xref: /freebsd/sys/vm/vm_object.c (revision efedac28c7ce4bef9d9b09b7901f1cb35babc2c4)
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/conf.h>
70 #include <sys/cpuset.h>
71 #include <sys/ipc.h>
72 #include <sys/jail.h>
73 #include <sys/limits.h>
74 #include <sys/lock.h>
75 #include <sys/mman.h>
76 #include <sys/mount.h>
77 #include <sys/kernel.h>
78 #include <sys/mutex.h>
79 #include <sys/pctrie.h>
80 #include <sys/proc.h>
81 #include <sys/refcount.h>
82 #include <sys/shm.h>
83 #include <sys/sx.h>
84 #include <sys/sysctl.h>
85 #include <sys/resourcevar.h>
86 #include <sys/refcount.h>
87 #include <sys/rwlock.h>
88 #include <sys/user.h>
89 #include <sys/vnode.h>
90 #include <sys/vmmeter.h>
91 
92 #include <vm/vm.h>
93 #include <vm/vm_param.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_pager.h>
100 #include <vm/vm_phys.h>
101 #include <vm/vm_pagequeue.h>
102 #include <vm/swap_pager.h>
103 #include <vm/vm_kern.h>
104 #include <vm/vm_extern.h>
105 #include <vm/vm_radix.h>
106 #include <vm/vm_reserv.h>
107 #include <vm/uma.h>
108 
109 static int old_msync;
110 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
111     "Use old (insecure) msync behavior");
112 
113 static int	vm_object_page_collect_flush(struct pctrie_iter *pages,
114 		    vm_page_t p, int pagerflags, int flags, boolean_t *allclean,
115 		    boolean_t *eio);
116 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
117 		    boolean_t *allclean);
118 static void	vm_object_backing_remove(vm_object_t object);
119 
120 /*
121  *	Virtual memory objects maintain the actual data
122  *	associated with allocated virtual memory.  A given
123  *	page of memory exists within exactly one object.
124  *
125  *	An object is only deallocated when all "references"
126  *	are given up.  Only one "reference" to a given
127  *	region of an object should be writeable.
128  *
129  *	Associated with each object is a list of all resident
130  *	memory pages belonging to that object; this list is
131  *	maintained by the "vm_page" module, and locked by the object's
132  *	lock.
133  *
134  *	Each object also records a "pager" routine which is
135  *	used to retrieve (and store) pages to the proper backing
136  *	storage.  In addition, objects may be backed by other
137  *	objects from which they were virtual-copied.
138  *
139  *	The only items within the object structure which are
140  *	modified after time of creation are:
141  *		reference count		locked by object's lock
142  *		pager routine		locked by object's lock
143  *
144  */
145 
146 struct object_q vm_object_list;
147 struct mtx vm_object_list_mtx;	/* lock for object list and count */
148 
149 struct vm_object kernel_object_store;
150 
151 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
152     "VM object stats");
153 
154 static COUNTER_U64_DEFINE_EARLY(object_collapses);
155 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
156     &object_collapses,
157     "VM object collapses");
158 
159 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
160 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
161     &object_bypasses,
162     "VM object bypasses");
163 
164 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
165 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
166     &object_collapse_waits,
167     "Number of sleeps for collapse");
168 
169 static uma_zone_t obj_zone;
170 
171 static int vm_object_zinit(void *mem, int size, int flags);
172 
173 #ifdef INVARIANTS
174 static void vm_object_zdtor(void *mem, int size, void *arg);
175 
176 static void
177 vm_object_zdtor(void *mem, int size, void *arg)
178 {
179 	vm_object_t object;
180 
181 	object = (vm_object_t)mem;
182 	KASSERT(object->ref_count == 0,
183 	    ("object %p ref_count = %d", object, object->ref_count));
184 	KASSERT(vm_radix_is_empty(&object->rtree),
185 	    ("object %p has resident pages in its trie", object));
186 #if VM_NRESERVLEVEL > 0
187 	KASSERT(LIST_EMPTY(&object->rvq),
188 	    ("object %p has reservations",
189 	    object));
190 #endif
191 	KASSERT(!vm_object_busied(object),
192 	    ("object %p busy = %d", object, blockcount_read(&object->busy)));
193 	KASSERT(object->resident_page_count == 0,
194 	    ("object %p resident_page_count = %d",
195 	    object, object->resident_page_count));
196 	KASSERT(atomic_load_int(&object->shadow_count) == 0,
197 	    ("object %p shadow_count = %d",
198 	    object, atomic_load_int(&object->shadow_count)));
199 	KASSERT(object->type == OBJT_DEAD,
200 	    ("object %p has non-dead type %d",
201 	    object, object->type));
202 	KASSERT(object->charge == 0 && object->cred == NULL,
203 	    ("object %p has non-zero charge %ju (%p)",
204 	    object, (uintmax_t)object->charge, object->cred));
205 }
206 #endif
207 
208 static int
209 vm_object_zinit(void *mem, int size, int flags)
210 {
211 	vm_object_t object;
212 
213 	object = (vm_object_t)mem;
214 	rw_init_flags(&object->lock, "vmobject", RW_DUPOK | RW_NEW);
215 
216 	/* These are true for any object that has been freed */
217 	object->type = OBJT_DEAD;
218 	vm_radix_init(&object->rtree);
219 	refcount_init(&object->ref_count, 0);
220 	blockcount_init(&object->paging_in_progress);
221 	blockcount_init(&object->busy);
222 	object->resident_page_count = 0;
223 	atomic_store_int(&object->shadow_count, 0);
224 	object->flags = OBJ_DEAD;
225 
226 	mtx_lock(&vm_object_list_mtx);
227 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
228 	mtx_unlock(&vm_object_list_mtx);
229 	return (0);
230 }
231 
232 static void
233 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
234     vm_object_t object, void *handle)
235 {
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 (!vm_radix_is_empty(&object->rtree))
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_single_page removes a pageable page from the object,
880  *	and removes it from the paging queues and frees it, if it is not wired.
881  *	It is invoked via callback from vm_object_terminate_pages.
882  */
883 static void
884 vm_object_terminate_single_page(vm_page_t p, void *objectv)
885 {
886 	vm_object_t object __diagused = objectv;
887 
888 	vm_page_assert_unbusied(p);
889 	KASSERT(p->object == object &&
890 	    (p->ref_count & VPRC_OBJREF) != 0,
891 	    ("%s: page %p is inconsistent", __func__, p));
892 	p->object = NULL;
893 	if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
894 		KASSERT((object->flags & OBJ_UNMANAGED) != 0 ||
895 		    vm_page_astate_load(p).queue != PQ_NONE,
896 		    ("%s: page %p does not belong to a queue", __func__, p));
897 		VM_CNT_INC(v_pfree);
898 		vm_page_free(p);
899 	}
900 }
901 
902 /*
903  *	vm_object_terminate_pages removes any remaining pageable pages
904  *	from the object and resets the object to an empty state.
905  */
906 static void
907 vm_object_terminate_pages(vm_object_t object)
908 {
909 	VM_OBJECT_ASSERT_WLOCKED(object);
910 
911 	/*
912 	 * If the object contained any pages, then reset it to an empty state.
913 	 * Rather than incrementally removing each page from the object, the
914 	 * page and object are reset to any empty state.
915 	 */
916 	if (object->resident_page_count == 0)
917 		return;
918 
919 	vm_radix_reclaim_callback(&object->rtree,
920 	    vm_object_terminate_single_page, object);
921 	object->resident_page_count = 0;
922 	if (object->type == OBJT_VNODE)
923 		vdrop(object->handle);
924 }
925 
926 /*
927  *	vm_object_terminate actually destroys the specified object, freeing
928  *	up all previously used resources.
929  *
930  *	The object must be locked.
931  *	This routine may block.
932  */
933 void
934 vm_object_terminate(vm_object_t object)
935 {
936 
937 	VM_OBJECT_ASSERT_WLOCKED(object);
938 	KASSERT((object->flags & OBJ_DEAD) != 0,
939 	    ("terminating non-dead obj %p", object));
940 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
941 	    ("terminating collapsing obj %p", object));
942 	KASSERT(object->backing_object == NULL,
943 	    ("terminating shadow obj %p", object));
944 
945 	/*
946 	 * Wait for the pageout daemon and other current users to be
947 	 * done with the object.  Note that new paging_in_progress
948 	 * users can come after this wait, but they must check
949 	 * OBJ_DEAD flag set (without unlocking the object), and avoid
950 	 * the object being terminated.
951 	 */
952 	vm_object_pip_wait(object, "objtrm");
953 
954 	KASSERT(object->ref_count == 0,
955 	    ("vm_object_terminate: object with references, ref_count=%d",
956 	    object->ref_count));
957 
958 	if ((object->flags & OBJ_PG_DTOR) == 0)
959 		vm_object_terminate_pages(object);
960 
961 #if VM_NRESERVLEVEL > 0
962 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
963 		vm_reserv_break_all(object);
964 #endif
965 
966 	KASSERT(object->cred == NULL || (object->flags & OBJ_SWAP) != 0,
967 	    ("%s: non-swap obj %p has cred", __func__, object));
968 
969 	/*
970 	 * Let the pager know object is dead.
971 	 */
972 	vm_pager_deallocate(object);
973 	VM_OBJECT_WUNLOCK(object);
974 
975 	vm_object_destroy(object);
976 }
977 
978 /*
979  * Make the page read-only so that we can clear the object flags.  However, if
980  * this is a nosync mmap then the object is likely to stay dirty so do not
981  * mess with the page and do not clear the object flags.  Returns TRUE if the
982  * page should be flushed, and FALSE otherwise.
