xref: /freebsd/sys/vm/vm_object.c (revision 8154df37a55948aca8358ade68b659d1aa71b755)
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/limits.h>
77 #include <sys/lock.h>
78 #include <sys/mman.h>
79 #include <sys/mount.h>
80 #include <sys/kernel.h>
81 #include <sys/pctrie.h>
82 #include <sys/sysctl.h>
83 #include <sys/mutex.h>
84 #include <sys/proc.h>		/* for curproc, pageproc */
85 #include <sys/refcount.h>
86 #include <sys/socket.h>
87 #include <sys/resourcevar.h>
88 #include <sys/refcount.h>
89 #include <sys/rwlock.h>
90 #include <sys/user.h>
91 #include <sys/vnode.h>
92 #include <sys/vmmeter.h>
93 #include <sys/sx.h>
94 
95 #include <vm/vm.h>
96 #include <vm/vm_param.h>
97 #include <vm/pmap.h>
98 #include <vm/vm_map.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_pageout.h>
102 #include <vm/vm_pager.h>
103 #include <vm/vm_phys.h>
104 #include <vm/vm_pagequeue.h>
105 #include <vm/swap_pager.h>
106 #include <vm/vm_kern.h>
107 #include <vm/vm_extern.h>
108 #include <vm/vm_radix.h>
109 #include <vm/vm_reserv.h>
110 #include <vm/uma.h>
111 
112 static int old_msync;
113 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
114     "Use old (insecure) msync behavior");
115 
116 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
117 		    int pagerflags, int flags, boolean_t *allclean,
118 		    boolean_t *eio);
119 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
120 		    boolean_t *allclean);
121 static void	vm_object_backing_remove(vm_object_t object);
122 
123 /*
124  *	Virtual memory objects maintain the actual data
125  *	associated with allocated virtual memory.  A given
126  *	page of memory exists within exactly one object.
127  *
128  *	An object is only deallocated when all "references"
129  *	are given up.  Only one "reference" to a given
130  *	region of an object should be writeable.
131  *
132  *	Associated with each object is a list of all resident
133  *	memory pages belonging to that object; this list is
134  *	maintained by the "vm_page" module, and locked by the object's
135  *	lock.
136  *
137  *	Each object also records a "pager" routine which is
138  *	used to retrieve (and store) pages to the proper backing
139  *	storage.  In addition, objects may be backed by other
140  *	objects from which they were virtual-copied.
141  *
142  *	The only items within the object structure which are
143  *	modified after time of creation are:
144  *		reference count		locked by object's lock
145  *		pager routine		locked by object's lock
146  *
147  */
148 
149 struct object_q vm_object_list;
150 struct mtx vm_object_list_mtx;	/* lock for object list and count */
151 
152 struct vm_object kernel_object_store;
153 
154 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
155     "VM object stats");
156 
157 static COUNTER_U64_DEFINE_EARLY(object_collapses);
158 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
159     &object_collapses,
160     "VM object collapses");
161 
162 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
163 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
164     &object_bypasses,
165     "VM object bypasses");
166 
167 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
168 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
169     &object_collapse_waits,
170     "Number of sleeps for collapse");
171 
172 static uma_zone_t obj_zone;
173 
174 static int vm_object_zinit(void *mem, int size, int flags);
175 
176 #ifdef INVARIANTS
177 static void vm_object_zdtor(void *mem, int size, void *arg);
178 
179 static void
180 vm_object_zdtor(void *mem, int size, void *arg)
181 {
182 	vm_object_t object;
183 
184 	object = (vm_object_t)mem;
185 	KASSERT(object->ref_count == 0,
186 	    ("object %p ref_count = %d", object, object->ref_count));
187 	KASSERT(TAILQ_EMPTY(&object->memq),
188 	    ("object %p has resident pages in its memq", object));
189 	KASSERT(vm_radix_is_empty(&object->rtree),
190 	    ("object %p has resident pages in its trie", object));
191 #if VM_NRESERVLEVEL > 0
192 	KASSERT(LIST_EMPTY(&object->rvq),
193 	    ("object %p has reservations",
194 	    object));
195 #endif
196 	KASSERT(!vm_object_busied(object),
197 	    ("object %p busy = %d", object, blockcount_read(&object->busy)));
198 	KASSERT(object->resident_page_count == 0,
199 	    ("object %p resident_page_count = %d",
200 	    object, object->resident_page_count));
201 	KASSERT(atomic_load_int(&object->shadow_count) == 0,
202 	    ("object %p shadow_count = %d",
203 	    object, atomic_load_int(&object->shadow_count)));
204 	KASSERT(object->type == OBJT_DEAD,
205 	    ("object %p has non-dead type %d",
206 	    object, object->type));
207 }
208 #endif
209 
210 static int
211 vm_object_zinit(void *mem, int size, int flags)
212 {
213 	vm_object_t object;
214 
215 	object = (vm_object_t)mem;
216 	rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
217 
218 	/* These are true for any object that has been freed */
219 	object->type = OBJT_DEAD;
220 	vm_radix_init(&object->rtree);
221 	refcount_init(&object->ref_count, 0);
222 	blockcount_init(&object->paging_in_progress);
223 	blockcount_init(&object->busy);
224 	object->resident_page_count = 0;
225 	atomic_store_int(&object->shadow_count, 0);
226 	object->flags = OBJ_DEAD;
227 
228 	mtx_lock(&vm_object_list_mtx);
229 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
230 	mtx_unlock(&vm_object_list_mtx);
231 	return (0);
232 }
233 
234 static void
235 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
236     vm_object_t object, void *handle)
237 {
238 
239 	TAILQ_INIT(&object->memq);
240 	LIST_INIT(&object->shadow_head);
241 
242 	object->type = type;
243 	object->flags = flags;
244 	if ((flags & OBJ_SWAP) != 0)
245 		pctrie_init(&object->un_pager.swp.swp_blks);
246 
247 	/*
248 	 * Ensure that swap_pager_swapoff() iteration over object_list
249 	 * sees up to date type and pctrie head if it observed
250 	 * non-dead object.
251 	 */
252 	atomic_thread_fence_rel();
253 
254 	object->pg_color = 0;
255 	object->size = size;
256 	object->domain.dr_policy = NULL;
257 	object->generation = 1;
258 	object->cleangeneration = 1;
259 	refcount_init(&object->ref_count, 1);
260 	object->memattr = VM_MEMATTR_DEFAULT;
261 	object->cred = NULL;
262 	object->charge = 0;
263 	object->handle = handle;
264 	object->backing_object = NULL;
265 	object->backing_object_offset = (vm_ooffset_t) 0;
266 #if VM_NRESERVLEVEL > 0
267 	LIST_INIT(&object->rvq);
268 #endif
269 	umtx_shm_object_init(object);
270 }
271 
272 /*
273  *	vm_object_init:
274  *
275  *	Initialize the VM objects module.
276  */
277 void
278 vm_object_init(void)
279 {
280 	TAILQ_INIT(&vm_object_list);
281 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
282 
283 	rw_init(&kernel_object->lock, "kernel vm object");
284 	_vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
285 	    VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
286 #if VM_NRESERVLEVEL > 0
287 	kernel_object->flags |= OBJ_COLORED;
288 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
289 #endif
290 	kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
291 
292 	/*
293 	 * The lock portion of struct vm_object must be type stable due
294 	 * to vm_pageout_fallback_object_lock locking a vm object
295 	 * without holding any references to it.
296 	 *
297 	 * paging_in_progress is valid always.  Lockless references to
298 	 * the objects may acquire pip and then check OBJ_DEAD.
299 	 */
300 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
301 #ifdef INVARIANTS
302 	    vm_object_zdtor,
303 #else
304 	    NULL,
305 #endif
306 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
307 
308 	vm_radix_zinit();
309 }
310 
311 void
312 vm_object_clear_flag(vm_object_t object, u_short bits)
313 {
314 
315 	VM_OBJECT_ASSERT_WLOCKED(object);
316 	object->flags &= ~bits;
317 }
318 
319 /*
320  *	Sets the default memory attribute for the specified object.  Pages
321  *	that are allocated to this object are by default assigned this memory
322  *	attribute.
323  *
324  *	Presently, this function must be called before any pages are allocated
325  *	to the object.  In the future, this requirement may be relaxed for
326  *	"default" and "swap" objects.
327  */
328 int
329 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
330 {
331 
332 	VM_OBJECT_ASSERT_WLOCKED(object);
333 
334 	if (object->type == OBJT_DEAD)
335 		return (KERN_INVALID_ARGUMENT);
336 	if (!TAILQ_EMPTY(&object->memq))
337 		return (KERN_FAILURE);
338 
339 	object->memattr = memattr;
340 	return (KERN_SUCCESS);
341 }
342 
343 void
344 vm_object_pip_add(vm_object_t object, short i)
345 {
346 
347 	if (i > 0)
348 		blockcount_acquire(&object->paging_in_progress, i);
349 }
350 
351 void
352 vm_object_pip_wakeup(vm_object_t object)
353 {
354 
355 	vm_object_pip_wakeupn(object, 1);
356 }
357 
358 void
359 vm_object_pip_wakeupn(vm_object_t object, short i)
360 {
361 
362 	if (i > 0)
363 		blockcount_release(&object->paging_in_progress, i);
364 }
365 
366 /*
367  * Atomically drop the object lock and wait for pip to drain.  This protects
368  * from sleep/wakeup races due to identity changes.  The lock is not re-acquired
369  * on return.
370  */
371 static void
372 vm_object_pip_sleep(vm_object_t object, const char *waitid)
373 {
374 
375 	(void)blockcount_sleep(&object->paging_in_progress, &object->lock,
376 	    waitid, PVM | PDROP);
377 }
378 
379 void
380 vm_object_pip_wait(vm_object_t object, const char *waitid)
381 {
382 
383 	VM_OBJECT_ASSERT_WLOCKED(object);
384 
385 	blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
386 	    PVM);
387 }
388 
389 void
390 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
391 {
392 
393 	VM_OBJECT_ASSERT_UNLOCKED(object);
394 
395 	blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
396 }
397 
398 /*
399  *	vm_object_allocate:
400  *
401  *	Returns a new object with the given size.
402  */
403 vm_object_t
404 vm_object_allocate(objtype_t type, vm_pindex_t size)
405 {
406 	vm_object_t object;
407 	u_short flags;
408 
409 	switch (type) {
410 	case OBJT_DEAD:
411 		panic("vm_object_allocate: can't create OBJT_DEAD");
412 	case OBJT_DEFAULT:
413 		flags = OBJ_COLORED;
414 		break;
415 	case OBJT_SWAP:
416 		flags = OBJ_COLORED | OBJ_SWAP;
417 		break;
418 	case OBJT_DEVICE:
419 	case OBJT_SG:
420 		flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
421 		break;
422 	case OBJT_MGTDEVICE:
423 		flags = OBJ_FICTITIOUS;
424 		break;
425 	case OBJT_PHYS:
426 		flags = OBJ_UNMANAGED;
427 		break;
428 	case OBJT_VNODE:
429 		flags = 0;
430 		break;
431 	default:
432 		panic("vm_object_allocate: type %d is undefined or dynamic",
433 		    type);
434 	}
435 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
436 	_vm_object_allocate(type, size, flags, object, NULL);
437 
438 	return (object);
439 }
440 
441 vm_object_t
442 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags)
443 {
444 	vm_object_t object;
445 
446 	MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */);
447 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
448 	_vm_object_allocate(dyntype, size, flags, object, NULL);
449 
450 	return (object);
451 }
452 
453 /*
454  *	vm_object_allocate_anon:
455  *
456  *	Returns a new default object of the given size and marked as
457  *	anonymous memory for special split/collapse handling.  Color
458  *	to be initialized by the caller.
