NetBSD/sys/uvm/uvm_loan.c

/*	$NetBSD: uvm_loan.c,v 1.50 2004/03/24 07:50:48 junyoung Exp $	*/

/*
 *
 * Copyright (c) 1997 Charles D. Cranor and Washington University.
 * 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, 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 Charles D. Cranor and
 *      Washington University.
 * 4. The name of the author may not be used to endorse or promote products
 *    derived from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``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 AUTHOR 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.
 *
 * from: Id: uvm_loan.c,v 1.1.6.4 1998/02/06 05:08:43 chs Exp
 */

/*
 * uvm_loan.c: page loanout handler
 */

#include <sys/cdefs.h>
__KERNEL_RCSID(0, "$NetBSD: uvm_loan.c,v 1.50 2004/03/24 07:50:48 junyoung Exp $");

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kernel.h>
#include <sys/proc.h>
#include <sys/malloc.h>
#include <sys/mman.h>

#include <uvm/uvm.h>

/*
 * "loaned" pages are pages which are (read-only, copy-on-write) loaned
 * from the VM system to other parts of the kernel.   this allows page
 * copying to be avoided (e.g. you can loan pages from objs/anons to
 * the mbuf system).
 *
 * there are 3 types of loans possible:
 *  O->K  uvm_object page to wired kernel page (e.g. mbuf data area)
 *  A->K  anon page to wired kernel page (e.g. mbuf data area)
 *  O->A  uvm_object to anon loan (e.g. vnode page to an anon)
 * note that it possible to have an O page loaned to both an A and K
 * at the same time.
 *
 * loans are tracked by pg->loan_count.  an O->A page will have both
 * a uvm_object and a vm_anon, but PQ_ANON will not be set.   this sort
 * of page is considered "owned" by the uvm_object (not the anon).
 *
 * each loan of a page to the kernel bumps the pg->wire_count.  the
 * kernel mappings for these pages will be read-only and wired.  since
 * the page will also be wired, it will not be a candidate for pageout,
 * and thus will never be pmap_page_protect()'d with VM_PROT_NONE.  a
 * write fault in the kernel to one of these pages will not cause
 * copy-on-write.  instead, the page fault is considered fatal.  this
 * is because the kernel mapping will have no way to look up the
 * object/anon which the page is owned by.  this is a good side-effect,
 * since a kernel write to a loaned page is an error.
 *
 * owners that want to free their pages and discover that they are
 * loaned out simply "disown" them (the page becomes an orphan).  these
 * pages should be freed when the last loan is dropped.   in some cases
 * an anon may "adopt" an orphaned page.
 *
 * locking: to read pg->loan_count either the owner or the page queues
 * must be locked.   to modify pg->loan_count, both the owner of the page
 * and the PQs must be locked.   pg->flags is (as always) locked by
 * the owner of the page.
 *
 * note that locking from the "loaned" side is tricky since the object
 * getting the loaned page has no reference to the page's owner and thus
 * the owner could "die" at any time.   in order to prevent the owner
 * from dying the page queues should be locked.   this forces us to sometimes
 * use "try" locking.
 *
 * loans are typically broken by the following events:
 *  1. user-level xwrite fault to a loaned page
 *  2. pageout of clean+inactive O->A loaned page
 *  3. owner frees page (e.g. pager flush)
 *
 * note that loaning a page causes all mappings of the page to become
 * read-only (via pmap_page_protect).   this could have an unexpected
 * effect on normal "wired" pages if one is not careful (XXX).
 */

/*
 * local prototypes
 */

static int	uvm_loananon(struct uvm_faultinfo *, void ***,
			     int, struct vm_anon *);
static int	uvm_loanentry(struct uvm_faultinfo *, void ***, int);
static int	uvm_loanuobj(struct uvm_faultinfo *, void ***,
			     int, vaddr_t);
static int	uvm_loanzero(struct uvm_faultinfo *, void ***, int);
static void	uvm_unloananon(struct vm_anon **, int);
static void	uvm_unloanpage(struct vm_page **, int);
static void	uvm_loanpage(struct vm_page **, int);


