xref: /dragonfly/sys/vm/swap_pager.c (revision 956939d5)

/*
 * Copyright (c) 1998,2004 The DragonFly Project.  All rights reserved.
 *
 * This code is derived from software contributed to The DragonFly Project
 * by Matthew Dillon <dillon@backplane.com>
 *
 * 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, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in
 *    the documentation and/or other materials provided with the
 *    distribution.
 * 3. Neither the name of The DragonFly Project nor the names of its
 *    contributors may be used to endorse or promote products derived
 *    from this software without specific, prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 *
 * Copyright (c) 1994 John S. Dyson
 * Copyright (c) 1990 University of Utah.
 * Copyright (c) 1991, 1993
 *	The Regents of the University of California.  All rights reserved.
 *
 * This code is derived from software contributed to Berkeley by
 * the Systems Programming Group of the University of Utah Computer
 * Science Department.
 *
 * 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, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 * 3. All advertising materials mentioning features or use of this software
 *    must display the following acknowledgement:
 *	This product includes software developed by the University of
 *	California, Berkeley and its contributors.
 * 4. Neither the name of the University nor the names of its contributors
 *    may be used to endorse or promote products derived from this software
 *    without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 *
 *				New Swap System
 *				Matthew Dillon
 *
 * Radix Bitmap 'blists'.
 *
 *	- The new swapper uses the new radix bitmap code.  This should scale
 *	  to arbitrarily small or arbitrarily large swap spaces and an almost
 *	  arbitrary degree of fragmentation.
 *
 * Features:
 *
 *	- on the fly reallocation of swap during putpages.  The new system
 *	  does not try to keep previously allocated swap blocks for dirty
 *	  pages.
 *
 *	- on the fly deallocation of swap
 *
 *	- No more garbage collection required.  Unnecessarily allocated swap
 *	  blocks only exist for dirty vm_page_t's now and these are already
 *	  cycled (in a high-load system) by the pager.  We also do on-the-fly
 *	  removal of invalidated swap blocks when a page is destroyed
 *	  or renamed.
 *
 * from: Utah $Hdr: swap_pager.c 1.4 91/04/30$
 *
 *	@(#)swap_pager.c	8.9 (Berkeley) 3/21/94
 *
 * $FreeBSD: src/sys/vm/swap_pager.c,v 1.130.2.12 2002/08/31 21:15:55 dillon Exp $
 * $DragonFly: src/sys/vm/swap_pager.c,v 1.32 2008/07/01 02:02:56 dillon Exp $
 */

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/conf.h>
#include <sys/kernel.h>
#include <sys/proc.h>
#include <sys/buf.h>
#include <sys/vnode.h>
#include <sys/malloc.h>
#include <sys/vmmeter.h>
#include <sys/sysctl.h>
#include <sys/blist.h>
#include <sys/lock.h>
#include <sys/thread2.h>

#ifndef MAX_PAGEOUT_CLUSTER
#define MAX_PAGEOUT_CLUSTER 16
#endif

#define SWB_NPAGES	MAX_PAGEOUT_CLUSTER

#include "opt_swap.h"
#include <vm/vm.h>
#include <vm/vm_object.h>
#include <vm/vm_page.h>
#include <vm/vm_pager.h>
#include <vm/vm_pageout.h>
#include <vm/swap_pager.h>
#include <vm/vm_extern.h>
#include <vm/vm_zone.h>

#include <sys/buf2.h>
#include <vm/vm_page2.h>

#define SWM_FREE	0x02	/* free, period			*/
#define SWM_POP		0x04	/* pop out			*/

#define SWBIO_READ	0x01
#define SWBIO_WRITE	0x02
#define SWBIO_SYNC	0x04

/*
 * vm_swap_size is in page-sized chunks now.  It was DEV_BSIZE'd chunks
 * in the old system.
 */

extern int vm_swap_size;	/* number of free swap blocks, in pages */

int swap_pager_full;		/* swap space exhaustion (task killing) */
static int swap_pager_almost_full; /* swap space exhaustion (w/ hysteresis)*/
static int nsw_rcount;		/* free read buffers			*/
static int nsw_wcount_sync;	/* limit write buffers / synchronous	*/
static int nsw_wcount_async;	/* limit write buffers / asynchronous	*/
static int nsw_wcount_async_max;/* assigned maximum			*/
static int nsw_cluster_max;	/* maximum VOP I/O allowed		*/
static int sw_alloc_interlock;	/* swap pager allocation interlock	*/

struct blist *swapblist;
static struct swblock **swhash;
static int swhash_mask;
static int swap_async_max = 4;	/* maximum in-progress async I/O's	*/

extern struct vnode *swapdev_vp;	/* from vm_swap.c */

SYSCTL_INT(_vm, OID_AUTO, swap_async_max,
        CTLFLAG_RW, &swap_async_max, 0, "Maximum running async swap ops");

/*
 * "named" and "unnamed" anon region objects.  Try to reduce the overhead
 * of searching a named list by hashing it just a little.
 */

#define NOBJLISTS		8

#define NOBJLIST(handle)	\
	(&swap_pager_object_list[((int)(intptr_t)handle >> 4) & (NOBJLISTS-1)])

static struct pagerlst	swap_pager_object_list[NOBJLISTS];
struct pagerlst		swap_pager_un_object_list;
vm_zone_t		swap_zone;

/*
 * pagerops for OBJT_SWAP - "swap pager".  Some ops are also global procedure
 * calls hooked from other parts of the VM system and do not appear here.
 * (see vm/swap_pager.h).
 */

static vm_object_t
		swap_pager_alloc (void *handle, off_t size,
				  vm_prot_t prot, off_t offset);
static void	swap_pager_dealloc (vm_object_t object);
static int	swap_pager_getpages (vm_object_t, vm_page_t *, int, int);
static void	swap_pager_init (void);
static void	swap_pager_unswapped (vm_page_t);
static void	swap_pager_strategy (vm_object_t, struct bio *);
static void	swap_chain_iodone(struct bio *biox);

struct pagerops swappagerops = {
	swap_pager_init,	/* early system initialization of pager	*/
	swap_pager_alloc,	/* allocate an OBJT_SWAP object		*/
	swap_pager_dealloc,	/* deallocate an OBJT_SWAP object	*/
	swap_pager_getpages,	/* pagein				*/
	swap_pager_putpages,	/* pageout				*/
	swap_pager_haspage,	/* get backing store status for page	*/
	swap_pager_unswapped,	/* remove swap related to page		*/
	swap_pager_strategy	/* pager strategy call			*/
};

/*
 * dmmax is in page-sized chunks with the new swap system.  It was
 * dev-bsized chunks in the old.  dmmax is always a power of 2.
 *
 * swap_*() routines are externally accessible.  swp_*() routines are
 * internal.
 */

int dmmax;
static int dmmax_mask;
int nswap_lowat = 128;		/* in pages, swap_pager_almost_full warn */
int nswap_hiwat = 512;		/* in pages, swap_pager_almost_full warn */

static __inline void	swp_sizecheck (void);
static void	swp_pager_async_iodone (struct bio *bio);

/*
 * Swap bitmap functions
 */

static __inline void	swp_pager_freeswapspace (daddr_t blk, int npages);
static __inline daddr_t	swp_pager_getswapspace (int npages);

/*
 * Metadata functions
 */

static void swp_pager_meta_build (vm_object_t, vm_pindex_t, daddr_t);
static void swp_pager_meta_free (vm_object_t, vm_pindex_t, daddr_t);
static void swp_pager_meta_free_all (vm_object_t);
static daddr_t swp_pager_meta_ctl (vm_object_t, vm_pindex_t, int);