459  */
460 vm_object_t
461 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
462     struct ucred *cred, vm_size_t charge)
463 {
464 	vm_object_t handle, object;
465 
466 	if (backing_object == NULL)
467 		handle = NULL;
468 	else if ((backing_object->flags & OBJ_ANON) != 0)
469 		handle = backing_object->handle;
470 	else
471 		handle = backing_object;
472 	object = uma_zalloc(obj_zone, M_WAITOK);
473 	_vm_object_allocate(OBJT_DEFAULT, size, OBJ_ANON | OBJ_ONEMAPPING,
474 	    object, handle);
475 	object->cred = cred;
476 	object->charge = cred != NULL ? charge : 0;
477 	return (object);
478 }
479 
480 static void
481 vm_object_reference_vnode(vm_object_t object)
482 {
483 	u_int old;
484 
485 	/*
486 	 * vnode objects need the lock for the first reference
487 	 * to serialize with vnode_object_deallocate().
488 	 */
489 	if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
490 		VM_OBJECT_RLOCK(object);
491 		old = refcount_acquire(&object->ref_count);
492 		if (object->type == OBJT_VNODE && old == 0)
493 			vref(object->handle);
494 		VM_OBJECT_RUNLOCK(object);
495 	}
496 }
497 
498 /*
499  *	vm_object_reference:
500  *
501  *	Acquires a reference to the given object.
502  */
503 void
504 vm_object_reference(vm_object_t object)
505 {
506 
507 	if (object == NULL)
508 		return;
509 
510 	if (object->type == OBJT_VNODE)
511 		vm_object_reference_vnode(object);
512 	else
513 		refcount_acquire(&object->ref_count);
514 	KASSERT((object->flags & OBJ_DEAD) == 0,
515 	    ("vm_object_reference: Referenced dead object."));
516 }
517 
518 /*
519  *	vm_object_reference_locked:
520  *
521  *	Gets another reference to the given object.
522  *
523  *	The object must be locked.
524  */
525 void
526 vm_object_reference_locked(vm_object_t object)
527 {
528 	u_int old;
529 
530 	VM_OBJECT_ASSERT_LOCKED(object);
531 	old = refcount_acquire(&object->ref_count);
532 	if (object->type == OBJT_VNODE && old == 0)
533 		vref(object->handle);
534 	KASSERT((object->flags & OBJ_DEAD) == 0,
535 	    ("vm_object_reference: Referenced dead object."));
536 }
537 
538 /*
539  * Handle deallocating an object of type OBJT_VNODE.
540  */
541 static void
542 vm_object_deallocate_vnode(vm_object_t object)
543 {
544 	struct vnode *vp = (struct vnode *) object->handle;
545 	bool last;
546 
547 	KASSERT(object->type == OBJT_VNODE,
548 	    ("vm_object_deallocate_vnode: not a vnode object"));
549 	KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
550 
551 	/* Object lock to protect handle lookup. */
552 	last = refcount_release(&object->ref_count);
553 	VM_OBJECT_RUNLOCK(object);
554 
555 	if (!last)
556 		return;
557 
558 	if (!umtx_shm_vnobj_persistent)
559 		umtx_shm_object_terminated(object);
560 
561 	/* vrele may need the vnode lock. */
562 	vrele(vp);
563 }
564 
565 /*
566  * We dropped a reference on an object and discovered that it had a
567  * single remaining shadow.  This is a sibling of the reference we
568  * dropped.  Attempt to collapse the sibling and backing object.
569  */
570 static vm_object_t
571 vm_object_deallocate_anon(vm_object_t backing_object)
572 {
573 	vm_object_t object;
574 
575 	/* Fetch the final shadow.  */
576 	object = LIST_FIRST(&backing_object->shadow_head);
577 	KASSERT(object != NULL &&
578 	    atomic_load_int(&backing_object->shadow_count) == 1,
579 	    ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
580 	    backing_object->ref_count,
581 	    atomic_load_int(&backing_object->shadow_count)));
582 	KASSERT((object->flags & OBJ_ANON) != 0,
583 	    ("invalid shadow object %p", object));
584 
585 	if (!VM_OBJECT_TRYWLOCK(object)) {
586 		/*
587 		 * Prevent object from disappearing since we do not have a
588 		 * reference.
589 		 */
590 		vm_object_pip_add(object, 1);
591 		VM_OBJECT_WUNLOCK(backing_object);
592 		VM_OBJECT_WLOCK(object);
593 		vm_object_pip_wakeup(object);
594 	} else
595 		VM_OBJECT_WUNLOCK(backing_object);
596 
597 	/*
598 	 * Check for a collapse/terminate race with the last reference holder.
599 	 */
600 	if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
601 	    !refcount_acquire_if_not_zero(&object->ref_count)) {
602 		VM_OBJECT_WUNLOCK(object);
603 		return (NULL);
604 	}
605 	backing_object = object->backing_object;
606 	if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
607 		vm_object_collapse(object);
608 	VM_OBJECT_WUNLOCK(object);
609 
610 	return (object);
611 }
612 
613 /*
614  *	vm_object_deallocate:
615  *
616  *	Release a reference to the specified object,
617  *	gained either through a vm_object_allocate
618  *	or a vm_object_reference call.  When all references
619  *	are gone, storage associated with this object
620  *	may be relinquished.
621  *
622  *	No object may be locked.
623  */
624 void
625 vm_object_deallocate(vm_object_t object)
626 {
627 	vm_object_t temp;
628 	bool released;
629 
630 	while (object != NULL) {
631 		/*
632 		 * If the reference count goes to 0 we start calling
633 		 * vm_object_terminate() on the object chain.  A ref count
634 		 * of 1 may be a special case depending on the shadow count
635 		 * being 0 or 1.  These cases require a write lock on the
636 		 * object.
637 		 */
638 		if ((object->flags & OBJ_ANON) == 0)
639 			released = refcount_release_if_gt(&object->ref_count, 1);
640 		else
641 			released = refcount_release_if_gt(&object->ref_count, 2);
642 		if (released)
643 			return;
644 
645 		if (object->type == OBJT_VNODE) {
646 			VM_OBJECT_RLOCK(object);
647 			if (object->type == OBJT_VNODE) {
648 				vm_object_deallocate_vnode(object);
649 				return;
650 			}
651 			VM_OBJECT_RUNLOCK(object);
652 		}
653 
654 		VM_OBJECT_WLOCK(object);
655 		KASSERT(object->ref_count > 0,
656 		    ("vm_object_deallocate: object deallocated too many times: %d",
657 		    object->type));
658 
659 		/*
660 		 * If this is not the final reference to an anonymous
661 		 * object we may need to collapse the shadow chain.
662 		 */
663 		if (!refcount_release(&object->ref_count)) {
664 			if (object->ref_count > 1 ||
665 			    atomic_load_int(&object->shadow_count) == 0) {
666 				if ((object->flags & OBJ_ANON) != 0 &&
667 				    object->ref_count == 1)
668 					vm_object_set_flag(object,
669 					    OBJ_ONEMAPPING);
670 				VM_OBJECT_WUNLOCK(object);
671 				return;
672 			}
673 
674 			/* Handle collapsing last ref on anonymous objects. */
675 			object = vm_object_deallocate_anon(object);
676 			continue;
677 		}
678 
679 		/*
680 		 * Handle the final reference to an object.  We restart
681 		 * the loop with the backing object to avoid recursion.
682 		 */
683 		umtx_shm_object_terminated(object);
684 		temp = object->backing_object;
685 		if (temp != NULL) {
686 			KASSERT(object->type == OBJT_DEFAULT ||
687 			    object->type == OBJT_SWAP,
688 			    ("shadowed tmpfs v_object 2 %p", object));
689 			vm_object_backing_remove(object);
690 		}
691 
692 		KASSERT((object->flags & OBJ_DEAD) == 0,
693 		    ("vm_object_deallocate: Terminating dead object."));
694 		vm_object_set_flag(object, OBJ_DEAD);
695 		vm_object_terminate(object);
696 		object = temp;
697 	}
698 }
699 
700 /*
701  *	vm_object_destroy removes the object from the global object list
702  *      and frees the space for the object.
703  */
704 void
705 vm_object_destroy(vm_object_t object)
706 {
707 
708 	/*
709 	 * Release the allocation charge.
710 	 */
711 	if (object->cred != NULL) {
712 		swap_release_by_cred(object->charge, object->cred);
713 		object->charge = 0;
714 		crfree(object->cred);
715 		object->cred = NULL;
716 	}
717 
718 	/*
719 	 * Free the space for the object.
720 	 */
721 	uma_zfree(obj_zone, object);
722 }
723 
724 static void
725 vm_object_sub_shadow(vm_object_t object)
726 {
727 	KASSERT(object->shadow_count >= 1,
728 	    ("object %p sub_shadow count zero", object));
729 	atomic_subtract_int(&object->shadow_count, 1);
730 }
731 
732 static void
733 vm_object_backing_remove_locked(vm_object_t object)
734 {
735 	vm_object_t backing_object;
736 
737 	backing_object = object->backing_object;
738 	VM_OBJECT_ASSERT_WLOCKED(object);
739 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
740 
741 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
742 	    ("vm_object_backing_remove: Removing collapsing object."));
743 
744 	vm_object_sub_shadow(backing_object);
745 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
746 		LIST_REMOVE(object, shadow_list);
747 		vm_object_clear_flag(object, OBJ_SHADOWLIST);
748 	}
749 	object->backing_object = NULL;
750 }
751 
752 static void
753 vm_object_backing_remove(vm_object_t object)
754 {
755 	vm_object_t backing_object;
756 
757 	VM_OBJECT_ASSERT_WLOCKED(object);
758 
759 	backing_object = object->backing_object;
760 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
761 		VM_OBJECT_WLOCK(backing_object);
762 		vm_object_backing_remove_locked(object);
763 		VM_OBJECT_WUNLOCK(backing_object);
764 	} else {
765 		object->backing_object = NULL;
766 		vm_object_sub_shadow(backing_object);
767 	}
768 }
769 
770 static void
771 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
772 {
773 
774 	VM_OBJECT_ASSERT_WLOCKED(object);
775 
776 	atomic_add_int(&backing_object->shadow_count, 1);
777 	if ((backing_object->flags & OBJ_ANON) != 0) {
778 		VM_OBJECT_ASSERT_WLOCKED(backing_object);
779 		LIST_INSERT_HEAD(&backing_object->shadow_head, object,
780 		    shadow_list);
781 		vm_object_set_flag(object, OBJ_SHADOWLIST);
782 	}
783 	object->backing_object = backing_object;
784 }
785 
786 static void
787 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
788 {
789 
790 	VM_OBJECT_ASSERT_WLOCKED(object);
791 
792 	if ((backing_object->flags & OBJ_ANON) != 0) {
793 		VM_OBJECT_WLOCK(backing_object);
794 		vm_object_backing_insert_locked(object, backing_object);
795 		VM_OBJECT_WUNLOCK(backing_object);
796 	} else {
797 		object->backing_object = backing_object;
798 		atomic_add_int(&backing_object->shadow_count, 1);
799 	}
800 }
801 
802 /*
803  * Insert an object into a backing_object's shadow list with an additional
804  * reference to the backing_object added.
805  */
806 static void
807 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
808 {
809 
810 	VM_OBJECT_ASSERT_WLOCKED(object);
811 
812 	if ((backing_object->flags & OBJ_ANON) != 0) {
813 		VM_OBJECT_WLOCK(backing_object);
814 		KASSERT((backing_object->flags & OBJ_DEAD) == 0,
815 		    ("shadowing dead anonymous object"));
816 		vm_object_reference_locked(backing_object);
817 		vm_object_backing_insert_locked(object, backing_object);
818 		vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
819 		VM_OBJECT_WUNLOCK(backing_object);
820 	} else {
821 		vm_object_reference(backing_object);
822 		atomic_add_int(&backing_object->shadow_count, 1);
823 		object->backing_object = backing_object;
824 	}
825 }
826 
827 /*
828  * Transfer a backing reference from backing_object to object.
