xref: /freebsd/sys/vm/vm_object.c (revision ba3c1f5972d7b90feb6e6da47905ff2757e0fe57)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	from: @(#)vm_object.c	8.5 (Berkeley) 3/22/94
35  *
36  *
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *	Virtual memory object module.
65  */
66 
67 #include <sys/cdefs.h>
68 __FBSDID("$FreeBSD$");
69 
70 #include "opt_vm.h"
71 
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/blockcount.h>
75 #include <sys/cpuset.h>
76 #include <sys/limits.h>
77 #include <sys/lock.h>
78 #include <sys/mman.h>
79 #include <sys/mount.h>
80 #include <sys/kernel.h>
81 #include <sys/pctrie.h>
82 #include <sys/sysctl.h>
83 #include <sys/mutex.h>
84 #include <sys/proc.h>		/* for curproc, pageproc */
85 #include <sys/refcount.h>
86 #include <sys/socket.h>
87 #include <sys/resourcevar.h>
88 #include <sys/refcount.h>
89 #include <sys/rwlock.h>
90 #include <sys/user.h>
91 #include <sys/vnode.h>
92 #include <sys/vmmeter.h>
93 #include <sys/sx.h>
94 
95 #include <vm/vm.h>
96 #include <vm/vm_param.h>
97 #include <vm/pmap.h>
98 #include <vm/vm_map.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_pageout.h>
102 #include <vm/vm_pager.h>
103 #include <vm/vm_phys.h>
104 #include <vm/vm_pagequeue.h>
105 #include <vm/swap_pager.h>
106 #include <vm/vm_kern.h>
107 #include <vm/vm_extern.h>
108 #include <vm/vm_radix.h>
109 #include <vm/vm_reserv.h>
110 #include <vm/uma.h>
111 
112 static int old_msync;
113 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
114     "Use old (insecure) msync behavior");
115 
116 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
117 		    int pagerflags, int flags, boolean_t *allclean,
118 		    boolean_t *eio);
119 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
120 		    boolean_t *allclean);
121 static void	vm_object_backing_remove(vm_object_t object);
122 
123 /*
124  *	Virtual memory objects maintain the actual data
125  *	associated with allocated virtual memory.  A given
126  *	page of memory exists within exactly one object.
127  *
128  *	An object is only deallocated when all "references"
129  *	are given up.  Only one "reference" to a given
130  *	region of an object should be writeable.
131  *
132  *	Associated with each object is a list of all resident
133  *	memory pages belonging to that object; this list is
134  *	maintained by the "vm_page" module, and locked by the object's
135  *	lock.
136  *
137  *	Each object also records a "pager" routine which is
138  *	used to retrieve (and store) pages to the proper backing
139  *	storage.  In addition, objects may be backed by other
140  *	objects from which they were virtual-copied.
141  *
142  *	The only items within the object structure which are
143  *	modified after time of creation are:
144  *		reference count		locked by object's lock
145  *		pager routine		locked by object's lock
146  *
147  */
148 
149 struct object_q vm_object_list;
150 struct mtx vm_object_list_mtx;	/* lock for object list and count */
151 
152 struct vm_object kernel_object_store;
153 
154 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
155     "VM object stats");
156 
157 static COUNTER_U64_DEFINE_EARLY(object_collapses);
158 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
159     &object_collapses,
160     "VM object collapses");
161 
162 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
163 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
164     &object_bypasses,
165     "VM object bypasses");
166 
167 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
168 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
169     &object_collapse_waits,
170     "Number of sleeps for collapse");
171 
172 static uma_zone_t obj_zone;
173 
174 static int vm_object_zinit(void *mem, int size, int flags);
175 
176 #ifdef INVARIANTS
177 static void vm_object_zdtor(void *mem, int size, void *arg);
178 
179 static void
180 vm_object_zdtor(void *mem, int size, void *arg)
181 {
182 	vm_object_t object;
183 
184 	object = (vm_object_t)mem;
185 	KASSERT(object->ref_count == 0,
186 	    ("object %p ref_count = %d", object, object->ref_count));
187 	KASSERT(TAILQ_EMPTY(&object->memq),
188 	    ("object %p has resident pages in its memq", object));
189 	KASSERT(vm_radix_is_empty(&object->rtree),
190 	    ("object %p has resident pages in its trie", object));
191 #if VM_NRESERVLEVEL > 0
192 	KASSERT(LIST_EMPTY(&object->rvq),
193 	    ("object %p has reservations",
194 	    object));
195 #endif
196 	KASSERT(!vm_object_busied(object),
197 	    ("object %p busy = %d", object, blockcount_read(&object->busy)));
198 	KASSERT(object->resident_page_count == 0,
199 	    ("object %p resident_page_count = %d",
200 	    object, object->resident_page_count));
201 	KASSERT(atomic_load_int(&object->shadow_count) == 0,
202 	    ("object %p shadow_count = %d",
203 	    object, atomic_load_int(&object->shadow_count)));
204 	KASSERT(object->type == OBJT_DEAD,
205 	    ("object %p has non-dead type %d",
206 	    object, object->type));
207 	KASSERT(object->charge == 0 && object->cred == NULL,
208 	    ("object %p has non-zero charge %ju (%p)",
209 	    object, (uintmax_t)object->charge, object->cred));
210 }
211 #endif
212 
213 static int
214 vm_object_zinit(void *mem, int size, int flags)
215 {
216 	vm_object_t object;
217 
218 	object = (vm_object_t)mem;
219 	rw_init_flags(&object->lock, "vmobject", RW_DUPOK | RW_NEW);
220 
221 	/* These are true for any object that has been freed */
222 	object->type = OBJT_DEAD;
223 	vm_radix_init(&object->rtree);
224 	refcount_init(&object->ref_count, 0);
225 	blockcount_init(&object->paging_in_progress);
226 	blockcount_init(&object->busy);
227 	object->resident_page_count = 0;
228 	atomic_store_int(&object->shadow_count, 0);
229 	object->flags = OBJ_DEAD;
230 
231 	mtx_lock(&vm_object_list_mtx);
232 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
233 	mtx_unlock(&vm_object_list_mtx);
234 	return (0);
235 }
236 
237 static void
238 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
239     vm_object_t object, void *handle)
240 {
241 
242 	TAILQ_INIT(&object->memq);
243 	LIST_INIT(&object->shadow_head);
244 
245 	object->type = type;
246 	object->flags = flags;
247 	if ((flags & OBJ_SWAP) != 0) {
248 		pctrie_init(&object->un_pager.swp.swp_blks);
249 		object->un_pager.swp.writemappings = 0;
250 	}
251 
252 	/*
253 	 * Ensure that swap_pager_swapoff() iteration over object_list
254 	 * sees up to date type and pctrie head if it observed
255 	 * non-dead object.
256 	 */
257 	atomic_thread_fence_rel();
258 
259 	object->pg_color = 0;
260 	object->size = size;
261 	object->domain.dr_policy = NULL;
262 	object->generation = 1;
263 	object->cleangeneration = 1;
264 	refcount_init(&object->ref_count, 1);
265 	object->memattr = VM_MEMATTR_DEFAULT;
266 	object->cred = NULL;
267 	object->charge = 0;
268 	object->handle = handle;
269 	object->backing_object = NULL;
270 	object->backing_object_offset = (vm_ooffset_t) 0;
271 #if VM_NRESERVLEVEL > 0
272 	LIST_INIT(&object->rvq);
273 #endif
274 	umtx_shm_object_init(object);
275 }
276 
277 /*
278  *	vm_object_init:
279  *
280  *	Initialize the VM objects module.
281  */
282 void
283 vm_object_init(void)
284 {
285 	TAILQ_INIT(&vm_object_list);
286 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
287 
288 	rw_init(&kernel_object->lock, "kernel vm object");
289 	vm_radix_init(&kernel_object->rtree);
290 	_vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
291 	    VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
292 #if VM_NRESERVLEVEL > 0
293 	kernel_object->flags |= OBJ_COLORED;
294 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
295 #endif
296 	kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
297 
298 	/*
299 	 * The lock portion of struct vm_object must be type stable due
300 	 * to vm_pageout_fallback_object_lock locking a vm object
301 	 * without holding any references to it.
302 	 *
303 	 * paging_in_progress is valid always.  Lockless references to
304 	 * the objects may acquire pip and then check OBJ_DEAD.
305 	 */
306 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
307 #ifdef INVARIANTS
308 	    vm_object_zdtor,
309 #else
310 	    NULL,
311 #endif
312 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
313 
314 	vm_radix_zinit();
315 }
316 
317 void
318 vm_object_clear_flag(vm_object_t object, u_short bits)
319 {
320 
321 	VM_OBJECT_ASSERT_WLOCKED(object);
322 	object->flags &= ~bits;
323 }
324 
325 /*
326  *	Sets the default memory attribute for the specified object.  Pages
327  *	that are allocated to this object are by default assigned this memory
328  *	attribute.
329  *
330  *	Presently, this function must be called before any pages are allocated
331  *	to the object.  In the future, this requirement may be relaxed for
332  *	"default" and "swap" objects.
333  */
334 int
335 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
336 {
337 
338 	VM_OBJECT_ASSERT_WLOCKED(object);
339 
340 	if (object->type == OBJT_DEAD)
341 		return (KERN_INVALID_ARGUMENT);
342 	if (!TAILQ_EMPTY(&object->memq))
343 		return (KERN_FAILURE);
344 
345 	object->memattr = memattr;
346 	return (KERN_SUCCESS);
347 }
348 
349 void
350 vm_object_pip_add(vm_object_t object, short i)
351 {
352 
353 	if (i > 0)
354 		blockcount_acquire(&object->paging_in_progress, i);
355 }
356 
357 void
358 vm_object_pip_wakeup(vm_object_t object)
359 {
360 
361 	vm_object_pip_wakeupn(object, 1);
362 }
363 
364 void
365 vm_object_pip_wakeupn(vm_object_t object, short i)
366 {
367 
368 	if (i > 0)
369 		blockcount_release(&object->paging_in_progress, i);
370 }
371 
372 /*
373  * Atomically drop the object lock and wait for pip to drain.  This protects
374  * from sleep/wakeup races due to identity changes.  The lock is not re-acquired
375  * on return.
376  */
377 static void
378 vm_object_pip_sleep(vm_object_t object, const char *waitid)
379 {
380 
381 	(void)blockcount_sleep(&object->paging_in_progress, &object->lock,
382 	    waitid, PVM | PDROP);
383 }
384 
385 void
386 vm_object_pip_wait(vm_object_t object, const char *waitid)
387 {
388 
389 	VM_OBJECT_ASSERT_WLOCKED(object);
390 
391 	blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
392 	    PVM);
393 }
394 
395 void
396 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
397 {
398 
399 	VM_OBJECT_ASSERT_UNLOCKED(object);
400 
401 	blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
402 }
403 
404 /*
405  *	vm_object_allocate:
406  *
407  *	Returns a new object with the given size.