/*
 * SWP_SIZECHECK() -	update swap_pager_full indication
 *
 *	update the swap_pager_almost_full indication and warn when we are
 *	about to run out of swap space, using lowat/hiwat hysteresis.
 *
 *	Clear swap_pager_full ( task killing ) indication when lowat is met.
 *
 *	No restrictions on call
 *	This routine may not block.
 *	This routine must be called at splvm()
 */

static __inline void
swp_sizecheck(void)
{
	if (vm_swap_size < nswap_lowat) {
		if (swap_pager_almost_full == 0) {
			kprintf("swap_pager: out of swap space\n");
			swap_pager_almost_full = 1;
		}
	} else {
		swap_pager_full = 0;
		if (vm_swap_size > nswap_hiwat)
			swap_pager_almost_full = 0;
	}
}

/*
 * SWAP_PAGER_INIT() -	initialize the swap pager!
 *
 *	Expected to be started from system init.  NOTE:  This code is run
 *	before much else so be careful what you depend on.  Most of the VM
 *	system has yet to be initialized at this point.
 */

static void
swap_pager_init(void)
{
	/*
	 * Initialize object lists
	 */
	int i;

	for (i = 0; i < NOBJLISTS; ++i)
		TAILQ_INIT(&swap_pager_object_list[i]);
	TAILQ_INIT(&swap_pager_un_object_list);

	/*
	 * Device Stripe, in PAGE_SIZE'd blocks
	 */

	dmmax = SWB_NPAGES * 2;
	dmmax_mask = ~(dmmax - 1);
}

/*
 * SWAP_PAGER_SWAP_INIT() - swap pager initialization from pageout process
 *
 *	Expected to be started from pageout process once, prior to entering
 *	its main loop.
 */

void
swap_pager_swap_init(void)
{
	int n, n2;

	/*
	 * Number of in-transit swap bp operations.  Don't
	 * exhaust the pbufs completely.  Make sure we
	 * initialize workable values (0 will work for hysteresis
	 * but it isn't very efficient).
	 *
	 * The nsw_cluster_max is constrained by the number of pages an XIO
	 * holds, i.e., (MAXPHYS/PAGE_SIZE) and our locally defined
	 * MAX_PAGEOUT_CLUSTER.   Also be aware that swap ops are
	 * constrained by the swap device interleave stripe size.
	 *
	 * Currently we hardwire nsw_wcount_async to 4.  This limit is
	 * designed to prevent other I/O from having high latencies due to
	 * our pageout I/O.  The value 4 works well for one or two active swap
	 * devices but is probably a little low if you have more.  Even so,
	 * a higher value would probably generate only a limited improvement
	 * with three or four active swap devices since the system does not
	 * typically have to pageout at extreme bandwidths.   We will want
	 * at least 2 per swap devices, and 4 is a pretty good value if you
	 * have one NFS swap device due to the command/ack latency over NFS.
	 * So it all works out pretty well.
	 */

	nsw_cluster_max = min((MAXPHYS/PAGE_SIZE), MAX_PAGEOUT_CLUSTER);

	nsw_rcount = (nswbuf + 1) / 2;
	nsw_wcount_sync = (nswbuf + 3) / 4;
	nsw_wcount_async = 4;
	nsw_wcount_async_max = nsw_wcount_async;

	/*
	 * The zone is dynamically allocated so generally size it to
	 * maxswzone (32MB to 512MB of KVM).  Set a minimum size based
	 * on physical memory of around 8x (each swblock can hold 16 pages).
	 *
	 * With the advent of SSDs (vs HDs) the practical (swap:memory) ratio
	 * has increased dramatically.
	 */
	n = vmstats.v_page_count / 2;
	if (maxswzone && n < maxswzone / sizeof(struct swblock))
		n = maxswzone / sizeof(struct swblock);
	n2 = n;

	do {
		swap_zone = zinit(
			"SWAPMETA",
			sizeof(struct swblock),
			n,
			ZONE_INTERRUPT,
			1);
		if (swap_zone != NULL)
			break;
		/*
		 * if the allocation failed, try a zone two thirds the
		 * size of the previous attempt.
		 */
		n -= ((n + 2) / 3);
	} while (n > 0);

	if (swap_zone == NULL)
		panic("swap_pager_swap_init: swap_zone == NULL");
	if (n2 != n)
		kprintf("Swap zone entries reduced from %d to %d.\n", n2, n);
	n2 = n;

	/*
	 * Initialize our meta-data hash table.  The swapper does not need to
	 * be quite as efficient as the VM system, so we do not use an
	 * oversized hash table.
	 *
	 * 	n: 		size of hash table, must be power of 2
	 *	swhash_mask:	hash table index mask
	 */

	for (n = 1; n < n2 / 8; n *= 2)
		;

	swhash = kmalloc(sizeof(struct swblock *) * n, M_VMPGDATA,
	    M_WAITOK | M_ZERO);

	swhash_mask = n - 1;
}

/*
 * SWAP_PAGER_ALLOC() -	allocate a new OBJT_SWAP VM object and instantiate
 *			its metadata structures.
 *
 *	This routine is called from the mmap and fork code to create a new
 *	OBJT_SWAP object.  We do this by creating an OBJT_DEFAULT object
 *	and then converting it with swp_pager_meta_build().
 *
 *	This routine may block in vm_object_allocate() and create a named
 *	object lookup race, so we must interlock.   We must also run at
 *	splvm() for the object lookup to handle races with interrupts, but
 *	we do not have to maintain splvm() in between the lookup and the
 *	add because (I believe) it is not possible to attempt to create
 *	a new swap object w/handle when a default object with that handle
 *	already exists.
 */

static vm_object_t
swap_pager_alloc(void *handle, off_t size, vm_prot_t prot, off_t offset)
{
	vm_object_t object;

	if (handle) {
		/*
		 * Reference existing named region or allocate new one.  There
		 * should not be a race here against swp_pager_meta_build()
		 * as called from vm_page_remove() in regards to the lookup
		 * of the handle.
		 */

		while (sw_alloc_interlock) {
			sw_alloc_interlock = -1;
			tsleep(&sw_alloc_interlock, 0, "swpalc", 0);
		}
		sw_alloc_interlock = 1;

		object = vm_pager_object_lookup(NOBJLIST(handle), handle);

		if (object != NULL) {
			vm_object_reference(object);
		} else {
			object = vm_object_allocate(OBJT_DEFAULT,
				OFF_TO_IDX(offset + PAGE_MASK + size));
			object->handle = handle;

			swp_pager_meta_build(object, 0, SWAPBLK_NONE);
		}

		if (sw_alloc_interlock < 0)
			wakeup(&sw_alloc_interlock);

		sw_alloc_interlock = 0;
	} else {
		object = vm_object_allocate(OBJT_DEFAULT,
			OFF_TO_IDX(offset + PAGE_MASK + size));

		swp_pager_meta_build(object, 0, SWAPBLK_NONE);
	}

	return (object);
}

/*
 * SWAP_PAGER_DEALLOC() -	remove swap metadata from object
 *
 *	The swap backing for the object is destroyed.  The code is
 *	designed such that we can reinstantiate it later, but this
 *	routine is typically called only when the entire object is
 *	about to be destroyed.
 *
 *	This routine may block, but no longer does.
 *
 *	The object must be locked or unreferenceable.
 */

static void
swap_pager_dealloc(vm_object_t object)
{
	/*
	 * Remove from list right away so lookups will fail if we block for
	 * pageout completion.
	 */

	if (object->handle == NULL) {
		TAILQ_REMOVE(&swap_pager_un_object_list, object, pager_object_list);
	} else {
		TAILQ_REMOVE(NOBJLIST(object->handle), object, pager_object_list);
	}

	vm_object_pip_wait(object, "swpdea");

	/*
	 * Free all remaining metadata.  We only bother to free it from
	 * the swap meta data.  We do not attempt to free swapblk's still
	 * associated with vm_page_t's for this object.  We do not care
	 * if paging is still in progress on some objects.
	 */
	crit_enter();
	swp_pager_meta_free_all(object);
	crit_exit();
}