/*
 * inlines
 */

/*
 * uvm_loanentry: loan out pages in a map entry (helper fn for uvm_loan())
 *
 * => "ufi" is the result of a successful map lookup (meaning that
 *	on entry the map is locked by the caller)
 * => we may unlock and then relock the map if needed (for I/O)
 * => we put our output result in "output"
 * => we always return with the map unlocked
 * => possible return values:
 *	-1 == error, map is unlocked
 *	 0 == map relock error (try again!), map is unlocked
 *	>0 == number of pages we loaned, map is unlocked
 */

static __inline int
uvm_loanentry(ufi, output, flags)
	struct uvm_faultinfo *ufi;
	void ***output;
	int flags;
{
	vaddr_t curaddr = ufi->orig_rvaddr;
	vsize_t togo = ufi->size;
	struct vm_aref *aref = &ufi->entry->aref;
	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
	struct vm_anon *anon;
	int rv, result = 0;

	/*
	 * lock us the rest of the way down (we unlock before return)
	 */
	if (aref->ar_amap)
		amap_lock(aref->ar_amap);
	if (uobj)
		simple_lock(&uobj->vmobjlock);

	/*
	 * loop until done
	 */
	while (togo) {

		/*
		 * find the page we want.   check the anon layer first.
		 */

		if (aref->ar_amap) {
			anon = amap_lookup(aref, curaddr - ufi->entry->start);
		} else {
			anon = NULL;
		}

		/* locked: map, amap, uobj */
		if (anon) {
			rv = uvm_loananon(ufi, output, flags, anon);
		} else if (uobj) {
			rv = uvm_loanuobj(ufi, output, flags, curaddr);
		} else if (UVM_ET_ISCOPYONWRITE(ufi->entry)) {
			rv = uvm_loanzero(ufi, output, flags);
		} else {
			uvmfault_unlockall(ufi, aref->ar_amap, uobj, NULL);
			rv = -1;
		}
		/* locked: if (rv > 0) => map, amap, uobj  [o.w. unlocked] */

		/* total failure */
		if (rv < 0)
			return (-1);

		/* relock failed, need to do another lookup */
		if (rv == 0)
			return (result);

		/*
		 * got it... advance to next page
		 */

		result++;
		togo -= PAGE_SIZE;
		curaddr += PAGE_SIZE;
	}

	/*
	 * unlock what we locked, unlock the maps and return
	 */

	if (aref->ar_amap)
		amap_unlock(aref->ar_amap);
	if (uobj)
		simple_unlock(&uobj->vmobjlock);
	uvmfault_unlockmaps(ufi, FALSE);
	return (result);
}

/*
 * normal functions
 */

/*
 * uvm_loan: loan pages in a map out to anons or to the kernel
 *
 * => map should be unlocked
 * => start and len should be multiples of PAGE_SIZE
 * => result is either an array of anon's or vm_pages (depending on flags)
 * => flag values: UVM_LOAN_TOANON - loan to anons
 *                 UVM_LOAN_TOPAGE - loan to wired kernel page
 *    one and only one of these flags must be set!
 * => returns 0 (success), or an appropriate error number
 */

int
uvm_loan(map, start, len, v, flags)
	struct vm_map *map;
	vaddr_t start;
	vsize_t len;
	void *v;
	int flags;
{
	struct uvm_faultinfo ufi;
	void **result, **output;
	int rv, error;

	/*
	 * ensure that one and only one of the flags is set
	 */

	KASSERT(((flags & UVM_LOAN_TOANON) == 0) ^
		((flags & UVM_LOAN_TOPAGE) == 0));
	KASSERT((map->flags & VM_MAP_INTRSAFE) == 0);

	/*
	 * "output" is a pointer to the current place to put the loaned page.
	 */

	result = v;
	output = &result[0];	/* start at the beginning ... */

	/*
	 * while we've got pages to do
	 */

	while (len > 0) {

		/*
		 * fill in params for a call to uvmfault_lookup
		 */

		ufi.orig_map = map;
		ufi.orig_rvaddr = start;
		ufi.orig_size = len;

