xref: /freebsd/sys/vm/vm_object.c (revision 716fd348e01c5f2ba125f878a634a753436c2994)
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 		object->flags &= ~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 		object->flags |= 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 			result->flags |= source->flags & OBJ_COLORED;
1502 			result->pg_color = (source->pg_color +
1503 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1504 			    1)) - 1);
1505 #endif
1506 		}
1507 		VM_OBJECT_WUNLOCK(result);
1508 	}
1509 
1510 	/*
1511 	 * Return the new things
1512 	 */
1513 	*offset = 0;
1514 	*object = result;
1515 }
1516 
1517 /*
1518  *	vm_object_split:
1519  *
1520  * Split the pages in a map entry into a new object.  This affords
1521  * easier removal of unused pages, and keeps object inheritance from
1522  * being a negative impact on memory usage.
1523  */
1524 void
1525 vm_object_split(vm_map_entry_t entry)
1526 {
1527 	vm_page_t m, m_busy, m_next;
1528 	vm_object_t orig_object, new_object, backing_object;
1529 	vm_pindex_t idx, offidxstart;
1530 	vm_size_t size;
1531 
1532 	orig_object = entry->object.vm_object;
1533 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1534 	    ("vm_object_split:  Splitting object with multiple mappings."));
1535 	if ((orig_object->flags & OBJ_ANON) == 0)
1536 		return;
1537 	if (orig_object->ref_count <= 1)
1538 		return;
1539 	VM_OBJECT_WUNLOCK(orig_object);
1540 
1541 	offidxstart = OFF_TO_IDX(entry->offset);
1542 	size = atop(entry->end - entry->start);
1543 
1544 	/*
1545 	 * If swap_pager_copy() is later called, it will convert new_object
1546 	 * into a swap object.
1547 	 */
1548 	new_object = vm_object_allocate_anon(size, orig_object,
1549 	    orig_object->cred, ptoa(size));
1550 
1551 	/*
1552 	 * We must wait for the orig_object to complete any in-progress
1553 	 * collapse so that the swap blocks are stable below.  The
1554 	 * additional reference on backing_object by new object will
1555 	 * prevent further collapse operations until split completes.
1556 	 */
1557 	VM_OBJECT_WLOCK(orig_object);
1558 	vm_object_collapse_wait(orig_object);
1559 
1560 	/*
1561 	 * At this point, the new object is still private, so the order in
1562 	 * which the original and new objects are locked does not matter.
1563 	 */
1564 	VM_OBJECT_WLOCK(new_object);
1565 	new_object->domain = orig_object->domain;
1566 	backing_object = orig_object->backing_object;
1567 	if (backing_object != NULL) {
1568 		vm_object_backing_insert_ref(new_object, backing_object);
1569 		new_object->backing_object_offset =
1570 		    orig_object->backing_object_offset + entry->offset;
1571 	}
1572 	if (orig_object->cred != NULL) {
1573 		crhold(orig_object->cred);
1574 		KASSERT(orig_object->charge >= ptoa(size),
1575 		    ("orig_object->charge < 0"));
1576 		orig_object->charge -= ptoa(size);
1577 	}
1578 
1579 	/*
1580 	 * Mark the split operation so that swap_pager_getpages() knows
1581 	 * that the object is in transition.
1582 	 */
1583 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1584 	m_busy = NULL;
1585 #ifdef INVARIANTS
1586 	idx = 0;
1587 #endif
1588 retry:
1589 	m = vm_page_find_least(orig_object, offidxstart);
1590 	KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1591 	    ("%s: object %p was repopulated", __func__, orig_object));
1592 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1593 	    m = m_next) {
1594 		m_next = TAILQ_NEXT(m, listq);
1595 
1596 		/*
1597 		 * We must wait for pending I/O to complete before we can
1598 		 * rename the page.
1599 		 *
1600 		 * We do not have to VM_PROT_NONE the page as mappings should
1601 		 * not be changed by this operation.
1602 		 */
1603 		if (vm_page_tryxbusy(m) == 0) {
1604 			VM_OBJECT_WUNLOCK(new_object);
1605 			if (vm_page_busy_sleep(m, "spltwt", 0))
1606 				VM_OBJECT_WLOCK(orig_object);
1607 			VM_OBJECT_WLOCK(new_object);
1608 			goto retry;
1609 		}
1610 
1611 		/*
1612 		 * The page was left invalid.  Likely placed there by
1613 		 * an incomplete fault.  Just remove and ignore.
1614 		 */
1615 		if (vm_page_none_valid(m)) {
1616 			if (vm_page_remove(m))
1617 				vm_page_free(m);
1618 			continue;
1619 		}
1620 
1621 		/* vm_page_rename() will dirty the page. */
1622 		if (vm_page_rename(m, new_object, idx)) {
1623 			vm_page_xunbusy(m);
1624 			VM_OBJECT_WUNLOCK(new_object);
1625 			VM_OBJECT_WUNLOCK(orig_object);
1626 			vm_radix_wait();
1627 			VM_OBJECT_WLOCK(orig_object);
1628 			VM_OBJECT_WLOCK(new_object);
1629 			goto retry;
1630 		}
1631 
1632 #if VM_NRESERVLEVEL > 0
1633 		/*
1634 		 * If some of the reservation's allocated pages remain with
1635 		 * the original object, then transferring the reservation to
1636 		 * the new object is neither particularly beneficial nor
1637 		 * particularly harmful as compared to leaving the reservation
1638 		 * with the original object.  If, however, all of the
1639 		 * reservation's allocated pages are transferred to the new
1640 		 * object, then transferring the reservation is typically
1641 		 * beneficial.  Determining which of these two cases applies
1642 		 * would be more costly than unconditionally renaming the
1643 		 * reservation.
1644 		 */
1645 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1646 #endif
1647 
1648 		/*
1649 		 * orig_object's type may change while sleeping, so keep track
1650 		 * of the beginning of the busied range.