/************************************************************************
 *			SWAP PAGER BITMAP ROUTINES			*
 ************************************************************************/

/*
 * SWP_PAGER_GETSWAPSPACE() -	allocate raw swap space
 *
 *	Allocate swap for the requested number of pages.  The starting
 *	swap block number (a page index) is returned or SWAPBLK_NONE
 *	if the allocation failed.
 *
 *	Also has the side effect of advising that somebody made a mistake
 *	when they configured swap and didn't configure enough.
 *
 *	Must be called at splvm() to avoid races with bitmap frees from
 *	vm_page_remove() aka swap_pager_page_removed().
 *
 *	This routine may not block
 *	This routine must be called at splvm().
 */

static __inline daddr_t
swp_pager_getswapspace(int npages)
{
	daddr_t blk;

	if ((blk = blist_alloc(swapblist, npages)) == SWAPBLK_NONE) {
		if (swap_pager_full != 2) {
			kprintf("swap_pager_getswapspace: failed\n");
			swap_pager_full = 2;
			swap_pager_almost_full = 1;
		}
	} else {
		vm_swap_size -= npages;
		swp_sizecheck();
	}
	return(blk);
}

/*
 * SWP_PAGER_FREESWAPSPACE() -	free raw swap space
 *
 *	This routine returns the specified swap blocks back to the bitmap.
 *
 *	Note:  This routine may not block (it could in the old swap code),
 *	and through the use of the new blist routines it does not block.
 *
 *	We must be called at splvm() to avoid races with bitmap frees from
 *	vm_page_remove() aka swap_pager_page_removed().
 *
 *	This routine may not block
 *	This routine must be called at splvm().
 */

static __inline void
swp_pager_freeswapspace(daddr_t blk, int npages)
{
	blist_free(swapblist, blk, npages);
	vm_swap_size += npages;
	swp_sizecheck();
}

/*
 * SWAP_PAGER_FREESPACE() -	frees swap blocks associated with a page
 *				range within an object.
 *
 *	This is a globally accessible routine.
 *
 *	This routine removes swapblk assignments from swap metadata.
 *
 *	The external callers of this routine typically have already destroyed
 *	or renamed vm_page_t's associated with this range in the object so
 *	we should be ok.
 *
 *	This routine may be called at any spl.  We up our spl to splvm temporarily
 *	in order to perform the metadata removal.
 */

void
swap_pager_freespace(vm_object_t object, vm_pindex_t start, vm_size_t size)
{
	crit_enter();
	swp_pager_meta_free(object, start, size);
	crit_exit();
}

/*
 * SWAP_PAGER_RESERVE() - reserve swap blocks in object
 *
 *	Assigns swap blocks to the specified range within the object.  The
 *	swap blocks are not zerod.  Any previous swap assignment is destroyed.
 *
 *	Returns 0 on success, -1 on failure.
 */

int
swap_pager_reserve(vm_object_t object, vm_pindex_t start, vm_size_t size)
{
	int n = 0;
	daddr_t blk = SWAPBLK_NONE;
	vm_pindex_t beg = start;	/* save start index */

	crit_enter();
	while (size) {
		if (n == 0) {
			n = BLIST_MAX_ALLOC;
			while ((blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE) {
				n >>= 1;
				if (n == 0) {
					swp_pager_meta_free(object, beg, start - beg);
					crit_exit();
					return(-1);
				}
			}
		}
		swp_pager_meta_build(object, start, blk);
		--size;
		++start;
		++blk;
		--n;
	}
	swp_pager_meta_free(object, start, n);
	crit_exit();
	return(0);
}

/*
 * SWAP_PAGER_COPY() -  copy blocks from source pager to destination pager
 *			and destroy the source.
 *
 *	Copy any valid swapblks from the source to the destination.  In
 *	cases where both the source and destination have a valid swapblk,
 *	we keep the destination's.
 *
 *	This routine is allowed to block.  It may block allocating metadata
 *	indirectly through swp_pager_meta_build() or if paging is still in
 *	progress on the source.
 *
 *	This routine can be called at any spl
 *
 *	XXX vm_page_collapse() kinda expects us not to block because we
 *	supposedly do not need to allocate memory, but for the moment we
 *	*may* have to get a little memory from the zone allocator, but
 *	it is taken from the interrupt memory.  We should be ok.
 *
 *	The source object contains no vm_page_t's (which is just as well)
 *
 *	The source object is of type OBJT_SWAP.
 *
 *	The source and destination objects must be locked or
 *	inaccessible (XXX are they ?)
 */

void
swap_pager_copy(vm_object_t srcobject, vm_object_t dstobject,
    vm_pindex_t offset, int destroysource)
{
	vm_pindex_t i;

	crit_enter();

	/*
	 * If destroysource is set, we remove the source object from the
	 * swap_pager internal queue now.
	 */

	if (destroysource) {
		if (srcobject->handle == NULL) {
			TAILQ_REMOVE(
			    &swap_pager_un_object_list,
			    srcobject,
			    pager_object_list
			);
		} else {
			TAILQ_REMOVE(
			    NOBJLIST(srcobject->handle),
			    srcobject,
			    pager_object_list
			);
		}
	}

	/*
	 * transfer source to destination.
	 */

	for (i = 0; i < dstobject->size; ++i) {
		daddr_t dstaddr;

		/*
		 * Locate (without changing) the swapblk on the destination,
		 * unless it is invalid in which case free it silently, or
		 * if the destination is a resident page, in which case the
		 * source is thrown away.
		 */

		dstaddr = swp_pager_meta_ctl(dstobject, i, 0);

		if (dstaddr == SWAPBLK_NONE) {
			/*
			 * Destination has no swapblk and is not resident,
			 * copy source.
			 */
			daddr_t srcaddr;

			srcaddr = swp_pager_meta_ctl(
			    srcobject,
			    i + offset,
			    SWM_POP
			);

			if (srcaddr != SWAPBLK_NONE)
				swp_pager_meta_build(dstobject, i, srcaddr);
		} else {
			/*
			 * Destination has valid swapblk or it is represented
			 * by a resident page.  We destroy the sourceblock.
			 */

			swp_pager_meta_ctl(srcobject, i + offset, SWM_FREE);
		}
	}

	/*
	 * Free left over swap blocks in source.
	 *
	 * We have to revert the type to OBJT_DEFAULT so we do not accidently
	 * double-remove the object from the swap queues.
	 */

	if (destroysource) {
		swp_pager_meta_free_all(srcobject);
		/*
		 * Reverting the type is not necessary, the caller is going
		 * to destroy srcobject directly, but I'm doing it here
		 * for consistency since we've removed the object from its
		 * queues.
		 */
		srcobject->type = OBJT_DEFAULT;
	}
	crit_exit();
}

/*
 * SWAP_PAGER_HASPAGE() -	determine if we have good backing store for
 *				the requested page.
 *
 *	We determine whether good backing store exists for the requested
 *	page and return TRUE if it does, FALSE if it doesn't.
 *
 *	If TRUE, we also try to determine how much valid, contiguous backing
 *	store exists before and after the requested page within a reasonable
 *	distance.  We do not try to restrict it to the swap device stripe
 *	(that is handled in getpages/putpages).  It probably isn't worth
 *	doing here.
 */

boolean_t
swap_pager_haspage(vm_object_t object, vm_pindex_t pindex, int *before,
    int *after)
{
	daddr_t blk0;

	/*
	 * do we have good backing store at the requested index ?
	 */

	crit_enter();
	blk0 = swp_pager_meta_ctl(object, pindex, 0);

	if (blk0 == SWAPBLK_NONE) {
		crit_exit();
		if (before)
			*before = 0;
		if (after)
			*after = 0;
		return (FALSE);
	}

