xref: /freebsd/sys/vm/vm_object.c (revision 6137b5f7b8c183ee8806d79b3f1d8e5e3ddb3df3)
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  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39  *
40  * Permission to use, copy, modify and distribute this software and
41  * its documentation is hereby granted, provided that both the copyright
42  * notice and this permission notice appear in all copies of the
43  * software, derivative works or modified versions, and any portions
44  * thereof, and that both notices appear in supporting documentation.
45  *
46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49  *
50  * Carnegie Mellon requests users of this software to return to
51  *
52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
53  *  School of Computer Science
54  *  Carnegie Mellon University
55  *  Pittsburgh PA 15213-3890
56  *
57  * any improvements or extensions that they make and grant Carnegie the
58  * rights to redistribute these changes.
59  */
60 
61 /*
62  *	Virtual memory object module.
63  */
64 
65 #include "opt_vm.h"
66 
67 #include <sys/systm.h>
68 #include <sys/blockcount.h>
69 #include <sys/cpuset.h>
70 #include <sys/limits.h>
71 #include <sys/lock.h>
72 #include <sys/mman.h>
73 #include <sys/mount.h>
74 #include <sys/kernel.h>
75 #include <sys/mutex.h>
76 #include <sys/pctrie.h>
77 #include <sys/proc.h>
78 #include <sys/refcount.h>
79 #include <sys/sx.h>
80 #include <sys/sysctl.h>
81 #include <sys/resourcevar.h>
82 #include <sys/refcount.h>
83 #include <sys/rwlock.h>
84 #include <sys/user.h>
85 #include <sys/vnode.h>
86 #include <sys/vmmeter.h>
87 
88 #include <vm/vm.h>
89 #include <vm/vm_param.h>
90 #include <vm/pmap.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_pageout.h>
95 #include <vm/vm_pager.h>
96 #include <vm/vm_phys.h>
97 #include <vm/vm_pagequeue.h>
98 #include <vm/swap_pager.h>
99 #include <vm/vm_kern.h>
100 #include <vm/vm_extern.h>
101 #include <vm/vm_radix.h>
102 #include <vm/vm_reserv.h>
103 #include <vm/uma.h>
104 
105 static int old_msync;
106 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
107     "Use old (insecure) msync behavior");
108 
109 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
110 		    int pagerflags, int flags, boolean_t *allclean,
111 		    boolean_t *eio);
112 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
113 		    boolean_t *allclean);
114 static void	vm_object_backing_remove(vm_object_t object);
115 
116 /*
117  *	Virtual memory objects maintain the actual data
118  *	associated with allocated virtual memory.  A given
119  *	page of memory exists within exactly one object.
120  *
121  *	An object is only deallocated when all "references"
122  *	are given up.  Only one "reference" to a given
123  *	region of an object should be writeable.
124  *
125  *	Associated with each object is a list of all resident
126  *	memory pages belonging to that object; this list is
127  *	maintained by the "vm_page" module, and locked by the object's
128  *	lock.
129  *
130  *	Each object also records a "pager" routine which is
131  *	used to retrieve (and store) pages to the proper backing
132  *	storage.  In addition, objects may be backed by other
133  *	objects from which they were virtual-copied.
134  *
135  *	The only items within the object structure which are
136  *	modified after time of creation are:
137  *		reference count		locked by object's lock
138  *		pager routine		locked by object's lock
139  *
140  */
141 
142 struct object_q vm_object_list;
143 struct mtx vm_object_list_mtx;	/* lock for object list and count */
144 
145 struct vm_object kernel_object_store;
146 
147 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
148     "VM object stats");
149 
150 static COUNTER_U64_DEFINE_EARLY(object_collapses);
151 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
152     &object_collapses,
153     "VM object collapses");
154 
155 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
156 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
157     &object_bypasses,
158     "VM object bypasses");
159 
160 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
161 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
162     &object_collapse_waits,
163     "Number of sleeps for collapse");
164 
165 static uma_zone_t obj_zone;
166 
167 static int vm_object_zinit(void *mem, int size, int flags);
168 
169 #ifdef INVARIANTS
170 static void vm_object_zdtor(void *mem, int size, void *arg);
171 
172 static void
173 vm_object_zdtor(void *mem, int size, void *arg)
174 {
175 	vm_object_t object;
176 
177 	object = (vm_object_t)mem;
178 	KASSERT(object->ref_count == 0,
179 	    ("object %p ref_count = %d", object, object->ref_count));
180 	KASSERT(TAILQ_EMPTY(&object->memq),
181 	    ("object %p has resident pages in its memq", object));
182 	KASSERT(vm_radix_is_empty(&object->rtree),
183 	    ("object %p has resident pages in its trie", object));
184 #if VM_NRESERVLEVEL > 0
185 	KASSERT(LIST_EMPTY(&object->rvq),
186 	    ("object %p has reservations",
187 	    object));
188 #endif
189 	KASSERT(!vm_object_busied(object),
190 	    ("object %p busy = %d", object, blockcount_read(&object->busy)));
191 	KASSERT(object->resident_page_count == 0,
192 	    ("object %p resident_page_count = %d",
193 	    object, object->resident_page_count));
194 	KASSERT(atomic_load_int(&object->shadow_count) == 0,
195 	    ("object %p shadow_count = %d",
196 	    object, atomic_load_int(&object->shadow_count)));
197 	KASSERT(object->type == OBJT_DEAD,
198 	    ("object %p has non-dead type %d",
199 	    object, object->type));
200 	KASSERT(object->charge == 0 && object->cred == NULL,
201 	    ("object %p has non-zero charge %ju (%p)",
202 	    object, (uintmax_t)object->charge, object->cred));
203 }
204 #endif
205 
206 static int
207 vm_object_zinit(void *mem, int size, int flags)
208 {
209 	vm_object_t object;
210 
211 	object = (vm_object_t)mem;
212 	rw_init_flags(&object->lock, "vmobject", RW_DUPOK | RW_NEW);
213 
214 	/* These are true for any object that has been freed */
215 	object->type = OBJT_DEAD;
216 	vm_radix_init(&object->rtree);
217 	refcount_init(&object->ref_count, 0);
218 	blockcount_init(&object->paging_in_progress);
219 	blockcount_init(&object->busy);
220 	object->resident_page_count = 0;
221 	atomic_store_int(&object->shadow_count, 0);
222 	object->flags = OBJ_DEAD;
223 
224 	mtx_lock(&vm_object_list_mtx);
225 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
226 	mtx_unlock(&vm_object_list_mtx);
227 	return (0);
228 }
229 
230 static void
231 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
232     vm_object_t object, void *handle)
233 {
234 
235 	TAILQ_INIT(&object->memq);
236 	LIST_INIT(&object->shadow_head);
237 
238 	object->type = type;
239 	object->flags = flags;
240 	if ((flags & OBJ_SWAP) != 0) {
241 		pctrie_init(&object->un_pager.swp.swp_blks);
242 		object->un_pager.swp.writemappings = 0;
243 	}
244 
245 	/*
246 	 * Ensure that swap_pager_swapoff() iteration over object_list
247 	 * sees up to date type and pctrie head if it observed
248 	 * non-dead object.
249 	 */
250 	atomic_thread_fence_rel();
251 
252 	object->pg_color = 0;
253 	object->size = size;
254 	object->domain.dr_policy = NULL;
255 	object->generation = 1;
256 	object->cleangeneration = 1;
257 	refcount_init(&object->ref_count, 1);
258 	object->memattr = VM_MEMATTR_DEFAULT;
259 	object->cred = NULL;
260 	object->charge = 0;
261 	object->handle = handle;
262 	object->backing_object = NULL;
263 	object->backing_object_offset = (vm_ooffset_t) 0;
264 #if VM_NRESERVLEVEL > 0
265 	LIST_INIT(&object->rvq);
266 #endif
267 	umtx_shm_object_init(object);
268 }
269 
270 /*
271  *	vm_object_init:
272  *
273  *	Initialize the VM objects module.
274  */
275 void
276 vm_object_init(void)
277 {
278 	TAILQ_INIT(&vm_object_list);
279 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
280 
281 	rw_init(&kernel_object->lock, "kernel vm object");
282 	vm_radix_init(&kernel_object->rtree);
283 	_vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
284 	    VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
285 #if VM_NRESERVLEVEL > 0
286 	kernel_object->flags |= OBJ_COLORED;
287 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
288 #endif
289 	kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
290 
291 	/*
292 	 * The lock portion of struct vm_object must be type stable due
293 	 * to vm_pageout_fallback_object_lock locking a vm object
294 	 * without holding any references to it.
295 	 *
296 	 * paging_in_progress is valid always.  Lockless references to
297 	 * the objects may acquire pip and then check OBJ_DEAD.
298 	 */
299 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
300 #ifdef INVARIANTS
301 	    vm_object_zdtor,
302 #else
303 	    NULL,
304 #endif
305 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
306 
307 	vm_radix_zinit();
308 }
309 
310 void
311 vm_object_clear_flag(vm_object_t object, u_short bits)
312 {
313 
314 	VM_OBJECT_ASSERT_WLOCKED(object);
315 	object->flags &= ~bits;
316 }
317 
318 /*
319  *	Sets the default memory attribute for the specified object.  Pages
320  *	that are allocated to this object are by default assigned this memory
321  *	attribute.
322  *
323  *	Presently, this function must be called before any pages are allocated
324  *	to the object.  In the future, this requirement may be relaxed for
325  *	"default" and "swap" objects.
326  */
327 int
328 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
329 {
330 
331 	VM_OBJECT_ASSERT_WLOCKED(object);
332 
333 	if (object->type == OBJT_DEAD)
334 		return (KERN_INVALID_ARGUMENT);
335 	if (!TAILQ_EMPTY(&object->memq))
336 		return (KERN_FAILURE);
337 
338 	object->memattr = memattr;
339 	return (KERN_SUCCESS);
340 }
341 
342 void
343 vm_object_pip_add(vm_object_t object, short i)
344 {
345 
346 	if (i > 0)
347 		blockcount_acquire(&object->paging_in_progress, i);
348 }
349 
350 void
351 vm_object_pip_wakeup(vm_object_t object)
352 {
353 
354 	vm_object_pip_wakeupn(object, 1);
355 }
356 
357 void
358 vm_object_pip_wakeupn(vm_object_t object, short i)
359 {
360 
361 	if (i > 0)
362 		blockcount_release(&object->paging_in_progress, i);
363 }
364 
365 /*
366  * Atomically drop the object lock and wait for pip to drain.  This protects
367  * from sleep/wakeup races due to identity changes.  The lock is not re-acquired
368  * on return.
369  */
370 static void
371 vm_object_pip_sleep(vm_object_t object, const char *waitid)
372 {
373 
374 	(void)blockcount_sleep(&object->paging_in_progress, &object->lock,
375 	    waitid, PVM | PDROP);
376 }
377 
378 void
379 vm_object_pip_wait(vm_object_t object, const char *waitid)
380 {
381 
382 	VM_OBJECT_ASSERT_WLOCKED(object);
383 
384 	blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
385 	    PVM);
386 }
387 
388 void
389 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
390 {
391 
392 	VM_OBJECT_ASSERT_UNLOCKED(object);
393 
394 	blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
395 }
396 
397 /*
398  *	vm_object_allocate:
399  *
400  *	Returns a new object with the given size.