		/*
		 * do the lookup, the only time this will fail is if we hit on
		 * an unmapped region (an error)
		 */

		if (!uvmfault_lookup(&ufi, FALSE)) {
			error = ENOENT;
			goto fail;
		}

		/*
		 * map now locked.  now do the loanout...
		 */

		rv = uvm_loanentry(&ufi, &output, flags);
		if (rv < 0) {
			/* all unlocked due to error */
			error = EINVAL;
			goto fail;
		}

		/*
		 * done!  the map is unlocked.  advance, if possible.
		 *
		 * XXXCDC: could be recoded to hold the map lock with
		 *	   smarter code (but it only happens on map entry
		 *	   boundaries, so it isn't that bad).
		 */

		if (rv) {
			rv <<= PAGE_SHIFT;
			len -= rv;
			start += rv;
		}
	}
	return 0;

fail:
	/*
	 * failed to complete loans.  drop any loans and return failure code.
	 * map is already unlocked.
	 */

	if (output - result) {
		if (flags & UVM_LOAN_TOANON) {
			uvm_unloananon((struct vm_anon **)result,
			    output - result);
		} else {
			uvm_unloanpage((struct vm_page **)result,
			    output - result);
		}
	}
	return (error);
}

/*
 * uvm_loananon: loan a page from an anon out
 *
 * => called with map, amap, uobj locked
 * => return value:
 *	-1 = fatal error, everything is unlocked, abort.
 *	 0 = lookup in ufi went stale, everything unlocked, relookup and
 *		try again
 *	 1 = got it, everything still locked
 */

int
uvm_loananon(ufi, output, flags, anon)
	struct uvm_faultinfo *ufi;
	void ***output;
	int flags;
	struct vm_anon *anon;
{
	struct vm_page *pg;
	int error;

	/*
	 * if we are loaning to "another" anon then it is easy, we just
	 * bump the reference count on the current anon and return a
	 * pointer to it (it becomes copy-on-write shared).
	 */

	if (flags & UVM_LOAN_TOANON) {
		simple_lock(&anon->an_lock);
		pg = anon->u.an_page;
		if (pg && (pg->pqflags & PQ_ANON) != 0 && anon->an_ref == 1) {
			pmap_page_protect(pg, VM_PROT_READ);
		}
		anon->an_ref++;
		**output = anon;
		(*output)++;
		simple_unlock(&anon->an_lock);
		return (1);
	}

	/*
	 * we are loaning to a kernel-page.   we need to get the page
	 * resident so we can wire it.   uvmfault_anonget will handle
	 * this for us.
	 */

	simple_lock(&anon->an_lock);
	error = uvmfault_anonget(ufi, ufi->entry->aref.ar_amap, anon);

	/*
	 * if we were unable to get the anon, then uvmfault_anonget has
	 * unlocked everything and returned an error code.
	 */

	if (error) {

		/* need to refault (i.e. refresh our lookup) ? */
		if (error == ERESTART) {
			return (0);
		}

		/* "try again"?   sleep a bit and retry ... */
		if (error == EAGAIN) {
			tsleep(&lbolt, PVM, "loanagain", 0);
			return (0);
		}

		/* otherwise flag it as an error */
		return (-1);
	}

	/*
	 * we have the page and its owner locked: do the loan now.
	 */

	pg = anon->u.an_page;
	uvm_lock_pageq();
	KASSERT(pg->wire_count == 0);
	if (pg->loan_count == 0) {
		pmap_page_protect(pg, VM_PROT_READ);
	}
	pg->loan_count++;
	uvm_pagedequeue(pg);
	uvm_unlock_pageq();
	**output = pg;
	(*output)++;

	/* unlock anon and return success */
	if (pg->uobject)	/* XXXCDC: what if this is our uobj? bad */
		simple_unlock(&pg->uobject->vmobjlock);
	simple_unlock(&anon->an_lock);
	return (1);
}