	/*
	 * find backwards-looking contiguous good backing store
	 */

	if (before != NULL) {
		int i;

		for (i = 1; i < (SWB_NPAGES/2); ++i) {
			daddr_t blk;

			if (i > pindex)
				break;
			blk = swp_pager_meta_ctl(object, pindex - i, 0);
			if (blk != blk0 - i)
				break;
		}
		*before = (i - 1);
	}

	/*
	 * find forward-looking contiguous good backing store
	 */

	if (after != NULL) {
		int i;

		for (i = 1; i < (SWB_NPAGES/2); ++i) {
			daddr_t blk;

			blk = swp_pager_meta_ctl(object, pindex + i, 0);
			if (blk != blk0 + i)
				break;
		}
		*after = (i - 1);
	}
	crit_exit();
	return (TRUE);
}

/*
 * SWAP_PAGER_PAGE_UNSWAPPED() - remove swap backing store related to page
 *
 *	This removes any associated swap backing store, whether valid or
 *	not, from the page.
 *
 *	This routine is typically called when a page is made dirty, at
 *	which point any associated swap can be freed.  MADV_FREE also
 *	calls us in a special-case situation
 *
 *	NOTE!!!  If the page is clean and the swap was valid, the caller
 *	should make the page dirty before calling this routine.  This routine
 *	does NOT change the m->dirty status of the page.  Also: MADV_FREE
 *	depends on it.
 *
 *	This routine may not block
 *	This routine must be called at splvm()
 */

static void
swap_pager_unswapped(vm_page_t m)
{
	swp_pager_meta_ctl(m->object, m->pindex, SWM_FREE);
}

/*
 * SWAP_PAGER_STRATEGY() - read, write, free blocks
 *
 *	This implements the vm_pager_strategy() interface to swap and allows
 *	other parts of the system to directly access swap as backing store
 *	through vm_objects of type OBJT_SWAP.  This is intended to be a
 *	cacheless interface ( i.e. caching occurs at higher levels ).
 *	Therefore we do not maintain any resident pages.  All I/O goes
 *	directly to and from the swap device.
 *
 *	We currently attempt to run I/O synchronously or asynchronously as
 *	the caller requests.  This isn't perfect because we loose error
 *	sequencing when we run multiple ops in parallel to satisfy a request.
 *	But this is swap, so we let it all hang out.
 */

static void
swap_pager_strategy(vm_object_t object, struct bio *bio)
{
	struct buf *bp = bio->bio_buf;
	struct bio *nbio;
	vm_pindex_t start;
	vm_pindex_t biox_blkno = 0;
	int count;
	char *data;
	struct bio *biox;
	struct buf *bufx;
	struct bio_track *track;

	/*
	 * tracking for swapdev vnode I/Os
	 */
	if (bp->b_cmd == BUF_CMD_READ)
		track = &swapdev_vp->v_track_read;
	else
		track = &swapdev_vp->v_track_write;

	if (bp->b_bcount & PAGE_MASK) {
		bp->b_error = EINVAL;
		bp->b_flags |= B_ERROR | B_INVAL;
		biodone(bio);
		kprintf("swap_pager_strategy: bp %p offset %lld size %d, "
			"not page bounded\n",
			bp, (long long)bio->bio_offset, (int)bp->b_bcount);
		return;
	}

	/*
	 * Clear error indication, initialize page index, count, data pointer.
	 */
	bp->b_error = 0;
	bp->b_flags &= ~B_ERROR;
	bp->b_resid = bp->b_bcount;

	start = (vm_pindex_t)(bio->bio_offset >> PAGE_SHIFT);
	count = howmany(bp->b_bcount, PAGE_SIZE);
	data = bp->b_data;

	/*
	 * Deal with BUF_CMD_FREEBLKS
	 */
	if (bp->b_cmd == BUF_CMD_FREEBLKS) {
		/*
		 * FREE PAGE(s) - destroy underlying swap that is no longer
		 *		  needed.
		 */
		swp_pager_meta_free(object, start, count);
		bp->b_resid = 0;
		biodone(bio);
		return;
	}

	/*
	 * We need to be able to create a new cluster of I/O's.  We cannot
	 * use the caller fields of the passed bio so push a new one.
	 *
	 * Because nbio is just a placeholder for the cluster links,
	 * we can biodone() the original bio instead of nbio to make
	 * things a bit more efficient.
	 */
	nbio = push_bio(bio);
	nbio->bio_offset = bio->bio_offset;
	nbio->bio_caller_info1.cluster_head = NULL;
	nbio->bio_caller_info2.cluster_tail = NULL;

	biox = NULL;
	bufx = NULL;

	/*
	 * Execute read or write
	 */
	while (count > 0) {
		daddr_t blk;

		/*
		 * Obtain block.  If block not found and writing, allocate a
		 * new block and build it into the object.
		 */
		blk = swp_pager_meta_ctl(object, start, 0);
		if ((blk == SWAPBLK_NONE) && bp->b_cmd != BUF_CMD_READ) {
			blk = swp_pager_getswapspace(1);
			if (blk == SWAPBLK_NONE) {
				bp->b_error = ENOMEM;
				bp->b_flags |= B_ERROR;
				break;
			}
			swp_pager_meta_build(object, start, blk);
		}

		/*
		 * Do we have to flush our current collection?  Yes if:
		 *
		 *	- no swap block at this index
		 *	- swap block is not contiguous
		 *	- we cross a physical disk boundry in the
		 *	  stripe.
		 */
		if (
		    biox && (biox_blkno + btoc(bufx->b_bcount) != blk ||
		     ((biox_blkno ^ blk) & dmmax_mask)
		    )
		) {
			if (bp->b_cmd == BUF_CMD_READ) {
				++mycpu->gd_cnt.v_swapin;
				mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
			} else {
				++mycpu->gd_cnt.v_swapout;
				mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
				bufx->b_dirtyend = bufx->b_bcount;
			}

			/*
			 * Finished with this buf.
			 */
			KKASSERT(bufx->b_bcount != 0);
			if (bufx->b_cmd != BUF_CMD_READ)
				bufx->b_dirtyend = bufx->b_bcount;
			biox = NULL;
			bufx = NULL;
		}

		/*
		 * Add new swapblk to biox, instantiating biox if necessary.
		 * Zero-fill reads are able to take a shortcut.
		 */
		if (blk == SWAPBLK_NONE) {
			/*
			 * We can only get here if we are reading.  Since
			 * we are at splvm() we can safely modify b_resid,
			 * even if chain ops are in progress.
			 */
			bzero(data, PAGE_SIZE);
			bp->b_resid -= PAGE_SIZE;
		} else {
			if (biox == NULL) {
				/* XXX chain count > 4, wait to <= 4 */

				bufx = getpbuf(NULL);
				biox = &bufx->b_bio1;
				cluster_append(nbio, bufx);
				bufx->b_flags |= (bufx->b_flags & B_ORDERED);
				bufx->b_cmd = bp->b_cmd;
				biox->bio_done = swap_chain_iodone;
				biox->bio_offset = (off_t)blk << PAGE_SHIFT;
				biox->bio_caller_info1.cluster_parent = nbio;
				biox_blkno = blk;
				bufx->b_bcount = 0;
				bufx->b_data = data;
			}
			bufx->b_bcount += PAGE_SIZE;
		}
		--count;
		++start;
		data += PAGE_SIZE;
	}

	/*
	 *  Flush out last buffer
	 */
	if (biox) {
		if (bufx->b_cmd == BUF_CMD_READ) {
			++mycpu->gd_cnt.v_swapin;
			mycpu->gd_cnt.v_swappgsin += btoc(bufx->b_bcount);
		} else {
			++mycpu->gd_cnt.v_swapout;
			mycpu->gd_cnt.v_swappgsout += btoc(bufx->b_bcount);
			bufx->b_dirtyend = bufx->b_bcount;
		}
		KKASSERT(bufx->b_bcount);
		if (bufx->b_cmd != BUF_CMD_READ)
			bufx->b_dirtyend = bufx->b_bcount;
		/* biox, bufx = NULL */
	}

