);
			continue;
		}

		/*
		 * Start freeing the bp.  This is somewhat involved.  nbp
		 * remains valid only for BQUEUE_EMPTY[KVA] bp's.  Buffers
		 * on the clean list must be disassociated from their 
		 * current vnode.  Buffers on the empty[kva] lists have
		 * already been disassociated.
		 *
		 * b_refs is checked after locking along with queue changes.
		 * We must check here to deal with zero->nonzero transitions
		 * made by the owner of the buffer lock, which is used by
		 * VFS's to hold the buffer while issuing an unlocked
		 * uiomove()s.  We cannot invalidate the buffer's pages
		 * for this case.  Once we successfully lock a buffer the
		 * only 0->1 transitions of b_refs will occur via findblk().
		 *
		 * We must also check for queue changes after successful
		 * locking as the current lock holder may dispose of the
		 * buffer and change its queue.
		 */
		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
			spin_unlock(&bufqspin);
			tsleep(&bd_request, 0, "gnbxxx", (hz + 99) / 100);
			goto restart;
		}
		if (bp->b_qindex != qindex || bp->b_refs) {
			spin_unlock(&bufqspin);
			BUF_UNLOCK(bp);
			goto restart;
		}
		bremfree_locked(bp);
		spin_unlock(&bufqspin);

		/*
		 * Dependancies must be handled before we disassociate the
		 * vnode.
		 *
		 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
		 * be immediately disassociated.  HAMMER then becomes
		 * responsible for releasing the buffer.
		 *
		 * NOTE: bufqspin is UNLOCKED now.
		 */
		if (LIST_FIRST(&bp->b_dep) != NULL) {
			buf_deallocate(bp);
			if (bp->b_flags & B_LOCKED) {
				bqrelse(bp);
				goto restart;
			}
			KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
		}

		if (qindex == BQUEUE_CLEAN) {
			if (bp->b_flags & B_VMIO)
				vfs_vmio_release(bp);
			if (bp->b_vp)
				brelvp(bp);
		}

		/*
		 * NOTE:  nbp is now entirely invalid.  We can only restart
		 * the scan from this point on.
		 *
		 * Get the rest of the buffer freed up.  b_kva* is still
		 * valid after this operation.
		 */
		KASSERT(bp->b_vp == NULL,
			("bp3 %p flags %08x vnode %p qindex %d "
			 "unexpectededly still associated!",
			 bp, bp->b_flags, bp->b_vp, qindex));
		KKASSERT((bp->b_flags & B_HASHED) == 0);

		/*
		 * critical section protection is not required when
		 * scrapping a buffer's contents because it is already 
		 * wired.
		 */
		if (bp->b_bufsize)
			allocbuf(bp, 0);

		bp->b_flags = B_BNOCLIP;
		bp->b_cmd = BUF_CMD_DONE;
		bp->b_vp = NULL;
		bp->b_error = 0;
		bp->b_resid = 0;
		bp->b_bcount = 0;
		bp->b_xio.xio_npages = 0;
		bp->b_dirtyoff = bp->b_dirtyend = 0;
		bp->b_act_count = ACT_INIT;
		reinitbufbio(bp);
		KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
		buf_dep_init(bp);
		if (blkflags & GETBLK_BHEAVY)
			bp->b_flags |= B_HEAVY;

		/*
		 * If we are defragging then free the buffer.
		 */
		if (defrag) {
			bp->b_flags |= B_INVAL;
			bfreekva(bp);
			brelse(bp);
			defrag = 0;
			goto restart;
		}

		/*
		 * If we are overcomitted then recover the buffer and its
		 * KVM space.  This occurs in rare situations when multiple
		 * processes are blocked in getnewbuf() or allocbuf().
		 */
		if (bufspace >= hibufspace)
			flushingbufs = 1;
		if (flushingbufs && bp->b_kvasize != 0) {
			bp->b_flags |= B_INVAL;
			bfreekva(bp);
			brelse(bp);
			goto restart;
		}
		if (bufspace < lobufspace)
			flushingbufs = 0;

		/*
		 * b_refs can transition to a non-zero value while we hold
		 * the buffer locked due to a findblk().  Our brelvp() above
		 * interlocked any future possible transitions due to
		 * findblk()s.
		 *
		 * If we find b_refs to be non-zero we can destroy the
		 * buffer's contents but we cannot yet reuse the buffer.
		 */
		if (bp->b_refs) {
			bp->b_flags |= B_INVAL;
			bfreekva(bp);
			brelse(bp);
			goto restart;
		}
		break;
		/* NOT REACHED, bufqspin not held */
	}