/*
 * uvm_loanpage: loan out pages to kernel (->K)
 *
 * => page's owner should be locked.
 * => caller should own the pages.
 */
static void
uvm_loanpage(pgpp, npages)
	struct vm_page **pgpp;
	int npages;
{
	int i;

	for (i = 0; i < npages; i++) {
		struct vm_page *pg = pgpp[i];

		KASSERT(pg->uobject != NULL);
		KASSERT(!(pg->flags & (PG_RELEASED|PG_PAGEOUT)));
		LOCK_ASSERT(simple_lock_held(&pg->uobject->vmobjlock));
		KASSERT(pg->flags & PG_BUSY);

		uvm_lock_pageq();
		if (pg->loan_count == 0) {
			pmap_page_protect(pg, VM_PROT_READ);
		}
		pg->loan_count++;
		uvm_pagedequeue(pg);
		uvm_unlock_pageq();
		if (pg->flags & PG_WANTED) {
			wakeup(pg);
		}
		pg->flags &= ~(PG_WANTED|PG_BUSY);
		UVM_PAGE_OWN(pg, NULL);
	}
}

/*
 * XXX UBC temp limit
 * number of pages to get at once.
 * should be <= MAX_READ_AHEAD in genfs_vnops.c
 */
#define	UVM_LOAN_GET_CHUNK	16

/*
 * uvm_loanuobjpages: loan pages from a uobj out (O->K)
 *
 * => uobj shouldn't be locked.  (we'll lock it)
 */
int
uvm_loanuobjpages(uobj, pgoff, orignpages, origpgpp)
	struct uvm_object *uobj;
	voff_t pgoff;
	int orignpages;
	struct vm_page **origpgpp;
{
	int ndone;
	struct vm_page **pgpp;
	int error;
	int i;
	struct simplelock *slock;

	pgpp = origpgpp;
	for (ndone = 0; ndone < orignpages; ) {
		int npages;
		int npendloan = 0xdead; /* XXX gcc */
reget:
		npages = MIN(UVM_LOAN_GET_CHUNK, orignpages - ndone);
		simple_lock(&uobj->vmobjlock);
		error = (*uobj->pgops->pgo_get)(uobj,
		    pgoff + (ndone << PAGE_SHIFT), pgpp, &npages, 0,
		    VM_PROT_READ, 0, PGO_SYNCIO);
		if (error == EAGAIN) {
			tsleep(&lbolt, PVM, "nfsread", 0);
			continue;
		}
		if (error) {
			uvm_unloan(origpgpp, ndone, UVM_LOAN_TOPAGE);
			return error;
		}

		KASSERT(npages > 0);

		/* loan and unbusy pages */
		slock = NULL;
		for (i = 0; i < npages; i++) {
			struct simplelock *nextslock; /* slock for next page */
			struct vm_page *pg = *pgpp;

			/* XXX assuming that the page is owned by uobj */
			KASSERT(pg->uobject != NULL);
			nextslock = &pg->uobject->vmobjlock;

			if (slock != nextslock) {
				if (slock) {
					KASSERT(npendloan > 0);
					uvm_loanpage(pgpp - npendloan,
					    npendloan);
					simple_unlock(slock);
				}
				slock = nextslock;
				simple_lock(slock);
				npendloan = 0;
			}

			if (pg->flags & PG_RELEASED) {
				/*
				 * release pages and try again.
				 */
				simple_unlock(slock);
				for (; i < npages; i++) {
					pg = pgpp[i];
					slock = &pg->uobject->vmobjlock;

					simple_lock(slock);
					uvm_lock_pageq();
					uvm_page_unbusy(&pg, 1);
					uvm_unlock_pageq();
					simple_unlock(slock);
				}
				goto reget;
			}

			npendloan++;
			pgpp++;
			ndone++;
			KASSERT(pgpp - origpgpp == ndone);
		}
		KASSERT(slock != NULL);
		KASSERT(npendloan > 0);
		uvm_loanpage(pgpp - npendloan, npendloan);
		simple_unlock(slock);
	}

