xref: /freebsd/sys/vm/vm_object.c (revision fbf2a778760f24565eab950ced3e2ad1df4cf844)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	from: @(#)vm_object.c	8.5 (Berkeley) 3/22/94
35  *
36  *
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *	Virtual memory object module.
65  */
66 
67 #include <sys/cdefs.h>
68 __FBSDID("$FreeBSD$");
69 
70 #include "opt_vm.h"
71 
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/blockcount.h>
75 #include <sys/cpuset.h>
76 #include <sys/lock.h>
77 #include <sys/mman.h>
78 #include <sys/mount.h>
79 #include <sys/kernel.h>
80 #include <sys/pctrie.h>
81 #include <sys/sysctl.h>
82 #include <sys/mutex.h>
83 #include <sys/proc.h>		/* for curproc, pageproc */
84 #include <sys/refcount.h>
85 #include <sys/socket.h>
86 #include <sys/resourcevar.h>
87 #include <sys/refcount.h>
88 #include <sys/rwlock.h>
89 #include <sys/user.h>
90 #include <sys/vnode.h>
91 #include <sys/vmmeter.h>
92 #include <sys/sx.h>
93 
94 #include <vm/vm.h>
95 #include <vm/vm_param.h>
96 #include <vm/pmap.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_pageout.h>
101 #include <vm/vm_pager.h>
102 #include <vm/vm_phys.h>
103 #include <vm/vm_pagequeue.h>
104 #include <vm/swap_pager.h>
105 #include <vm/vm_kern.h>
106 #include <vm/vm_extern.h>
107 #include <vm/vm_radix.h>
108 #include <vm/vm_reserv.h>
109 #include <vm/uma.h>
110 
111 static int old_msync;
112 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
113     "Use old (insecure) msync behavior");
114 
115 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
116 		    int pagerflags, int flags, boolean_t *allclean,
117 		    boolean_t *eio);
118 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
119 		    boolean_t *allclean);
120 static void	vm_object_backing_remove(vm_object_t object);
121 
122 /*
123  *	Virtual memory objects maintain the actual data
124  *	associated with allocated virtual memory.  A given
125  *	page of memory exists within exactly one object.
126  *
127  *	An object is only deallocated when all "references"
128  *	are given up.  Only one "reference" to a given
129  *	region of an object should be writeable.
130  *
131  *	Associated with each object is a list of all resident
132  *	memory pages belonging to that object; this list is
133  *	maintained by the "vm_page" module, and locked by the object's
134  *	lock.
135  *
136  *	Each object also records a "pager" routine which is
137  *	used to retrieve (and store) pages to the proper backing
138  *	storage.  In addition, objects may be backed by other
139  *	objects from which they were virtual-copied.
140  *
141  *	The only items within the object structure which are
142  *	modified after time of creation are:
143  *		reference count		locked by object's lock
144  *		pager routine		locked by object's lock
145  *
146  */
147 
148 struct object_q vm_object_list;
149 struct mtx vm_object_list_mtx;	/* lock for object list and count */
150 
151 struct vm_object kernel_object_store;
152 
153 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
154     "VM object stats");
155 
156 static COUNTER_U64_DEFINE_EARLY(object_collapses);
157 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
158     &object_collapses,
159     "VM object collapses");
160 
161 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
162 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
163     &object_bypasses,
164     "VM object bypasses");
165 
166 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
167 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
168     &object_collapse_waits,
169     "Number of sleeps for collapse");
170 
171 static uma_zone_t obj_zone;
172 
173 static int vm_object_zinit(void *mem, int size, int flags);
174 
175 #ifdef INVARIANTS
176 static void vm_object_zdtor(void *mem, int size, void *arg);
177 
178 static void
179 vm_object_zdtor(void *mem, int size, void *arg)
180 {
181 	vm_object_t object;
182 
183 	object = (vm_object_t)mem;
184 	KASSERT(object->ref_count == 0,
185 	    ("object %p ref_count = %d", object, object->ref_count));
186 	KASSERT(TAILQ_EMPTY(&object->memq),
187 	    ("object %p has resident pages in its memq", object));
188 	KASSERT(vm_radix_is_empty(&object->rtree),
189 	    ("object %p has resident pages in its trie", object));
190 #if VM_NRESERVLEVEL > 0
191 	KASSERT(LIST_EMPTY(&object->rvq),
192 	    ("object %p has reservations",
193 	    object));
194 #endif
195 	KASSERT(!vm_object_busied(object),
196 	    ("object %p busy = %d", object, blockcount_read(&object->busy)));
197 	KASSERT(object->resident_page_count == 0,
198 	    ("object %p resident_page_count = %d",
199 	    object, object->resident_page_count));
200 	KASSERT(object->shadow_count == 0,
201 	    ("object %p shadow_count = %d",
202 	    object, object->shadow_count));
203 	KASSERT(object->type == OBJT_DEAD,
204 	    ("object %p has non-dead type %d",
205 	    object, object->type));
206 }
207 #endif
208 
209 static int
210 vm_object_zinit(void *mem, int size, int flags)
211 {
212 	vm_object_t object;
213 
214 	object = (vm_object_t)mem;
215 	rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
216 
217 	/* These are true for any object that has been freed */
218 	object->type = OBJT_DEAD;
219 	vm_radix_init(&object->rtree);
220 	refcount_init(&object->ref_count, 0);
221 	blockcount_init(&object->paging_in_progress);
222 	blockcount_init(&object->busy);
223 	object->resident_page_count = 0;
224 	object->shadow_count = 0;
225 	object->flags = OBJ_DEAD;
226 
227 	mtx_lock(&vm_object_list_mtx);
228 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
229 	mtx_unlock(&vm_object_list_mtx);
230 	return (0);
231 }
232 
233 static void
234 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
235     vm_object_t object, void *handle)
236 {
237 
238 	TAILQ_INIT(&object->memq);
239 	LIST_INIT(&object->shadow_head);
240 
241 	object->type = type;
242 	if (type == OBJT_SWAP)
243 		pctrie_init(&object->un_pager.swp.swp_blks);
244 
245 	/*
246 	 * Ensure that swap_pager_swapoff() iteration over object_list
247 	 * sees up to date type and pctrie head if it observed
248 	 * non-dead object.
249 	 */
250 	atomic_thread_fence_rel();
251 
252 	object->pg_color = 0;
253 	object->flags = flags;
254 	object->size = size;
255 	object->domain.dr_policy = NULL;
256 	object->generation = 1;
257 	object->cleangeneration = 1;
258 	refcount_init(&object->ref_count, 1);
259 	object->memattr = VM_MEMATTR_DEFAULT;
260 	object->cred = NULL;
261 	object->charge = 0;
262 	object->handle = handle;
263 	object->backing_object = NULL;
264 	object->backing_object_offset = (vm_ooffset_t) 0;
265 #if VM_NRESERVLEVEL > 0
266 	LIST_INIT(&object->rvq);
267 #endif
268 	umtx_shm_object_init(object);
269 }
270 
271 /*
272  *	vm_object_init:
273  *
274  *	Initialize the VM objects module.
275  */
276 void
277 vm_object_init(void)
278 {
279 	TAILQ_INIT(&vm_object_list);
280 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
281 
282 	rw_init(&kernel_object->lock, "kernel vm object");
283 	_vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
284 	    VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
285 #if VM_NRESERVLEVEL > 0
286 	kernel_object->flags |= OBJ_COLORED;
287 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
288 #endif
289 
290 	/*
291 	 * The lock portion of struct vm_object must be type stable due
292 	 * to vm_pageout_fallback_object_lock locking a vm object
293 	 * without holding any references to it.
294 	 *
295 	 * paging_in_progress is valid always.  Lockless references to
296 	 * the objects may acquire pip and then check OBJ_DEAD.
297 	 */
298 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
299 #ifdef INVARIANTS
300 	    vm_object_zdtor,
301 #else
302 	    NULL,
303 #endif
304 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
305 
306 	vm_radix_zinit();
307 }
308 
309 void
310 vm_object_clear_flag(vm_object_t object, u_short bits)
311 {
312 
313 	VM_OBJECT_ASSERT_WLOCKED(object);
314 	object->flags &= ~bits;
315 }
316 
317 /*
318  *	Sets the default memory attribute for the specified object.  Pages
319  *	that are allocated to this object are by default assigned this memory
320  *	attribute.
321  *
322  *	Presently, this function must be called before any pages are allocated
323  *	to the object.  In the future, this requirement may be relaxed for
324  *	"default" and "swap" objects.
325  */
326 int
327 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
328 {
329 
330 	VM_OBJECT_ASSERT_WLOCKED(object);
331 	switch (object->type) {
332 	case OBJT_DEFAULT:
333 	case OBJT_DEVICE:
334 	case OBJT_MGTDEVICE:
335 	case OBJT_PHYS:
336 	case OBJT_SG:
337 	case OBJT_SWAP:
338 	case OBJT_VNODE:
339 		if (!TAILQ_EMPTY(&object->memq))
340 			return (KERN_FAILURE);
341 		break;
342 	case OBJT_DEAD:
343 		return (KERN_INVALID_ARGUMENT);
344 	default:
345 		panic("vm_object_set_memattr: object %p is of undefined type",
346 		    object);
347 	}
348 	object->memattr = memattr;
349 	return (KERN_SUCCESS);
350 }
351 
352 void
353 vm_object_pip_add(vm_object_t object, short i)
354 {
355 
356 	if (i > 0)
357 		blockcount_acquire(&object->paging_in_progress, i);
358 }
359 
360 void
361 vm_object_pip_wakeup(vm_object_t object)
362 {
363 
364 	vm_object_pip_wakeupn(object, 1);
365 }
366 
367 void
368 vm_object_pip_wakeupn(vm_object_t object, short i)
369 {
370 
371 	if (i > 0)
372 		blockcount_release(&object->paging_in_progress, i);
373 }
374 
375 /*
376  * Atomically drop the object lock and wait for pip to drain.  This protects
377  * from sleep/wakeup races due to identity changes.  The lock is not re-acquired
378  * on return.
379  */
380 static void
381 vm_object_pip_sleep(vm_object_t object, const char *waitid)
382 {
383 
384 	(void)blockcount_sleep(&object->paging_in_progress, &object->lock,
385 	    waitid, PVM | PDROP);
386 }
387 
388 void
389 vm_object_pip_wait(vm_object_t object, const char *waitid)
390 {
391 
392 	VM_OBJECT_ASSERT_WLOCKED(object);
393 
394 	blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
395 	    PVM);
396 }
397 
398 void
399 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
400 {
401 
402 	VM_OBJECT_ASSERT_UNLOCKED(object);
403 
404 	blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
405 }
406 
407 /*
408  *	vm_object_allocate:
409  *
410  *	Returns a new object with the given size.
411  */
412 vm_object_t
413 vm_object_allocate(objtype_t type, vm_pindex_t size)
414 {
415 	vm_object_t object;
416 	u_short flags;
417 
418 	switch (type) {
419 	case OBJT_DEAD:
420 		panic("vm_object_allocate: can't create OBJT_DEAD");
421 	case OBJT_DEFAULT:
422 	case OBJT_SWAP:
423 		flags = OBJ_COLORED;
424 		break;
425 	case OBJT_DEVICE:
426 	case OBJT_SG:
427 		flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
428 		break;
429 	case OBJT_MGTDEVICE:
430 		flags = OBJ_FICTITIOUS;
431 		break;
432 	case OBJT_PHYS:
433 		flags = OBJ_UNMANAGED;
434 		break;
435 	case OBJT_VNODE:
436 		flags = 0;
437 		break;
438 	default:
439 		panic("vm_object_allocate: type %d is undefined", type);
440 	}
441 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
442 	_vm_object_allocate(type, size, flags, object, NULL);
443 
444 	return (object);
445 }
446 
447 /*
448  *	vm_object_allocate_anon:
449  *
450  *	Returns a new default object of the given size and marked as
451  *	anonymous memory for special split/collapse handling.  Color
452  *	to be initialized by the caller.