401  */
402 vm_object_t
403 vm_object_allocate(objtype_t type, vm_pindex_t size)
404 {
405 	vm_object_t object;
406 	u_short flags;
407 
408 	switch (type) {
409 	case OBJT_DEAD:
410 		panic("vm_object_allocate: can't create OBJT_DEAD");
411 	case OBJT_SWAP:
412 		flags = OBJ_COLORED | OBJ_SWAP;
413 		break;
414 	case OBJT_DEVICE:
415 	case OBJT_SG:
416 		flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
417 		break;
418 	case OBJT_MGTDEVICE:
419 		flags = OBJ_FICTITIOUS;
420 		break;
421 	case OBJT_PHYS:
422 		flags = OBJ_UNMANAGED;
423 		break;
424 	case OBJT_VNODE:
425 		flags = 0;
426 		break;
427 	default:
428 		panic("vm_object_allocate: type %d is undefined or dynamic",
429 		    type);
430 	}
431 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
432 	_vm_object_allocate(type, size, flags, object, NULL);
433 
434 	return (object);
435 }
436 
437 vm_object_t
438 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags)
439 {
440 	vm_object_t object;
441 
442 	MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */);
443 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
444 	_vm_object_allocate(dyntype, size, flags, object, NULL);
445 
446 	return (object);
447 }
448 
449 /*
450  *	vm_object_allocate_anon:
451  *
452  *	Returns a new default object of the given size and marked as
453  *	anonymous memory for special split/collapse handling.  Color
454  *	to be initialized by the caller.
455  */
456 vm_object_t
457 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
458     struct ucred *cred, vm_size_t charge)
459 {
460 	vm_object_t handle, object;
461 
462 	if (backing_object == NULL)
463 		handle = NULL;
464 	else if ((backing_object->flags & OBJ_ANON) != 0)
465 		handle = backing_object->handle;
466 	else
467 		handle = backing_object;
468 	object = uma_zalloc(obj_zone, M_WAITOK);
469 	_vm_object_allocate(OBJT_SWAP, size,
470 	    OBJ_ANON | OBJ_ONEMAPPING | OBJ_SWAP, object, handle);
471 	object->cred = cred;
472 	object->charge = cred != NULL ? charge : 0;
473 	return (object);
474 }
475 
476 static void
477 vm_object_reference_vnode(vm_object_t object)
478 {
479 	u_int old;
480 
481 	/*
482 	 * vnode objects need the lock for the first reference
483 	 * to serialize with vnode_object_deallocate().
484 	 */
485 	if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
486 		VM_OBJECT_RLOCK(object);
487 		old = refcount_acquire(&object->ref_count);
488 		if (object->type == OBJT_VNODE && old == 0)
489 			vref(object->handle);
490 		VM_OBJECT_RUNLOCK(object);
491 	}
492 }
493 
494 /*
495  *	vm_object_reference:
496  *
497  *	Acquires a reference to the given object.
498  */
499 void
500 vm_object_reference(vm_object_t object)
501 {
502 
503 	if (object == NULL)
504 		return;
505 
506 	if (object->type == OBJT_VNODE)
507 		vm_object_reference_vnode(object);
508 	else
509 		refcount_acquire(&object->ref_count);
510 	KASSERT((object->flags & OBJ_DEAD) == 0,
511 	    ("vm_object_reference: Referenced dead object."));
512 }
513 
514 /*
515  *	vm_object_reference_locked:
516  *
517  *	Gets another reference to the given object.
518  *
519  *	The object must be locked.
520  */
521 void
522 vm_object_reference_locked(vm_object_t object)
523 {
524 	u_int old;
525 
526 	VM_OBJECT_ASSERT_LOCKED(object);
527 	old = refcount_acquire(&object->ref_count);
528 	if (object->type == OBJT_VNODE && old == 0)
529 		vref(object->handle);
530 	KASSERT((object->flags & OBJ_DEAD) == 0,
531 	    ("vm_object_reference: Referenced dead object."));
532 }
533 
534 /*
535  * Handle deallocating an object of type OBJT_VNODE.
536  */
537 static void
538 vm_object_deallocate_vnode(vm_object_t object)
539 {
540 	struct vnode *vp = (struct vnode *) object->handle;
541 	bool last;
542 
543 	KASSERT(object->type == OBJT_VNODE,
544 	    ("vm_object_deallocate_vnode: not a vnode object"));
545 	KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
546 
547 	/* Object lock to protect handle lookup. */
548 	last = refcount_release(&object->ref_count);
549 	VM_OBJECT_RUNLOCK(object);
550 
551 	if (!last)
552 		return;
553 
554 	if (!umtx_shm_vnobj_persistent)
555 		umtx_shm_object_terminated(object);
556 
557 	/* vrele may need the vnode lock. */
558 	vrele(vp);
559 }
560 
561 /*
562  * We dropped a reference on an object and discovered that it had a
563  * single remaining shadow.  This is a sibling of the reference we
564  * dropped.  Attempt to collapse the sibling and backing object.
565  */
566 static vm_object_t
567 vm_object_deallocate_anon(vm_object_t backing_object)
568 {
569 	vm_object_t object;
570 
571 	/* Fetch the final shadow.  */
572 	object = LIST_FIRST(&backing_object->shadow_head);
573 	KASSERT(object != NULL &&
574 	    atomic_load_int(&backing_object->shadow_count) == 1,
575 	    ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
576 	    backing_object->ref_count,
577 	    atomic_load_int(&backing_object->shadow_count)));
578 	KASSERT((object->flags & OBJ_ANON) != 0,
579 	    ("invalid shadow object %p", object));
580 
581 	if (!VM_OBJECT_TRYWLOCK(object)) {
582 		/*
583 		 * Prevent object from disappearing since we do not have a
584 		 * reference.
585 		 */
586 		vm_object_pip_add(object, 1);
587 		VM_OBJECT_WUNLOCK(backing_object);
588 		VM_OBJECT_WLOCK(object);
589 		vm_object_pip_wakeup(object);
590 	} else
591 		VM_OBJECT_WUNLOCK(backing_object);
592 
593 	/*
594 	 * Check for a collapse/terminate race with the last reference holder.
595 	 */
596 	if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
597 	    !refcount_acquire_if_not_zero(&object->ref_count)) {
598 		VM_OBJECT_WUNLOCK(object);
599 		return (NULL);
600 	}
601 	backing_object = object->backing_object;
602 	if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
603 		vm_object_collapse(object);
604 	VM_OBJECT_WUNLOCK(object);
605 
606 	return (object);
607 }
608 
609 /*
610  *	vm_object_deallocate:
611  *
612  *	Release a reference to the specified object,
613  *	gained either through a vm_object_allocate
614  *	or a vm_object_reference call.  When all references
615  *	are gone, storage associated with this object
616  *	may be relinquished.
617  *
618  *	No object may be locked.
619  */
620 void
621 vm_object_deallocate(vm_object_t object)
622 {
623 	vm_object_t temp;
624 	bool released;
625 
626 	while (object != NULL) {
627 		/*
628 		 * If the reference count goes to 0 we start calling
629 		 * vm_object_terminate() on the object chain.  A ref count
630 		 * of 1 may be a special case depending on the shadow count
631 		 * being 0 or 1.  These cases require a write lock on the
632 		 * object.
633 		 */
634 		if ((object->flags & OBJ_ANON) == 0)
635 			released = refcount_release_if_gt(&object->ref_count, 1);
636 		else
637 			released = refcount_release_if_gt(&object->ref_count, 2);
638 		if (released)
639 			return;
640 
641 		if (object->type == OBJT_VNODE) {
642 			VM_OBJECT_RLOCK(object);
643 			if (object->type == OBJT_VNODE) {
644 				vm_object_deallocate_vnode(object);
645 				return;
646 			}
647 			VM_OBJECT_RUNLOCK(object);
648 		}
649 
650 		VM_OBJECT_WLOCK(object);
651 		KASSERT(object->ref_count > 0,
652 		    ("vm_object_deallocate: object deallocated too many times: %d",
653 		    object->type));
654 
655 		/*
656 		 * If this is not the final reference to an anonymous
657 		 * object we may need to collapse the shadow chain.
658 		 */
659 		if (!refcount_release(&object->ref_count)) {
660 			if (object->ref_count > 1 ||
661 			    atomic_load_int(&object->shadow_count) == 0) {
662 				if ((object->flags & OBJ_ANON) != 0 &&
663 				    object->ref_count == 1)
664 					vm_object_set_flag(object,
665 					    OBJ_ONEMAPPING);
666 				VM_OBJECT_WUNLOCK(object);
667 				return;
668 			}
669 
670 			/* Handle collapsing last ref on anonymous objects. */
671 			object = vm_object_deallocate_anon(object);
672 			continue;
673 		}
674 
675 		/*
676 		 * Handle the final reference to an object.  We restart
677 		 * the loop with the backing object to avoid recursion.
678 		 */
679 		umtx_shm_object_terminated(object);
680 		temp = object->backing_object;
681 		if (temp != NULL) {
682 			KASSERT(object->type == OBJT_SWAP,
683 			    ("shadowed tmpfs v_object 2 %p", object));
684 			vm_object_backing_remove(object);
685 		}
686 
687 		KASSERT((object->flags & OBJ_DEAD) == 0,
688 		    ("vm_object_deallocate: Terminating dead object."));
689 		vm_object_set_flag(object, OBJ_DEAD);
690 		vm_object_terminate(object);
691 		object = temp;
692 	}
693 }
694 
695 void
696 vm_object_destroy(vm_object_t object)
697 {
698 	uma_zfree(obj_zone, object);
699 }
700 
701 static void
702 vm_object_sub_shadow(vm_object_t object)
703 {
704 	KASSERT(object->shadow_count >= 1,
705 	    ("object %p sub_shadow count zero", object));
706 	atomic_subtract_int(&object->shadow_count, 1);
707 }
708 
709 static void
710 vm_object_backing_remove_locked(vm_object_t object)
711 {
712 	vm_object_t backing_object;
713 
714 	backing_object = object->backing_object;
715 	VM_OBJECT_ASSERT_WLOCKED(object);
716 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
717 
718 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
719 	    ("vm_object_backing_remove: Removing collapsing object."));
720 
721 	vm_object_sub_shadow(backing_object);
722 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
723 		LIST_REMOVE(object, shadow_list);
724 		vm_object_clear_flag(object, OBJ_SHADOWLIST);
725 	}
726 	object->backing_object = NULL;
727 }
728 
729 static void
730 vm_object_backing_remove(vm_object_t object)
731 {
732 	vm_object_t backing_object;
733 
734 	VM_OBJECT_ASSERT_WLOCKED(object);
735 
736 	backing_object = object->backing_object;
737 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
738 		VM_OBJECT_WLOCK(backing_object);
739 		vm_object_backing_remove_locked(object);
740 		VM_OBJECT_WUNLOCK(backing_object);
741 	} else {
742 		object->backing_object = NULL;
743 		vm_object_sub_shadow(backing_object);
744 	}
745 }
746 
747 static void
748 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
749 {
750 
751 	VM_OBJECT_ASSERT_WLOCKED(object);
752 
753 	atomic_add_int(&backing_object->shadow_count, 1);
754 	if ((backing_object->flags & OBJ_ANON) != 0) {
755 		VM_OBJECT_ASSERT_WLOCKED(backing_object);
756 		LIST_INSERT_HEAD(&backing_object->shadow_head, object,
757 		    shadow_list);
758 		vm_object_set_flag(object, OBJ_SHADOWLIST);
759 	}
760 	object->backing_object = backing_object;
761 }
762 
763 static void
764 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
765 {
766 
767 	VM_OBJECT_ASSERT_WLOCKED(object);
768 
769 	if ((backing_object->flags & OBJ_ANON) != 0) {
770 		VM_OBJECT_WLOCK(backing_object);
771 		vm_object_backing_insert_locked(object, backing_object);
772 		VM_OBJECT_WUNLOCK(backing_object);
773 	} else {
774 		object->backing_object = backing_object;
775 		atomic_add_int(&backing_object->shadow_count, 1);
776 	}
777 }
778 
779 /*
780  * Insert an object into a backing_object's shadow list with an additional
781  * reference to the backing_object added.