1651 		 */
1652 		if (orig_object->type != OBJT_SWAP)
1653 			vm_page_xunbusy(m);
1654 		else if (m_busy == NULL)
1655 			m_busy = m;
1656 	}
1657 	if ((orig_object->flags & OBJ_SWAP) != 0) {
1658 		/*
1659 		 * swap_pager_copy() can sleep, in which case the orig_object's
1660 		 * and new_object's locks are released and reacquired.
1661 		 */
1662 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1663 		if (m_busy != NULL)
1664 			TAILQ_FOREACH_FROM(m_busy, &new_object->memq, listq)
1665 				vm_page_xunbusy(m_busy);
1666 	}
1667 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1668 	VM_OBJECT_WUNLOCK(orig_object);
1669 	VM_OBJECT_WUNLOCK(new_object);
1670 	entry->object.vm_object = new_object;
1671 	entry->offset = 0LL;
1672 	vm_object_deallocate(orig_object);
1673 	VM_OBJECT_WLOCK(new_object);
1674 }
1675 
1676 static vm_page_t
1677 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1678 {
1679 	vm_object_t backing_object;
1680 
1681 	VM_OBJECT_ASSERT_WLOCKED(object);
1682 	backing_object = object->backing_object;
1683 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1684 
1685 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1686 	    ("invalid ownership %p %p %p", p, object, backing_object));
1687 	/* The page is only NULL when rename fails. */
1688 	if (p == NULL) {
1689 		VM_OBJECT_WUNLOCK(object);
1690 		VM_OBJECT_WUNLOCK(backing_object);
1691 		vm_radix_wait();
1692 		VM_OBJECT_WLOCK(object);
1693 	} else if (p->object == object) {
1694 		VM_OBJECT_WUNLOCK(backing_object);
1695 		if (vm_page_busy_sleep(p, "vmocol", 0))
1696 			VM_OBJECT_WLOCK(object);
1697 	} else {
1698 		VM_OBJECT_WUNLOCK(object);
1699 		if (!vm_page_busy_sleep(p, "vmocol", 0))
1700 			VM_OBJECT_WUNLOCK(backing_object);
1701 		VM_OBJECT_WLOCK(object);
1702 	}
1703 	VM_OBJECT_WLOCK(backing_object);
1704 	return (TAILQ_FIRST(&backing_object->memq));
1705 }
1706 
1707 static bool
1708 vm_object_scan_all_shadowed(vm_object_t object)
1709 {
1710 	vm_object_t backing_object;
1711 	vm_page_t p, pp;
1712 	vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1713 
1714 	VM_OBJECT_ASSERT_WLOCKED(object);
1715 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1716 
1717 	backing_object = object->backing_object;
1718 
1719 	if ((backing_object->flags & OBJ_ANON) == 0)
1720 		return (false);
1721 
1722 	pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1723 	p = vm_page_find_least(backing_object, pi);
1724 	ps = swap_pager_find_least(backing_object, pi);
1725 
1726 	/*
1727 	 * Only check pages inside the parent object's range and
1728 	 * inside the parent object's mapping of the backing object.
1729 	 */
1730 	for (;; pi++) {
1731 		if (p != NULL && p->pindex < pi)
1732 			p = TAILQ_NEXT(p, listq);
1733 		if (ps < pi)
1734 			ps = swap_pager_find_least(backing_object, pi);
1735 		if (p == NULL && ps >= backing_object->size)
1736 			break;
1737 		else if (p == NULL)
1738 			pi = ps;
1739 		else
1740 			pi = MIN(p->pindex, ps);
1741 
1742 		new_pindex = pi - backing_offset_index;
1743 		if (new_pindex >= object->size)
1744 			break;
1745 
1746 		if (p != NULL) {
1747 			/*
1748 			 * If the backing object page is busy a
1749 			 * grandparent or older page may still be
1750 			 * undergoing CoW.  It is not safe to collapse
1751 			 * the backing object until it is quiesced.
1752 			 */
1753 			if (vm_page_tryxbusy(p) == 0)
1754 				return (false);
1755 
1756 			/*
1757 			 * We raced with the fault handler that left
1758 			 * newly allocated invalid page on the object
1759 			 * queue and retried.
1760 			 */
1761 			if (!vm_page_all_valid(p))
1762 				goto unbusy_ret;
1763 		}
1764 
1765 		/*
1766 		 * See if the parent has the page or if the parent's object
1767 		 * pager has the page.  If the parent has the page but the page
1768 		 * is not valid, the parent's object pager must have the page.
1769 		 *
1770 		 * If this fails, the parent does not completely shadow the
1771 		 * object and we might as well give up now.
1772 		 */
1773 		pp = vm_page_lookup(object, new_pindex);
1774 
1775 		/*
1776 		 * The valid check here is stable due to object lock
1777 		 * being required to clear valid and initiate paging.
1778 		 * Busy of p disallows fault handler to validate pp.
1779 		 */
1780 		if ((pp == NULL || vm_page_none_valid(pp)) &&
1781 		    !vm_pager_has_page(object, new_pindex, NULL, NULL))
1782 			goto unbusy_ret;
1783 		if (p != NULL)
1784 			vm_page_xunbusy(p);
1785 	}
1786 	return (true);
1787 
1788 unbusy_ret:
1789 	if (p != NULL)
1790 		vm_page_xunbusy(p);
1791 	return (false);
1792 }
1793 
1794 static void
1795 vm_object_collapse_scan(vm_object_t object)
1796 {
1797 	vm_object_t backing_object;
1798 	vm_page_t next, p, pp;
1799 	vm_pindex_t backing_offset_index, new_pindex;
1800 
1801 	VM_OBJECT_ASSERT_WLOCKED(object);
1802 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1803 
1804 	backing_object = object->backing_object;
1805 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1806 
1807 	/*
1808 	 * Our scan
1809 	 */
1810 	for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1811 		next = TAILQ_NEXT(p, listq);
1812 		new_pindex = p->pindex - backing_offset_index;
1813 
1814 		/*
1815 		 * Check for busy page
1816 		 */
1817 		if (vm_page_tryxbusy(p) == 0) {
1818 			next = vm_object_collapse_scan_wait(object, p);
1819 			continue;
1820 		}
1821 
1822 		KASSERT(object->backing_object == backing_object,
1823 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1824 		    object->backing_object, backing_object));
1825 		KASSERT(p->object == backing_object,
1826 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1827 		    p->object, backing_object));
1828 
1829 		if (p->pindex < backing_offset_index ||
1830 		    new_pindex >= object->size) {
1831 			vm_pager_freespace(backing_object, p->pindex, 1);
1832 
1833 			KASSERT(!pmap_page_is_mapped(p),
1834 			    ("freeing mapped page %p", p));
1835 			if (vm_page_remove(p))
1836 				vm_page_free(p);
1837 			continue;
1838 		}
1839 
1840 		if (!vm_page_all_valid(p)) {
1841 			KASSERT(!pmap_page_is_mapped(p),
1842 			    ("freeing mapped page %p", p));
1843 			if (vm_page_remove(p))
1844 				vm_page_free(p);
1845 			continue;
1846 		}
1847 
1848 		pp = vm_page_lookup(object, new_pindex);
1849 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1850 			vm_page_xunbusy(p);
1851 			/*
1852 			 * The page in the parent is busy and possibly not
1853 			 * (yet) valid.  Until its state is finalized by the
1854 			 * busy bit owner, we can't tell whether it shadows the
1855 			 * original page.