829  */
830 static void
831 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
832 {
833 	vm_object_t new_backing_object;
834 
835 	/*
836 	 * Note that the reference to backing_object->backing_object
837 	 * moves from within backing_object to within object.
838 	 */
839 	vm_object_backing_remove_locked(object);
840 	new_backing_object = backing_object->backing_object;
841 	if (new_backing_object == NULL)
842 		return;
843 	if ((new_backing_object->flags & OBJ_ANON) != 0) {
844 		VM_OBJECT_WLOCK(new_backing_object);
845 		vm_object_backing_remove_locked(backing_object);
846 		vm_object_backing_insert_locked(object, new_backing_object);
847 		VM_OBJECT_WUNLOCK(new_backing_object);
848 	} else {
849 		/*
850 		 * shadow_count for new_backing_object is left
851 		 * unchanged, its reference provided by backing_object
852 		 * is replaced by object.
853 		 */
854 		object->backing_object = new_backing_object;
855 		backing_object->backing_object = NULL;
856 	}
857 }
858 
859 /*
860  * Wait for a concurrent collapse to settle.
861  */
862 static void
863 vm_object_collapse_wait(vm_object_t object)
864 {
865 
866 	VM_OBJECT_ASSERT_WLOCKED(object);
867 
868 	while ((object->flags & OBJ_COLLAPSING) != 0) {
869 		vm_object_pip_wait(object, "vmcolwait");
870 		counter_u64_add(object_collapse_waits, 1);
871 	}
872 }
873 
874 /*
875  * Waits for a backing object to clear a pending collapse and returns
876  * it locked if it is an ANON object.
877  */
878 static vm_object_t
879 vm_object_backing_collapse_wait(vm_object_t object)
880 {
881 	vm_object_t backing_object;
882 
883 	VM_OBJECT_ASSERT_WLOCKED(object);
884 
885 	for (;;) {
886 		backing_object = object->backing_object;
887 		if (backing_object == NULL ||
888 		    (backing_object->flags & OBJ_ANON) == 0)
889 			return (NULL);
890 		VM_OBJECT_WLOCK(backing_object);
891 		if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
892 			break;
893 		VM_OBJECT_WUNLOCK(object);
894 		vm_object_pip_sleep(backing_object, "vmbckwait");
895 		counter_u64_add(object_collapse_waits, 1);
896 		VM_OBJECT_WLOCK(object);
897 	}
898 	return (backing_object);
899 }
900 
901 /*
902  *	vm_object_terminate_pages removes any remaining pageable pages
903  *	from the object and resets the object to an empty state.
904  */
905 static void
906 vm_object_terminate_pages(vm_object_t object)
907 {
908 	vm_page_t p, p_next;
909 
910 	VM_OBJECT_ASSERT_WLOCKED(object);
911 
912 	/*
913 	 * Free any remaining pageable pages.  This also removes them from the
914 	 * paging queues.  However, don't free wired pages, just remove them
915 	 * from the object.  Rather than incrementally removing each page from
916 	 * the object, the page and object are reset to any empty state.
917 	 */
918 	TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
919 		vm_page_assert_unbusied(p);
920 		KASSERT(p->object == object &&
921 		    (p->ref_count & VPRC_OBJREF) != 0,
922 		    ("vm_object_terminate_pages: page %p is inconsistent", p));
923 
924 		p->object = NULL;
925 		if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
926 			VM_CNT_INC(v_pfree);
927 			vm_page_free(p);
928 		}
929 	}
930 
931 	/*
932 	 * If the object contained any pages, then reset it to an empty state.
933 	 * None of the object's fields, including "resident_page_count", were
934 	 * modified by the preceding loop.
935 	 */
936 	if (object->resident_page_count != 0) {
937 		vm_radix_reclaim_allnodes(&object->rtree);
938 		TAILQ_INIT(&object->memq);
939 		object->resident_page_count = 0;
940 		if (object->type == OBJT_VNODE)
941 			vdrop(object->handle);
942 	}
943 }
944 
945 /*
946  *	vm_object_terminate actually destroys the specified object, freeing
947  *	up all previously used resources.
948  *
949  *	The object must be locked.
950  *	This routine may block.
951  */
952 void
953 vm_object_terminate(vm_object_t object)
954 {
955 
956 	VM_OBJECT_ASSERT_WLOCKED(object);
957 	KASSERT((object->flags & OBJ_DEAD) != 0,
958 	    ("terminating non-dead obj %p", object));
959 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
960 	    ("terminating collapsing obj %p", object));
961 	KASSERT(object->backing_object == NULL,
962 	    ("terminating shadow obj %p", object));
963 
964 	/*
965 	 * Wait for the pageout daemon and other current users to be
966 	 * done with the object.  Note that new paging_in_progress
967 	 * users can come after this wait, but they must check
968 	 * OBJ_DEAD flag set (without unlocking the object), and avoid
969 	 * the object being terminated.
970 	 */
971 	vm_object_pip_wait(object, "objtrm");
972 
973 	KASSERT(object->ref_count == 0,
974 	    ("vm_object_terminate: object with references, ref_count=%d",
975 	    object->ref_count));
976 
977 	if ((object->flags & OBJ_PG_DTOR) == 0)
978 		vm_object_terminate_pages(object);
979 
980 #if VM_NRESERVLEVEL > 0
981 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
982 		vm_reserv_break_all(object);
983 #endif
984 
985 	KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
986 	    (object->flags & OBJ_SWAP) != 0,
987 	    ("%s: non-swap obj %p has cred", __func__, object));
988 
989 	/*
990 	 * Let the pager know object is dead.
991 	 */
992 	vm_pager_deallocate(object);
993 	VM_OBJECT_WUNLOCK(object);
994 
995 	vm_object_destroy(object);
996 }
997 
998 /*
999  * Make the page read-only so that we can clear the object flags.  However, if
1000  * this is a nosync mmap then the object is likely to stay dirty so do not
1001  * mess with the page and do not clear the object flags.  Returns TRUE if the
1002  * page should be flushed, and FALSE otherwise.
1003  */
1004 static boolean_t
1005 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
1006 {
1007 
1008 	vm_page_assert_busied(p);
1009 
1010 	/*
1011 	 * If we have been asked to skip nosync pages and this is a
1012 	 * nosync page, skip it.  Note that the object flags were not
1013 	 * cleared in this case so we do not have to set them.
1014 	 */
1015 	if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
1016 		*allclean = FALSE;
1017 		return (FALSE);
1018 	} else {
1019 		pmap_remove_write(p);
1020 		return (p->dirty != 0);
1021 	}
1022 }
1023 
1024 /*
1025  *	vm_object_page_clean
1026  *
1027  *	Clean all dirty pages in the specified range of object.  Leaves page
1028  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
1029  *	write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1030  *	leaving the object dirty.
1031  *
1032  *	For swap objects backing tmpfs regular files, do not flush anything,
1033  *	but remove write protection on the mapped pages to update mtime through
1034  *	mmaped writes.
1035  *
1036  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
1037  *	synchronous clustering mode implementation.
1038  *
1039  *	Odd semantics: if start == end, we clean everything.
1040  *
1041  *	The object must be locked.
1042  *
1043  *	Returns FALSE if some page from the range was not written, as
1044  *	reported by the pager, and TRUE otherwise.
1045  */
1046 boolean_t
1047 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1048     int flags)
1049 {
1050 	vm_page_t np, p;
1051 	vm_pindex_t pi, tend, tstart;
1052 	int curgeneration, n, pagerflags;
1053 	boolean_t eio, res, allclean;
1054 
1055 	VM_OBJECT_ASSERT_WLOCKED(object);
1056 
1057 	if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1058 		return (TRUE);
1059 
1060 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1061 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1062 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1063 
1064 	tstart = OFF_TO_IDX(start);
1065 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1066 	allclean = tstart == 0 && tend >= object->size;
1067 	res = TRUE;
1068 
1069 rescan:
1070 	curgeneration = object->generation;
1071 
1072 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1073 		pi = p->pindex;
1074 		if (pi >= tend)
1075 			break;
1076 		np = TAILQ_NEXT(p, listq);
1077 		if (vm_page_none_valid(p))
1078 			continue;
1079 		if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1080 			if (object->generation != curgeneration &&
1081 			    (flags & OBJPC_SYNC) != 0)
1082 				goto rescan;
1083 			np = vm_page_find_least(object, pi);
1084 			continue;
1085 		}
1086 		if (!vm_object_page_remove_write(p, flags, &allclean)) {
1087 			vm_page_xunbusy(p);
1088 			continue;
1089 		}
1090 		if (object->type == OBJT_VNODE) {
1091 			n = vm_object_page_collect_flush(object, p, pagerflags,
1092 			    flags, &allclean, &eio);
1093 			if (eio) {
1094 				res = FALSE;
1095 				allclean = FALSE;
1096 			}
1097 			if (object->generation != curgeneration &&
1098 			    (flags & OBJPC_SYNC) != 0)
1099 				goto rescan;
1100 
1101 			/*
1102 			 * If the VOP_PUTPAGES() did a truncated write, so
1103 			 * that even the first page of the run is not fully
1104 			 * written, vm_pageout_flush() returns 0 as the run
1105 			 * length.  Since the condition that caused truncated
1106 			 * write may be permanent, e.g. exhausted free space,
1107 			 * accepting n == 0 would cause an infinite loop.
1108 			 *
1109 			 * Forwarding the iterator leaves the unwritten page
1110 			 * behind, but there is not much we can do there if
1111 			 * filesystem refuses to write it.
1112 			 */
1113 			if (n == 0) {
1114 				n = 1;
1115 				allclean = FALSE;
1116 			}
1117 		} else {
1118 			n = 1;
1119 			vm_page_xunbusy(p);
1120 		}
1121 		np = vm_page_find_least(object, pi + n);
1122 	}
1123 #if 0
1124 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1125 #endif
1126 
1127 	/*
1128 	 * Leave updating cleangeneration for tmpfs objects to tmpfs
1129 	 * scan.  It needs to update mtime, which happens for other
1130 	 * filesystems during page writeouts.
1131 	 */
1132 	if (allclean && object->type == OBJT_VNODE)
1133 		object->cleangeneration = curgeneration;
1134 	return (res);
1135 }
1136 
1137 static int
1138 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1139     int flags, boolean_t *allclean, boolean_t *eio)
1140 {
1141 	vm_page_t ma[vm_pageout_page_count], p_first, tp;
1142 	int count, i, mreq, runlen;
1143 
1144 	vm_page_lock_assert(p, MA_NOTOWNED);
1145 	vm_page_assert_xbusied(p);
1146 	VM_OBJECT_ASSERT_WLOCKED(object);
1147 
1148 	count = 1;
1149 	mreq = 0;
1150 
1151 	for (tp = p; count < vm_pageout_page_count; count++) {
1152 		tp = vm_page_next(tp);
1153 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1154 			break;
1155 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1156 			vm_page_xunbusy(tp);
1157 			break;
1158 		}
1159 	}
1160 
1161 	for (p_first = p; count < vm_pageout_page_count; count++) {
1162 		tp = vm_page_prev(p_first);
1163 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1164 			break;
1165 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1166 			vm_page_xunbusy(tp);
1167 			break;
1168 		}
1169 		p_first = tp;
1170 		mreq++;
1171 	}
1172 
1173 	for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1174 		ma[i] = tp;
1175 
1176 	vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1177 	return (runlen);
1178 }
1179 
1180 /*
1181  * Note that there is absolutely no sense in writing out
1182  * anonymous objects, so we track down the vnode object
1183  * to write out.