408  */
409 vm_object_t
410 vm_object_allocate(objtype_t type, vm_pindex_t size)
411 {
412 	vm_object_t object;
413 	u_short flags;
414 
415 	switch (type) {
416 	case OBJT_DEAD:
417 		panic("vm_object_allocate: can't create OBJT_DEAD");
418 	case OBJT_SWAP:
419 		flags = OBJ_COLORED | OBJ_SWAP;
420 		break;
421 	case OBJT_DEVICE:
422 	case OBJT_SG:
423 		flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
424 		break;
425 	case OBJT_MGTDEVICE:
426 		flags = OBJ_FICTITIOUS;
427 		break;
428 	case OBJT_PHYS:
429 		flags = OBJ_UNMANAGED;
430 		break;
431 	case OBJT_VNODE:
432 		flags = 0;
433 		break;
434 	default:
435 		panic("vm_object_allocate: type %d is undefined or dynamic",
436 		    type);
437 	}
438 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
439 	_vm_object_allocate(type, size, flags, object, NULL);
440 
441 	return (object);
442 }
443 
444 vm_object_t
445 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags)
446 {
447 	vm_object_t object;
448 
449 	MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */);
450 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
451 	_vm_object_allocate(dyntype, size, flags, object, NULL);
452 
453 	return (object);
454 }
455 
456 /*
457  *	vm_object_allocate_anon:
458  *
459  *	Returns a new default object of the given size and marked as
460  *	anonymous memory for special split/collapse handling.  Color
461  *	to be initialized by the caller.
462  */
463 vm_object_t
464 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
465     struct ucred *cred, vm_size_t charge)
466 {
467 	vm_object_t handle, object;
468 
469 	if (backing_object == NULL)
470 		handle = NULL;
471 	else if ((backing_object->flags & OBJ_ANON) != 0)
472 		handle = backing_object->handle;
473 	else
474 		handle = backing_object;
475 	object = uma_zalloc(obj_zone, M_WAITOK);
476 	_vm_object_allocate(OBJT_SWAP, size,
477 	    OBJ_ANON | OBJ_ONEMAPPING | OBJ_SWAP, object, handle);
478 	object->cred = cred;
479 	object->charge = cred != NULL ? charge : 0;
480 	return (object);
481 }
482 
483 static void
484 vm_object_reference_vnode(vm_object_t object)
485 {
486 	u_int old;
487 
488 	/*
489 	 * vnode objects need the lock for the first reference
490 	 * to serialize with vnode_object_deallocate().
491 	 */
492 	if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
493 		VM_OBJECT_RLOCK(object);
494 		old = refcount_acquire(&object->ref_count);
495 		if (object->type == OBJT_VNODE && old == 0)
496 			vref(object->handle);
497 		VM_OBJECT_RUNLOCK(object);
498 	}
499 }
500 
501 /*
502  *	vm_object_reference:
503  *
504  *	Acquires a reference to the given object.
505  */
506 void
507 vm_object_reference(vm_object_t object)
508 {
509 
510 	if (object == NULL)
511 		return;
512 
513 	if (object->type == OBJT_VNODE)
514 		vm_object_reference_vnode(object);
515 	else
516 		refcount_acquire(&object->ref_count);
517 	KASSERT((object->flags & OBJ_DEAD) == 0,
518 	    ("vm_object_reference: Referenced dead object."));
519 }
520 
521 /*
522  *	vm_object_reference_locked:
523  *
524  *	Gets another reference to the given object.
525  *
526  *	The object must be locked.
527  */
528 void
529 vm_object_reference_locked(vm_object_t object)
530 {
531 	u_int old;
532 
533 	VM_OBJECT_ASSERT_LOCKED(object);
534 	old = refcount_acquire(&object->ref_count);
535 	if (object->type == OBJT_VNODE && old == 0)
536 		vref(object->handle);
537 	KASSERT((object->flags & OBJ_DEAD) == 0,
538 	    ("vm_object_reference: Referenced dead object."));
539 }
540 
541 /*
542  * Handle deallocating an object of type OBJT_VNODE.
543  */
544 static void
545 vm_object_deallocate_vnode(vm_object_t object)
546 {
547 	struct vnode *vp = (struct vnode *) object->handle;
548 	bool last;
549 
550 	KASSERT(object->type == OBJT_VNODE,
551 	    ("vm_object_deallocate_vnode: not a vnode object"));
552 	KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
553 
554 	/* Object lock to protect handle lookup. */
555 	last = refcount_release(&object->ref_count);
556 	VM_OBJECT_RUNLOCK(object);
557 
558 	if (!last)
559 		return;
560 
561 	if (!umtx_shm_vnobj_persistent)
562 		umtx_shm_object_terminated(object);
563 
564 	/* vrele may need the vnode lock. */
565 	vrele(vp);
566 }
567 
568 /*
569  * We dropped a reference on an object and discovered that it had a
570  * single remaining shadow.  This is a sibling of the reference we
571  * dropped.  Attempt to collapse the sibling and backing object.
572  */
573 static vm_object_t
574 vm_object_deallocate_anon(vm_object_t backing_object)
575 {
576 	vm_object_t object;
577 
578 	/* Fetch the final shadow.  */
579 	object = LIST_FIRST(&backing_object->shadow_head);
580 	KASSERT(object != NULL &&
581 	    atomic_load_int(&backing_object->shadow_count) == 1,
582 	    ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
583 	    backing_object->ref_count,
584 	    atomic_load_int(&backing_object->shadow_count)));
585 	KASSERT((object->flags & OBJ_ANON) != 0,
586 	    ("invalid shadow object %p", object));
587 
588 	if (!VM_OBJECT_TRYWLOCK(object)) {
589 		/*
590 		 * Prevent object from disappearing since we do not have a
591 		 * reference.
592 		 */
593 		vm_object_pip_add(object, 1);
594 		VM_OBJECT_WUNLOCK(backing_object);
595 		VM_OBJECT_WLOCK(object);
596 		vm_object_pip_wakeup(object);
597 	} else
598 		VM_OBJECT_WUNLOCK(backing_object);
599 
600 	/*
601 	 * Check for a collapse/terminate race with the last reference holder.
602 	 */
603 	if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
604 	    !refcount_acquire_if_not_zero(&object->ref_count)) {
605 		VM_OBJECT_WUNLOCK(object);
606 		return (NULL);
607 	}
608 	backing_object = object->backing_object;
609 	if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
610 		vm_object_collapse(object);
611 	VM_OBJECT_WUNLOCK(object);
612 
613 	return (object);
614 }
615 
616 /*
617  *	vm_object_deallocate:
618  *
619  *	Release a reference to the specified object,
620  *	gained either through a vm_object_allocate
621  *	or a vm_object_reference call.  When all references
622  *	are gone, storage associated with this object
623  *	may be relinquished.
624  *
625  *	No object may be locked.
626  */
627 void
628 vm_object_deallocate(vm_object_t object)
629 {
630 	vm_object_t temp;
631 	bool released;
632 
633 	while (object != NULL) {
634 		/*
635 		 * If the reference count goes to 0 we start calling
636 		 * vm_object_terminate() on the object chain.  A ref count
637 		 * of 1 may be a special case depending on the shadow count
638 		 * being 0 or 1.  These cases require a write lock on the
639 		 * object.
640 		 */
641 		if ((object->flags & OBJ_ANON) == 0)
642 			released = refcount_release_if_gt(&object->ref_count, 1);
643 		else
644 			released = refcount_release_if_gt(&object->ref_count, 2);
645 		if (released)
646 			return;
647 
648 		if (object->type == OBJT_VNODE) {
649 			VM_OBJECT_RLOCK(object);
650 			if (object->type == OBJT_VNODE) {
651 				vm_object_deallocate_vnode(object);
652 				return;
653 			}
654 			VM_OBJECT_RUNLOCK(object);
655 		}
656 
657 		VM_OBJECT_WLOCK(object);
658 		KASSERT(object->ref_count > 0,
659 		    ("vm_object_deallocate: object deallocated too many times: %d",
660 		    object->type));
661 
662 		/*
663 		 * If this is not the final reference to an anonymous
664 		 * object we may need to collapse the shadow chain.
665 		 */
666 		if (!refcount_release(&object->ref_count)) {
667 			if (object->ref_count > 1 ||
668 			    atomic_load_int(&object->shadow_count) == 0) {
669 				if ((object->flags & OBJ_ANON) != 0 &&
670 				    object->ref_count == 1)
671 					vm_object_set_flag(object,
672 					    OBJ_ONEMAPPING);
673 				VM_OBJECT_WUNLOCK(object);
674 				return;
675 			}
676 
677 			/* Handle collapsing last ref on anonymous objects. */
678 			object = vm_object_deallocate_anon(object);
679 			continue;
680 		}
681 
682 		/*
683 		 * Handle the final reference to an object.  We restart
684 		 * the loop with the backing object to avoid recursion.
685 		 */
686 		umtx_shm_object_terminated(object);
687 		temp = object->backing_object;
688 		if (temp != NULL) {
689 			KASSERT(object->type == OBJT_SWAP,
690 			    ("shadowed tmpfs v_object 2 %p", object));
691 			vm_object_backing_remove(object);
692 		}
693 
694 		KASSERT((object->flags & OBJ_DEAD) == 0,
695 		    ("vm_object_deallocate: Terminating dead object."));
696 		vm_object_set_flag(object, OBJ_DEAD);
697 		vm_object_terminate(object);
698 		object = temp;
699 	}
700 }
701 
702 void
703 vm_object_destroy(vm_object_t object)
704 {
705 	uma_zfree(obj_zone, object);
706 }
707 
708 static void
709 vm_object_sub_shadow(vm_object_t object)
710 {
711 	KASSERT(object->shadow_count >= 1,
712 	    ("object %p sub_shadow count zero", object));
713 	atomic_subtract_int(&object->shadow_count, 1);
714 }
715 
716 static void
717 vm_object_backing_remove_locked(vm_object_t object)
718 {
719 	vm_object_t backing_object;
720 
721 	backing_object = object->backing_object;
722 	VM_OBJECT_ASSERT_WLOCKED(object);
723 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
724 
725 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
726 	    ("vm_object_backing_remove: Removing collapsing object."));
727 
728 	vm_object_sub_shadow(backing_object);
729 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
730 		LIST_REMOVE(object, shadow_list);
731 		vm_object_clear_flag(object, OBJ_SHADOWLIST);
732 	}
733 	object->backing_object = NULL;
734 }
735 
736 static void
737 vm_object_backing_remove(vm_object_t object)
738 {
739 	vm_object_t backing_object;
740 
741 	VM_OBJECT_ASSERT_WLOCKED(object);
742 
743 	backing_object = object->backing_object;
744 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
745 		VM_OBJECT_WLOCK(backing_object);
746 		vm_object_backing_remove_locked(object);
747 		VM_OBJECT_WUNLOCK(backing_object);
748 	} else {
749 		object->backing_object = NULL;
750 		vm_object_sub_shadow(backing_object);
751 	}
752 }
753 
754 static void
755 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
756 {
757 
758 	VM_OBJECT_ASSERT_WLOCKED(object);
759 
760 	atomic_add_int(&backing_object->shadow_count, 1);
761 	if ((backing_object->flags & OBJ_ANON) != 0) {
762 		VM_OBJECT_ASSERT_WLOCKED(backing_object);
763 		LIST_INSERT_HEAD(&backing_object->shadow_head, object,
764 		    shadow_list);
765 		vm_object_set_flag(object, OBJ_SHADOWLIST);
766 	}
767 	object->backing_object = backing_object;
768 }
769 
770 static void
771 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
772 {
773 
774 	VM_OBJECT_ASSERT_WLOCKED(object);
775 
776 	if ((backing_object->flags & OBJ_ANON) != 0) {
777 		VM_OBJECT_WLOCK(backing_object);
778 		vm_object_backing_insert_locked(object, backing_object);
779 		VM_OBJECT_WUNLOCK(backing_object);
780 	} else {
781 		object->backing_object = backing_object;
782 		atomic_add_int(&backing_object->shadow_count, 1);
783 	}
784 }
785 
786 /*
787  * Insert an object into a backing_object's shadow list with an additional
788  * reference to the backing_object added.