1856 			 */
1857 			next = vm_object_collapse_scan_wait(object, pp);
1858 			continue;
1859 		}
1860 
1861 		if (pp != NULL && vm_page_none_valid(pp)) {
1862 			/*
1863 			 * The page was invalid in the parent.  Likely placed
1864 			 * there by an incomplete fault.  Just remove and
1865 			 * ignore.  p can replace it.
1866 			 */
1867 			if (vm_page_remove(pp))
1868 				vm_page_free(pp);
1869 			pp = NULL;
1870 		}
1871 
1872 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1873 			NULL)) {
1874 			/*
1875 			 * The page already exists in the parent OR swap exists
1876 			 * for this location in the parent.  Leave the parent's
1877 			 * page alone.  Destroy the original page from the
1878 			 * backing object.
1879 			 */
1880 			vm_pager_freespace(backing_object, p->pindex, 1);
1881 			KASSERT(!pmap_page_is_mapped(p),
1882 			    ("freeing mapped page %p", p));
1883 			if (vm_page_remove(p))
1884 				vm_page_free(p);
1885 			if (pp != NULL)
1886 				vm_page_xunbusy(pp);
1887 			continue;
1888 		}
1889 
1890 		/*
1891 		 * Page does not exist in parent, rename the page from the
1892 		 * backing object to the main object.
1893 		 *
1894 		 * If the page was mapped to a process, it can remain mapped
1895 		 * through the rename.  vm_page_rename() will dirty the page.
1896 		 */
1897 		if (vm_page_rename(p, object, new_pindex)) {
1898 			vm_page_xunbusy(p);
1899 			next = vm_object_collapse_scan_wait(object, NULL);
1900 			continue;
1901 		}
1902 
1903 		/* Use the old pindex to free the right page. */
1904 		vm_pager_freespace(backing_object, new_pindex +
1905 		    backing_offset_index, 1);
1906 
1907 #if VM_NRESERVLEVEL > 0
1908 		/*
1909 		 * Rename the reservation.
1910 		 */
1911 		vm_reserv_rename(p, object, backing_object,
1912 		    backing_offset_index);
1913 #endif
1914 		vm_page_xunbusy(p);
1915 	}
1916 	return;
1917 }
1918 
1919 /*
1920  *	vm_object_collapse:
1921  *
1922  *	Collapse an object with the object backing it.
1923  *	Pages in the backing object are moved into the
1924  *	parent, and the backing object is deallocated.
1925  */
1926 void
1927 vm_object_collapse(vm_object_t object)
1928 {
1929 	vm_object_t backing_object, new_backing_object;
1930 
1931 	VM_OBJECT_ASSERT_WLOCKED(object);
1932 
1933 	while (TRUE) {
1934 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1935 		    ("collapsing invalid object"));
1936 
1937 		/*
1938 		 * Wait for the backing_object to finish any pending
1939 		 * collapse so that the caller sees the shortest possible
1940 		 * shadow chain.
1941 		 */
1942 		backing_object = vm_object_backing_collapse_wait(object);
1943 		if (backing_object == NULL)
1944 			return;
1945 
1946 		KASSERT(object->ref_count > 0 &&
1947 		    object->ref_count > atomic_load_int(&object->shadow_count),
1948 		    ("collapse with invalid ref %d or shadow %d count.",
1949 		    object->ref_count, atomic_load_int(&object->shadow_count)));
1950 		KASSERT((backing_object->flags &
1951 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1952 		    ("vm_object_collapse: Backing object already collapsing."));
1953 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1954 		    ("vm_object_collapse: object is already collapsing."));
1955 
1956 		/*
1957 		 * We know that we can either collapse the backing object if
1958 		 * the parent is the only reference to it, or (perhaps) have
1959 		 * the parent bypass the object if the parent happens to shadow
1960 		 * all the resident pages in the entire backing object.
1961 		 */
1962 		if (backing_object->ref_count == 1) {
1963 			KASSERT(atomic_load_int(&backing_object->shadow_count)
1964 			    == 1,
1965 			    ("vm_object_collapse: shadow_count: %d",
1966 			    atomic_load_int(&backing_object->shadow_count)));
1967 			vm_object_pip_add(object, 1);
1968 			vm_object_set_flag(object, OBJ_COLLAPSING);
1969 			vm_object_pip_add(backing_object, 1);
1970 			vm_object_set_flag(backing_object, OBJ_DEAD);
1971 
1972 			/*
1973 			 * If there is exactly one reference to the backing
1974 			 * object, we can collapse it into the parent.
1975 			 */
1976 			vm_object_collapse_scan(object);
1977 
1978 #if VM_NRESERVLEVEL > 0
1979 			/*
1980 			 * Break any reservations from backing_object.