782  */
783 static void
784 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
785 {
786 
787 	VM_OBJECT_ASSERT_WLOCKED(object);
788 
789 	if ((backing_object->flags & OBJ_ANON) != 0) {
790 		VM_OBJECT_WLOCK(backing_object);
791 		KASSERT((backing_object->flags & OBJ_DEAD) == 0,
792 		    ("shadowing dead anonymous object"));
793 		vm_object_reference_locked(backing_object);
794 		vm_object_backing_insert_locked(object, backing_object);
795 		vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
796 		VM_OBJECT_WUNLOCK(backing_object);
797 	} else {
798 		vm_object_reference(backing_object);
799 		atomic_add_int(&backing_object->shadow_count, 1);
800 		object->backing_object = backing_object;
801 	}
802 }
803 
804 /*
805  * Transfer a backing reference from backing_object to object.
806  */
807 static void
808 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
809 {
810 	vm_object_t new_backing_object;
811 
812 	/*
813 	 * Note that the reference to backing_object->backing_object
814 	 * moves from within backing_object to within object.
815 	 */
816 	vm_object_backing_remove_locked(object);
817 	new_backing_object = backing_object->backing_object;
818 	if (new_backing_object == NULL)
819 		return;
820 	if ((new_backing_object->flags & OBJ_ANON) != 0) {
821 		VM_OBJECT_WLOCK(new_backing_object);
822 		vm_object_backing_remove_locked(backing_object);
823 		vm_object_backing_insert_locked(object, new_backing_object);
824 		VM_OBJECT_WUNLOCK(new_backing_object);
825 	} else {
826 		/*
827 		 * shadow_count for new_backing_object is left
828 		 * unchanged, its reference provided by backing_object
829 		 * is replaced by object.
830 		 */
831 		object->backing_object = new_backing_object;
832 		backing_object->backing_object = NULL;
833 	}
834 }
835 
836 /*
837  * Wait for a concurrent collapse to settle.
838  */
839 static void
840 vm_object_collapse_wait(vm_object_t object)
841 {
842 
843 	VM_OBJECT_ASSERT_WLOCKED(object);
844 
845 	while ((object->flags & OBJ_COLLAPSING) != 0) {
846 		vm_object_pip_wait(object, "vmcolwait");
847 		counter_u64_add(object_collapse_waits, 1);
848 	}
849 }
850 
851 /*
852  * Waits for a backing object to clear a pending collapse and returns
853  * it locked if it is an ANON object.
854  */
855 static vm_object_t
856 vm_object_backing_collapse_wait(vm_object_t object)
857 {
858 	vm_object_t backing_object;
859 
860 	VM_OBJECT_ASSERT_WLOCKED(object);
861 
862 	for (;;) {
863 		backing_object = object->backing_object;
864 		if (backing_object == NULL ||
865 		    (backing_object->flags & OBJ_ANON) == 0)
866 			return (NULL);
867 		VM_OBJECT_WLOCK(backing_object);
868 		if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
869 			break;
870 		VM_OBJECT_WUNLOCK(object);
871 		vm_object_pip_sleep(backing_object, "vmbckwait");
872 		counter_u64_add(object_collapse_waits, 1);
873 		VM_OBJECT_WLOCK(object);
874 	}
875 	return (backing_object);
876 }
877 
878 /*
879  *	vm_object_terminate_pages removes any remaining pageable pages
880  *	from the object and resets the object to an empty state.
881  */
882 static void
883 vm_object_terminate_pages(vm_object_t object)
884 {
885 	vm_page_t p, p_next;
886 
887 	VM_OBJECT_ASSERT_WLOCKED(object);
888 
889 	/*
890 	 * Free any remaining pageable pages.  This also removes them from the
891 	 * paging queues.  However, don't free wired pages, just remove them
892 	 * from the object.  Rather than incrementally removing each page from
893 	 * the object, the page and object are reset to any empty state.
894 	 */
895 	TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
896 		vm_page_assert_unbusied(p);
897 		KASSERT(p->object == object &&
898 		    (p->ref_count & VPRC_OBJREF) != 0,
899 		    ("vm_object_terminate_pages: page %p is inconsistent", p));
900 
901 		p->object = NULL;
902 		if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
903 			VM_CNT_INC(v_pfree);
904 			vm_page_free(p);
905 		}
906 	}
907 
908 	/*
909 	 * If the object contained any pages, then reset it to an empty state.
910 	 * None of the object's fields, including "resident_page_count", were
911 	 * modified by the preceding loop.
912 	 */
913 	if (object->resident_page_count != 0) {
914 		vm_radix_reclaim_allnodes(&object->rtree);
915 		TAILQ_INIT(&object->memq);
916 		object->resident_page_count = 0;
917 		if (object->type == OBJT_VNODE)
918 			vdrop(object->handle);
919 	}
920 }
921 
922 /*
923  *	vm_object_terminate actually destroys the specified object, freeing
924  *	up all previously used resources.
925  *
926  *	The object must be locked.
927  *	This routine may block.
928  */
929 void
930 vm_object_terminate(vm_object_t object)
931 {
932 
933 	VM_OBJECT_ASSERT_WLOCKED(object);
934 	KASSERT((object->flags & OBJ_DEAD) != 0,
935 	    ("terminating non-dead obj %p", object));
936 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
937 	    ("terminating collapsing obj %p", object));
938 	KASSERT(object->backing_object == NULL,
939 	    ("terminating shadow obj %p", object));
940 
941 	/*
942 	 * Wait for the pageout daemon and other current users to be
943 	 * done with the object.  Note that new paging_in_progress
944 	 * users can come after this wait, but they must check
945 	 * OBJ_DEAD flag set (without unlocking the object), and avoid
946 	 * the object being terminated.
947 	 */
948 	vm_object_pip_wait(object, "objtrm");
949 
950 	KASSERT(object->ref_count == 0,
951 	    ("vm_object_terminate: object with references, ref_count=%d",
952 	    object->ref_count));
953 
954 	if ((object->flags & OBJ_PG_DTOR) == 0)
955 		vm_object_terminate_pages(object);
956 
957 #if VM_NRESERVLEVEL > 0
958 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
959 		vm_reserv_break_all(object);
960 #endif
961 
962 	KASSERT(object->cred == NULL || (object->flags & OBJ_SWAP) != 0,
963 	    ("%s: non-swap obj %p has cred", __func__, object));
964 
965 	/*
966 	 * Let the pager know object is dead.
967 	 */
968 	vm_pager_deallocate(object);
969 	VM_OBJECT_WUNLOCK(object);
970 
971 	vm_object_destroy(object);
972 }
973 
974 /*
975  * Make the page read-only so that we can clear the object flags.  However, if
976  * this is a nosync mmap then the object is likely to stay dirty so do not
977  * mess with the page and do not clear the object flags.  Returns TRUE if the
978  * page should be flushed, and FALSE otherwise.
979  */
980 static boolean_t
981 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
982 {
983 
984 	vm_page_assert_busied(p);
985 
986 	/*
987 	 * If we have been asked to skip nosync pages and this is a
988 	 * nosync page, skip it.  Note that the object flags were not
989 	 * cleared in this case so we do not have to set them.
990 	 */
991 	if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
992 		*allclean = FALSE;
993 		return (FALSE);
994 	} else {
995 		pmap_remove_write(p);
996 		return (p->dirty != 0);
997 	}
998 }
999 
1000 /*
1001  *	vm_object_page_clean
1002  *
1003  *	Clean all dirty pages in the specified range of object.  Leaves page
1004  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
1005  *	write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1006  *	leaving the object dirty.
1007  *
1008  *	For swap objects backing tmpfs regular files, do not flush anything,
1009  *	but remove write protection on the mapped pages to update mtime through
1010  *	mmaped writes.
1011  *
1012  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
1013  *	synchronous clustering mode implementation.
1014  *
1015  *	Odd semantics: if start == end, we clean everything.
1016  *
1017  *	The object must be locked.
1018  *
1019  *	Returns FALSE if some page from the range was not written, as
1020  *	reported by the pager, and TRUE otherwise.
1021  */
1022 boolean_t
1023 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1024     int flags)
1025 {
1026 	vm_page_t np, p;
1027 	vm_pindex_t pi, tend, tstart;
1028 	int curgeneration, n, pagerflags;
1029 	boolean_t eio, res, allclean;
1030 
1031 	VM_OBJECT_ASSERT_WLOCKED(object);
1032 
1033 	if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1034 		return (TRUE);
1035 
1036 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1037 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1038 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1039 
1040 	tstart = OFF_TO_IDX(start);
1041 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1042 	allclean = tstart == 0 && tend >= object->size;
1043 	res = TRUE;
1044 
1045 rescan:
1046 	curgeneration = object->generation;
1047 
1048 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1049 		pi = p->pindex;
1050 		if (pi >= tend)
1051 			break;
1052 		np = TAILQ_NEXT(p, listq);
1053 		if (vm_page_none_valid(p))
1054 			continue;
1055 		if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1056 			if (object->generation != curgeneration &&
1057 			    (flags & OBJPC_SYNC) != 0)
1058 				goto rescan;
1059 			np = vm_page_find_least(object, pi);
1060 			continue;
1061 		}
1062 		if (!vm_object_page_remove_write(p, flags, &allclean)) {
1063 			vm_page_xunbusy(p);
1064 			continue;
1065 		}
1066 		if (object->type == OBJT_VNODE) {
1067 			n = vm_object_page_collect_flush(object, p, pagerflags,
1068 			    flags, &allclean, &eio);
1069 			if (eio) {
1070 				res = FALSE;
1071 				allclean = FALSE;
1072 			}
1073 			if (object->generation != curgeneration &&
1074 			    (flags & OBJPC_SYNC) != 0)
1075 				goto rescan;
1076 
1077 			/*
1078 			 * If the VOP_PUTPAGES() did a truncated write, so
1079 			 * that even the first page of the run is not fully
1080 			 * written, vm_pageout_flush() returns 0 as the run
1081 			 * length.  Since the condition that caused truncated
1082 			 * write may be permanent, e.g. exhausted free space,
1083 			 * accepting n == 0 would cause an infinite loop.
1084 			 *
1085 			 * Forwarding the iterator leaves the unwritten page
1086 			 * behind, but there is not much we can do there if
1087 			 * filesystem refuses to write it.
1088 			 */
1089 			if (n == 0) {
1090 				n = 1;
1091 				allclean = FALSE;
1092 			}
1093 		} else {
1094 			n = 1;
1095 			vm_page_xunbusy(p);
1096 		}
1097 		np = vm_page_find_least(object, pi + n);
1098 	}
1099 #if 0
1100 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1101 #endif
1102 
1103 	/*
1104 	 * Leave updating cleangeneration for tmpfs objects to tmpfs
1105 	 * scan.  It needs to update mtime, which happens for other
1106 	 * filesystems during page writeouts.
1107 	 */
1108 	if (allclean && object->type == OBJT_VNODE)
1109 		object->cleangeneration = curgeneration;
1110 	return (res);
1111 }
1112 
1113 static int
1114 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1115     int flags, boolean_t *allclean, boolean_t *eio)
1116 {
1117 	vm_page_t ma[vm_pageout_page_count], p_first, tp;
1118 	int count, i, mreq, runlen;
1119 
1120 	vm_page_lock_assert(p, MA_NOTOWNED);
1121 	vm_page_assert_xbusied(p);
1122 	VM_OBJECT_ASSERT_WLOCKED(object);
1123 
1124 	count = 1;
1125 	mreq = 0;
1126 
1127 	for (tp = p; count < vm_pageout_page_count; count++) {
1128 		tp = vm_page_next(tp);
1129 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1130 			break;
1131 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1132 			vm_page_xunbusy(tp);
1133 			break;
1134 		}
1135 	}
1136 
1137 	for (p_first = p; count < vm_pageout_page_count; count++) {
1138 		tp = vm_page_prev(p_first);
1139 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1140 			break;
1141 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1142 			vm_page_xunbusy(tp);
1143 			break;
1144 		}
1145 		p_first = tp;
1146 		mreq++;
1147 	}
1148 
1149 	for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1150 		ma[i] = tp;
1151 
1152 	vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1153 	return (runlen);
1154 }
1155 
1156 /*
1157  * Note that there is absolutely no sense in writing out
1158  * anonymous objects, so we track down the vnode object
1159  * to write out.