789  */
790 static void
791 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
792 {
793 
794 	VM_OBJECT_ASSERT_WLOCKED(object);
795 
796 	if ((backing_object->flags & OBJ_ANON) != 0) {
797 		VM_OBJECT_WLOCK(backing_object);
798 		KASSERT((backing_object->flags & OBJ_DEAD) == 0,
799 		    ("shadowing dead anonymous object"));
800 		vm_object_reference_locked(backing_object);
801 		vm_object_backing_insert_locked(object, backing_object);
802 		vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
803 		VM_OBJECT_WUNLOCK(backing_object);
804 	} else {
805 		vm_object_reference(backing_object);
806 		atomic_add_int(&backing_object->shadow_count, 1);
807 		object->backing_object = backing_object;
808 	}
809 }
810 
811 /*
812  * Transfer a backing reference from backing_object to object.
813  */
814 static void
815 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
816 {
817 	vm_object_t new_backing_object;
818 
819 	/*
820 	 * Note that the reference to backing_object->backing_object
821 	 * moves from within backing_object to within object.
822 	 */
823 	vm_object_backing_remove_locked(object);
824 	new_backing_object = backing_object->backing_object;
825 	if (new_backing_object == NULL)
826 		return;
827 	if ((new_backing_object->flags & OBJ_ANON) != 0) {
828 		VM_OBJECT_WLOCK(new_backing_object);
829 		vm_object_backing_remove_locked(backing_object);
830 		vm_object_backing_insert_locked(object, new_backing_object);
831 		VM_OBJECT_WUNLOCK(new_backing_object);
832 	} else {
833 		/*
834 		 * shadow_count for new_backing_object is left
835 		 * unchanged, its reference provided by backing_object
836 		 * is replaced by object.
837 		 */
838 		object->backing_object = new_backing_object;
839 		backing_object->backing_object = NULL;
840 	}
841 }
842 
843 /*
844  * Wait for a concurrent collapse to settle.
845  */
846 static void
847 vm_object_collapse_wait(vm_object_t object)
848 {
849 
850 	VM_OBJECT_ASSERT_WLOCKED(object);
851 
852 	while ((object->flags & OBJ_COLLAPSING) != 0) {
853 		vm_object_pip_wait(object, "vmcolwait");
854 		counter_u64_add(object_collapse_waits, 1);
855 	}
856 }
857 
858 /*
859  * Waits for a backing object to clear a pending collapse and returns
860  * it locked if it is an ANON object.
861  */
862 static vm_object_t
863 vm_object_backing_collapse_wait(vm_object_t object)
864 {
865 	vm_object_t backing_object;
866 
867 	VM_OBJECT_ASSERT_WLOCKED(object);
868 
869 	for (;;) {
870 		backing_object = object->backing_object;
871 		if (backing_object == NULL ||
872 		    (backing_object->flags & OBJ_ANON) == 0)
873 			return (NULL);
874 		VM_OBJECT_WLOCK(backing_object);
875 		if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
876 			break;
877 		VM_OBJECT_WUNLOCK(object);
878 		vm_object_pip_sleep(backing_object, "vmbckwait");
879 		counter_u64_add(object_collapse_waits, 1);
880 		VM_OBJECT_WLOCK(object);
881 	}
882 	return (backing_object);
883 }
884 
885 /*
886  *	vm_object_terminate_pages removes any remaining pageable pages
887  *	from the object and resets the object to an empty state.
888  */
889 static void
890 vm_object_terminate_pages(vm_object_t object)
891 {
892 	vm_page_t p, p_next;
893 
894 	VM_OBJECT_ASSERT_WLOCKED(object);
895 
896 	/*
897 	 * Free any remaining pageable pages.  This also removes them from the
898 	 * paging queues.  However, don't free wired pages, just remove them
899 	 * from the object.  Rather than incrementally removing each page from
900 	 * the object, the page and object are reset to any empty state.
901 	 */
902 	TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
903 		vm_page_assert_unbusied(p);
904 		KASSERT(p->object == object &&
905 		    (p->ref_count & VPRC_OBJREF) != 0,
906 		    ("vm_object_terminate_pages: page %p is inconsistent", p));
907 
908 		p->object = NULL;
909 		if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
910 			VM_CNT_INC(v_pfree);
911 			vm_page_free(p);
912 		}
913 	}
914 
915 	/*
916 	 * If the object contained any pages, then reset it to an empty state.
917 	 * None of the object's fields, including "resident_page_count", were
918 	 * modified by the preceding loop.
919 	 */
920 	if (object->resident_page_count != 0) {
921 		vm_radix_reclaim_allnodes(&object->rtree);
922 		TAILQ_INIT(&object->memq);
923 		object->resident_page_count = 0;
924 		if (object->type == OBJT_VNODE)
925 			vdrop(object->handle);
926 	}
927 }
928 
929 /*
930  *	vm_object_terminate actually destroys the specified object, freeing
931  *	up all previously used resources.
932  *
933  *	The object must be locked.
934  *	This routine may block.
935  */
936 void
937 vm_object_terminate(vm_object_t object)
938 {
939 
940 	VM_OBJECT_ASSERT_WLOCKED(object);
941 	KASSERT((object->flags & OBJ_DEAD) != 0,
942 	    ("terminating non-dead obj %p", object));
943 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
944 	    ("terminating collapsing obj %p", object));
945 	KASSERT(object->backing_object == NULL,
946 	    ("terminating shadow obj %p", object));
947 
948 	/*
949 	 * Wait for the pageout daemon and other current users to be
950 	 * done with the object.  Note that new paging_in_progress
951 	 * users can come after this wait, but they must check
952 	 * OBJ_DEAD flag set (without unlocking the object), and avoid
953 	 * the object being terminated.
954 	 */
955 	vm_object_pip_wait(object, "objtrm");
956 
957 	KASSERT(object->ref_count == 0,
958 	    ("vm_object_terminate: object with references, ref_count=%d",
959 	    object->ref_count));
960 
961 	if ((object->flags & OBJ_PG_DTOR) == 0)
962 		vm_object_terminate_pages(object);
963 
964 #if VM_NRESERVLEVEL > 0
965 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
966 		vm_reserv_break_all(object);
967 #endif
968 
969 	KASSERT(object->cred == NULL || (object->flags & OBJ_SWAP) != 0,
970 	    ("%s: non-swap obj %p has cred", __func__, object));
971 
972 	/*
973 	 * Let the pager know object is dead.
974 	 */
975 	vm_pager_deallocate(object);
976 	VM_OBJECT_WUNLOCK(object);
977 
978 	vm_object_destroy(object);
979 }
980 
981 /*
982  * Make the page read-only so that we can clear the object flags.  However, if
983  * this is a nosync mmap then the object is likely to stay dirty so do not
984  * mess with the page and do not clear the object flags.  Returns TRUE if the
985  * page should be flushed, and FALSE otherwise.
986  */
987 static boolean_t
988 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
989 {
990 
991 	vm_page_assert_busied(p);
992 
993 	/*
994 	 * If we have been asked to skip nosync pages and this is a
995 	 * nosync page, skip it.  Note that the object flags were not
996 	 * cleared in this case so we do not have to set them.
997 	 */
998 	if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
999 		*allclean = FALSE;
1000 		return (FALSE);
1001 	} else {
1002 		pmap_remove_write(p);
1003 		return (p->dirty != 0);
1004 	}
1005 }
1006 
1007 /*
1008  *	vm_object_page_clean
1009  *
1010  *	Clean all dirty pages in the specified range of object.  Leaves page
1011  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
1012  *	write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1013  *	leaving the object dirty.
1014  *
1015  *	For swap objects backing tmpfs regular files, do not flush anything,
1016  *	but remove write protection on the mapped pages to update mtime through
1017  *	mmaped writes.
1018  *
1019  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
1020  *	synchronous clustering mode implementation.
1021  *
1022  *	Odd semantics: if start == end, we clean everything.
1023  *
1024  *	The object must be locked.
1025  *
1026  *	Returns FALSE if some page from the range was not written, as
1027  *	reported by the pager, and TRUE otherwise.
1028  */
1029 boolean_t
1030 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1031     int flags)
1032 {
1033 	vm_page_t np, p;
1034 	vm_pindex_t pi, tend, tstart;
1035 	int curgeneration, n, pagerflags;
1036 	boolean_t eio, res, allclean;
1037 
1038 	VM_OBJECT_ASSERT_WLOCKED(object);
1039 
1040 	if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1041 		return (TRUE);
1042 
1043 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1044 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1045 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1046 
1047 	tstart = OFF_TO_IDX(start);
1048 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1049 	allclean = tstart == 0 && tend >= object->size;
1050 	res = TRUE;
1051 
1052 rescan:
1053 	curgeneration = object->generation;
1054 
1055 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1056 		pi = p->pindex;
1057 		if (pi >= tend)
1058 			break;
1059 		np = TAILQ_NEXT(p, listq);
1060 		if (vm_page_none_valid(p))
1061 			continue;
1062 		if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1063 			if (object->generation != curgeneration &&
1064 			    (flags & OBJPC_SYNC) != 0)
1065 				goto rescan;
1066 			np = vm_page_find_least(object, pi);
1067 			continue;
1068 		}
1069 		if (!vm_object_page_remove_write(p, flags, &allclean)) {
1070 			vm_page_xunbusy(p);
1071 			continue;
1072 		}
1073 		if (object->type == OBJT_VNODE) {
1074 			n = vm_object_page_collect_flush(object, p, pagerflags,
1075 			    flags, &allclean, &eio);
1076 			if (eio) {
1077 				res = FALSE;
1078 				allclean = FALSE;
1079 			}
1080 			if (object->generation != curgeneration &&
1081 			    (flags & OBJPC_SYNC) != 0)
1082 				goto rescan;
1083 
1084 			/*
1085 			 * If the VOP_PUTPAGES() did a truncated write, so
1086 			 * that even the first page of the run is not fully
1087 			 * written, vm_pageout_flush() returns 0 as the run
1088 			 * length.  Since the condition that caused truncated
1089 			 * write may be permanent, e.g. exhausted free space,
1090 			 * accepting n == 0 would cause an infinite loop.
1091 			 *
1092 			 * Forwarding the iterator leaves the unwritten page
1093 			 * behind, but there is not much we can do there if
1094 			 * filesystem refuses to write it.
1095 			 */
1096 			if (n == 0) {
1097 				n = 1;
1098 				allclean = FALSE;
1099 			}
1100 		} else {
1101 			n = 1;
1102 			vm_page_xunbusy(p);
1103 		}
1104 		np = vm_page_find_least(object, pi + n);
1105 	}
1106 #if 0
1107 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1108 #endif
1109 
1110 	/*
1111 	 * Leave updating cleangeneration for tmpfs objects to tmpfs
1112 	 * scan.  It needs to update mtime, which happens for other
1113 	 * filesystems during page writeouts.
1114 	 */
1115 	if (allclean && object->type == OBJT_VNODE)
1116 		object->cleangeneration = curgeneration;
1117 	return (res);
1118 }
1119 
1120 static int
1121 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1122     int flags, boolean_t *allclean, boolean_t *eio)
1123 {
1124 	vm_page_t ma[vm_pageout_page_count], p_first, tp;
1125 	int count, i, mreq, runlen;
1126 
1127 	vm_page_lock_assert(p, MA_NOTOWNED);
1128 	vm_page_assert_xbusied(p);
1129 	VM_OBJECT_ASSERT_WLOCKED(object);
1130 
1131 	count = 1;
1132 	mreq = 0;
1133 
1134 	for (tp = p; count < vm_pageout_page_count; count++) {
1135 		tp = vm_page_next(tp);
1136 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1137 			break;
1138 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1139 			vm_page_xunbusy(tp);
1140 			break;
1141 		}
1142 	}
1143 
1144 	for (p_first = p; count < vm_pageout_page_count; count++) {
1145 		tp = vm_page_prev(p_first);
1146 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1147 			break;
1148 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1149 			vm_page_xunbusy(tp);
1150 			break;
1151 		}
1152 		p_first = tp;
1153 		mreq++;
1154 	}
1155 
1156 	for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1157 		ma[i] = tp;
1158 
1159 	vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1160 	return (runlen);
1161 }
1162 
1163 /*
1164  * Note that there is absolutely no sense in writing out
1165  * anonymous objects, so we track down the vnode object
1166  * to write out.