1981 			 */
1982 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1983 				vm_reserv_break_all(backing_object);
1984 #endif
1985 
1986 			/*
1987 			 * Move the pager from backing_object to object.
1988 			 */
1989 			if ((backing_object->flags & OBJ_SWAP) != 0) {
1990 				/*
1991 				 * swap_pager_copy() can sleep, in which case
1992 				 * the backing_object's and object's locks are
1993 				 * released and reacquired.
1994 				 * Since swap_pager_copy() is being asked to
1995 				 * destroy backing_object, it will change the
1996 				 * type to OBJT_DEFAULT.
1997 				 */
1998 				swap_pager_copy(
1999 				    backing_object,
2000 				    object,
2001 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
2002 			}
2003 
2004 			/*
2005 			 * Object now shadows whatever backing_object did.
2006 			 */
2007 			vm_object_clear_flag(object, OBJ_COLLAPSING);
2008 			vm_object_backing_transfer(object, backing_object);
2009 			object->backing_object_offset +=
2010 			    backing_object->backing_object_offset;
2011 			VM_OBJECT_WUNLOCK(object);
2012 			vm_object_pip_wakeup(object);
2013 
2014 			/*
2015 			 * Discard backing_object.
2016 			 *
2017 			 * Since the backing object has no pages, no pager left,
2018 			 * and no object references within it, all that is
2019 			 * necessary is to dispose of it.
2020 			 */
2021 			KASSERT(backing_object->ref_count == 1, (
2022 "backing_object %p was somehow re-referenced during collapse!",
2023 			    backing_object));
2024 			vm_object_pip_wakeup(backing_object);
2025 			(void)refcount_release(&backing_object->ref_count);
2026 			vm_object_terminate(backing_object);
2027 			counter_u64_add(object_collapses, 1);
2028 			VM_OBJECT_WLOCK(object);
2029 		} else {
2030 			/*
2031 			 * If we do not entirely shadow the backing object,
2032 			 * there is nothing we can do so we give up.
2033 			 *
2034 			 * The object lock and backing_object lock must not
2035 			 * be dropped during this sequence.
2036 			 */
2037 			if (!vm_object_scan_all_shadowed(object)) {
2038 				VM_OBJECT_WUNLOCK(backing_object);
2039 				break;
2040 			}
2041 
2042 			/*
2043 			 * Make the parent shadow the next object in the
2044 			 * chain.  Deallocating backing_object will not remove
2045 			 * it, since its reference count is at least 2.
2046 			 */
2047 			vm_object_backing_remove_locked(object);
2048 			new_backing_object = backing_object->backing_object;
2049 			if (new_backing_object != NULL) {
2050 				vm_object_backing_insert_ref(object,
2051 				    new_backing_object);
2052 				object->backing_object_offset +=
2053 				    backing_object->backing_object_offset;
2054 			}
2055 
2056 			/*
2057 			 * Drop the reference count on backing_object. Since
2058 			 * its ref_count was at least 2, it will not vanish.
2059 			 */
2060 			(void)refcount_release(&backing_object->ref_count);
2061 			KASSERT(backing_object->ref_count >= 1, (
2062 "backing_object %p was somehow dereferenced during collapse!",
2063 			    backing_object));
2064 			VM_OBJECT_WUNLOCK(backing_object);
2065 			counter_u64_add(object_bypasses, 1);
2066 		}
2067 
2068 		/*
2069 		 * Try again with this object's new backing object.
2070 		 */
2071 	}
2072 }
2073 
2074 /*
2075  *	vm_object_page_remove:
2076  *
2077  *	For the given object, either frees or invalidates each of the
2078  *	specified pages.  In general, a page is freed.  However, if a page is
2079  *	wired for any reason other than the existence of a managed, wired
2080  *	mapping, then it may be invalidated but not removed from the object.
2081  *	Pages are specified by the given range ["start", "end") and the option
2082  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
2083  *	extends from "start" to the end of the object.  If the option
2084  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2085  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
2086  *	specified, then the pages within the specified range must have no
2087  *	mappings.  Otherwise, if this option is not specified, any mappings to
2088  *	the specified pages are removed before the pages are freed or
2089  *	invalidated.
2090  *
2091  *	In general, this operation should only be performed on objects that
2092  *	contain managed pages.  There are, however, two exceptions.  First, it
2093  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
2094  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2095  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
2096  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
2097  *
2098  *	The object must be locked.
2099  */
2100 void
2101 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2102     int options)
2103 {
2104 	vm_page_t p, next;
2105 
2106 	VM_OBJECT_ASSERT_WLOCKED(object);
2107 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2108 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2109 	    ("vm_object_page_remove: illegal options for object %p", object));
2110 	if (object->resident_page_count == 0)
2111 		return;
2112 	vm_object_pip_add(object, 1);
2113 again:
2114 	p = vm_page_find_least(object, start);
2115 
2116 	/*
2117 	 * Here, the variable "p" is either (1) the page with the least pindex
2118 	 * greater than or equal to the parameter "start" or (2) NULL.
2119 	 */
2120 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2121 		next = TAILQ_NEXT(p, listq);
2122 
2123 		/*
2124 		 * Skip invalid pages if asked to do so.  Try to avoid acquiring
2125 		 * the busy lock, as some consumers rely on this to avoid
2126 		 * deadlocks.
2127 		 *
2128 		 * A thread may concurrently transition the page from invalid to
2129 		 * valid using only the busy lock, so the result of this check
2130 		 * is immediately stale.  It is up to consumers to handle this,
2131 		 * for instance by ensuring that all invalid->valid transitions
2132 		 * happen with a mutex held, as may be possible for a
2133 		 * filesystem.
2134 		 */
2135 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2136 			continue;
2137 
2138 		/*
2139 		 * If the page is wired for any reason besides the existence
2140 		 * of managed, wired mappings, then it cannot be freed.  For
2141 		 * example, fictitious pages, which represent device memory,
2142 		 * are inherently wired and cannot be freed.  They can,
2143 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2144 		 * not specified.