	/*
	 * If we exhausted our list, sleep as appropriate.  We may have to
	 * wakeup various daemons and write out some dirty buffers.
	 *
	 * Generally we are sleeping due to insufficient buffer space.
	 *
	 * NOTE: bufqspin is held if bp is NULL, else it is not held.
	 */
	if (bp == NULL) {
		int flags;
		char *waitmsg;

		spin_unlock(&bufqspin);
		if (defrag) {
			flags = VFS_BIO_NEED_BUFSPACE;
			waitmsg = "nbufkv";
		} else if (bufspace >= hibufspace) {
			waitmsg = "nbufbs";
			flags = VFS_BIO_NEED_BUFSPACE;
		} else {
			waitmsg = "newbuf";
			flags = VFS_BIO_NEED_ANY;
		}

		bd_speedup();	/* heeeelp */
		spin_lock(&bufcspin);
		needsbuffer |= flags;
		while (needsbuffer & flags) {
			if (ssleep(&needsbuffer, &bufcspin,
				   slpflags, waitmsg, slptimeo)) {
				spin_unlock(&bufcspin);
				return (NULL);
			}
		}
		spin_unlock(&bufcspin);
	} else {
		/*
		 * We finally have a valid bp.  We aren't quite out of the
		 * woods, we still have to reserve kva space.  In order
		 * to keep fragmentation sane we only allocate kva in
		 * BKVASIZE chunks.
		 *
		 * (bufqspin is not held)
		 */
		maxsize = (maxsize + BKVAMASK) & ~BKVAMASK;

		if (maxsize != bp->b_kvasize) {
			vm_offset_t addr = 0;
			int count;

			bfreekva(bp);

			count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
			vm_map_lock(&buffer_map);

			if (vm_map_findspace(&buffer_map,
				    vm_map_min(&buffer_map), maxsize,
				    maxsize, 0, &addr)) {
				/*
				 * Uh oh.  Buffer map is too fragmented.  We
				 * must defragment the map.
				 */
				vm_map_unlock(&buffer_map);
				vm_map_entry_release(count);
				++bufdefragcnt;
				defrag = 1;
				bp->b_flags |= B_INVAL;
				brelse(bp);
				goto restart;
			}
			if (addr) {
				vm_map_insert(&buffer_map, &count,
					NULL, 0,
					addr, addr + maxsize,
					VM_MAPTYPE_NORMAL,
					VM_PROT_ALL, VM_PROT_ALL,
					MAP_NOFAULT);

				bp->b_kvabase = (caddr_t) addr;
				bp->b_kvasize = maxsize;
				bufspace += bp->b_kvasize;
				++bufreusecnt;
			}
			vm_map_unlock(&buffer_map);
			vm_map_entry_release(count);
		}
		bp->b_data = bp->b_kvabase;
	}
	return(bp);
}

/*
 * This routine is called in an emergency to recover VM pages from the
 * buffer cache by cashing in clean buffers.  The idea is to recover
 * enough pages to be able to satisfy a stuck bio_page_alloc().
 *
 * MPSAFE
 */
static int
recoverbufpages(void)
{
	struct buf *bp;
	int bytes = 0;

	++recoverbufcalls;

	spin_lock(&bufqspin);
	while (bytes < MAXBSIZE) {
		bp = TAILQ_FIRST(&bufqueues[BQUEUE_CLEAN]);
		if (bp == NULL)
			break;

		/*
		 * BQUEUE_CLEAN - B_AGE special case.  If not set the bp
		 * cycles through the queue twice before being selected.
		 */
		if ((bp->b_flags & B_AGE) == 0 && TAILQ_NEXT(bp, b_freelist)) {
			bp->b_flags |= B_AGE;
			TAILQ_REMOVE(&bufqueues[BQUEUE_CLEAN], bp, b_freelist);
			TAILQ_INSERT_TAIL(&bufqueues[BQUEUE_CLEAN],
					  bp, b_freelist);
			continue;
		}

		/*
		 * Sanity Checks
		 */
		KKASSERT(bp->b_qindex == BQUEUE_CLEAN);
		KKASSERT((bp->b_flags & B_DELWRI) == 0);

		/*
		 * Start freeing the bp.  This is somewhat involved.
		 *
		 * Buffers on the clean list must be disassociated from
		 * their current vnode
		 */

		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
			kprintf("recoverbufpages: warning, locked buf %p, "
				"race corrected\n",
				bp);
			ssleep(&bd_request, &bufqspin, 0, "gnbxxx", hz / 100);
			continue;
		}
		if (bp->b_qindex != BQUEUE_CLEAN) {
			kprintf("recoverbufpages: warning, BUF_LOCK blocked "
				"unexpectedly on buf %p index %d, race "
				"corrected\n",
				bp, bp->b_qindex);
			BUF_UNLOCK(bp);
			continue;
		}
		bremfree_locked(bp);
		spin_unlock(&bufqspin);

		/*
		 * Dependancies must be handled before we disassociate the
		 * vnode.
		 *
		 * NOTE: HAMMER will set B_LOCKED if the buffer cannot
		 * be immediately disassociated.  HAMMER then becomes
		 * responsible for releasing the buffer.
		 */
		if (LIST_FIRST(&bp->b_dep) != NULL) {
			buf_deallocate(bp);
			if (bp->b_flags & B_LOCKED) {
				bqrelse(bp);
				spin_lock(&bufqspin);
				continue;
			}
			KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
		}

		bytes += bp->b_bufsize;

		if (bp->b_flags & B_VMIO) {
			bp->b_flags |= B_DIRECT;    /* try to free pages */
			vfs_vmio_release(bp);
		}
		if (bp->b_vp)
			brelvp(bp);

		KKASSERT(bp->b_vp == NULL);
		KKASSERT((bp->b_flags & B_HASHED) == 0);