453  */
454 vm_object_t
455 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
456     struct ucred *cred, vm_size_t charge)
457 {
458 	vm_object_t handle, object;
459 
460 	if (backing_object == NULL)
461 		handle = NULL;
462 	else if ((backing_object->flags & OBJ_ANON) != 0)
463 		handle = backing_object->handle;
464 	else
465 		handle = backing_object;
466 	object = uma_zalloc(obj_zone, M_WAITOK);
467 	_vm_object_allocate(OBJT_DEFAULT, size, OBJ_ANON | OBJ_ONEMAPPING,
468 	    object, handle);
469 	object->cred = cred;
470 	object->charge = cred != NULL ? charge : 0;
471 	return (object);
472 }
473 
474 static void
475 vm_object_reference_vnode(vm_object_t object)
476 {
477 	u_int old;
478 
479 	/*
480 	 * vnode objects need the lock for the first reference
481 	 * to serialize with vnode_object_deallocate().
482 	 */
483 	if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
484 		VM_OBJECT_RLOCK(object);
485 		old = refcount_acquire(&object->ref_count);
486 		if (object->type == OBJT_VNODE && old == 0)
487 			vref(object->handle);
488 		VM_OBJECT_RUNLOCK(object);
489 	}
490 }
491 
492 /*
493  *	vm_object_reference:
494  *
495  *	Acquires a reference to the given object.
496  */
497 void
498 vm_object_reference(vm_object_t object)
499 {
500 
501 	if (object == NULL)
502 		return;
503 
504 	if (object->type == OBJT_VNODE)
505 		vm_object_reference_vnode(object);
506 	else
507 		refcount_acquire(&object->ref_count);
508 	KASSERT((object->flags & OBJ_DEAD) == 0,
509 	    ("vm_object_reference: Referenced dead object."));
510 }
511 
512 /*
513  *	vm_object_reference_locked:
514  *
515  *	Gets another reference to the given object.
516  *
517  *	The object must be locked.
518  */
519 void
520 vm_object_reference_locked(vm_object_t object)
521 {
522 	u_int old;
523 
524 	VM_OBJECT_ASSERT_LOCKED(object);
525 	old = refcount_acquire(&object->ref_count);
526 	if (object->type == OBJT_VNODE && old == 0)
527 		vref(object->handle);
528 	KASSERT((object->flags & OBJ_DEAD) == 0,
529 	    ("vm_object_reference: Referenced dead object."));
530 }
531 
532 /*
533  * Handle deallocating an object of type OBJT_VNODE.
534  */
535 static void
536 vm_object_deallocate_vnode(vm_object_t object)
537 {
538 	struct vnode *vp = (struct vnode *) object->handle;
539 	bool last;
540 
541 	KASSERT(object->type == OBJT_VNODE,
542 	    ("vm_object_deallocate_vnode: not a vnode object"));
543 	KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
544 
545 	/* Object lock to protect handle lookup. */
546 	last = refcount_release(&object->ref_count);
547 	VM_OBJECT_RUNLOCK(object);
548 
549 	if (!last)
550 		return;
551 
552 	if (!umtx_shm_vnobj_persistent)
553 		umtx_shm_object_terminated(object);
554 
555 	/* vrele may need the vnode lock. */
556 	vrele(vp);
557 }
558 
559 /*
560  * We dropped a reference on an object and discovered that it had a
561  * single remaining shadow.  This is a sibling of the reference we
562  * dropped.  Attempt to collapse the sibling and backing object.
563  */
564 static vm_object_t
565 vm_object_deallocate_anon(vm_object_t backing_object)
566 {
567 	vm_object_t object;
568 
569 	/* Fetch the final shadow.  */
570 	object = LIST_FIRST(&backing_object->shadow_head);
571 	KASSERT(object != NULL && backing_object->shadow_count == 1,
572 	    ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
573 	    backing_object->ref_count, backing_object->shadow_count));
574 	KASSERT((object->flags & (OBJ_TMPFS_NODE | OBJ_ANON)) == OBJ_ANON,
575 	    ("invalid shadow object %p", object));
576 
577 	if (!VM_OBJECT_TRYWLOCK(object)) {
578 		/*
579 		 * Prevent object from disappearing since we do not have a
580 		 * reference.
581 		 */
582 		vm_object_pip_add(object, 1);
583 		VM_OBJECT_WUNLOCK(backing_object);
584 		VM_OBJECT_WLOCK(object);
585 		vm_object_pip_wakeup(object);
586 	} else
587 		VM_OBJECT_WUNLOCK(backing_object);
588 
589 	/*
590 	 * Check for a collapse/terminate race with the last reference holder.
591 	 */
592 	if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
593 	    !refcount_acquire_if_not_zero(&object->ref_count)) {
594 		VM_OBJECT_WUNLOCK(object);
595 		return (NULL);
596 	}
597 	backing_object = object->backing_object;
598 	if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
599 		vm_object_collapse(object);
600 	VM_OBJECT_WUNLOCK(object);
601 
602 	return (object);
603 }
604 
605 /*
606  *	vm_object_deallocate:
607  *
608  *	Release a reference to the specified object,
609  *	gained either through a vm_object_allocate
610  *	or a vm_object_reference call.  When all references
611  *	are gone, storage associated with this object
612  *	may be relinquished.
613  *
614  *	No object may be locked.
615  */
616 void
617 vm_object_deallocate(vm_object_t object)
618 {
619 	vm_object_t temp;
620 	bool released;
621 
622 	while (object != NULL) {
623 		/*
624 		 * If the reference count goes to 0 we start calling
625 		 * vm_object_terminate() on the object chain.  A ref count
626 		 * of 1 may be a special case depending on the shadow count
627 		 * being 0 or 1.  These cases require a write lock on the
628 		 * object.
629 		 */
630 		if ((object->flags & OBJ_ANON) == 0)
631 			released = refcount_release_if_gt(&object->ref_count, 1);
632 		else
633 			released = refcount_release_if_gt(&object->ref_count, 2);
634 		if (released)
635 			return;
636 
637 		if (object->type == OBJT_VNODE) {
638 			VM_OBJECT_RLOCK(object);
639 			if (object->type == OBJT_VNODE) {
640 				vm_object_deallocate_vnode(object);
641 				return;
642 			}
643 			VM_OBJECT_RUNLOCK(object);
644 		}
645 
646 		VM_OBJECT_WLOCK(object);
647 		KASSERT(object->ref_count > 0,
648 		    ("vm_object_deallocate: object deallocated too many times: %d",
649 		    object->type));
650 
651 		/*
652 		 * If this is not the final reference to an anonymous
653 		 * object we may need to collapse the shadow chain.
654 		 */
655 		if (!refcount_release(&object->ref_count)) {
656 			if (object->ref_count > 1 ||
657 			    object->shadow_count == 0) {
658 				if ((object->flags & OBJ_ANON) != 0 &&
659 				    object->ref_count == 1)
660 					vm_object_set_flag(object,
661 					    OBJ_ONEMAPPING);
662 				VM_OBJECT_WUNLOCK(object);
663 				return;
664 			}
665 
666 			/* Handle collapsing last ref on anonymous objects. */
667 			object = vm_object_deallocate_anon(object);
668 			continue;
669 		}
670 
671 		/*
672 		 * Handle the final reference to an object.  We restart
673 		 * the loop with the backing object to avoid recursion.
674 		 */
675 		umtx_shm_object_terminated(object);
676 		temp = object->backing_object;
677 		if (temp != NULL) {
678 			KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
679 			    ("shadowed tmpfs v_object 2 %p", object));
680 			vm_object_backing_remove(object);
681 		}
682 
683 		KASSERT((object->flags & OBJ_DEAD) == 0,
684 		    ("vm_object_deallocate: Terminating dead object."));
685 		vm_object_set_flag(object, OBJ_DEAD);
686 		vm_object_terminate(object);
687 		object = temp;
688 	}
689 }
690 
691 /*
692  *	vm_object_destroy removes the object from the global object list
693  *      and frees the space for the object.
694  */
695 void
696 vm_object_destroy(vm_object_t object)
697 {
698 
699 	/*
700 	 * Release the allocation charge.
701 	 */
702 	if (object->cred != NULL) {
703 		swap_release_by_cred(object->charge, object->cred);
704 		object->charge = 0;
705 		crfree(object->cred);
706 		object->cred = NULL;
707 	}
708 
709 	/*
710 	 * Free the space for the object.
711 	 */
712 	uma_zfree(obj_zone, object);
713 }
714 
715 static void
716 vm_object_backing_remove_locked(vm_object_t object)
717 {
718 	vm_object_t backing_object;
719 
720 	backing_object = object->backing_object;
721 	VM_OBJECT_ASSERT_WLOCKED(object);
722 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
723 
724 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
725 	    ("vm_object_backing_remove: Removing collapsing object."));
726 
727 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
728 		LIST_REMOVE(object, shadow_list);
729 		backing_object->shadow_count--;
730 		object->flags &= ~OBJ_SHADOWLIST;
731 	}
732 	object->backing_object = NULL;
733 }
734 
735 static void
736 vm_object_backing_remove(vm_object_t object)
737 {
738 	vm_object_t backing_object;
739 
740 	VM_OBJECT_ASSERT_WLOCKED(object);
741 
742 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
743 		backing_object = object->backing_object;
744 		VM_OBJECT_WLOCK(backing_object);
745 		vm_object_backing_remove_locked(object);
746 		VM_OBJECT_WUNLOCK(backing_object);
747 	} else
748 		object->backing_object = NULL;
749 }
750 
751 static void
752 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
753 {
754 
755 	VM_OBJECT_ASSERT_WLOCKED(object);
756 
757 	if ((backing_object->flags & OBJ_ANON) != 0) {
758 		VM_OBJECT_ASSERT_WLOCKED(backing_object);
759 		LIST_INSERT_HEAD(&backing_object->shadow_head, object,
760 		    shadow_list);
761 		backing_object->shadow_count++;
762 		object->flags |= OBJ_SHADOWLIST;
763 	}
764 	object->backing_object = backing_object;
765 }
766 
767 static void
768 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
769 {
770 
771 	VM_OBJECT_ASSERT_WLOCKED(object);
772 
773 	if ((backing_object->flags & OBJ_ANON) != 0) {
774 		VM_OBJECT_WLOCK(backing_object);
775 		vm_object_backing_insert_locked(object, backing_object);
776 		VM_OBJECT_WUNLOCK(backing_object);
777 	} else
778 		object->backing_object = backing_object;
779 }
780 
781 /*
782  * Insert an object into a backing_object's shadow list with an additional
783  * reference to the backing_object added.
784  */
785 static void
786 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
787 {
788 
789 	VM_OBJECT_ASSERT_WLOCKED(object);
790 
791 	if ((backing_object->flags & OBJ_ANON) != 0) {
792 		VM_OBJECT_WLOCK(backing_object);
793 		KASSERT((backing_object->flags & OBJ_DEAD) == 0,
794 		    ("shadowing dead anonymous object"));
795 		vm_object_reference_locked(backing_object);
796 		vm_object_backing_insert_locked(object, backing_object);
797 		vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
798 		VM_OBJECT_WUNLOCK(backing_object);
799 	} else {
800 		vm_object_reference(backing_object);
801 		object->backing_object = backing_object;
802 	}
803 }
804 
805 /*
806  * Transfer a backing reference from backing_object to object.