983  */
984 static boolean_t
985 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
986 {
987 
988 	vm_page_assert_busied(p);
989 
990 	/*
991 	 * If we have been asked to skip nosync pages and this is a
992 	 * nosync page, skip it.  Note that the object flags were not
993 	 * cleared in this case so we do not have to set them.
994 	 */
995 	if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
996 		*allclean = FALSE;
997 		return (FALSE);
998 	} else {
999 		pmap_remove_write(p);
1000 		return (p->dirty != 0);
1001 	}
1002 }
1003 
1004 /*
1005  *	vm_object_page_clean
1006  *
1007  *	Clean all dirty pages in the specified range of object.  Leaves page
1008  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
1009  *	write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1010  *	leaving the object dirty.
1011  *
1012  *	For swap objects backing tmpfs regular files, do not flush anything,
1013  *	but remove write protection on the mapped pages to update mtime through
1014  *	mmaped writes.
1015  *
1016  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
1017  *	synchronous clustering mode implementation.
1018  *
1019  *	Odd semantics: if start == end, we clean everything.
1020  *
1021  *	The object must be locked.
1022  *
1023  *	Returns FALSE if some page from the range was not written, as
1024  *	reported by the pager, and TRUE otherwise.
1025  */
1026 boolean_t
1027 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1028     int flags)
1029 {
1030 	struct pctrie_iter pages;
1031 	vm_page_t np, p;
1032 	vm_pindex_t pi, tend, tstart;
1033 	int curgeneration, n, pagerflags;
1034 	boolean_t eio, res, allclean;
1035 
1036 	VM_OBJECT_ASSERT_WLOCKED(object);
1037 
1038 	if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1039 		return (TRUE);
1040 
1041 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1042 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1043 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1044 
1045 	tstart = OFF_TO_IDX(start);
1046 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1047 	allclean = tstart == 0 && tend >= object->size;
1048 	res = TRUE;
1049 	vm_page_iter_init(&pages, object);
1050 
1051 rescan:
1052 	curgeneration = object->generation;
1053 
1054 	for (p = vm_radix_iter_lookup_ge(&pages, tstart); p != NULL; p = np) {
1055 		pi = p->pindex;
1056 		if (pi >= tend)
1057 			break;
1058 		if (vm_page_none_valid(p)) {
1059 			np = vm_radix_iter_step(&pages);
1060 			continue;
1061 		}
1062 		if (!vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL)) {
1063 			pctrie_iter_reset(&pages);
1064 			if (object->generation != curgeneration &&
1065 			    (flags & OBJPC_SYNC) != 0)
1066 				goto rescan;
1067 			np = vm_radix_iter_lookup_ge(&pages, pi);
1068 			continue;
1069 		}
1070 		if (!vm_object_page_remove_write(p, flags, &allclean)) {
1071 			np = vm_radix_iter_step(&pages);
1072 			vm_page_xunbusy(p);
1073 			continue;
1074 		}
1075 		if (object->type == OBJT_VNODE) {
1076 			n = vm_object_page_collect_flush(&pages, p, pagerflags,
1077 			    flags, &allclean, &eio);
1078 			pctrie_iter_reset(&pages);
1079 			if (eio) {
1080 				res = FALSE;
1081 				allclean = FALSE;
1082 			}
1083 			if (object->generation != curgeneration &&
1084 			    (flags & OBJPC_SYNC) != 0)
1085 				goto rescan;
1086 
1087 			/*
1088 			 * If the VOP_PUTPAGES() did a truncated write, so
1089 			 * that even the first page of the run is not fully
1090 			 * written, vm_pageout_flush() returns 0 as the run
1091 			 * length.  Since the condition that caused truncated
1092 			 * write may be permanent, e.g. exhausted free space,
1093 			 * accepting n == 0 would cause an infinite loop.
1094 			 *
1095 			 * Forwarding the iterator leaves the unwritten page
1096 			 * behind, but there is not much we can do there if
1097 			 * filesystem refuses to write it.
1098 			 */
1099 			if (n == 0) {
1100 				n = 1;
1101 				allclean = FALSE;
1102 			}
1103 		} else {
1104 			n = 1;
1105 			vm_page_xunbusy(p);
1106 		}
1107 		np = vm_radix_iter_lookup_ge(&pages, pi + n);
1108 	}
1109 #if 0
1110 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1111 #endif
1112 
1113 	/*
1114 	 * Leave updating cleangeneration for tmpfs objects to tmpfs
1115 	 * scan.  It needs to update mtime, which happens for other
1116 	 * filesystems during page writeouts.
1117 	 */
1118 	if (allclean && object->type == OBJT_VNODE)
1119 		object->cleangeneration = curgeneration;
1120 	return (res);
1121 }
1122 
1123 static int
1124 vm_object_page_collect_flush(struct pctrie_iter *pages, vm_page_t p,
1125     int pagerflags, int flags, boolean_t *allclean, boolean_t *eio)
1126 {
1127 	vm_page_t ma[2 * vm_pageout_page_count - 1];
1128 	int base, count, runlen;
1129 
1130 	vm_page_lock_assert(p, MA_NOTOWNED);
1131 	vm_page_assert_xbusied(p);
1132 	base = nitems(ma) / 2;
1133 	ma[base] = p;
1134 	for (count = 1; count < vm_pageout_page_count; count++) {
1135 		p = vm_radix_iter_next(pages);
1136 		if (p == NULL || vm_page_tryxbusy(p) == 0)
1137 			break;
1138 		if (!vm_object_page_remove_write(p, flags, allclean)) {
1139 			vm_page_xunbusy(p);
1140 			break;
1141 		}
1142 		ma[base + count] = p;
1143 	}
1144 
1145 	pages->index = ma[base]->pindex;
1146 	for (; count < vm_pageout_page_count; count++) {
1147 		p = vm_radix_iter_prev(pages);
1148 		if (p == NULL || vm_page_tryxbusy(p) == 0)
1149 			break;
1150 		if (!vm_object_page_remove_write(p, flags, allclean)) {
1151 			vm_page_xunbusy(p);
1152 			break;
1153 		}
1154 		ma[--base] = p;
1155 	}
1156 
1157 	vm_pageout_flush(&ma[base], count, pagerflags, nitems(ma) / 2 - base,
1158 	    &runlen, eio);
1159 	return (runlen);
1160 }
1161 
1162 /*
1163  * Note that there is absolutely no sense in writing out
1164  * anonymous objects, so we track down the vnode object
1165  * to write out.
1166  * We invalidate (remove) all pages from the address space
1167  * for semantic correctness.
1168  *
1169  * If the backing object is a device object with unmanaged pages, then any
1170  * mappings to the specified range of pages must be removed before this
1171  * function is called.
1172  *
1173  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1174  * may start out with a NULL object.
1175  */
1176 boolean_t
1177 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1178     boolean_t syncio, boolean_t invalidate)
1179 {
1180 	vm_object_t backing_object;
1181 	struct vnode *vp;
1182 	struct mount *mp;
1183 	int error, flags, fsync_after;
1184 	boolean_t res;
1185 
1186 	if (object == NULL)
1187 		return (TRUE);
1188 	res = TRUE;
1189 	error = 0;
1190 	VM_OBJECT_WLOCK(object);
1191 	while ((backing_object = object->backing_object) != NULL) {
1192 		VM_OBJECT_WLOCK(backing_object);
1193 		offset += object->backing_object_offset;
1194 		VM_OBJECT_WUNLOCK(object);
1195 		object = backing_object;
1196 		if (object->size < OFF_TO_IDX(offset + size))
1197 			size = IDX_TO_OFF(object->size) - offset;
1198 	}
1199 	/*
1200 	 * Flush pages if writing is allowed, invalidate them
1201 	 * if invalidation requested.  Pages undergoing I/O
1202 	 * will be ignored by vm_object_page_remove().
1203 	 *
1204 	 * We cannot lock the vnode and then wait for paging
1205 	 * to complete without deadlocking against vm_fault.
1206 	 * Instead we simply call vm_object_page_remove() and
1207 	 * allow it to block internally on a page-by-page
1208 	 * basis when it encounters pages undergoing async
1209 	 * I/O.
1210 	 */
1211 	if (object->type == OBJT_VNODE &&
1212 	    vm_object_mightbedirty(object) != 0 &&
1213 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1214 		VM_OBJECT_WUNLOCK(object);
1215 		(void)vn_start_write(vp, &mp, V_WAIT);
1216 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1217 		if (syncio && !invalidate && offset == 0 &&
1218 		    atop(size) == object->size) {
1219 			/*
1220 			 * If syncing the whole mapping of the file,
1221 			 * it is faster to schedule all the writes in
1222 			 * async mode, also allowing the clustering,
1223 			 * and then wait for i/o to complete.
1224 			 */
1225 			flags = 0;
1226 			fsync_after = TRUE;
1227 		} else {
1228 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1229 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1230 			fsync_after = FALSE;
1231 		}
1232 		VM_OBJECT_WLOCK(object);
1233 		res = vm_object_page_clean(object, offset, offset + size,
1234 		    flags);
1235 		VM_OBJECT_WUNLOCK(object);
1236 		if (fsync_after) {
1237 			for (;;) {
1238 				error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1239 				if (error != ERELOOKUP)
1240 					break;
1241 
1242 				/*
1243 				 * Allow SU/bufdaemon to handle more
1244 				 * dependencies in the meantime.