	return 0;
}

/*
 * uvm_loanuobj: loan a page from a uobj out
 *
 * => called with map, amap, uobj locked
 * => return value:
 *	-1 = fatal error, everything is unlocked, abort.
 *	 0 = lookup in ufi went stale, everything unlocked, relookup and
 *		try again
 *	 1 = got it, everything still locked
 */

static int
uvm_loanuobj(ufi, output, flags, va)
	struct uvm_faultinfo *ufi;
	void ***output;
	int flags;
	vaddr_t va;
{
	struct vm_amap *amap = ufi->entry->aref.ar_amap;
	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
	struct vm_page *pg;
	struct vm_anon *anon;
	int error, npages;
	boolean_t locked;

	/*
	 * first we must make sure the page is resident.
	 *
	 * XXXCDC: duplicate code with uvm_fault().
	 */

	if (uobj->pgops->pgo_get) {	/* try locked pgo_get */
		npages = 1;
		pg = NULL;
		error = (*uobj->pgops->pgo_get)(uobj,
		    va - ufi->entry->start + ufi->entry->offset,
		    &pg, &npages, 0, VM_PROT_READ, MADV_NORMAL, PGO_LOCKED);
	} else {
		error = EIO;		/* must have pgo_get op */
	}

	/*
	 * check the result of the locked pgo_get.  if there is a problem,
	 * then we fail the loan.
	 */

	if (error && error != EBUSY) {
		uvmfault_unlockall(ufi, amap, uobj, NULL);
		return (-1);
	}

	/*
	 * if we need to unlock for I/O, do so now.
	 */

	if (error == EBUSY) {
		uvmfault_unlockall(ufi, amap, NULL, NULL);

		/* locked: uobj */
		npages = 1;
		error = (*uobj->pgops->pgo_get)(uobj,
		    va - ufi->entry->start + ufi->entry->offset,
		    &pg, &npages, 0, VM_PROT_READ, MADV_NORMAL, PGO_SYNCIO);
		/* locked: <nothing> */

		if (error) {
			if (error == EAGAIN) {
				tsleep(&lbolt, PVM, "fltagain2", 0);
				return (0);
			}
			return (-1);
		}

		/*
		 * pgo_get was a success.   attempt to relock everything.
		 */

		locked = uvmfault_relock(ufi);
		if (locked && amap)
			amap_lock(amap);
		simple_lock(&uobj->vmobjlock);

		/*
		 * verify that the page has not be released and re-verify
		 * that amap slot is still free.   if there is a problem we
		 * drop our lock (thus force a lookup refresh/retry).
		 */

		if ((pg->flags & PG_RELEASED) != 0 ||
		    (locked && amap && amap_lookup(&ufi->entry->aref,
		    ufi->orig_rvaddr - ufi->entry->start))) {
			if (locked)
				uvmfault_unlockall(ufi, amap, NULL, NULL);
			locked = FALSE;
		}

		/*
		 * didn't get the lock?   release the page and retry.
		 */

		if (locked == FALSE) {
			if (pg->flags & PG_WANTED) {
				wakeup(pg);
			}
			if (pg->flags & PG_RELEASED) {
				uvm_lock_pageq();
				uvm_pagefree(pg);
				uvm_unlock_pageq();
				return (0);
			}
			uvm_lock_pageq();
			uvm_pageactivate(pg);
			uvm_unlock_pageq();
			pg->flags &= ~(PG_BUSY|PG_WANTED);
			UVM_PAGE_OWN(pg, NULL);
			simple_unlock(&uobj->vmobjlock);
			return (0);
		}
	}

	/*
	 * at this point we have the page we want ("pg") marked PG_BUSY for us
	 * and we have all data structures locked.  do the loanout.  page can
	 * not be PG_RELEASED (we caught this above).
	 */

	if ((flags & UVM_LOAN_TOANON) == 0) {
		uvm_loanpage(&pg, 1);
		**output = pg;
		(*output)++;
		return (1);
	}