807  */
808 static void
809 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
810 {
811 	vm_object_t new_backing_object;
812 
813 	/*
814 	 * Note that the reference to backing_object->backing_object
815 	 * moves from within backing_object to within object.
816 	 */
817 	vm_object_backing_remove_locked(object);
818 	new_backing_object = backing_object->backing_object;
819 	if (new_backing_object == NULL)
820 		return;
821 	if ((new_backing_object->flags & OBJ_ANON) != 0) {
822 		VM_OBJECT_WLOCK(new_backing_object);
823 		vm_object_backing_remove_locked(backing_object);
824 		vm_object_backing_insert_locked(object, new_backing_object);
825 		VM_OBJECT_WUNLOCK(new_backing_object);
826 	} else {
827 		object->backing_object = new_backing_object;
828 		backing_object->backing_object = NULL;
829 	}
830 }
831 
832 /*
833  * Wait for a concurrent collapse to settle.
834  */
835 static void
836 vm_object_collapse_wait(vm_object_t object)
837 {
838 
839 	VM_OBJECT_ASSERT_WLOCKED(object);
840 
841 	while ((object->flags & OBJ_COLLAPSING) != 0) {
842 		vm_object_pip_wait(object, "vmcolwait");
843 		counter_u64_add(object_collapse_waits, 1);
844 	}
845 }
846 
847 /*
848  * Waits for a backing object to clear a pending collapse and returns
849  * it locked if it is an ANON object.
850  */
851 static vm_object_t
852 vm_object_backing_collapse_wait(vm_object_t object)
853 {
854 	vm_object_t backing_object;
855 
856 	VM_OBJECT_ASSERT_WLOCKED(object);
857 
858 	for (;;) {
859 		backing_object = object->backing_object;
860 		if (backing_object == NULL ||
861 		    (backing_object->flags & OBJ_ANON) == 0)
862 			return (NULL);
863 		VM_OBJECT_WLOCK(backing_object);
864 		if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
865 			break;
866 		VM_OBJECT_WUNLOCK(object);
867 		vm_object_pip_sleep(backing_object, "vmbckwait");
868 		counter_u64_add(object_collapse_waits, 1);
869 		VM_OBJECT_WLOCK(object);
870 	}
871 	return (backing_object);
872 }
873 
874 /*
875  *	vm_object_terminate_pages removes any remaining pageable pages
876  *	from the object and resets the object to an empty state.
877  */
878 static void
879 vm_object_terminate_pages(vm_object_t object)
880 {
881 	vm_page_t p, p_next;
882 
883 	VM_OBJECT_ASSERT_WLOCKED(object);
884 
885 	/*
886 	 * Free any remaining pageable pages.  This also removes them from the
887 	 * paging queues.  However, don't free wired pages, just remove them
888 	 * from the object.  Rather than incrementally removing each page from
889 	 * the object, the page and object are reset to any empty state.
890 	 */
891 	TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
892 		vm_page_assert_unbusied(p);
893 		KASSERT(p->object == object &&
894 		    (p->ref_count & VPRC_OBJREF) != 0,
895 		    ("vm_object_terminate_pages: page %p is inconsistent", p));
896 
897 		p->object = NULL;
898 		if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
899 			VM_CNT_INC(v_pfree);
900 			vm_page_free(p);
901 		}
902 	}
903 
904 	/*
905 	 * If the object contained any pages, then reset it to an empty state.
906 	 * None of the object's fields, including "resident_page_count", were
907 	 * modified by the preceding loop.
908 	 */
909 	if (object->resident_page_count != 0) {
910 		vm_radix_reclaim_allnodes(&object->rtree);
911 		TAILQ_INIT(&object->memq);
912 		object->resident_page_count = 0;
913 		if (object->type == OBJT_VNODE)
914 			vdrop(object->handle);
915 	}
916 }
917 
918 /*
919  *	vm_object_terminate actually destroys the specified object, freeing
920  *	up all previously used resources.
921  *
922  *	The object must be locked.
923  *	This routine may block.
924  */
925 void
926 vm_object_terminate(vm_object_t object)
927 {
928 
929 	VM_OBJECT_ASSERT_WLOCKED(object);
930 	KASSERT((object->flags & OBJ_DEAD) != 0,
931 	    ("terminating non-dead obj %p", object));
932 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
933 	    ("terminating collapsing obj %p", object));
934 	KASSERT(object->backing_object == NULL,
935 	    ("terminating shadow obj %p", object));
936 
937 	/*
938 	 * Wait for the pageout daemon and other current users to be
939 	 * done with the object.  Note that new paging_in_progress
940 	 * users can come after this wait, but they must check
941 	 * OBJ_DEAD flag set (without unlocking the object), and avoid
942 	 * the object being terminated.
943 	 */
944 	vm_object_pip_wait(object, "objtrm");
945 
946 	KASSERT(object->ref_count == 0,
947 	    ("vm_object_terminate: object with references, ref_count=%d",
948 	    object->ref_count));
949 
950 	if ((object->flags & OBJ_PG_DTOR) == 0)
951 		vm_object_terminate_pages(object);
952 
953 #if VM_NRESERVLEVEL > 0
954 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
955 		vm_reserv_break_all(object);
956 #endif
957 
958 	KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
959 	    object->type == OBJT_SWAP,
960 	    ("%s: non-swap obj %p has cred", __func__, object));
961 
962 	/*
963 	 * Let the pager know object is dead.
964 	 */
965 	vm_pager_deallocate(object);
966 	VM_OBJECT_WUNLOCK(object);
967 
968 	vm_object_destroy(object);
969 }
970 
971 /*
972  * Make the page read-only so that we can clear the object flags.  However, if
973  * this is a nosync mmap then the object is likely to stay dirty so do not
974  * mess with the page and do not clear the object flags.  Returns TRUE if the
975  * page should be flushed, and FALSE otherwise.
976  */
977 static boolean_t
978 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
979 {
980 
981 	vm_page_assert_busied(p);
982 
983 	/*
984 	 * If we have been asked to skip nosync pages and this is a
985 	 * nosync page, skip it.  Note that the object flags were not
986 	 * cleared in this case so we do not have to set them.
987 	 */
988 	if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
989 		*allclean = FALSE;
990 		return (FALSE);
991 	} else {
992 		pmap_remove_write(p);
993 		return (p->dirty != 0);
994 	}
995 }
996 
997 /*
998  *	vm_object_page_clean
999  *
1000  *	Clean all dirty pages in the specified range of object.  Leaves page
1001  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
1002  *	write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1003  *	leaving the object dirty.
1004  *
1005  *	For swap objects backing tmpfs regular files, do not flush anything,
1006  *	but remove write protection on the mapped pages to update mtime through
1007  *	mmaped writes.
1008  *
1009  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
1010  *	synchronous clustering mode implementation.
1011  *
1012  *	Odd semantics: if start == end, we clean everything.
1013  *
1014  *	The object must be locked.
1015  *
1016  *	Returns FALSE if some page from the range was not written, as
1017  *	reported by the pager, and TRUE otherwise.
1018  */
1019 boolean_t
1020 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1021     int flags)
1022 {
1023 	vm_page_t np, p;
1024 	vm_pindex_t pi, tend, tstart;
1025 	int curgeneration, n, pagerflags;
1026 	boolean_t eio, res, allclean;
1027 
1028 	VM_OBJECT_ASSERT_WLOCKED(object);
1029 
1030 	if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1031 		return (TRUE);
1032 
1033 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1034 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1035 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1036 
1037 	tstart = OFF_TO_IDX(start);
1038 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1039 	allclean = tstart == 0 && tend >= object->size;
1040 	res = TRUE;
1041 
1042 rescan:
1043 	curgeneration = object->generation;
1044 
1045 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1046 		pi = p->pindex;
1047 		if (pi >= tend)
1048 			break;
1049 		np = TAILQ_NEXT(p, listq);
1050 		if (vm_page_none_valid(p))
1051 			continue;
1052 		if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1053 			if (object->generation != curgeneration &&
1054 			    (flags & OBJPC_SYNC) != 0)
1055 				goto rescan;
1056 			np = vm_page_find_least(object, pi);
1057 			continue;
1058 		}
1059 		if (!vm_object_page_remove_write(p, flags, &allclean)) {
1060 			vm_page_xunbusy(p);
1061 			continue;
1062 		}
1063 		if (object->type == OBJT_VNODE) {
1064 			n = vm_object_page_collect_flush(object, p, pagerflags,
1065 			    flags, &allclean, &eio);
1066 			if (eio) {
1067 				res = FALSE;
1068 				allclean = FALSE;
1069 			}
1070 			if (object->generation != curgeneration &&
1071 			    (flags & OBJPC_SYNC) != 0)
1072 				goto rescan;
1073 
1074 			/*
1075 			 * If the VOP_PUTPAGES() did a truncated write, so
1076 			 * that even the first page of the run is not fully
1077 			 * written, vm_pageout_flush() returns 0 as the run
1078 			 * length.  Since the condition that caused truncated
1079 			 * write may be permanent, e.g. exhausted free space,
1080 			 * accepting n == 0 would cause an infinite loop.
1081 			 *
1082 			 * Forwarding the iterator leaves the unwritten page
1083 			 * behind, but there is not much we can do there if
1084 			 * filesystem refuses to write it.
1085 			 */
1086 			if (n == 0) {
1087 				n = 1;
1088 				allclean = FALSE;
1089 			}
1090 		} else {
1091 			n = 1;
1092 			vm_page_xunbusy(p);
1093 		}
1094 		np = vm_page_find_least(object, pi + n);
1095 	}
1096 #if 0
1097 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1098 #endif
1099 
1100 	/*
1101 	 * Leave updating cleangeneration for tmpfs objects to tmpfs
1102 	 * scan.  It needs to update mtime, which happens for other
1103 	 * filesystems during page writeouts.
1104 	 */
1105 	if (allclean && object->type == OBJT_VNODE)
1106 		object->cleangeneration = curgeneration;
1107 	return (res);
1108 }
1109 
1110 static int
1111 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1112     int flags, boolean_t *allclean, boolean_t *eio)
1113 {
1114 	vm_page_t ma[vm_pageout_page_count], p_first, tp;
1115 	int count, i, mreq, runlen;
1116 
1117 	vm_page_lock_assert(p, MA_NOTOWNED);
1118 	vm_page_assert_xbusied(p);
1119 	VM_OBJECT_ASSERT_WLOCKED(object);
1120 
1121 	count = 1;
1122 	mreq = 0;
1123 
1124 	for (tp = p; count < vm_pageout_page_count; count++) {
1125 		tp = vm_page_next(tp);
1126 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1127 			break;
1128 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1129 			vm_page_xunbusy(tp);
1130 			break;
1131 		}
1132 	}
1133 
1134 	for (p_first = p; count < vm_pageout_page_count; count++) {
1135 		tp = vm_page_prev(p_first);
1136 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1137 			break;
1138 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1139 			vm_page_xunbusy(tp);
1140 			break;
1141 		}
1142 		p_first = tp;
1143 		mreq++;
1144 	}
1145 
1146 	for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1147 		ma[i] = tp;
1148 
1149 	vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1150 	return (runlen);
1151 }
1152 
1153 /*
1154  * Note that there is absolutely no sense in writing out
1155  * anonymous objects, so we track down the vnode object
1156  * to write out.
1157  * We invalidate (remove) all pages from the address space
1158  * for semantic correctness.
1159  *
1160  * If the backing object is a device object with unmanaged pages, then any
1161  * mappings to the specified range of pages must be removed before this
1162  * function is called.