	/*
	 * Now initiate all the I/O.  Be careful looping on our chain as
	 * I/O's may complete while we are still initiating them.
	 */
	nbio->bio_caller_info2.cluster_tail = NULL;
	bufx = nbio->bio_caller_info1.cluster_head;

	while (bufx) {
		biox = &bufx->b_bio1;
		BUF_KERNPROC(bufx);
		bufx = bufx->b_cluster_next;
		vn_strategy(swapdev_vp, biox);
	}

	/*
	 * Completion of the cluster will also call biodone_chain(nbio).
	 * We never call biodone(nbio) so we don't have to worry about
	 * setting up a bio_done callback.  It's handled in the sub-IO.
	 */
	/**/
}

static void
swap_chain_iodone(struct bio *biox)
{
	struct buf **nextp;
	struct buf *bufx;	/* chained sub-buffer */
	struct bio *nbio;	/* parent nbio with chain glue */
	struct buf *bp;		/* original bp associated with nbio */
	int chain_empty;

	bufx = biox->bio_buf;
	nbio = biox->bio_caller_info1.cluster_parent;
	bp = nbio->bio_buf;

	/*
	 * Update the original buffer
	 */
        KKASSERT(bp != NULL);
	if (bufx->b_flags & B_ERROR) {
		atomic_set_int(&bufx->b_flags, B_ERROR);
		bp->b_error = bufx->b_error;
	} else if (bufx->b_resid != 0) {
		atomic_set_int(&bufx->b_flags, B_ERROR);
		bp->b_error = EINVAL;
	} else {
		atomic_subtract_int(&bp->b_resid, bufx->b_bcount);
	}

	/*
	 * Remove us from the chain.
	 */
	spin_lock_wr(&bp->b_lock.lk_spinlock);
	nextp = &nbio->bio_caller_info1.cluster_head;
	while (*nextp != bufx) {
		KKASSERT(*nextp != NULL);
		nextp = &(*nextp)->b_cluster_next;
	}
	*nextp = bufx->b_cluster_next;
	chain_empty = (nbio->bio_caller_info1.cluster_head == NULL);
	spin_unlock_wr(&bp->b_lock.lk_spinlock);

	/*
	 * Clean up bufx.  If the chain is now empty we finish out
	 * the parent.  Note that we may be racing other completions
	 * so we must use the chain_empty status from above.
	 */
	if (chain_empty) {
		if (bp->b_resid != 0 && !(bp->b_flags & B_ERROR)) {
			atomic_set_int(&bp->b_flags, B_ERROR);
			bp->b_error = EINVAL;
		}
		biodone_chain(nbio);
        }
        relpbuf(bufx, NULL);
}

/*
 * SWAP_PAGER_GETPAGES() - bring pages in from swap
 *
 *	Attempt to retrieve (m, count) pages from backing store, but make
 *	sure we retrieve at least m[reqpage].  We try to load in as large
 *	a chunk surrounding m[reqpage] as is contiguous in swap and which
 *	belongs to the same object.
 *
 *	The code is designed for asynchronous operation and
 *	immediate-notification of 'reqpage' but tends not to be
 *	used that way.  Please do not optimize-out this algorithmic
 *	feature, I intend to improve on it in the future.
 *
 *	The parent has a single vm_object_pip_add() reference prior to
 *	calling us and we should return with the same.
 *
 *	The parent has BUSY'd the pages.  We should return with 'm'
 *	left busy, but the others adjusted.
 */

static int
swap_pager_getpages(vm_object_t object, vm_page_t *m, int count, int reqpage)
{
	struct buf *bp;
	struct bio *bio;
	vm_page_t mreq;
	int i;
	int j;
	daddr_t blk;
	vm_offset_t kva;
	vm_pindex_t lastpindex;

	mreq = m[reqpage];

	if (mreq->object != object) {
		panic("swap_pager_getpages: object mismatch %p/%p",
		    object,
		    mreq->object
		);
	}

	/*
	 * Calculate range to retrieve.  The pages have already been assigned
	 * their swapblks.  We require a *contiguous* range that falls entirely
	 * within a single device stripe.   If we do not supply it, bad things
	 * happen.  Note that blk, iblk & jblk can be SWAPBLK_NONE, but the
	 * loops are set up such that the case(s) are handled implicitly.
	 *
	 * The swp_*() calls must be made at splvm().  vm_page_free() does
	 * not need to be, but it will go a little faster if it is.
	 */
	crit_enter();
	blk = swp_pager_meta_ctl(mreq->object, mreq->pindex, 0);

	for (i = reqpage - 1; i >= 0; --i) {
		daddr_t iblk;

		iblk = swp_pager_meta_ctl(m[i]->object, m[i]->pindex, 0);
		if (blk != iblk + (reqpage - i))
			break;
		if ((blk ^ iblk) & dmmax_mask)
			break;
	}
	++i;

	for (j = reqpage + 1; j < count; ++j) {
		daddr_t jblk;

		jblk = swp_pager_meta_ctl(m[j]->object, m[j]->pindex, 0);
		if (blk != jblk - (j - reqpage))
			break;
		if ((blk ^ jblk) & dmmax_mask)
			break;
	}

	/*
	 * free pages outside our collection range.   Note: we never free
	 * mreq, it must remain busy throughout.
	 */

	{
		int k;

		for (k = 0; k < i; ++k)
			vm_page_free(m[k]);
		for (k = j; k < count; ++k)
			vm_page_free(m[k]);
	}
	crit_exit();


	/*
	 * Return VM_PAGER_FAIL if we have nothing to do.  Return mreq
	 * still busy, but the others unbusied.
	 */

	if (blk == SWAPBLK_NONE)
		return(VM_PAGER_FAIL);

	/*
	 * Get a swap buffer header to perform the IO
	 */

	bp = getpbuf(&nsw_rcount);
	bio = &bp->b_bio1;
	kva = (vm_offset_t) bp->b_data;

	/*
	 * map our page(s) into kva for input
	 */

	pmap_qenter(kva, m + i, j - i);

	bp->b_data = (caddr_t) kva;
	bp->b_bcount = PAGE_SIZE * (j - i);
	bio->bio_done = swp_pager_async_iodone;
	bio->bio_offset = (off_t)(blk - (reqpage - i)) << PAGE_SHIFT;
	bio->bio_driver_info = (void *)(intptr_t)(reqpage - i);
	bio->bio_caller_info1.index = SWBIO_READ;

	{
		int k;

		for (k = i; k < j; ++k) {
			bp->b_xio.xio_pages[k - i] = m[k];
			vm_page_flag_set(m[k], PG_SWAPINPROG);
		}
	}
	bp->b_xio.xio_npages = j - i;

	mycpu->gd_cnt.v_swapin++;
	mycpu->gd_cnt.v_swappgsin += bp->b_xio.xio_npages;

	/*
	 * We still hold the lock on mreq, and our automatic completion routine
	 * does not remove it.
	 */

	vm_object_pip_add(mreq->object, bp->b_xio.xio_npages);
	lastpindex = m[j-1]->pindex;

	/*
	 * perform the I/O.  NOTE!!!  bp cannot be considered valid after
	 * this point because we automatically release it on completion.
	 * Instead, we look at the one page we are interested in which we
	 * still hold a lock on even through the I/O completion.
	 *
	 * The other pages in our m[] array are also released on completion,
	 * so we cannot assume they are valid anymore either.
	 */

	bp->b_cmd = BUF_CMD_READ;
	BUF_KERNPROC(bp);
	vn_strategy(swapdev_vp, bio);