1184  * We invalidate (remove) all pages from the address space
1185  * for semantic correctness.
1186  *
1187  * If the backing object is a device object with unmanaged pages, then any
1188  * mappings to the specified range of pages must be removed before this
1189  * function is called.
1190  *
1191  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1192  * may start out with a NULL object.
1193  */
1194 boolean_t
1195 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1196     boolean_t syncio, boolean_t invalidate)
1197 {
1198 	vm_object_t backing_object;
1199 	struct vnode *vp;
1200 	struct mount *mp;
1201 	int error, flags, fsync_after;
1202 	boolean_t res;
1203 
1204 	if (object == NULL)
1205 		return (TRUE);
1206 	res = TRUE;
1207 	error = 0;
1208 	VM_OBJECT_WLOCK(object);
1209 	while ((backing_object = object->backing_object) != NULL) {
1210 		VM_OBJECT_WLOCK(backing_object);
1211 		offset += object->backing_object_offset;
1212 		VM_OBJECT_WUNLOCK(object);
1213 		object = backing_object;
1214 		if (object->size < OFF_TO_IDX(offset + size))
1215 			size = IDX_TO_OFF(object->size) - offset;
1216 	}
1217 	/*
1218 	 * Flush pages if writing is allowed, invalidate them
1219 	 * if invalidation requested.  Pages undergoing I/O
1220 	 * will be ignored by vm_object_page_remove().
1221 	 *
1222 	 * We cannot lock the vnode and then wait for paging
1223 	 * to complete without deadlocking against vm_fault.
1224 	 * Instead we simply call vm_object_page_remove() and
1225 	 * allow it to block internally on a page-by-page
1226 	 * basis when it encounters pages undergoing async
1227 	 * I/O.
1228 	 */
1229 	if (object->type == OBJT_VNODE &&
1230 	    vm_object_mightbedirty(object) != 0 &&
1231 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1232 		VM_OBJECT_WUNLOCK(object);
1233 		(void) vn_start_write(vp, &mp, V_WAIT);
1234 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1235 		if (syncio && !invalidate && offset == 0 &&
1236 		    atop(size) == object->size) {
1237 			/*
1238 			 * If syncing the whole mapping of the file,
1239 			 * it is faster to schedule all the writes in
1240 			 * async mode, also allowing the clustering,
1241 			 * and then wait for i/o to complete.
1242 			 */
1243 			flags = 0;
1244 			fsync_after = TRUE;
1245 		} else {
1246 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1247 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1248 			fsync_after = FALSE;
1249 		}
1250 		VM_OBJECT_WLOCK(object);
1251 		res = vm_object_page_clean(object, offset, offset + size,
1252 		    flags);
1253 		VM_OBJECT_WUNLOCK(object);
1254 		if (fsync_after)
1255 			error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1256 		VOP_UNLOCK(vp);
1257 		vn_finished_write(mp);
1258 		if (error != 0)
1259 			res = FALSE;
1260 		VM_OBJECT_WLOCK(object);
1261 	}
1262 	if ((object->type == OBJT_VNODE ||
1263 	     object->type == OBJT_DEVICE) && invalidate) {
1264 		if (object->type == OBJT_DEVICE)
1265 			/*
1266 			 * The option OBJPR_NOTMAPPED must be passed here
1267 			 * because vm_object_page_remove() cannot remove
1268 			 * unmanaged mappings.
1269 			 */
1270 			flags = OBJPR_NOTMAPPED;
1271 		else if (old_msync)
1272 			flags = 0;
1273 		else
1274 			flags = OBJPR_CLEANONLY;
1275 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1276 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1277 	}
1278 	VM_OBJECT_WUNLOCK(object);
1279 	return (res);
1280 }
1281 
1282 /*
1283  * Determine whether the given advice can be applied to the object.  Advice is
1284  * not applied to unmanaged pages since they never belong to page queues, and
1285  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1286  * have been mapped at most once.
1287  */
1288 static bool
1289 vm_object_advice_applies(vm_object_t object, int advice)
1290 {
1291 
1292 	if ((object->flags & OBJ_UNMANAGED) != 0)
1293 		return (false);
1294 	if (advice != MADV_FREE)
1295 		return (true);
1296 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1297 	    (OBJ_ONEMAPPING | OBJ_ANON));
1298 }
1299 
1300 static void
1301 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1302     vm_size_t size)
1303 {
1304 
1305 	if (advice == MADV_FREE)
1306 		vm_pager_freespace(object, pindex, size);
1307 }
1308 
1309 /*
1310  *	vm_object_madvise:
1311  *
1312  *	Implements the madvise function at the object/page level.
1313  *
1314  *	MADV_WILLNEED	(any object)
1315  *
1316  *	    Activate the specified pages if they are resident.
1317  *
1318  *	MADV_DONTNEED	(any object)
1319  *
1320  *	    Deactivate the specified pages if they are resident.
1321  *
1322  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1323  *			 OBJ_ONEMAPPING only)
1324  *
1325  *	    Deactivate and clean the specified pages if they are
1326  *	    resident.  This permits the process to reuse the pages
1327  *	    without faulting or the kernel to reclaim the pages
1328  *	    without I/O.
1329  */
1330 void
1331 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1332     int advice)
1333 {
1334 	vm_pindex_t tpindex;
1335 	vm_object_t backing_object, tobject;
1336 	vm_page_t m, tm;
1337 
1338 	if (object == NULL)
1339 		return;
1340 
1341 relookup:
1342 	VM_OBJECT_WLOCK(object);
1343 	if (!vm_object_advice_applies(object, advice)) {
1344 		VM_OBJECT_WUNLOCK(object);
1345 		return;
1346 	}
1347 	for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1348 		tobject = object;
1349 
1350 		/*
1351 		 * If the next page isn't resident in the top-level object, we
1352 		 * need to search the shadow chain.  When applying MADV_FREE, we
1353 		 * take care to release any swap space used to store
1354 		 * non-resident pages.
1355 		 */
1356 		if (m == NULL || pindex < m->pindex) {
1357 			/*
1358 			 * Optimize a common case: if the top-level object has
1359 			 * no backing object, we can skip over the non-resident
1360 			 * range in constant time.
1361 			 */
1362 			if (object->backing_object == NULL) {
1363 				tpindex = (m != NULL && m->pindex < end) ?
1364 				    m->pindex : end;
1365 				vm_object_madvise_freespace(object, advice,
1366 				    pindex, tpindex - pindex);
1367 				if ((pindex = tpindex) == end)
1368 					break;
1369 				goto next_page;
1370 			}
1371 
1372 			tpindex = pindex;
1373 			do {
1374 				vm_object_madvise_freespace(tobject, advice,
1375 				    tpindex, 1);
1376 				/*
1377 				 * Prepare to search the next object in the
1378 				 * chain.
1379 				 */
1380 				backing_object = tobject->backing_object;
1381 				if (backing_object == NULL)
1382 					goto next_pindex;
1383 				VM_OBJECT_WLOCK(backing_object);
1384 				tpindex +=
1385 				    OFF_TO_IDX(tobject->backing_object_offset);
1386 				if (tobject != object)
1387 					VM_OBJECT_WUNLOCK(tobject);
1388 				tobject = backing_object;
1389 				if (!vm_object_advice_applies(tobject, advice))
1390 					goto next_pindex;
1391 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1392 			    NULL);
1393 		} else {
1394 next_page:
1395 			tm = m;
1396 			m = TAILQ_NEXT(m, listq);
1397 		}
1398 
1399 		/*
1400 		 * If the page is not in a normal state, skip it.  The page
1401 		 * can not be invalidated while the object lock is held.
1402 		 */
1403 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1404 			goto next_pindex;
1405 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1406 		    ("vm_object_madvise: page %p is fictitious", tm));
1407 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1408 		    ("vm_object_madvise: page %p is not managed", tm));
1409 		if (vm_page_tryxbusy(tm) == 0) {
1410 			if (object != tobject)
1411 				VM_OBJECT_WUNLOCK(object);
1412 			if (advice == MADV_WILLNEED) {
1413 				/*
1414 				 * Reference the page before unlocking and
1415 				 * sleeping so that the page daemon is less
1416 				 * likely to reclaim it.
1417 				 */
1418 				vm_page_aflag_set(tm, PGA_REFERENCED);
1419 			}
1420 			if (!vm_page_busy_sleep(tm, "madvpo", 0))
1421 				VM_OBJECT_WUNLOCK(tobject);
1422   			goto relookup;
1423 		}
1424 		vm_page_advise(tm, advice);
1425 		vm_page_xunbusy(tm);
1426 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1427 next_pindex:
1428 		if (tobject != object)
1429 			VM_OBJECT_WUNLOCK(tobject);
1430 	}
1431 	VM_OBJECT_WUNLOCK(object);
1432 }
1433 
1434 /*
1435  *	vm_object_shadow:
1436  *
1437  *	Create a new object which is backed by the
1438  *	specified existing object range.  The source
1439  *	object reference is deallocated.
1440  *
1441  *	The new object and offset into that object
1442  *	are returned in the source parameters.
1443  */
1444 void
1445 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1446     struct ucred *cred, bool shared)
1447 {
1448 	vm_object_t source;
1449 	vm_object_t result;
1450 
1451 	source = *object;
1452 
1453 	/*
1454 	 * Don't create the new object if the old object isn't shared.
1455 	 *
1456 	 * If we hold the only reference we can guarantee that it won't
1457 	 * increase while we have the map locked.  Otherwise the race is
1458 	 * harmless and we will end up with an extra shadow object that
1459 	 * will be collapsed later.
1460 	 */
1461 	if (source != NULL && source->ref_count == 1 &&
1462 	    (source->flags & OBJ_ANON) != 0)
1463 		return;
1464 
1465 	/*
1466 	 * Allocate a new object with the given length.
1467 	 */
1468 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1469 
1470 	/*
1471 	 * Store the offset into the source object, and fix up the offset into
1472 	 * the new object.
1473 	 */
1474 	result->backing_object_offset = *offset;
1475 
1476 	if (shared || source != NULL) {
1477 		VM_OBJECT_WLOCK(result);
1478 
1479 		/*
1480 		 * The new object shadows the source object, adding a
1481 		 * reference to it.  Our caller changes his reference
1482 		 * to point to the new object, removing a reference to
1483 		 * the source object.  Net result: no change of
1484 		 * reference count, unless the caller needs to add one
1485 		 * more reference due to forking a shared map entry.
1486 		 */
1487 		if (shared) {
1488 			vm_object_reference_locked(result);
1489 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1490 		}
1491 
1492 		/*
1493 		 * Try to optimize the result object's page color when
1494 		 * shadowing in order to maintain page coloring
1495 		 * consistency in the combined shadowed object.
1496 		 */
1497 		if (source != NULL) {
1498 			vm_object_backing_insert(result, source);
1499 			result->domain = source->domain;
1500 #if VM_NRESERVLEVEL > 0
1501 			vm_object_set_flag(result,
1502 			    (source->flags & OBJ_COLORED));
1503 			result->pg_color = (source->pg_color +
1504 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1505 			    1)) - 1);
1506 #endif
1507 		}
1508 		VM_OBJECT_WUNLOCK(result);
1509 	}
1510 
1511 	/*
1512 	 * Return the new things
1513 	 */
1514 	*offset = 0;
1515 	*object = result;
1516 }
1517 
1518 /*
1519  *	vm_object_split:
1520  *
1521  * Split the pages in a map entry into a new object.  This affords
1522  * easier removal of unused pages, and keeps object inheritance from
1523  * being a negative impact on memory usage.