		/*
		 * critical section protection is not required when
		 * scrapping a buffer's contents because it is already 
		 * wired.
		 */
		if (bp->b_bufsize)
			allocbuf(bp, 0);

		bp->b_flags = B_BNOCLIP;
		bp->b_cmd = BUF_CMD_DONE;
		bp->b_vp = NULL;
		bp->b_error = 0;
		bp->b_resid = 0;
		bp->b_bcount = 0;
		bp->b_xio.xio_npages = 0;
		bp->b_dirtyoff = bp->b_dirtyend = 0;
		reinitbufbio(bp);
		KKASSERT(LIST_FIRST(&bp->b_dep) == NULL);
		buf_dep_init(bp);
		bp->b_flags |= B_INVAL;
		/* bfreekva(bp); */
		brelse(bp);
		spin_lock(&bufqspin);
	}
	spin_unlock(&bufqspin);
	return(bytes);
}

/*
 * buf_daemon:
 *
 *	Buffer flushing daemon.  Buffers are normally flushed by the
 *	update daemon but if it cannot keep up this process starts to
 *	take the load in an attempt to prevent getnewbuf() from blocking.
 *
 *	Once a flush is initiated it does not stop until the number
 *	of buffers falls below lodirtybuffers, but we will wake up anyone
 *	waiting at the mid-point.
 */

static struct kproc_desc buf_kp = {
	"bufdaemon",
	buf_daemon,
	&bufdaemon_td
};
SYSINIT(bufdaemon, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
	kproc_start, &buf_kp)

static struct kproc_desc bufhw_kp = {
	"bufdaemon_hw",
	buf_daemon_hw,
	&bufdaemonhw_td
};
SYSINIT(bufdaemon_hw, SI_SUB_KTHREAD_BUF, SI_ORDER_FIRST,
	kproc_start, &bufhw_kp)

/*
 * MPSAFE thread
 */
static void
buf_daemon(void)
{
	int limit;

	/*
	 * This process needs to be suspended prior to shutdown sync.
	 */
	EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
			      bufdaemon_td, SHUTDOWN_PRI_LAST);
	curthread->td_flags |= TDF_SYSTHREAD;

	/*
	 * This process is allowed to take the buffer cache to the limit
	 */
	for (;;) {
		kproc_suspend_loop();

		/*
		 * Do the flush as long as the number of dirty buffers
		 * (including those running) exceeds lodirtybufspace.
		 *
		 * When flushing limit running I/O to hirunningspace
		 * Do the flush.  Limit the amount of in-transit I/O we
		 * allow to build up, otherwise we would completely saturate
		 * the I/O system.  Wakeup any waiting processes before we
		 * normally would so they can run in parallel with our drain.
		 *
		 * Our aggregate normal+HW lo water mark is lodirtybufspace,
		 * but because we split the operation into two threads we
		 * have to cut it in half for each thread.
		 */
		waitrunningbufspace();
		limit = lodirtybufspace / 2;
		while (runningbufspace + dirtybufspace > limit ||
		       dirtybufcount - dirtybufcounthw >= nbuf / 2) {
			if (flushbufqueues(BQUEUE_DIRTY) == 0)
				break;
			if (runningbufspace < hirunningspace)
				continue;
			waitrunningbufspace();
		}

		/*
		 * We reached our low water mark, reset the
		 * request and sleep until we are needed again.
		 * The sleep is just so the suspend code works.
		 */
		spin_lock(&bufcspin);
		if (bd_request == 0)
			ssleep(&bd_request, &bufcspin, 0, "psleep", hz);
		bd_request = 0;
		spin_unlock(&bufcspin);
	}
}

/*
 * MPSAFE thread
 */
static void
buf_daemon_hw(void)
{
	int limit;

	/*
	 * This process needs to be suspended prior to shutdown sync.
	 */
	EVENTHANDLER_REGISTER(shutdown_pre_sync, shutdown_kproc,
			      bufdaemonhw_td, SHUTDOWN_PRI_LAST);
	curthread->td_flags |= TDF_SYSTHREAD;

	/*
	 * This process is allowed to take the buffer cache to the limit
	 */
	for (;;) {
		kproc_suspend_loop();

		/*
		 * Do the flush.  Limit the amount of in-transit I/O we
		 * allow to build up, otherwise we would completely saturate
		 * the I/O system.  Wakeup any waiting processes before we
		 * normally would so they can run in parallel with our drain.
		 *
		 * Once we decide to flush push the queued I/O up to
		 * hirunningspace in order to trigger bursting by the bioq
		 * subsystem.
		 *
		 * Our aggregate normal+HW lo water mark is lodirtybufspace,
		 * but because we split the operation into two threads we
		 * have to cut it in half for each thread.
		 */
		waitrunningbufspace();
		limit = lodirtybufspace / 2;
		while (runningbufspace + dirtybufspacehw > limit ||
		       dirtybufcounthw >= nbuf / 2) {
			if (flushbufqueues(BQUEUE_DIRTY_HW) == 0)
				break;
			if (runningbufspace < hirunningspace)
				continue;
			waitrunningbufspace();
		}

		/*
		 * We reached our low water mark, reset the
		 * request and sleep until we are needed again.
		 * The sleep is just so the suspend code works.
		 */
		spin_lock(&bufcspin);
		if (bd_request_hw == 0)
			ssleep(&bd_request_hw, &bufcspin, 0, "psleep", hz);
		bd_request_hw = 0;
		spin_unlock(&bufcspin);
	}
}