	/*
	 * must be a loan to an anon.   check to see if there is already
	 * an anon associated with this page.  if so, then just return
	 * a reference to this object.   the page should already be
	 * mapped read-only because it is already on loan.
	 */

	if (pg->uanon) {
		anon = pg->uanon;
		simple_lock(&anon->an_lock);
		anon->an_ref++;
		simple_unlock(&anon->an_lock);
		if (pg->flags & PG_WANTED) {
			wakeup(pg);
		}
		pg->flags &= ~(PG_WANTED|PG_BUSY);
		UVM_PAGE_OWN(pg, NULL);
		**output = anon;
		(*output)++;
		return (1);
	}

	/*
	 * need to allocate a new anon
	 */

	anon = uvm_analloc();
	if (anon == NULL) {
		if (pg->flags & PG_WANTED) {
			wakeup(pg);
		}
		pg->flags &= ~(PG_WANTED|PG_BUSY);
		UVM_PAGE_OWN(pg, NULL);
		uvmfault_unlockall(ufi, amap, uobj, NULL);
		return (-1);
	}
	anon->u.an_page = pg;
	pg->uanon = anon;
	uvm_lock_pageq();
	if (pg->loan_count == 0) {
		pmap_page_protect(pg, VM_PROT_READ);
	}
	pg->loan_count++;
	uvm_pageactivate(pg);
	uvm_unlock_pageq();
	if (pg->flags & PG_WANTED) {
		wakeup(pg);
	}
	pg->flags &= ~(PG_WANTED|PG_BUSY);
	UVM_PAGE_OWN(pg, NULL);
	simple_unlock(&anon->an_lock);
	**output = anon;
	(*output)++;
	return (1);
}

/*
 * uvm_loanzero: loan a zero-fill page out
 *
 * => called with map, amap, uobj locked
 * => return value:
 *	-1 = fatal error, everything is unlocked, abort.
 *	 0 = lookup in ufi went stale, everything unlocked, relookup and
 *		try again
 *	 1 = got it, everything still locked
 */

static struct uvm_object uvm_loanzero_object;

static int
uvm_loanzero(ufi, output, flags)
	struct uvm_faultinfo *ufi;
	void ***output;
	int flags;
{
	struct vm_anon *anon;
	struct vm_page *pg;
	struct uvm_object *uobj = ufi->entry->object.uvm_obj;
	struct vm_amap *amap = ufi->entry->aref.ar_amap;

again:
	simple_lock(&uvm_loanzero_object.vmobjlock);

	/*
	 * first, get ahold of our single zero page.
	 */

	if (__predict_false((pg =
			     TAILQ_FIRST(&uvm_loanzero_object.memq)) == NULL)) {
		while ((pg = uvm_pagealloc(&uvm_loanzero_object, 0, NULL,
					   UVM_PGA_ZERO)) == NULL) {
			simple_unlock(&uvm_loanzero_object.vmobjlock);
			uvmfault_unlockall(ufi, amap, uobj, NULL);
			uvm_wait("loanzero");
			if (!uvmfault_relock(ufi)) {
				return (0);
			}
			if (amap) {
				amap_lock(amap);
			}
			if (uobj) {
				simple_lock(&uobj->vmobjlock);
			}
			goto again;
		}

		/* got a zero'd page. */
		pg->flags &= ~(PG_WANTED|PG_BUSY|PG_FAKE);
		pg->flags |= PG_RDONLY;
		uvm_lock_pageq();
		uvm_pageactivate(pg);
		uvm_unlock_pageq();
		UVM_PAGE_OWN(pg, NULL);
	}

	if ((flags & UVM_LOAN_TOANON) == 0) {	/* loaning to kernel-page */
		uvm_lock_pageq();
		pg->loan_count++;
		uvm_pagedequeue(pg);
		uvm_unlock_pageq();
		simple_unlock(&uvm_loanzero_object.vmobjlock);
		**output = pg;
		(*output)++;
		return (1);
	}