1163  *
1164  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1165  * may start out with a NULL object.
1166  */
1167 boolean_t
1168 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1169     boolean_t syncio, boolean_t invalidate)
1170 {
1171 	vm_object_t backing_object;
1172 	struct vnode *vp;
1173 	struct mount *mp;
1174 	int error, flags, fsync_after;
1175 	boolean_t res;
1176 
1177 	if (object == NULL)
1178 		return (TRUE);
1179 	res = TRUE;
1180 	error = 0;
1181 	VM_OBJECT_WLOCK(object);
1182 	while ((backing_object = object->backing_object) != NULL) {
1183 		VM_OBJECT_WLOCK(backing_object);
1184 		offset += object->backing_object_offset;
1185 		VM_OBJECT_WUNLOCK(object);
1186 		object = backing_object;
1187 		if (object->size < OFF_TO_IDX(offset + size))
1188 			size = IDX_TO_OFF(object->size) - offset;
1189 	}
1190 	/*
1191 	 * Flush pages if writing is allowed, invalidate them
1192 	 * if invalidation requested.  Pages undergoing I/O
1193 	 * will be ignored by vm_object_page_remove().
1194 	 *
1195 	 * We cannot lock the vnode and then wait for paging
1196 	 * to complete without deadlocking against vm_fault.
1197 	 * Instead we simply call vm_object_page_remove() and
1198 	 * allow it to block internally on a page-by-page
1199 	 * basis when it encounters pages undergoing async
1200 	 * I/O.
1201 	 */
1202 	if (object->type == OBJT_VNODE &&
1203 	    vm_object_mightbedirty(object) != 0 &&
1204 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1205 		VM_OBJECT_WUNLOCK(object);
1206 		(void) vn_start_write(vp, &mp, V_WAIT);
1207 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1208 		if (syncio && !invalidate && offset == 0 &&
1209 		    atop(size) == object->size) {
1210 			/*
1211 			 * If syncing the whole mapping of the file,
1212 			 * it is faster to schedule all the writes in
1213 			 * async mode, also allowing the clustering,
1214 			 * and then wait for i/o to complete.
1215 			 */
1216 			flags = 0;
1217 			fsync_after = TRUE;
1218 		} else {
1219 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1220 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1221 			fsync_after = FALSE;
1222 		}
1223 		VM_OBJECT_WLOCK(object);
1224 		res = vm_object_page_clean(object, offset, offset + size,
1225 		    flags);
1226 		VM_OBJECT_WUNLOCK(object);
1227 		if (fsync_after)
1228 			error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1229 		VOP_UNLOCK(vp);
1230 		vn_finished_write(mp);
1231 		if (error != 0)
1232 			res = FALSE;
1233 		VM_OBJECT_WLOCK(object);
1234 	}
1235 	if ((object->type == OBJT_VNODE ||
1236 	     object->type == OBJT_DEVICE) && invalidate) {
1237 		if (object->type == OBJT_DEVICE)
1238 			/*
1239 			 * The option OBJPR_NOTMAPPED must be passed here
1240 			 * because vm_object_page_remove() cannot remove
1241 			 * unmanaged mappings.
1242 			 */
1243 			flags = OBJPR_NOTMAPPED;
1244 		else if (old_msync)
1245 			flags = 0;
1246 		else
1247 			flags = OBJPR_CLEANONLY;
1248 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1249 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1250 	}
1251 	VM_OBJECT_WUNLOCK(object);
1252 	return (res);
1253 }
1254 
1255 /*
1256  * Determine whether the given advice can be applied to the object.  Advice is
1257  * not applied to unmanaged pages since they never belong to page queues, and
1258  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1259  * have been mapped at most once.
1260  */
1261 static bool
1262 vm_object_advice_applies(vm_object_t object, int advice)
1263 {
1264 
1265 	if ((object->flags & OBJ_UNMANAGED) != 0)
1266 		return (false);
1267 	if (advice != MADV_FREE)
1268 		return (true);
1269 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1270 	    (OBJ_ONEMAPPING | OBJ_ANON));
1271 }
1272 
1273 static void
1274 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1275     vm_size_t size)
1276 {
1277 
1278 	if (advice == MADV_FREE && object->type == OBJT_SWAP)
1279 		swap_pager_freespace(object, pindex, size);
1280 }
1281 
1282 /*
1283  *	vm_object_madvise:
1284  *
1285  *	Implements the madvise function at the object/page level.
1286  *
1287  *	MADV_WILLNEED	(any object)
1288  *
1289  *	    Activate the specified pages if they are resident.
1290  *
1291  *	MADV_DONTNEED	(any object)
1292  *
1293  *	    Deactivate the specified pages if they are resident.
1294  *
1295  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1296  *			 OBJ_ONEMAPPING only)
1297  *
1298  *	    Deactivate and clean the specified pages if they are
1299  *	    resident.  This permits the process to reuse the pages
1300  *	    without faulting or the kernel to reclaim the pages
1301  *	    without I/O.
1302  */
1303 void
1304 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1305     int advice)
1306 {
1307 	vm_pindex_t tpindex;
1308 	vm_object_t backing_object, tobject;
1309 	vm_page_t m, tm;
1310 
1311 	if (object == NULL)
1312 		return;
1313 
1314 relookup:
1315 	VM_OBJECT_WLOCK(object);
1316 	if (!vm_object_advice_applies(object, advice)) {
1317 		VM_OBJECT_WUNLOCK(object);
1318 		return;
1319 	}
1320 	for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1321 		tobject = object;
1322 
1323 		/*
1324 		 * If the next page isn't resident in the top-level object, we
1325 		 * need to search the shadow chain.  When applying MADV_FREE, we
1326 		 * take care to release any swap space used to store
1327 		 * non-resident pages.
1328 		 */
1329 		if (m == NULL || pindex < m->pindex) {
1330 			/*
1331 			 * Optimize a common case: if the top-level object has
1332 			 * no backing object, we can skip over the non-resident
1333 			 * range in constant time.
1334 			 */
1335 			if (object->backing_object == NULL) {
1336 				tpindex = (m != NULL && m->pindex < end) ?
1337 				    m->pindex : end;
1338 				vm_object_madvise_freespace(object, advice,
1339 				    pindex, tpindex - pindex);
1340 				if ((pindex = tpindex) == end)
1341 					break;
1342 				goto next_page;
1343 			}
1344 
1345 			tpindex = pindex;
1346 			do {
1347 				vm_object_madvise_freespace(tobject, advice,
1348 				    tpindex, 1);
1349 				/*
1350 				 * Prepare to search the next object in the
1351 				 * chain.
1352 				 */
1353 				backing_object = tobject->backing_object;
1354 				if (backing_object == NULL)
1355 					goto next_pindex;
1356 				VM_OBJECT_WLOCK(backing_object);
1357 				tpindex +=
1358 				    OFF_TO_IDX(tobject->backing_object_offset);
1359 				if (tobject != object)
1360 					VM_OBJECT_WUNLOCK(tobject);
1361 				tobject = backing_object;
1362 				if (!vm_object_advice_applies(tobject, advice))
1363 					goto next_pindex;
1364 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1365 			    NULL);
1366 		} else {
1367 next_page:
1368 			tm = m;
1369 			m = TAILQ_NEXT(m, listq);
1370 		}
1371 
1372 		/*
1373 		 * If the page is not in a normal state, skip it.  The page
1374 		 * can not be invalidated while the object lock is held.
1375 		 */
1376 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1377 			goto next_pindex;
1378 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1379 		    ("vm_object_madvise: page %p is fictitious", tm));
1380 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1381 		    ("vm_object_madvise: page %p is not managed", tm));
1382 		if (vm_page_tryxbusy(tm) == 0) {
1383 			if (object != tobject)
1384 				VM_OBJECT_WUNLOCK(object);
1385 			if (advice == MADV_WILLNEED) {
1386 				/*
1387 				 * Reference the page before unlocking and
1388 				 * sleeping so that the page daemon is less
1389 				 * likely to reclaim it.
1390 				 */
1391 				vm_page_aflag_set(tm, PGA_REFERENCED);
1392 			}
1393 			vm_page_busy_sleep(tm, "madvpo", false);
1394   			goto relookup;
1395 		}
1396 		vm_page_advise(tm, advice);
1397 		vm_page_xunbusy(tm);
1398 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1399 next_pindex:
1400 		if (tobject != object)
1401 			VM_OBJECT_WUNLOCK(tobject);
1402 	}
1403 	VM_OBJECT_WUNLOCK(object);
1404 }
1405 
1406 /*
1407  *	vm_object_shadow:
1408  *
1409  *	Create a new object which is backed by the
1410  *	specified existing object range.  The source
1411  *	object reference is deallocated.
1412  *
1413  *	The new object and offset into that object
1414  *	are returned in the source parameters.
1415  */
1416 void
1417 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1418     struct ucred *cred, bool shared)
1419 {
1420 	vm_object_t source;
1421 	vm_object_t result;
1422 
1423 	source = *object;
1424 
1425 	/*
1426 	 * Don't create the new object if the old object isn't shared.
1427 	 *
1428 	 * If we hold the only reference we can guarantee that it won't
1429 	 * increase while we have the map locked.  Otherwise the race is
1430 	 * harmless and we will end up with an extra shadow object that
1431 	 * will be collapsed later.
1432 	 */
1433 	if (source != NULL && source->ref_count == 1 &&
1434 	    (source->flags & OBJ_ANON) != 0)
1435 		return;
1436 
1437 	/*
1438 	 * Allocate a new object with the given length.
1439 	 */
1440 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1441 
1442 	/*
1443 	 * Store the offset into the source object, and fix up the offset into
1444 	 * the new object.
1445 	 */
1446 	result->backing_object_offset = *offset;
1447 
1448 	if (shared || source != NULL) {
1449 		VM_OBJECT_WLOCK(result);
1450 
1451 		/*
1452 		 * The new object shadows the source object, adding a
1453 		 * reference to it.  Our caller changes his reference
1454 		 * to point to the new object, removing a reference to
1455 		 * the source object.  Net result: no change of
1456 		 * reference count, unless the caller needs to add one
1457 		 * more reference due to forking a shared map entry.
1458 		 */
1459 		if (shared) {
1460 			vm_object_reference_locked(result);
1461 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1462 		}
1463 
1464 		/*
1465 		 * Try to optimize the result object's page color when
1466 		 * shadowing in order to maintain page coloring
1467 		 * consistency in the combined shadowed object.
1468 		 */
1469 		if (source != NULL) {
1470 			vm_object_backing_insert(result, source);
1471 			result->domain = source->domain;
1472 #if VM_NRESERVLEVEL > 0
1473 			result->flags |= source->flags & OBJ_COLORED;
1474 			result->pg_color = (source->pg_color +
1475 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1476 			    1)) - 1);
1477 #endif
1478 		}
1479 		VM_OBJECT_WUNLOCK(result);
1480 	}
1481 
1482 	/*
1483 	 * Return the new things
1484 	 */
1485 	*offset = 0;
1486 	*object = result;
1487 }
1488 
1489 /*
1490  *	vm_object_split:
1491  *
1492  * Split the pages in a map entry into a new object.  This affords
1493  * easier removal of unused pages, and keeps object inheritance from
1494  * being a negative impact on memory usage.