1524  */
1525 void
1526 vm_object_split(vm_map_entry_t entry)
1527 {
1528 	vm_page_t m, m_busy, m_next;
1529 	vm_object_t orig_object, new_object, backing_object;
1530 	vm_pindex_t idx, offidxstart;
1531 	vm_size_t size;
1532 
1533 	orig_object = entry->object.vm_object;
1534 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1535 	    ("vm_object_split:  Splitting object with multiple mappings."));
1536 	if ((orig_object->flags & OBJ_ANON) == 0)
1537 		return;
1538 	if (orig_object->ref_count <= 1)
1539 		return;
1540 	VM_OBJECT_WUNLOCK(orig_object);
1541 
1542 	offidxstart = OFF_TO_IDX(entry->offset);
1543 	size = atop(entry->end - entry->start);
1544 
1545 	/*
1546 	 * If swap_pager_copy() is later called, it will convert new_object
1547 	 * into a swap object.
1548 	 */
1549 	new_object = vm_object_allocate_anon(size, orig_object,
1550 	    orig_object->cred, ptoa(size));
1551 
1552 	/*
1553 	 * We must wait for the orig_object to complete any in-progress
1554 	 * collapse so that the swap blocks are stable below.  The
1555 	 * additional reference on backing_object by new object will
1556 	 * prevent further collapse operations until split completes.
1557 	 */
1558 	VM_OBJECT_WLOCK(orig_object);
1559 	vm_object_collapse_wait(orig_object);
1560 
1561 	/*
1562 	 * At this point, the new object is still private, so the order in
1563 	 * which the original and new objects are locked does not matter.
1564 	 */
1565 	VM_OBJECT_WLOCK(new_object);
1566 	new_object->domain = orig_object->domain;
1567 	backing_object = orig_object->backing_object;
1568 	if (backing_object != NULL) {
1569 		vm_object_backing_insert_ref(new_object, backing_object);
1570 		new_object->backing_object_offset =
1571 		    orig_object->backing_object_offset + entry->offset;
1572 	}
1573 	if (orig_object->cred != NULL) {
1574 		crhold(orig_object->cred);
1575 		KASSERT(orig_object->charge >= ptoa(size),
1576 		    ("orig_object->charge < 0"));
1577 		orig_object->charge -= ptoa(size);
1578 	}
1579 
1580 	/*
1581 	 * Mark the split operation so that swap_pager_getpages() knows
1582 	 * that the object is in transition.
1583 	 */
1584 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1585 	m_busy = NULL;
1586 #ifdef INVARIANTS
1587 	idx = 0;
1588 #endif
1589 retry:
1590 	m = vm_page_find_least(orig_object, offidxstart);
1591 	KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1592 	    ("%s: object %p was repopulated", __func__, orig_object));
1593 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1594 	    m = m_next) {
1595 		m_next = TAILQ_NEXT(m, listq);
1596 
1597 		/*
1598 		 * We must wait for pending I/O to complete before we can
1599 		 * rename the page.
1600 		 *
1601 		 * We do not have to VM_PROT_NONE the page as mappings should
1602 		 * not be changed by this operation.
1603 		 */
1604 		if (vm_page_tryxbusy(m) == 0) {
1605 			VM_OBJECT_WUNLOCK(new_object);
1606 			if (vm_page_busy_sleep(m, "spltwt", 0))
1607 				VM_OBJECT_WLOCK(orig_object);
1608 			VM_OBJECT_WLOCK(new_object);
1609 			goto retry;
1610 		}
1611 
1612 		/*
1613 		 * The page was left invalid.  Likely placed there by
1614 		 * an incomplete fault.  Just remove and ignore.
1615 		 */
1616 		if (vm_page_none_valid(m)) {
1617 			if (vm_page_remove(m))
1618 				vm_page_free(m);
1619 			continue;
1620 		}
1621 
1622 		/* vm_page_rename() will dirty the page. */
1623 		if (vm_page_rename(m, new_object, idx)) {
1624 			vm_page_xunbusy(m);
1625 			VM_OBJECT_WUNLOCK(new_object);
1626 			VM_OBJECT_WUNLOCK(orig_object);
1627 			vm_radix_wait();
1628 			VM_OBJECT_WLOCK(orig_object);
1629 			VM_OBJECT_WLOCK(new_object);
1630 			goto retry;
1631 		}
1632 
1633 #if VM_NRESERVLEVEL > 0
1634 		/*
1635 		 * If some of the reservation's allocated pages remain with
1636 		 * the original object, then transferring the reservation to
1637 		 * the new object is neither particularly beneficial nor
1638 		 * particularly harmful as compared to leaving the reservation
1639 		 * with the original object.  If, however, all of the
1640 		 * reservation's allocated pages are transferred to the new
1641 		 * object, then transferring the reservation is typically
1642 		 * beneficial.  Determining which of these two cases applies
1643 		 * would be more costly than unconditionally renaming the
1644 		 * reservation.
1645 		 */
1646 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1647 #endif
1648 
1649 		/*
1650 		 * orig_object's type may change while sleeping, so keep track
1651 		 * of the beginning of the busied range.
1652 		 */
1653 		if (orig_object->type != OBJT_SWAP)
1654 			vm_page_xunbusy(m);
1655 		else if (m_busy == NULL)
1656 			m_busy = m;
1657 	}
1658 	if ((orig_object->flags & OBJ_SWAP) != 0) {
1659 		/*
1660 		 * swap_pager_copy() can sleep, in which case the orig_object's
1661 		 * and new_object's locks are released and reacquired.
1662 		 */
1663 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1664 		if (m_busy != NULL)
1665 			TAILQ_FOREACH_FROM(m_busy, &new_object->memq, listq)
1666 				vm_page_xunbusy(m_busy);
1667 	}
1668 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1669 	VM_OBJECT_WUNLOCK(orig_object);
1670 	VM_OBJECT_WUNLOCK(new_object);
1671 	entry->object.vm_object = new_object;
1672 	entry->offset = 0LL;
1673 	vm_object_deallocate(orig_object);
1674 	VM_OBJECT_WLOCK(new_object);
1675 }
1676 
1677 static vm_page_t
1678 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1679 {
1680 	vm_object_t backing_object;
1681 
1682 	VM_OBJECT_ASSERT_WLOCKED(object);
1683 	backing_object = object->backing_object;
1684 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1685 
1686 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1687 	    ("invalid ownership %p %p %p", p, object, backing_object));
1688 	/* The page is only NULL when rename fails. */
1689 	if (p == NULL) {
1690 		VM_OBJECT_WUNLOCK(object);
1691 		VM_OBJECT_WUNLOCK(backing_object);
1692 		vm_radix_wait();
1693 		VM_OBJECT_WLOCK(object);
1694 	} else if (p->object == object) {
1695 		VM_OBJECT_WUNLOCK(backing_object);
1696 		if (vm_page_busy_sleep(p, "vmocol", 0))
1697 			VM_OBJECT_WLOCK(object);
1698 	} else {
1699 		VM_OBJECT_WUNLOCK(object);
1700 		if (!vm_page_busy_sleep(p, "vmocol", 0))
1701 			VM_OBJECT_WUNLOCK(backing_object);
1702 		VM_OBJECT_WLOCK(object);
1703 	}
1704 	VM_OBJECT_WLOCK(backing_object);
1705 	return (TAILQ_FIRST(&backing_object->memq));
1706 }
1707 
1708 static bool
1709 vm_object_scan_all_shadowed(vm_object_t object)
1710 {
1711 	vm_object_t backing_object;
1712 	vm_page_t p, pp;
1713 	vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1714 
1715 	VM_OBJECT_ASSERT_WLOCKED(object);
1716 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1717 
1718 	backing_object = object->backing_object;
1719 
1720 	if ((backing_object->flags & OBJ_ANON) == 0)
1721 		return (false);
1722 
1723 	pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1724 	p = vm_page_find_least(backing_object, pi);
1725 	ps = swap_pager_find_least(backing_object, pi);
1726 
1727 	/*
1728 	 * Only check pages inside the parent object's range and
1729 	 * inside the parent object's mapping of the backing object.
1730 	 */
1731 	for (;; pi++) {
1732 		if (p != NULL && p->pindex < pi)
1733 			p = TAILQ_NEXT(p, listq);
1734 		if (ps < pi)
1735 			ps = swap_pager_find_least(backing_object, pi);
1736 		if (p == NULL && ps >= backing_object->size)
1737 			break;
1738 		else if (p == NULL)
1739 			pi = ps;
1740 		else
1741 			pi = MIN(p->pindex, ps);
1742 
1743 		new_pindex = pi - backing_offset_index;
1744 		if (new_pindex >= object->size)
1745 			break;
1746 
1747 		if (p != NULL) {
1748 			/*
1749 			 * If the backing object page is busy a
1750 			 * grandparent or older page may still be
1751 			 * undergoing CoW.  It is not safe to collapse
1752 			 * the backing object until it is quiesced.
1753 			 */
1754 			if (vm_page_tryxbusy(p) == 0)
1755 				return (false);
1756 
1757 			/*
1758 			 * We raced with the fault handler that left
1759 			 * newly allocated invalid page on the object
1760 			 * queue and retried.
1761 			 */
1762 			if (!vm_page_all_valid(p))
1763 				goto unbusy_ret;
1764 		}
1765 
1766 		/*
1767 		 * See if the parent has the page or if the parent's object
1768 		 * pager has the page.  If the parent has the page but the page
1769 		 * is not valid, the parent's object pager must have the page.
1770 		 *
1771 		 * If this fails, the parent does not completely shadow the
1772 		 * object and we might as well give up now.
1773 		 */
1774 		pp = vm_page_lookup(object, new_pindex);
1775 
1776 		/*
1777 		 * The valid check here is stable due to object lock
1778 		 * being required to clear valid and initiate paging.
1779 		 * Busy of p disallows fault handler to validate pp.
1780 		 */
1781 		if ((pp == NULL || vm_page_none_valid(pp)) &&
1782 		    !vm_pager_has_page(object, new_pindex, NULL, NULL))
1783 			goto unbusy_ret;
1784 		if (p != NULL)
1785 			vm_page_xunbusy(p);
1786 	}
1787 	return (true);
1788 
1789 unbusy_ret:
1790 	if (p != NULL)
1791 		vm_page_xunbusy(p);
1792 	return (false);
1793 }
1794 
1795 static void
1796 vm_object_collapse_scan(vm_object_t object)
1797 {
1798 	vm_object_t backing_object;
1799 	vm_page_t next, p, pp;
1800 	vm_pindex_t backing_offset_index, new_pindex;
1801 
1802 	VM_OBJECT_ASSERT_WLOCKED(object);
1803 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1804 
1805 	backing_object = object->backing_object;
1806 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1807 
1808 	/*
1809 	 * Our scan
1810 	 */
1811 	for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1812 		next = TAILQ_NEXT(p, listq);
1813 		new_pindex = p->pindex - backing_offset_index;
1814 
1815 		/*
1816 		 * Check for busy page
1817 		 */
1818 		if (vm_page_tryxbusy(p) == 0) {
1819 			next = vm_object_collapse_scan_wait(object, p);
1820 			continue;
1821 		}
1822 
1823 		KASSERT(object->backing_object == backing_object,
1824 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1825 		    object->backing_object, backing_object));
1826 		KASSERT(p->object == backing_object,
1827 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1828 		    p->object, backing_object));
1829 
1830 		if (p->pindex < backing_offset_index ||
1831 		    new_pindex >= object->size) {
1832 			vm_pager_freespace(backing_object, p->pindex, 1);
1833 
1834 			KASSERT(!pmap_page_is_mapped(p),
1835 			    ("freeing mapped page %p", p));
1836 			if (vm_page_remove(p))
1837 				vm_page_free(p);
1838 			continue;
1839 		}
1840 
1841 		if (!vm_page_all_valid(p)) {
1842 			KASSERT(!pmap_page_is_mapped(p),
1843 			    ("freeing mapped page %p", p));
1844 			if (vm_page_remove(p))
1845 				vm_page_free(p);
1846 			continue;
1847 		}
1848 
1849 		pp = vm_page_lookup(object, new_pindex);
1850 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1851 			vm_page_xunbusy(p);
1852 			/*
1853 			 * The page in the parent is busy and possibly not
1854 			 * (yet) valid.  Until its state is finalized by the
1855 			 * busy bit owner, we can't tell whether it shadows the
1856 			 * original page.