1160  * We invalidate (remove) all pages from the address space
1161  * for semantic correctness.
1162  *
1163  * If the backing object is a device object with unmanaged pages, then any
1164  * mappings to the specified range of pages must be removed before this
1165  * function is called.
1166  *
1167  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1168  * may start out with a NULL object.
1169  */
1170 boolean_t
1171 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1172     boolean_t syncio, boolean_t invalidate)
1173 {
1174 	vm_object_t backing_object;
1175 	struct vnode *vp;
1176 	struct mount *mp;
1177 	int error, flags, fsync_after;
1178 	boolean_t res;
1179 
1180 	if (object == NULL)
1181 		return (TRUE);
1182 	res = TRUE;
1183 	error = 0;
1184 	VM_OBJECT_WLOCK(object);
1185 	while ((backing_object = object->backing_object) != NULL) {
1186 		VM_OBJECT_WLOCK(backing_object);
1187 		offset += object->backing_object_offset;
1188 		VM_OBJECT_WUNLOCK(object);
1189 		object = backing_object;
1190 		if (object->size < OFF_TO_IDX(offset + size))
1191 			size = IDX_TO_OFF(object->size) - offset;
1192 	}
1193 	/*
1194 	 * Flush pages if writing is allowed, invalidate them
1195 	 * if invalidation requested.  Pages undergoing I/O
1196 	 * will be ignored by vm_object_page_remove().
1197 	 *
1198 	 * We cannot lock the vnode and then wait for paging
1199 	 * to complete without deadlocking against vm_fault.
1200 	 * Instead we simply call vm_object_page_remove() and
1201 	 * allow it to block internally on a page-by-page
1202 	 * basis when it encounters pages undergoing async
1203 	 * I/O.
1204 	 */
1205 	if (object->type == OBJT_VNODE &&
1206 	    vm_object_mightbedirty(object) != 0 &&
1207 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1208 		VM_OBJECT_WUNLOCK(object);
1209 		(void)vn_start_write(vp, &mp, V_WAIT);
1210 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1211 		if (syncio && !invalidate && offset == 0 &&
1212 		    atop(size) == object->size) {
1213 			/*
1214 			 * If syncing the whole mapping of the file,
1215 			 * it is faster to schedule all the writes in
1216 			 * async mode, also allowing the clustering,
1217 			 * and then wait for i/o to complete.
1218 			 */
1219 			flags = 0;
1220 			fsync_after = TRUE;
1221 		} else {
1222 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1223 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1224 			fsync_after = FALSE;
1225 		}
1226 		VM_OBJECT_WLOCK(object);
1227 		res = vm_object_page_clean(object, offset, offset + size,
1228 		    flags);
1229 		VM_OBJECT_WUNLOCK(object);
1230 		if (fsync_after) {
1231 			for (;;) {
1232 				error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1233 				if (error != ERELOOKUP)
1234 					break;
1235 
1236 				/*
1237 				 * Allow SU/bufdaemon to handle more
1238 				 * dependencies in the meantime.
1239 				 */
1240 				VOP_UNLOCK(vp);
1241 				vn_finished_write(mp);
1242 
1243 				(void)vn_start_write(vp, &mp, V_WAIT);
1244 				vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1245 			}
1246 		}
1247 		VOP_UNLOCK(vp);
1248 		vn_finished_write(mp);
1249 		if (error != 0)
1250 			res = FALSE;
1251 		VM_OBJECT_WLOCK(object);
1252 	}
1253 	if ((object->type == OBJT_VNODE ||
1254 	     object->type == OBJT_DEVICE) && invalidate) {
1255 		if (object->type == OBJT_DEVICE)
1256 			/*
1257 			 * The option OBJPR_NOTMAPPED must be passed here
1258 			 * because vm_object_page_remove() cannot remove
1259 			 * unmanaged mappings.
1260 			 */
1261 			flags = OBJPR_NOTMAPPED;
1262 		else if (old_msync)
1263 			flags = 0;
1264 		else
1265 			flags = OBJPR_CLEANONLY;
1266 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1267 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1268 	}
1269 	VM_OBJECT_WUNLOCK(object);
1270 	return (res);
1271 }
1272 
1273 /*
1274  * Determine whether the given advice can be applied to the object.  Advice is
1275  * not applied to unmanaged pages since they never belong to page queues, and
1276  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1277  * have been mapped at most once.
1278  */
1279 static bool
1280 vm_object_advice_applies(vm_object_t object, int advice)
1281 {
1282 
1283 	if ((object->flags & OBJ_UNMANAGED) != 0)
1284 		return (false);
1285 	if (advice != MADV_FREE)
1286 		return (true);
1287 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1288 	    (OBJ_ONEMAPPING | OBJ_ANON));
1289 }
1290 
1291 static void
1292 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1293     vm_size_t size)
1294 {
1295 
1296 	if (advice == MADV_FREE)
1297 		vm_pager_freespace(object, pindex, size);
1298 }
1299 
1300 /*
1301  *	vm_object_madvise:
1302  *
1303  *	Implements the madvise function at the object/page level.
1304  *
1305  *	MADV_WILLNEED	(any object)
1306  *
1307  *	    Activate the specified pages if they are resident.
1308  *
1309  *	MADV_DONTNEED	(any object)
1310  *
1311  *	    Deactivate the specified pages if they are resident.
1312  *
1313  *	MADV_FREE	(OBJT_SWAP objects, OBJ_ONEMAPPING only)
1314  *
1315  *	    Deactivate and clean the specified pages if they are
1316  *	    resident.  This permits the process to reuse the pages
1317  *	    without faulting or the kernel to reclaim the pages
1318  *	    without I/O.
1319  */
1320 void
1321 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1322     int advice)
1323 {
1324 	vm_pindex_t tpindex;
1325 	vm_object_t backing_object, tobject;
1326 	vm_page_t m, tm;
1327 
1328 	if (object == NULL)
1329 		return;
1330 
1331 relookup:
1332 	VM_OBJECT_WLOCK(object);
1333 	if (!vm_object_advice_applies(object, advice)) {
1334 		VM_OBJECT_WUNLOCK(object);
1335 		return;
1336 	}
1337 	for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1338 		tobject = object;
1339 
1340 		/*
1341 		 * If the next page isn't resident in the top-level object, we
1342 		 * need to search the shadow chain.  When applying MADV_FREE, we
1343 		 * take care to release any swap space used to store
1344 		 * non-resident pages.
1345 		 */
1346 		if (m == NULL || pindex < m->pindex) {
1347 			/*
1348 			 * Optimize a common case: if the top-level object has
1349 			 * no backing object, we can skip over the non-resident
1350 			 * range in constant time.
1351 			 */
1352 			if (object->backing_object == NULL) {
1353 				tpindex = (m != NULL && m->pindex < end) ?
1354 				    m->pindex : end;
1355 				vm_object_madvise_freespace(object, advice,
1356 				    pindex, tpindex - pindex);
1357 				if ((pindex = tpindex) == end)
1358 					break;
1359 				goto next_page;
1360 			}
1361 
1362 			tpindex = pindex;
1363 			do {
1364 				vm_object_madvise_freespace(tobject, advice,
1365 				    tpindex, 1);
1366 				/*
1367 				 * Prepare to search the next object in the
1368 				 * chain.
1369 				 */
1370 				backing_object = tobject->backing_object;
1371 				if (backing_object == NULL)
1372 					goto next_pindex;
1373 				VM_OBJECT_WLOCK(backing_object);
1374 				tpindex +=
1375 				    OFF_TO_IDX(tobject->backing_object_offset);
1376 				if (tobject != object)
1377 					VM_OBJECT_WUNLOCK(tobject);
1378 				tobject = backing_object;
1379 				if (!vm_object_advice_applies(tobject, advice))
1380 					goto next_pindex;
1381 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1382 			    NULL);
1383 		} else {
1384 next_page:
1385 			tm = m;
1386 			m = TAILQ_NEXT(m, listq);
1387 		}
1388 
1389 		/*
1390 		 * If the page is not in a normal state, skip it.  The page
1391 		 * can not be invalidated while the object lock is held.
1392 		 */
1393 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1394 			goto next_pindex;
1395 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1396 		    ("vm_object_madvise: page %p is fictitious", tm));
1397 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1398 		    ("vm_object_madvise: page %p is not managed", tm));
1399 		if (vm_page_tryxbusy(tm) == 0) {
1400 			if (object != tobject)
1401 				VM_OBJECT_WUNLOCK(object);
1402 			if (advice == MADV_WILLNEED) {
1403 				/*
1404 				 * Reference the page before unlocking and
1405 				 * sleeping so that the page daemon is less
1406 				 * likely to reclaim it.
1407 				 */
1408 				vm_page_aflag_set(tm, PGA_REFERENCED);
1409 			}
1410 			if (!vm_page_busy_sleep(tm, "madvpo", 0))
1411 				VM_OBJECT_WUNLOCK(tobject);
1412   			goto relookup;
1413 		}
1414 		vm_page_advise(tm, advice);
1415 		vm_page_xunbusy(tm);
1416 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1417 next_pindex:
1418 		if (tobject != object)
1419 			VM_OBJECT_WUNLOCK(tobject);
1420 	}
1421 	VM_OBJECT_WUNLOCK(object);
1422 }
1423 
1424 /*
1425  *	vm_object_shadow:
1426  *
1427  *	Create a new object which is backed by the
1428  *	specified existing object range.  The source
1429  *	object reference is deallocated.
1430  *
1431  *	The new object and offset into that object
1432  *	are returned in the source parameters.
1433  */
1434 void
1435 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1436     struct ucred *cred, bool shared)
1437 {
1438 	vm_object_t source;
1439 	vm_object_t result;
1440 
1441 	source = *object;
1442 
1443 	/*
1444 	 * Don't create the new object if the old object isn't shared.
1445 	 *
1446 	 * If we hold the only reference we can guarantee that it won't
1447 	 * increase while we have the map locked.  Otherwise the race is
1448 	 * harmless and we will end up with an extra shadow object that
1449 	 * will be collapsed later.
1450 	 */
1451 	if (source != NULL && source->ref_count == 1 &&
1452 	    (source->flags & OBJ_ANON) != 0)
1453 		return;
1454 
1455 	/*
1456 	 * Allocate a new object with the given length.
1457 	 */
1458 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1459 
1460 	/*
1461 	 * Store the offset into the source object, and fix up the offset into
1462 	 * the new object.
1463 	 */
1464 	result->backing_object_offset = *offset;
1465 
1466 	if (shared || source != NULL) {
1467 		VM_OBJECT_WLOCK(result);
1468 
1469 		/*
1470 		 * The new object shadows the source object, adding a
1471 		 * reference to it.  Our caller changes his reference
1472 		 * to point to the new object, removing a reference to
1473 		 * the source object.  Net result: no change of
1474 		 * reference count, unless the caller needs to add one
1475 		 * more reference due to forking a shared map entry.
1476 		 */
1477 		if (shared) {
1478 			vm_object_reference_locked(result);
1479 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1480 		}
1481 
1482 		/*
1483 		 * Try to optimize the result object's page color when
1484 		 * shadowing in order to maintain page coloring
1485 		 * consistency in the combined shadowed object.