	/*
	 * loaning to an anon.  check to see if there is already an anon
	 * associated with this page.  if so, then just return a reference
	 * to this object.
	 */

	if (pg->uanon) {
		anon = pg->uanon;
		simple_lock(&anon->an_lock);
		anon->an_ref++;
		simple_unlock(&anon->an_lock);
		simple_unlock(&uvm_loanzero_object.vmobjlock);
		**output = anon;
		(*output)++;
		return (1);
	}

	/*
	 * need to allocate a new anon
	 */

	anon = uvm_analloc();
	if (anon == NULL) {
		/* out of swap causes us to fail */
		simple_unlock(&uvm_loanzero_object.vmobjlock);
		uvmfault_unlockall(ufi, amap, uobj, NULL);
		return (-1);
	}
	anon->u.an_page = pg;
	pg->uanon = anon;
	uvm_lock_pageq();
	pg->loan_count++;
	uvm_pageactivate(pg);
	uvm_unlock_pageq();
	simple_unlock(&uvm_loanzero_object.vmobjlock);
	**output = anon;
	(*output)++;
	return (1);
}


/*
 * uvm_unloananon: kill loans on anons (basically a normal ref drop)
 *
 * => we expect all our resources to be unlocked
 */

static void
uvm_unloananon(aloans, nanons)
	struct vm_anon **aloans;
	int nanons;
{
	struct vm_anon *anon;

	while (nanons-- > 0) {
		int refs;

		anon = *aloans++;
		simple_lock(&anon->an_lock);
		refs = --anon->an_ref;
		simple_unlock(&anon->an_lock);

		if (refs == 0) {
			uvm_anfree(anon);
		}
	}
}

/*
 * uvm_unloanpage: kill loans on pages loaned out to the kernel
 *
 * => we expect all our resources to be unlocked
 */

static void
uvm_unloanpage(ploans, npages)
	struct vm_page **ploans;
	int npages;
{
	struct vm_page *pg;
	struct simplelock *slock;

	uvm_lock_pageq();
	while (npages-- > 0) {
		pg = *ploans++;

		/*
		 * do a little dance to acquire the object or anon lock
		 * as appropriate.  we are locking in the wrong order,
		 * so we have to do a try-lock here.
		 */

		slock = NULL;
		while (pg->uobject != NULL || pg->uanon != NULL) {
			if (pg->uobject != NULL) {
				slock = &pg->uobject->vmobjlock;
			} else {
				slock = &pg->uanon->an_lock;
			}
			if (simple_lock_try(slock)) {
				break;
			}
			uvm_unlock_pageq();
			uvm_lock_pageq();
			slock = NULL;
		}

		/*
		 * drop our loan.  if page is owned by an anon but
		 * PQ_ANON is not set, the page was loaned to the anon
		 * from an object which dropped ownership, so resolve
		 * this by turning the anon's loan into real ownership
		 * (ie. decrement loan_count again and set PQ_ANON).
		 * after all this, if there are no loans left, put the
		 * page back a paging queue (if the page is owned by
		 * an anon) or free it (if the page is now unowned).
		 */

		KASSERT(pg->loan_count > 0);
		pg->loan_count--;
		if (pg->uobject == NULL && pg->uanon != NULL &&
		    (pg->pqflags & PQ_ANON) == 0) {
			KASSERT(pg->loan_count > 0);
			pg->loan_count--;
			pg->pqflags |= PQ_ANON;
		}
		if (pg->loan_count == 0) {
			if (pg->uobject == NULL && pg->uanon == NULL) {
				KASSERT((pg->flags & PG_BUSY) == 0);
				uvm_pagefree(pg);
			} else {
				uvm_pageactivate(pg);
			}
		} else if (pg->loan_count == 1 && pg->uobject != NULL &&
			   pg->uanon != NULL) {
			uvm_pageactivate(pg);
		}
		if (slock != NULL) {
			simple_unlock(slock);
		}
	}
	uvm_unlock_pageq();
}

/*
 * uvm_unloan: kill loans on pages or anons.
 */

void
uvm_unloan(void *v, int npages, int flags)
{
	if (flags & UVM_LOAN_TOANON) {
		uvm_unloananon(v, npages);
	} else {
		uvm_unloanpage(v, npages);
	}
}