1857 			 */
1858 			next = vm_object_collapse_scan_wait(object, pp);
1859 			continue;
1860 		}
1861 
1862 		if (pp != NULL && vm_page_none_valid(pp)) {
1863 			/*
1864 			 * The page was invalid in the parent.  Likely placed
1865 			 * there by an incomplete fault.  Just remove and
1866 			 * ignore.  p can replace it.
1867 			 */
1868 			if (vm_page_remove(pp))
1869 				vm_page_free(pp);
1870 			pp = NULL;
1871 		}
1872 
1873 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1874 			NULL)) {
1875 			/*
1876 			 * The page already exists in the parent OR swap exists
1877 			 * for this location in the parent.  Leave the parent's
1878 			 * page alone.  Destroy the original page from the
1879 			 * backing object.
1880 			 */
1881 			vm_pager_freespace(backing_object, p->pindex, 1);
1882 			KASSERT(!pmap_page_is_mapped(p),
1883 			    ("freeing mapped page %p", p));
1884 			if (vm_page_remove(p))
1885 				vm_page_free(p);
1886 			if (pp != NULL)
1887 				vm_page_xunbusy(pp);
1888 			continue;
1889 		}
1890 
1891 		/*
1892 		 * Page does not exist in parent, rename the page from the
1893 		 * backing object to the main object.
1894 		 *
1895 		 * If the page was mapped to a process, it can remain mapped
1896 		 * through the rename.  vm_page_rename() will dirty the page.
1897 		 */
1898 		if (vm_page_rename(p, object, new_pindex)) {
1899 			vm_page_xunbusy(p);
1900 			next = vm_object_collapse_scan_wait(object, NULL);
1901 			continue;
1902 		}
1903 
1904 		/* Use the old pindex to free the right page. */
1905 		vm_pager_freespace(backing_object, new_pindex +
1906 		    backing_offset_index, 1);
1907 
1908 #if VM_NRESERVLEVEL > 0
1909 		/*
1910 		 * Rename the reservation.
1911 		 */
1912 		vm_reserv_rename(p, object, backing_object,
1913 		    backing_offset_index);
1914 #endif
1915 		vm_page_xunbusy(p);
1916 	}
1917 	return;
1918 }
1919 
1920 /*
1921  *	vm_object_collapse:
1922  *
1923  *	Collapse an object with the object backing it.
1924  *	Pages in the backing object are moved into the
1925  *	parent, and the backing object is deallocated.
1926  */
1927 void
1928 vm_object_collapse(vm_object_t object)
1929 {
1930 	vm_object_t backing_object, new_backing_object;
1931 
1932 	VM_OBJECT_ASSERT_WLOCKED(object);
1933 
1934 	while (TRUE) {
1935 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1936 		    ("collapsing invalid object"));
1937 
1938 		/*
1939 		 * Wait for the backing_object to finish any pending
1940 		 * collapse so that the caller sees the shortest possible
1941 		 * shadow chain.
1942 		 */
1943 		backing_object = vm_object_backing_collapse_wait(object);
1944 		if (backing_object == NULL)
1945 			return;
1946 
1947 		KASSERT(object->ref_count > 0 &&
1948 		    object->ref_count > atomic_load_int(&object->shadow_count),
1949 		    ("collapse with invalid ref %d or shadow %d count.",
1950 		    object->ref_count, atomic_load_int(&object->shadow_count)));
1951 		KASSERT((backing_object->flags &
1952 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1953 		    ("vm_object_collapse: Backing object already collapsing."));
1954 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1955 		    ("vm_object_collapse: object is already collapsing."));
1956 
1957 		/*
1958 		 * We know that we can either collapse the backing object if
1959 		 * the parent is the only reference to it, or (perhaps) have
1960 		 * the parent bypass the object if the parent happens to shadow
1961 		 * all the resident pages in the entire backing object.
1962 		 */
1963 		if (backing_object->ref_count == 1) {
1964 			KASSERT(atomic_load_int(&backing_object->shadow_count)
1965 			    == 1,
1966 			    ("vm_object_collapse: shadow_count: %d",
1967 			    atomic_load_int(&backing_object->shadow_count)));
1968 			vm_object_pip_add(object, 1);
1969 			vm_object_set_flag(object, OBJ_COLLAPSING);
1970 			vm_object_pip_add(backing_object, 1);
1971 			vm_object_set_flag(backing_object, OBJ_DEAD);
1972 
1973 			/*
1974 			 * If there is exactly one reference to the backing
1975 			 * object, we can collapse it into the parent.
1976 			 */
1977 			vm_object_collapse_scan(object);
1978 
1979 #if VM_NRESERVLEVEL > 0
1980 			/*
1981 			 * Break any reservations from backing_object.
1982 			 */
1983 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1984 				vm_reserv_break_all(backing_object);
1985 #endif
1986 
1987 			/*
1988 			 * Move the pager from backing_object to object.
1989 			 */
1990 			if ((backing_object->flags & OBJ_SWAP) != 0) {
1991 				/*
1992 				 * swap_pager_copy() can sleep, in which case
1993 				 * the backing_object's and object's locks are
1994 				 * released and reacquired.
1995 				 * Since swap_pager_copy() is being asked to
1996 				 * destroy backing_object, it will change the
1997 				 * type to OBJT_DEFAULT.
1998 				 */
1999 				swap_pager_copy(
2000 				    backing_object,
2001 				    object,
2002 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
2003 			}
2004 
2005 			/*
2006 			 * Object now shadows whatever backing_object did.
2007 			 */
2008 			vm_object_clear_flag(object, OBJ_COLLAPSING);
2009 			vm_object_backing_transfer(object, backing_object);
2010 			object->backing_object_offset +=
2011 			    backing_object->backing_object_offset;
2012 			VM_OBJECT_WUNLOCK(object);
2013 			vm_object_pip_wakeup(object);
2014 
2015 			/*
2016 			 * Discard backing_object.
2017 			 *
2018 			 * Since the backing object has no pages, no pager left,
2019 			 * and no object references within it, all that is
2020 			 * necessary is to dispose of it.
2021 			 */
2022 			KASSERT(backing_object->ref_count == 1, (
2023 "backing_object %p was somehow re-referenced during collapse!",
2024 			    backing_object));
2025 			vm_object_pip_wakeup(backing_object);
2026 			(void)refcount_release(&backing_object->ref_count);
2027 			vm_object_terminate(backing_object);
2028 			counter_u64_add(object_collapses, 1);
2029 			VM_OBJECT_WLOCK(object);
2030 		} else {
2031 			/*
2032 			 * If we do not entirely shadow the backing object,
2033 			 * there is nothing we can do so we give up.
2034 			 *
2035 			 * The object lock and backing_object lock must not
2036 			 * be dropped during this sequence.
2037 			 */
2038 			if (!vm_object_scan_all_shadowed(object)) {
2039 				VM_OBJECT_WUNLOCK(backing_object);
2040 				break;
2041 			}
2042 
2043 			/*
2044 			 * Make the parent shadow the next object in the
2045 			 * chain.  Deallocating backing_object will not remove
2046 			 * it, since its reference count is at least 2.
2047 			 */
2048 			vm_object_backing_remove_locked(object);
2049 			new_backing_object = backing_object->backing_object;
2050 			if (new_backing_object != NULL) {
2051 				vm_object_backing_insert_ref(object,
2052 				    new_backing_object);
2053 				object->backing_object_offset +=
2054 				    backing_object->backing_object_offset;
2055 			}
2056 
2057 			/*
2058 			 * Drop the reference count on backing_object. Since
2059 			 * its ref_count was at least 2, it will not vanish.
2060 			 */
2061 			(void)refcount_release(&backing_object->ref_count);
2062 			KASSERT(backing_object->ref_count >= 1, (
2063 "backing_object %p was somehow dereferenced during collapse!",
2064 			    backing_object));
2065 			VM_OBJECT_WUNLOCK(backing_object);
2066 			counter_u64_add(object_bypasses, 1);
2067 		}
2068 
2069 		/*
2070 		 * Try again with this object's new backing object.
2071 		 */
2072 	}
2073 }
2074 
2075 /*
2076  *	vm_object_page_remove:
2077  *
2078  *	For the given object, either frees or invalidates each of the
2079  *	specified pages.  In general, a page is freed.  However, if a page is
2080  *	wired for any reason other than the existence of a managed, wired
2081  *	mapping, then it may be invalidated but not removed from the object.
2082  *	Pages are specified by the given range ["start", "end") and the option
2083  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
2084  *	extends from "start" to the end of the object.  If the option
2085  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2086  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
2087  *	specified, then the pages within the specified range must have no
2088  *	mappings.  Otherwise, if this option is not specified, any mappings to
2089  *	the specified pages are removed before the pages are freed or
2090  *	invalidated.
2091  *
2092  *	In general, this operation should only be performed on objects that
2093  *	contain managed pages.  There are, however, two exceptions.  First, it
2094  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
2095  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2096  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
2097  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
2098  *
2099  *	The object must be locked.
2100  */
2101 void
2102 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2103     int options)
2104 {
2105 	vm_page_t p, next;
2106 
2107 	VM_OBJECT_ASSERT_WLOCKED(object);
2108 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2109 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2110 	    ("vm_object_page_remove: illegal options for object %p", object));
2111 	if (object->resident_page_count == 0)
2112 		return;
2113 	vm_object_pip_add(object, 1);
2114 again:
2115 	p = vm_page_find_least(object, start);
2116 
2117 	/*
2118 	 * Here, the variable "p" is either (1) the page with the least pindex
2119 	 * greater than or equal to the parameter "start" or (2) NULL.
2120 	 */
2121 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2122 		next = TAILQ_NEXT(p, listq);
2123 
2124 		/*
2125 		 * Skip invalid pages if asked to do so.  Try to avoid acquiring
2126 		 * the busy lock, as some consumers rely on this to avoid
2127 		 * deadlocks.
2128 		 *
2129 		 * A thread may concurrently transition the page from invalid to
2130 		 * valid using only the busy lock, so the result of this check
2131 		 * is immediately stale.  It is up to consumers to handle this,
2132 		 * for instance by ensuring that all invalid->valid transitions
2133 		 * happen with a mutex held, as may be possible for a
2134 		 * filesystem.
2135 		 */
2136 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2137 			continue;
2138 
2139 		/*
2140 		 * If the page is wired for any reason besides the existence
2141 		 * of managed, wired mappings, then it cannot be freed.  For
2142 		 * example, fictitious pages, which represent device memory,
2143 		 * are inherently wired and cannot be freed.  They can,
2144 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2145 		 * not specified.