1167  * We invalidate (remove) all pages from the address space
1168  * for semantic correctness.
1169  *
1170  * If the backing object is a device object with unmanaged pages, then any
1171  * mappings to the specified range of pages must be removed before this
1172  * function is called.
1173  *
1174  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1175  * may start out with a NULL object.
1176  */
1177 boolean_t
1178 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1179     boolean_t syncio, boolean_t invalidate)
1180 {
1181 	vm_object_t backing_object;
1182 	struct vnode *vp;
1183 	struct mount *mp;
1184 	int error, flags, fsync_after;
1185 	boolean_t res;
1186 
1187 	if (object == NULL)
1188 		return (TRUE);
1189 	res = TRUE;
1190 	error = 0;
1191 	VM_OBJECT_WLOCK(object);
1192 	while ((backing_object = object->backing_object) != NULL) {
1193 		VM_OBJECT_WLOCK(backing_object);
1194 		offset += object->backing_object_offset;
1195 		VM_OBJECT_WUNLOCK(object);
1196 		object = backing_object;
1197 		if (object->size < OFF_TO_IDX(offset + size))
1198 			size = IDX_TO_OFF(object->size) - offset;
1199 	}
1200 	/*
1201 	 * Flush pages if writing is allowed, invalidate them
1202 	 * if invalidation requested.  Pages undergoing I/O
1203 	 * will be ignored by vm_object_page_remove().
1204 	 *
1205 	 * We cannot lock the vnode and then wait for paging
1206 	 * to complete without deadlocking against vm_fault.
1207 	 * Instead we simply call vm_object_page_remove() and
1208 	 * allow it to block internally on a page-by-page
1209 	 * basis when it encounters pages undergoing async
1210 	 * I/O.
1211 	 */
1212 	if (object->type == OBJT_VNODE &&
1213 	    vm_object_mightbedirty(object) != 0 &&
1214 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1215 		VM_OBJECT_WUNLOCK(object);
1216 		(void)vn_start_write(vp, &mp, V_WAIT);
1217 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1218 		if (syncio && !invalidate && offset == 0 &&
1219 		    atop(size) == object->size) {
1220 			/*
1221 			 * If syncing the whole mapping of the file,
1222 			 * it is faster to schedule all the writes in
1223 			 * async mode, also allowing the clustering,
1224 			 * and then wait for i/o to complete.
1225 			 */
1226 			flags = 0;
1227 			fsync_after = TRUE;
1228 		} else {
1229 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1230 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1231 			fsync_after = FALSE;
1232 		}
1233 		VM_OBJECT_WLOCK(object);
1234 		res = vm_object_page_clean(object, offset, offset + size,
1235 		    flags);
1236 		VM_OBJECT_WUNLOCK(object);
1237 		if (fsync_after) {
1238 			for (;;) {
1239 				error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1240 				if (error != ERELOOKUP)
1241 					break;
1242 
1243 				/*
1244 				 * Allow SU/bufdaemon to handle more
1245 				 * dependencies in the meantime.
1246 				 */
1247 				VOP_UNLOCK(vp);
1248 				vn_finished_write(mp);
1249 
1250 				(void)vn_start_write(vp, &mp, V_WAIT);
1251 				vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1252 			}
1253 		}
1254 		VOP_UNLOCK(vp);
1255 		vn_finished_write(mp);
1256 		if (error != 0)
1257 			res = FALSE;
1258 		VM_OBJECT_WLOCK(object);
1259 	}
1260 	if ((object->type == OBJT_VNODE ||
1261 	     object->type == OBJT_DEVICE) && invalidate) {
1262 		if (object->type == OBJT_DEVICE)
1263 			/*
1264 			 * The option OBJPR_NOTMAPPED must be passed here
1265 			 * because vm_object_page_remove() cannot remove
1266 			 * unmanaged mappings.
1267 			 */
1268 			flags = OBJPR_NOTMAPPED;
1269 		else if (old_msync)
1270 			flags = 0;
1271 		else
1272 			flags = OBJPR_CLEANONLY;
1273 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1274 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1275 	}
1276 	VM_OBJECT_WUNLOCK(object);
1277 	return (res);
1278 }
1279 
1280 /*
1281  * Determine whether the given advice can be applied to the object.  Advice is
1282  * not applied to unmanaged pages since they never belong to page queues, and
1283  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1284  * have been mapped at most once.
1285  */
1286 static bool
1287 vm_object_advice_applies(vm_object_t object, int advice)
1288 {
1289 
1290 	if ((object->flags & OBJ_UNMANAGED) != 0)
1291 		return (false);
1292 	if (advice != MADV_FREE)
1293 		return (true);
1294 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1295 	    (OBJ_ONEMAPPING | OBJ_ANON));
1296 }
1297 
1298 static void
1299 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1300     vm_size_t size)
1301 {
1302 
1303 	if (advice == MADV_FREE)
1304 		vm_pager_freespace(object, pindex, size);
1305 }
1306 
1307 /*
1308  *	vm_object_madvise:
1309  *
1310  *	Implements the madvise function at the object/page level.
1311  *
1312  *	MADV_WILLNEED	(any object)
1313  *
1314  *	    Activate the specified pages if they are resident.
1315  *
1316  *	MADV_DONTNEED	(any object)
1317  *
1318  *	    Deactivate the specified pages if they are resident.
1319  *
1320  *	MADV_FREE	(OBJT_SWAP objects, OBJ_ONEMAPPING only)
1321  *
1322  *	    Deactivate and clean the specified pages if they are
1323  *	    resident.  This permits the process to reuse the pages
1324  *	    without faulting or the kernel to reclaim the pages
1325  *	    without I/O.
1326  */
1327 void
1328 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1329     int advice)
1330 {
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 relookup:
1339 	VM_OBJECT_WLOCK(object);
1340 	if (!vm_object_advice_applies(object, advice)) {
1341 		VM_OBJECT_WUNLOCK(object);
1342 		return;
1343 	}
1344 	for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1345 		tobject = object;
1346 
1347 		/*
1348 		 * If the next page isn't resident in the top-level object, we
1349 		 * need to search the shadow chain.  When applying MADV_FREE, we
1350 		 * take care to release any swap space used to store
1351 		 * non-resident pages.
1352 		 */
1353 		if (m == NULL || pindex < m->pindex) {
1354 			/*
1355 			 * Optimize a common case: if the top-level object has
1356 			 * no backing object, we can skip over the non-resident
1357 			 * range in constant time.
1358 			 */
1359 			if (object->backing_object == NULL) {
1360 				tpindex = (m != NULL && m->pindex < end) ?
1361 				    m->pindex : end;
1362 				vm_object_madvise_freespace(object, advice,
1363 				    pindex, tpindex - pindex);
1364 				if ((pindex = tpindex) == end)
1365 					break;
1366 				goto next_page;
1367 			}
1368 
1369 			tpindex = pindex;
1370 			do {
1371 				vm_object_madvise_freespace(tobject, advice,
1372 				    tpindex, 1);
1373 				/*
1374 				 * Prepare to search the next object in the
1375 				 * chain.
1376 				 */
1377 				backing_object = tobject->backing_object;
1378 				if (backing_object == NULL)
1379 					goto next_pindex;
1380 				VM_OBJECT_WLOCK(backing_object);
1381 				tpindex +=
1382 				    OFF_TO_IDX(tobject->backing_object_offset);
1383 				if (tobject != object)
1384 					VM_OBJECT_WUNLOCK(tobject);
1385 				tobject = backing_object;
1386 				if (!vm_object_advice_applies(tobject, advice))
1387 					goto next_pindex;
1388 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1389 			    NULL);
1390 		} else {
1391 next_page:
1392 			tm = m;
1393 			m = TAILQ_NEXT(m, listq);
1394 		}
1395 
1396 		/*
1397 		 * If the page is not in a normal state, skip it.  The page
1398 		 * can not be invalidated while the object lock is held.
1399 		 */
1400 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1401 			goto next_pindex;
1402 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1403 		    ("vm_object_madvise: page %p is fictitious", tm));
1404 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1405 		    ("vm_object_madvise: page %p is not managed", tm));
1406 		if (vm_page_tryxbusy(tm) == 0) {
1407 			if (object != tobject)
1408 				VM_OBJECT_WUNLOCK(object);
1409 			if (advice == MADV_WILLNEED) {
1410 				/*
1411 				 * Reference the page before unlocking and
1412 				 * sleeping so that the page daemon is less
1413 				 * likely to reclaim it.
1414 				 */
1415 				vm_page_aflag_set(tm, PGA_REFERENCED);
1416 			}
1417 			if (!vm_page_busy_sleep(tm, "madvpo", 0))
1418 				VM_OBJECT_WUNLOCK(tobject);
1419   			goto relookup;
1420 		}
1421 		vm_page_advise(tm, advice);
1422 		vm_page_xunbusy(tm);
1423 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1424 next_pindex:
1425 		if (tobject != object)
1426 			VM_OBJECT_WUNLOCK(tobject);
1427 	}
1428 	VM_OBJECT_WUNLOCK(object);
1429 }
1430 
1431 /*
1432  *	vm_object_shadow:
1433  *
1434  *	Create a new object which is backed by the
1435  *	specified existing object range.  The source
1436  *	object reference is deallocated.
1437  *
1438  *	The new object and offset into that object
1439  *	are returned in the source parameters.
1440  */
1441 void
1442 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1443     struct ucred *cred, bool shared)
1444 {
1445 	vm_object_t source;
1446 	vm_object_t result;
1447 
1448 	source = *object;
1449 
1450 	/*
1451 	 * Don't create the new object if the old object isn't shared.
1452 	 *
1453 	 * If we hold the only reference we can guarantee that it won't
1454 	 * increase while we have the map locked.  Otherwise the race is
1455 	 * harmless and we will end up with an extra shadow object that
1456 	 * will be collapsed later.
1457 	 */
1458 	if (source != NULL && source->ref_count == 1 &&
1459 	    (source->flags & OBJ_ANON) != 0)
1460 		return;
1461 
1462 	/*
1463 	 * Allocate a new object with the given length.
1464 	 */
1465 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1466 
1467 	/*
1468 	 * Store the offset into the source object, and fix up the offset into
1469 	 * the new object.
1470 	 */
1471 	result->backing_object_offset = *offset;
1472 
1473 	if (shared || source != NULL) {
1474 		VM_OBJECT_WLOCK(result);
1475 
1476 		/*
1477 		 * The new object shadows the source object, adding a
1478 		 * reference to it.  Our caller changes his reference
1479 		 * to point to the new object, removing a reference to
1480 		 * the source object.  Net result: no change of
1481 		 * reference count, unless the caller needs to add one
1482 		 * more reference due to forking a shared map entry.
1483 		 */
1484 		if (shared) {
1485 			vm_object_reference_locked(result);
1486 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1487 		}
1488 
1489 		/*
1490 		 * Try to optimize the result object's page color when
1491 		 * shadowing in order to maintain page coloring
1492 		 * consistency in the combined shadowed object.