/*
 * Minimal pager for uvm_loanzero_object.  We need to provide a "put"
 * method, because the page can end up on a paging queue, and the
 * page daemon will want to call pgo_put when it encounters the page
 * on the inactive list.
 */

static int
ulz_put(struct uvm_object *uobj, voff_t start, voff_t stop, int flags)
{
	struct vm_page *pg;

	KDASSERT(uobj == &uvm_loanzero_object);

	/*
	 * Don't need to do any work here if we're not freeing pages.
	 */

	if ((flags & PGO_FREE) == 0) {
		simple_unlock(&uobj->vmobjlock);
		return 0;
	}

	/*
	 * we don't actually want to ever free the uvm_loanzero_page, so
	 * just reactivate or dequeue it.
	 */

	pg = TAILQ_FIRST(&uobj->memq);
	KASSERT(pg != NULL);
	KASSERT(TAILQ_NEXT(pg, listq) == NULL);

	uvm_lock_pageq();
	if (pg->uanon)
		uvm_pageactivate(pg);
	else
		uvm_pagedequeue(pg);
	uvm_unlock_pageq();

	simple_unlock(&uobj->vmobjlock);
	return 0;
}

static struct uvm_pagerops ulz_pager = {
	NULL,		/* init */
	NULL,		/* reference */
	NULL,		/* detach */
	NULL,		/* fault */
	NULL,		/* get */
	ulz_put,	/* put */
};

/*
 * uvm_loan_init(): initialize the uvm_loan() facility.
 */

void
uvm_loan_init(void)
{

	simple_lock_init(&uvm_loanzero_object.vmobjlock);
	TAILQ_INIT(&uvm_loanzero_object.memq);
	uvm_loanzero_object.pgops = &ulz_pager;
}

/*
 * uvm_loanbreak: break loan on a uobj page
 *
 * => called with uobj locked
 * => the page should be busy
 * => return value:
 *	newly allocated page if succeeded
 */
struct vm_page *
uvm_loanbreak(struct vm_page *uobjpage)
{
	struct vm_page *pg;
#ifdef DIAGNOSTIC
	struct uvm_object *uobj = uobjpage->uobject;
#endif

	KASSERT(uobj != NULL);
	LOCK_ASSERT(simple_lock_held(&uobj->vmobjlock));
	KASSERT(uobjpage->flags & PG_BUSY);

	/* alloc new un-owned page */
	pg = uvm_pagealloc(NULL, 0, NULL, 0);
	if (pg == NULL)
		return NULL;

	/*
	 * copy the data from the old page to the new
	 * one and clear the fake/clean flags on the
	 * new page (keep it busy).  force a reload
	 * of the old page by clearing it from all
	 * pmaps.  then lock the page queues to
	 * rename the pages.
	 */

	uvm_pagecopy(uobjpage, pg);	/* old -> new */
	pg->flags &= ~(PG_FAKE|PG_CLEAN);
	pmap_page_protect(uobjpage, VM_PROT_NONE);
	if (uobjpage->flags & PG_WANTED)
		wakeup(uobjpage);
	/* uobj still locked */
	uobjpage->flags &= ~(PG_WANTED|PG_BUSY);
	UVM_PAGE_OWN(uobjpage, NULL);

	uvm_lock_pageq();

	/*
	 * replace uobjpage with new page.
	 */

	uvm_pagereplace(uobjpage, pg);

	/*
	 * if the page is no longer referenced by
	 * an anon (i.e. we are breaking an O->K
	 * loan), then remove it from any pageq's.
	 */
	if (uobjpage->uanon == NULL)
		uvm_pagedequeue(uobjpage);

	/*
	 * at this point we have absolutely no
	 * control over uobjpage
	 */

	/* install new page */
	uvm_pageactivate(pg);
	uvm_unlock_pageq();

	/*
	 * done!  loan is broken and "pg" is
	 * PG_BUSY.   it can now replace uobjpage.
	 */

	return pg;
}