1495  */
1496 void
1497 vm_object_split(vm_map_entry_t entry)
1498 {
1499 	vm_page_t m, m_busy, m_next;
1500 	vm_object_t orig_object, new_object, backing_object;
1501 	vm_pindex_t idx, offidxstart;
1502 	vm_size_t size;
1503 
1504 	orig_object = entry->object.vm_object;
1505 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1506 	    ("vm_object_split:  Splitting object with multiple mappings."));
1507 	if ((orig_object->flags & OBJ_ANON) == 0)
1508 		return;
1509 	if (orig_object->ref_count <= 1)
1510 		return;
1511 	VM_OBJECT_WUNLOCK(orig_object);
1512 
1513 	offidxstart = OFF_TO_IDX(entry->offset);
1514 	size = atop(entry->end - entry->start);
1515 
1516 	/*
1517 	 * If swap_pager_copy() is later called, it will convert new_object
1518 	 * into a swap object.
1519 	 */
1520 	new_object = vm_object_allocate_anon(size, orig_object,
1521 	    orig_object->cred, ptoa(size));
1522 
1523 	/*
1524 	 * We must wait for the orig_object to complete any in-progress
1525 	 * collapse so that the swap blocks are stable below.  The
1526 	 * additional reference on backing_object by new object will
1527 	 * prevent further collapse operations until split completes.
1528 	 */
1529 	VM_OBJECT_WLOCK(orig_object);
1530 	vm_object_collapse_wait(orig_object);
1531 
1532 	/*
1533 	 * At this point, the new object is still private, so the order in
1534 	 * which the original and new objects are locked does not matter.
1535 	 */
1536 	VM_OBJECT_WLOCK(new_object);
1537 	new_object->domain = orig_object->domain;
1538 	backing_object = orig_object->backing_object;
1539 	if (backing_object != NULL) {
1540 		vm_object_backing_insert_ref(new_object, backing_object);
1541 		new_object->backing_object_offset =
1542 		    orig_object->backing_object_offset + entry->offset;
1543 	}
1544 	if (orig_object->cred != NULL) {
1545 		crhold(orig_object->cred);
1546 		KASSERT(orig_object->charge >= ptoa(size),
1547 		    ("orig_object->charge < 0"));
1548 		orig_object->charge -= ptoa(size);
1549 	}
1550 
1551 	/*
1552 	 * Mark the split operation so that swap_pager_getpages() knows
1553 	 * that the object is in transition.
1554 	 */
1555 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1556 	m_busy = NULL;
1557 #ifdef INVARIANTS
1558 	idx = 0;
1559 #endif
1560 retry:
1561 	m = vm_page_find_least(orig_object, offidxstart);
1562 	KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1563 	    ("%s: object %p was repopulated", __func__, orig_object));
1564 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1565 	    m = m_next) {
1566 		m_next = TAILQ_NEXT(m, listq);
1567 
1568 		/*
1569 		 * We must wait for pending I/O to complete before we can
1570 		 * rename the page.
1571 		 *
1572 		 * We do not have to VM_PROT_NONE the page as mappings should
1573 		 * not be changed by this operation.
1574 		 */
1575 		if (vm_page_tryxbusy(m) == 0) {
1576 			VM_OBJECT_WUNLOCK(new_object);
1577 			vm_page_sleep_if_busy(m, "spltwt");
1578 			VM_OBJECT_WLOCK(new_object);
1579 			goto retry;
1580 		}
1581 
1582 		/*
1583 		 * The page was left invalid.  Likely placed there by
1584 		 * an incomplete fault.  Just remove and ignore.
1585 		 */
1586 		if (vm_page_none_valid(m)) {
1587 			if (vm_page_remove(m))
1588 				vm_page_free(m);
1589 			continue;
1590 		}
1591 
1592 		/* vm_page_rename() will dirty the page. */
1593 		if (vm_page_rename(m, new_object, idx)) {
1594 			vm_page_xunbusy(m);
1595 			VM_OBJECT_WUNLOCK(new_object);
1596 			VM_OBJECT_WUNLOCK(orig_object);
1597 			vm_radix_wait();
1598 			VM_OBJECT_WLOCK(orig_object);
1599 			VM_OBJECT_WLOCK(new_object);
1600 			goto retry;
1601 		}
1602 
1603 #if VM_NRESERVLEVEL > 0
1604 		/*
1605 		 * If some of the reservation's allocated pages remain with
1606 		 * the original object, then transferring the reservation to
1607 		 * the new object is neither particularly beneficial nor
1608 		 * particularly harmful as compared to leaving the reservation
1609 		 * with the original object.  If, however, all of the
1610 		 * reservation's allocated pages are transferred to the new
1611 		 * object, then transferring the reservation is typically
1612 		 * beneficial.  Determining which of these two cases applies
1613 		 * would be more costly than unconditionally renaming the
1614 		 * reservation.
1615 		 */
1616 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1617 #endif
1618 
1619 		/*
1620 		 * orig_object's type may change while sleeping, so keep track
1621 		 * of the beginning of the busied range.
1622 		 */
1623 		if (orig_object->type != OBJT_SWAP)
1624 			vm_page_xunbusy(m);
1625 		else if (m_busy == NULL)
1626 			m_busy = m;
1627 	}
1628 	if (orig_object->type == OBJT_SWAP) {
1629 		/*
1630 		 * swap_pager_copy() can sleep, in which case the orig_object's
1631 		 * and new_object's locks are released and reacquired.
1632 		 */
1633 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1634 		if (m_busy != NULL)
1635 			TAILQ_FOREACH_FROM(m_busy, &new_object->memq, listq)
1636 				vm_page_xunbusy(m_busy);
1637 	}
1638 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1639 	VM_OBJECT_WUNLOCK(orig_object);
1640 	VM_OBJECT_WUNLOCK(new_object);
1641 	entry->object.vm_object = new_object;
1642 	entry->offset = 0LL;
1643 	vm_object_deallocate(orig_object);
1644 	VM_OBJECT_WLOCK(new_object);
1645 }
1646 
1647 static vm_page_t
1648 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1649 {
1650 	vm_object_t backing_object;
1651 
1652 	VM_OBJECT_ASSERT_WLOCKED(object);
1653 	backing_object = object->backing_object;
1654 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1655 
1656 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1657 	    ("invalid ownership %p %p %p", p, object, backing_object));
1658 	/* The page is only NULL when rename fails. */
1659 	if (p == NULL) {
1660 		VM_OBJECT_WUNLOCK(object);
1661 		VM_OBJECT_WUNLOCK(backing_object);
1662 		vm_radix_wait();
1663 	} else {
1664 		if (p->object == object)
1665 			VM_OBJECT_WUNLOCK(backing_object);
1666 		else
1667 			VM_OBJECT_WUNLOCK(object);
1668 		vm_page_busy_sleep(p, "vmocol", false);
1669 	}
1670 	VM_OBJECT_WLOCK(object);
1671 	VM_OBJECT_WLOCK(backing_object);
1672 	return (TAILQ_FIRST(&backing_object->memq));
1673 }
1674 
1675 static bool
1676 vm_object_scan_all_shadowed(vm_object_t object)
1677 {
1678 	vm_object_t backing_object;
1679 	vm_page_t p, pp;
1680 	vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1681 
1682 	VM_OBJECT_ASSERT_WLOCKED(object);
1683 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1684 
1685 	backing_object = object->backing_object;
1686 
1687 	if ((backing_object->flags & OBJ_ANON) == 0)
1688 		return (false);
1689 
1690 	pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1691 	p = vm_page_find_least(backing_object, pi);
1692 	ps = swap_pager_find_least(backing_object, pi);
1693 
1694 	/*
1695 	 * Only check pages inside the parent object's range and
1696 	 * inside the parent object's mapping of the backing object.
1697 	 */
1698 	for (;; pi++) {
1699 		if (p != NULL && p->pindex < pi)
1700 			p = TAILQ_NEXT(p, listq);
1701 		if (ps < pi)
1702 			ps = swap_pager_find_least(backing_object, pi);
1703 		if (p == NULL && ps >= backing_object->size)
1704 			break;
1705 		else if (p == NULL)
1706 			pi = ps;
1707 		else
1708 			pi = MIN(p->pindex, ps);
1709 
1710 		new_pindex = pi - backing_offset_index;
1711 		if (new_pindex >= object->size)
1712 			break;
1713 
1714 		if (p != NULL) {
1715 			/*
1716 			 * If the backing object page is busy a
1717 			 * grandparent or older page may still be
1718 			 * undergoing CoW.  It is not safe to collapse
1719 			 * the backing object until it is quiesced.
1720 			 */
1721 			if (vm_page_tryxbusy(p) == 0)
1722 				return (false);
1723 
1724 			/*
1725 			 * We raced with the fault handler that left
1726 			 * newly allocated invalid page on the object
1727 			 * queue and retried.
1728 			 */
1729 			if (!vm_page_all_valid(p))
1730 				goto unbusy_ret;
1731 		}
1732 
1733 		/*
1734 		 * See if the parent has the page or if the parent's object
1735 		 * pager has the page.  If the parent has the page but the page
1736 		 * is not valid, the parent's object pager must have the page.
1737 		 *
1738 		 * If this fails, the parent does not completely shadow the
1739 		 * object and we might as well give up now.
1740 		 */
1741 		pp = vm_page_lookup(object, new_pindex);
1742 
1743 		/*
1744 		 * The valid check here is stable due to object lock
1745 		 * being required to clear valid and initiate paging.
1746 		 * Busy of p disallows fault handler to validate pp.
1747 		 */
1748 		if ((pp == NULL || vm_page_none_valid(pp)) &&
1749 		    !vm_pager_has_page(object, new_pindex, NULL, NULL))
1750 			goto unbusy_ret;
1751 		if (p != NULL)
1752 			vm_page_xunbusy(p);
1753 	}
1754 	return (true);
1755 
1756 unbusy_ret:
1757 	if (p != NULL)
1758 		vm_page_xunbusy(p);
1759 	return (false);
1760 }
1761 
1762 static void
1763 vm_object_collapse_scan(vm_object_t object)
1764 {
1765 	vm_object_t backing_object;
1766 	vm_page_t next, p, pp;
1767 	vm_pindex_t backing_offset_index, new_pindex;
1768 
1769 	VM_OBJECT_ASSERT_WLOCKED(object);
1770 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1771 
1772 	backing_object = object->backing_object;
1773 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1774 
1775 	/*
1776 	 * Our scan
1777 	 */
1778 	for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1779 		next = TAILQ_NEXT(p, listq);
1780 		new_pindex = p->pindex - backing_offset_index;
1781 
1782 		/*
1783 		 * Check for busy page
1784 		 */
1785 		if (vm_page_tryxbusy(p) == 0) {
1786 			next = vm_object_collapse_scan_wait(object, p);
1787 			continue;
1788 		}
1789 
1790 		KASSERT(object->backing_object == backing_object,
1791 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1792 		    object->backing_object, backing_object));
1793 		KASSERT(p->object == backing_object,
1794 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1795 		    p->object, backing_object));
1796 
1797 		if (p->pindex < backing_offset_index ||
1798 		    new_pindex >= object->size) {
1799 			if (backing_object->type == OBJT_SWAP)
1800 				swap_pager_freespace(backing_object, p->pindex,
1801 				    1);
1802 
1803 			KASSERT(!pmap_page_is_mapped(p),
1804 			    ("freeing mapped page %p", p));
1805 			if (vm_page_remove(p))
1806 				vm_page_free(p);
1807 			continue;
1808 		}
1809 
1810 		if (!vm_page_all_valid(p)) {
1811 			KASSERT(!pmap_page_is_mapped(p),
1812 			    ("freeing mapped page %p", p));
1813 			if (vm_page_remove(p))
1814 				vm_page_free(p);
1815 			continue;
1816 		}
1817 
1818 		pp = vm_page_lookup(object, new_pindex);
1819 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1820 			vm_page_xunbusy(p);
1821 			/*
1822 			 * The page in the parent is busy and possibly not
1823 			 * (yet) valid.  Until its state is finalized by the
1824 			 * busy bit owner, we can't tell whether it shadows the
1825 			 * original page.