1486 		 */
1487 		if (source != NULL) {
1488 			vm_object_backing_insert(result, source);
1489 			result->domain = source->domain;
1490 #if VM_NRESERVLEVEL > 0
1491 			vm_object_set_flag(result,
1492 			    (source->flags & OBJ_COLORED));
1493 			result->pg_color = (source->pg_color +
1494 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1495 			    1)) - 1);
1496 #endif
1497 		}
1498 		VM_OBJECT_WUNLOCK(result);
1499 	}
1500 
1501 	/*
1502 	 * Return the new things
1503 	 */
1504 	*offset = 0;
1505 	*object = result;
1506 }
1507 
1508 /*
1509  *	vm_object_split:
1510  *
1511  * Split the pages in a map entry into a new object.  This affords
1512  * easier removal of unused pages, and keeps object inheritance from
1513  * being a negative impact on memory usage.
1514  */
1515 void
1516 vm_object_split(vm_map_entry_t entry)
1517 {
1518 	vm_page_t m, m_next;
1519 	vm_object_t orig_object, new_object, backing_object;
1520 	vm_pindex_t idx, offidxstart;
1521 	vm_size_t size;
1522 
1523 	orig_object = entry->object.vm_object;
1524 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1525 	    ("vm_object_split:  Splitting object with multiple mappings."));
1526 	if ((orig_object->flags & OBJ_ANON) == 0)
1527 		return;
1528 	if (orig_object->ref_count <= 1)
1529 		return;
1530 	VM_OBJECT_WUNLOCK(orig_object);
1531 
1532 	offidxstart = OFF_TO_IDX(entry->offset);
1533 	size = atop(entry->end - entry->start);
1534 
1535 	new_object = vm_object_allocate_anon(size, orig_object,
1536 	    orig_object->cred, ptoa(size));
1537 
1538 	/*
1539 	 * We must wait for the orig_object to complete any in-progress
1540 	 * collapse so that the swap blocks are stable below.  The
1541 	 * additional reference on backing_object by new object will
1542 	 * prevent further collapse operations until split completes.
1543 	 */
1544 	VM_OBJECT_WLOCK(orig_object);
1545 	vm_object_collapse_wait(orig_object);
1546 
1547 	/*
1548 	 * At this point, the new object is still private, so the order in
1549 	 * which the original and new objects are locked does not matter.
1550 	 */
1551 	VM_OBJECT_WLOCK(new_object);
1552 	new_object->domain = orig_object->domain;
1553 	backing_object = orig_object->backing_object;
1554 	if (backing_object != NULL) {
1555 		vm_object_backing_insert_ref(new_object, backing_object);
1556 		new_object->backing_object_offset =
1557 		    orig_object->backing_object_offset + entry->offset;
1558 	}
1559 	if (orig_object->cred != NULL) {
1560 		crhold(orig_object->cred);
1561 		KASSERT(orig_object->charge >= ptoa(size),
1562 		    ("orig_object->charge < 0"));
1563 		orig_object->charge -= ptoa(size);
1564 	}
1565 
1566 	/*
1567 	 * Mark the split operation so that swap_pager_getpages() knows
1568 	 * that the object is in transition.
1569 	 */
1570 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1571 #ifdef INVARIANTS
1572 	idx = 0;
1573 #endif
1574 retry:
1575 	m = vm_page_find_least(orig_object, offidxstart);
1576 	KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1577 	    ("%s: object %p was repopulated", __func__, orig_object));
1578 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1579 	    m = m_next) {
1580 		m_next = TAILQ_NEXT(m, listq);
1581 
1582 		/*
1583 		 * We must wait for pending I/O to complete before we can
1584 		 * rename the page.
1585 		 *
1586 		 * We do not have to VM_PROT_NONE the page as mappings should
1587 		 * not be changed by this operation.
1588 		 */
1589 		if (vm_page_tryxbusy(m) == 0) {
1590 			VM_OBJECT_WUNLOCK(new_object);
1591 			if (vm_page_busy_sleep(m, "spltwt", 0))
1592 				VM_OBJECT_WLOCK(orig_object);
1593 			VM_OBJECT_WLOCK(new_object);
1594 			goto retry;
1595 		}
1596 
1597 		/*
1598 		 * The page was left invalid.  Likely placed there by
1599 		 * an incomplete fault.  Just remove and ignore.
1600 		 */
1601 		if (vm_page_none_valid(m)) {
1602 			if (vm_page_remove(m))
1603 				vm_page_free(m);
1604 			continue;
1605 		}
1606 
1607 		/* vm_page_rename() will dirty the page. */
1608 		if (vm_page_rename(m, new_object, idx)) {
1609 			vm_page_xunbusy(m);
1610 			VM_OBJECT_WUNLOCK(new_object);
1611 			VM_OBJECT_WUNLOCK(orig_object);
1612 			vm_radix_wait();
1613 			VM_OBJECT_WLOCK(orig_object);
1614 			VM_OBJECT_WLOCK(new_object);
1615 			goto retry;
1616 		}
1617 
1618 #if VM_NRESERVLEVEL > 0
1619 		/*
1620 		 * If some of the reservation's allocated pages remain with
1621 		 * the original object, then transferring the reservation to
1622 		 * the new object is neither particularly beneficial nor
1623 		 * particularly harmful as compared to leaving the reservation
1624 		 * with the original object.  If, however, all of the
1625 		 * reservation's allocated pages are transferred to the new
1626 		 * object, then transferring the reservation is typically
1627 		 * beneficial.  Determining which of these two cases applies
1628 		 * would be more costly than unconditionally renaming the
1629 		 * reservation.
1630 		 */
1631 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1632 #endif
1633 	}
1634 
1635 	/*
1636 	 * swap_pager_copy() can sleep, in which case the orig_object's
1637 	 * and new_object's locks are released and reacquired.
1638 	 */
1639 	swap_pager_copy(orig_object, new_object, offidxstart, 0);
1640 
1641 	TAILQ_FOREACH(m, &new_object->memq, listq)
1642 		vm_page_xunbusy(m);
1643 
1644 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1645 	VM_OBJECT_WUNLOCK(orig_object);
1646 	VM_OBJECT_WUNLOCK(new_object);
1647 	entry->object.vm_object = new_object;
1648 	entry->offset = 0LL;
1649 	vm_object_deallocate(orig_object);
1650 	VM_OBJECT_WLOCK(new_object);
1651 }
1652 
1653 static vm_page_t
1654 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1655 {
1656 	vm_object_t backing_object;
1657 
1658 	VM_OBJECT_ASSERT_WLOCKED(object);
1659 	backing_object = object->backing_object;
1660 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1661 
1662 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1663 	    ("invalid ownership %p %p %p", p, object, backing_object));
1664 	/* The page is only NULL when rename fails. */
1665 	if (p == NULL) {
1666 		VM_OBJECT_WUNLOCK(object);
1667 		VM_OBJECT_WUNLOCK(backing_object);
1668 		vm_radix_wait();
1669 		VM_OBJECT_WLOCK(object);
1670 	} else if (p->object == object) {
1671 		VM_OBJECT_WUNLOCK(backing_object);
1672 		if (vm_page_busy_sleep(p, "vmocol", 0))
1673 			VM_OBJECT_WLOCK(object);
1674 	} else {
1675 		VM_OBJECT_WUNLOCK(object);
1676 		if (!vm_page_busy_sleep(p, "vmocol", 0))
1677 			VM_OBJECT_WUNLOCK(backing_object);
1678 		VM_OBJECT_WLOCK(object);
1679 	}
1680 	VM_OBJECT_WLOCK(backing_object);
1681 	return (TAILQ_FIRST(&backing_object->memq));
1682 }
1683 
1684 static bool
1685 vm_object_scan_all_shadowed(vm_object_t object)
1686 {
1687 	vm_object_t backing_object;
1688 	vm_page_t p, pp;
1689 	vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1690 
1691 	VM_OBJECT_ASSERT_WLOCKED(object);
1692 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1693 
1694 	backing_object = object->backing_object;
1695 
1696 	if ((backing_object->flags & OBJ_ANON) == 0)
1697 		return (false);
1698 
1699 	pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1700 	p = vm_page_find_least(backing_object, pi);
1701 	ps = swap_pager_find_least(backing_object, pi);
1702 
1703 	/*
1704 	 * Only check pages inside the parent object's range and
1705 	 * inside the parent object's mapping of the backing object.
1706 	 */
1707 	for (;; pi++) {
1708 		if (p != NULL && p->pindex < pi)
1709 			p = TAILQ_NEXT(p, listq);
1710 		if (ps < pi)
1711 			ps = swap_pager_find_least(backing_object, pi);
1712 		if (p == NULL && ps >= backing_object->size)
1713 			break;
1714 		else if (p == NULL)
1715 			pi = ps;
1716 		else
1717 			pi = MIN(p->pindex, ps);
1718 
1719 		new_pindex = pi - backing_offset_index;
1720 		if (new_pindex >= object->size)
1721 			break;
1722 
1723 		if (p != NULL) {
1724 			/*
1725 			 * If the backing object page is busy a
1726 			 * grandparent or older page may still be
1727 			 * undergoing CoW.  It is not safe to collapse
1728 			 * the backing object until it is quiesced.
1729 			 */
1730 			if (vm_page_tryxbusy(p) == 0)
1731 				return (false);
1732 
1733 			/*
1734 			 * We raced with the fault handler that left
1735 			 * newly allocated invalid page on the object
1736 			 * queue and retried.
1737 			 */
1738 			if (!vm_page_all_valid(p))
1739 				goto unbusy_ret;
1740 		}
1741 
1742 		/*
1743 		 * See if the parent has the page or if the parent's object
1744 		 * pager has the page.  If the parent has the page but the page
1745 		 * is not valid, the parent's object pager must have the page.
1746 		 *
1747 		 * If this fails, the parent does not completely shadow the
1748 		 * object and we might as well give up now.
1749 		 */
1750 		pp = vm_page_lookup(object, new_pindex);
1751 
1752 		/*
1753 		 * The valid check here is stable due to object lock
1754 		 * being required to clear valid and initiate paging.
1755 		 * Busy of p disallows fault handler to validate pp.
1756 		 */
1757 		if ((pp == NULL || vm_page_none_valid(pp)) &&
1758 		    !vm_pager_has_page(object, new_pindex, NULL, NULL))
1759 			goto unbusy_ret;
1760 		if (p != NULL)
1761 			vm_page_xunbusy(p);
1762 	}
1763 	return (true);
1764 
1765 unbusy_ret:
1766 	if (p != NULL)
1767 		vm_page_xunbusy(p);
1768 	return (false);
1769 }
1770 
1771 static void
1772 vm_object_collapse_scan(vm_object_t object)
1773 {
1774 	vm_object_t backing_object;
1775 	vm_page_t next, p, pp;
1776 	vm_pindex_t backing_offset_index, new_pindex;
1777 
1778 	VM_OBJECT_ASSERT_WLOCKED(object);
1779 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1780 
1781 	backing_object = object->backing_object;
1782 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1783 
1784 	/*
1785 	 * Our scan
1786 	 */
1787 	for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1788 		next = TAILQ_NEXT(p, listq);
1789 		new_pindex = p->pindex - backing_offset_index;
1790 
1791 		/*
1792 		 * Check for busy page
1793 		 */
1794 		if (vm_page_tryxbusy(p) == 0) {
1795 			next = vm_object_collapse_scan_wait(object, p);
1796 			continue;
1797 		}
1798 
1799 		KASSERT(object->backing_object == backing_object,
1800 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1801 		    object->backing_object, backing_object));
1802 		KASSERT(p->object == backing_object,
1803 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1804 		    p->object, backing_object));
1805 
1806 		if (p->pindex < backing_offset_index ||
1807 		    new_pindex >= object->size) {
1808 			vm_pager_freespace(backing_object, p->pindex, 1);
1809 
1810 			KASSERT(!pmap_page_is_mapped(p),
1811 			    ("freeing mapped page %p", p));
1812 			if (vm_page_remove(p))
1813 				vm_page_free(p);
1814 			continue;
1815 		}
1816 
1817 		if (!vm_page_all_valid(p)) {
1818 			KASSERT(!pmap_page_is_mapped(p),
1819 			    ("freeing mapped page %p", p));
1820 			if (vm_page_remove(p))
1821 				vm_page_free(p);
1822 			continue;
1823 		}
1824 
1825 		pp = vm_page_lookup(object, new_pindex);
1826 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1827 			vm_page_xunbusy(p);
1828 			/*
1829 			 * The page in the parent is busy and possibly not
1830 			 * (yet) valid.  Until its state is finalized by the
1831 			 * busy bit owner, we can't tell whether it shadows the
1832 			 * original page.