	/*
	 * wait for the page we want to complete.  PG_SWAPINPROG is always
	 * cleared on completion.  If an I/O error occurs, SWAPBLK_NONE
	 * is set in the meta-data.
	 */

	crit_enter();

	while ((mreq->flags & PG_SWAPINPROG) != 0) {
		vm_page_flag_set(mreq, PG_WANTED | PG_REFERENCED);
		mycpu->gd_cnt.v_intrans++;
		if (tsleep(mreq, 0, "swread", hz*20)) {
			kprintf(
			    "swap_pager: indefinite wait buffer: "
				" offset: %lld, size: %ld\n",
			    (long long)bio->bio_offset,
			    (long)bp->b_bcount
			);
		}
	}

	crit_exit();

	/*
	 * mreq is left bussied after completion, but all the other pages
	 * are freed.  If we had an unrecoverable read error the page will
	 * not be valid.
	 */

	if (mreq->valid != VM_PAGE_BITS_ALL) {
		return(VM_PAGER_ERROR);
	} else {
		return(VM_PAGER_OK);
	}

	/*
	 * A final note: in a low swap situation, we cannot deallocate swap
	 * and mark a page dirty here because the caller is likely to mark
	 * the page clean when we return, causing the page to possibly revert
	 * to all-zero's later.
	 */
}

/*
 *	swap_pager_putpages:
 *
 *	Assign swap (if necessary) and initiate I/O on the specified pages.
 *
 *	We support both OBJT_DEFAULT and OBJT_SWAP objects.  DEFAULT objects
 *	are automatically converted to SWAP objects.
 *
 *	In a low memory situation we may block in vn_strategy(), but the new
 *	vm_page reservation system coupled with properly written VFS devices
 *	should ensure that no low-memory deadlock occurs.  This is an area
 *	which needs work.
 *
 *	The parent has N vm_object_pip_add() references prior to
 *	calling us and will remove references for rtvals[] that are
 *	not set to VM_PAGER_PEND.  We need to remove the rest on I/O
 *	completion.
 *
 *	The parent has soft-busy'd the pages it passes us and will unbusy
 *	those whos rtvals[] entry is not set to VM_PAGER_PEND on return.
 *	We need to unbusy the rest on I/O completion.
 */
void
swap_pager_putpages(vm_object_t object, vm_page_t *m, int count,
		    boolean_t sync, int *rtvals)
{
	int i;
	int n = 0;

	if (count && m[0]->object != object) {
		panic("swap_pager_getpages: object mismatch %p/%p",
		    object,
		    m[0]->object
		);
	}

	/*
	 * Step 1
	 *
	 * Turn object into OBJT_SWAP
	 * check for bogus sysops
	 * force sync if not pageout process
	 */

	if (object->type != OBJT_SWAP)
		swp_pager_meta_build(object, 0, SWAPBLK_NONE);

	if (curthread != pagethread)
		sync = TRUE;

	/*
	 * Step 2
	 *
	 * Update nsw parameters from swap_async_max sysctl values.
	 * Do not let the sysop crash the machine with bogus numbers.
	 */

	if (swap_async_max != nsw_wcount_async_max) {
		int n;

		/*
		 * limit range
		 */
		if ((n = swap_async_max) > nswbuf / 2)
			n = nswbuf / 2;
		if (n < 1)
			n = 1;
		swap_async_max = n;

		/*
		 * Adjust difference ( if possible ).  If the current async
		 * count is too low, we may not be able to make the adjustment
		 * at this time.
		 */
		crit_enter();
		n -= nsw_wcount_async_max;
		if (nsw_wcount_async + n >= 0) {
			nsw_wcount_async += n;
			nsw_wcount_async_max += n;
			wakeup(&nsw_wcount_async);
		}
		crit_exit();
	}

	/*
	 * Step 3
	 *
	 * Assign swap blocks and issue I/O.  We reallocate swap on the fly.
	 * The page is left dirty until the pageout operation completes
	 * successfully.
	 */

	for (i = 0; i < count; i += n) {
		struct buf *bp;
		struct bio *bio;
		daddr_t blk;
		int j;

		/*
		 * Maximum I/O size is limited by a number of factors.
		 */

		n = min(BLIST_MAX_ALLOC, count - i);
		n = min(n, nsw_cluster_max);

		crit_enter();

		/*
		 * Get biggest block of swap we can.  If we fail, fall
		 * back and try to allocate a smaller block.  Don't go
		 * overboard trying to allocate space if it would overly
		 * fragment swap.
		 */
		while (
		    (blk = swp_pager_getswapspace(n)) == SWAPBLK_NONE &&
		    n > 4
		) {
			n >>= 1;
		}
		if (blk == SWAPBLK_NONE) {
			for (j = 0; j < n; ++j)
				rtvals[i+j] = VM_PAGER_FAIL;
			crit_exit();
			continue;
		}

		/*
		 * The I/O we are constructing cannot cross a physical
		 * disk boundry in the swap stripe.  Note: we are still
		 * at splvm().
		 */
		if ((blk ^ (blk + n)) & dmmax_mask) {
			j = ((blk + dmmax) & dmmax_mask) - blk;
			swp_pager_freeswapspace(blk + j, n - j);
			n = j;
		}

		/*
		 * All I/O parameters have been satisfied, build the I/O
		 * request and assign the swap space.
		 */

		if (sync == TRUE)
			bp = getpbuf(&nsw_wcount_sync);
		else
			bp = getpbuf(&nsw_wcount_async);
		bio = &bp->b_bio1;

		pmap_qenter((vm_offset_t)bp->b_data, &m[i], n);

		bp->b_bcount = PAGE_SIZE * n;
		bio->bio_offset = (off_t)blk << PAGE_SHIFT;

		for (j = 0; j < n; ++j) {
			vm_page_t mreq = m[i+j];

			swp_pager_meta_build(
			    mreq->object,
			    mreq->pindex,
			    blk + j
			);
			vm_page_dirty(mreq);
			rtvals[i+j] = VM_PAGER_OK;

			vm_page_flag_set(mreq, PG_SWAPINPROG);
			bp->b_xio.xio_pages[j] = mreq;
		}
		bp->b_xio.xio_npages = n;

		mycpu->gd_cnt.v_swapout++;
		mycpu->gd_cnt.v_swappgsout += bp->b_xio.xio_npages;

		crit_exit();

		bp->b_dirtyoff = 0;		/* req'd for NFS */
		bp->b_dirtyend = bp->b_bcount;	/* req'd for NFS */
		bp->b_cmd = BUF_CMD_WRITE;
		bio->bio_caller_info1.index = SWBIO_WRITE;

		/*
		 * asynchronous
		 */
		if (sync == FALSE) {
			bio->bio_done = swp_pager_async_iodone;
			BUF_KERNPROC(bp);
			vn_strategy(swapdev_vp, bio);

			for (j = 0; j < n; ++j)
				rtvals[i+j] = VM_PAGER_PEND;
			continue;
		}

		/*
		 * Issue synchrnously.
		 *
		 * Wait for the sync I/O to complete, then update rtvals.
		 * We just set the rtvals[] to VM_PAGER_PEND so we can call
		 * our async completion routine at the end, thus avoiding a
		 * double-free.
		 */
		bio->bio_caller_info1.index |= SWBIO_SYNC;
		bio->bio_done = biodone_sync;
		bio->bio_flags |= BIO_SYNC;
		vn_strategy(swapdev_vp, bio);
		biowait(bio, "swwrt");

		for (j = 0; j < n; ++j)
			rtvals[i+j] = VM_PAGER_PEND;