1826 			 */
1827 			next = vm_object_collapse_scan_wait(object, pp);
1828 			continue;
1829 		}
1830 
1831 		if (pp != NULL && vm_page_none_valid(pp)) {
1832 			/*
1833 			 * The page was invalid in the parent.  Likely placed
1834 			 * there by an incomplete fault.  Just remove and
1835 			 * ignore.  p can replace it.
1836 			 */
1837 			if (vm_page_remove(pp))
1838 				vm_page_free(pp);
1839 			pp = NULL;
1840 		}
1841 
1842 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1843 			NULL)) {
1844 			/*
1845 			 * The page already exists in the parent OR swap exists
1846 			 * for this location in the parent.  Leave the parent's
1847 			 * page alone.  Destroy the original page from the
1848 			 * backing object.
1849 			 */
1850 			if (backing_object->type == OBJT_SWAP)
1851 				swap_pager_freespace(backing_object, p->pindex,
1852 				    1);
1853 			KASSERT(!pmap_page_is_mapped(p),
1854 			    ("freeing mapped page %p", p));
1855 			if (vm_page_remove(p))
1856 				vm_page_free(p);
1857 			if (pp != NULL)
1858 				vm_page_xunbusy(pp);
1859 			continue;
1860 		}
1861 
1862 		/*
1863 		 * Page does not exist in parent, rename the page from the
1864 		 * backing object to the main object.
1865 		 *
1866 		 * If the page was mapped to a process, it can remain mapped
1867 		 * through the rename.  vm_page_rename() will dirty the page.
1868 		 */
1869 		if (vm_page_rename(p, object, new_pindex)) {
1870 			vm_page_xunbusy(p);
1871 			next = vm_object_collapse_scan_wait(object, NULL);
1872 			continue;
1873 		}
1874 
1875 		/* Use the old pindex to free the right page. */
1876 		if (backing_object->type == OBJT_SWAP)
1877 			swap_pager_freespace(backing_object,
1878 			    new_pindex + backing_offset_index, 1);
1879 
1880 #if VM_NRESERVLEVEL > 0
1881 		/*
1882 		 * Rename the reservation.
1883 		 */
1884 		vm_reserv_rename(p, object, backing_object,
1885 		    backing_offset_index);
1886 #endif
1887 		vm_page_xunbusy(p);
1888 	}
1889 	return;
1890 }
1891 
1892 /*
1893  *	vm_object_collapse:
1894  *
1895  *	Collapse an object with the object backing it.
1896  *	Pages in the backing object are moved into the
1897  *	parent, and the backing object is deallocated.
1898  */
1899 void
1900 vm_object_collapse(vm_object_t object)
1901 {
1902 	vm_object_t backing_object, new_backing_object;
1903 
1904 	VM_OBJECT_ASSERT_WLOCKED(object);
1905 
1906 	while (TRUE) {
1907 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1908 		    ("collapsing invalid object"));
1909 
1910 		/*
1911 		 * Wait for the backing_object to finish any pending
1912 		 * collapse so that the caller sees the shortest possible
1913 		 * shadow chain.
1914 		 */
1915 		backing_object = vm_object_backing_collapse_wait(object);
1916 		if (backing_object == NULL)
1917 			return;
1918 
1919 		KASSERT(object->ref_count > 0 &&
1920 		    object->ref_count > object->shadow_count,
1921 		    ("collapse with invalid ref %d or shadow %d count.",
1922 		    object->ref_count, object->shadow_count));
1923 		KASSERT((backing_object->flags &
1924 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1925 		    ("vm_object_collapse: Backing object already collapsing."));
1926 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1927 		    ("vm_object_collapse: object is already collapsing."));
1928 
1929 		/*
1930 		 * We know that we can either collapse the backing object if
1931 		 * the parent is the only reference to it, or (perhaps) have
1932 		 * the parent bypass the object if the parent happens to shadow
1933 		 * all the resident pages in the entire backing object.
1934 		 */
1935 		if (backing_object->ref_count == 1) {
1936 			KASSERT(backing_object->shadow_count == 1,
1937 			    ("vm_object_collapse: shadow_count: %d",
1938 			    backing_object->shadow_count));
1939 			vm_object_pip_add(object, 1);
1940 			vm_object_set_flag(object, OBJ_COLLAPSING);
1941 			vm_object_pip_add(backing_object, 1);
1942 			vm_object_set_flag(backing_object, OBJ_DEAD);
1943 
1944 			/*
1945 			 * If there is exactly one reference to the backing
1946 			 * object, we can collapse it into the parent.
1947 			 */
1948 			vm_object_collapse_scan(object);
1949 
1950 #if VM_NRESERVLEVEL > 0
1951 			/*
1952 			 * Break any reservations from backing_object.
1953 			 */
1954 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1955 				vm_reserv_break_all(backing_object);
1956 #endif
1957 
1958 			/*
1959 			 * Move the pager from backing_object to object.
1960 			 */
1961 			if (backing_object->type == OBJT_SWAP) {
1962 				/*
1963 				 * swap_pager_copy() can sleep, in which case
1964 				 * the backing_object's and object's locks are
1965 				 * released and reacquired.
1966 				 * Since swap_pager_copy() is being asked to
1967 				 * destroy backing_object, it will change the
1968 				 * type to OBJT_DEFAULT.
1969 				 */
1970 				swap_pager_copy(
1971 				    backing_object,
1972 				    object,
1973 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1974 			}
1975 
1976 			/*
1977 			 * Object now shadows whatever backing_object did.
1978 			 */
1979 			vm_object_clear_flag(object, OBJ_COLLAPSING);
1980 			vm_object_backing_transfer(object, backing_object);
1981 			object->backing_object_offset +=
1982 			    backing_object->backing_object_offset;
1983 			VM_OBJECT_WUNLOCK(object);
1984 			vm_object_pip_wakeup(object);
1985 
1986 			/*
1987 			 * Discard backing_object.
1988 			 *
1989 			 * Since the backing object has no pages, no pager left,
1990 			 * and no object references within it, all that is
1991 			 * necessary is to dispose of it.
1992 			 */
1993 			KASSERT(backing_object->ref_count == 1, (
1994 "backing_object %p was somehow re-referenced during collapse!",
1995 			    backing_object));
1996 			vm_object_pip_wakeup(backing_object);
1997 			(void)refcount_release(&backing_object->ref_count);
1998 			vm_object_terminate(backing_object);
1999 			counter_u64_add(object_collapses, 1);
2000 			VM_OBJECT_WLOCK(object);
2001 		} else {
2002 			/*
2003 			 * If we do not entirely shadow the backing object,
2004 			 * there is nothing we can do so we give up.
2005 			 *
2006 			 * The object lock and backing_object lock must not
2007 			 * be dropped during this sequence.
2008 			 */
2009 			if (!vm_object_scan_all_shadowed(object)) {
2010 				VM_OBJECT_WUNLOCK(backing_object);
2011 				break;
2012 			}
2013 
2014 			/*
2015 			 * Make the parent shadow the next object in the
2016 			 * chain.  Deallocating backing_object will not remove
2017 			 * it, since its reference count is at least 2.
2018 			 */
2019 			vm_object_backing_remove_locked(object);
2020 			new_backing_object = backing_object->backing_object;
2021 			if (new_backing_object != NULL) {
2022 				vm_object_backing_insert_ref(object,
2023 				    new_backing_object);
2024 				object->backing_object_offset +=
2025 				    backing_object->backing_object_offset;
2026 			}
2027 
2028 			/*
2029 			 * Drop the reference count on backing_object. Since
2030 			 * its ref_count was at least 2, it will not vanish.
2031 			 */
2032 			(void)refcount_release(&backing_object->ref_count);
2033 			KASSERT(backing_object->ref_count >= 1, (
2034 "backing_object %p was somehow dereferenced during collapse!",
2035 			    backing_object));
2036 			VM_OBJECT_WUNLOCK(backing_object);
2037 			counter_u64_add(object_bypasses, 1);
2038 		}
2039 
2040 		/*
2041 		 * Try again with this object's new backing object.
2042 		 */
2043 	}
2044 }
2045 
2046 /*
2047  *	vm_object_page_remove:
2048  *
2049  *	For the given object, either frees or invalidates each of the
2050  *	specified pages.  In general, a page is freed.  However, if a page is
2051  *	wired for any reason other than the existence of a managed, wired
2052  *	mapping, then it may be invalidated but not removed from the object.
2053  *	Pages are specified by the given range ["start", "end") and the option
2054  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
2055  *	extends from "start" to the end of the object.  If the option
2056  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2057  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
2058  *	specified, then the pages within the specified range must have no
2059  *	mappings.  Otherwise, if this option is not specified, any mappings to
2060  *	the specified pages are removed before the pages are freed or
2061  *	invalidated.
2062  *
2063  *	In general, this operation should only be performed on objects that
2064  *	contain managed pages.  There are, however, two exceptions.  First, it
2065  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
2066  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2067  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
2068  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
2069  *
2070  *	The object must be locked.
2071  */
2072 void
2073 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2074     int options)
2075 {
2076 	vm_page_t p, next;
2077 
2078 	VM_OBJECT_ASSERT_WLOCKED(object);
2079 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2080 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2081 	    ("vm_object_page_remove: illegal options for object %p", object));
2082 	if (object->resident_page_count == 0)
2083 		return;
2084 	vm_object_pip_add(object, 1);
2085 again:
2086 	p = vm_page_find_least(object, start);
2087 
2088 	/*
2089 	 * Here, the variable "p" is either (1) the page with the least pindex
2090 	 * greater than or equal to the parameter "start" or (2) NULL.
2091 	 */
2092 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2093 		next = TAILQ_NEXT(p, listq);
2094 
2095 		/*
2096 		 * If the page is wired for any reason besides the existence
2097 		 * of managed, wired mappings, then it cannot be freed.  For
2098 		 * example, fictitious pages, which represent device memory,
2099 		 * are inherently wired and cannot be freed.  They can,
2100 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2101 		 * not specified.
2102 		 */
2103 		if (vm_page_tryxbusy(p) == 0) {
2104 			vm_page_sleep_if_busy(p, "vmopar");
2105 			goto again;
2106 		}
2107 		if (vm_page_wired(p)) {
2108 wired:
2109 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2110 			    object->ref_count != 0)
2111 				pmap_remove_all(p);
2112 			if ((options & OBJPR_CLEANONLY) == 0) {
2113 				vm_page_invalid(p);
2114 				vm_page_undirty(p);
2115 			}
2116 			vm_page_xunbusy(p);
2117 			continue;
2118 		}
2119 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2120 		    ("vm_object_page_remove: page %p is fictitious", p));
2121 		if ((options & OBJPR_CLEANONLY) != 0 &&
2122 		    !vm_page_none_valid(p)) {
2123 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2124 			    object->ref_count != 0 &&
2125 			    !vm_page_try_remove_write(p))
2126 				goto wired;
2127 			if (p->dirty != 0) {
2128 				vm_page_xunbusy(p);
2129 				continue;
2130 			}
2131 		}
2132 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2133 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2134 			goto wired;
2135 		vm_page_free(p);
2136 	}
2137 	vm_object_pip_wakeup(object);
2138 
2139 	if (object->type == OBJT_SWAP) {
2140 		if (end == 0)
2141 			end = object->size;
2142 		swap_pager_freespace(object, start, end - start);
2143 	}
2144 }
2145 
2146 /*
2147  *	vm_object_page_noreuse:
2148  *
2149  *	For the given object, attempt to move the specified pages to
2150  *	the head of the inactive queue.  This bypasses regular LRU
2151  *	operation and allows the pages to be reused quickly under memory
2152  *	pressure.  If a page is wired for any reason, then it will not
2153  *	be queued.  Pages are specified by the range ["start", "end").
