vm_object.c revision 121866
1/*
2 * Copyright (c) 1991, 1993
3 *	The Regents of the University of California.  All rights reserved.
4 *
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 *    notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 *    notice, this list of conditions and the following disclaimer in the
15 *    documentation and/or other materials provided with the distribution.
16 * 3. All advertising materials mentioning features or use of this software
17 *    must display the following acknowledgement:
18 *	This product includes software developed by the University of
19 *	California, Berkeley and its contributors.
20 * 4. Neither the name of the University nor the names of its contributors
21 *    may be used to endorse or promote products derived from this software
22 *    without specific prior written permission.
23 *
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * SUCH DAMAGE.
35 *
36 *	from: @(#)vm_object.c	8.5 (Berkeley) 3/22/94
37 *
38 *
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
41 *
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43 *
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
49 *
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53 *
54 * Carnegie Mellon requests users of this software to return to
55 *
56 *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57 *  School of Computer Science
58 *  Carnegie Mellon University
59 *  Pittsburgh PA 15213-3890
60 *
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
63 */
64
65/*
66 *	Virtual memory object module.
67 */
68
69#include <sys/cdefs.h>
70__FBSDID("$FreeBSD: head/sys/vm/vm_object.c 121866 2003-11-01 23:06:41Z alc $");
71
72#include <sys/param.h>
73#include <sys/systm.h>
74#include <sys/lock.h>
75#include <sys/mman.h>
76#include <sys/mount.h>
77#include <sys/kernel.h>
78#include <sys/sysctl.h>
79#include <sys/mutex.h>
80#include <sys/proc.h>		/* for curproc, pageproc */
81#include <sys/socket.h>
82#include <sys/vnode.h>
83#include <sys/vmmeter.h>
84#include <sys/sx.h>
85
86#include <vm/vm.h>
87#include <vm/vm_param.h>
88#include <vm/pmap.h>
89#include <vm/vm_map.h>
90#include <vm/vm_object.h>
91#include <vm/vm_page.h>
92#include <vm/vm_pageout.h>
93#include <vm/vm_pager.h>
94#include <vm/swap_pager.h>
95#include <vm/vm_kern.h>
96#include <vm/vm_extern.h>
97#include <vm/uma.h>
98
99#define EASY_SCAN_FACTOR       8
100
101#define MSYNC_FLUSH_HARDSEQ	0x01
102#define MSYNC_FLUSH_SOFTSEQ	0x02
103
104/*
105 * msync / VM object flushing optimizations
106 */
107static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ;
108SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags,
109        CTLFLAG_RW, &msync_flush_flags, 0, "");
110
111static void	vm_object_qcollapse(vm_object_t object);
112static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
113
114/*
115 *	Virtual memory objects maintain the actual data
116 *	associated with allocated virtual memory.  A given
117 *	page of memory exists within exactly one object.
118 *
119 *	An object is only deallocated when all "references"
120 *	are given up.  Only one "reference" to a given
121 *	region of an object should be writeable.
122 *
123 *	Associated with each object is a list of all resident
124 *	memory pages belonging to that object; this list is
125 *	maintained by the "vm_page" module, and locked by the object's
126 *	lock.
127 *
128 *	Each object also records a "pager" routine which is
129 *	used to retrieve (and store) pages to the proper backing
130 *	storage.  In addition, objects may be backed by other
131 *	objects from which they were virtual-copied.
132 *
133 *	The only items within the object structure which are
134 *	modified after time of creation are:
135 *		reference count		locked by object's lock
136 *		pager routine		locked by object's lock
137 *
138 */
139
140struct object_q vm_object_list;
141struct mtx vm_object_list_mtx;	/* lock for object list and count */
142
143struct vm_object kernel_object_store;
144struct vm_object kmem_object_store;
145
146static long object_collapses;
147static long object_bypasses;
148static int next_index;
149static uma_zone_t obj_zone;
150#define VM_OBJECTS_INIT 256
151
152static void vm_object_zinit(void *mem, int size);
153
154#ifdef INVARIANTS
155static void vm_object_zdtor(void *mem, int size, void *arg);
156
157static void
158vm_object_zdtor(void *mem, int size, void *arg)
159{
160	vm_object_t object;
161
162	object = (vm_object_t)mem;
163	KASSERT(TAILQ_EMPTY(&object->memq),
164	    ("object %p has resident pages",
165	    object));
166	KASSERT(object->paging_in_progress == 0,
167	    ("object %p paging_in_progress = %d",
168	    object, object->paging_in_progress));
169	KASSERT(object->resident_page_count == 0,
170	    ("object %p resident_page_count = %d",
171	    object, object->resident_page_count));
172	KASSERT(object->shadow_count == 0,
173	    ("object %p shadow_count = %d",
174	    object, object->shadow_count));
175}
176#endif
177
178static void
179vm_object_zinit(void *mem, int size)
180{
181	vm_object_t object;
182
183	object = (vm_object_t)mem;
184	bzero(&object->mtx, sizeof(object->mtx));
185	VM_OBJECT_LOCK_INIT(object);
186
187	/* These are true for any object that has been freed */
188	object->paging_in_progress = 0;
189	object->resident_page_count = 0;
190	object->shadow_count = 0;
191}
192
193void
194_vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
195{
196	int incr;
197
198	TAILQ_INIT(&object->memq);
199	LIST_INIT(&object->shadow_head);
200
201	object->root = NULL;
202	object->type = type;
203	object->size = size;
204	object->generation = 1;
205	object->ref_count = 1;
206	object->flags = 0;
207	if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
208		object->flags = OBJ_ONEMAPPING;
209	if (size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
210		incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
211	else
212		incr = size;
213	do
214		object->pg_color = next_index;
215	while (!atomic_cmpset_int(&next_index, object->pg_color,
216				  (object->pg_color + incr) & PQ_L2_MASK));
217	object->handle = NULL;
218	object->backing_object = NULL;
219	object->backing_object_offset = (vm_ooffset_t) 0;
220
221	mtx_lock(&vm_object_list_mtx);
222	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
223	mtx_unlock(&vm_object_list_mtx);
224}
225
226/*
227 *	vm_object_init:
228 *
229 *	Initialize the VM objects module.
230 */
231void
232vm_object_init(void)
233{
234	TAILQ_INIT(&vm_object_list);
235	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
236
237	VM_OBJECT_LOCK_INIT(&kernel_object_store);
238	_vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
239	    kernel_object);
240
241	/*
242	 * The kmem object's mutex is given a unique name, instead of
243	 * "vm object", to avoid false reports of lock-order reversal
244	 * with a system map mutex.
245	 */
246	mtx_init(VM_OBJECT_MTX(kmem_object), "kmem object", NULL, MTX_DEF);
247	_vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
248	    kmem_object);
249
250	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
251#ifdef INVARIANTS
252	    vm_object_zdtor,
253#else
254	    NULL,
255#endif
256	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
257	uma_prealloc(obj_zone, VM_OBJECTS_INIT);
258}
259
260void
261vm_object_clear_flag(vm_object_t object, u_short bits)
262{
263
264	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
265	object->flags &= ~bits;
266}
267
268void
269vm_object_pip_add(vm_object_t object, short i)
270{
271
272	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
273	object->paging_in_progress += i;
274}
275
276void
277vm_object_pip_subtract(vm_object_t object, short i)
278{
279
280	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
281	object->paging_in_progress -= i;
282}
283
284void
285vm_object_pip_wakeup(vm_object_t object)
286{
287
288	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
289	object->paging_in_progress--;
290	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
291		vm_object_clear_flag(object, OBJ_PIPWNT);
292		wakeup(object);
293	}
294}
295
296void
297vm_object_pip_wakeupn(vm_object_t object, short i)
298{
299
300	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
301	if (i)
302		object->paging_in_progress -= i;
303	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
304		vm_object_clear_flag(object, OBJ_PIPWNT);
305		wakeup(object);
306	}
307}
308
309void
310vm_object_pip_wait(vm_object_t object, char *waitid)
311{
312
313	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
314	while (object->paging_in_progress) {
315		object->flags |= OBJ_PIPWNT;
316		msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
317	}
318}
319
320/*
321 *	vm_object_allocate_wait
322 *
323 *	Return a new object with the given size, and give the user the
324 *	option of waiting for it to complete or failing if the needed
325 *	memory isn't available.