/*
 * flushbufqueues:
 *
 *	Try to flush a buffer in the dirty queue.  We must be careful to
 *	free up B_INVAL buffers instead of write them, which NFS is 
 *	particularly sensitive to.
 *
 *	B_RELBUF may only be set by VFSs.  We do set B_AGE to indicate
 *	that we really want to try to get the buffer out and reuse it
 *	due to the write load on the machine.
 *
 *	We must lock the buffer in order to check its validity before we
 *	can mess with its contents.  bufqspin isn't enough.
 */
static int
flushbufqueues(bufq_type_t q)
{
	struct buf *bp;
	int r = 0;
	int spun;

	spin_lock(&bufqspin);
	spun = 1;

	bp = TAILQ_FIRST(&bufqueues[q]);
	while (bp) {
		if ((bp->b_flags & B_DELWRI) == 0) {
			kprintf("Unexpected clean buffer %p\n", bp);
			bp = TAILQ_NEXT(bp, b_freelist);
			continue;
		}
		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
			bp = TAILQ_NEXT(bp, b_freelist);
			continue;
		}
		KKASSERT(bp->b_qindex == q);

		/*
		 * Must recheck B_DELWRI after successfully locking
		 * the buffer.
		 */
		if ((bp->b_flags & B_DELWRI) == 0) {
			BUF_UNLOCK(bp);
			bp = TAILQ_NEXT(bp, b_freelist);
			continue;
		}

		if (bp->b_flags & B_INVAL) {
			_bremfree(bp);
			spin_unlock(&bufqspin);
			spun = 0;
			brelse(bp);
			++r;
			break;
		}

		spin_unlock(&bufqspin);
		spun = 0;

		if (LIST_FIRST(&bp->b_dep) != NULL &&
		    (bp->b_flags & B_DEFERRED) == 0 &&
		    buf_countdeps(bp, 0)) {
			spin_lock(&bufqspin);
			spun = 1;
			TAILQ_REMOVE(&bufqueues[q], bp, b_freelist);
			TAILQ_INSERT_TAIL(&bufqueues[q], bp, b_freelist);
			bp->b_flags |= B_DEFERRED;
			BUF_UNLOCK(bp);
			bp = TAILQ_FIRST(&bufqueues[q]);
			continue;
		}

		/*
		 * If the buffer has a dependancy, buf_checkwrite() must
		 * also return 0 for us to be able to initate the write.
		 *
		 * If the buffer is flagged B_ERROR it may be requeued
		 * over and over again, we try to avoid a live lock.
		 *
		 * NOTE: buf_checkwrite is MPSAFE.
		 */
		if (LIST_FIRST(&bp->b_dep) != NULL && buf_checkwrite(bp)) {
			bremfree(bp);
			brelse(bp);
		} else if (bp->b_flags & B_ERROR) {
			tsleep(bp, 0, "bioer", 1);
			bp->b_flags &= ~B_AGE;
			vfs_bio_awrite(bp);
		} else {
			bp->b_flags |= B_AGE;
			vfs_bio_awrite(bp);
		}
		++r;
		break;
	}
	if (spun)
		spin_unlock(&bufqspin);
	return (r);
}

/*
 * inmem:
 *
 *	Returns true if no I/O is needed to access the associated VM object.
 *	This is like findblk except it also hunts around in the VM system for
 *	the data.
 *
 *	Note that we ignore vm_page_free() races from interrupts against our
 *	lookup, since if the caller is not protected our return value will not
 *	be any more valid then otherwise once we exit the critical section.
 */
int
inmem(struct vnode *vp, off_t loffset)
{
	vm_object_t obj;
	vm_offset_t toff, tinc, size;
	vm_page_t m;

	if (findblk(vp, loffset, FINDBLK_TEST))
		return 1;
	if (vp->v_mount == NULL)
		return 0;
	if ((obj = vp->v_object) == NULL)
		return 0;

	size = PAGE_SIZE;
	if (size > vp->v_mount->mnt_stat.f_iosize)
		size = vp->v_mount->mnt_stat.f_iosize;

	for (toff = 0; toff < vp->v_mount->mnt_stat.f_iosize; toff += tinc) {
		lwkt_gettoken(&vm_token);
		m = vm_page_lookup(obj, OFF_TO_IDX(loffset + toff));
		lwkt_reltoken(&vm_token);
		if (m == NULL)
			return 0;
		tinc = size;
		if (tinc > PAGE_SIZE - ((toff + loffset) & PAGE_MASK))
			tinc = PAGE_SIZE - ((toff + loffset) & PAGE_MASK);
		if (vm_page_is_valid(m,
		    (vm_offset_t) ((toff + loffset) & PAGE_MASK), tinc) == 0)
			return 0;
	}
	return 1;
}