		/*
		 * Now that we are through with the bp, we can call the
		 * normal async completion, which frees everything up.
		 */
		swp_pager_async_iodone(bio);
	}
}

void
swap_pager_newswap(void)
{
	swp_sizecheck();
}

/*
 *	swp_pager_async_iodone:
 *
 *	Completion routine for asynchronous reads and writes from/to swap.
 *	Also called manually by synchronous code to finish up a bp.
 *
 *	For READ operations, the pages are PG_BUSY'd.  For WRITE operations,
 *	the pages are vm_page_t->busy'd.  For READ operations, we PG_BUSY
 *	unbusy all pages except the 'main' request page.  For WRITE
 *	operations, we vm_page_t->busy'd unbusy all pages ( we can do this
 *	because we marked them all VM_PAGER_PEND on return from putpages ).
 *
 *	This routine may not block.
 */
static void
swp_pager_async_iodone(struct bio *bio)
{
	struct buf *bp = bio->bio_buf;
	vm_object_t object = NULL;
	int i;
	int *nswptr;

	/*
	 * report error
	 */
	if (bp->b_flags & B_ERROR) {
		kprintf(
		    "swap_pager: I/O error - %s failed; offset %lld,"
			"size %ld, error %d\n",
		    ((bio->bio_caller_info1.index & SWBIO_READ) ?
			"pagein" : "pageout"),
		    (long long)bio->bio_offset,
		    (long)bp->b_bcount,
		    bp->b_error
		);
	}

	/*
	 * set object, raise to splvm().
	 */
	if (bp->b_xio.xio_npages)
		object = bp->b_xio.xio_pages[0]->object;
	crit_enter();

	/*
	 * remove the mapping for kernel virtual
	 */
	pmap_qremove((vm_offset_t)bp->b_data, bp->b_xio.xio_npages);

	/*
	 * cleanup pages.  If an error occurs writing to swap, we are in
	 * very serious trouble.  If it happens to be a disk error, though,
	 * we may be able to recover by reassigning the swap later on.  So
	 * in this case we remove the m->swapblk assignment for the page
	 * but do not free it in the rlist.  The errornous block(s) are thus
	 * never reallocated as swap.  Redirty the page and continue.
	 */
	for (i = 0; i < bp->b_xio.xio_npages; ++i) {
		vm_page_t m = bp->b_xio.xio_pages[i];

		vm_page_flag_clear(m, PG_SWAPINPROG);

		if (bp->b_flags & B_ERROR) {
			/*
			 * If an error occurs I'd love to throw the swapblk
			 * away without freeing it back to swapspace, so it
			 * can never be used again.  But I can't from an
			 * interrupt.
			 */

			if (bio->bio_caller_info1.index & SWBIO_READ) {
				/*
				 * When reading, reqpage needs to stay
				 * locked for the parent, but all other
				 * pages can be freed.  We still want to
				 * wakeup the parent waiting on the page,
				 * though.  ( also: pg_reqpage can be -1 and
				 * not match anything ).
				 *
				 * We have to wake specifically requested pages
				 * up too because we cleared PG_SWAPINPROG and
				 * someone may be waiting for that.
				 *
				 * NOTE: for reads, m->dirty will probably
				 * be overridden by the original caller of
				 * getpages so don't play cute tricks here.
				 *
				 * NOTE: We can't actually free the page from
				 * here, because this is an interrupt.  It
				 * is not legal to mess with object->memq
				 * from an interrupt.  Deactivate the page
				 * instead.
				 */

				m->valid = 0;
				vm_page_flag_clear(m, PG_ZERO);

				/*
				 * bio_driver_info holds the requested page
				 * index.
				 */
				if (i != (int)(intptr_t)bio->bio_driver_info) {
					vm_page_deactivate(m);
					vm_page_wakeup(m);
				} else {
					vm_page_flash(m);
				}
				/*
				 * If i == bp->b_pager.pg_reqpage, do not wake
				 * the page up.  The caller needs to.
				 */
			} else {
				/*
				 * If a write error occurs, reactivate page
				 * so it doesn't clog the inactive list,
				 * then finish the I/O.
				 */
				vm_page_dirty(m);
				vm_page_activate(m);
				vm_page_io_finish(m);
			}
		} else if (bio->bio_caller_info1.index & SWBIO_READ) {
			/*
			 * NOTE: for reads, m->dirty will probably be
			 * overridden by the original caller of getpages so
			 * we cannot set them in order to free the underlying
			 * swap in a low-swap situation.  I don't think we'd
			 * want to do that anyway, but it was an optimization
			 * that existed in the old swapper for a time before
			 * it got ripped out due to precisely this problem.
			 *
			 * clear PG_ZERO in page.
			 *
			 * If not the requested page then deactivate it.
			 *
			 * Note that the requested page, reqpage, is left
			 * busied, but we still have to wake it up.  The
			 * other pages are released (unbusied) by
			 * vm_page_wakeup().  We do not set reqpage's
			 * valid bits here, it is up to the caller.
			 */

			/*
			 * NOTE: can't call pmap_clear_modify(m) from an
			 * interrupt thread, the pmap code may have to map
			 * non-kernel pmaps and currently asserts the case.
			 */
			/*pmap_clear_modify(m);*/
			m->valid = VM_PAGE_BITS_ALL;
			vm_page_undirty(m);
			vm_page_flag_clear(m, PG_ZERO);

			/*
			 * We have to wake specifically requested pages
			 * up too because we cleared PG_SWAPINPROG and
			 * could be waiting for it in getpages.  However,
			 * be sure to not unbusy getpages specifically
			 * requested page - getpages expects it to be
			 * left busy.
			 *
			 * bio_driver_info holds the requested page
			 */
			if (i != (int)(intptr_t)bio->bio_driver_info) {
				vm_page_deactivate(m);
				vm_page_wakeup(m);
			} else {
				vm_page_flash(m);
			}
		} else {
			/*
			 * Mark the page clean but do not mess with the
			 * pmap-layer's modified state.  That state should
			 * also be clear since the caller protected the
			 * page VM_PROT_READ, but allow the case.
			 *
			 * We are in an interrupt, avoid pmap operations.
			 *
			 * If we have a severe page deficit, deactivate the
			 * page.  Do not try to cache it (which would also
			 * involve a pmap op), because the page might still
			 * be read-heavy.
			 */
			vm_page_undirty(m);
			vm_page_io_finish(m);
			if (vm_page_count_severe())
				vm_page_deactivate(m);
#if 0
			if (!vm_page_count_severe() || !vm_page_try_to_cache(m))
				vm_page_protect(m, VM_PROT_READ);
#endif
		}
	}

	/*
	 * adjust pip.  NOTE: the original parent may still have its own
	 * pip refs on the object.
	 */

	if (object)
		vm_object_pip_wakeupn(object, bp->b_xio.xio_npages);

	/*
	 * Release the physical I/O buffer.
	 *
	 * NOTE: Due to synchronous operations in the write case b_cmd may
	 *	 already be set to BUF_CMD_DONE and BIO_SYNC may have already
	 *	 been cleared.
	 */
	if (bio->bio_caller_info1.index & SWBIO_READ)
		nswptr = &nsw_rcount;
	else if (bio->bio_caller_info1.index & SWBIO_SYNC)
		nswptr = &nsw_wcount_sync;
	else
		nswptr = &nsw_wcount_async;
	bp->b_cmd = BUF_CMD_DONE;
	relpbuf(bp, nswptr);
	crit_exit();
}

/************************************************************************
 *				SWAP META DATA 				*
 ************************************************************************
 *
 *	These routines manipulate the swap metadata stored in the
 *	OBJT_SWAP object.  All swp_*() routines must be called at
 *	splvm() because swap can be freed up by the low level vm_page
 *	code which might be called from interrupts beyond what splbio() covers.
 *
 *	Swap metadata is implemented with a global hash and not directly
 *	linked into the object.  Instead the object simply contains
 *	appropriate tracking counters.
 */

/*
 * SWP_PAGER_HASH() -	hash swap meta data
 *
 *	This is an inline helper function which hashes the swapblk given
 *	the object and page index.  It returns a pointer to a pointer
 *	to the object, or a pointer to a NULL pointer if it could not
 *	find a swapblk.
 *
 *	This routine must be called at splvm().
 */

static __inline struct swblock **
swp_pager_hash(vm_object_t object, vm_pindex_t index)
{
	struct swblock **pswap;
	struct swblock *swap;

	index &= ~SWAP_META_MASK;
	pswap = &swhash[(index ^ (int)(intptr_t)object) & swhash_mask];

	while ((swap = *pswap) != NULL) {
		if (swap->swb_object == object &&
		    swap->swb_index == index
		) {
			break;
		}
		pswap = &swap->swb_hnext;
	}
	return(pswap);
}

/*
 * SWP_PAGER_META_BUILD() -	add swap block to swap meta data for object
 *
 *	We first convert the object to a swap object if it is a default
 *	object.
 *
 *	The specified swapblk is added to the object's swap metadata.  If
 *	the swapblk is not valid, it is freed instead.  Any previously
 *	assigned swapblk is freed.
 *
 *	This routine must be called at splvm(), except when used to convert
 *	an OBJT_DEFAULT object into an OBJT_SWAP object.