1245 				 */
1246 				VOP_UNLOCK(vp);
1247 				vn_finished_write(mp);
1248 
1249 				(void)vn_start_write(vp, &mp, V_WAIT);
1250 				vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1251 			}
1252 		}
1253 		VOP_UNLOCK(vp);
1254 		vn_finished_write(mp);
1255 		if (error != 0)
1256 			res = FALSE;
1257 		VM_OBJECT_WLOCK(object);
1258 	}
1259 	if ((object->type == OBJT_VNODE ||
1260 	     object->type == OBJT_DEVICE) && invalidate) {
1261 		if (object->type == OBJT_DEVICE)
1262 			/*
1263 			 * The option OBJPR_NOTMAPPED must be passed here
1264 			 * because vm_object_page_remove() cannot remove
1265 			 * unmanaged mappings.
1266 			 */
1267 			flags = OBJPR_NOTMAPPED;
1268 		else if (old_msync)
1269 			flags = 0;
1270 		else
1271 			flags = OBJPR_CLEANONLY;
1272 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1273 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1274 	}
1275 	VM_OBJECT_WUNLOCK(object);
1276 	return (res);
1277 }
1278 
1279 /*
1280  * Determine whether the given advice can be applied to the object.  Advice is
1281  * not applied to unmanaged pages since they never belong to page queues, and
1282  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1283  * have been mapped at most once.
1284  */
1285 static bool
1286 vm_object_advice_applies(vm_object_t object, int advice)
1287 {
1288 
1289 	if ((object->flags & OBJ_UNMANAGED) != 0)
1290 		return (false);
1291 	if (advice != MADV_FREE)
1292 		return (true);
1293 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1294 	    (OBJ_ONEMAPPING | OBJ_ANON));
1295 }
1296 
1297 static void
1298 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1299     vm_size_t size)
1300 {
1301 
1302 	if (advice == MADV_FREE)
1303 		vm_pager_freespace(object, pindex, size);
1304 }
1305 
1306 /*
1307  *	vm_object_madvise:
1308  *
1309  *	Implements the madvise function at the object/page level.
1310  *
1311  *	MADV_WILLNEED	(any object)
1312  *
1313  *	    Activate the specified pages if they are resident.
1314  *
1315  *	MADV_DONTNEED	(any object)
1316  *
1317  *	    Deactivate the specified pages if they are resident.
1318  *
1319  *	MADV_FREE	(OBJT_SWAP objects, OBJ_ONEMAPPING only)
1320  *
1321  *	    Deactivate and clean the specified pages if they are
1322  *	    resident.  This permits the process to reuse the pages
1323  *	    without faulting or the kernel to reclaim the pages
1324  *	    without I/O.
1325  */
1326 void
1327 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1328     int advice)
1329 {
1330 	struct pctrie_iter pages;
1331 	vm_pindex_t tpindex;
1332 	vm_object_t backing_object, tobject;
1333 	vm_page_t m, tm;
1334 
1335 	if (object == NULL)
1336 		return;
1337 
1338 	vm_page_iter_init(&pages, object);
1339 relookup:
1340 	VM_OBJECT_WLOCK(object);
1341 	if (!vm_object_advice_applies(object, advice)) {
1342 		VM_OBJECT_WUNLOCK(object);
1343 		return;
1344 	}
1345 	for (m = vm_radix_iter_lookup_ge(&pages, pindex); pindex < end;
1346 	    pindex++) {
1347 		tobject = object;
1348 
1349 		/*
1350 		 * If the next page isn't resident in the top-level object, we
1351 		 * need to search the shadow chain.  When applying MADV_FREE, we
1352 		 * take care to release any swap space used to store
1353 		 * non-resident pages.
1354 		 */
1355 		if (m == NULL || pindex < m->pindex) {
1356 			/*
1357 			 * Optimize a common case: if the top-level object has
1358 			 * no backing object, we can skip over the non-resident
1359 			 * range in constant time.
1360 			 */
1361 			if (object->backing_object == NULL) {
1362 				tpindex = (m != NULL && m->pindex < end) ?
1363 				    m->pindex : end;
1364 				vm_object_madvise_freespace(object, advice,
1365 				    pindex, tpindex - pindex);
1366 				if ((pindex = tpindex) == end)
1367 					break;
1368 				goto next_page;
1369 			}
1370 
1371 			tpindex = pindex;
1372 			do {
1373 				vm_object_madvise_freespace(tobject, advice,
1374 				    tpindex, 1);
1375 				/*
1376 				 * Prepare to search the next object in the
1377 				 * chain.
1378 				 */
1379 				backing_object = tobject->backing_object;
1380 				if (backing_object == NULL)
1381 					goto next_pindex;
1382 				VM_OBJECT_WLOCK(backing_object);
1383 				tpindex +=
1384 				    OFF_TO_IDX(tobject->backing_object_offset);
1385 				if (tobject != object)
1386 					VM_OBJECT_WUNLOCK(tobject);
1387 				tobject = backing_object;
1388 				if (!vm_object_advice_applies(tobject, advice))
1389 					goto next_pindex;
1390 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1391 			    NULL);
1392 		} else {
1393 next_page:
1394 			tm = m;
1395 			m = vm_radix_iter_step(&pages);
1396 		}
1397 
1398 		/*
1399 		 * If the page is not in a normal state, skip it.  The page
1400 		 * can not be invalidated while the object lock is held.
1401 		 */
1402 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1403 			goto next_pindex;
1404 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1405 		    ("vm_object_madvise: page %p is fictitious", tm));
1406 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1407 		    ("vm_object_madvise: page %p is not managed", tm));
1408 		if (vm_page_tryxbusy(tm) == 0) {
1409 			if (object != tobject)
1410 				VM_OBJECT_WUNLOCK(object);
1411 			if (advice == MADV_WILLNEED) {
1412 				/*
1413 				 * Reference the page before unlocking and
1414 				 * sleeping so that the page daemon is less
1415 				 * likely to reclaim it.
1416 				 */
1417 				vm_page_aflag_set(tm, PGA_REFERENCED);
1418 			}
1419 			if (!vm_page_busy_sleep(tm, "madvpo", 0))
1420 				VM_OBJECT_WUNLOCK(tobject);
1421 			pctrie_iter_reset(&pages);
1422   			goto relookup;
1423 		}
1424 		vm_page_advise(tm, advice);
1425 		vm_page_xunbusy(tm);
1426 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1427 next_pindex:
1428 		if (tobject != object)
1429 			VM_OBJECT_WUNLOCK(tobject);
1430 	}
1431 	VM_OBJECT_WUNLOCK(object);
1432 }
1433 
1434 /*
1435  *	vm_object_shadow:
1436  *
1437  *	Create a new object which is backed by the
1438  *	specified existing object range.  The source
1439  *	object reference is deallocated.
1440  *
1441  *	The new object and offset into that object
1442  *	are returned in the source parameters.
1443  */
1444 void
1445 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1446     struct ucred *cred, bool shared)
1447 {
1448 	vm_object_t source;
1449 	vm_object_t result;
1450 
1451 	source = *object;
1452 
1453 	/*
1454 	 * Don't create the new object if the old object isn't shared.
1455 	 *
1456 	 * If we hold the only reference we can guarantee that it won't
1457 	 * increase while we have the map locked.  Otherwise the race is
1458 	 * harmless and we will end up with an extra shadow object that
1459 	 * will be collapsed later.
1460 	 */
1461 	if (source != NULL && source->ref_count == 1 &&
1462 	    (source->flags & OBJ_ANON) != 0)
1463 		return;
1464 
1465 	/*
1466 	 * Allocate a new object with the given length.
1467 	 */
1468 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1469 
1470 	/*
1471 	 * Store the offset into the source object, and fix up the offset into
1472 	 * the new object.
1473 	 */
1474 	result->backing_object_offset = *offset;
1475 
1476 	if (shared || source != NULL) {
1477 		VM_OBJECT_WLOCK(result);
1478 
1479 		/*
1480 		 * The new object shadows the source object, adding a
1481 		 * reference to it.  Our caller changes his reference
1482 		 * to point to the new object, removing a reference to
1483 		 * the source object.  Net result: no change of
1484 		 * reference count, unless the caller needs to add one
1485 		 * more reference due to forking a shared map entry.
1486 		 */
1487 		if (shared) {
1488 			vm_object_reference_locked(result);
1489 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1490 		}
1491 
1492 		/*
1493 		 * Try to optimize the result object's page color when
1494 		 * shadowing in order to maintain page coloring
1495 		 * consistency in the combined shadowed object.
1496 		 */
1497 		if (source != NULL) {
1498 			vm_object_backing_insert(result, source);
1499 			result->domain = source->domain;
1500 #if VM_NRESERVLEVEL > 0
1501 			vm_object_set_flag(result,
1502 			    (source->flags & OBJ_COLORED));
1503 			result->pg_color = (source->pg_color +
1504 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1505 			    1)) - 1);
1506 #endif
1507 		}
1508 		VM_OBJECT_WUNLOCK(result);
1509 	}
1510 
1511 	/*
1512 	 * Return the new things
1513 	 */
1514 	*offset = 0;
1515 	*object = result;
1516 }
1517 
1518 /*
1519  *	vm_object_split:
1520  *
1521  * Split the pages in a map entry into a new object.  This affords
1522  * easier removal of unused pages, and keeps object inheritance from
1523  * being a negative impact on memory usage.
1524  */
1525 void
1526 vm_object_split(vm_map_entry_t entry)
1527 {
1528 	struct pctrie_iter pages;
1529 	vm_page_t m;
1530 	vm_object_t orig_object, new_object, backing_object;
1531 	vm_pindex_t offidxstart;
1532 	vm_size_t size;
1533 
1534 	orig_object = entry->object.vm_object;
1535 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1536 	    ("vm_object_split:  Splitting object with multiple mappings."));
1537 	if ((orig_object->flags & OBJ_ANON) == 0)
1538 		return;
1539 	if (orig_object->ref_count <= 1)
1540 		return;
1541 	VM_OBJECT_WUNLOCK(orig_object);
1542 
1543 	offidxstart = OFF_TO_IDX(entry->offset);
1544 	size = atop(entry->end - entry->start);
1545 
1546 	new_object = vm_object_allocate_anon(size, orig_object,
1547 	    orig_object->cred, ptoa(size));
1548 
1549 	/*
1550 	 * We must wait for the orig_object to complete any in-progress
1551 	 * collapse so that the swap blocks are stable below.  The
1552 	 * additional reference on backing_object by new object will
1553 	 * prevent further collapse operations until split completes.