1833 			 */
1834 			next = vm_object_collapse_scan_wait(object, pp);
1835 			continue;
1836 		}
1837 
1838 		if (pp != NULL && vm_page_none_valid(pp)) {
1839 			/*
1840 			 * The page was invalid in the parent.  Likely placed
1841 			 * there by an incomplete fault.  Just remove and
1842 			 * ignore.  p can replace it.
1843 			 */
1844 			if (vm_page_remove(pp))
1845 				vm_page_free(pp);
1846 			pp = NULL;
1847 		}
1848 
1849 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1850 			NULL)) {
1851 			/*
1852 			 * The page already exists in the parent OR swap exists
1853 			 * for this location in the parent.  Leave the parent's
1854 			 * page alone.  Destroy the original page from the
1855 			 * backing object.
1856 			 */
1857 			vm_pager_freespace(backing_object, p->pindex, 1);
1858 			KASSERT(!pmap_page_is_mapped(p),
1859 			    ("freeing mapped page %p", p));
1860 			if (vm_page_remove(p))
1861 				vm_page_free(p);
1862 			if (pp != NULL)
1863 				vm_page_xunbusy(pp);
1864 			continue;
1865 		}
1866 
1867 		/*
1868 		 * Page does not exist in parent, rename the page from the
1869 		 * backing object to the main object.
1870 		 *
1871 		 * If the page was mapped to a process, it can remain mapped
1872 		 * through the rename.  vm_page_rename() will dirty the page.
1873 		 */
1874 		if (vm_page_rename(p, object, new_pindex)) {
1875 			vm_page_xunbusy(p);
1876 			next = vm_object_collapse_scan_wait(object, NULL);
1877 			continue;
1878 		}
1879 
1880 		/* Use the old pindex to free the right page. */
1881 		vm_pager_freespace(backing_object, new_pindex +
1882 		    backing_offset_index, 1);
1883 
1884 #if VM_NRESERVLEVEL > 0
1885 		/*
1886 		 * Rename the reservation.
1887 		 */
1888 		vm_reserv_rename(p, object, backing_object,
1889 		    backing_offset_index);
1890 #endif
1891 		vm_page_xunbusy(p);
1892 	}
1893 	return;
1894 }
1895 
1896 /*
1897  *	vm_object_collapse:
1898  *
1899  *	Collapse an object with the object backing it.
1900  *	Pages in the backing object are moved into the
1901  *	parent, and the backing object is deallocated.
1902  */
1903 void
1904 vm_object_collapse(vm_object_t object)
1905 {
1906 	vm_object_t backing_object, new_backing_object;
1907 
1908 	VM_OBJECT_ASSERT_WLOCKED(object);
1909 
1910 	while (TRUE) {
1911 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1912 		    ("collapsing invalid object"));
1913 
1914 		/*
1915 		 * Wait for the backing_object to finish any pending
1916 		 * collapse so that the caller sees the shortest possible
1917 		 * shadow chain.
1918 		 */
1919 		backing_object = vm_object_backing_collapse_wait(object);
1920 		if (backing_object == NULL)
1921 			return;
1922 
1923 		KASSERT(object->ref_count > 0 &&
1924 		    object->ref_count > atomic_load_int(&object->shadow_count),
1925 		    ("collapse with invalid ref %d or shadow %d count.",
1926 		    object->ref_count, atomic_load_int(&object->shadow_count)));
1927 		KASSERT((backing_object->flags &
1928 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1929 		    ("vm_object_collapse: Backing object already collapsing."));
1930 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1931 		    ("vm_object_collapse: object is already collapsing."));
1932 
1933 		/*
1934 		 * We know that we can either collapse the backing object if
1935 		 * the parent is the only reference to it, or (perhaps) have
1936 		 * the parent bypass the object if the parent happens to shadow
1937 		 * all the resident pages in the entire backing object.
1938 		 */
1939 		if (backing_object->ref_count == 1) {
1940 			KASSERT(atomic_load_int(&backing_object->shadow_count)
1941 			    == 1,
1942 			    ("vm_object_collapse: shadow_count: %d",
1943 			    atomic_load_int(&backing_object->shadow_count)));
1944 			vm_object_pip_add(object, 1);
1945 			vm_object_set_flag(object, OBJ_COLLAPSING);
1946 			vm_object_pip_add(backing_object, 1);
1947 			vm_object_set_flag(backing_object, OBJ_DEAD);
1948 
1949 			/*
1950 			 * If there is exactly one reference to the backing
1951 			 * object, we can collapse it into the parent.
1952 			 */
1953 			vm_object_collapse_scan(object);
1954 
1955 #if VM_NRESERVLEVEL > 0
1956 			/*
1957 			 * Break any reservations from backing_object.
1958 			 */
1959 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1960 				vm_reserv_break_all(backing_object);
1961 #endif
1962 
1963 			/*
1964 			 * Move the pager from backing_object to object.
1965 			 *
1966 			 * swap_pager_copy() can sleep, in which case the
1967 			 * backing_object's and object's locks are released and
1968 			 * reacquired.
1969 			 */
1970 			swap_pager_copy(backing_object, object,
1971 			    OFF_TO_IDX(object->backing_object_offset), TRUE);
1972 
1973 			/*
1974 			 * Object now shadows whatever backing_object did.
1975 			 */
1976 			vm_object_clear_flag(object, OBJ_COLLAPSING);
1977 			vm_object_backing_transfer(object, backing_object);
1978 			object->backing_object_offset +=
1979 			    backing_object->backing_object_offset;
1980 			VM_OBJECT_WUNLOCK(object);
1981 			vm_object_pip_wakeup(object);
1982 
1983 			/*
1984 			 * Discard backing_object.
1985 			 *
1986 			 * Since the backing object has no pages, no pager left,
1987 			 * and no object references within it, all that is
1988 			 * necessary is to dispose of it.
1989 			 */
1990 			KASSERT(backing_object->ref_count == 1, (
1991 "backing_object %p was somehow re-referenced during collapse!",
1992 			    backing_object));
1993 			vm_object_pip_wakeup(backing_object);
1994 			(void)refcount_release(&backing_object->ref_count);
1995 			umtx_shm_object_terminated(backing_object);
1996 			vm_object_terminate(backing_object);
1997 			counter_u64_add(object_collapses, 1);
1998 			VM_OBJECT_WLOCK(object);
1999 		} else {
2000 			/*
2001 			 * If we do not entirely shadow the backing object,
2002 			 * there is nothing we can do so we give up.
2003 			 *
2004 			 * The object lock and backing_object lock must not
2005 			 * be dropped during this sequence.
2006 			 */
2007 			if (!vm_object_scan_all_shadowed(object)) {
2008 				VM_OBJECT_WUNLOCK(backing_object);
2009 				break;
2010 			}
2011 
2012 			/*
2013 			 * Make the parent shadow the next object in the
2014 			 * chain.  Deallocating backing_object will not remove
2015 			 * it, since its reference count is at least 2.
2016 			 */
2017 			vm_object_backing_remove_locked(object);
2018 			new_backing_object = backing_object->backing_object;
2019 			if (new_backing_object != NULL) {
2020 				vm_object_backing_insert_ref(object,
2021 				    new_backing_object);
2022 				object->backing_object_offset +=
2023 				    backing_object->backing_object_offset;
2024 			}
2025 
2026 			/*
2027 			 * Drop the reference count on backing_object. Since
2028 			 * its ref_count was at least 2, it will not vanish.
2029 			 */
2030 			(void)refcount_release(&backing_object->ref_count);
2031 			KASSERT(backing_object->ref_count >= 1, (
2032 "backing_object %p was somehow dereferenced during collapse!",
2033 			    backing_object));
2034 			VM_OBJECT_WUNLOCK(backing_object);
2035 			counter_u64_add(object_bypasses, 1);
2036 		}
2037 
2038 		/*
2039 		 * Try again with this object's new backing object.
2040 		 */
2041 	}
2042 }
2043 
2044 /*
2045  *	vm_object_page_remove:
2046  *
2047  *	For the given object, either frees or invalidates each of the
2048  *	specified pages.  In general, a page is freed.  However, if a page is
2049  *	wired for any reason other than the existence of a managed, wired
2050  *	mapping, then it may be invalidated but not removed from the object.
2051  *	Pages are specified by the given range ["start", "end") and the option
2052  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
2053  *	extends from "start" to the end of the object.  If the option
2054  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2055  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
2056  *	specified, then the pages within the specified range must have no
2057  *	mappings.  Otherwise, if this option is not specified, any mappings to
2058  *	the specified pages are removed before the pages are freed or
2059  *	invalidated.
2060  *
2061  *	In general, this operation should only be performed on objects that
2062  *	contain managed pages.  There are, however, two exceptions.  First, it
2063  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
2064  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2065  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
2066  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
2067  *
2068  *	The object must be locked.
2069  */
2070 void
2071 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2072     int options)
2073 {
2074 	vm_page_t p, next;
2075 
2076 	VM_OBJECT_ASSERT_WLOCKED(object);
2077 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2078 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2079 	    ("vm_object_page_remove: illegal options for object %p", object));
2080 	if (object->resident_page_count == 0)
2081 		return;
2082 	vm_object_pip_add(object, 1);
2083 again:
2084 	p = vm_page_find_least(object, start);
2085 
2086 	/*
2087 	 * Here, the variable "p" is either (1) the page with the least pindex
2088 	 * greater than or equal to the parameter "start" or (2) NULL.
2089 	 */
2090 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2091 		next = TAILQ_NEXT(p, listq);
2092 
2093 		/*
2094 		 * Skip invalid pages if asked to do so.  Try to avoid acquiring
2095 		 * the busy lock, as some consumers rely on this to avoid
2096 		 * deadlocks.
2097 		 *
2098 		 * A thread may concurrently transition the page from invalid to
2099 		 * valid using only the busy lock, so the result of this check
2100 		 * is immediately stale.  It is up to consumers to handle this,
2101 		 * for instance by ensuring that all invalid->valid transitions
2102 		 * happen with a mutex held, as may be possible for a
2103 		 * filesystem.
2104 		 */
2105 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2106 			continue;
2107 
2108 		/*
2109 		 * If the page is wired for any reason besides the existence
2110 		 * of managed, wired mappings, then it cannot be freed.  For
2111 		 * example, fictitious pages, which represent device memory,
2112 		 * are inherently wired and cannot be freed.  They can,
2113 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2114 		 * not specified.