1493 		 */
1494 		if (source != NULL) {
1495 			vm_object_backing_insert(result, source);
1496 			result->domain = source->domain;
1497 #if VM_NRESERVLEVEL > 0
1498 			vm_object_set_flag(result,
1499 			    (source->flags & OBJ_COLORED));
1500 			result->pg_color = (source->pg_color +
1501 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1502 			    1)) - 1);
1503 #endif
1504 		}
1505 		VM_OBJECT_WUNLOCK(result);
1506 	}
1507 
1508 	/*
1509 	 * Return the new things
1510 	 */
1511 	*offset = 0;
1512 	*object = result;
1513 }
1514 
1515 /*
1516  *	vm_object_split:
1517  *
1518  * Split the pages in a map entry into a new object.  This affords
1519  * easier removal of unused pages, and keeps object inheritance from
1520  * being a negative impact on memory usage.
1521  */
1522 void
1523 vm_object_split(vm_map_entry_t entry)
1524 {
1525 	vm_page_t m, m_next;
1526 	vm_object_t orig_object, new_object, backing_object;
1527 	vm_pindex_t idx, offidxstart;
1528 	vm_size_t size;
1529 
1530 	orig_object = entry->object.vm_object;
1531 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1532 	    ("vm_object_split:  Splitting object with multiple mappings."));
1533 	if ((orig_object->flags & OBJ_ANON) == 0)
1534 		return;
1535 	if (orig_object->ref_count <= 1)
1536 		return;
1537 	VM_OBJECT_WUNLOCK(orig_object);
1538 
1539 	offidxstart = OFF_TO_IDX(entry->offset);
1540 	size = atop(entry->end - entry->start);
1541 
1542 	new_object = vm_object_allocate_anon(size, orig_object,
1543 	    orig_object->cred, ptoa(size));
1544 
1545 	/*
1546 	 * We must wait for the orig_object to complete any in-progress
1547 	 * collapse so that the swap blocks are stable below.  The
1548 	 * additional reference on backing_object by new object will
1549 	 * prevent further collapse operations until split completes.
1550 	 */
1551 	VM_OBJECT_WLOCK(orig_object);
1552 	vm_object_collapse_wait(orig_object);
1553 
1554 	/*
1555 	 * At this point, the new object is still private, so the order in
1556 	 * which the original and new objects are locked does not matter.
1557 	 */
1558 	VM_OBJECT_WLOCK(new_object);
1559 	new_object->domain = orig_object->domain;
1560 	backing_object = orig_object->backing_object;
1561 	if (backing_object != NULL) {
1562 		vm_object_backing_insert_ref(new_object, backing_object);
1563 		new_object->backing_object_offset =
1564 		    orig_object->backing_object_offset + entry->offset;
1565 	}
1566 	if (orig_object->cred != NULL) {
1567 		crhold(orig_object->cred);
1568 		KASSERT(orig_object->charge >= ptoa(size),
1569 		    ("orig_object->charge < 0"));
1570 		orig_object->charge -= ptoa(size);
1571 	}
1572 
1573 	/*
1574 	 * Mark the split operation so that swap_pager_getpages() knows
1575 	 * that the object is in transition.
1576 	 */
1577 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1578 #ifdef INVARIANTS
1579 	idx = 0;
1580 #endif
1581 retry:
1582 	m = vm_page_find_least(orig_object, offidxstart);
1583 	KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1584 	    ("%s: object %p was repopulated", __func__, orig_object));
1585 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1586 	    m = m_next) {
1587 		m_next = TAILQ_NEXT(m, listq);
1588 
1589 		/*
1590 		 * We must wait for pending I/O to complete before we can
1591 		 * rename the page.
1592 		 *
1593 		 * We do not have to VM_PROT_NONE the page as mappings should
1594 		 * not be changed by this operation.
1595 		 */
1596 		if (vm_page_tryxbusy(m) == 0) {
1597 			VM_OBJECT_WUNLOCK(new_object);
1598 			if (vm_page_busy_sleep(m, "spltwt", 0))
1599 				VM_OBJECT_WLOCK(orig_object);
1600 			VM_OBJECT_WLOCK(new_object);
1601 			goto retry;
1602 		}
1603 
1604 		/*
1605 		 * The page was left invalid.  Likely placed there by
1606 		 * an incomplete fault.  Just remove and ignore.
1607 		 */
1608 		if (vm_page_none_valid(m)) {
1609 			if (vm_page_remove(m))
1610 				vm_page_free(m);
1611 			continue;
1612 		}
1613 
1614 		/* vm_page_rename() will dirty the page. */
1615 		if (vm_page_rename(m, new_object, idx)) {
1616 			vm_page_xunbusy(m);
1617 			VM_OBJECT_WUNLOCK(new_object);
1618 			VM_OBJECT_WUNLOCK(orig_object);
1619 			vm_radix_wait();
1620 			VM_OBJECT_WLOCK(orig_object);
1621 			VM_OBJECT_WLOCK(new_object);
1622 			goto retry;
1623 		}
1624 
1625 #if VM_NRESERVLEVEL > 0
1626 		/*
1627 		 * If some of the reservation's allocated pages remain with
1628 		 * the original object, then transferring the reservation to
1629 		 * the new object is neither particularly beneficial nor
1630 		 * particularly harmful as compared to leaving the reservation
1631 		 * with the original object.  If, however, all of the
1632 		 * reservation's allocated pages are transferred to the new
1633 		 * object, then transferring the reservation is typically
1634 		 * beneficial.  Determining which of these two cases applies
1635 		 * would be more costly than unconditionally renaming the
1636 		 * reservation.
1637 		 */
1638 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1639 #endif
1640 	}
1641 
1642 	/*
1643 	 * swap_pager_copy() can sleep, in which case the orig_object's
1644 	 * and new_object's locks are released and reacquired.
1645 	 */
1646 	swap_pager_copy(orig_object, new_object, offidxstart, 0);
1647 
1648 	TAILQ_FOREACH(m, &new_object->memq, listq)
1649 		vm_page_xunbusy(m);
1650 
1651 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1652 	VM_OBJECT_WUNLOCK(orig_object);
1653 	VM_OBJECT_WUNLOCK(new_object);
1654 	entry->object.vm_object = new_object;
1655 	entry->offset = 0LL;
1656 	vm_object_deallocate(orig_object);
1657 	VM_OBJECT_WLOCK(new_object);
1658 }
1659 
1660 static vm_page_t
1661 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1662 {
1663 	vm_object_t backing_object;
1664 
1665 	VM_OBJECT_ASSERT_WLOCKED(object);
1666 	backing_object = object->backing_object;
1667 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1668 
1669 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1670 	    ("invalid ownership %p %p %p", p, object, backing_object));
1671 	/* The page is only NULL when rename fails. */
1672 	if (p == NULL) {
1673 		VM_OBJECT_WUNLOCK(object);
1674 		VM_OBJECT_WUNLOCK(backing_object);
1675 		vm_radix_wait();
1676 		VM_OBJECT_WLOCK(object);
1677 	} else if (p->object == object) {
1678 		VM_OBJECT_WUNLOCK(backing_object);
1679 		if (vm_page_busy_sleep(p, "vmocol", 0))
1680 			VM_OBJECT_WLOCK(object);
1681 	} else {
1682 		VM_OBJECT_WUNLOCK(object);
1683 		if (!vm_page_busy_sleep(p, "vmocol", 0))
1684 			VM_OBJECT_WUNLOCK(backing_object);
1685 		VM_OBJECT_WLOCK(object);
1686 	}
1687 	VM_OBJECT_WLOCK(backing_object);
1688 	return (TAILQ_FIRST(&backing_object->memq));
1689 }
1690 
1691 static bool
1692 vm_object_scan_all_shadowed(vm_object_t object)
1693 {
1694 	vm_object_t backing_object;
1695 	vm_page_t p, pp;
1696 	vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1697 
1698 	VM_OBJECT_ASSERT_WLOCKED(object);
1699 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1700 
1701 	backing_object = object->backing_object;
1702 
1703 	if ((backing_object->flags & OBJ_ANON) == 0)
1704 		return (false);
1705 
1706 	pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1707 	p = vm_page_find_least(backing_object, pi);
1708 	ps = swap_pager_find_least(backing_object, pi);
1709 
1710 	/*
1711 	 * Only check pages inside the parent object's range and
1712 	 * inside the parent object's mapping of the backing object.
1713 	 */
1714 	for (;; pi++) {
1715 		if (p != NULL && p->pindex < pi)
1716 			p = TAILQ_NEXT(p, listq);
1717 		if (ps < pi)
1718 			ps = swap_pager_find_least(backing_object, pi);
1719 		if (p == NULL && ps >= backing_object->size)
1720 			break;
1721 		else if (p == NULL)
1722 			pi = ps;
1723 		else
1724 			pi = MIN(p->pindex, ps);
1725 
1726 		new_pindex = pi - backing_offset_index;
1727 		if (new_pindex >= object->size)
1728 			break;
1729 
1730 		if (p != NULL) {
1731 			/*
1732 			 * If the backing object page is busy a
1733 			 * grandparent or older page may still be
1734 			 * undergoing CoW.  It is not safe to collapse
1735 			 * the backing object until it is quiesced.
1736 			 */
1737 			if (vm_page_tryxbusy(p) == 0)
1738 				return (false);
1739 
1740 			/*
1741 			 * We raced with the fault handler that left
1742 			 * newly allocated invalid page on the object
1743 			 * queue and retried.
1744 			 */
1745 			if (!vm_page_all_valid(p))
1746 				goto unbusy_ret;
1747 		}
1748 
1749 		/*
1750 		 * See if the parent has the page or if the parent's object
1751 		 * pager has the page.  If the parent has the page but the page
1752 		 * is not valid, the parent's object pager must have the page.
1753 		 *
1754 		 * If this fails, the parent does not completely shadow the
1755 		 * object and we might as well give up now.
1756 		 */
1757 		pp = vm_page_lookup(object, new_pindex);
1758 
1759 		/*
1760 		 * The valid check here is stable due to object lock
1761 		 * being required to clear valid and initiate paging.
1762 		 * Busy of p disallows fault handler to validate pp.
1763 		 */
1764 		if ((pp == NULL || vm_page_none_valid(pp)) &&
1765 		    !vm_pager_has_page(object, new_pindex, NULL, NULL))
1766 			goto unbusy_ret;
1767 		if (p != NULL)
1768 			vm_page_xunbusy(p);
1769 	}
1770 	return (true);
1771 
1772 unbusy_ret:
1773 	if (p != NULL)
1774 		vm_page_xunbusy(p);
1775 	return (false);
1776 }
1777 
1778 static void
1779 vm_object_collapse_scan(vm_object_t object)
1780 {
1781 	vm_object_t backing_object;
1782 	vm_page_t next, p, pp;
1783 	vm_pindex_t backing_offset_index, new_pindex;
1784 
1785 	VM_OBJECT_ASSERT_WLOCKED(object);
1786 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1787 
1788 	backing_object = object->backing_object;
1789 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1790 
1791 	/*
1792 	 * Our scan
1793 	 */
1794 	for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1795 		next = TAILQ_NEXT(p, listq);
1796 		new_pindex = p->pindex - backing_offset_index;
1797 
1798 		/*
1799 		 * Check for busy page
1800 		 */
1801 		if (vm_page_tryxbusy(p) == 0) {
1802 			next = vm_object_collapse_scan_wait(object, p);
1803 			continue;
1804 		}
1805 
1806 		KASSERT(object->backing_object == backing_object,
1807 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1808 		    object->backing_object, backing_object));
1809 		KASSERT(p->object == backing_object,
1810 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1811 		    p->object, backing_object));
1812 
1813 		if (p->pindex < backing_offset_index ||
1814 		    new_pindex >= object->size) {
1815 			vm_pager_freespace(backing_object, p->pindex, 1);
1816 
1817 			KASSERT(!pmap_page_is_mapped(p),
1818 			    ("freeing mapped page %p", p));
1819 			if (vm_page_remove(p))
1820 				vm_page_free(p);
1821 			continue;
1822 		}
1823 
1824 		if (!vm_page_all_valid(p)) {
1825 			KASSERT(!pmap_page_is_mapped(p),
1826 			    ("freeing mapped page %p", p));
1827 			if (vm_page_remove(p))
1828 				vm_page_free(p);
1829 			continue;
1830 		}
1831 
1832 		pp = vm_page_lookup(object, new_pindex);
1833 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1834 			vm_page_xunbusy(p);
1835 			/*
1836 			 * The page in the parent is busy and possibly not
1837 			 * (yet) valid.  Until its state is finalized by the
1838 			 * busy bit owner, we can't tell whether it shadows the
1839 			 * original page.