1554 	 */
1555 	VM_OBJECT_WLOCK(orig_object);
1556 	vm_object_collapse_wait(orig_object);
1557 
1558 	/*
1559 	 * At this point, the new object is still private, so the order in
1560 	 * which the original and new objects are locked does not matter.
1561 	 */
1562 	VM_OBJECT_WLOCK(new_object);
1563 	new_object->domain = orig_object->domain;
1564 	backing_object = orig_object->backing_object;
1565 	if (backing_object != NULL) {
1566 		vm_object_backing_insert_ref(new_object, backing_object);
1567 		new_object->backing_object_offset =
1568 		    orig_object->backing_object_offset + entry->offset;
1569 	}
1570 	if (orig_object->cred != NULL) {
1571 		crhold(orig_object->cred);
1572 		KASSERT(orig_object->charge >= ptoa(size),
1573 		    ("orig_object->charge < 0"));
1574 		orig_object->charge -= ptoa(size);
1575 	}
1576 
1577 	/*
1578 	 * Mark the split operation so that swap_pager_getpages() knows
1579 	 * that the object is in transition.
1580 	 */
1581 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1582 	vm_page_iter_limit_init(&pages, orig_object, offidxstart + size);
1583 retry:
1584 	KASSERT(pctrie_iter_is_reset(&pages),
1585 	    ("%s: pctrie_iter not reset for retry", __func__));
1586 	for (m = vm_radix_iter_lookup_ge(&pages, offidxstart); m != NULL;
1587 	    m = vm_radix_iter_step(&pages)) {
1588 		/*
1589 		 * We must wait for pending I/O to complete before we can
1590 		 * rename the page.
1591 		 *
1592 		 * We do not have to VM_PROT_NONE the page as mappings should
1593 		 * not be changed by this operation.
1594 		 */
1595 		if (vm_page_tryxbusy(m) == 0) {
1596 			VM_OBJECT_WUNLOCK(new_object);
1597 			if (vm_page_busy_sleep(m, "spltwt", 0))
1598 				VM_OBJECT_WLOCK(orig_object);
1599 			pctrie_iter_reset(&pages);
1600 			VM_OBJECT_WLOCK(new_object);
1601 			goto retry;
1602 		}
1603 
1604 		/*
1605 		 * If the page was left invalid, it was likely placed there by
1606 		 * an incomplete fault.  Just remove and ignore.
1607 		 *
1608 		 * One other possibility is that the map entry is wired, in
1609 		 * which case we must hang on to the page to avoid leaking it,
1610 		 * as the map entry owns the wiring.  This case can arise if the
1611 		 * backing object is truncated by the pager.
1612 		 */
1613 		if (vm_page_none_valid(m) && entry->wired_count == 0) {
1614 			if (vm_page_iter_remove(&pages, m))
1615 				vm_page_free(m);
1616 			continue;
1617 		}
1618 
1619 		/* vm_page_iter_rename() will dirty the page if it is valid. */
1620 		if (!vm_page_iter_rename(&pages, m, new_object, m->pindex -
1621 		    offidxstart)) {
1622 			vm_page_xunbusy(m);
1623 			VM_OBJECT_WUNLOCK(new_object);
1624 			VM_OBJECT_WUNLOCK(orig_object);
1625 			vm_radix_wait();
1626 			pctrie_iter_reset(&pages);
1627 			VM_OBJECT_WLOCK(orig_object);
1628 			VM_OBJECT_WLOCK(new_object);
1629 			goto retry;
1630 		}
1631 
1632 #if VM_NRESERVLEVEL > 0
1633 		/*
1634 		 * If some of the reservation's allocated pages remain with
1635 		 * the original object, then transferring the reservation to
1636 		 * the new object is neither particularly beneficial nor
1637 		 * particularly harmful as compared to leaving the reservation
1638 		 * with the original object.  If, however, all of the
1639 		 * reservation's allocated pages are transferred to the new
1640 		 * object, then transferring the reservation is typically
1641 		 * beneficial.  Determining which of these two cases applies
1642 		 * would be more costly than unconditionally renaming the
1643 		 * reservation.
1644 		 */
1645 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1646 #endif
1647 	}
1648 
1649 	/*
1650 	 * swap_pager_copy() can sleep, in which case the orig_object's
1651 	 * and new_object's locks are released and reacquired.
1652 	 */
1653 	swap_pager_copy(orig_object, new_object, offidxstart, 0);
1654 	vm_page_iter_init(&pages, new_object);
1655 	VM_RADIX_FOREACH(m, &pages)
1656 		vm_page_xunbusy(m);
1657 
1658 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1659 	VM_OBJECT_WUNLOCK(orig_object);
1660 	VM_OBJECT_WUNLOCK(new_object);
1661 	entry->object.vm_object = new_object;
1662 	entry->offset = 0LL;
1663 	vm_object_deallocate(orig_object);
1664 	VM_OBJECT_WLOCK(new_object);
1665 }
1666 
1667 static vm_page_t
1668 vm_object_collapse_scan_wait(struct pctrie_iter *pages, vm_object_t object,
1669     vm_page_t p)
1670 {
1671 	vm_object_t backing_object;
1672 
1673 	VM_OBJECT_ASSERT_WLOCKED(object);
1674 	backing_object = object->backing_object;
1675 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1676 
1677 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1678 	    ("invalid ownership %p %p %p", p, object, backing_object));
1679 	/* The page is only NULL when rename fails. */
1680 	if (p == NULL) {
1681 		VM_OBJECT_WUNLOCK(object);
1682 		VM_OBJECT_WUNLOCK(backing_object);
1683 		vm_radix_wait();
1684 		VM_OBJECT_WLOCK(object);
1685 	} else if (p->object == object) {
1686 		VM_OBJECT_WUNLOCK(backing_object);
1687 		if (vm_page_busy_sleep(p, "vmocol", 0))
1688 			VM_OBJECT_WLOCK(object);
1689 	} else {
1690 		VM_OBJECT_WUNLOCK(object);
1691 		if (!vm_page_busy_sleep(p, "vmocol", 0))
1692 			VM_OBJECT_WUNLOCK(backing_object);
1693 		VM_OBJECT_WLOCK(object);
1694 	}
1695 	VM_OBJECT_WLOCK(backing_object);
1696 	vm_page_iter_init(pages, backing_object);
1697 	return (vm_radix_iter_lookup_ge(pages, 0));
1698 }
1699 
1700 static void
1701 vm_object_collapse_scan(vm_object_t object)
1702 {
1703 	struct pctrie_iter pages;
1704 	vm_object_t backing_object;
1705 	vm_page_t next, p, pp;
1706 	vm_pindex_t backing_offset_index, new_pindex;
1707 
1708 	VM_OBJECT_ASSERT_WLOCKED(object);
1709 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1710 
1711 	backing_object = object->backing_object;
1712 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1713 
1714 	/*
1715 	 * Our scan
1716 	 */
1717 	vm_page_iter_init(&pages, backing_object);
1718 	for (p = vm_radix_iter_lookup_ge(&pages, 0); p != NULL; p = next) {
1719 		/*
1720 		 * Check for busy page
1721 		 */
1722 		if (vm_page_tryxbusy(p) == 0) {
1723 			next = vm_object_collapse_scan_wait(&pages, object, p);
1724 			continue;
1725 		}
1726 
1727 		KASSERT(object->backing_object == backing_object,
1728 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1729 		    object->backing_object, backing_object));
1730 		KASSERT(p->object == backing_object,
1731 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1732 		    p->object, backing_object));
1733 
1734 		if (p->pindex < backing_offset_index || object->size <=
1735 		    (new_pindex = p->pindex - backing_offset_index)) {
1736 			vm_pager_freespace(backing_object, p->pindex, 1);
1737 
1738 			KASSERT(!pmap_page_is_mapped(p),
1739 			    ("freeing mapped page %p", p));
1740 			if (vm_page_iter_remove(&pages, p))
1741 				vm_page_free(p);
1742 			next = vm_radix_iter_step(&pages);
1743 			continue;
1744 		}
1745 
1746 		if (!vm_page_all_valid(p)) {
1747 			KASSERT(!pmap_page_is_mapped(p),
1748 			    ("freeing mapped page %p", p));
1749 			if (vm_page_iter_remove(&pages, p))
1750 				vm_page_free(p);
1751 			next = vm_radix_iter_step(&pages);
1752 			continue;
1753 		}
1754 
1755 		pp = vm_page_lookup(object, new_pindex);
1756 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1757 			vm_page_xunbusy(p);
1758 			/*
1759 			 * The page in the parent is busy and possibly not
1760 			 * (yet) valid.  Until its state is finalized by the
1761 			 * busy bit owner, we can't tell whether it shadows the
1762 			 * original page.
1763 			 */
1764 			next = vm_object_collapse_scan_wait(&pages, object, pp);
1765 			continue;
1766 		}
1767 
1768 		if (pp != NULL && vm_page_none_valid(pp)) {
1769 			/*
1770 			 * The page was invalid in the parent.  Likely placed
1771 			 * there by an incomplete fault.  Just remove and
1772 			 * ignore.  p can replace it.