2145 		 */
2146 		if (vm_page_tryxbusy(p) == 0) {
2147 			if (vm_page_busy_sleep(p, "vmopar", 0))
2148 				VM_OBJECT_WLOCK(object);
2149 			goto again;
2150 		}
2151 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2152 			vm_page_xunbusy(p);
2153 			continue;
2154 		}
2155 		if (vm_page_wired(p)) {
2156 wired:
2157 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2158 			    object->ref_count != 0)
2159 				pmap_remove_all(p);
2160 			if ((options & OBJPR_CLEANONLY) == 0) {
2161 				vm_page_invalid(p);
2162 				vm_page_undirty(p);
2163 			}
2164 			vm_page_xunbusy(p);
2165 			continue;
2166 		}
2167 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2168 		    ("vm_object_page_remove: page %p is fictitious", p));
2169 		if ((options & OBJPR_CLEANONLY) != 0 &&
2170 		    !vm_page_none_valid(p)) {
2171 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2172 			    object->ref_count != 0 &&
2173 			    !vm_page_try_remove_write(p))
2174 				goto wired;
2175 			if (p->dirty != 0) {
2176 				vm_page_xunbusy(p);
2177 				continue;
2178 			}
2179 		}
2180 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2181 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2182 			goto wired;
2183 		vm_page_free(p);
2184 	}
2185 	vm_object_pip_wakeup(object);
2186 
2187 	vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2188 	    start);
2189 }
2190 
2191 /*
2192  *	vm_object_page_noreuse:
2193  *
2194  *	For the given object, attempt to move the specified pages to
2195  *	the head of the inactive queue.  This bypasses regular LRU
2196  *	operation and allows the pages to be reused quickly under memory
2197  *	pressure.  If a page is wired for any reason, then it will not
2198  *	be queued.  Pages are specified by the range ["start", "end").
2199  *	As a special case, if "end" is zero, then the range extends from
2200  *	"start" to the end of the object.
2201  *
2202  *	This operation should only be performed on objects that
2203  *	contain non-fictitious, managed pages.
2204  *
2205  *	The object must be locked.
2206  */
2207 void
2208 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2209 {
2210 	vm_page_t p, next;
2211 
2212 	VM_OBJECT_ASSERT_LOCKED(object);
2213 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2214 	    ("vm_object_page_noreuse: illegal object %p", object));
2215 	if (object->resident_page_count == 0)
2216 		return;
2217 	p = vm_page_find_least(object, start);
2218 
2219 	/*
2220 	 * Here, the variable "p" is either (1) the page with the least pindex
2221 	 * greater than or equal to the parameter "start" or (2) NULL.
2222 	 */
2223 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2224 		next = TAILQ_NEXT(p, listq);
2225 		vm_page_deactivate_noreuse(p);
2226 	}
2227 }
2228 
2229 /*
2230  *	Populate the specified range of the object with valid pages.  Returns
2231  *	TRUE if the range is successfully populated and FALSE otherwise.
2232  *
2233  *	Note: This function should be optimized to pass a larger array of
2234  *	pages to vm_pager_get_pages() before it is applied to a non-
2235  *	OBJT_DEVICE object.
2236  *
2237  *	The object must be locked.
2238  */
2239 boolean_t
2240 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2241 {
2242 	vm_page_t m;
2243 	vm_pindex_t pindex;
2244 	int rv;
2245 
2246 	VM_OBJECT_ASSERT_WLOCKED(object);
2247 	for (pindex = start; pindex < end; pindex++) {
2248 		rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2249 		if (rv != VM_PAGER_OK)
2250 			break;
2251 
2252 		/*
2253 		 * Keep "m" busy because a subsequent iteration may unlock
2254 		 * the object.
2255 		 */
2256 	}
2257 	if (pindex > start) {
2258 		m = vm_page_lookup(object, start);
2259 		while (m != NULL && m->pindex < pindex) {
2260 			vm_page_xunbusy(m);
2261 			m = TAILQ_NEXT(m, listq);
2262 		}
2263 	}
2264 	return (pindex == end);
2265 }
2266 
2267 /*
2268  *	Routine:	vm_object_coalesce
2269  *	Function:	Coalesces two objects backing up adjoining
2270  *			regions of memory into a single object.
2271  *
2272  *	returns TRUE if objects were combined.
2273  *
2274  *	NOTE:	Only works at the moment if the second object is NULL -
2275  *		if it's not, which object do we lock first?
2276  *
2277  *	Parameters:
2278  *		prev_object	First object to coalesce
2279  *		prev_offset	Offset into prev_object
2280  *		prev_size	Size of reference to prev_object
2281  *		next_size	Size of reference to the second object
2282  *		reserved	Indicator that extension region has
2283  *				swap accounted for
2284  *
2285  *	Conditions:
2286  *	The object must *not* be locked.
2287  */
2288 boolean_t
2289 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2290     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2291 {
2292 	vm_pindex_t next_pindex;
2293 
2294 	if (prev_object == NULL)
2295 		return (TRUE);
2296 	if ((prev_object->flags & OBJ_ANON) == 0)
2297 		return (FALSE);
2298 
2299 	VM_OBJECT_WLOCK(prev_object);
2300 	/*
2301 	 * Try to collapse the object first.
2302 	 */
2303 	vm_object_collapse(prev_object);
2304 
2305 	/*
2306 	 * Can't coalesce if: . more than one reference . paged out . shadows
2307 	 * another object . has a copy elsewhere (any of which mean that the
2308 	 * pages not mapped to prev_entry may be in use anyway)
2309 	 */
2310 	if (prev_object->backing_object != NULL) {
2311 		VM_OBJECT_WUNLOCK(prev_object);
2312 		return (FALSE);
2313 	}
2314 
2315 	prev_size >>= PAGE_SHIFT;
2316 	next_size >>= PAGE_SHIFT;
2317 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2318 
2319 	if (prev_object->ref_count > 1 &&
2320 	    prev_object->size != next_pindex &&
2321 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2322 		VM_OBJECT_WUNLOCK(prev_object);
2323 		return (FALSE);
2324 	}
2325 
2326 	/*
2327 	 * Account for the charge.