2154  *	As a special case, if "end" is zero, then the range extends from
2155  *	"start" to the end of the object.
2156  *
2157  *	This operation should only be performed on objects that
2158  *	contain non-fictitious, managed pages.
2159  *
2160  *	The object must be locked.
2161  */
2162 void
2163 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2164 {
2165 	vm_page_t p, next;
2166 
2167 	VM_OBJECT_ASSERT_LOCKED(object);
2168 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2169 	    ("vm_object_page_noreuse: illegal object %p", object));
2170 	if (object->resident_page_count == 0)
2171 		return;
2172 	p = vm_page_find_least(object, start);
2173 
2174 	/*
2175 	 * Here, the variable "p" is either (1) the page with the least pindex
2176 	 * greater than or equal to the parameter "start" or (2) NULL.
2177 	 */
2178 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2179 		next = TAILQ_NEXT(p, listq);
2180 		vm_page_deactivate_noreuse(p);
2181 	}
2182 }
2183 
2184 /*
2185  *	Populate the specified range of the object with valid pages.  Returns
2186  *	TRUE if the range is successfully populated and FALSE otherwise.
2187  *
2188  *	Note: This function should be optimized to pass a larger array of
2189  *	pages to vm_pager_get_pages() before it is applied to a non-
2190  *	OBJT_DEVICE object.
2191  *
2192  *	The object must be locked.
2193  */
2194 boolean_t
2195 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2196 {
2197 	vm_page_t m;
2198 	vm_pindex_t pindex;
2199 	int rv;
2200 
2201 	VM_OBJECT_ASSERT_WLOCKED(object);
2202 	for (pindex = start; pindex < end; pindex++) {
2203 		rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2204 		if (rv != VM_PAGER_OK)
2205 			break;
2206 
2207 		/*
2208 		 * Keep "m" busy because a subsequent iteration may unlock
2209 		 * the object.
2210 		 */
2211 	}
2212 	if (pindex > start) {
2213 		m = vm_page_lookup(object, start);
2214 		while (m != NULL && m->pindex < pindex) {
2215 			vm_page_xunbusy(m);
2216 			m = TAILQ_NEXT(m, listq);
2217 		}
2218 	}
2219 	return (pindex == end);
2220 }
2221 
2222 /*
2223  *	Routine:	vm_object_coalesce
2224  *	Function:	Coalesces two objects backing up adjoining
2225  *			regions of memory into a single object.
2226  *
2227  *	returns TRUE if objects were combined.
2228  *
2229  *	NOTE:	Only works at the moment if the second object is NULL -
2230  *		if it's not, which object do we lock first?
2231  *
2232  *	Parameters:
2233  *		prev_object	First object to coalesce
2234  *		prev_offset	Offset into prev_object
2235  *		prev_size	Size of reference to prev_object
2236  *		next_size	Size of reference to the second object
2237  *		reserved	Indicator that extension region has
2238  *				swap accounted for
2239  *
2240  *	Conditions:
2241  *	The object must *not* be locked.
2242  */
2243 boolean_t
2244 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2245     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2246 {
2247 	vm_pindex_t next_pindex;
2248 
2249 	if (prev_object == NULL)
2250 		return (TRUE);
2251 	if ((prev_object->flags & OBJ_ANON) == 0)
2252 		return (FALSE);
2253 
2254 	VM_OBJECT_WLOCK(prev_object);
2255 	/*
2256 	 * Try to collapse the object first.
2257 	 */
2258 	vm_object_collapse(prev_object);
2259 
2260 	/*
2261 	 * Can't coalesce if: . more than one reference . paged out . shadows
2262 	 * another object . has a copy elsewhere (any of which mean that the
2263 	 * pages not mapped to prev_entry may be in use anyway)
2264 	 */
2265 	if (prev_object->backing_object != NULL) {
2266 		VM_OBJECT_WUNLOCK(prev_object);
2267 		return (FALSE);
2268 	}
2269 
2270 	prev_size >>= PAGE_SHIFT;
2271 	next_size >>= PAGE_SHIFT;
2272 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2273 
2274 	if (prev_object->ref_count > 1 &&
2275 	    prev_object->size != next_pindex &&
2276 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2277 		VM_OBJECT_WUNLOCK(prev_object);
2278 		return (FALSE);
2279 	}
2280 
2281 	/*
2282 	 * Account for the charge.
2283 	 */
2284 	if (prev_object->cred != NULL) {
2285 		/*
2286 		 * If prev_object was charged, then this mapping,
2287 		 * although not charged now, may become writable
2288 		 * later. Non-NULL cred in the object would prevent
2289 		 * swap reservation during enabling of the write
2290 		 * access, so reserve swap now. Failed reservation
2291 		 * cause allocation of the separate object for the map
2292 		 * entry, and swap reservation for this entry is
2293 		 * managed in appropriate time.
2294 		 */
2295 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2296 		    prev_object->cred)) {
2297 			VM_OBJECT_WUNLOCK(prev_object);
2298 			return (FALSE);
2299 		}
2300 		prev_object->charge += ptoa(next_size);
2301 	}
2302 
2303 	/*
2304 	 * Remove any pages that may still be in the object from a previous
2305 	 * deallocation.
2306 	 */
2307 	if (next_pindex < prev_object->size) {
2308 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2309 		    next_size, 0);
2310 #if 0
2311 		if (prev_object->cred != NULL) {
2312 			KASSERT(prev_object->charge >=
2313 			    ptoa(prev_object->size - next_pindex),
2314 			    ("object %p overcharged 1 %jx %jx", prev_object,
2315 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2316 			prev_object->charge -= ptoa(prev_object->size -
2317 			    next_pindex);
2318 		}
2319 #endif
2320 	}
2321 
2322 	/*
2323 	 * Extend the object if necessary.
2324 	 */
2325 	if (next_pindex + next_size > prev_object->size)
2326 		prev_object->size = next_pindex + next_size;
2327 
2328 	VM_OBJECT_WUNLOCK(prev_object);
2329 	return (TRUE);
2330 }
2331 
2332 void
2333 vm_object_set_writeable_dirty(vm_object_t object)
2334 {
2335 
2336 	/* Only set for vnodes & tmpfs */
2337 	if (object->type != OBJT_VNODE &&
2338 	    (object->flags & OBJ_TMPFS_NODE) == 0)
2339 		return;
2340 	atomic_add_int(&object->generation, 1);
2341 }
2342 
2343 /*
2344  *	vm_object_unwire:
2345  *
2346  *	For each page offset within the specified range of the given object,
2347  *	find the highest-level page in the shadow chain and unwire it.  A page
2348  *	must exist at every page offset, and the highest-level page must be
2349  *	wired.
2350  */
2351 void
2352 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2353     uint8_t queue)
2354 {
2355 	vm_object_t tobject, t1object;
2356 	vm_page_t m, tm;
2357 	vm_pindex_t end_pindex, pindex, tpindex;
2358 	int depth, locked_depth;
2359 
2360 	KASSERT((offset & PAGE_MASK) == 0,
2361 	    ("vm_object_unwire: offset is not page aligned"));
2362 	KASSERT((length & PAGE_MASK) == 0,
2363 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2364 	/* The wired count of a fictitious page never changes. */
2365 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2366 		return;
2367 	pindex = OFF_TO_IDX(offset);
2368 	end_pindex = pindex + atop(length);
2369 again:
2370 	locked_depth = 1;
2371 	VM_OBJECT_RLOCK(object);
2372 	m = vm_page_find_least(object, pindex);
2373 	while (pindex < end_pindex) {
2374 		if (m == NULL || pindex < m->pindex) {
2375 			/*
2376 			 * The first object in the shadow chain doesn't
2377 			 * contain a page at the current index.  Therefore,
2378 			 * the page must exist in a backing object.
2379 			 */
2380 			tobject = object;
2381 			tpindex = pindex;
2382 			depth = 0;
2383 			do {
2384 				tpindex +=
2385 				    OFF_TO_IDX(tobject->backing_object_offset);
2386 				tobject = tobject->backing_object;
2387 				KASSERT(tobject != NULL,
2388 				    ("vm_object_unwire: missing page"));
2389 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2390 					goto next_page;
2391 				depth++;
2392 				if (depth == locked_depth) {
2393 					locked_depth++;
2394 					VM_OBJECT_RLOCK(tobject);
2395 				}
2396 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2397 			    NULL);
2398 		} else {
2399 			tm = m;
2400 			m = TAILQ_NEXT(m, listq);
2401 		}
2402 		if (vm_page_trysbusy(tm) == 0) {
2403 			for (tobject = object; locked_depth >= 1;
2404 			    locked_depth--) {
2405 				t1object = tobject->backing_object;
2406 				if (tm->object != tobject)
2407 					VM_OBJECT_RUNLOCK(tobject);
2408 				tobject = t1object;
2409 			}
2410 			vm_page_busy_sleep(tm, "unwbo", true);
2411 			goto again;
2412 		}
2413 		vm_page_unwire(tm, queue);
2414 		vm_page_sunbusy(tm);
2415 next_page:
2416 		pindex++;
2417 	}
2418 	/* Release the accumulated object locks. */
2419 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2420 		t1object = tobject->backing_object;
2421 		VM_OBJECT_RUNLOCK(tobject);
2422 		tobject = t1object;
2423 	}
2424 }
2425 
2426 /*
2427  * Return the vnode for the given object, or NULL if none exists.
2428  * For tmpfs objects, the function may return NULL if there is
2429  * no vnode allocated at the time of the call.
2430  */
2431 struct vnode *
2432 vm_object_vnode(vm_object_t object)
2433 {
2434 	struct vnode *vp;
2435 
2436 	VM_OBJECT_ASSERT_LOCKED(object);
2437 	if (object->type == OBJT_VNODE) {
2438 		vp = object->handle;
2439 		KASSERT(vp != NULL, ("%s: OBJT_VNODE has no vnode", __func__));
2440 	} else if (object->type == OBJT_SWAP &&
2441 	    (object->flags & OBJ_TMPFS) != 0) {
2442 		vp = object->un_pager.swp.swp_tmpfs;
2443 		KASSERT(vp != NULL, ("%s: OBJT_TMPFS has no vnode", __func__));
2444 	} else {
2445 		vp = NULL;
2446 	}
2447 	return (vp);
2448 }
2449 
2450 /*
2451  * Busy the vm object.  This prevents new pages belonging to the object from
2452  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2453  * for checking page state before proceeding.
2454  */
2455 void
2456 vm_object_busy(vm_object_t obj)
2457 {
2458 
2459 	VM_OBJECT_ASSERT_LOCKED(obj);
2460 
2461 	blockcount_acquire(&obj->busy, 1);
2462 	/* The fence is required to order loads of page busy. */
2463 	atomic_thread_fence_acq_rel();
2464 }
2465 
2466 void
2467 vm_object_unbusy(vm_object_t obj)
2468 {
2469 
2470 	blockcount_release(&obj->busy, 1);
2471 }
2472 
2473 void
2474 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2475 {
2476 
2477 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2478 
2479 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2480 }
2481 
2482 /*
2483  * Return the kvme type of the given object.
2484  * If vpp is not NULL, set it to the object's vm_object_vnode() or NULL.