2146 		 */
2147 		if (vm_page_tryxbusy(p) == 0) {
2148 			if (vm_page_busy_sleep(p, "vmopar", 0))
2149 				VM_OBJECT_WLOCK(object);
2150 			goto again;
2151 		}
2152 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2153 			vm_page_xunbusy(p);
2154 			continue;
2155 		}
2156 		if (vm_page_wired(p)) {
2157 wired:
2158 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2159 			    object->ref_count != 0)
2160 				pmap_remove_all(p);
2161 			if ((options & OBJPR_CLEANONLY) == 0) {
2162 				vm_page_invalid(p);
2163 				vm_page_undirty(p);
2164 			}
2165 			vm_page_xunbusy(p);
2166 			continue;
2167 		}
2168 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2169 		    ("vm_object_page_remove: page %p is fictitious", p));
2170 		if ((options & OBJPR_CLEANONLY) != 0 &&
2171 		    !vm_page_none_valid(p)) {
2172 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2173 			    object->ref_count != 0 &&
2174 			    !vm_page_try_remove_write(p))
2175 				goto wired;
2176 			if (p->dirty != 0) {
2177 				vm_page_xunbusy(p);
2178 				continue;
2179 			}
2180 		}
2181 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2182 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2183 			goto wired;
2184 		vm_page_free(p);
2185 	}
2186 	vm_object_pip_wakeup(object);
2187 
2188 	vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2189 	    start);
2190 }
2191 
2192 /*
2193  *	vm_object_page_noreuse:
2194  *
2195  *	For the given object, attempt to move the specified pages to
2196  *	the head of the inactive queue.  This bypasses regular LRU
2197  *	operation and allows the pages to be reused quickly under memory
2198  *	pressure.  If a page is wired for any reason, then it will not
2199  *	be queued.  Pages are specified by the range ["start", "end").
2200  *	As a special case, if "end" is zero, then the range extends from
2201  *	"start" to the end of the object.
2202  *
2203  *	This operation should only be performed on objects that
2204  *	contain non-fictitious, managed pages.
2205  *
2206  *	The object must be locked.
2207  */
2208 void
2209 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2210 {
2211 	vm_page_t p, next;
2212 
2213 	VM_OBJECT_ASSERT_LOCKED(object);
2214 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2215 	    ("vm_object_page_noreuse: illegal object %p", object));
2216 	if (object->resident_page_count == 0)
2217 		return;
2218 	p = vm_page_find_least(object, start);
2219 
2220 	/*
2221 	 * Here, the variable "p" is either (1) the page with the least pindex
2222 	 * greater than or equal to the parameter "start" or (2) NULL.
2223 	 */
2224 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2225 		next = TAILQ_NEXT(p, listq);
2226 		vm_page_deactivate_noreuse(p);
2227 	}
2228 }
2229 
2230 /*
2231  *	Populate the specified range of the object with valid pages.  Returns
2232  *	TRUE if the range is successfully populated and FALSE otherwise.
2233  *
2234  *	Note: This function should be optimized to pass a larger array of
2235  *	pages to vm_pager_get_pages() before it is applied to a non-
2236  *	OBJT_DEVICE object.
2237  *
2238  *	The object must be locked.
2239  */
2240 boolean_t
2241 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2242 {
2243 	vm_page_t m;
2244 	vm_pindex_t pindex;
2245 	int rv;
2246 
2247 	VM_OBJECT_ASSERT_WLOCKED(object);
2248 	for (pindex = start; pindex < end; pindex++) {
2249 		rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2250 		if (rv != VM_PAGER_OK)
2251 			break;
2252 
2253 		/*
2254 		 * Keep "m" busy because a subsequent iteration may unlock
2255 		 * the object.
2256 		 */
2257 	}
2258 	if (pindex > start) {
2259 		m = vm_page_lookup(object, start);
2260 		while (m != NULL && m->pindex < pindex) {
2261 			vm_page_xunbusy(m);
2262 			m = TAILQ_NEXT(m, listq);
2263 		}
2264 	}
2265 	return (pindex == end);
2266 }
2267 
2268 /*
2269  *	Routine:	vm_object_coalesce
2270  *	Function:	Coalesces two objects backing up adjoining
2271  *			regions of memory into a single object.
2272  *
2273  *	returns TRUE if objects were combined.
2274  *
2275  *	NOTE:	Only works at the moment if the second object is NULL -
2276  *		if it's not, which object do we lock first?
2277  *
2278  *	Parameters:
2279  *		prev_object	First object to coalesce
2280  *		prev_offset	Offset into prev_object
2281  *		prev_size	Size of reference to prev_object
2282  *		next_size	Size of reference to the second object
2283  *		reserved	Indicator that extension region has
2284  *				swap accounted for
2285  *
2286  *	Conditions:
2287  *	The object must *not* be locked.
2288  */
2289 boolean_t
2290 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2291     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2292 {
2293 	vm_pindex_t next_pindex;
2294 
2295 	if (prev_object == NULL)
2296 		return (TRUE);
2297 	if ((prev_object->flags & OBJ_ANON) == 0)
2298 		return (FALSE);
2299 
2300 	VM_OBJECT_WLOCK(prev_object);
2301 	/*
2302 	 * Try to collapse the object first.
2303 	 */
2304 	vm_object_collapse(prev_object);
2305 
2306 	/*
2307 	 * Can't coalesce if: . more than one reference . paged out . shadows
2308 	 * another object . has a copy elsewhere (any of which mean that the
2309 	 * pages not mapped to prev_entry may be in use anyway)
2310 	 */
2311 	if (prev_object->backing_object != NULL) {
2312 		VM_OBJECT_WUNLOCK(prev_object);
2313 		return (FALSE);
2314 	}
2315 
2316 	prev_size >>= PAGE_SHIFT;
2317 	next_size >>= PAGE_SHIFT;
2318 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2319 
2320 	if (prev_object->ref_count > 1 &&
2321 	    prev_object->size != next_pindex &&
2322 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2323 		VM_OBJECT_WUNLOCK(prev_object);
2324 		return (FALSE);
2325 	}
2326 
2327 	/*
2328 	 * Account for the charge.
2329 	 */
2330 	if (prev_object->cred != NULL) {
2331 		/*
2332 		 * If prev_object was charged, then this mapping,
2333 		 * although not charged now, may become writable
2334 		 * later. Non-NULL cred in the object would prevent
2335 		 * swap reservation during enabling of the write
2336 		 * access, so reserve swap now. Failed reservation
2337 		 * cause allocation of the separate object for the map
2338 		 * entry, and swap reservation for this entry is
2339 		 * managed in appropriate time.
2340 		 */
2341 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2342 		    prev_object->cred)) {
2343 			VM_OBJECT_WUNLOCK(prev_object);
2344 			return (FALSE);
2345 		}
2346 		prev_object->charge += ptoa(next_size);
2347 	}
2348 
2349 	/*
2350 	 * Remove any pages that may still be in the object from a previous
2351 	 * deallocation.
2352 	 */
2353 	if (next_pindex < prev_object->size) {
2354 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2355 		    next_size, 0);
2356 #if 0
2357 		if (prev_object->cred != NULL) {
2358 			KASSERT(prev_object->charge >=
2359 			    ptoa(prev_object->size - next_pindex),
2360 			    ("object %p overcharged 1 %jx %jx", prev_object,
2361 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2362 			prev_object->charge -= ptoa(prev_object->size -
2363 			    next_pindex);
2364 		}
2365 #endif
2366 	}
2367 
2368 	/*
2369 	 * Extend the object if necessary.
2370 	 */
2371 	if (next_pindex + next_size > prev_object->size)
2372 		prev_object->size = next_pindex + next_size;
2373 
2374 	VM_OBJECT_WUNLOCK(prev_object);
2375 	return (TRUE);
2376 }
2377 
2378 void
2379 vm_object_set_writeable_dirty_(vm_object_t object)
2380 {
2381 	atomic_add_int(&object->generation, 1);
2382 }
2383 
2384 bool
2385 vm_object_mightbedirty_(vm_object_t object)
2386 {
2387 	return (object->generation != object->cleangeneration);
2388 }
2389 
2390 /*
2391  *	vm_object_unwire:
2392  *
2393  *	For each page offset within the specified range of the given object,
2394  *	find the highest-level page in the shadow chain and unwire it.  A page
2395  *	must exist at every page offset, and the highest-level page must be
2396  *	wired.
2397  */
2398 void
2399 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2400     uint8_t queue)
2401 {
2402 	vm_object_t tobject, t1object;
2403 	vm_page_t m, tm;
2404 	vm_pindex_t end_pindex, pindex, tpindex;
2405 	int depth, locked_depth;
2406 
2407 	KASSERT((offset & PAGE_MASK) == 0,
2408 	    ("vm_object_unwire: offset is not page aligned"));
2409 	KASSERT((length & PAGE_MASK) == 0,
2410 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2411 	/* The wired count of a fictitious page never changes. */
2412 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2413 		return;
2414 	pindex = OFF_TO_IDX(offset);
2415 	end_pindex = pindex + atop(length);
2416 again:
2417 	locked_depth = 1;
2418 	VM_OBJECT_RLOCK(object);
2419 	m = vm_page_find_least(object, pindex);
2420 	while (pindex < end_pindex) {
2421 		if (m == NULL || pindex < m->pindex) {
2422 			/*
2423 			 * The first object in the shadow chain doesn't
2424 			 * contain a page at the current index.  Therefore,
2425 			 * the page must exist in a backing object.
2426 			 */
2427 			tobject = object;
2428 			tpindex = pindex;
2429 			depth = 0;
2430 			do {
2431 				tpindex +=
2432 				    OFF_TO_IDX(tobject->backing_object_offset);
2433 				tobject = tobject->backing_object;
2434 				KASSERT(tobject != NULL,
2435 				    ("vm_object_unwire: missing page"));
2436 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2437 					goto next_page;
2438 				depth++;
2439 				if (depth == locked_depth) {
2440 					locked_depth++;
2441 					VM_OBJECT_RLOCK(tobject);
2442 				}
2443 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2444 			    NULL);
2445 		} else {
2446 			tm = m;
2447 			m = TAILQ_NEXT(m, listq);
2448 		}
2449 		if (vm_page_trysbusy(tm) == 0) {
2450 			for (tobject = object; locked_depth >= 1;
2451 			    locked_depth--) {
2452 				t1object = tobject->backing_object;
2453 				if (tm->object != tobject)
2454 					VM_OBJECT_RUNLOCK(tobject);
2455 				tobject = t1object;
2456 			}
2457 			tobject = tm->object;
2458 			if (!vm_page_busy_sleep(tm, "unwbo",
2459 			    VM_ALLOC_IGN_SBUSY))
2460 				VM_OBJECT_RUNLOCK(tobject);
2461 			goto again;
2462 		}
2463 		vm_page_unwire(tm, queue);
2464 		vm_page_sunbusy(tm);
2465 next_page:
2466 		pindex++;
2467 	}
2468 	/* Release the accumulated object locks. */
2469 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2470 		t1object = tobject->backing_object;
2471 		VM_OBJECT_RUNLOCK(tobject);
2472 		tobject = t1object;
2473 	}
2474 }
2475 
2476 /*
2477  * Return the vnode for the given object, or NULL if none exists.
2478  * For tmpfs objects, the function may return NULL if there is
2479  * no vnode allocated at the time of the call.
2480  */
2481 struct vnode *
2482 vm_object_vnode(vm_object_t object)
2483 {
2484 	struct vnode *vp;
2485 
2486 	VM_OBJECT_ASSERT_LOCKED(object);
2487 	vm_pager_getvp(object, &vp, NULL);
2488 	return (vp);
2489 }
2490 
2491 /*
2492  * Busy the vm object.  This prevents new pages belonging to the object from
2493  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2494  * for checking page state before proceeding.
2495  */
2496 void
2497 vm_object_busy(vm_object_t obj)
2498 {
2499 
2500 	VM_OBJECT_ASSERT_LOCKED(obj);
2501 
2502 	blockcount_acquire(&obj->busy, 1);
2503 	/* The fence is required to order loads of page busy. */
2504 	atomic_thread_fence_acq_rel();
2505 }
2506 
2507 void
2508 vm_object_unbusy(vm_object_t obj)
2509 {
2510 
2511 	blockcount_release(&obj->busy, 1);
2512 }
2513 
2514 void
2515 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2516 {
2517 
2518 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2519 
2520 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2521 }
2522 
2523 /*
2524  * This function aims to determine if the object is mapped,
2525  * specifically, if it is referenced by a vm_map_entry.  Because
2526  * objects occasionally acquire transient references that do not
2527  * represent a mapping, the method used here is inexact.  However, it
2528  * has very low overhead and is good enough for the advisory
2529  * vm.vmtotal sysctl.