/*
 * findblk:
 *
 *	Locate and return the specified buffer.  Unless flagged otherwise,
 *	a locked buffer will be returned if it exists or NULL if it does not.
 *
 *	findblk()'d buffers are still on the bufqueues and if you intend
 *	to use your (locked NON-TEST) buffer you need to bremfree(bp)
 *	and possibly do other stuff to it.
 *
 *	FINDBLK_TEST	- Do not lock the buffer.  The caller is responsible
 *			  for locking the buffer and ensuring that it remains
 *			  the desired buffer after locking.
 *
 *	FINDBLK_NBLOCK	- Lock the buffer non-blocking.  If we are unable
 *			  to acquire the lock we return NULL, even if the
 *			  buffer exists.
 *
 *	FINDBLK_REF	- Returns the buffer ref'd, which prevents normal
 *			  reuse by getnewbuf() but does not prevent
 *			  disassociation (B_INVAL).  Used to avoid deadlocks
 *			  against random (vp,loffset)s due to reassignment.
 *
 *	(0)		- Lock the buffer blocking.
 *
 * MPSAFE
 */
struct buf *
findblk(struct vnode *vp, off_t loffset, int flags)
{
	struct buf *bp;
	int lkflags;

	lkflags = LK_EXCLUSIVE;
	if (flags & FINDBLK_NBLOCK)
		lkflags |= LK_NOWAIT;

	for (;;) {
		/*
		 * Lookup.  Ref the buf while holding v_token to prevent
		 * reuse (but does not prevent diassociation).
		 */
		lwkt_gettoken(&vp->v_token);
		bp = buf_rb_hash_RB_LOOKUP(&vp->v_rbhash_tree, loffset);
		if (bp == NULL) {
			lwkt_reltoken(&vp->v_token);
			return(NULL);
		}
		bqhold(bp);
		lwkt_reltoken(&vp->v_token);

		/*
		 * If testing only break and return bp, do not lock.
		 */
		if (flags & FINDBLK_TEST)
			break;

		/*
		 * Lock the buffer, return an error if the lock fails.
		 * (only FINDBLK_NBLOCK can cause the lock to fail).
		 */
		if (BUF_LOCK(bp, lkflags)) {
			atomic_subtract_int(&bp->b_refs, 1);
			/* bp = NULL; not needed */
			return(NULL);
		}

		/*
		 * Revalidate the locked buf before allowing it to be
		 * returned.
		 */
		if (bp->b_vp == vp && bp->b_loffset == loffset)
			break;
		atomic_subtract_int(&bp->b_refs, 1);
		BUF_UNLOCK(bp);
	}

	/*
	 * Success
	 */
	if ((flags & FINDBLK_REF) == 0)
		atomic_subtract_int(&bp->b_refs, 1);
	return(bp);
}

/*
 * getcacheblk:
 *
 *	Similar to getblk() except only returns the buffer if it is
 *	B_CACHE and requires no other manipulation.  Otherwise NULL
 *	is returned.
 *
 *	If B_RAM is set the buffer might be just fine, but we return
 *	NULL anyway because we want the code to fall through to the
 *	cluster read.  Otherwise read-ahead breaks.
 *
 *	If blksize is 0 the buffer cache buffer must already be fully
 *	cached.
 *
 *	If blksize is non-zero getblk() will be used, allowing a buffer
 *	to be reinstantiated from its VM backing store.  The buffer must
 *	still be fully cached after reinstantiation to be returned.
 */
struct buf *
getcacheblk(struct vnode *vp, off_t loffset, int blksize)
{
	struct buf *bp;

	if (blksize) {
		bp = getblk(vp, loffset, blksize, 0, 0);
		if (bp) {
			if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
			    B_CACHE) {
				bp->b_flags &= ~B_AGE;
			} else {
				brelse(bp);
				bp = NULL;
			}
		}
	} else {
		bp = findblk(vp, loffset, 0);
		if (bp) {
			if ((bp->b_flags & (B_INVAL | B_CACHE | B_RAM)) ==
			    B_CACHE) {
				bp->b_flags &= ~B_AGE;
				bremfree(bp);
			} else {
				BUF_UNLOCK(bp);
				bp = NULL;
			}
		}
	}
	return (bp);
}

/*
 * getblk:
 *
 *	Get a block given a specified block and offset into a file/device.
 * 	B_INVAL may or may not be set on return.  The caller should clear
 *	B_INVAL prior to initiating a READ.
 *
 *	IT IS IMPORTANT TO UNDERSTAND THAT IF YOU CALL GETBLK() AND B_CACHE
 *	IS NOT SET, YOU MUST INITIALIZE THE RETURNED BUFFER, ISSUE A READ,
 *	OR SET B_INVAL BEFORE RETIRING IT.  If you retire a getblk'd buffer
 *	without doing any of those things the system will likely believe
 *	the buffer to be valid (especially if it is not B_VMIO), and the
 *	next getblk() will return the buffer with B_CACHE set.
 *
 *	For a non-VMIO buffer, B_CACHE is set to the opposite of B_INVAL for
 *	an existing buffer.
 *
 *	For a VMIO buffer, B_CACHE is modified according to the backing VM.
 *	If getblk()ing a previously 0-sized invalid buffer, B_CACHE is set
 *	and then cleared based on the backing VM.  If the previous buffer is
 *	non-0-sized but invalid, B_CACHE will be cleared.
 *
 *	If getblk() must create a new buffer, the new buffer is returned with
 *	both B_INVAL and B_CACHE clear unless it is a VMIO buffer, in which
 *	case it is returned with B_INVAL clear and B_CACHE set based on the
 *	backing VM.
 *
 *	getblk() also forces a bwrite() for any B_DELWRI buffer whos
 *	B_CACHE bit is clear.
 *	
 *	What this means, basically, is that the caller should use B_CACHE to
 *	determine whether the buffer is fully valid or not and should clear
 *	B_INVAL prior to issuing a read.  If the caller intends to validate
 *	the buffer by loading its data area with something, the caller needs
 *	to clear B_INVAL.  If the caller does this without issuing an I/O, 
 *	the caller should set B_CACHE ( as an optimization ), else the caller
 *	should issue the I/O and biodone() will set B_CACHE if the I/O was
 *	a write attempt or if it was a successfull read.  If the caller 
 *	intends to issue a READ, the caller must clear B_INVAL and B_ERROR
 *	prior to issuing the READ.  biodone() will *not* clear B_INVAL.
 *
 *	getblk flags:
 *
 *	GETBLK_PCATCH - catch signal if blocked, can cause NULL return
 *	GETBLK_BHEAVY - heavy-weight buffer cache buffer
 *
 * MPALMOSTSAFE
 */
struct buf *
getblk(struct vnode *vp, off_t loffset, int size, int blkflags, int slptimeo)
{
	struct buf *bp;
	int slpflags = (blkflags & GETBLK_PCATCH) ? PCATCH : 0;
	int error;
	int lkflags;