1840 			 */
1841 			next = vm_object_collapse_scan_wait(object, pp);
1842 			continue;
1843 		}
1844 
1845 		if (pp != NULL && vm_page_none_valid(pp)) {
1846 			/*
1847 			 * The page was invalid in the parent.  Likely placed
1848 			 * there by an incomplete fault.  Just remove and
1849 			 * ignore.  p can replace it.
1850 			 */
1851 			if (vm_page_remove(pp))
1852 				vm_page_free(pp);
1853 			pp = NULL;
1854 		}
1855 
1856 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1857 			NULL)) {
1858 			/*
1859 			 * The page already exists in the parent OR swap exists
1860 			 * for this location in the parent.  Leave the parent's
1861 			 * page alone.  Destroy the original page from the
1862 			 * backing object.
1863 			 */
1864 			vm_pager_freespace(backing_object, p->pindex, 1);
1865 			KASSERT(!pmap_page_is_mapped(p),
1866 			    ("freeing mapped page %p", p));
1867 			if (vm_page_remove(p))
1868 				vm_page_free(p);
1869 			if (pp != NULL)
1870 				vm_page_xunbusy(pp);
1871 			continue;
1872 		}
1873 
1874 		/*
1875 		 * Page does not exist in parent, rename the page from the
1876 		 * backing object to the main object.
1877 		 *
1878 		 * If the page was mapped to a process, it can remain mapped
1879 		 * through the rename.  vm_page_rename() will dirty the page.
1880 		 */
1881 		if (vm_page_rename(p, object, new_pindex)) {
1882 			vm_page_xunbusy(p);
1883 			next = vm_object_collapse_scan_wait(object, NULL);
1884 			continue;
1885 		}
1886 
1887 		/* Use the old pindex to free the right page. */
1888 		vm_pager_freespace(backing_object, new_pindex +
1889 		    backing_offset_index, 1);
1890 
1891 #if VM_NRESERVLEVEL > 0
1892 		/*
1893 		 * Rename the reservation.
1894 		 */
1895 		vm_reserv_rename(p, object, backing_object,
1896 		    backing_offset_index);
1897 #endif
1898 		vm_page_xunbusy(p);
1899 	}
1900 	return;
1901 }
1902 
1903 /*
1904  *	vm_object_collapse:
1905  *
1906  *	Collapse an object with the object backing it.
1907  *	Pages in the backing object are moved into the
1908  *	parent, and the backing object is deallocated.
1909  */
1910 void
1911 vm_object_collapse(vm_object_t object)
1912 {
1913 	vm_object_t backing_object, new_backing_object;
1914 
1915 	VM_OBJECT_ASSERT_WLOCKED(object);
1916 
1917 	while (TRUE) {
1918 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1919 		    ("collapsing invalid object"));
1920 
1921 		/*
1922 		 * Wait for the backing_object to finish any pending
1923 		 * collapse so that the caller sees the shortest possible
1924 		 * shadow chain.
1925 		 */
1926 		backing_object = vm_object_backing_collapse_wait(object);
1927 		if (backing_object == NULL)
1928 			return;
1929 
1930 		KASSERT(object->ref_count > 0 &&
1931 		    object->ref_count > atomic_load_int(&object->shadow_count),
1932 		    ("collapse with invalid ref %d or shadow %d count.",
1933 		    object->ref_count, atomic_load_int(&object->shadow_count)));
1934 		KASSERT((backing_object->flags &
1935 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1936 		    ("vm_object_collapse: Backing object already collapsing."));
1937 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1938 		    ("vm_object_collapse: object is already collapsing."));
1939 
1940 		/*
1941 		 * We know that we can either collapse the backing object if
1942 		 * the parent is the only reference to it, or (perhaps) have
1943 		 * the parent bypass the object if the parent happens to shadow
1944 		 * all the resident pages in the entire backing object.
1945 		 */
1946 		if (backing_object->ref_count == 1) {
1947 			KASSERT(atomic_load_int(&backing_object->shadow_count)
1948 			    == 1,
1949 			    ("vm_object_collapse: shadow_count: %d",
1950 			    atomic_load_int(&backing_object->shadow_count)));
1951 			vm_object_pip_add(object, 1);
1952 			vm_object_set_flag(object, OBJ_COLLAPSING);
1953 			vm_object_pip_add(backing_object, 1);
1954 			vm_object_set_flag(backing_object, OBJ_DEAD);
1955 
1956 			/*
1957 			 * If there is exactly one reference to the backing
1958 			 * object, we can collapse it into the parent.
1959 			 */
1960 			vm_object_collapse_scan(object);
1961 
1962 #if VM_NRESERVLEVEL > 0
1963 			/*
1964 			 * Break any reservations from backing_object.
1965 			 */
1966 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1967 				vm_reserv_break_all(backing_object);
1968 #endif
1969 
1970 			/*
1971 			 * Move the pager from backing_object to object.
1972 			 *
1973 			 * swap_pager_copy() can sleep, in which case the
1974 			 * backing_object's and object's locks are released and
1975 			 * reacquired.
1976 			 */
1977 			swap_pager_copy(backing_object, object,
1978 			    OFF_TO_IDX(object->backing_object_offset), TRUE);
1979 
1980 			/*
1981 			 * Object now shadows whatever backing_object did.
1982 			 */
1983 			vm_object_clear_flag(object, OBJ_COLLAPSING);
1984 			vm_object_backing_transfer(object, backing_object);
1985 			object->backing_object_offset +=
1986 			    backing_object->backing_object_offset;
1987 			VM_OBJECT_WUNLOCK(object);
1988 			vm_object_pip_wakeup(object);
1989 
1990 			/*
1991 			 * Discard backing_object.
1992 			 *
1993 			 * Since the backing object has no pages, no pager left,
1994 			 * and no object references within it, all that is
1995 			 * necessary is to dispose of it.
1996 			 */
1997 			KASSERT(backing_object->ref_count == 1, (
1998 "backing_object %p was somehow re-referenced during collapse!",
1999 			    backing_object));
2000 			vm_object_pip_wakeup(backing_object);
2001 			(void)refcount_release(&backing_object->ref_count);
2002 			vm_object_terminate(backing_object);
2003 			counter_u64_add(object_collapses, 1);
2004 			VM_OBJECT_WLOCK(object);
2005 		} else {
2006 			/*
2007 			 * If we do not entirely shadow the backing object,
2008 			 * there is nothing we can do so we give up.
2009 			 *
2010 			 * The object lock and backing_object lock must not
2011 			 * be dropped during this sequence.
2012 			 */
2013 			if (!vm_object_scan_all_shadowed(object)) {
2014 				VM_OBJECT_WUNLOCK(backing_object);
2015 				break;
2016 			}
2017 
2018 			/*
2019 			 * Make the parent shadow the next object in the
2020 			 * chain.  Deallocating backing_object will not remove
2021 			 * it, since its reference count is at least 2.
2022 			 */
2023 			vm_object_backing_remove_locked(object);
2024 			new_backing_object = backing_object->backing_object;
2025 			if (new_backing_object != NULL) {
2026 				vm_object_backing_insert_ref(object,
2027 				    new_backing_object);
2028 				object->backing_object_offset +=
2029 				    backing_object->backing_object_offset;
2030 			}
2031 
2032 			/*
2033 			 * Drop the reference count on backing_object. Since
2034 			 * its ref_count was at least 2, it will not vanish.
2035 			 */
2036 			(void)refcount_release(&backing_object->ref_count);
2037 			KASSERT(backing_object->ref_count >= 1, (
2038 "backing_object %p was somehow dereferenced during collapse!",
2039 			    backing_object));
2040 			VM_OBJECT_WUNLOCK(backing_object);
2041 			counter_u64_add(object_bypasses, 1);
2042 		}
2043 
2044 		/*
2045 		 * Try again with this object's new backing object.
2046 		 */
2047 	}
2048 }
2049 
2050 /*
2051  *	vm_object_page_remove:
2052  *
2053  *	For the given object, either frees or invalidates each of the
2054  *	specified pages.  In general, a page is freed.  However, if a page is
2055  *	wired for any reason other than the existence of a managed, wired
2056  *	mapping, then it may be invalidated but not removed from the object.
2057  *	Pages are specified by the given range ["start", "end") and the option
2058  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
2059  *	extends from "start" to the end of the object.  If the option
2060  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2061  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
2062  *	specified, then the pages within the specified range must have no
2063  *	mappings.  Otherwise, if this option is not specified, any mappings to
2064  *	the specified pages are removed before the pages are freed or
2065  *	invalidated.
2066  *
2067  *	In general, this operation should only be performed on objects that
2068  *	contain managed pages.  There are, however, two exceptions.  First, it
2069  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
2070  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2071  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
2072  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
2073  *
2074  *	The object must be locked.
2075  */
2076 void
2077 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2078     int options)
2079 {
2080 	vm_page_t p, next;
2081 
2082 	VM_OBJECT_ASSERT_WLOCKED(object);
2083 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2084 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2085 	    ("vm_object_page_remove: illegal options for object %p", object));
2086 	if (object->resident_page_count == 0)
2087 		return;
2088 	vm_object_pip_add(object, 1);
2089 again:
2090 	p = vm_page_find_least(object, start);
2091 
2092 	/*
2093 	 * Here, the variable "p" is either (1) the page with the least pindex
2094 	 * greater than or equal to the parameter "start" or (2) NULL.
2095 	 */
2096 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2097 		next = TAILQ_NEXT(p, listq);
2098 
2099 		/*
2100 		 * Skip invalid pages if asked to do so.  Try to avoid acquiring
2101 		 * the busy lock, as some consumers rely on this to avoid
2102 		 * deadlocks.
2103 		 *
2104 		 * A thread may concurrently transition the page from invalid to
2105 		 * valid using only the busy lock, so the result of this check
2106 		 * is immediately stale.  It is up to consumers to handle this,
2107 		 * for instance by ensuring that all invalid->valid transitions
2108 		 * happen with a mutex held, as may be possible for a
2109 		 * filesystem.
2110 		 */
2111 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2112 			continue;
2113 
2114 		/*
2115 		 * If the page is wired for any reason besides the existence
2116 		 * of managed, wired mappings, then it cannot be freed.  For
2117 		 * example, fictitious pages, which represent device memory,
2118 		 * are inherently wired and cannot be freed.  They can,
2119 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2120 		 * not specified.