2485  */
2486 int
2487 vm_object_kvme_type(vm_object_t object, struct vnode **vpp)
2488 {
2489 
2490 	VM_OBJECT_ASSERT_LOCKED(object);
2491 	if (vpp != NULL)
2492 		*vpp = vm_object_vnode(object);
2493 	switch (object->type) {
2494 	case OBJT_DEFAULT:
2495 		return (KVME_TYPE_DEFAULT);
2496 	case OBJT_VNODE:
2497 		return (KVME_TYPE_VNODE);
2498 	case OBJT_SWAP:
2499 		if ((object->flags & OBJ_TMPFS_NODE) != 0)
2500 			return (KVME_TYPE_VNODE);
2501 		return (KVME_TYPE_SWAP);
2502 	case OBJT_DEVICE:
2503 		return (KVME_TYPE_DEVICE);
2504 	case OBJT_PHYS:
2505 		return (KVME_TYPE_PHYS);
2506 	case OBJT_DEAD:
2507 		return (KVME_TYPE_DEAD);
2508 	case OBJT_SG:
2509 		return (KVME_TYPE_SG);
2510 	case OBJT_MGTDEVICE:
2511 		return (KVME_TYPE_MGTDEVICE);
2512 	default:
2513 		return (KVME_TYPE_UNKNOWN);
2514 	}
2515 }
2516 
2517 static int
2518 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2519 {
2520 	struct kinfo_vmobject *kvo;
2521 	char *fullpath, *freepath;
2522 	struct vnode *vp;
2523 	struct vattr va;
2524 	vm_object_t obj;
2525 	vm_page_t m;
2526 	int count, error;
2527 
2528 	if (req->oldptr == NULL) {
2529 		/*
2530 		 * If an old buffer has not been provided, generate an
2531 		 * estimate of the space needed for a subsequent call.
2532 		 */
2533 		mtx_lock(&vm_object_list_mtx);
2534 		count = 0;
2535 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2536 			if (obj->type == OBJT_DEAD)
2537 				continue;
2538 			count++;
2539 		}
2540 		mtx_unlock(&vm_object_list_mtx);
2541 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2542 		    count * 11 / 10));
2543 	}
2544 
2545 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK);
2546 	error = 0;
2547 
2548 	/*
2549 	 * VM objects are type stable and are never removed from the
2550 	 * list once added.  This allows us to safely read obj->object_list
2551 	 * after reacquiring the VM object lock.
2552 	 */
2553 	mtx_lock(&vm_object_list_mtx);
2554 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2555 		if (obj->type == OBJT_DEAD)
2556 			continue;
2557 		VM_OBJECT_RLOCK(obj);
2558 		if (obj->type == OBJT_DEAD) {
2559 			VM_OBJECT_RUNLOCK(obj);
2560 			continue;
2561 		}
2562 		mtx_unlock(&vm_object_list_mtx);
2563 		kvo->kvo_size = ptoa(obj->size);
2564 		kvo->kvo_resident = obj->resident_page_count;
2565 		kvo->kvo_ref_count = obj->ref_count;
2566 		kvo->kvo_shadow_count = obj->shadow_count;
2567 		kvo->kvo_memattr = obj->memattr;
2568 		kvo->kvo_active = 0;
2569 		kvo->kvo_inactive = 0;
2570 		TAILQ_FOREACH(m, &obj->memq, listq) {
2571 			/*
2572 			 * A page may belong to the object but be
2573 			 * dequeued and set to PQ_NONE while the
2574 			 * object lock is not held.  This makes the
2575 			 * reads of m->queue below racy, and we do not
2576 			 * count pages set to PQ_NONE.  However, this
2577 			 * sysctl is only meant to give an
2578 			 * approximation of the system anyway.
2579 			 */
2580 			if (m->a.queue == PQ_ACTIVE)
2581 				kvo->kvo_active++;
2582 			else if (m->a.queue == PQ_INACTIVE)
2583 				kvo->kvo_inactive++;
2584 		}
2585 
2586 		kvo->kvo_vn_fileid = 0;
2587 		kvo->kvo_vn_fsid = 0;
2588 		kvo->kvo_vn_fsid_freebsd11 = 0;
2589 		freepath = NULL;
2590 		fullpath = "";
2591 		kvo->kvo_type = vm_object_kvme_type(obj, &vp);
2592 		if (vp != NULL)
2593 			vref(vp);
2594 		VM_OBJECT_RUNLOCK(obj);
2595 		if (vp != NULL) {
2596 			vn_fullpath(vp, &fullpath, &freepath);
2597 			vn_lock(vp, LK_SHARED | LK_RETRY);
2598 			if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2599 				kvo->kvo_vn_fileid = va.va_fileid;
2600 				kvo->kvo_vn_fsid = va.va_fsid;
2601 				kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2602 								/* truncate */
2603 			}
2604 			vput(vp);
2605 		}
2606 
2607 		strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2608 		if (freepath != NULL)
2609 			free(freepath, M_TEMP);
2610 
2611 		/* Pack record size down */
2612 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2613 		    + strlen(kvo->kvo_path) + 1;
2614 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2615 		    sizeof(uint64_t));
2616 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2617 		mtx_lock(&vm_object_list_mtx);
2618 		if (error)
2619 			break;
2620 	}
2621 	mtx_unlock(&vm_object_list_mtx);
2622 	free(kvo, M_TEMP);
2623 	return (error);
2624 }
2625 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2626     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2627     "List of VM objects");
2628 
2629 #include "opt_ddb.h"
2630 #ifdef DDB
2631 #include <sys/kernel.h>
2632 
2633 #include <sys/cons.h>
2634 
2635 #include <ddb/ddb.h>
2636 
2637 static int
2638 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2639 {
2640 	vm_map_t tmpm;
2641 	vm_map_entry_t tmpe;
2642 	vm_object_t obj;
2643 
2644 	if (map == 0)
2645 		return 0;
2646 
2647 	if (entry == 0) {
2648 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2649 			if (_vm_object_in_map(map, object, tmpe)) {
2650 				return 1;
2651 			}
2652 		}
2653 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2654 		tmpm = entry->object.sub_map;
2655 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2656 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2657 				return 1;
2658 			}
2659 		}
2660 	} else if ((obj = entry->object.vm_object) != NULL) {
2661 		for (; obj; obj = obj->backing_object)
2662 			if (obj == object) {
2663 				return 1;
2664 			}
2665 	}
2666 	return 0;
2667 }
2668 
2669 static int
2670 vm_object_in_map(vm_object_t object)
2671 {
2672 	struct proc *p;
2673 
2674 	/* sx_slock(&allproc_lock); */
2675 	FOREACH_PROC_IN_SYSTEM(p) {
2676 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2677 			continue;
2678 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2679 			/* sx_sunlock(&allproc_lock); */
2680 			return 1;
2681 		}
2682 	}
2683 	/* sx_sunlock(&allproc_lock); */
2684 	if (_vm_object_in_map(kernel_map, object, 0))
2685 		return 1;
2686 	return 0;
2687 }
2688 
2689 DB_SHOW_COMMAND(vmochk, vm_object_check)
2690 {
2691 	vm_object_t object;
2692 
2693 	/*
2694 	 * make sure that internal objs are in a map somewhere
2695 	 * and none have zero ref counts.
2696 	 */
2697 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2698 		if ((object->flags & OBJ_ANON) != 0) {
2699 			if (object->ref_count == 0) {
2700 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2701 					(long)object->size);
2702 			}
2703 			if (!vm_object_in_map(object)) {
2704 				db_printf(
2705 			"vmochk: internal obj is not in a map: "
2706 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2707 				    object->ref_count, (u_long)object->size,
2708 				    (u_long)object->size,
2709 				    (void *)object->backing_object);
2710 			}
2711 		}
2712 		if (db_pager_quit)
2713 			return;
2714 	}
2715 }
2716 
2717 /*
2718  *	vm_object_print:	[ debug ]
2719  */
2720 DB_SHOW_COMMAND(object, vm_object_print_static)
2721 {
2722 	/* XXX convert args. */
2723 	vm_object_t object = (vm_object_t)addr;
2724 	boolean_t full = have_addr;
2725 
2726 	vm_page_t p;
2727 
2728 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2729 #define	count	was_count
2730 
2731 	int count;
2732 
2733 	if (object == NULL)
2734 		return;
2735 
2736 	db_iprintf(
2737 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2738 	    object, (int)object->type, (uintmax_t)object->size,
2739 	    object->resident_page_count, object->ref_count, object->flags,
2740 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2741 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2742 	    object->shadow_count,
2743 	    object->backing_object ? object->backing_object->ref_count : 0,
2744 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2745 
2746 	if (!full)
2747 		return;
2748 
2749 	db_indent += 2;
2750 	count = 0;
2751 	TAILQ_FOREACH(p, &object->memq, listq) {
2752 		if (count == 0)
2753 			db_iprintf("memory:=");
2754 		else if (count == 6) {
2755 			db_printf("\n");
2756 			db_iprintf(" ...");
2757 			count = 0;
2758 		} else
2759 			db_printf(",");
2760 		count++;
2761 
2762 		db_printf("(off=0x%jx,page=0x%jx)",
2763 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2764 
2765 		if (db_pager_quit)
2766 			break;
2767 	}
2768 	if (count != 0)
2769 		db_printf("\n");
2770 	db_indent -= 2;
2771 }
2772 
2773 /* XXX. */
2774 #undef count
2775 
2776 /* XXX need this non-static entry for calling from vm_map_print. */
2777 void
2778 vm_object_print(
2779         /* db_expr_t */ long addr,
2780 	boolean_t have_addr,
2781 	/* db_expr_t */ long count,
2782 	char *modif)
2783 {
2784 	vm_object_print_static(addr, have_addr, count, modif);
2785 }
2786 
2787 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2788 {
2789 	vm_object_t object;
2790 	vm_pindex_t fidx;
2791 	vm_paddr_t pa;
2792 	vm_page_t m, prev_m;
2793 	int rcount, nl, c;
2794 
2795 	nl = 0;
2796 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2797 		db_printf("new object: %p\n", (void *)object);
2798 		if (nl > 18) {
2799 			c = cngetc();
2800 			if (c != ' ')
2801 				return;
2802 			nl = 0;
2803 		}
2804 		nl++;
2805 		rcount = 0;
2806 		fidx = 0;
2807 		pa = -1;
2808 		TAILQ_FOREACH(m, &object->memq, listq) {
2809 			if (m->pindex > 128)
2810 				break;
2811 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2812 			    prev_m->pindex + 1 != m->pindex) {
2813 				if (rcount) {
2814 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2815 						(long)fidx, rcount, (long)pa);
2816 					if (nl > 18) {
2817 						c = cngetc();
2818 						if (c != ' ')
2819 							return;
2820 						nl = 0;
2821 					}
2822 					nl++;
2823 					rcount = 0;
2824 				}
2825 			}
2826 			if (rcount &&
2827 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2828 				++rcount;
2829 				continue;
2830 			}
2831 			if (rcount) {
2832 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2833 					(long)fidx, rcount, (long)pa);
2834 				if (nl > 18) {
2835 					c = cngetc();
2836 					if (c != ' ')
2837 						return;
2838 					nl = 0;
2839 				}
2840 				nl++;
2841 			}
2842 			fidx = m->pindex;
2843 			pa = VM_PAGE_TO_PHYS(m);
2844 			rcount = 1;
2845 		}
2846 		if (rcount) {
2847 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2848 				(long)fidx, rcount, (long)pa);
2849 			if (nl > 18) {
2850 				c = cngetc();
2851 				if (c != ' ')
2852 					return;
2853 				nl = 0;
2854 			}
2855 			nl++;
2856 		}
2857 	}
2858 }
2859 #endif /* DDB */
2860