	if (size > MAXBSIZE)
		panic("getblk: size(%d) > MAXBSIZE(%d)", size, MAXBSIZE);
	if (vp->v_object == NULL)
		panic("getblk: vnode %p has no object!", vp);

loop:
	if ((bp = findblk(vp, loffset, FINDBLK_REF | FINDBLK_TEST)) != NULL) {
		/*
		 * The buffer was found in the cache, but we need to lock it.
		 * We must acquire a ref on the bp to prevent reuse, but
		 * this will not prevent disassociation (brelvp()) so we
		 * must recheck (vp,loffset) after acquiring the lock.
		 *
		 * Without the ref the buffer could potentially be reused
		 * before we acquire the lock and create a deadlock
		 * situation between the thread trying to reuse the buffer
		 * and us due to the fact that we would wind up blocking
		 * on a random (vp,loffset).
		 */
		if (BUF_LOCK(bp, LK_EXCLUSIVE | LK_NOWAIT)) {
			if (blkflags & GETBLK_NOWAIT) {
				bqdrop(bp);
				return(NULL);
			}
			lkflags = LK_EXCLUSIVE | LK_SLEEPFAIL;
			if (blkflags & GETBLK_PCATCH)
				lkflags |= LK_PCATCH;
			error = BUF_TIMELOCK(bp, lkflags, "getblk", slptimeo);
			if (error) {
				bqdrop(bp);
				if (error == ENOLCK)
					goto loop;
				return (NULL);
			}
			/* buffer may have changed on us */
		}
		bqdrop(bp);

		/*
		 * Once the buffer has been locked, make sure we didn't race
		 * a buffer recyclement.  Buffers that are no longer hashed
		 * will have b_vp == NULL, so this takes care of that check
		 * as well.
		 */
		if (bp->b_vp != vp || bp->b_loffset != loffset) {
			kprintf("Warning buffer %p (vp %p loffset %lld) "
				"was recycled\n",
				bp, vp, (long long)loffset);
			BUF_UNLOCK(bp);
			goto loop;
		}

		/*
		 * If SZMATCH any pre-existing buffer must be of the requested
		 * size or NULL is returned.  The caller absolutely does not
		 * want getblk() to bwrite() the buffer on a size mismatch.
		 */
		if ((blkflags & GETBLK_SZMATCH) && size != bp->b_bcount) {
			BUF_UNLOCK(bp);
			return(NULL);
		}

		/*
		 * All vnode-based buffers must be backed by a VM object.
		 */
		KKASSERT(bp->b_flags & B_VMIO);
		KKASSERT(bp->b_cmd == BUF_CMD_DONE);
		bp->b_flags &= ~B_AGE;

		/*
		 * Make sure that B_INVAL buffers do not have a cached
		 * block number translation.
		 */
		if ((bp->b_flags & B_INVAL) && (bp->b_bio2.bio_offset != NOOFFSET)) {
			kprintf("Warning invalid buffer %p (vp %p loffset %lld)"
				" did not have cleared bio_offset cache\n",
				bp, vp, (long long)loffset);
			clearbiocache(&bp->b_bio2);
		}

		/*
		 * The buffer is locked.  B_CACHE is cleared if the buffer is 
		 * invalid.
		 */
		if (bp->b_flags & B_INVAL)
			bp->b_flags &= ~B_CACHE;
		bremfree(bp);

		/*
		 * Any size inconsistancy with a dirty buffer or a buffer
		 * with a softupdates dependancy must be resolved.  Resizing
		 * the buffer in such circumstances can lead to problems.
		 *
		 * Dirty or dependant buffers are written synchronously.
		 * Other types of buffers are simply released and
		 * reconstituted as they may be backed by valid, dirty VM
		 * pages (but not marked B_DELWRI).
		 *
		 * NFS NOTE: NFS buffers which straddle EOF are oddly-sized
		 * and may be left over from a prior truncation (and thus
		 * no longer represent the actual EOF point), so we
		 * definitely do not want to B_NOCACHE the backing store.
		 */
		if (size != bp->b_bcount) {
			if (bp->b_flags & B_DELWRI) {
				bp->b_flags |= B_RELBUF;
				bwrite(bp);
			} else if (LIST_FIRST(&bp->b_dep)) {
				bp->b_flags |= B_RELBUF;
				bwrite(bp);
			} else {
				bp->b_flags |= B_RELBUF;
				brelse(bp);
			}
			goto loop;
		}
		KKASSERT(size <= bp->b_kvasize);
		KASSERT(bp->b_loffset != NOOFFSET, 
			("getblk: no buffer offset"));