2328 	 */
2329 	if (prev_object->cred != NULL) {
2330 		/*
2331 		 * If prev_object was charged, then this mapping,
2332 		 * although not charged now, may become writable
2333 		 * later. Non-NULL cred in the object would prevent
2334 		 * swap reservation during enabling of the write
2335 		 * access, so reserve swap now. Failed reservation
2336 		 * cause allocation of the separate object for the map
2337 		 * entry, and swap reservation for this entry is
2338 		 * managed in appropriate time.
2339 		 */
2340 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2341 		    prev_object->cred)) {
2342 			VM_OBJECT_WUNLOCK(prev_object);
2343 			return (FALSE);
2344 		}
2345 		prev_object->charge += ptoa(next_size);
2346 	}
2347 
2348 	/*
2349 	 * Remove any pages that may still be in the object from a previous
2350 	 * deallocation.
2351 	 */
2352 	if (next_pindex < prev_object->size) {
2353 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2354 		    next_size, 0);
2355 #if 0
2356 		if (prev_object->cred != NULL) {
2357 			KASSERT(prev_object->charge >=
2358 			    ptoa(prev_object->size - next_pindex),
2359 			    ("object %p overcharged 1 %jx %jx", prev_object,
2360 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2361 			prev_object->charge -= ptoa(prev_object->size -
2362 			    next_pindex);
2363 		}
2364 #endif
2365 	}
2366 
2367 	/*
2368 	 * Extend the object if necessary.
2369 	 */
2370 	if (next_pindex + next_size > prev_object->size)
2371 		prev_object->size = next_pindex + next_size;
2372 
2373 	VM_OBJECT_WUNLOCK(prev_object);
2374 	return (TRUE);
2375 }
2376 
2377 void
2378 vm_object_set_writeable_dirty_(vm_object_t object)
2379 {
2380 	atomic_add_int(&object->generation, 1);
2381 }
2382 
2383 bool
2384 vm_object_mightbedirty_(vm_object_t object)
2385 {
2386 	return (object->generation != object->cleangeneration);
2387 }
2388 
2389 /*
2390  *	vm_object_unwire:
2391  *
2392  *	For each page offset within the specified range of the given object,
2393  *	find the highest-level page in the shadow chain and unwire it.  A page
2394  *	must exist at every page offset, and the highest-level page must be
2395  *	wired.
2396  */
2397 void
2398 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2399     uint8_t queue)
2400 {
2401 	vm_object_t tobject, t1object;
2402 	vm_page_t m, tm;
2403 	vm_pindex_t end_pindex, pindex, tpindex;
2404 	int depth, locked_depth;
2405 
2406 	KASSERT((offset & PAGE_MASK) == 0,
2407 	    ("vm_object_unwire: offset is not page aligned"));
2408 	KASSERT((length & PAGE_MASK) == 0,
2409 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2410 	/* The wired count of a fictitious page never changes. */
2411 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2412 		return;
2413 	pindex = OFF_TO_IDX(offset);
2414 	end_pindex = pindex + atop(length);
2415 again:
2416 	locked_depth = 1;
2417 	VM_OBJECT_RLOCK(object);
2418 	m = vm_page_find_least(object, pindex);
2419 	while (pindex < end_pindex) {
2420 		if (m == NULL || pindex < m->pindex) {
2421 			/*
2422 			 * The first object in the shadow chain doesn't
2423 			 * contain a page at the current index.  Therefore,
2424 			 * the page must exist in a backing object.
2425 			 */
2426 			tobject = object;
2427 			tpindex = pindex;
2428 			depth = 0;
2429 			do {
2430 				tpindex +=
2431 				    OFF_TO_IDX(tobject->backing_object_offset);
2432 				tobject = tobject->backing_object;
2433 				KASSERT(tobject != NULL,
2434 				    ("vm_object_unwire: missing page"));
2435 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2436 					goto next_page;
2437 				depth++;
2438 				if (depth == locked_depth) {
2439 					locked_depth++;
2440 					VM_OBJECT_RLOCK(tobject);
2441 				}
2442 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2443 			    NULL);
2444 		} else {
2445 			tm = m;
2446 			m = TAILQ_NEXT(m, listq);
2447 		}
2448 		if (vm_page_trysbusy(tm) == 0) {
2449 			for (tobject = object; locked_depth >= 1;
2450 			    locked_depth--) {
2451 				t1object = tobject->backing_object;
2452 				if (tm->object != tobject)
2453 					VM_OBJECT_RUNLOCK(tobject);
2454 				tobject = t1object;
2455 			}
2456 			tobject = tm->object;
2457 			if (!vm_page_busy_sleep(tm, "unwbo",
2458 			    VM_ALLOC_IGN_SBUSY))
2459 				VM_OBJECT_RUNLOCK(tobject);
2460 			goto again;
2461 		}
2462 		vm_page_unwire(tm, queue);
2463 		vm_page_sunbusy(tm);
2464 next_page:
2465 		pindex++;
2466 	}
2467 	/* Release the accumulated object locks. */
2468 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2469 		t1object = tobject->backing_object;
2470 		VM_OBJECT_RUNLOCK(tobject);
2471 		tobject = t1object;
2472 	}
2473 }
2474 
2475 /*
2476  * Return the vnode for the given object, or NULL if none exists.
2477  * For tmpfs objects, the function may return NULL if there is
2478  * no vnode allocated at the time of the call.
2479  */
2480 struct vnode *
2481 vm_object_vnode(vm_object_t object)
2482 {
2483 	struct vnode *vp;
2484 
2485 	VM_OBJECT_ASSERT_LOCKED(object);
2486 	vm_pager_getvp(object, &vp, NULL);
2487 	return (vp);
2488 }
2489 
2490 /*
2491  * Busy the vm object.  This prevents new pages belonging to the object from
2492  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2493  * for checking page state before proceeding.