1773 			 */
1774 			if (vm_page_remove(pp))
1775 				vm_page_free(pp);
1776 			pp = NULL;
1777 		}
1778 
1779 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1780 			NULL)) {
1781 			/*
1782 			 * The page already exists in the parent OR swap exists
1783 			 * for this location in the parent.  Leave the parent's
1784 			 * page alone.  Destroy the original page from the
1785 			 * backing object.
1786 			 */
1787 			vm_pager_freespace(backing_object, p->pindex, 1);
1788 			KASSERT(!pmap_page_is_mapped(p),
1789 			    ("freeing mapped page %p", p));
1790 			if (pp != NULL)
1791 				vm_page_xunbusy(pp);
1792 			if (vm_page_iter_remove(&pages, p))
1793 				vm_page_free(p);
1794 			next = vm_radix_iter_step(&pages);
1795 			continue;
1796 		}
1797 
1798 		/*
1799 		 * Page does not exist in parent, rename the page from the
1800 		 * backing object to the main object.
1801 		 *
1802 		 * If the page was mapped to a process, it can remain mapped
1803 		 * through the rename.  vm_page_iter_rename() will dirty the
1804 		 * page.
1805 		 */
1806 		if (!vm_page_iter_rename(&pages, p, object, new_pindex)) {
1807 			vm_page_xunbusy(p);
1808 			next = vm_object_collapse_scan_wait(&pages, object,
1809 			    NULL);
1810 			continue;
1811 		}
1812 
1813 		/* Use the old pindex to free the right page. */
1814 		vm_pager_freespace(backing_object, new_pindex +
1815 		    backing_offset_index, 1);
1816 
1817 #if VM_NRESERVLEVEL > 0
1818 		/*
1819 		 * Rename the reservation.
1820 		 */
1821 		vm_reserv_rename(p, object, backing_object,
1822 		    backing_offset_index);
1823 #endif
1824 		vm_page_xunbusy(p);
1825 		next = vm_radix_iter_step(&pages);
1826 	}
1827 	return;
1828 }
1829 
1830 /*
1831  *	vm_object_collapse:
1832  *
1833  *	Collapse an object with the object backing it.
1834  *	Pages in the backing object are moved into the
1835  *	parent, and the backing object is deallocated.
1836  */
1837 void
1838 vm_object_collapse(vm_object_t object)
1839 {
1840 	vm_object_t backing_object, new_backing_object;
1841 
1842 	VM_OBJECT_ASSERT_WLOCKED(object);
1843 
1844 	while (TRUE) {
1845 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1846 		    ("collapsing invalid object"));
1847 
1848 		/*
1849 		 * Wait for the backing_object to finish any pending
1850 		 * collapse so that the caller sees the shortest possible
1851 		 * shadow chain.
1852 		 */
1853 		backing_object = vm_object_backing_collapse_wait(object);
1854 		if (backing_object == NULL)
1855 			return;
1856 
1857 		KASSERT(object->ref_count > 0 &&
1858 		    object->ref_count > atomic_load_int(&object->shadow_count),
1859 		    ("collapse with invalid ref %d or shadow %d count.",
1860 		    object->ref_count, atomic_load_int(&object->shadow_count)));
1861 		KASSERT((backing_object->flags &
1862 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1863 		    ("vm_object_collapse: Backing object already collapsing."));
1864 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1865 		    ("vm_object_collapse: object is already collapsing."));
1866 
1867 		/*
1868 		 * We know that we can either collapse the backing object if
1869 		 * the parent is the only reference to it, or (perhaps) have
1870 		 * the parent bypass the object if the parent happens to shadow
1871 		 * all the resident pages in the entire backing object.
1872 		 */
1873 		if (backing_object->ref_count == 1) {
1874 			KASSERT(atomic_load_int(&backing_object->shadow_count)
1875 			    == 1,
1876 			    ("vm_object_collapse: shadow_count: %d",
1877 			    atomic_load_int(&backing_object->shadow_count)));
1878 			vm_object_pip_add(object, 1);
1879 			vm_object_set_flag(object, OBJ_COLLAPSING);
1880 			vm_object_pip_add(backing_object, 1);
1881 			vm_object_set_flag(backing_object, OBJ_DEAD);
1882 
1883 			/*
1884 			 * If there is exactly one reference to the backing
1885 			 * object, we can collapse it into the parent.
1886 			 */
1887 			vm_object_collapse_scan(object);
1888 
1889 			/*
1890 			 * Move the pager from backing_object to object.
1891 			 *
1892 			 * swap_pager_copy() can sleep, in which case the
1893 			 * backing_object's and object's locks are released and
1894 			 * reacquired.
1895 			 */
1896 			swap_pager_copy(backing_object, object,
1897 			    OFF_TO_IDX(object->backing_object_offset), TRUE);
1898 
1899 			/*
1900 			 * Object now shadows whatever backing_object did.
1901 			 */
1902 			vm_object_clear_flag(object, OBJ_COLLAPSING);
1903 			vm_object_backing_transfer(object, backing_object);
1904 			object->backing_object_offset +=
1905 			    backing_object->backing_object_offset;
1906 			VM_OBJECT_WUNLOCK(object);
1907 			vm_object_pip_wakeup(object);
1908 
1909 			/*
1910 			 * Discard backing_object.
1911 			 *
1912 			 * Since the backing object has no pages, no pager left,
1913 			 * and no object references within it, all that is
1914 			 * necessary is to dispose of it.
1915 			 */
1916 			KASSERT(backing_object->ref_count == 1, (
1917 "backing_object %p was somehow re-referenced during collapse!",
1918 			    backing_object));
1919 			vm_object_pip_wakeup(backing_object);
1920 			(void)refcount_release(&backing_object->ref_count);
1921 			umtx_shm_object_terminated(backing_object);
1922 			vm_object_terminate(backing_object);
1923 			counter_u64_add(object_collapses, 1);
1924 			VM_OBJECT_WLOCK(object);
1925 		} else {
1926 			/*
1927 			 * If we do not entirely shadow the backing object,
1928 			 * there is nothing we can do so we give up.
1929 			 *
1930 			 * The object lock and backing_object lock must not
1931 			 * be dropped during this sequence.
1932 			 */
1933 			if (!swap_pager_scan_all_shadowed(object)) {
1934 				VM_OBJECT_WUNLOCK(backing_object);
1935 				break;
1936 			}
1937 
1938 			/*
1939 			 * Make the parent shadow the next object in the
1940 			 * chain.  Deallocating backing_object will not remove
1941 			 * it, since its reference count is at least 2.
1942 			 */
1943 			vm_object_backing_remove_locked(object);
1944 			new_backing_object = backing_object->backing_object;
1945 			if (new_backing_object != NULL) {
1946 				vm_object_backing_insert_ref(object,
1947 				    new_backing_object);
1948 				object->backing_object_offset +=
1949 				    backing_object->backing_object_offset;
1950 			}
1951 
1952 			/*
1953 			 * Drop the reference count on backing_object. Since
1954 			 * its ref_count was at least 2, it will not vanish.
1955 			 */
1956 			(void)refcount_release(&backing_object->ref_count);
1957 			KASSERT(backing_object->ref_count >= 1, (
1958 "backing_object %p was somehow dereferenced during collapse!",
1959 			    backing_object));
1960 			VM_OBJECT_WUNLOCK(backing_object);
1961 			counter_u64_add(object_bypasses, 1);
1962 		}
1963 
1964 		/*
1965 		 * Try again with this object's new backing object.
1966 		 */
1967 	}
1968 }
1969 
1970 /*
1971  *	vm_object_page_remove:
1972  *
1973  *	For the given object, either frees or invalidates each of the
1974  *	specified pages.  In general, a page is freed.  However, if a page is
1975  *	wired for any reason other than the existence of a managed, wired
1976  *	mapping, then it may be invalidated but not removed from the object.
1977  *	Pages are specified by the given range ["start", "end") and the option
1978  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
1979  *	extends from "start" to the end of the object.  If the option
1980  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1981  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
1982  *	specified, then the pages within the specified range must have no
1983  *	mappings.  Otherwise, if this option is not specified, any mappings to
1984  *	the specified pages are removed before the pages are freed or
1985  *	invalidated.
1986  *
1987  *	In general, this operation should only be performed on objects that
1988  *	contain managed pages.  There are, however, two exceptions.  First, it
1989  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
1990  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1991  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
1992  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
1993  *
1994  *	The object must be locked.
1995  */
1996 void
1997 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1998     int options)
1999 {
2000 	struct pctrie_iter pages;
2001 	vm_page_t p;
2002 
2003 	VM_OBJECT_ASSERT_WLOCKED(object);
2004 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2005 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2006 	    ("vm_object_page_remove: illegal options for object %p", object));
2007 	if (object->resident_page_count == 0)
2008 		return;
2009 	vm_object_pip_add(object, 1);
2010 	vm_page_iter_limit_init(&pages, object, end);
2011 again:
2012 	KASSERT(pctrie_iter_is_reset(&pages),
2013 	    ("%s: pctrie_iter not reset for retry", __func__));
2014 	for (p = vm_radix_iter_lookup_ge(&pages, start); p != NULL;
2015 	     p = vm_radix_iter_step(&pages)) {
2016 		/*
2017 		 * Skip invalid pages if asked to do so.  Try to avoid acquiring
2018 		 * the busy lock, as some consumers rely on this to avoid
2019 		 * deadlocks.
2020 		 *
2021 		 * A thread may concurrently transition the page from invalid to
2022 		 * valid using only the busy lock, so the result of this check
2023 		 * is immediately stale.  It is up to consumers to handle this,
2024 		 * for instance by ensuring that all invalid->valid transitions
2025 		 * happen with a mutex held, as may be possible for a
2026 		 * filesystem.