 */

static void
swp_pager_meta_build(
	vm_object_t object,
	vm_pindex_t index,
	daddr_t swapblk
) {
	struct swblock *swap;
	struct swblock **pswap;

	/*
	 * Convert default object to swap object if necessary
	 */

	if (object->type != OBJT_SWAP) {
		object->type = OBJT_SWAP;
		object->un_pager.swp.swp_bcount = 0;

		if (object->handle != NULL) {
			TAILQ_INSERT_TAIL(
			    NOBJLIST(object->handle),
			    object,
			    pager_object_list
			);
		} else {
			TAILQ_INSERT_TAIL(
			    &swap_pager_un_object_list,
			    object,
			    pager_object_list
			);
		}
	}

	/*
	 * Locate hash entry.  If not found create, but if we aren't adding
	 * anything just return.  If we run out of space in the map we wait
	 * and, since the hash table may have changed, retry.
	 */

retry:
	pswap = swp_pager_hash(object, index);

	if ((swap = *pswap) == NULL) {
		int i;

		if (swapblk == SWAPBLK_NONE)
			return;

		swap = *pswap = zalloc(swap_zone);
		if (swap == NULL) {
			vm_wait(0);
			goto retry;
		}
		swap->swb_hnext = NULL;
		swap->swb_object = object;
		swap->swb_index = index & ~SWAP_META_MASK;
		swap->swb_count = 0;

		++object->un_pager.swp.swp_bcount;

		for (i = 0; i < SWAP_META_PAGES; ++i)
			swap->swb_pages[i] = SWAPBLK_NONE;
	}

	/*
	 * Delete prior contents of metadata
	 */

	index &= SWAP_META_MASK;

	if (swap->swb_pages[index] != SWAPBLK_NONE) {
		swp_pager_freeswapspace(swap->swb_pages[index], 1);
		--swap->swb_count;
	}

	/*
	 * Enter block into metadata
	 */

	swap->swb_pages[index] = swapblk;
	if (swapblk != SWAPBLK_NONE)
		++swap->swb_count;
}

/*
 * SWP_PAGER_META_FREE() - free a range of blocks in the object's swap metadata
 *
 *	The requested range of blocks is freed, with any associated swap
 *	returned to the swap bitmap.
 *
 *	This routine will free swap metadata structures as they are cleaned
 *	out.  This routine does *NOT* operate on swap metadata associated
 *	with resident pages.
 *
 *	This routine must be called at splvm()
 */

static void
swp_pager_meta_free(vm_object_t object, vm_pindex_t index, daddr_t count)
{
	if (object->type != OBJT_SWAP)
		return;

	while (count > 0) {
		struct swblock **pswap;
		struct swblock *swap;

		pswap = swp_pager_hash(object, index);

		if ((swap = *pswap) != NULL) {
			daddr_t v = swap->swb_pages[index & SWAP_META_MASK];

			if (v != SWAPBLK_NONE) {
				swp_pager_freeswapspace(v, 1);
				swap->swb_pages[index & SWAP_META_MASK] =
					SWAPBLK_NONE;
				if (--swap->swb_count == 0) {
					*pswap = swap->swb_hnext;
					zfree(swap_zone, swap);
					--object->un_pager.swp.swp_bcount;
				}
			}
			--count;
			++index;
		} else {
			int n = SWAP_META_PAGES - (index & SWAP_META_MASK);
			count -= n;
			index += n;
		}
	}
}

/*
 * SWP_PAGER_META_FREE_ALL() - destroy all swap metadata associated with object
 *
 *	This routine locates and destroys all swap metadata associated with
 *	an object.
 *
 *	This routine must be called at splvm()
 */

static void
swp_pager_meta_free_all(vm_object_t object)
{
	daddr_t index = 0;

	if (object->type != OBJT_SWAP)
		return;

	while (object->un_pager.swp.swp_bcount) {
		struct swblock **pswap;
		struct swblock *swap;

		pswap = swp_pager_hash(object, index);
		if ((swap = *pswap) != NULL) {
			int i;

			for (i = 0; i < SWAP_META_PAGES; ++i) {
				daddr_t v = swap->swb_pages[i];
				if (v != SWAPBLK_NONE) {
					--swap->swb_count;
					swp_pager_freeswapspace(v, 1);
				}
			}
			if (swap->swb_count != 0)
				panic("swap_pager_meta_free_all: swb_count != 0");
			*pswap = swap->swb_hnext;
			zfree(swap_zone, swap);
			--object->un_pager.swp.swp_bcount;
		}
		index += SWAP_META_PAGES;
		if (index > 0x20000000)
			panic("swp_pager_meta_free_all: failed to locate all swap meta blocks");
	}
}

/*
 * SWP_PAGER_METACTL() -  misc control of swap and vm_page_t meta data.
 *
 *	This routine is capable of looking up, popping, or freeing
 *	swapblk assignments in the swap meta data or in the vm_page_t.
 *	The routine typically returns the swapblk being looked-up, or popped,
 *	or SWAPBLK_NONE if the block was freed, or SWAPBLK_NONE if the block
 *	was invalid.  This routine will automatically free any invalid
 *	meta-data swapblks.
 *
 *	It is not possible to store invalid swapblks in the swap meta data
 *	(other then a literal 'SWAPBLK_NONE'), so we don't bother checking.
 *
 *	When acting on a busy resident page and paging is in progress, we
 *	have to wait until paging is complete but otherwise can act on the
 *	busy page.
 *
 *	This routine must be called at splvm().
 *
 *	SWM_FREE	remove and free swap block from metadata
 *	SWM_POP		remove from meta data but do not free.. pop it out
 */

static daddr_t
swp_pager_meta_ctl(
	vm_object_t object,
	vm_pindex_t index,
	int flags
) {
	struct swblock **pswap;
	struct swblock *swap;
	daddr_t r1;

	/*
	 * The meta data only exists of the object is OBJT_SWAP
	 * and even then might not be allocated yet.
	 */

	if (object->type != OBJT_SWAP)
		return(SWAPBLK_NONE);

	r1 = SWAPBLK_NONE;
	pswap = swp_pager_hash(object, index);

	if ((swap = *pswap) != NULL) {
		index &= SWAP_META_MASK;
		r1 = swap->swb_pages[index];

		if (r1 != SWAPBLK_NONE) {
			if (flags & SWM_FREE) {
				swp_pager_freeswapspace(r1, 1);
				r1 = SWAPBLK_NONE;
			}
			if (flags & (SWM_FREE|SWM_POP)) {
				swap->swb_pages[index] = SWAPBLK_NONE;
				if (--swap->swb_count == 0) {
					*pswap = swap->swb_hnext;
					zfree(swap_zone, swap);
					--object->un_pager.swp.swp_bcount;
				}
			}
		}
	}
	return(r1);
}