2115 		 */
2116 		if (vm_page_tryxbusy(p) == 0) {
2117 			if (vm_page_busy_sleep(p, "vmopar", 0))
2118 				VM_OBJECT_WLOCK(object);
2119 			goto again;
2120 		}
2121 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2122 			vm_page_xunbusy(p);
2123 			continue;
2124 		}
2125 		if (vm_page_wired(p)) {
2126 wired:
2127 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2128 			    object->ref_count != 0)
2129 				pmap_remove_all(p);
2130 			if ((options & OBJPR_CLEANONLY) == 0) {
2131 				vm_page_invalid(p);
2132 				vm_page_undirty(p);
2133 			}
2134 			vm_page_xunbusy(p);
2135 			continue;
2136 		}
2137 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2138 		    ("vm_object_page_remove: page %p is fictitious", p));
2139 		if ((options & OBJPR_CLEANONLY) != 0 &&
2140 		    !vm_page_none_valid(p)) {
2141 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2142 			    object->ref_count != 0 &&
2143 			    !vm_page_try_remove_write(p))
2144 				goto wired;
2145 			if (p->dirty != 0) {
2146 				vm_page_xunbusy(p);
2147 				continue;
2148 			}
2149 		}
2150 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2151 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2152 			goto wired;
2153 		vm_page_free(p);
2154 	}
2155 	vm_object_pip_wakeup(object);
2156 
2157 	vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2158 	    start);
2159 }
2160 
2161 /*
2162  *	vm_object_page_noreuse:
2163  *
2164  *	For the given object, attempt to move the specified pages to
2165  *	the head of the inactive queue.  This bypasses regular LRU
2166  *	operation and allows the pages to be reused quickly under memory
2167  *	pressure.  If a page is wired for any reason, then it will not
2168  *	be queued.  Pages are specified by the range ["start", "end").
2169  *	As a special case, if "end" is zero, then the range extends from
2170  *	"start" to the end of the object.
2171  *
2172  *	This operation should only be performed on objects that
2173  *	contain non-fictitious, managed pages.
2174  *
2175  *	The object must be locked.
2176  */
2177 void
2178 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2179 {
2180 	vm_page_t p, next;
2181 
2182 	VM_OBJECT_ASSERT_LOCKED(object);
2183 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2184 	    ("vm_object_page_noreuse: illegal object %p", object));
2185 	if (object->resident_page_count == 0)
2186 		return;
2187 	p = vm_page_find_least(object, start);
2188 
2189 	/*
2190 	 * Here, the variable "p" is either (1) the page with the least pindex
2191 	 * greater than or equal to the parameter "start" or (2) NULL.
2192 	 */
2193 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2194 		next = TAILQ_NEXT(p, listq);
2195 		vm_page_deactivate_noreuse(p);
2196 	}
2197 }
2198 
2199 /*
2200  *	Populate the specified range of the object with valid pages.  Returns
2201  *	TRUE if the range is successfully populated and FALSE otherwise.
2202  *
2203  *	Note: This function should be optimized to pass a larger array of
2204  *	pages to vm_pager_get_pages() before it is applied to a non-
2205  *	OBJT_DEVICE object.
2206  *
2207  *	The object must be locked.
2208  */
2209 boolean_t
2210 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2211 {
2212 	vm_page_t m;
2213 	vm_pindex_t pindex;
2214 	int rv;
2215 
2216 	VM_OBJECT_ASSERT_WLOCKED(object);
2217 	for (pindex = start; pindex < end; pindex++) {
2218 		rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2219 		if (rv != VM_PAGER_OK)
2220 			break;
2221 
2222 		/*
2223 		 * Keep "m" busy because a subsequent iteration may unlock
2224 		 * the object.
2225 		 */
2226 	}
2227 	if (pindex > start) {
2228 		m = vm_page_lookup(object, start);
2229 		while (m != NULL && m->pindex < pindex) {
2230 			vm_page_xunbusy(m);
2231 			m = TAILQ_NEXT(m, listq);
2232 		}
2233 	}
2234 	return (pindex == end);
2235 }
2236 
2237 /*
2238  *	Routine:	vm_object_coalesce
2239  *	Function:	Coalesces two objects backing up adjoining
2240  *			regions of memory into a single object.
2241  *
2242  *	returns TRUE if objects were combined.
2243  *
2244  *	NOTE:	Only works at the moment if the second object is NULL -
2245  *		if it's not, which object do we lock first?
2246  *
2247  *	Parameters:
2248  *		prev_object	First object to coalesce
2249  *		prev_offset	Offset into prev_object
2250  *		prev_size	Size of reference to prev_object
2251  *		next_size	Size of reference to the second object
2252  *		reserved	Indicator that extension region has
2253  *				swap accounted for
2254  *
2255  *	Conditions:
2256  *	The object must *not* be locked.
2257  */
2258 boolean_t
2259 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2260     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2261 {
2262 	vm_pindex_t next_pindex;
2263 
2264 	if (prev_object == NULL)
2265 		return (TRUE);
2266 	if ((prev_object->flags & OBJ_ANON) == 0)
2267 		return (FALSE);
2268 
2269 	VM_OBJECT_WLOCK(prev_object);
2270 	/*
2271 	 * Try to collapse the object first.
2272 	 */
2273 	vm_object_collapse(prev_object);
2274 
2275 	/*
2276 	 * Can't coalesce if: . more than one reference . paged out . shadows
2277 	 * another object . has a copy elsewhere (any of which mean that the
2278 	 * pages not mapped to prev_entry may be in use anyway)
2279 	 */
2280 	if (prev_object->backing_object != NULL) {
2281 		VM_OBJECT_WUNLOCK(prev_object);
2282 		return (FALSE);
2283 	}
2284 
2285 	prev_size >>= PAGE_SHIFT;
2286 	next_size >>= PAGE_SHIFT;
2287 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2288 
2289 	if (prev_object->ref_count > 1 &&
2290 	    prev_object->size != next_pindex &&
2291 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2292 		VM_OBJECT_WUNLOCK(prev_object);
2293 		return (FALSE);
2294 	}
2295 
2296 	/*
2297 	 * Account for the charge.
2298 	 */
2299 	if (prev_object->cred != NULL) {
2300 		/*
2301 		 * If prev_object was charged, then this mapping,
2302 		 * although not charged now, may become writable
2303 		 * later. Non-NULL cred in the object would prevent
2304 		 * swap reservation during enabling of the write
2305 		 * access, so reserve swap now. Failed reservation
2306 		 * cause allocation of the separate object for the map
2307 		 * entry, and swap reservation for this entry is
2308 		 * managed in appropriate time.
2309 		 */
2310 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2311 		    prev_object->cred)) {
2312 			VM_OBJECT_WUNLOCK(prev_object);
2313 			return (FALSE);
2314 		}
2315 		prev_object->charge += ptoa(next_size);
2316 	}
2317 
2318 	/*
2319 	 * Remove any pages that may still be in the object from a previous
2320 	 * deallocation.
2321 	 */
2322 	if (next_pindex < prev_object->size) {
2323 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2324 		    next_size, 0);
2325 #if 0
2326 		if (prev_object->cred != NULL) {
2327 			KASSERT(prev_object->charge >=
2328 			    ptoa(prev_object->size - next_pindex),
2329 			    ("object %p overcharged 1 %jx %jx", prev_object,
2330 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2331 			prev_object->charge -= ptoa(prev_object->size -
2332 			    next_pindex);
2333 		}
2334 #endif
2335 	}
2336 
2337 	/*
2338 	 * Extend the object if necessary.
2339 	 */
2340 	if (next_pindex + next_size > prev_object->size)
2341 		prev_object->size = next_pindex + next_size;
2342 
2343 	VM_OBJECT_WUNLOCK(prev_object);
2344 	return (TRUE);
2345 }
2346 
2347 void
2348 vm_object_set_writeable_dirty_(vm_object_t object)
2349 {
2350 	atomic_add_int(&object->generation, 1);
2351 }
2352 
2353 bool
2354 vm_object_mightbedirty_(vm_object_t object)
2355 {
2356 	return (object->generation != object->cleangeneration);
2357 }
2358 
2359 /*
2360  *	vm_object_unwire:
2361  *
2362  *	For each page offset within the specified range of the given object,
2363  *	find the highest-level page in the shadow chain and unwire it.  A page
2364  *	must exist at every page offset, and the highest-level page must be
2365  *	wired.
2366  */
2367 void
2368 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2369     uint8_t queue)
2370 {
2371 	vm_object_t tobject, t1object;
2372 	vm_page_t m, tm;
2373 	vm_pindex_t end_pindex, pindex, tpindex;
2374 	int depth, locked_depth;
2375 
2376 	KASSERT((offset & PAGE_MASK) == 0,
2377 	    ("vm_object_unwire: offset is not page aligned"));
2378 	KASSERT((length & PAGE_MASK) == 0,
2379 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2380 	/* The wired count of a fictitious page never changes. */
2381 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2382 		return;
2383 	pindex = OFF_TO_IDX(offset);
2384 	end_pindex = pindex + atop(length);
2385 again:
2386 	locked_depth = 1;
2387 	VM_OBJECT_RLOCK(object);
2388 	m = vm_page_find_least(object, pindex);
2389 	while (pindex < end_pindex) {
2390 		if (m == NULL || pindex < m->pindex) {
2391 			/*
2392 			 * The first object in the shadow chain doesn't
2393 			 * contain a page at the current index.  Therefore,
2394 			 * the page must exist in a backing object.
2395 			 */
2396 			tobject = object;
2397 			tpindex = pindex;
2398 			depth = 0;
2399 			do {
2400 				tpindex +=
2401 				    OFF_TO_IDX(tobject->backing_object_offset);
2402 				tobject = tobject->backing_object;
2403 				KASSERT(tobject != NULL,
2404 				    ("vm_object_unwire: missing page"));
2405 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2406 					goto next_page;
2407 				depth++;
2408 				if (depth == locked_depth) {
2409 					locked_depth++;
2410 					VM_OBJECT_RLOCK(tobject);
2411 				}
2412 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2413 			    NULL);
2414 		} else {
2415 			tm = m;
2416 			m = TAILQ_NEXT(m, listq);
2417 		}
2418 		if (vm_page_trysbusy(tm) == 0) {
2419 			for (tobject = object; locked_depth >= 1;
2420 			    locked_depth--) {
2421 				t1object = tobject->backing_object;
2422 				if (tm->object != tobject)
2423 					VM_OBJECT_RUNLOCK(tobject);
2424 				tobject = t1object;
2425 			}
2426 			tobject = tm->object;
2427 			if (!vm_page_busy_sleep(tm, "unwbo",
2428 			    VM_ALLOC_IGN_SBUSY))
2429 				VM_OBJECT_RUNLOCK(tobject);
2430 			goto again;
2431 		}
2432 		vm_page_unwire(tm, queue);
2433 		vm_page_sunbusy(tm);
2434 next_page:
2435 		pindex++;
2436 	}
2437 	/* Release the accumulated object locks. */
2438 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2439 		t1object = tobject->backing_object;
2440 		VM_OBJECT_RUNLOCK(tobject);
2441 		tobject = t1object;
2442 	}
2443 }
2444 
2445 /*
2446  * Return the vnode for the given object, or NULL if none exists.
2447  * For tmpfs objects, the function may return NULL if there is
2448  * no vnode allocated at the time of the call.
2449  */
2450 struct vnode *
2451 vm_object_vnode(vm_object_t object)
2452 {
2453 	struct vnode *vp;
2454 
2455 	VM_OBJECT_ASSERT_LOCKED(object);
2456 	vm_pager_getvp(object, &vp, NULL);
2457 	return (vp);
2458 }
2459 
2460 /*
2461  * Busy the vm object.  This prevents new pages belonging to the object from
2462  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2463  * for checking page state before proceeding.