2121 		 */
2122 		if (vm_page_tryxbusy(p) == 0) {
2123 			if (vm_page_busy_sleep(p, "vmopar", 0))
2124 				VM_OBJECT_WLOCK(object);
2125 			goto again;
2126 		}
2127 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2128 			vm_page_xunbusy(p);
2129 			continue;
2130 		}
2131 		if (vm_page_wired(p)) {
2132 wired:
2133 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2134 			    object->ref_count != 0)
2135 				pmap_remove_all(p);
2136 			if ((options & OBJPR_CLEANONLY) == 0) {
2137 				vm_page_invalid(p);
2138 				vm_page_undirty(p);
2139 			}
2140 			vm_page_xunbusy(p);
2141 			continue;
2142 		}
2143 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2144 		    ("vm_object_page_remove: page %p is fictitious", p));
2145 		if ((options & OBJPR_CLEANONLY) != 0 &&
2146 		    !vm_page_none_valid(p)) {
2147 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2148 			    object->ref_count != 0 &&
2149 			    !vm_page_try_remove_write(p))
2150 				goto wired;
2151 			if (p->dirty != 0) {
2152 				vm_page_xunbusy(p);
2153 				continue;
2154 			}
2155 		}
2156 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2157 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2158 			goto wired;
2159 		vm_page_free(p);
2160 	}
2161 	vm_object_pip_wakeup(object);
2162 
2163 	vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2164 	    start);
2165 }
2166 
2167 /*
2168  *	vm_object_page_noreuse:
2169  *
2170  *	For the given object, attempt to move the specified pages to
2171  *	the head of the inactive queue.  This bypasses regular LRU
2172  *	operation and allows the pages to be reused quickly under memory
2173  *	pressure.  If a page is wired for any reason, then it will not
2174  *	be queued.  Pages are specified by the range ["start", "end").
2175  *	As a special case, if "end" is zero, then the range extends from
2176  *	"start" to the end of the object.
2177  *
2178  *	This operation should only be performed on objects that
2179  *	contain non-fictitious, managed pages.
2180  *
2181  *	The object must be locked.
2182  */
2183 void
2184 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2185 {
2186 	vm_page_t p, next;
2187 
2188 	VM_OBJECT_ASSERT_LOCKED(object);
2189 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2190 	    ("vm_object_page_noreuse: illegal object %p", object));
2191 	if (object->resident_page_count == 0)
2192 		return;
2193 	p = vm_page_find_least(object, start);
2194 
2195 	/*
2196 	 * Here, the variable "p" is either (1) the page with the least pindex
2197 	 * greater than or equal to the parameter "start" or (2) NULL.
2198 	 */
2199 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2200 		next = TAILQ_NEXT(p, listq);
2201 		vm_page_deactivate_noreuse(p);
2202 	}
2203 }
2204 
2205 /*
2206  *	Populate the specified range of the object with valid pages.  Returns
2207  *	TRUE if the range is successfully populated and FALSE otherwise.
2208  *
2209  *	Note: This function should be optimized to pass a larger array of
2210  *	pages to vm_pager_get_pages() before it is applied to a non-
2211  *	OBJT_DEVICE object.
2212  *
2213  *	The object must be locked.
2214  */
2215 boolean_t
2216 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2217 {
2218 	vm_page_t m;
2219 	vm_pindex_t pindex;
2220 	int rv;
2221 
2222 	VM_OBJECT_ASSERT_WLOCKED(object);
2223 	for (pindex = start; pindex < end; pindex++) {
2224 		rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2225 		if (rv != VM_PAGER_OK)
2226 			break;
2227 
2228 		/*
2229 		 * Keep "m" busy because a subsequent iteration may unlock
2230 		 * the object.
2231 		 */
2232 	}
2233 	if (pindex > start) {
2234 		m = vm_page_lookup(object, start);
2235 		while (m != NULL && m->pindex < pindex) {
2236 			vm_page_xunbusy(m);
2237 			m = TAILQ_NEXT(m, listq);
2238 		}
2239 	}
2240 	return (pindex == end);
2241 }
2242 
2243 /*
2244  *	Routine:	vm_object_coalesce
2245  *	Function:	Coalesces two objects backing up adjoining
2246  *			regions of memory into a single object.
2247  *
2248  *	returns TRUE if objects were combined.
2249  *
2250  *	NOTE:	Only works at the moment if the second object is NULL -
2251  *		if it's not, which object do we lock first?
2252  *
2253  *	Parameters:
2254  *		prev_object	First object to coalesce
2255  *		prev_offset	Offset into prev_object
2256  *		prev_size	Size of reference to prev_object
2257  *		next_size	Size of reference to the second object
2258  *		reserved	Indicator that extension region has
2259  *				swap accounted for
2260  *
2261  *	Conditions:
2262  *	The object must *not* be locked.
2263  */
2264 boolean_t
2265 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2266     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2267 {
2268 	vm_pindex_t next_pindex;
2269 
2270 	if (prev_object == NULL)
2271 		return (TRUE);
2272 	if ((prev_object->flags & OBJ_ANON) == 0)
2273 		return (FALSE);
2274 
2275 	VM_OBJECT_WLOCK(prev_object);
2276 	/*
2277 	 * Try to collapse the object first.
2278 	 */
2279 	vm_object_collapse(prev_object);
2280 
2281 	/*
2282 	 * Can't coalesce if: . more than one reference . paged out . shadows
2283 	 * another object . has a copy elsewhere (any of which mean that the
2284 	 * pages not mapped to prev_entry may be in use anyway)
2285 	 */
2286 	if (prev_object->backing_object != NULL) {
2287 		VM_OBJECT_WUNLOCK(prev_object);
2288 		return (FALSE);
2289 	}
2290 
2291 	prev_size >>= PAGE_SHIFT;
2292 	next_size >>= PAGE_SHIFT;
2293 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2294 
2295 	if (prev_object->ref_count > 1 &&
2296 	    prev_object->size != next_pindex &&
2297 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2298 		VM_OBJECT_WUNLOCK(prev_object);
2299 		return (FALSE);
2300 	}
2301 
2302 	/*
2303 	 * Account for the charge.
2304 	 */
2305 	if (prev_object->cred != NULL) {
2306 		/*
2307 		 * If prev_object was charged, then this mapping,
2308 		 * although not charged now, may become writable
2309 		 * later. Non-NULL cred in the object would prevent
2310 		 * swap reservation during enabling of the write
2311 		 * access, so reserve swap now. Failed reservation
2312 		 * cause allocation of the separate object for the map
2313 		 * entry, and swap reservation for this entry is
2314 		 * managed in appropriate time.
2315 		 */
2316 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2317 		    prev_object->cred)) {
2318 			VM_OBJECT_WUNLOCK(prev_object);
2319 			return (FALSE);
2320 		}
2321 		prev_object->charge += ptoa(next_size);
2322 	}
2323 
2324 	/*
2325 	 * Remove any pages that may still be in the object from a previous
2326 	 * deallocation.
2327 	 */
2328 	if (next_pindex < prev_object->size) {
2329 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2330 		    next_size, 0);
2331 #if 0
2332 		if (prev_object->cred != NULL) {
2333 			KASSERT(prev_object->charge >=
2334 			    ptoa(prev_object->size - next_pindex),
2335 			    ("object %p overcharged 1 %jx %jx", prev_object,
2336 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2337 			prev_object->charge -= ptoa(prev_object->size -
2338 			    next_pindex);
2339 		}
2340 #endif
2341 	}
2342 
2343 	/*
2344 	 * Extend the object if necessary.
2345 	 */
2346 	if (next_pindex + next_size > prev_object->size)
2347 		prev_object->size = next_pindex + next_size;
2348 
2349 	VM_OBJECT_WUNLOCK(prev_object);
2350 	return (TRUE);
2351 }
2352 
2353 void
2354 vm_object_set_writeable_dirty_(vm_object_t object)
2355 {
2356 	atomic_add_int(&object->generation, 1);
2357 }
2358 
2359 bool
2360 vm_object_mightbedirty_(vm_object_t object)
2361 {
2362 	return (object->generation != object->cleangeneration);
2363 }
2364 
2365 /*
2366  *	vm_object_unwire:
2367  *
2368  *	For each page offset within the specified range of the given object,
2369  *	find the highest-level page in the shadow chain and unwire it.  A page
2370  *	must exist at every page offset, and the highest-level page must be
2371  *	wired.
2372  */
2373 void
2374 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2375     uint8_t queue)
2376 {
2377 	vm_object_t tobject, t1object;
2378 	vm_page_t m, tm;
2379 	vm_pindex_t end_pindex, pindex, tpindex;
2380 	int depth, locked_depth;
2381 
2382 	KASSERT((offset & PAGE_MASK) == 0,
2383 	    ("vm_object_unwire: offset is not page aligned"));
2384 	KASSERT((length & PAGE_MASK) == 0,
2385 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2386 	/* The wired count of a fictitious page never changes. */
2387 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2388 		return;
2389 	pindex = OFF_TO_IDX(offset);
2390 	end_pindex = pindex + atop(length);
2391 again:
2392 	locked_depth = 1;
2393 	VM_OBJECT_RLOCK(object);
2394 	m = vm_page_find_least(object, pindex);
2395 	while (pindex < end_pindex) {
2396 		if (m == NULL || pindex < m->pindex) {
2397 			/*
2398 			 * The first object in the shadow chain doesn't
2399 			 * contain a page at the current index.  Therefore,
2400 			 * the page must exist in a backing object.
2401 			 */
2402 			tobject = object;
2403 			tpindex = pindex;
2404 			depth = 0;
2405 			do {
2406 				tpindex +=
2407 				    OFF_TO_IDX(tobject->backing_object_offset);
2408 				tobject = tobject->backing_object;
2409 				KASSERT(tobject != NULL,
2410 				    ("vm_object_unwire: missing page"));
2411 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2412 					goto next_page;
2413 				depth++;
2414 				if (depth == locked_depth) {
2415 					locked_depth++;
2416 					VM_OBJECT_RLOCK(tobject);
2417 				}
2418 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2419 			    NULL);
2420 		} else {
2421 			tm = m;
2422 			m = TAILQ_NEXT(m, listq);
2423 		}
2424 		if (vm_page_trysbusy(tm) == 0) {
2425 			for (tobject = object; locked_depth >= 1;
2426 			    locked_depth--) {
2427 				t1object = tobject->backing_object;
2428 				if (tm->object != tobject)
2429 					VM_OBJECT_RUNLOCK(tobject);
2430 				tobject = t1object;
2431 			}
2432 			tobject = tm->object;
2433 			if (!vm_page_busy_sleep(tm, "unwbo",
2434 			    VM_ALLOC_IGN_SBUSY))
2435 				VM_OBJECT_RUNLOCK(tobject);
2436 			goto again;
2437 		}
2438 		vm_page_unwire(tm, queue);
2439 		vm_page_sunbusy(tm);
2440 next_page:
2441 		pindex++;
2442 	}
2443 	/* Release the accumulated object locks. */
2444 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2445 		t1object = tobject->backing_object;
2446 		VM_OBJECT_RUNLOCK(tobject);
2447 		tobject = t1object;
2448 	}
2449 }
2450 
2451 /*
2452  * Return the vnode for the given object, or NULL if none exists.
2453  * For tmpfs objects, the function may return NULL if there is
2454  * no vnode allocated at the time of the call.
2455  */
2456 struct vnode *
2457 vm_object_vnode(vm_object_t object)
2458 {
2459 	struct vnode *vp;
2460 
2461 	VM_OBJECT_ASSERT_LOCKED(object);
2462 	vm_pager_getvp(object, &vp, NULL);
2463 	return (vp);
2464 }
2465 
2466 /*
2467  * Busy the vm object.  This prevents new pages belonging to the object from
2468  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2469  * for checking page state before proceeding.