2027 		 */
2028 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2029 			continue;
2030 
2031 		/*
2032 		 * If the page is wired for any reason besides the existence
2033 		 * of managed, wired mappings, then it cannot be freed.  For
2034 		 * example, fictitious pages, which represent device memory,
2035 		 * are inherently wired and cannot be freed.  They can,
2036 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2037 		 * not specified.
2038 		 */
2039 		if (vm_page_tryxbusy(p) == 0) {
2040 			if (vm_page_busy_sleep(p, "vmopar", 0))
2041 				VM_OBJECT_WLOCK(object);
2042 			pctrie_iter_reset(&pages);
2043 			goto again;
2044 		}
2045 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2046 			vm_page_xunbusy(p);
2047 			continue;
2048 		}
2049 		if (vm_page_wired(p)) {
2050 wired:
2051 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2052 			    object->ref_count != 0)
2053 				pmap_remove_all(p);
2054 			if ((options & OBJPR_CLEANONLY) == 0) {
2055 				vm_page_invalid(p);
2056 				vm_page_undirty(p);
2057 			}
2058 			vm_page_xunbusy(p);
2059 			continue;
2060 		}
2061 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2062 		    ("vm_object_page_remove: page %p is fictitious", p));
2063 		if ((options & OBJPR_CLEANONLY) != 0 &&
2064 		    !vm_page_none_valid(p)) {
2065 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2066 			    object->ref_count != 0 &&
2067 			    !vm_page_try_remove_write(p))
2068 				goto wired;
2069 			if (p->dirty != 0) {
2070 				vm_page_xunbusy(p);
2071 				continue;
2072 			}
2073 		}
2074 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2075 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2076 			goto wired;
2077 		vm_page_iter_free(&pages, p);
2078 	}
2079 	vm_object_pip_wakeup(object);
2080 
2081 	vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2082 	    start);
2083 }
2084 
2085 /*
2086  *	vm_object_page_noreuse:
2087  *
2088  *	For the given object, attempt to move the specified pages to
2089  *	the head of the inactive queue.  This bypasses regular LRU
2090  *	operation and allows the pages to be reused quickly under memory
2091  *	pressure.  If a page is wired for any reason, then it will not
2092  *	be queued.  Pages are specified by the range ["start", "end").
2093  *	As a special case, if "end" is zero, then the range extends from
2094  *	"start" to the end of the object.
2095  *
2096  *	This operation should only be performed on objects that
2097  *	contain non-fictitious, managed pages.
2098  *
2099  *	The object must be locked.
2100  */
2101 void
2102 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2103 {
2104 	struct pctrie_iter pages;
2105 	vm_page_t p;
2106 
2107 	VM_OBJECT_ASSERT_LOCKED(object);
2108 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2109 	    ("vm_object_page_noreuse: illegal object %p", object));
2110 	if (object->resident_page_count == 0)
2111 		return;
2112 
2113 	vm_page_iter_limit_init(&pages, object, end);
2114 	VM_RADIX_FOREACH_FROM(p, &pages, start)
2115 		vm_page_deactivate_noreuse(p);
2116 }
2117 
2118 /*
2119  *	Populate the specified range of the object with valid pages.  Returns
2120  *	TRUE if the range is successfully populated and FALSE otherwise.
2121  *
2122  *	Note: This function should be optimized to pass a larger array of
2123  *	pages to vm_pager_get_pages() before it is applied to a non-
2124  *	OBJT_DEVICE object.
2125  *
2126  *	The object must be locked.
2127  */
2128 boolean_t
2129 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2130 {
2131 	struct pctrie_iter pages;
2132 	vm_page_t m;
2133 	vm_pindex_t pindex;
2134 	int rv;
2135 
2136 	vm_page_iter_init(&pages, object);
2137 	VM_OBJECT_ASSERT_WLOCKED(object);
2138 	for (pindex = start; pindex < end; pindex++) {
2139 		rv = vm_page_grab_valid_iter(&m, object, pindex,
2140 		    VM_ALLOC_NORMAL, &pages);
2141 		if (rv != VM_PAGER_OK)
2142 			break;
2143 
2144 		/*
2145 		 * Keep "m" busy because a subsequent iteration may unlock
2146 		 * the object.
2147 		 */
2148 	}
2149 	if (pindex > start) {
2150 		pages.limit = pindex;
2151 		VM_RADIX_FORALL_FROM(m, &pages, start)
2152 			vm_page_xunbusy(m);
2153 	}
2154 	return (pindex == end);
2155 }
2156 
2157 /*
2158  *	Routine:	vm_object_coalesce
2159  *	Function:	Coalesces two objects backing up adjoining
2160  *			regions of memory into a single object.
2161  *
2162  *	returns TRUE if objects were combined.
2163  *
2164  *	NOTE:	Only works at the moment if the second object is NULL -
2165  *		if it's not, which object do we lock first?
2166  *
2167  *	Parameters:
2168  *		prev_object	First object to coalesce
2169  *		prev_offset	Offset into prev_object
2170  *		prev_size	Size of reference to prev_object
2171  *		next_size	Size of reference to the second object
2172  *		reserved	Indicator that extension region has
2173  *				swap accounted for
2174  *
2175  *	Conditions:
2176  *	The object must *not* be locked.
2177  */
2178 boolean_t
2179 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2180     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2181 {
2182 	vm_pindex_t next_pindex;
2183 
2184 	if (prev_object == NULL)
2185 		return (TRUE);
2186 	if ((prev_object->flags & OBJ_ANON) == 0)
2187 		return (FALSE);
2188 
2189 	VM_OBJECT_WLOCK(prev_object);
2190 	/*
2191 	 * Try to collapse the object first.
2192 	 */
2193 	vm_object_collapse(prev_object);
2194 
2195 	/*
2196 	 * Can't coalesce if: . more than one reference . paged out . shadows
2197 	 * another object . has a copy elsewhere (any of which mean that the
2198 	 * pages not mapped to prev_entry may be in use anyway)
2199 	 */
2200 	if (prev_object->backing_object != NULL) {
2201 		VM_OBJECT_WUNLOCK(prev_object);
2202 		return (FALSE);
2203 	}
2204 
2205 	prev_size >>= PAGE_SHIFT;
2206 	next_size >>= PAGE_SHIFT;
2207 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2208 
2209 	if (prev_object->ref_count > 1 &&
2210 	    prev_object->size != next_pindex &&
2211 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2212 		VM_OBJECT_WUNLOCK(prev_object);
2213 		return (FALSE);
2214 	}
2215 
2216 	/*
2217 	 * Account for the charge.
2218 	 */
2219 	if (prev_object->cred != NULL) {
2220 		/*
2221 		 * If prev_object was charged, then this mapping,
2222 		 * although not charged now, may become writable
2223 		 * later. Non-NULL cred in the object would prevent
2224 		 * swap reservation during enabling of the write
2225 		 * access, so reserve swap now. Failed reservation
2226 		 * cause allocation of the separate object for the map
2227 		 * entry, and swap reservation for this entry is
2228 		 * managed in appropriate time.
2229 		 */
2230 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2231 		    prev_object->cred)) {
2232 			VM_OBJECT_WUNLOCK(prev_object);
2233 			return (FALSE);
2234 		}
2235 		prev_object->charge += ptoa(next_size);
2236 	}
2237 
2238 	/*
2239 	 * Remove any pages that may still be in the object from a previous
2240 	 * deallocation.
2241 	 */
2242 	if (next_pindex < prev_object->size) {
2243 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2244 		    next_size, 0);
2245 #if 0
2246 		if (prev_object->cred != NULL) {
2247 			KASSERT(prev_object->charge >=
2248 			    ptoa(prev_object->size - next_pindex),
2249 			    ("object %p overcharged 1 %jx %jx", prev_object,
2250 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2251 			prev_object->charge -= ptoa(prev_object->size -
2252 			    next_pindex);
2253 		}
2254 #endif
2255 	}
2256 
2257 	/*
2258 	 * Extend the object if necessary.
2259 	 */
2260 	if (next_pindex + next_size > prev_object->size)
2261 		prev_object->size = next_pindex + next_size;
2262 
2263 	VM_OBJECT_WUNLOCK(prev_object);
2264 	return (TRUE);
2265 }
2266 
2267 /*
2268  * Fill in the m_dst array with up to *rbehind optional pages before m_src[0]
2269  * and up to *rahead optional pages after m_src[count - 1].  In both cases, stop
2270  * the filling-in short on encountering a cached page, an object boundary limit,
2271  * or an allocation error.  Update *rbehind and *rahead to indicate the number
2272  * of pages allocated.  Copy elements of m_src into array elements from
2273  * m_dst[*rbehind] to m_dst[*rbehind + count -1].