2530  */
2531 bool
2532 vm_object_is_active(vm_object_t obj)
2533 {
2534 
2535 	return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2536 }
2537 
2538 static int
2539 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2540 {
2541 	struct kinfo_vmobject *kvo;
2542 	char *fullpath, *freepath;
2543 	struct vnode *vp;
2544 	struct vattr va;
2545 	vm_object_t obj;
2546 	vm_page_t m;
2547 	u_long sp;
2548 	int count, error;
2549 
2550 	if (req->oldptr == NULL) {
2551 		/*
2552 		 * If an old buffer has not been provided, generate an
2553 		 * estimate of the space needed for a subsequent call.
2554 		 */
2555 		mtx_lock(&vm_object_list_mtx);
2556 		count = 0;
2557 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2558 			if (obj->type == OBJT_DEAD)
2559 				continue;
2560 			count++;
2561 		}
2562 		mtx_unlock(&vm_object_list_mtx);
2563 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2564 		    count * 11 / 10));
2565 	}
2566 
2567 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK);
2568 	error = 0;
2569 
2570 	/*
2571 	 * VM objects are type stable and are never removed from the
2572 	 * list once added.  This allows us to safely read obj->object_list
2573 	 * after reacquiring the VM object lock.
2574 	 */
2575 	mtx_lock(&vm_object_list_mtx);
2576 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2577 		if (obj->type == OBJT_DEAD ||
2578 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2579 			continue;
2580 		VM_OBJECT_RLOCK(obj);
2581 		if (obj->type == OBJT_DEAD ||
2582 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2583 			VM_OBJECT_RUNLOCK(obj);
2584 			continue;
2585 		}
2586 		mtx_unlock(&vm_object_list_mtx);
2587 		kvo->kvo_size = ptoa(obj->size);
2588 		kvo->kvo_resident = obj->resident_page_count;
2589 		kvo->kvo_ref_count = obj->ref_count;
2590 		kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2591 		kvo->kvo_memattr = obj->memattr;
2592 		kvo->kvo_active = 0;
2593 		kvo->kvo_inactive = 0;
2594 		if (!swap_only) {
2595 			TAILQ_FOREACH(m, &obj->memq, listq) {
2596 				/*
2597 				 * A page may belong to the object but be
2598 				 * dequeued and set to PQ_NONE while the
2599 				 * object lock is not held.  This makes the
2600 				 * reads of m->queue below racy, and we do not
2601 				 * count pages set to PQ_NONE.  However, this
2602 				 * sysctl is only meant to give an
2603 				 * approximation of the system anyway.
2604 				 */
2605 				if (m->a.queue == PQ_ACTIVE)
2606 					kvo->kvo_active++;
2607 				else if (m->a.queue == PQ_INACTIVE)
2608 					kvo->kvo_inactive++;
2609 			}
2610 		}
2611 
2612 		kvo->kvo_vn_fileid = 0;
2613 		kvo->kvo_vn_fsid = 0;
2614 		kvo->kvo_vn_fsid_freebsd11 = 0;
2615 		freepath = NULL;
2616 		fullpath = "";
2617 		vp = NULL;
2618 		kvo->kvo_type = vm_object_kvme_type(obj, swap_only ? NULL : &vp);
2619 		if (vp != NULL) {
2620 			vref(vp);
2621 		} else if ((obj->flags & OBJ_ANON) != 0) {
2622 			MPASS(kvo->kvo_type == KVME_TYPE_DEFAULT ||
2623 			    kvo->kvo_type == KVME_TYPE_SWAP);
2624 			kvo->kvo_me = (uintptr_t)obj;
2625 			/* tmpfs objs are reported as vnodes */
2626 			kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2627 			sp = swap_pager_swapped_pages(obj);
2628 			kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2629 		}
2630 		VM_OBJECT_RUNLOCK(obj);
2631 		if (vp != NULL) {
2632 			vn_fullpath(vp, &fullpath, &freepath);
2633 			vn_lock(vp, LK_SHARED | LK_RETRY);
2634 			if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2635 				kvo->kvo_vn_fileid = va.va_fileid;
2636 				kvo->kvo_vn_fsid = va.va_fsid;
2637 				kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2638 								/* truncate */
2639 			}
2640 			vput(vp);
2641 		}
2642 
2643 		strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2644 		if (freepath != NULL)
2645 			free(freepath, M_TEMP);
2646 
2647 		/* Pack record size down */
2648 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2649 		    + strlen(kvo->kvo_path) + 1;
2650 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2651 		    sizeof(uint64_t));
2652 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2653 		maybe_yield();
2654 		mtx_lock(&vm_object_list_mtx);
2655 		if (error)
2656 			break;
2657 	}
2658 	mtx_unlock(&vm_object_list_mtx);
2659 	free(kvo, M_TEMP);
2660 	return (error);
2661 }
2662 
2663 static int
2664 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2665 {
2666 	return (vm_object_list_handler(req, false));
2667 }
2668 
2669 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2670     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2671     "List of VM objects");
2672 
2673 static int
2674 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2675 {
2676 	return (vm_object_list_handler(req, true));
2677 }
2678 
2679 /*
2680  * This sysctl returns list of the anonymous or swap objects. Intent
2681  * is to provide stripped optimized list useful to analyze swap use.
2682  * Since technically non-swap (default) objects participate in the
2683  * shadow chains, and are converted to swap type as needed by swap
2684  * pager, we must report them.
2685  */
2686 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2687     CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2688     sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2689     "List of swap VM objects");
2690 
2691 #include "opt_ddb.h"
2692 #ifdef DDB
2693 #include <sys/kernel.h>
2694 
2695 #include <sys/cons.h>
2696 
2697 #include <ddb/ddb.h>
2698 
2699 static int
2700 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2701 {
2702 	vm_map_t tmpm;
2703 	vm_map_entry_t tmpe;
2704 	vm_object_t obj;
2705 
2706 	if (map == 0)
2707 		return 0;
2708 
2709 	if (entry == 0) {
2710 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2711 			if (_vm_object_in_map(map, object, tmpe)) {
2712 				return 1;
2713 			}
2714 		}
2715 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2716 		tmpm = entry->object.sub_map;
2717 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2718 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2719 				return 1;
2720 			}
2721 		}
2722 	} else if ((obj = entry->object.vm_object) != NULL) {
2723 		for (; obj; obj = obj->backing_object)
2724 			if (obj == object) {
2725 				return 1;
2726 			}
2727 	}
2728 	return 0;
2729 }
2730 
2731 static int
2732 vm_object_in_map(vm_object_t object)
2733 {
2734 	struct proc *p;
2735 
2736 	/* sx_slock(&allproc_lock); */
2737 	FOREACH_PROC_IN_SYSTEM(p) {
2738 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2739 			continue;
2740 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2741 			/* sx_sunlock(&allproc_lock); */
2742 			return 1;
2743 		}
2744 	}
2745 	/* sx_sunlock(&allproc_lock); */
2746 	if (_vm_object_in_map(kernel_map, object, 0))
2747 		return 1;
2748 	return 0;
2749 }
2750 
2751 DB_SHOW_COMMAND(vmochk, vm_object_check)
2752 {
2753 	vm_object_t object;
2754 
2755 	/*
2756 	 * make sure that internal objs are in a map somewhere
2757 	 * and none have zero ref counts.
2758 	 */
2759 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2760 		if ((object->flags & OBJ_ANON) != 0) {
2761 			if (object->ref_count == 0) {
2762 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2763 					(long)object->size);
2764 			}
2765 			if (!vm_object_in_map(object)) {
2766 				db_printf(
2767 			"vmochk: internal obj is not in a map: "
2768 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2769 				    object->ref_count, (u_long)object->size,
2770 				    (u_long)object->size,
2771 				    (void *)object->backing_object);
2772 			}
2773 		}
2774 		if (db_pager_quit)
2775 			return;
2776 	}
2777 }
2778 
2779 /*
2780  *	vm_object_print:	[ debug ]
2781  */
2782 DB_SHOW_COMMAND(object, vm_object_print_static)
2783 {
2784 	/* XXX convert args. */
2785 	vm_object_t object = (vm_object_t)addr;
2786 	boolean_t full = have_addr;
2787 
2788 	vm_page_t p;
2789 
2790 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2791 #define	count	was_count
2792 
2793 	int count;
2794 
2795 	if (object == NULL)
2796 		return;
2797 
2798 	db_iprintf(
2799 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2800 	    object, (int)object->type, (uintmax_t)object->size,
2801 	    object->resident_page_count, object->ref_count, object->flags,
2802 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2803 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2804 	    atomic_load_int(&object->shadow_count),
2805 	    object->backing_object ? object->backing_object->ref_count : 0,
2806 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2807 
2808 	if (!full)
2809 		return;
2810 
2811 	db_indent += 2;
2812 	count = 0;
2813 	TAILQ_FOREACH(p, &object->memq, listq) {
2814 		if (count == 0)
2815 			db_iprintf("memory:=");
2816 		else if (count == 6) {
2817 			db_printf("\n");
2818 			db_iprintf(" ...");
2819 			count = 0;
2820 		} else
2821 			db_printf(",");
2822 		count++;
2823 
2824 		db_printf("(off=0x%jx,page=0x%jx)",
2825 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2826 
2827 		if (db_pager_quit)
2828 			break;
2829 	}
2830 	if (count != 0)
2831 		db_printf("\n");
2832 	db_indent -= 2;
2833 }
2834 
2835 /* XXX. */
2836 #undef count
2837 
2838 /* XXX need this non-static entry for calling from vm_map_print. */
2839 void
2840 vm_object_print(
2841         /* db_expr_t */ long addr,
2842 	boolean_t have_addr,
2843 	/* db_expr_t */ long count,
2844 	char *modif)
2845 {
2846 	vm_object_print_static(addr, have_addr, count, modif);
2847 }
2848 
2849 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2850 {
2851 	vm_object_t object;
2852 	vm_pindex_t fidx;
2853 	vm_paddr_t pa;
2854 	vm_page_t m, prev_m;
2855 	int rcount;
2856 
2857 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2858 		db_printf("new object: %p\n", (void *)object);
2859 		if (db_pager_quit)
2860 			return;
2861 
2862 		rcount = 0;
2863 		fidx = 0;
2864 		pa = -1;
2865 		TAILQ_FOREACH(m, &object->memq, listq) {
2866 			if (m->pindex > 128)
2867 				break;
2868 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2869 			    prev_m->pindex + 1 != m->pindex) {
2870 				if (rcount) {
2871 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2872 						(long)fidx, rcount, (long)pa);
2873 					if (db_pager_quit)
2874 						return;
2875 					rcount = 0;
2876 				}
2877 			}
2878 			if (rcount &&
2879 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2880 				++rcount;
2881 				continue;
2882 			}
2883 			if (rcount) {
2884 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2885 					(long)fidx, rcount, (long)pa);
2886 				if (db_pager_quit)
2887 					return;
2888 			}
2889 			fidx = m->pindex;
2890 			pa = VM_PAGE_TO_PHYS(m);
2891 			rcount = 1;
2892 		}
2893 		if (rcount) {
2894 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2895 				(long)fidx, rcount, (long)pa);
2896 			if (db_pager_quit)
2897 				return;
2898 		}
2899 	}
2900 }
2901 #endif /* DDB */
2902