2470  */
2471 void
2472 vm_object_busy(vm_object_t obj)
2473 {
2474 
2475 	VM_OBJECT_ASSERT_LOCKED(obj);
2476 
2477 	blockcount_acquire(&obj->busy, 1);
2478 	/* The fence is required to order loads of page busy. */
2479 	atomic_thread_fence_acq_rel();
2480 }
2481 
2482 void
2483 vm_object_unbusy(vm_object_t obj)
2484 {
2485 
2486 	blockcount_release(&obj->busy, 1);
2487 }
2488 
2489 void
2490 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2491 {
2492 
2493 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2494 
2495 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2496 }
2497 
2498 /*
2499  * This function aims to determine if the object is mapped,
2500  * specifically, if it is referenced by a vm_map_entry.  Because
2501  * objects occasionally acquire transient references that do not
2502  * represent a mapping, the method used here is inexact.  However, it
2503  * has very low overhead and is good enough for the advisory
2504  * vm.vmtotal sysctl.
2505  */
2506 bool
2507 vm_object_is_active(vm_object_t obj)
2508 {
2509 
2510 	return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2511 }
2512 
2513 static int
2514 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2515 {
2516 	struct kinfo_vmobject *kvo;
2517 	char *fullpath, *freepath;
2518 	struct vnode *vp;
2519 	struct vattr va;
2520 	vm_object_t obj;
2521 	vm_page_t m;
2522 	u_long sp;
2523 	int count, error;
2524 
2525 	if (req->oldptr == NULL) {
2526 		/*
2527 		 * If an old buffer has not been provided, generate an
2528 		 * estimate of the space needed for a subsequent call.
2529 		 */
2530 		mtx_lock(&vm_object_list_mtx);
2531 		count = 0;
2532 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2533 			if (obj->type == OBJT_DEAD)
2534 				continue;
2535 			count++;
2536 		}
2537 		mtx_unlock(&vm_object_list_mtx);
2538 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2539 		    count * 11 / 10));
2540 	}
2541 
2542 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO);
2543 	error = 0;
2544 
2545 	/*
2546 	 * VM objects are type stable and are never removed from the
2547 	 * list once added.  This allows us to safely read obj->object_list
2548 	 * after reacquiring the VM object lock.
2549 	 */
2550 	mtx_lock(&vm_object_list_mtx);
2551 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2552 		if (obj->type == OBJT_DEAD ||
2553 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2554 			continue;
2555 		VM_OBJECT_RLOCK(obj);
2556 		if (obj->type == OBJT_DEAD ||
2557 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2558 			VM_OBJECT_RUNLOCK(obj);
2559 			continue;
2560 		}
2561 		mtx_unlock(&vm_object_list_mtx);
2562 		kvo->kvo_size = ptoa(obj->size);
2563 		kvo->kvo_resident = obj->resident_page_count;
2564 		kvo->kvo_ref_count = obj->ref_count;
2565 		kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2566 		kvo->kvo_memattr = obj->memattr;
2567 		kvo->kvo_active = 0;
2568 		kvo->kvo_inactive = 0;
2569 		if (!swap_only) {
2570 			TAILQ_FOREACH(m, &obj->memq, listq) {
2571 				/*
2572 				 * A page may belong to the object but be
2573 				 * dequeued and set to PQ_NONE while the
2574 				 * object lock is not held.  This makes the
2575 				 * reads of m->queue below racy, and we do not
2576 				 * count pages set to PQ_NONE.  However, this
2577 				 * sysctl is only meant to give an
2578 				 * approximation of the system anyway.
2579 				 */
2580 				if (m->a.queue == PQ_ACTIVE)
2581 					kvo->kvo_active++;
2582 				else if (m->a.queue == PQ_INACTIVE)
2583 					kvo->kvo_inactive++;
2584 			}
2585 		}
2586 
2587 		kvo->kvo_vn_fileid = 0;
2588 		kvo->kvo_vn_fsid = 0;
2589 		kvo->kvo_vn_fsid_freebsd11 = 0;
2590 		freepath = NULL;
2591 		fullpath = "";
2592 		vp = NULL;
2593 		kvo->kvo_type = vm_object_kvme_type(obj, swap_only ? NULL : &vp);
2594 		if (vp != NULL) {
2595 			vref(vp);
2596 		} else if ((obj->flags & OBJ_ANON) != 0) {
2597 			MPASS(kvo->kvo_type == KVME_TYPE_SWAP);
2598 			kvo->kvo_me = (uintptr_t)obj;
2599 			/* tmpfs objs are reported as vnodes */
2600 			kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2601 			sp = swap_pager_swapped_pages(obj);
2602 			kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2603 		}
2604 		VM_OBJECT_RUNLOCK(obj);
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 		}
2616 
2617 		strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2618 		if (freepath != NULL)
2619 			free(freepath, M_TEMP);
2620 
2621 		/* Pack record size down */
2622 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2623 		    + strlen(kvo->kvo_path) + 1;
2624 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2625 		    sizeof(uint64_t));
2626 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2627 		maybe_yield();
2628 		mtx_lock(&vm_object_list_mtx);
2629 		if (error)
2630 			break;
2631 	}
2632 	mtx_unlock(&vm_object_list_mtx);
2633 	free(kvo, M_TEMP);
2634 	return (error);
2635 }
2636 
2637 static int
2638 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2639 {
2640 	return (vm_object_list_handler(req, false));
2641 }
2642 
2643 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2644     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2645     "List of VM objects");
2646 
2647 static int
2648 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2649 {
2650 	return (vm_object_list_handler(req, true));
2651 }
2652 
2653 /*
2654  * This sysctl returns list of the anonymous or swap objects. Intent
2655  * is to provide stripped optimized list useful to analyze swap use.
2656  * Since technically non-swap (default) objects participate in the
2657  * shadow chains, and are converted to swap type as needed by swap
2658  * pager, we must report them.
2659  */
2660 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2661     CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2662     sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2663     "List of swap VM objects");
2664 
2665 #include "opt_ddb.h"
2666 #ifdef DDB
2667 #include <sys/kernel.h>
2668 
2669 #include <sys/cons.h>
2670 
2671 #include <ddb/ddb.h>
2672 
2673 static int
2674 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2675 {
2676 	vm_map_t tmpm;
2677 	vm_map_entry_t tmpe;
2678 	vm_object_t obj;
2679 
2680 	if (map == 0)
2681 		return 0;
2682 
2683 	if (entry == 0) {
2684 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2685 			if (_vm_object_in_map(map, object, tmpe)) {
2686 				return 1;
2687 			}
2688 		}
2689 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2690 		tmpm = entry->object.sub_map;
2691 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2692 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2693 				return 1;
2694 			}
2695 		}
2696 	} else if ((obj = entry->object.vm_object) != NULL) {
2697 		for (; obj; obj = obj->backing_object)
2698 			if (obj == object) {
2699 				return 1;
2700 			}
2701 	}
2702 	return 0;
2703 }
2704 
2705 static int
2706 vm_object_in_map(vm_object_t object)
2707 {
2708 	struct proc *p;
2709 
2710 	/* sx_slock(&allproc_lock); */
2711 	FOREACH_PROC_IN_SYSTEM(p) {
2712 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2713 			continue;
2714 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2715 			/* sx_sunlock(&allproc_lock); */
2716 			return 1;
2717 		}
2718 	}
2719 	/* sx_sunlock(&allproc_lock); */
2720 	if (_vm_object_in_map(kernel_map, object, 0))
2721 		return 1;
2722 	return 0;
2723 }
2724 
2725 DB_SHOW_COMMAND_FLAGS(vmochk, vm_object_check, DB_CMD_MEMSAFE)
2726 {
2727 	vm_object_t object;
2728 
2729 	/*
2730 	 * make sure that internal objs are in a map somewhere
2731 	 * and none have zero ref counts.
2732 	 */
2733 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2734 		if ((object->flags & OBJ_ANON) != 0) {
2735 			if (object->ref_count == 0) {
2736 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2737 					(long)object->size);
2738 			}
2739 			if (!vm_object_in_map(object)) {
2740 				db_printf(
2741 			"vmochk: internal obj is not in a map: "
2742 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2743 				    object->ref_count, (u_long)object->size,
2744 				    (u_long)object->size,
2745 				    (void *)object->backing_object);
2746 			}
2747 		}
2748 		if (db_pager_quit)
2749 			return;
2750 	}
2751 }
2752 
2753 /*
2754  *	vm_object_print:	[ debug ]
2755  */
2756 DB_SHOW_COMMAND(object, vm_object_print_static)
2757 {
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(
2773 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2774 	    object, (int)object->type, (uintmax_t)object->size,
2775 	    object->resident_page_count, object->ref_count, object->flags,
2776 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2777 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2778 	    atomic_load_int(&object->shadow_count),
2779 	    object->backing_object ? object->backing_object->ref_count : 0,
2780 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2781 
2782 	if (!full)
2783 		return;
2784 
2785 	db_indent += 2;
2786 	count = 0;
2787 	TAILQ_FOREACH(p, &object->memq, listq) {
2788 		if (count == 0)
2789 			db_iprintf("memory:=");
2790 		else if (count == 6) {
2791 			db_printf("\n");
2792 			db_iprintf(" ...");
2793 			count = 0;
2794 		} else
2795 			db_printf(",");
2796 		count++;
2797 
2798 		db_printf("(off=0x%jx,page=0x%jx)",
2799 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2800 
2801 		if (db_pager_quit)
2802 			break;
2803 	}
2804 	if (count != 0)
2805 		db_printf("\n");
2806 	db_indent -= 2;
2807 }
2808 
2809 /* XXX. */
2810 #undef count
2811 
2812 /* XXX need this non-static entry for calling from vm_map_print. */
2813 void
2814 vm_object_print(
2815         /* db_expr_t */ long addr,
2816 	boolean_t have_addr,
2817 	/* db_expr_t */ long count,
2818 	char *modif)
2819 {
2820 	vm_object_print_static(addr, have_addr, count, modif);
2821 }
2822 
2823 DB_SHOW_COMMAND_FLAGS(vmopag, vm_object_print_pages, DB_CMD_MEMSAFE)
2824 {
2825 	vm_object_t object;
2826 	vm_pindex_t fidx;
2827 	vm_paddr_t pa;
2828 	vm_page_t m, prev_m;
2829 	int rcount;
2830 
2831 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2832 		db_printf("new object: %p\n", (void *)object);
2833 		if (db_pager_quit)
2834 			return;
2835 
2836 		rcount = 0;
2837 		fidx = 0;
2838 		pa = -1;
2839 		TAILQ_FOREACH(m, &object->memq, listq) {
2840 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2841 			    prev_m->pindex + 1 != m->pindex) {
2842 				if (rcount) {
2843 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2844 						(long)fidx, rcount, (long)pa);
2845 					if (db_pager_quit)
2846 						return;
2847 					rcount = 0;
2848 				}
2849 			}
2850 			if (rcount &&
2851 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2852 				++rcount;
2853 				continue;
2854 			}
2855 			if (rcount) {
2856 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2857 					(long)fidx, rcount, (long)pa);
2858 				if (db_pager_quit)
2859 					return;
2860 			}
2861 			fidx = m->pindex;
2862 			pa = VM_PAGE_TO_PHYS(m);
2863 			rcount = 1;
2864 		}
2865 		if (rcount) {
2866 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2867 				(long)fidx, rcount, (long)pa);
2868 			if (db_pager_quit)
2869 				return;
2870 		}
2871 	}
2872 }
2873 #endif /* DDB */
2874