326 */
327vm_object_t
328vm_object_allocate_wait(objtype_t type, vm_pindex_t size, int flags)
329{
330	vm_object_t result;
331
332	result = (vm_object_t) uma_zalloc(obj_zone, flags);
333
334	if (result != NULL)
335		_vm_object_allocate(type, size, result);
336
337	return (result);
338}
339
340/*
341 *	vm_object_allocate:
342 *
343 *	Returns a new object with the given size.
344 */
345vm_object_t
346vm_object_allocate(objtype_t type, vm_pindex_t size)
347{
348	return(vm_object_allocate_wait(type, size, M_WAITOK));
349}
350
351
352/*
353 *	vm_object_reference:
354 *
355 *	Gets another reference to the given object.  Note: OBJ_DEAD
356 *	objects can be referenced during final cleaning.
357 */
358void
359vm_object_reference(vm_object_t object)
360{
361	struct vnode *vp;
362	int flags;
363
364	if (object == NULL)
365		return;
366	VM_OBJECT_LOCK(object);
367	object->ref_count++;
368	if (object->type == OBJT_VNODE) {
369		vp = object->handle;
370		VI_LOCK(vp);
371		VM_OBJECT_UNLOCK(object);
372		for (flags = LK_INTERLOCK; vget(vp, flags, curthread);
373		     flags = 0)
374			printf("vm_object_reference: delay in vget\n");
375	} else
376		VM_OBJECT_UNLOCK(object);
377}
378
379/*
380 * Handle deallocating an object of type OBJT_VNODE.
381 */
382void
383vm_object_vndeallocate(vm_object_t object)
384{
385	struct vnode *vp = (struct vnode *) object->handle;
386
387	GIANT_REQUIRED;
388	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
389	KASSERT(object->type == OBJT_VNODE,
390	    ("vm_object_vndeallocate: not a vnode object"));
391	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
392#ifdef INVARIANTS
393	if (object->ref_count == 0) {
394		vprint("vm_object_vndeallocate", vp);
395		panic("vm_object_vndeallocate: bad object reference count");
396	}
397#endif
398
399	object->ref_count--;
400	if (object->ref_count == 0) {
401		mp_fixme("Unlocked vflag access.");
402		vp->v_vflag &= ~VV_TEXT;
403	}
404	VM_OBJECT_UNLOCK(object);
405	/*
406	 * vrele may need a vop lock
407	 */
408	vrele(vp);
409}
410
411/*
412 *	vm_object_deallocate:
413 *
414 *	Release a reference to the specified object,
415 *	gained either through a vm_object_allocate
416 *	or a vm_object_reference call.  When all references
417 *	are gone, storage associated with this object
418 *	may be relinquished.
419 *
420 *	No object may be locked.
421 */
422void
423vm_object_deallocate(vm_object_t object)
424{
425	vm_object_t temp;
426
427	if (object != kmem_object)
428		mtx_lock(&Giant);
429	while (object != NULL) {
430		VM_OBJECT_LOCK(object);
431		if (object->type == OBJT_VNODE) {
432			vm_object_vndeallocate(object);
433			goto done;
434		}
435
436		KASSERT(object->ref_count != 0,
437			("vm_object_deallocate: object deallocated too many times: %d", object->type));
438
439		/*
440		 * If the reference count goes to 0 we start calling
441		 * vm_object_terminate() on the object chain.
442		 * A ref count of 1 may be a special case depending on the
443		 * shadow count being 0 or 1.
444		 */
445		object->ref_count--;
446		if (object->ref_count > 1) {
447			VM_OBJECT_UNLOCK(object);
448			goto done;
449		} else if (object->ref_count == 1) {
450			if (object->shadow_count == 0) {
451				vm_object_set_flag(object, OBJ_ONEMAPPING);
452			} else if ((object->shadow_count == 1) &&
453			    (object->handle == NULL) &&
454			    (object->type == OBJT_DEFAULT ||
455			     object->type == OBJT_SWAP)) {
456				vm_object_t robject;
457
458				robject = LIST_FIRST(&object->shadow_head);
459				KASSERT(robject != NULL,
460				    ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
461					 object->ref_count,
462					 object->shadow_count));
463				if (!VM_OBJECT_TRYLOCK(robject)) {
464					/*
465					 * Avoid a potential deadlock.
466					 */
467					object->ref_count++;
468					VM_OBJECT_UNLOCK(object);
469					continue;
470				}
471				if ((robject->handle == NULL) &&
472				    (robject->type == OBJT_DEFAULT ||
473				     robject->type == OBJT_SWAP)) {
474
475					robject->ref_count++;
476retry:
477					if (robject->paging_in_progress) {
478						VM_OBJECT_UNLOCK(object);
479						vm_object_pip_wait(robject,
480						    "objde1");
481						VM_OBJECT_LOCK(object);
482						goto retry;
483					} else if (object->paging_in_progress) {
484						VM_OBJECT_UNLOCK(robject);
485						object->flags |= OBJ_PIPWNT;
486						msleep(object,
487						    VM_OBJECT_MTX(object),
488						    PDROP | PVM, "objde2", 0);
489						VM_OBJECT_LOCK(robject);
490						VM_OBJECT_LOCK(object);
491						goto retry;
492					}
493					VM_OBJECT_UNLOCK(object);
494					if (robject->ref_count == 1) {
495						robject->ref_count--;
496						object = robject;
497						goto doterm;
498					}
499					object = robject;
500					vm_object_collapse(object);
501					VM_OBJECT_UNLOCK(object);
502					continue;
503				}
504				VM_OBJECT_UNLOCK(robject);
505			}
506			VM_OBJECT_UNLOCK(object);
507			goto done;
508		}
509doterm:
510		temp = object->backing_object;
511		if (temp != NULL) {
512			VM_OBJECT_LOCK(temp);
513			LIST_REMOVE(object, shadow_list);
514			temp->shadow_count--;
515			temp->generation++;
516			VM_OBJECT_UNLOCK(temp);
517			object->backing_object = NULL;
518		}
519		/*
520		 * Don't double-terminate, we could be in a termination
521		 * recursion due to the terminate having to sync data
522		 * to disk.
523		 */
524		if ((object->flags & OBJ_DEAD) == 0)
525			vm_object_terminate(object);
526		else
527			VM_OBJECT_UNLOCK(object);
528		object = temp;
529	}
530done:
531	if (object != kmem_object)
532		mtx_unlock(&Giant);
533}
534
535/*
536 *	vm_object_terminate actually destroys the specified object, freeing
537 *	up all previously used resources.
538 *
539 *	The object must be locked.
540 *	This routine may block.
541 */
542void
543vm_object_terminate(vm_object_t object)
544{
545	vm_page_t p;
546	int s;
547
548	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
549
550	/*
551	 * Make sure no one uses us.
552	 */
553	vm_object_set_flag(object, OBJ_DEAD);
554
555	/*
556	 * wait for the pageout daemon to be done with the object
557	 */
558	vm_object_pip_wait(object, "objtrm");
559
560	KASSERT(!object->paging_in_progress,
561		("vm_object_terminate: pageout in progress"));
562
563	/*
564	 * Clean and free the pages, as appropriate. All references to the
565	 * object are gone, so we don't need to lock it.
566	 */
567	if (object->type == OBJT_VNODE) {
568		struct vnode *vp = (struct vnode *)object->handle;
569
570		/*
571		 * Clean pages and flush buffers.