2464  */
2465 void
2466 vm_object_busy(vm_object_t obj)
2467 {
2468 
2469 	VM_OBJECT_ASSERT_LOCKED(obj);
2470 
2471 	blockcount_acquire(&obj->busy, 1);
2472 	/* The fence is required to order loads of page busy. */
2473 	atomic_thread_fence_acq_rel();
2474 }
2475 
2476 void
2477 vm_object_unbusy(vm_object_t obj)
2478 {
2479 
2480 	blockcount_release(&obj->busy, 1);
2481 }
2482 
2483 void
2484 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2485 {
2486 
2487 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2488 
2489 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2490 }
2491 
2492 /*
2493  * This function aims to determine if the object is mapped,
2494  * specifically, if it is referenced by a vm_map_entry.  Because
2495  * objects occasionally acquire transient references that do not
2496  * represent a mapping, the method used here is inexact.  However, it
2497  * has very low overhead and is good enough for the advisory
2498  * vm.vmtotal sysctl.
2499  */
2500 bool
2501 vm_object_is_active(vm_object_t obj)
2502 {
2503 
2504 	return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2505 }
2506 
2507 static int
2508 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2509 {
2510 	struct kinfo_vmobject *kvo;
2511 	char *fullpath, *freepath;
2512 	struct vnode *vp;
2513 	struct vattr va;
2514 	vm_object_t obj;
2515 	vm_page_t m;
2516 	u_long sp;
2517 	int count, error;
2518 
2519 	if (req->oldptr == NULL) {
2520 		/*
2521 		 * If an old buffer has not been provided, generate an
2522 		 * estimate of the space needed for a subsequent call.
2523 		 */
2524 		mtx_lock(&vm_object_list_mtx);
2525 		count = 0;
2526 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2527 			if (obj->type == OBJT_DEAD)
2528 				continue;
2529 			count++;
2530 		}
2531 		mtx_unlock(&vm_object_list_mtx);
2532 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2533 		    count * 11 / 10));
2534 	}
2535 
2536 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO);
2537 	error = 0;
2538 
2539 	/*
2540 	 * VM objects are type stable and are never removed from the
2541 	 * list once added.  This allows us to safely read obj->object_list
2542 	 * after reacquiring the VM object lock.
2543 	 */
2544 	mtx_lock(&vm_object_list_mtx);
2545 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2546 		if (obj->type == OBJT_DEAD ||
2547 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2548 			continue;
2549 		VM_OBJECT_RLOCK(obj);
2550 		if (obj->type == OBJT_DEAD ||
2551 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2552 			VM_OBJECT_RUNLOCK(obj);
2553 			continue;
2554 		}
2555 		mtx_unlock(&vm_object_list_mtx);
2556 		kvo->kvo_size = ptoa(obj->size);
2557 		kvo->kvo_resident = obj->resident_page_count;
2558 		kvo->kvo_ref_count = obj->ref_count;
2559 		kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2560 		kvo->kvo_memattr = obj->memattr;
2561 		kvo->kvo_active = 0;
2562 		kvo->kvo_inactive = 0;
2563 		if (!swap_only) {
2564 			TAILQ_FOREACH(m, &obj->memq, listq) {
2565 				/*
2566 				 * A page may belong to the object but be
2567 				 * dequeued and set to PQ_NONE while the
2568 				 * object lock is not held.  This makes the
2569 				 * reads of m->queue below racy, and we do not
2570 				 * count pages set to PQ_NONE.  However, this
2571 				 * sysctl is only meant to give an
2572 				 * approximation of the system anyway.
2573 				 */
2574 				if (m->a.queue == PQ_ACTIVE)
2575 					kvo->kvo_active++;
2576 				else if (m->a.queue == PQ_INACTIVE)
2577 					kvo->kvo_inactive++;
2578 			}
2579 		}
2580 
2581 		kvo->kvo_vn_fileid = 0;
2582 		kvo->kvo_vn_fsid = 0;
2583 		kvo->kvo_vn_fsid_freebsd11 = 0;
2584 		freepath = NULL;
2585 		fullpath = "";
2586 		vp = NULL;
2587 		kvo->kvo_type = vm_object_kvme_type(obj, swap_only ? NULL : &vp);
2588 		if (vp != NULL) {
2589 			vref(vp);
2590 		} else if ((obj->flags & OBJ_ANON) != 0) {
2591 			MPASS(kvo->kvo_type == KVME_TYPE_SWAP);
2592 			kvo->kvo_me = (uintptr_t)obj;
2593 			/* tmpfs objs are reported as vnodes */
2594 			kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2595 			sp = swap_pager_swapped_pages(obj);
2596 			kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2597 		}
2598 		VM_OBJECT_RUNLOCK(obj);
2599 		if (vp != NULL) {
2600 			vn_fullpath(vp, &fullpath, &freepath);
2601 			vn_lock(vp, LK_SHARED | LK_RETRY);
2602 			if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2603 				kvo->kvo_vn_fileid = va.va_fileid;
2604 				kvo->kvo_vn_fsid = va.va_fsid;
2605 				kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2606 								/* truncate */
2607 			}
2608 			vput(vp);
2609 		}
2610 
2611 		strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2612 		free(freepath, M_TEMP);
2613 
2614 		/* Pack record size down */
2615 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2616 		    + strlen(kvo->kvo_path) + 1;
2617 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2618 		    sizeof(uint64_t));
2619 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2620 		maybe_yield();
2621 		mtx_lock(&vm_object_list_mtx);
2622 		if (error)
2623 			break;
2624 	}
2625 	mtx_unlock(&vm_object_list_mtx);
2626 	free(kvo, M_TEMP);
2627 	return (error);
2628 }
2629 
2630 static int
2631 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2632 {
2633 	return (vm_object_list_handler(req, false));
2634 }
2635 
2636 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2637     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2638     "List of VM objects");
2639 
2640 static int
2641 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2642 {
2643 	return (vm_object_list_handler(req, true));
2644 }
2645 
2646 /*
2647  * This sysctl returns list of the anonymous or swap objects. Intent
2648  * is to provide stripped optimized list useful to analyze swap use.
2649  * Since technically non-swap (default) objects participate in the
2650  * shadow chains, and are converted to swap type as needed by swap
2651  * pager, we must report them.
2652  */
2653 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2654     CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2655     sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2656     "List of swap VM objects");
2657 
2658 #include "opt_ddb.h"
2659 #ifdef DDB
2660 #include <sys/kernel.h>
2661 
2662 #include <sys/cons.h>
2663 
2664 #include <ddb/ddb.h>
2665 
2666 static int
2667 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2668 {
2669 	vm_map_t tmpm;
2670 	vm_map_entry_t tmpe;
2671 	vm_object_t obj;
2672 
2673 	if (map == 0)
2674 		return 0;
2675 
2676 	if (entry == 0) {
2677 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2678 			if (_vm_object_in_map(map, object, tmpe)) {
2679 				return 1;
2680 			}
2681 		}
2682 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2683 		tmpm = entry->object.sub_map;
2684 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2685 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2686 				return 1;
2687 			}
2688 		}
2689 	} else if ((obj = entry->object.vm_object) != NULL) {
2690 		for (; obj; obj = obj->backing_object)
2691 			if (obj == object) {
2692 				return 1;
2693 			}
2694 	}
2695 	return 0;
2696 }
2697 
2698 static int
2699 vm_object_in_map(vm_object_t object)
2700 {
2701 	struct proc *p;
2702 
2703 	/* sx_slock(&allproc_lock); */
2704 	FOREACH_PROC_IN_SYSTEM(p) {
2705 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2706 			continue;
2707 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2708 			/* sx_sunlock(&allproc_lock); */
2709 			return 1;
2710 		}
2711 	}
2712 	/* sx_sunlock(&allproc_lock); */
2713 	if (_vm_object_in_map(kernel_map, object, 0))
2714 		return 1;
2715 	return 0;
2716 }
2717 
2718 DB_SHOW_COMMAND_FLAGS(vmochk, vm_object_check, DB_CMD_MEMSAFE)
2719 {
2720 	vm_object_t object;
2721 
2722 	/*
2723 	 * make sure that internal objs are in a map somewhere
2724 	 * and none have zero ref counts.
2725 	 */
2726 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2727 		if ((object->flags & OBJ_ANON) != 0) {
2728 			if (object->ref_count == 0) {
2729 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2730 					(long)object->size);
2731 			}
2732 			if (!vm_object_in_map(object)) {
2733 				db_printf(
2734 			"vmochk: internal obj is not in a map: "
2735 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2736 				    object->ref_count, (u_long)object->size,
2737 				    (u_long)object->size,
2738 				    (void *)object->backing_object);
2739 			}
2740 		}
2741 		if (db_pager_quit)
2742 			return;
2743 	}
2744 }
2745 
2746 /*
2747  *	vm_object_print:	[ debug ]
2748  */
2749 DB_SHOW_COMMAND(object, vm_object_print_static)
2750 {
2751 	/* XXX convert args. */
2752 	vm_object_t object = (vm_object_t)addr;
2753 	boolean_t full = have_addr;
2754 
2755 	vm_page_t p;
2756 
2757 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2758 #define	count	was_count
2759 
2760 	int count;
2761 
2762 	if (object == NULL)
2763 		return;
2764 
2765 	db_iprintf(
2766 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2767 	    object, (int)object->type, (uintmax_t)object->size,
2768 	    object->resident_page_count, object->ref_count, object->flags,
2769 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2770 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2771 	    atomic_load_int(&object->shadow_count),
2772 	    object->backing_object ? object->backing_object->ref_count : 0,
2773 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2774 
2775 	if (!full)
2776 		return;
2777 
2778 	db_indent += 2;
2779 	count = 0;
2780 	TAILQ_FOREACH(p, &object->memq, listq) {
2781 		if (count == 0)
2782 			db_iprintf("memory:=");
2783 		else if (count == 6) {
2784 			db_printf("\n");
2785 			db_iprintf(" ...");
2786 			count = 0;
2787 		} else
2788 			db_printf(",");
2789 		count++;
2790 
2791 		db_printf("(off=0x%jx,page=0x%jx)",
2792 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2793 
2794 		if (db_pager_quit)
2795 			break;
2796 	}
2797 	if (count != 0)
2798 		db_printf("\n");
2799 	db_indent -= 2;
2800 }
2801 
2802 /* XXX. */
2803 #undef count
2804 
2805 /* XXX need this non-static entry for calling from vm_map_print. */
2806 void
2807 vm_object_print(
2808         /* db_expr_t */ long addr,
2809 	boolean_t have_addr,
2810 	/* db_expr_t */ long count,
2811 	char *modif)
2812 {
2813 	vm_object_print_static(addr, have_addr, count, modif);
2814 }
2815 
2816 DB_SHOW_COMMAND_FLAGS(vmopag, vm_object_print_pages, DB_CMD_MEMSAFE)
2817 {
2818 	vm_object_t object;
2819 	vm_pindex_t fidx;
2820 	vm_paddr_t pa;
2821 	vm_page_t m, prev_m;
2822 	int rcount;
2823 
2824 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2825 		db_printf("new object: %p\n", (void *)object);
2826 		if (db_pager_quit)
2827 			return;
2828 
2829 		rcount = 0;
2830 		fidx = 0;
2831 		pa = -1;
2832 		TAILQ_FOREACH(m, &object->memq, listq) {
2833 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2834 			    prev_m->pindex + 1 != m->pindex) {
2835 				if (rcount) {
2836 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2837 						(long)fidx, rcount, (long)pa);
2838 					if (db_pager_quit)
2839 						return;
2840 					rcount = 0;
2841 				}
2842 			}
2843 			if (rcount &&
2844 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2845 				++rcount;
2846 				continue;
2847 			}
2848 			if (rcount) {
2849 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2850 					(long)fidx, rcount, (long)pa);
2851 				if (db_pager_quit)
2852 					return;
2853 			}
2854 			fidx = m->pindex;
2855 			pa = VM_PAGE_TO_PHYS(m);
2856 			rcount = 1;
2857 		}
2858 		if (rcount) {
2859 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2860 				(long)fidx, rcount, (long)pa);
2861 			if (db_pager_quit)
2862 				return;
2863 		}
2864 	}
2865 }
2866 #endif /* DDB */
2867