2274  */
2275 void
2276 vm_object_prepare_buf_pages(vm_object_t object, vm_page_t *ma_dst, int count,
2277     int *rbehind, int *rahead, vm_page_t *ma_src)
2278 {
2279 	struct pctrie_iter pages;
2280 	vm_pindex_t pindex;
2281 	vm_page_t m, mpred, msucc;
2282 
2283 	vm_page_iter_init(&pages, object);
2284 	VM_OBJECT_ASSERT_LOCKED(object);
2285 	if (*rbehind != 0) {
2286 		m = ma_src[0];
2287 		pindex = m->pindex;
2288 		mpred = vm_radix_iter_lookup_lt(&pages, pindex);
2289 		*rbehind = MIN(*rbehind,
2290 		    pindex - (mpred != NULL ? mpred->pindex + 1 : 0));
2291 		for (int i = 0; i < *rbehind; i++) {
2292 			m = vm_page_alloc_iter(object, pindex - i - 1,
2293 			    VM_ALLOC_NORMAL, &pages);
2294 			if (m == NULL) {
2295 				/* Shift the array. */
2296 				for (int j = 0; j < i; j++)
2297 					ma_dst[j] = ma_dst[j + *rbehind - i];
2298 				*rbehind = i;
2299 				*rahead = 0;
2300 				break;
2301 			}
2302 			ma_dst[*rbehind - i - 1] = m;
2303 		}
2304 	}
2305 	for (int i = 0; i < count; i++)
2306 		ma_dst[*rbehind + i] = ma_src[i];
2307 	if (*rahead != 0) {
2308 		m = ma_src[count - 1];
2309 		pindex = m->pindex + 1;
2310 		msucc = vm_radix_iter_lookup_ge(&pages, pindex);
2311 		*rahead = MIN(*rahead,
2312 		    (msucc != NULL ? msucc->pindex : object->size) - pindex);
2313 		for (int i = 0; i < *rahead; i++) {
2314 			m = vm_page_alloc_iter(object, pindex + i,
2315 			    VM_ALLOC_NORMAL, &pages);
2316 			if (m == NULL) {
2317 				*rahead = i;
2318 				break;
2319 			}
2320 			ma_dst[*rbehind + count + i] = m;
2321 		}
2322 	}
2323 }
2324 
2325 void
2326 vm_object_set_writeable_dirty_(vm_object_t object)
2327 {
2328 	atomic_add_int(&object->generation, 1);
2329 }
2330 
2331 bool
2332 vm_object_mightbedirty_(vm_object_t object)
2333 {
2334 	return (object->generation != object->cleangeneration);
2335 }
2336 
2337 /*
2338  *	vm_object_unwire:
2339  *
2340  *	For each page offset within the specified range of the given object,
2341  *	find the highest-level page in the shadow chain and unwire it.  A page
2342  *	must exist at every page offset, and the highest-level page must be
2343  *	wired.
2344  */
2345 void
2346 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2347     uint8_t queue)
2348 {
2349 	struct pctrie_iter pages;
2350 	vm_object_t tobject, t1object;
2351 	vm_page_t m, tm;
2352 	vm_pindex_t end_pindex, pindex, tpindex;
2353 	int depth, locked_depth;
2354 
2355 	KASSERT((offset & PAGE_MASK) == 0,
2356 	    ("vm_object_unwire: offset is not page aligned"));
2357 	KASSERT((length & PAGE_MASK) == 0,
2358 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2359 	/* The wired count of a fictitious page never changes. */
2360 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2361 		return;
2362 	pindex = OFF_TO_IDX(offset);
2363 	end_pindex = pindex + atop(length);
2364 	vm_page_iter_init(&pages, object);
2365 again:
2366 	locked_depth = 1;
2367 	VM_OBJECT_RLOCK(object);
2368 	m = vm_radix_iter_lookup_ge(&pages, pindex);
2369 	while (pindex < end_pindex) {
2370 		if (m == NULL || pindex < m->pindex) {
2371 			/*
2372 			 * The first object in the shadow chain doesn't
2373 			 * contain a page at the current index.  Therefore,
2374 			 * the page must exist in a backing object.
2375 			 */
2376 			tobject = object;
2377 			tpindex = pindex;
2378 			depth = 0;
2379 			do {
2380 				tpindex +=
2381 				    OFF_TO_IDX(tobject->backing_object_offset);
2382 				tobject = tobject->backing_object;
2383 				KASSERT(tobject != NULL,
2384 				    ("vm_object_unwire: missing page"));
2385 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2386 					goto next_page;
2387 				depth++;
2388 				if (depth == locked_depth) {
2389 					locked_depth++;
2390 					VM_OBJECT_RLOCK(tobject);
2391 				}
2392 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2393 			    NULL);
2394 		} else {
2395 			tm = m;
2396 			m = vm_radix_iter_step(&pages);
2397 		}
2398 		if (vm_page_trysbusy(tm) == 0) {
2399 			for (tobject = object; locked_depth >= 1;
2400 			    locked_depth--) {
2401 				t1object = tobject->backing_object;
2402 				if (tm->object != tobject)
2403 					VM_OBJECT_RUNLOCK(tobject);
2404 				tobject = t1object;
2405 			}
2406 			tobject = tm->object;
2407 			if (!vm_page_busy_sleep(tm, "unwbo",
2408 			    VM_ALLOC_IGN_SBUSY))
2409 				VM_OBJECT_RUNLOCK(tobject);
2410 			pctrie_iter_reset(&pages);
2411 			goto again;
2412 		}
2413 		vm_page_unwire(tm, queue);
2414 		vm_page_sunbusy(tm);
2415 next_page:
2416 		pindex++;
2417 	}
2418 	/* Release the accumulated object locks. */
2419 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2420 		t1object = tobject->backing_object;
2421 		VM_OBJECT_RUNLOCK(tobject);
2422 		tobject = t1object;
2423 	}
2424 }
2425 
2426 /*
2427  * Return the vnode for the given object, or NULL if none exists.
2428  * For tmpfs objects, the function may return NULL if there is
2429  * no vnode allocated at the time of the call.
2430  */
2431 struct vnode *
2432 vm_object_vnode(vm_object_t object)
2433 {
2434 	struct vnode *vp;
2435 
2436 	VM_OBJECT_ASSERT_LOCKED(object);
2437 	vm_pager_getvp(object, &vp, NULL);
2438 	return (vp);
2439 }
2440 
2441 /*
2442  * Busy the vm object.  This prevents new pages belonging to the object from
2443  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2444  * for checking page state before proceeding.
2445  */
2446 void
2447 vm_object_busy(vm_object_t obj)
2448 {
2449 
2450 	VM_OBJECT_ASSERT_LOCKED(obj);
2451 
2452 	blockcount_acquire(&obj->busy, 1);
2453 	/* The fence is required to order loads of page busy. */
2454 	atomic_thread_fence_acq_rel();
2455 }
2456 
2457 void
2458 vm_object_unbusy(vm_object_t obj)
2459 {
2460 
2461 	blockcount_release(&obj->busy, 1);
2462 }
2463 
2464 void
2465 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2466 {
2467 
2468 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2469 
2470 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2471 }
2472 
2473 /*
2474  * This function aims to determine if the object is mapped,
2475  * specifically, if it is referenced by a vm_map_entry.  Because
2476  * objects occasionally acquire transient references that do not
2477  * represent a mapping, the method used here is inexact.  However, it
2478  * has very low overhead and is good enough for the advisory
2479  * vm.vmtotal sysctl.
2480  */
2481 bool
2482 vm_object_is_active(vm_object_t obj)
2483 {
2484 
2485 	return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2486 }
2487 
2488 static int
2489 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2490 {
2491 	struct pctrie_iter pages;
2492 	struct kinfo_vmobject *kvo;
2493 	char *fullpath, *freepath;
2494 	struct vnode *vp;
2495 	struct vattr va;
2496 	vm_object_t obj;
2497 	vm_page_t m;
2498 	u_long sp;
2499 	int count, error;
2500 	key_t key;
2501 	unsigned short seq;
2502 	bool want_path;
2503 
2504 	if (req->oldptr == NULL) {
2505 		/*
2506 		 * If an old buffer has not been provided, generate an
2507 		 * estimate of the space needed for a subsequent call.
2508 		 */
2509 		mtx_lock(&vm_object_list_mtx);
2510 		count = 0;
2511 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2512 			if (obj->type == OBJT_DEAD)
2513 				continue;
2514 			count++;
2515 		}
2516 		mtx_unlock(&vm_object_list_mtx);
2517 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2518 		    count * 11 / 10));
2519 	}
2520 
2521 	want_path = !(swap_only || jailed(curthread->td_ucred));
2522 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO);
2523 	error = 0;
2524 
2525 	/*
2526 	 * VM objects are type stable and are never removed from the
2527 	 * list once added.  This allows us to safely read obj->object_list
2528 	 * after reacquiring the VM object lock.
2529 	 */
2530 	mtx_lock(&vm_object_list_mtx);
2531 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2532 		if (obj->type == OBJT_DEAD ||
2533 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2534 			continue;
2535 		VM_OBJECT_RLOCK(obj);
2536 		if (obj->type == OBJT_DEAD ||
2537 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2538 			VM_OBJECT_RUNLOCK(obj);
2539 			continue;
2540 		}
2541 		mtx_unlock(&vm_object_list_mtx);
2542 		kvo->kvo_size = ptoa(obj->size);
2543 		kvo->kvo_resident = obj->resident_page_count;
2544 		kvo->kvo_ref_count = obj->ref_count;
2545 		kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2546 		kvo->kvo_memattr = obj->memattr;
2547 		kvo->kvo_active = 0;
2548 		kvo->kvo_inactive = 0;
2549 		kvo->kvo_flags = 0;
2550 		if (!swap_only) {
2551 			vm_page_iter_init(&pages, obj);
2552 			VM_RADIX_FOREACH(m, &pages) {
2553 				/*
2554 				 * A page may belong to the object but be
2555 				 * dequeued and set to PQ_NONE while the
2556 				 * object lock is not held.  This makes the
2557 				 * reads of m->queue below racy, and we do not
2558 				 * count pages set to PQ_NONE.  However, this
2559 				 * sysctl is only meant to give an
2560 				 * approximation of the system anyway.