		/*
		 * A buffer with B_DELWRI set and B_CACHE clear must
		 * be committed before we can return the buffer in
		 * order to prevent the caller from issuing a read
		 * ( due to B_CACHE not being set ) and overwriting
		 * it.
		 *
		 * Most callers, including NFS and FFS, need this to
		 * operate properly either because they assume they
		 * can issue a read if B_CACHE is not set, or because
		 * ( for example ) an uncached B_DELWRI might loop due 
		 * to softupdates re-dirtying the buffer.  In the latter
		 * case, B_CACHE is set after the first write completes,
		 * preventing further loops.
		 *
		 * NOTE!  b*write() sets B_CACHE.  If we cleared B_CACHE
		 * above while extending the buffer, we cannot allow the
		 * buffer to remain with B_CACHE set after the write
		 * completes or it will represent a corrupt state.  To
		 * deal with this we set B_NOCACHE to scrap the buffer
		 * after the write.
		 *
		 * XXX Should this be B_RELBUF instead of B_NOCACHE?
		 *     I'm not even sure this state is still possible
		 *     now that getblk() writes out any dirty buffers
		 *     on size changes.
		 *
		 * We might be able to do something fancy, like setting
		 * B_CACHE in bwrite() except if B_DELWRI is already set,
		 * so the below call doesn't set B_CACHE, but that gets real
		 * confusing.  This is much easier.
		 */

		if ((bp->b_flags & (B_CACHE|B_DELWRI)) == B_DELWRI) {
			kprintf("getblk: Warning, bp %p loff=%jx DELWRI set "
				"and CACHE clear, b_flags %08x\n",
				bp, (intmax_t)bp->b_loffset, bp->b_flags);
			bp->b_flags |= B_NOCACHE;
			bwrite(bp);
			goto loop;
		}
	} else {
		/*
		 * Buffer is not in-core, create new buffer.  The buffer
		 * returned by getnewbuf() is locked.  Note that the returned
		 * buffer is also considered valid (not marked B_INVAL).
		 *
		 * Calculating the offset for the I/O requires figuring out
		 * the block size.  We use DEV_BSIZE for VBLK or VCHR and
		 * the mount's f_iosize otherwise.  If the vnode does not
		 * have an associated mount we assume that the passed size is 
		 * the block size.  
		 *
		 * Note that vn_isdisk() cannot be used here since it may
		 * return a failure for numerous reasons.   Note that the
		 * buffer size may be larger then the block size (the caller
		 * will use block numbers with the proper multiple).  Beware
		 * of using any v_* fields which are part of unions.  In
		 * particular, in DragonFly the mount point overloading 
		 * mechanism uses the namecache only and the underlying
		 * directory vnode is not a special case.
		 */
		int bsize, maxsize;

		if (vp->v_type == VBLK || vp->v_type == VCHR)
			bsize = DEV_BSIZE;
		else if (vp->v_mount)
			bsize = vp->v_mount->mnt_stat.f_iosize;
		else
			bsize = size;

		maxsize = size + (loffset & PAGE_MASK);
		maxsize = imax(maxsize, bsize);

		bp = getnewbuf(blkflags, slptimeo, size, maxsize);
		if (bp == NULL) {
			if (slpflags || slptimeo)
				return NULL;
			goto loop;
		}

		/*
		 * Atomically insert the buffer into the hash, so that it can
		 * be found by findblk().
		 *
		 * If bgetvp() returns non-zero a collision occured, and the
		 * bp will not be associated with the vnode.
		 *
		 * Make sure the translation layer has been cleared.
		 */
		bp->b_loffset = loffset;
		bp->b_bio2.bio_offset = NOOFFSET;
		/* bp->b_bio2.bio_next = NULL; */

		if (bgetvp(vp, bp, size)) {
			bp->b_flags |= B_INVAL;
			brelse(bp);
			goto loop;
		}

		/*
		 * All vnode-based buffers must be backed by a VM object.
		 */
		KKASSERT(vp->v_object != NULL);
		bp->b_flags |= B_VMIO;
		KKASSERT(bp->b_cmd == BUF_CMD_DONE);

		allocbuf(bp, size);
	}
	KKASSERT(dsched_is_clear_buf_priv(bp));
	return (bp);
}

/*
 * regetblk(bp)
 *
 * Reacquire a buffer that was previously released to the locked queue,
 * or reacquire a buffer which is interlocked by having bioops->io_deallocate
 * set B_LOCKED (which handles the acquisition race).
 *
 * To this end, either B_LOCKED must be set or the dependancy list must be
 * non-empty.
 *
 * MPSAFE
 */
void
regetblk(struct buf *bp)
{
	KKASSERT((bp->b_flags & B_LOCKED) || LIST_FIRST(&bp->b_dep) != NULL);
	BUF_LOCK(bp, LK_EXCLUSIVE | LK_RETRY);
	bremfree(bp);
}