2494  */
2495 void
2496 vm_object_busy(vm_object_t obj)
2497 {
2498 
2499 	VM_OBJECT_ASSERT_LOCKED(obj);
2500 
2501 	blockcount_acquire(&obj->busy, 1);
2502 	/* The fence is required to order loads of page busy. */
2503 	atomic_thread_fence_acq_rel();
2504 }
2505 
2506 void
2507 vm_object_unbusy(vm_object_t obj)
2508 {
2509 
2510 	blockcount_release(&obj->busy, 1);
2511 }
2512 
2513 void
2514 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2515 {
2516 
2517 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2518 
2519 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2520 }
2521 
2522 /*
2523  * This function aims to determine if the object is mapped,
2524  * specifically, if it is referenced by a vm_map_entry.  Because
2525  * objects occasionally acquire transient references that do not
2526  * represent a mapping, the method used here is inexact.  However, it
2527  * has very low overhead and is good enough for the advisory
2528  * vm.vmtotal sysctl.
2529  */
2530 bool
2531 vm_object_is_active(vm_object_t obj)
2532 {
2533 
2534 	return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2535 }
2536 
2537 static int
2538 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2539 {
2540 	struct kinfo_vmobject *kvo;
2541 	char *fullpath, *freepath;
2542 	struct vnode *vp;
2543 	struct vattr va;
2544 	vm_object_t obj;
2545 	vm_page_t m;
2546 	u_long sp;
2547 	int count, error;
2548 
2549 	if (req->oldptr == NULL) {
2550 		/*
2551 		 * If an old buffer has not been provided, generate an
2552 		 * estimate of the space needed for a subsequent call.
2553 		 */
2554 		mtx_lock(&vm_object_list_mtx);
2555 		count = 0;
2556 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2557 			if (obj->type == OBJT_DEAD)
2558 				continue;
2559 			count++;
2560 		}
2561 		mtx_unlock(&vm_object_list_mtx);
2562 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2563 		    count * 11 / 10));
2564 	}
2565 
2566 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK);
2567 	error = 0;
2568 
2569 	/*
2570 	 * VM objects are type stable and are never removed from the
2571 	 * list once added.  This allows us to safely read obj->object_list
2572 	 * after reacquiring the VM object lock.
2573 	 */
2574 	mtx_lock(&vm_object_list_mtx);
2575 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2576 		if (obj->type == OBJT_DEAD ||
2577 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2578 			continue;
2579 		VM_OBJECT_RLOCK(obj);
2580 		if (obj->type == OBJT_DEAD ||
2581 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2582 			VM_OBJECT_RUNLOCK(obj);
2583 			continue;
2584 		}
2585 		mtx_unlock(&vm_object_list_mtx);
2586 		kvo->kvo_size = ptoa(obj->size);
2587 		kvo->kvo_resident = obj->resident_page_count;
2588 		kvo->kvo_ref_count = obj->ref_count;
2589 		kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2590 		kvo->kvo_memattr = obj->memattr;
2591 		kvo->kvo_active = 0;
2592 		kvo->kvo_inactive = 0;
2593 		if (!swap_only) {
2594 			TAILQ_FOREACH(m, &obj->memq, listq) {
2595 				/*
2596 				 * A page may belong to the object but be
2597 				 * dequeued and set to PQ_NONE while the
2598 				 * object lock is not held.  This makes the
2599 				 * reads of m->queue below racy, and we do not
2600 				 * count pages set to PQ_NONE.  However, this
2601 				 * sysctl is only meant to give an
2602 				 * approximation of the system anyway.
2603 				 */
2604 				if (m->a.queue == PQ_ACTIVE)
2605 					kvo->kvo_active++;
2606 				else if (m->a.queue == PQ_INACTIVE)
2607 					kvo->kvo_inactive++;
2608 			}
2609 		}
2610 
2611 		kvo->kvo_vn_fileid = 0;
2612 		kvo->kvo_vn_fsid = 0;
2613 		kvo->kvo_vn_fsid_freebsd11 = 0;
2614 		freepath = NULL;
2615 		fullpath = "";
2616 		vp = NULL;
2617 		kvo->kvo_type = vm_object_kvme_type(obj, swap_only ? NULL : &vp);
2618 		if (vp != NULL) {
2619 			vref(vp);
2620 		} else if ((obj->flags & OBJ_ANON) != 0) {
2621 			MPASS(kvo->kvo_type == KVME_TYPE_DEFAULT ||
2622 			    kvo->kvo_type == KVME_TYPE_SWAP);
2623 			kvo->kvo_me = (uintptr_t)obj;
2624 			/* tmpfs objs are reported as vnodes */
2625 			kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2626 			sp = swap_pager_swapped_pages(obj);
2627 			kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2628 		}
2629 		VM_OBJECT_RUNLOCK(obj);
2630 		if (vp != NULL) {
2631 			vn_fullpath(vp, &fullpath, &freepath);
2632 			vn_lock(vp, LK_SHARED | LK_RETRY);
2633 			if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2634 				kvo->kvo_vn_fileid = va.va_fileid;
2635 				kvo->kvo_vn_fsid = va.va_fsid;
2636 				kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2637 								/* truncate */
2638 			}
2639 			vput(vp);
2640 		}
2641 
2642 		strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2643 		if (freepath != NULL)
2644 			free(freepath, M_TEMP);
2645 
2646 		/* Pack record size down */
2647 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2648 		    + strlen(kvo->kvo_path) + 1;
2649 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2650 		    sizeof(uint64_t));
2651 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2652 		maybe_yield();
2653 		mtx_lock(&vm_object_list_mtx);
2654 		if (error)
2655 			break;
2656 	}
2657 	mtx_unlock(&vm_object_list_mtx);
2658 	free(kvo, M_TEMP);
2659 	return (error);
2660 }
2661 
2662 static int
2663 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2664 {
2665 	return (vm_object_list_handler(req, false));
2666 }
2667 
2668 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2669     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2670     "List of VM objects");
2671 
2672 static int
2673 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2674 {
2675 	return (vm_object_list_handler(req, true));
2676 }
2677 
2678 /*
2679  * This sysctl returns list of the anonymous or swap objects. Intent
2680  * is to provide stripped optimized list useful to analyze swap use.
2681  * Since technically non-swap (default) objects participate in the
2682  * shadow chains, and are converted to swap type as needed by swap
2683  * pager, we must report them.