572		 */
573		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
574		VM_OBJECT_UNLOCK(object);
575
576		vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0);
577
578		VM_OBJECT_LOCK(object);
579	}
580
581	KASSERT(object->ref_count == 0,
582		("vm_object_terminate: object with references, ref_count=%d",
583		object->ref_count));
584
585	/*
586	 * Now free any remaining pages. For internal objects, this also
587	 * removes them from paging queues. Don't free wired pages, just
588	 * remove them from the object.
589	 */
590	s = splvm();
591	vm_page_lock_queues();
592	while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
593		KASSERT(!p->busy && (p->flags & PG_BUSY) == 0,
594			("vm_object_terminate: freeing busy page %p "
595			"p->busy = %d, p->flags %x\n", p, p->busy, p->flags));
596		if (p->wire_count == 0) {
597			vm_page_busy(p);
598			vm_page_free(p);
599			cnt.v_pfree++;
600		} else {
601			vm_page_busy(p);
602			vm_page_remove(p);
603		}
604	}
605	vm_page_unlock_queues();
606	splx(s);
607
608	/*
609	 * Let the pager know object is dead.
610	 */
611	vm_pager_deallocate(object);
612	VM_OBJECT_UNLOCK(object);
613
614	/*
615	 * Remove the object from the global object list.
616	 */
617	mtx_lock(&vm_object_list_mtx);
618	TAILQ_REMOVE(&vm_object_list, object, object_list);
619	mtx_unlock(&vm_object_list_mtx);
620
621	wakeup(object);
622
623	/*
624	 * Free the space for the object.
625	 */
626	uma_zfree(obj_zone, object);
627}
628
629/*
630 *	vm_object_page_clean
631 *
632 *	Clean all dirty pages in the specified range of object.  Leaves page
633 * 	on whatever queue it is currently on.   If NOSYNC is set then do not
634 *	write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
635 *	leaving the object dirty.
636 *
637 *	When stuffing pages asynchronously, allow clustering.  XXX we need a
638 *	synchronous clustering mode implementation.
639 *
640 *	Odd semantics: if start == end, we clean everything.
641 *
642 *	The object must be locked.
643 */
644void
645vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags)
646{
647	vm_page_t p, np;
648	vm_pindex_t tstart, tend;
649	vm_pindex_t pi;
650	int clearobjflags;
651	int pagerflags;
652	int curgeneration;
653
654	GIANT_REQUIRED;
655	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
656	if (object->type != OBJT_VNODE ||
657		(object->flags & OBJ_MIGHTBEDIRTY) == 0)
658		return;
659
660	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
661	pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
662
663	vm_object_set_flag(object, OBJ_CLEANING);
664
665	tstart = start;
666	if (end == 0) {
667		tend = object->size;
668	} else {
669		tend = end;
670	}
671
672	vm_page_lock_queues();
673	/*
674	 * If the caller is smart and only msync()s a range he knows is
675	 * dirty, we may be able to avoid an object scan.  This results in
676	 * a phenominal improvement in performance.  We cannot do this
677	 * as a matter of course because the object may be huge - e.g.
678	 * the size might be in the gigabytes or terrabytes.
679	 */
680	if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
681		vm_pindex_t tscan;
682		int scanlimit;
683		int scanreset;
684
685		scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
686		if (scanreset < 16)
687			scanreset = 16;
688		pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
689
690		scanlimit = scanreset;
691		tscan = tstart;
692		while (tscan < tend) {
693			curgeneration = object->generation;
694			p = vm_page_lookup(object, tscan);
695			if (p == NULL || p->valid == 0 ||
696			    (p->queue - p->pc) == PQ_CACHE) {
697				if (--scanlimit == 0)
698					break;
699				++tscan;
700				continue;
701			}
702			vm_page_test_dirty(p);
703			if ((p->dirty & p->valid) == 0) {
704				if (--scanlimit == 0)
705					break;
706				++tscan;
707				continue;
708			}
709			/*
710			 * If we have been asked to skip nosync pages and
711			 * this is a nosync page, we can't continue.
712			 */
713			if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
714				if (--scanlimit == 0)
715					break;
716				++tscan;
717				continue;
718			}
719			scanlimit = scanreset;
720
721			/*
722			 * This returns 0 if it was unable to busy the first
723			 * page (i.e. had to sleep).
724			 */
725			tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
726		}
727
728		/*
729		 * If everything was dirty and we flushed it successfully,
730		 * and the requested range is not the entire object, we
731		 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
732		 * return immediately.
733		 */
734		if (tscan >= tend && (tstart || tend < object->size)) {
735			vm_page_unlock_queues();
736			vm_object_clear_flag(object, OBJ_CLEANING);
737			return;
738		}
739		pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
740	}
741
742	/*
743	 * Generally set CLEANCHK interlock and make the page read-only so
744	 * we can then clear the object flags.
745	 *
746	 * However, if this is a nosync mmap then the object is likely to
747	 * stay dirty so do not mess with the page and do not clear the
748	 * object flags.
749	 */
750	clearobjflags = 1;
751	TAILQ_FOREACH(p, &object->memq, listq) {
752		vm_page_flag_set(p, PG_CLEANCHK);
753		if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
754			clearobjflags = 0;
755		else
756			pmap_page_protect(p, VM_PROT_READ);
757	}
758
759	if (clearobjflags && (tstart == 0) && (tend == object->size)) {
760		struct vnode *vp;
761
762		vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
763		if (object->type == OBJT_VNODE &&
764		    (vp = (struct vnode *)object->handle) != NULL) {
765			VI_LOCK(vp);
766			if (vp->v_iflag & VI_OBJDIRTY)
767				vp->v_iflag &= ~VI_OBJDIRTY;
768			VI_UNLOCK(vp);
769		}
770	}
771
772rescan:
773	curgeneration = object->generation;
774
775	for (p = TAILQ_FIRST(&object->memq); p; p = np) {
776		int n;
777
778		np = TAILQ_NEXT(p, listq);
779
780again:
781		pi = p->pindex;
782		if (((p->flags & PG_CLEANCHK) == 0) ||
783			(pi < tstart) || (pi >= tend) ||
784			(p->valid == 0) ||
785			((p->queue - p->pc) == PQ_CACHE)) {
786			vm_page_flag_clear(p, PG_CLEANCHK);
787			continue;
788		}
789
790		vm_page_test_dirty(p);
791		if ((p->dirty & p->valid) == 0) {
792			vm_page_flag_clear(p, PG_CLEANCHK);
793			continue;
794		}
795
796		/*
797		 * If we have been asked to skip nosync pages and this is a
798		 * nosync page, skip it.  Note that the object flags were
799		 * not cleared in this case so we do not have to set them.
800		 */
801		if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
802			vm_page_flag_clear(p, PG_CLEANCHK);
803			continue;
804		}
805
806		n = vm_object_page_collect_flush(object, p,
807			curgeneration, pagerflags);
808		if (n == 0)
809			goto rescan;
810
811		if (object->generation != curgeneration)
812			goto rescan;
813
814		/*
815		 * Try to optimize the next page.  If we can't we pick up
816		 * our (random) scan where we left off.