/*
 * geteblk:
 *
 *	Get an empty, disassociated buffer of given size.  The buffer is
 *	initially set to B_INVAL.
 *
 *	critical section protection is not required for the allocbuf()
 *	call because races are impossible here.
 *
 * MPALMOSTSAFE
 */
struct buf *
geteblk(int size)
{
	struct buf *bp;
	int maxsize;

	maxsize = (size + BKVAMASK) & ~BKVAMASK;

	while ((bp = getnewbuf(0, 0, size, maxsize)) == 0)
		;
	allocbuf(bp, size);
	bp->b_flags |= B_INVAL;	/* b_dep cleared by getnewbuf() */
	KKASSERT(dsched_is_clear_buf_priv(bp));
	return (bp);
}


/*
 * allocbuf:
 *
 *	This code constitutes the buffer memory from either anonymous system
 *	memory (in the case of non-VMIO operations) or from an associated
 *	VM object (in the case of VMIO operations).  This code is able to
 *	resize a buffer up or down.
 *
 *	Note that this code is tricky, and has many complications to resolve
 *	deadlock or inconsistant data situations.  Tread lightly!!! 
 *	There are B_CACHE and B_DELWRI interactions that must be dealt with by 
 *	the caller.  Calling this code willy nilly can result in the loss of
 *	data.
 *
 *	allocbuf() only adjusts B_CACHE for VMIO buffers.  getblk() deals with
 *	B_CACHE for the non-VMIO case.
 *
 *	This routine does not need to be called from a critical section but you
 *	must own the buffer.
 *
 * MPSAFE
 */
int
allocbuf(struct buf *bp, int size)
{
	int newbsize, mbsize;
	int i;

	if (BUF_REFCNT(bp) == 0)
		panic("allocbuf: buffer not busy");

	if (bp->b_kvasize < size)
		panic("allocbuf: buffer too small");

	if ((bp->b_flags & B_VMIO) == 0) {
		caddr_t origbuf;
		int origbufsize;
		/*
		 * Just get anonymous memory from the kernel.  Don't
		 * mess with B_CACHE.
		 */
		mbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
		if (bp->b_flags & B_MALLOC)
			newbsize = mbsize;
		else
			newbsize = round_page(size);

		if (newbsize < bp->b_bufsize) {
			/*
			 * Malloced buffers are not shrunk
			 */
			if (bp->b_flags & B_MALLOC) {
				if (newbsize) {
					bp->b_bcount = size;
				} else {
					kfree(bp->b_data, M_BIOBUF);
					if (bp->b_bufsize) {
						atomic_subtract_int(&bufmallocspace, bp->b_bufsize);
						bufspacewakeup();
						bp->b_bufsize = 0;
					}
					bp->b_data = bp->b_kvabase;
					bp->b_bcount = 0;
					bp->b_flags &= ~B_MALLOC;
				}
				return 1;
			}		
			vm_hold_free_pages(
			    bp,
			    (vm_offset_t) bp->b_data + newbsize,
			    (vm_offset_t) bp->b_data + bp->b_bufsize);
		} else if (newbsize > bp->b_bufsize) {
			/*
			 * We only use malloced memory on the first allocation.
			 * and revert to page-allocated memory when the buffer
			 * grows.
			 */
			if ((bufmallocspace < maxbufmallocspace) &&
				(bp->b_bufsize == 0) &&
				(mbsize <= PAGE_SIZE/2)) {

				bp->b_data = kmalloc(mbsize, M_BIOBUF, M_WAITOK);
				bp->b_bufsize = mbsize;
				bp->b_bcount = size;
				bp->b_flags |= B_MALLOC;
				atomic_add_int(&bufmallocspace, mbsize);
				return 1;
			}
			origbuf = NULL;
			origbufsize = 0;
			/*
			 * If the buffer is growing on its other-than-first
			 * allocation, then we revert to the page-allocation
			 * scheme.
			 */
			if (bp->b_flags & B_MALLOC) {
				origbuf = bp->b_data;
				origbufsize = bp->b_bufsize;
				bp->b_data = bp->b_kvabase;
				if (bp->b_bufsize) {
					atomic_subtract_int(&bufmallocspace,
							    bp->b_bufsize);
					bufspacewakeup();
					bp->b_bufsize = 0;
				}
				bp->b_flags &= ~B_MALLOC;
				newbsize = round_page(newbsize);
			}
			vm_hold_load_pages(
			    bp,
			    (vm_offset_t) bp->b_data + bp->b_bufsize,
			    (vm_offset_t) bp->b_data + newbsize);
			if (origbuf) {
				bcopy(origbuf, bp->b_data, origbufsize);
				kfree(origbuf, M_BIOBUF);
			}
		}
	} else {
		vm_page_t m;
		int desiredpages;

		newbsize = (size + DEV_BSIZE - 1) & ~(DEV_BSIZE - 1);
		desiredpages = ((int)(bp->b_loffset & PAGE_MASK) +
				newbsize + PAGE_MASK) >> PAGE_SHIFT;
		KKASSERT(desiredpages <= XIO_INTERNAL_PAGES);

		if (bp->b_flags & B_MALLOC)
			panic("allocbuf: VMIO buffer can't be malloced");
		/*
		 * Set B_CACHE initially if buffer is 0 length or will become
		 * 0-length.
		 */