2561 				 */
2562 				if (vm_page_active(m))
2563 					kvo->kvo_active++;
2564 				else if (vm_page_inactive(m))
2565 					kvo->kvo_inactive++;
2566 				else if (vm_page_in_laundry(m))
2567 					kvo->kvo_laundry++;
2568 			}
2569 		}
2570 
2571 		kvo->kvo_vn_fileid = 0;
2572 		kvo->kvo_vn_fsid = 0;
2573 		kvo->kvo_vn_fsid_freebsd11 = 0;
2574 		freepath = NULL;
2575 		fullpath = "";
2576 		vp = NULL;
2577 		kvo->kvo_type = vm_object_kvme_type(obj, want_path ? &vp :
2578 		    NULL);
2579 		if (vp != NULL) {
2580 			vref(vp);
2581 		} else if ((obj->flags & OBJ_ANON) != 0) {
2582 			MPASS(kvo->kvo_type == KVME_TYPE_SWAP);
2583 			kvo->kvo_me = (uintptr_t)obj;
2584 			/* tmpfs objs are reported as vnodes */
2585 			kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2586 			sp = swap_pager_swapped_pages(obj);
2587 			kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2588 		}
2589 		if (obj->type == OBJT_DEVICE || obj->type == OBJT_MGTDEVICE) {
2590 			cdev_pager_get_path(obj, kvo->kvo_path,
2591 			    sizeof(kvo->kvo_path));
2592 		}
2593 		VM_OBJECT_RUNLOCK(obj);
2594 		if ((obj->flags & OBJ_SYSVSHM) != 0) {
2595 			kvo->kvo_flags |= KVMO_FLAG_SYSVSHM;
2596 			shmobjinfo(obj, &key, &seq);
2597 			kvo->kvo_vn_fileid = key;
2598 			kvo->kvo_vn_fsid_freebsd11 = seq;
2599 		}
2600 		if ((obj->flags & OBJ_POSIXSHM) != 0) {
2601 			kvo->kvo_flags |= KVMO_FLAG_POSIXSHM;
2602 			shm_get_path(obj, kvo->kvo_path,
2603 			    sizeof(kvo->kvo_path));
2604 		}
2605 		if (vp != NULL) {
2606 			vn_fullpath(vp, &fullpath, &freepath);
2607 			vn_lock(vp, LK_SHARED | LK_RETRY);
2608 			if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2609 				kvo->kvo_vn_fileid = va.va_fileid;
2610 				kvo->kvo_vn_fsid = va.va_fsid;
2611 				kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2612 								/* truncate */
2613 			}
2614 			vput(vp);
2615 			strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2616 			free(freepath, M_TEMP);
2617 		}
2618 
2619 		/* Pack record size down */
2620 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2621 		    + strlen(kvo->kvo_path) + 1;
2622 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2623 		    sizeof(uint64_t));
2624 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2625 		maybe_yield();
2626 		mtx_lock(&vm_object_list_mtx);
2627 		if (error)
2628 			break;
2629 	}
2630 	mtx_unlock(&vm_object_list_mtx);
2631 	free(kvo, M_TEMP);
2632 	return (error);
2633 }
2634 
2635 static int
2636 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2637 {
2638 	return (vm_object_list_handler(req, false));
2639 }
2640 
2641 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2642     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2643     "List of VM objects");
2644 
2645 static int
2646 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2647 {
2648 	return (vm_object_list_handler(req, true));
2649 }
2650 
2651 /*
2652  * This sysctl returns list of the anonymous or swap objects. Intent
2653  * is to provide stripped optimized list useful to analyze swap use.
2654  * Since technically non-swap (default) objects participate in the
2655  * shadow chains, and are converted to swap type as needed by swap
2656  * pager, we must report them.
2657  */
2658 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2659     CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2660     sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2661     "List of swap VM objects");
2662 
2663 #include "opt_ddb.h"
2664 #ifdef DDB
2665 #include <sys/kernel.h>
2666 
2667 #include <sys/cons.h>
2668 
2669 #include <ddb/ddb.h>
2670 
2671 static int
2672 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2673 {
2674 	vm_map_t tmpm;
2675 	vm_map_entry_t tmpe;
2676 	vm_object_t obj;
2677 
2678 	if (map == 0)
2679 		return 0;
2680 
2681 	if (entry == 0) {
2682 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2683 			if (_vm_object_in_map(map, object, tmpe)) {
2684 				return 1;
2685 			}
2686 		}
2687 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2688 		tmpm = entry->object.sub_map;
2689 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2690 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2691 				return 1;
2692 			}
2693 		}
2694 	} else if ((obj = entry->object.vm_object) != NULL) {
2695 		for (; obj; obj = obj->backing_object)
2696 			if (obj == object) {
2697 				return 1;
2698 			}
2699 	}
2700 	return 0;
2701 }
2702 
2703 static int
2704 vm_object_in_map(vm_object_t object)
2705 {
2706 	struct proc *p;
2707 
2708 	/* sx_slock(&allproc_lock); */
2709 	FOREACH_PROC_IN_SYSTEM(p) {
2710 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2711 			continue;
2712 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2713 			/* sx_sunlock(&allproc_lock); */
2714 			return 1;
2715 		}
2716 	}
2717 	/* sx_sunlock(&allproc_lock); */
2718 	if (_vm_object_in_map(kernel_map, object, 0))
2719 		return 1;
2720 	return 0;
2721 }
2722 
2723 DB_SHOW_COMMAND_FLAGS(vmochk, vm_object_check, DB_CMD_MEMSAFE)
2724 {
2725 	vm_object_t object;
2726 
2727 	/*
2728 	 * make sure that internal objs are in a map somewhere
2729 	 * and none have zero ref counts.
2730 	 */
2731 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2732 		if ((object->flags & OBJ_ANON) != 0) {
2733 			if (object->ref_count == 0) {
2734 				db_printf(
2735 			"vmochk: internal obj has zero ref count: %lu\n",
2736 				    (u_long)object->size);
2737 			}
2738 			if (!vm_object_in_map(object)) {
2739 				db_printf(
2740 			"vmochk: internal obj is not in a map: "
2741 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2742 				    object->ref_count, (u_long)object->size,
2743 				    (u_long)object->size,
2744 				    (void *)object->backing_object);
2745 			}
2746 		}
2747 		if (db_pager_quit)
2748 			return;
2749 	}
2750 }
2751 
2752 /*
2753  *	vm_object_print:	[ debug ]
2754  */
2755 DB_SHOW_COMMAND(object, vm_object_print_static)
2756 {
2757 	struct pctrie_iter pages;
2758 	/* XXX convert args. */
2759 	vm_object_t object = (vm_object_t)addr;
2760 	boolean_t full = have_addr;
2761 
2762 	vm_page_t p;
2763 
2764 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2765 #define	count	was_count
2766 
2767 	int count;
2768 
2769 	if (object == NULL)
2770 		return;
2771 
2772 	db_iprintf("Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x",
2773 	    object, (int)object->type, (uintmax_t)object->size,
2774 	    object->resident_page_count, object->ref_count, object->flags);
2775 	db_iprintf(" ruid %d charge %jx\n",
2776 	    object->cred ? object->cred->cr_ruid : -1,
2777 	    (uintmax_t)object->charge);
2778 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2779 	    atomic_load_int(&object->shadow_count),
2780 	    object->backing_object ? object->backing_object->ref_count : 0,
2781 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2782 
2783 	if (!full)
2784 		return;
2785 
2786 	db_indent += 2;
2787 	count = 0;
2788 	vm_page_iter_init(&pages, object);
2789 	VM_RADIX_FOREACH(p, &pages) {
2790 		if (count == 0)
2791 			db_iprintf("memory:=");
2792 		else if (count == 6) {
2793 			db_printf("\n");
2794 			db_iprintf(" ...");
2795 			count = 0;
2796 		} else
2797 			db_printf(",");
2798 		count++;
2799 
2800 		db_printf("(off=0x%jx,page=0x%jx)",
2801 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2802 
2803 		if (db_pager_quit)
2804 			break;
2805 	}
2806 	if (count != 0)
2807 		db_printf("\n");
2808 	db_indent -= 2;
2809 }
2810 
2811 /* XXX. */
2812 #undef count
2813 
2814 /* XXX need this non-static entry for calling from vm_map_print. */
2815 void
2816 vm_object_print(
2817         /* db_expr_t */ long addr,
2818 	boolean_t have_addr,
2819 	/* db_expr_t */ long count,
2820 	char *modif)
2821 {
2822 	vm_object_print_static(addr, have_addr, count, modif);
2823 }
2824 
2825 DB_SHOW_COMMAND_FLAGS(vmopag, vm_object_print_pages, DB_CMD_MEMSAFE)
2826 {
2827 	struct pctrie_iter pages;
2828 	vm_object_t object;
2829 	vm_page_t m, start_m;
2830 	int rcount;
2831 
2832 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2833 		db_printf("new object: %p\n", (void *)object);
2834 		if (db_pager_quit)
2835 			return;
2836 		start_m = NULL;
2837 		vm_page_iter_init(&pages, object);
2838 		VM_RADIX_FOREACH(m, &pages) {
2839 			if (start_m == NULL) {
2840 				start_m = m;
2841 				rcount = 0;
2842 			} else if (start_m->pindex + rcount != m->pindex ||
2843 			    VM_PAGE_TO_PHYS(start_m)  + ptoa(rcount) !=
2844 			    VM_PAGE_TO_PHYS(m)) {
2845 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2846 				    (long)start_m->pindex, rcount,
2847 				    (long)VM_PAGE_TO_PHYS(start_m));
2848 				if (db_pager_quit)
2849 					return;
2850 				start_m = m;
2851 				rcount = 0;
2852 			}
2853 			rcount++;
2854 		}
2855 		if (start_m != NULL) {
2856 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2857 			    (long)start_m->pindex, rcount,
2858 			    (long)VM_PAGE_TO_PHYS(start_m));
2859 			if (db_pager_quit)
2860 				return;
2861 		}
2862 	}
2863 }
2864 #endif /* DDB */
2865