817		 */
818		if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
819			if ((p = vm_page_lookup(object, pi + n)) != NULL)
820				goto again;
821		}
822	}
823	vm_page_unlock_queues();
824#if 0
825	VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
826#endif
827
828	vm_object_clear_flag(object, OBJ_CLEANING);
829	return;
830}
831
832static int
833vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
834{
835	int runlen;
836	int s;
837	int maxf;
838	int chkb;
839	int maxb;
840	int i;
841	vm_pindex_t pi;
842	vm_page_t maf[vm_pageout_page_count];
843	vm_page_t mab[vm_pageout_page_count];
844	vm_page_t ma[vm_pageout_page_count];
845
846	s = splvm();
847	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
848	pi = p->pindex;
849	while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
850		vm_page_lock_queues();
851		if (object->generation != curgeneration) {
852			splx(s);
853			return(0);
854		}
855	}
856	maxf = 0;
857	for(i = 1; i < vm_pageout_page_count; i++) {
858		vm_page_t tp;
859
860		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
861			if ((tp->flags & PG_BUSY) ||
862				((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
863				 (tp->flags & PG_CLEANCHK) == 0) ||
864				(tp->busy != 0))
865				break;
866			if((tp->queue - tp->pc) == PQ_CACHE) {
867				vm_page_flag_clear(tp, PG_CLEANCHK);
868				break;
869			}
870			vm_page_test_dirty(tp);
871			if ((tp->dirty & tp->valid) == 0) {
872				vm_page_flag_clear(tp, PG_CLEANCHK);
873				break;
874			}
875			maf[ i - 1 ] = tp;
876			maxf++;
877			continue;
878		}
879		break;
880	}
881
882	maxb = 0;
883	chkb = vm_pageout_page_count -  maxf;
884	if (chkb) {
885		for(i = 1; i < chkb;i++) {
886			vm_page_t tp;
887
888			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
889				if ((tp->flags & PG_BUSY) ||
890					((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
891					 (tp->flags & PG_CLEANCHK) == 0) ||
892					(tp->busy != 0))
893					break;
894				if ((tp->queue - tp->pc) == PQ_CACHE) {
895					vm_page_flag_clear(tp, PG_CLEANCHK);
896					break;
897				}
898				vm_page_test_dirty(tp);
899				if ((tp->dirty & tp->valid) == 0) {
900					vm_page_flag_clear(tp, PG_CLEANCHK);
901					break;
902				}
903				mab[ i - 1 ] = tp;
904				maxb++;
905				continue;
906			}
907			break;
908		}
909	}
910
911	for(i = 0; i < maxb; i++) {
912		int index = (maxb - i) - 1;
913		ma[index] = mab[i];
914		vm_page_flag_clear(ma[index], PG_CLEANCHK);
915	}
916	vm_page_flag_clear(p, PG_CLEANCHK);
917	ma[maxb] = p;
918	for(i = 0; i < maxf; i++) {
919		int index = (maxb + i) + 1;
920		ma[index] = maf[i];
921		vm_page_flag_clear(ma[index], PG_CLEANCHK);
922	}
923	runlen = maxb + maxf + 1;
924
925	splx(s);
926	vm_pageout_flush(ma, runlen, pagerflags);
927	for (i = 0; i < runlen; i++) {
928		if (ma[i]->valid & ma[i]->dirty) {
929			pmap_page_protect(ma[i], VM_PROT_READ);
930			vm_page_flag_set(ma[i], PG_CLEANCHK);
931
932			/*
933			 * maxf will end up being the actual number of pages
934			 * we wrote out contiguously, non-inclusive of the
935			 * first page.  We do not count look-behind pages.
936			 */
937			if (i >= maxb + 1 && (maxf > i - maxb - 1))
938				maxf = i - maxb - 1;
939		}
940	}
941	return(maxf + 1);
942}
943
944/*
945 *	vm_object_madvise:
946 *
947 *	Implements the madvise function at the object/page level.
948 *
949 *	MADV_WILLNEED	(any object)
950 *
951 *	    Activate the specified pages if they are resident.
952 *
953 *	MADV_DONTNEED	(any object)
954 *
955 *	    Deactivate the specified pages if they are resident.
956 *
957 *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
958 *			 OBJ_ONEMAPPING only)
959 *
960 *	    Deactivate and clean the specified pages if they are
961 *	    resident.  This permits the process to reuse the pages
962 *	    without faulting or the kernel to reclaim the pages
963 *	    without I/O.
964 */
965void
966vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
967{
968	vm_pindex_t end, tpindex;
969	vm_object_t backing_object, tobject;
970	vm_page_t m;
971
972	if (object == NULL)
973		return;
974	end = pindex + count;
975	/*
976	 * Locate and adjust resident pages
977	 */
978	for (; pindex < end; pindex += 1) {
979relookup:
980		tobject = object;
981		tpindex = pindex;
982		VM_OBJECT_LOCK(tobject);
983shadowlookup:
984		/*
985		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
986		 * and those pages must be OBJ_ONEMAPPING.
987		 */
988		if (advise == MADV_FREE) {
989			if ((tobject->type != OBJT_DEFAULT &&
990			     tobject->type != OBJT_SWAP) ||
991			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
992				goto unlock_tobject;
993			}
994		}
995		m = vm_page_lookup(tobject, tpindex);
996		if (m == NULL) {
997			/*
998			 * There may be swap even if there is no backing page
999			 */
1000			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1001				swap_pager_freespace(tobject, tpindex, 1);
1002			/*
1003			 * next object
1004			 */
1005			backing_object = tobject->backing_object;
1006			if (backing_object == NULL)
1007				goto unlock_tobject;
1008			VM_OBJECT_LOCK(backing_object);
1009			VM_OBJECT_UNLOCK(tobject);
1010			tobject = backing_object;
1011			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1012			goto shadowlookup;
1013		}
1014		/*
1015		 * If the page is busy or not in a normal active state,
1016		 * we skip it.  If the page is not managed there are no
1017		 * page queues to mess with.  Things can break if we mess
1018		 * with pages in any of the below states.
1019		 */
1020		vm_page_lock_queues();
1021		if (m->hold_count ||
1022		    m->wire_count ||
1023		    (m->flags & PG_UNMANAGED) ||
1024		    m->valid != VM_PAGE_BITS_ALL) {
1025			vm_page_unlock_queues();
1026			goto unlock_tobject;
1027		}
1028 		if (vm_page_sleep_if_busy(m, TRUE, "madvpo")) {
1029			VM_OBJECT_UNLOCK(tobject);
1030  			goto relookup;
1031		}
1032		if (advise == MADV_WILLNEED) {
1033			vm_page_activate(m);
1034		} else if (advise == MADV_DONTNEED) {
1035			vm_page_dontneed(m);
1036		} else if (advise == MADV_FREE) {
1037			/*
1038			 * Mark the page clean.  This will allow the page
1039			 * to be freed up by the system.  However, such pages
1040			 * are often reused quickly by malloc()/free()
1041			 * so we do not do anything that would cause
1042			 * a page fault if we can help it.
1043			 *
1044			 * Specifically, we do not try to actually free
1045			 * the page now nor do we try to put it in the
1046			 * cache (which would cause a page fault on reuse).
1047			 *
1048			 * But we do make the page is freeable as we
1049			 * can without actually taking the step of unmapping
1050			 * it.
1051			 */
1052			pmap_clear_modify(m);
1053			m->dirty = 0;
1054			m->act_count = 0;
1055			vm_page_dontneed(m);
1056		}
1057		vm_page_unlock_queues();
1058		if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1059			swap_pager_freespace(tobject, tpindex, 1);
1060unlock_tobject:
1061		VM_OBJECT_UNLOCK(tobject);
1062	}
1063}
1064
1065/*
1066 *	vm_object_shadow:
1067 *
1068 *	Create a new object which is backed by the
1069 *	specified existing object range.  The source
1070 *	object reference is deallocated.
1071 *
1072 *	The new object and offset into that object
1073 *	are returned in the source parameters.
1074 */
1075void
1076vm_object_shadow(
1077	vm_object_t *object,	/* IN/OUT */
1078	vm_ooffset_t *offset,	/* IN/OUT */
1079	vm_size_t length)
1080{
1081	vm_object_t source;
1082	vm_object_t result;
1083
1084	source = *object;
1085
1086	/*
1087	 * Don't create the new object if the old object isn't shared.