2684  */
2685 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2686     CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2687     sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2688     "List of swap VM objects");
2689 
2690 #include "opt_ddb.h"
2691 #ifdef DDB
2692 #include <sys/kernel.h>
2693 
2694 #include <sys/cons.h>
2695 
2696 #include <ddb/ddb.h>
2697 
2698 static int
2699 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2700 {
2701 	vm_map_t tmpm;
2702 	vm_map_entry_t tmpe;
2703 	vm_object_t obj;
2704 
2705 	if (map == 0)
2706 		return 0;
2707 
2708 	if (entry == 0) {
2709 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2710 			if (_vm_object_in_map(map, object, tmpe)) {
2711 				return 1;
2712 			}
2713 		}
2714 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2715 		tmpm = entry->object.sub_map;
2716 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2717 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2718 				return 1;
2719 			}
2720 		}
2721 	} else if ((obj = entry->object.vm_object) != NULL) {
2722 		for (; obj; obj = obj->backing_object)
2723 			if (obj == object) {
2724 				return 1;
2725 			}
2726 	}
2727 	return 0;
2728 }
2729 
2730 static int
2731 vm_object_in_map(vm_object_t object)
2732 {
2733 	struct proc *p;
2734 
2735 	/* sx_slock(&allproc_lock); */
2736 	FOREACH_PROC_IN_SYSTEM(p) {
2737 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2738 			continue;
2739 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2740 			/* sx_sunlock(&allproc_lock); */
2741 			return 1;
2742 		}
2743 	}
2744 	/* sx_sunlock(&allproc_lock); */
2745 	if (_vm_object_in_map(kernel_map, object, 0))
2746 		return 1;
2747 	return 0;
2748 }
2749 
2750 DB_SHOW_COMMAND(vmochk, vm_object_check)
2751 {
2752 	vm_object_t object;
2753 
2754 	/*
2755 	 * make sure that internal objs are in a map somewhere
2756 	 * and none have zero ref counts.
2757 	 */
2758 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2759 		if ((object->flags & OBJ_ANON) != 0) {
2760 			if (object->ref_count == 0) {
2761 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2762 					(long)object->size);
2763 			}
2764 			if (!vm_object_in_map(object)) {
2765 				db_printf(
2766 			"vmochk: internal obj is not in a map: "
2767 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2768 				    object->ref_count, (u_long)object->size,
2769 				    (u_long)object->size,
2770 				    (void *)object->backing_object);
2771 			}
2772 		}
2773 		if (db_pager_quit)
2774 			return;
2775 	}
2776 }
2777 
2778 /*
2779  *	vm_object_print:	[ debug ]
2780  */
2781 DB_SHOW_COMMAND(object, vm_object_print_static)
2782 {
2783 	/* XXX convert args. */
2784 	vm_object_t object = (vm_object_t)addr;
2785 	boolean_t full = have_addr;
2786 
2787 	vm_page_t p;
2788 
2789 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2790 #define	count	was_count
2791 
2792 	int count;
2793 
2794 	if (object == NULL)
2795 		return;
2796 
2797 	db_iprintf(
2798 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2799 	    object, (int)object->type, (uintmax_t)object->size,
2800 	    object->resident_page_count, object->ref_count, object->flags,
2801 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2802 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2803 	    atomic_load_int(&object->shadow_count),
2804 	    object->backing_object ? object->backing_object->ref_count : 0,
2805 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2806 
2807 	if (!full)
2808 		return;
2809 
2810 	db_indent += 2;
2811 	count = 0;
2812 	TAILQ_FOREACH(p, &object->memq, listq) {
2813 		if (count == 0)
2814 			db_iprintf("memory:=");
2815 		else if (count == 6) {
2816 			db_printf("\n");
2817 			db_iprintf(" ...");
2818 			count = 0;
2819 		} else
2820 			db_printf(",");
2821 		count++;
2822 
2823 		db_printf("(off=0x%jx,page=0x%jx)",
2824 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2825 
2826 		if (db_pager_quit)
2827 			break;
2828 	}
2829 	if (count != 0)
2830 		db_printf("\n");
2831 	db_indent -= 2;
2832 }
2833 
2834 /* XXX. */
2835 #undef count
2836 
2837 /* XXX need this non-static entry for calling from vm_map_print. */
2838 void
2839 vm_object_print(
2840         /* db_expr_t */ long addr,
2841 	boolean_t have_addr,
2842 	/* db_expr_t */ long count,
2843 	char *modif)
2844 {
2845 	vm_object_print_static(addr, have_addr, count, modif);
2846 }
2847 
2848 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2849 {
2850 	vm_object_t object;
2851 	vm_pindex_t fidx;
2852 	vm_paddr_t pa;
2853 	vm_page_t m, prev_m;
2854 	int rcount;
2855 
2856 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2857 		db_printf("new object: %p\n", (void *)object);
2858 		if (db_pager_quit)
2859 			return;
2860 
2861 		rcount = 0;
2862 		fidx = 0;
2863 		pa = -1;
2864 		TAILQ_FOREACH(m, &object->memq, listq) {
2865 			if (m->pindex > 128)
2866 				break;
2867 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2868 			    prev_m->pindex + 1 != m->pindex) {
2869 				if (rcount) {
2870 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2871 						(long)fidx, rcount, (long)pa);
2872 					if (db_pager_quit)
2873 						return;
2874 					rcount = 0;
2875 				}
2876 			}
2877 			if (rcount &&
2878 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2879 				++rcount;
2880 				continue;
2881 			}
2882 			if (rcount) {
2883 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2884 					(long)fidx, rcount, (long)pa);
2885 				if (db_pager_quit)
2886 					return;
2887 			}
2888 			fidx = m->pindex;
2889 			pa = VM_PAGE_TO_PHYS(m);
2890 			rcount = 1;
2891 		}
2892 		if (rcount) {
2893 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2894 				(long)fidx, rcount, (long)pa);
2895 			if (db_pager_quit)
2896 				return;
2897 		}
2898 	}
2899 }
2900 #endif /* DDB */
2901