1088	 */
1089	if (source != NULL) {
1090		VM_OBJECT_LOCK(source);
1091		if (source->ref_count == 1 &&
1092		    source->handle == NULL &&
1093		    (source->type == OBJT_DEFAULT ||
1094		     source->type == OBJT_SWAP)) {
1095			VM_OBJECT_UNLOCK(source);
1096			return;
1097		}
1098		VM_OBJECT_UNLOCK(source);
1099	}
1100
1101	/*
1102	 * Allocate a new object with the given length.
1103	 */
1104	result = vm_object_allocate(OBJT_DEFAULT, length);
1105
1106	/*
1107	 * The new object shadows the source object, adding a reference to it.
1108	 * Our caller changes his reference to point to the new object,
1109	 * removing a reference to the source object.  Net result: no change
1110	 * of reference count.
1111	 *
1112	 * Try to optimize the result object's page color when shadowing
1113	 * in order to maintain page coloring consistency in the combined
1114	 * shadowed object.
1115	 */
1116	result->backing_object = source;
1117	if (source != NULL) {
1118		VM_OBJECT_LOCK(source);
1119		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1120		source->shadow_count++;
1121		source->generation++;
1122		if (length < source->size)
1123			length = source->size;
1124		if (length > PQ_L2_SIZE / 3 + PQ_PRIME1 ||
1125		    source->generation > 1)
1126			length = PQ_L2_SIZE / 3 + PQ_PRIME1;
1127		result->pg_color = (source->pg_color +
1128		    length * source->generation) & PQ_L2_MASK;
1129		VM_OBJECT_UNLOCK(source);
1130		next_index = (result->pg_color + PQ_L2_SIZE / 3 + PQ_PRIME1) &
1131		    PQ_L2_MASK;
1132	}
1133
1134	/*
1135	 * Store the offset into the source object, and fix up the offset into
1136	 * the new object.
1137	 */
1138	result->backing_object_offset = *offset;
1139
1140	/*
1141	 * Return the new things
1142	 */
1143	*offset = 0;
1144	*object = result;
1145}
1146
1147/*
1148 *	vm_object_split:
1149 *
1150 * Split the pages in a map entry into a new object.  This affords
1151 * easier removal of unused pages, and keeps object inheritance from
1152 * being a negative impact on memory usage.
1153 */
1154void
1155vm_object_split(vm_map_entry_t entry)
1156{
1157	vm_page_t m;
1158	vm_object_t orig_object, new_object, source;
1159	vm_offset_t s, e;
1160	vm_pindex_t offidxstart, offidxend;
1161	vm_size_t idx, size;
1162	vm_ooffset_t offset;
1163
1164	GIANT_REQUIRED;
1165
1166	orig_object = entry->object.vm_object;
1167	if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1168		return;
1169	if (orig_object->ref_count <= 1)
1170		return;
1171
1172	offset = entry->offset;
1173	s = entry->start;
1174	e = entry->end;
1175
1176	offidxstart = OFF_TO_IDX(offset);
1177	offidxend = offidxstart + OFF_TO_IDX(e - s);
1178	size = offidxend - offidxstart;
1179
1180	new_object = vm_pager_allocate(orig_object->type,
1181		NULL, IDX_TO_OFF(size), VM_PROT_ALL, 0LL);
1182	if (new_object == NULL)
1183		return;
1184
1185	source = orig_object->backing_object;
1186	if (source != NULL) {
1187		vm_object_reference(source);	/* Referenced by new_object */
1188		VM_OBJECT_LOCK(source);
1189		LIST_INSERT_HEAD(&source->shadow_head,
1190				  new_object, shadow_list);
1191		source->shadow_count++;
1192		source->generation++;
1193		vm_object_clear_flag(source, OBJ_ONEMAPPING);
1194		VM_OBJECT_UNLOCK(source);
1195		new_object->backing_object_offset =
1196			orig_object->backing_object_offset + offset;
1197		new_object->backing_object = source;
1198	}
1199	VM_OBJECT_LOCK(new_object);
1200	VM_OBJECT_LOCK(orig_object);
1201	for (idx = 0; idx < size; idx++) {
1202	retry:
1203		m = vm_page_lookup(orig_object, offidxstart + idx);
1204		if (m == NULL)
1205			continue;
1206
1207		/*
1208		 * We must wait for pending I/O to complete before we can
1209		 * rename the page.
1210		 *
1211		 * We do not have to VM_PROT_NONE the page as mappings should
1212		 * not be changed by this operation.
1213		 */
1214		vm_page_lock_queues();
1215		if ((m->flags & PG_BUSY) || m->busy) {
1216			vm_page_flag_set(m, PG_WANTED | PG_REFERENCED);
1217			VM_OBJECT_UNLOCK(orig_object);
1218			VM_OBJECT_UNLOCK(new_object);
1219			msleep(m, &vm_page_queue_mtx, PDROP | PVM, "spltwt", 0);
1220			VM_OBJECT_LOCK(new_object);
1221			VM_OBJECT_LOCK(orig_object);
1222			goto retry;
1223		}
1224		vm_page_busy(m);
1225		vm_page_rename(m, new_object, idx);
1226		/* page automatically made dirty by rename and cache handled */
1227		vm_page_busy(m);
1228		vm_page_unlock_queues();
1229	}
1230	if (orig_object->type == OBJT_SWAP) {
1231		/*
1232		 * swap_pager_copy() can sleep, in which case the orig_object's
1233		 * and new_object's locks are released and reacquired.
1234		 */
1235		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1236	}
1237	VM_OBJECT_UNLOCK(orig_object);
1238	vm_page_lock_queues();
1239	TAILQ_FOREACH(m, &new_object->memq, listq)
1240		vm_page_wakeup(m);
1241	vm_page_unlock_queues();
1242	VM_OBJECT_UNLOCK(new_object);
1243	entry->object.vm_object = new_object;
1244	entry->offset = 0LL;
1245	vm_object_deallocate(orig_object);
1246}
1247
1248#define	OBSC_TEST_ALL_SHADOWED	0x0001
1249#define	OBSC_COLLAPSE_NOWAIT	0x0002
1250#define	OBSC_COLLAPSE_WAIT	0x0004
1251
1252static int
1253vm_object_backing_scan(vm_object_t object, int op)
1254{
1255	int s;
1256	int r = 1;
1257	vm_page_t p;
1258	vm_object_t backing_object;
1259	vm_pindex_t backing_offset_index;
1260
1261	s = splvm();
1262	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1263	VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1264
1265	backing_object = object->backing_object;
1266	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1267
1268	/*
1269	 * Initial conditions
1270	 */
1271	if (op & OBSC_TEST_ALL_SHADOWED) {
1272		/*
1273		 * We do not want to have to test for the existence of
1274		 * swap pages in the backing object.  XXX but with the
1275		 * new swapper this would be pretty easy to do.
1276		 *
1277		 * XXX what about anonymous MAP_SHARED memory that hasn't
1278		 * been ZFOD faulted yet?  If we do not test for this, the
1279		 * shadow test may succeed! XXX
1280		 */
1281		if (backing_object->type != OBJT_DEFAULT) {
1282			splx(s);
1283			return (0);
1284		}
1285	}
1286	if (op & OBSC_COLLAPSE_WAIT) {
1287		vm_object_set_flag(backing_object, OBJ_DEAD);
1288	}
1289
1290	/*
1291	 * Our scan
1292	 */
1293	p = TAILQ_FIRST(&backing_object->memq);
1294	while (p) {
1295		vm_page_t next = TAILQ_NEXT(p, listq);
1296		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1297
1298		if (op & OBSC_TEST_ALL_SHADOWED) {
1299			vm_page_t pp;
1300
1301			/*
1302			 * Ignore pages outside the parent object's range
1303			 * and outside the parent object's mapping of the
1304			 * backing object.
1305			 *
1306			 * note that we do not busy the backing object's
1307			 * page.
1308			 */
1309			if (
1310			    p->pindex < backing_offset_index ||
1311			    new_pindex >= object->size
1312			) {
1313				p = next;
1314				continue;
1315			}
1316
1317			/*
1318			 * See if the parent has the page or if the parent's
1319			 * object pager has the page.  If the parent has the
1320			 * page but the page is not valid, the parent's
1321			 * object pager must have the page.
1322			 *
1323			 * If this fails, the parent does not completely shadow
1324			 * the object and we might as well give up now.
1325			 */
1326
1327			pp = vm_page_lookup(object, new_pindex);
1328			if (
1329			    (pp == NULL || pp->valid == 0) &&
1330			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1331			) {
1332				r = 0;
1333				break;
1334			}
1335		}
1336
1337		/*
1338		 * Check for busy page
1339		 */
1340		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1341			vm_page_t pp;
1342
1343			vm_page_lock_queues();
1344			if (op & OBSC_COLLAPSE_NOWAIT) {
1345				if ((p->flags & PG_BUSY) ||
1346				    !p->valid ||
1347				    p->hold_count ||
1348				    p->wire_count ||
1349				    p->busy) {
1350					vm_page_unlock_queues();
1351					p = next;
1352					continue;
1353				}
1354			} else if (op & OBSC_COLLAPSE_WAIT) {
1355				if ((p->flags & PG_BUSY) || p->busy) {
1356					vm_page_flag_set(p,
1357					    PG_WANTED | PG_REFERENCED);
1358					VM_OBJECT_UNLOCK(backing_object);
1359					VM_OBJECT_UNLOCK(object);
1360					msleep(p, &vm_page_queue_mtx,
1361					    PDROP | PVM, "vmocol", 0);
1362					VM_OBJECT_LOCK(object);
1363					VM_OBJECT_LOCK(backing_object);
1364					/*
1365					 * If we slept, anything could have
1366					 * happened.  Since the object is
1367					 * marked dead, the backing offset
1368					 * should not have changed so we
1369					 * just restart our scan.
1370					 */
1371					p = TAILQ_FIRST(&backing_object->memq);
1372					continue;
1373				}
1374			}
1375
1376			/*
1377			 * Busy the page
1378			 */
1379			vm_page_busy(p);
1380			vm_page_unlock_queues();
1381
1382			KASSERT(
1383			    p->object == backing_object,
1384			    ("vm_object_qcollapse(): object mismatch")
1385			);
1386
1387			/*
1388			 * Destroy any associated swap
1389			 */
1390			if (backing_object->type == OBJT_SWAP) {
1391				swap_pager_freespace(
1392				    backing_object,
1393				    p->pindex,
1394				    1
1395				);
1396			}
1397
1398			if (
1399			    p->pindex < backing_offset_index ||
1400			    new_pindex >= object->size
1401			) {
1402				/*
1403				 * Page is out of the parent object's range, we
1404				 * can simply destroy it.
1405				 */
1406				vm_page_lock_queues();
1407				pmap_remove_all(p);
1408				vm_page_free(p);
1409				vm_page_unlock_queues();
1410				p = next;
1411				continue;
1412			}
1413
1414			pp = vm_page_lookup(object, new_pindex);
1415			if (
1416			    pp != NULL ||
1417			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1418			) {
1419				/*
1420				 * page already exists in parent OR swap exists
1421				 * for this location in the parent.  Destroy
1422				 * the original page from the backing object.
1423				 *
1424				 * Leave the parent's page alone
1425				 */
1426				vm_page_lock_queues();
1427				pmap_remove_all(p);
1428				vm_page_free(p);
1429				vm_page_unlock_queues();
1430				p = next;
1431				continue;
1432			}
1433
1434			/*
1435			 * Page does not exist in parent, rename the
1436			 * page from the backing object to the main object.
1437			 *
1438			 * If the page was mapped to a process, it can remain
1439			 * mapped through the rename.
1440			 */
1441			vm_page_lock_queues();
1442			vm_page_rename(p, object, new_pindex);
1443			vm_page_unlock_queues();
1444			/* page automatically made dirty by rename */
1445		}
1446		p = next;
1447	}
1448	splx(s);
1449	return (r);
1450}
1451
1452
1453/*
1454 * this version of collapse allows the operation to occur earlier and
1455 * when paging_in_progress is true for an object...  This is not a complete
1456 * operation, but should plug 99.9% of the rest of the leaks.
1457 */
1458static void
1459vm_object_qcollapse(vm_object_t object)
1460{
1461	vm_object_t backing_object = object->backing_object;
1462
1463	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1464	VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1465
1466	if (backing_object->ref_count != 1)
1467		return;
1468
1469	backing_object->ref_count += 2;
1470
1471	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1472
1473	backing_object->ref_count -= 2;
1474}
1475
1476/*
1477 *	vm_object_collapse:
1478 *
1479 *	Collapse an object with the object backing it.
1480 *	Pages in the backing object are moved into the
1481 *	parent, and the backing object is deallocated.
1482 */
1483void
1484vm_object_collapse(vm_object_t object)
1485{
1486	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1487
1488	while (TRUE) {
1489		vm_object_t backing_object;
1490
1491		/*
1492		 * Verify that the conditions are right for collapse:
1493		 *
1494		 * The object exists and the backing object exists.
1495		 */
1496		if ((backing_object = object->backing_object) == NULL)
1497			break;
1498
1499		/*
1500		 * we check the backing object first, because it is most likely
1501		 * not collapsable.
1502		 */
1503		VM_OBJECT_LOCK(backing_object);
1504		if (backing_object->handle != NULL ||
1505		    (backing_object->type != OBJT_DEFAULT &&
1506		     backing_object->type != OBJT_SWAP) ||
1507		    (backing_object->flags & OBJ_DEAD) ||
1508		    object->handle != NULL ||
1509		    (object->type != OBJT_DEFAULT &&
1510		     object->type != OBJT_SWAP) ||
1511		    (object->flags & OBJ_DEAD)) {
1512			VM_OBJECT_UNLOCK(backing_object);
1513			break;
1514		}
1515
1516		if (
1517		    object->paging_in_progress != 0 ||
1518		    backing_object->paging_in_progress != 0
1519		) {
1520			vm_object_qcollapse(object);
1521			VM_OBJECT_UNLOCK(backing_object);
1522			break;
1523		}
1524		/*
1525		 * We know that we can either collapse the backing object (if
1526		 * the parent is the only reference to it) or (perhaps) have
1527		 * the parent bypass the object if the parent happens to shadow
1528		 * all the resident pages in the entire backing object.
1529		 *
1530		 * This is ignoring pager-backed pages such as swap pages.
1531		 * vm_object_backing_scan fails the shadowing test in this
1532		 * case.
1533		 */
1534		if (backing_object->ref_count == 1) {
1535			/*
1536			 * If there is exactly one reference to the backing
1537			 * object, we can collapse it into the parent.
1538			 */
1539			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1540
1541			/*
1542			 * Move the pager from backing_object to object.
1543			 */
1544			if (backing_object->type == OBJT_SWAP) {
1545				/*
1546				 * swap_pager_copy() can sleep, in which case
1547				 * the backing_object's and object's locks are
1548				 * released and reacquired.
1549				 */
1550				swap_pager_copy(
1551				    backing_object,
1552				    object,
1553				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1554			}
1555			/*
1556			 * Object now shadows whatever backing_object did.
1557			 * Note that the reference to
1558			 * backing_object->backing_object moves from within
1559			 * backing_object to within object.
1560			 */
1561			LIST_REMOVE(object, shadow_list);
1562			backing_object->shadow_count--;
1563			backing_object->generation++;
1564			if (backing_object->backing_object) {
1565				VM_OBJECT_LOCK(backing_object->backing_object);
1566				LIST_REMOVE(backing_object, shadow_list);
1567				LIST_INSERT_HEAD(
1568				    &backing_object->backing_object->shadow_head,
1569				    object, shadow_list);
1570				/*
1571				 * The shadow_count has not changed.
1572				 */
1573				backing_object->backing_object->generation++;
1574				VM_OBJECT_UNLOCK(backing_object->backing_object);
1575			}
1576			object->backing_object = backing_object->backing_object;
1577			object->backing_object_offset +=
1578			    backing_object->backing_object_offset;
1579/* XXX */		VM_OBJECT_UNLOCK(object);
1580
1581			/*
1582			 * Discard backing_object.
1583			 *
1584			 * Since the backing object has no pages, no pager left,
1585			 * and no object references within it, all that is
1586			 * necessary is to dispose of it.
1587			 */
1588			KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1589			VM_OBJECT_UNLOCK(backing_object);
1590
1591			mtx_lock(&vm_object_list_mtx);
1592			TAILQ_REMOVE(
1593			    &vm_object_list,
1594			    backing_object,
1595			    object_list
1596			);
1597			mtx_unlock(&vm_object_list_mtx);
1598
1599/* XXX */		VM_OBJECT_LOCK(object);
1600			uma_zfree(obj_zone, backing_object);
1601
1602			object_collapses++;
1603		} else {
1604			vm_object_t new_backing_object;
1605
1606			/*
1607			 * If we do not entirely shadow the backing object,
1608			 * there is nothing we can do so we give up.
1609			 */
1610			if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1611				VM_OBJECT_UNLOCK(backing_object);
1612				break;
1613			}
1614
1615			/*
1616			 * Make the parent shadow the next object in the
1617			 * chain.  Deallocating backing_object will not remove
1618			 * it, since its reference count is at least 2.
1619			 */
1620			LIST_REMOVE(object, shadow_list);
1621			backing_object->shadow_count--;
1622			backing_object->generation++;
1623
1624			new_backing_object = backing_object->backing_object;
1625			if ((object->backing_object = new_backing_object) != NULL) {
1626				vm_object_reference(new_backing_object);
1627				VM_OBJECT_LOCK(new_backing_object);
1628				LIST_INSERT_HEAD(
1629				    &new_backing_object->shadow_head,
1630				    object,
1631				    shadow_list
1632				);
1633				new_backing_object->shadow_count++;
1634				new_backing_object->generation++;
1635				VM_OBJECT_UNLOCK(new_backing_object);
1636				object->backing_object_offset +=
1637					backing_object->backing_object_offset;
1638			}
1639
1640			/*
1641			 * Drop the reference count on backing_object. Since
1642			 * its ref_count was at least 2, it will not vanish.
1643			 */
1644			backing_object->ref_count--;
1645			VM_OBJECT_UNLOCK(backing_object);
1646			object_bypasses++;
1647		}
1648
1649		/*
1650		 * Try again with this object's new backing object.
1651		 */
1652	}
1653}
1654
1655/*
1656 *	vm_object_page_remove:
1657 *
1658 *	Removes all physical pages in the given range from the
1659 *	object's list of pages.  If the range's end is zero, all
1660 *	physical pages from the range's start to the end of the object
1661 *	are deleted.
1662 *
1663 *	The object must be locked.
1664 */
1665void
1666vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1667    boolean_t clean_only)
1668{
1669	vm_page_t p, next;
1670
1671	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1672	if (object->resident_page_count == 0)
1673		return;
1674
1675	/*
1676	 * Since physically-backed objects do not use managed pages, we can't
1677	 * remove pages from the object (we must instead remove the page
1678	 * references, and then destroy the object).
1679	 */
1680	KASSERT(object->type != OBJT_PHYS,
1681	    ("attempt to remove pages from a physical object"));
1682
1683	vm_object_pip_add(object, 1);
1684again:
1685	vm_page_lock_queues();
1686	if ((p = TAILQ_FIRST(&object->memq)) != NULL) {
1687		if (p->pindex < start) {
1688			p = vm_page_splay(start, object->root);
1689			if ((object->root = p)->pindex < start)
1690				p = TAILQ_NEXT(p, listq);
1691		}
1692	}
1693	/*
1694	 * Assert: the variable p is either (1) the page with the
1695	 * least pindex greater than or equal to the parameter pindex
1696	 * or (2) NULL.
1697	 */
1698	for (;
1699	     p != NULL && (p->pindex < end || end == 0);
1700	     p = next) {
1701		next = TAILQ_NEXT(p, listq);
1702
1703		if (p->wire_count != 0) {
1704			pmap_remove_all(p);
1705			if (!clean_only)
1706				p->valid = 0;
1707			continue;
1708		}
1709		if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1710			goto again;
1711		if (clean_only && p->valid) {
1712			vm_page_test_dirty(p);
1713			if (p->valid & p->dirty)
1714				continue;
1715		}
1716		vm_page_busy(p);
1717		pmap_remove_all(p);
1718		vm_page_free(p);
1719	}
1720	vm_page_unlock_queues();
1721	vm_object_pip_wakeup(object);
1722}
1723
1724/*
1725 *	Routine:	vm_object_coalesce
1726 *	Function:	Coalesces two objects backing up adjoining
1727 *			regions of memory into a single object.
1728 *
1729 *	returns TRUE if objects were combined.
1730 *
1731 *	NOTE:	Only works at the moment if the second object is NULL -
1732 *		if it's not, which object do we lock first?
1733 *
1734 *	Parameters:
1735 *		prev_object	First object to coalesce
1736 *		prev_offset	Offset into prev_object
1737 *		next_object	Second object into coalesce
1738 *		next_offset	Offset into next_object
1739 *
1740 *		prev_size	Size of reference to prev_object
1741 *		next_size	Size of reference to next_object
1742 *
1743 *	Conditions:
1744 *	The object must *not* be locked.
1745 */
1746boolean_t
1747vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
1748	vm_size_t prev_size, vm_size_t next_size)
1749{
1750	vm_pindex_t next_pindex;
1751
1752	if (prev_object == NULL)
1753		return (TRUE);
1754	VM_OBJECT_LOCK(prev_object);
1755	if (prev_object->type != OBJT_DEFAULT &&
1756	    prev_object->type != OBJT_SWAP) {
1757		VM_OBJECT_UNLOCK(prev_object);
1758		return (FALSE);
1759	}
1760
1761	/*
1762	 * Try to collapse the object first
1763	 */
1764	vm_object_collapse(prev_object);
1765
1766	/*
1767	 * Can't coalesce if: . more than one reference . paged out . shadows
1768	 * another object . has a copy elsewhere (any of which mean that the
1769	 * pages not mapped to prev_entry may be in use anyway)
1770	 */
1771	if (prev_object->backing_object != NULL) {
1772		VM_OBJECT_UNLOCK(prev_object);
1773		return (FALSE);
1774	}
1775
1776	prev_size >>= PAGE_SHIFT;
1777	next_size >>= PAGE_SHIFT;
1778	next_pindex = prev_pindex + prev_size;
1779
1780	if ((prev_object->ref_count > 1) &&
1781	    (prev_object->size != next_pindex)) {
1782		VM_OBJECT_UNLOCK(prev_object);
1783		return (FALSE);
1784	}
1785
1786	/*
1787	 * Remove any pages that may still be in the object from a previous
1788	 * deallocation.
1789	 */
1790	if (next_pindex < prev_object->size) {
1791		vm_object_page_remove(prev_object,
1792				      next_pindex,
1793				      next_pindex + next_size, FALSE);
1794		if (prev_object->type == OBJT_SWAP)
1795			swap_pager_freespace(prev_object,
1796					     next_pindex, next_size);
1797	}
1798
1799	/*
1800	 * Extend the object if necessary.
1801	 */
1802	if (next_pindex + next_size > prev_object->size)
1803		prev_object->size = next_pindex + next_size;
1804
1805	VM_OBJECT_UNLOCK(prev_object);
1806	return (TRUE);
1807}
1808
1809void
1810vm_object_set_writeable_dirty(vm_object_t object)
1811{
1812	struct vnode *vp;
1813
1814	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1815	vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1816	if (object->type == OBJT_VNODE &&
1817	    (vp = (struct vnode *)object->handle) != NULL) {
1818		VI_LOCK(vp);
1819		if ((vp->v_iflag & VI_OBJDIRTY) == 0)
1820			vp->v_iflag |= VI_OBJDIRTY;
1821		VI_UNLOCK(vp);
1822	}
1823}
1824
1825#include "opt_ddb.h"
1826#ifdef DDB
1827#include <sys/kernel.h>
1828
1829#include <sys/cons.h>
1830
1831#include <ddb/ddb.h>
1832
1833static int
1834_vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
1835{
1836	vm_map_t tmpm;
1837	vm_map_entry_t tmpe;
1838	vm_object_t obj;
1839	int entcount;
1840
1841	if (map == 0)
1842		return 0;
1843
1844	if (entry == 0) {
1845		tmpe = map->header.next;
1846		entcount = map->nentries;
1847		while (entcount-- && (tmpe != &map->header)) {
1848			if (_vm_object_in_map(map, object, tmpe)) {
1849				return 1;
1850			}
1851			tmpe = tmpe->next;
1852		}
1853	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
1854		tmpm = entry->object.sub_map;
1855		tmpe = tmpm->header.next;
1856		entcount = tmpm->nentries;
1857		while (entcount-- && tmpe != &tmpm->header) {
1858			if (_vm_object_in_map(tmpm, object, tmpe)) {
1859				return 1;
1860			}
1861			tmpe = tmpe->next;
1862		}
1863	} else if ((obj = entry->object.vm_object) != NULL) {
1864		for (; obj; obj = obj->backing_object)
1865			if (obj == object) {
1866				return 1;
1867			}
1868	}
1869	return 0;
1870}
1871
1872static int
1873vm_object_in_map(vm_object_t object)
1874{
1875	struct proc *p;
1876
1877	/* sx_slock(&allproc_lock); */
1878	LIST_FOREACH(p, &allproc, p_list) {
1879		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1880			continue;
1881		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
1882			/* sx_sunlock(&allproc_lock); */
1883			return 1;
1884		}
1885	}
1886	/* sx_sunlock(&allproc_lock); */
1887	if (_vm_object_in_map(kernel_map, object, 0))
1888		return 1;
1889	if (_vm_object_in_map(kmem_map, object, 0))
1890		return 1;
1891	if (_vm_object_in_map(pager_map, object, 0))
1892		return 1;
1893	if (_vm_object_in_map(buffer_map, object, 0))
1894		return 1;
1895	return 0;
1896}
1897
1898DB_SHOW_COMMAND(vmochk, vm_object_check)
1899{
1900	vm_object_t object;
1901
1902	/*
1903	 * make sure that internal objs are in a map somewhere
1904	 * and none have zero ref counts.
1905	 */
1906	TAILQ_FOREACH(object, &vm_object_list, object_list) {
1907		if (object->handle == NULL &&
1908		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1909			if (object->ref_count == 0) {
1910				db_printf("vmochk: internal obj has zero ref count: %ld\n",
1911					(long)object->size);
1912			}
1913			if (!vm_object_in_map(object)) {
1914				db_printf(
1915			"vmochk: internal obj is not in a map: "
1916			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1917				    object->ref_count, (u_long)object->size,
1918				    (u_long)object->size,
1919				    (void *)object->backing_object);
1920			}
1921		}
1922	}
1923}
1924
1925/*
1926 *	vm_object_print:	[ debug ]
1927 */
1928DB_SHOW_COMMAND(object, vm_object_print_static)
1929{
1930	/* XXX convert args. */
1931	vm_object_t object = (vm_object_t)addr;
1932	boolean_t full = have_addr;
1933
1934	vm_page_t p;
1935
1936	/* XXX count is an (unused) arg.  Avoid shadowing it. */
1937#define	count	was_count
1938
1939	int count;
1940
1941	if (object == NULL)
1942		return;
1943
1944	db_iprintf(
1945	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x\n",
1946	    object, (int)object->type, (uintmax_t)object->size,
1947	    object->resident_page_count, object->ref_count, object->flags);
1948	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
1949	    object->shadow_count,
1950	    object->backing_object ? object->backing_object->ref_count : 0,
1951	    object->backing_object, (uintmax_t)object->backing_object_offset);
1952
1953	if (!full)
1954		return;
1955
1956	db_indent += 2;
1957	count = 0;
1958	TAILQ_FOREACH(p, &object->memq, listq) {
1959		if (count == 0)
1960			db_iprintf("memory:=");
1961		else if (count == 6) {
1962			db_printf("\n");
1963			db_iprintf(" ...");
1964			count = 0;
1965		} else
1966			db_printf(",");
1967		count++;
1968
1969		db_printf("(off=0x%jx,page=0x%jx)",
1970		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
1971	}
1972	if (count != 0)
1973		db_printf("\n");
1974	db_indent -= 2;
1975}
1976
1977/* XXX. */
1978#undef count
1979
1980/* XXX need this non-static entry for calling from vm_map_print. */
1981void
1982vm_object_print(
1983        /* db_expr_t */ long addr,
1984	boolean_t have_addr,
1985	/* db_expr_t */ long count,
1986	char *modif)
1987{
1988	vm_object_print_static(addr, have_addr, count, modif);
1989}
1990
1991DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1992{
1993	vm_object_t object;
1994	int nl = 0;
1995	int c;
1996
1997	TAILQ_FOREACH(object, &vm_object_list, object_list) {
1998		vm_pindex_t idx, fidx;
1999		vm_pindex_t osize;
2000		vm_paddr_t pa = -1, padiff;
2001		int rcount;
2002		vm_page_t m;
2003
2004		db_printf("new object: %p\n", (void *)object);
2005		if (nl > 18) {
2006			c = cngetc();
2007			if (c != ' ')
2008				return;
2009			nl = 0;
2010		}
2011		nl++;
2012		rcount = 0;
2013		fidx = 0;
2014		osize = object->size;
2015		if (osize > 128)
2016			osize = 128;
2017		for (idx = 0; idx < osize; idx++) {
2018			m = vm_page_lookup(object, idx);
2019			if (m == NULL) {
2020				if (rcount) {
2021					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2022						(long)fidx, rcount, (long)pa);
2023					if (nl > 18) {
2024						c = cngetc();
2025						if (c != ' ')
2026							return;
2027						nl = 0;
2028					}
2029					nl++;
2030					rcount = 0;
2031				}
2032				continue;
2033			}
2034
2035
2036			if (rcount &&
2037				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2038				++rcount;
2039				continue;
2040			}
2041			if (rcount) {
2042				padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
2043				padiff >>= PAGE_SHIFT;
2044				padiff &= PQ_L2_MASK;
2045				if (padiff == 0) {
2046					pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
2047					++rcount;
2048					continue;
2049				}
2050				db_printf(" index(%ld)run(%d)pa(0x%lx)",
2051					(long)fidx, rcount, (long)pa);
2052				db_printf("pd(%ld)\n", (long)padiff);
2053				if (nl > 18) {
2054					c = cngetc();
2055					if (c != ' ')
2056						return;
2057					nl = 0;
2058				}
2059				nl++;
2060			}
2061			fidx = idx;
2062			pa = VM_PAGE_TO_PHYS(m);
2063			rcount = 1;
2064		}
2065		if (rcount) {
2066			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2067				(long)fidx, rcount, (long)pa);
2068			if (nl > 18) {
2069				c = cngetc();
2070				if (c != ' ')
2071					return;
2072				nl = 0;
2073			}
2074			nl++;
2075		}
2076	}
2077}
2078#endif /* DDB */
2079