xref: /freebsd/sys/vm/vm_object.c (revision 1dfcff294e44d4b45813288ef4095c36abb22f0e)
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, "vm object", 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 
250 	/*
251 	 * Ensure that swap_pager_swapoff() iteration over object_list
252 	 * sees up to date type and pctrie head if it observed
253 	 * non-dead object.
254 	 */
255 	atomic_thread_fence_rel();
256 
257 	object->pg_color = 0;
258 	object->size = size;
259 	object->domain.dr_policy = NULL;
260 	object->generation = 1;
261 	object->cleangeneration = 1;
262 	refcount_init(&object->ref_count, 1);
263 	object->memattr = VM_MEMATTR_DEFAULT;
264 	object->cred = NULL;
265 	object->charge = 0;
266 	object->handle = handle;
267 	object->backing_object = NULL;
268 	object->backing_object_offset = (vm_ooffset_t) 0;
269 #if VM_NRESERVLEVEL > 0
270 	LIST_INIT(&object->rvq);
271 #endif
272 	umtx_shm_object_init(object);
273 }
274 
275 /*
276  *	vm_object_init:
277  *
278  *	Initialize the VM objects module.
279  */
280 void
281 vm_object_init(void)
282 {
283 	TAILQ_INIT(&vm_object_list);
284 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
285 
286 	rw_init(&kernel_object->lock, "kernel vm object");
287 	_vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
288 	    VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
289 #if VM_NRESERVLEVEL > 0
290 	kernel_object->flags |= OBJ_COLORED;
291 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
292 #endif
293 	kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
294 
295 	/*
296 	 * The lock portion of struct vm_object must be type stable due
297 	 * to vm_pageout_fallback_object_lock locking a vm object
298 	 * without holding any references to it.
299 	 *
300 	 * paging_in_progress is valid always.  Lockless references to
301 	 * the objects may acquire pip and then check OBJ_DEAD.
302 	 */
303 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
304 #ifdef INVARIANTS
305 	    vm_object_zdtor,
306 #else
307 	    NULL,
308 #endif
309 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
310 
311 	vm_radix_zinit();
312 }
313 
314 void
315 vm_object_clear_flag(vm_object_t object, u_short bits)
316 {
317 
318 	VM_OBJECT_ASSERT_WLOCKED(object);
319 	object->flags &= ~bits;
320 }
321 
322 /*
323  *	Sets the default memory attribute for the specified object.  Pages
324  *	that are allocated to this object are by default assigned this memory
325  *	attribute.
326  *
327  *	Presently, this function must be called before any pages are allocated
328  *	to the object.  In the future, this requirement may be relaxed for
329  *	"default" and "swap" objects.
330  */
331 int
332 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
333 {
334 
335 	VM_OBJECT_ASSERT_WLOCKED(object);
336 
337 	if (object->type == OBJT_DEAD)
338 		return (KERN_INVALID_ARGUMENT);
339 	if (!TAILQ_EMPTY(&object->memq))
340 		return (KERN_FAILURE);
341 
342 	object->memattr = memattr;
343 	return (KERN_SUCCESS);
344 }
345 
346 void
347 vm_object_pip_add(vm_object_t object, short i)
348 {
349 
350 	if (i > 0)
351 		blockcount_acquire(&object->paging_in_progress, i);
352 }
353 
354 void
355 vm_object_pip_wakeup(vm_object_t object)
356 {
357 
358 	vm_object_pip_wakeupn(object, 1);
359 }
360 
361 void
362 vm_object_pip_wakeupn(vm_object_t object, short i)
363 {
364 
365 	if (i > 0)
366 		blockcount_release(&object->paging_in_progress, i);
367 }
368 
369 /*
370  * Atomically drop the object lock and wait for pip to drain.  This protects
371  * from sleep/wakeup races due to identity changes.  The lock is not re-acquired
372  * on return.
373  */
374 static void
375 vm_object_pip_sleep(vm_object_t object, const char *waitid)
376 {
377 
378 	(void)blockcount_sleep(&object->paging_in_progress, &object->lock,
379 	    waitid, PVM | PDROP);
380 }
381 
382 void
383 vm_object_pip_wait(vm_object_t object, const char *waitid)
384 {
385 
386 	VM_OBJECT_ASSERT_WLOCKED(object);
387 
388 	blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
389 	    PVM);
390 }
391 
392 void
393 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
394 {
395 
396 	VM_OBJECT_ASSERT_UNLOCKED(object);
397 
398 	blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
399 }
400 
401 /*
402  *	vm_object_allocate:
403  *
404  *	Returns a new object with the given size.
405  */
406 vm_object_t
407 vm_object_allocate(objtype_t type, vm_pindex_t size)
408 {
409 	vm_object_t object;
410 	u_short flags;
411 
412 	switch (type) {
413 	case OBJT_DEAD:
414 		panic("vm_object_allocate: can't create OBJT_DEAD");
415 	case OBJT_DEFAULT:
416 		flags = OBJ_COLORED;
417 		break;
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_DEFAULT, size, OBJ_ANON | OBJ_ONEMAPPING,
477 	    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_DEFAULT ||
690 			    object->type == OBJT_SWAP,
691 			    ("shadowed tmpfs v_object 2 %p", object));
692 			vm_object_backing_remove(object);
693 		}
694 
695 		KASSERT((object->flags & OBJ_DEAD) == 0,
696 		    ("vm_object_deallocate: Terminating dead object."));
697 		vm_object_set_flag(object, OBJ_DEAD);
698 		vm_object_terminate(object);
699 		object = temp;
700 	}
701 }
702 
703 /*
704  *	vm_object_destroy removes the object from the global object list
705  *      and frees the space for the object.
706  */
707 void
708 vm_object_destroy(vm_object_t object)
709 {
710 
711 	/*
712 	 * Release the allocation charge.
713 	 */
714 	if (object->cred != NULL) {
715 		swap_release_by_cred(object->charge, object->cred);
716 		object->charge = 0;
717 		crfree(object->cred);
718 		object->cred = NULL;
719 	}
720 
721 	/*
722 	 * Free the space for the object.
723 	 */
724 	uma_zfree(obj_zone, object);
725 }
726 
727 static void
728 vm_object_sub_shadow(vm_object_t object)
729 {
730 	KASSERT(object->shadow_count >= 1,
731 	    ("object %p sub_shadow count zero", object));
732 	atomic_subtract_int(&object->shadow_count, 1);
733 }
734 
735 static void
736 vm_object_backing_remove_locked(vm_object_t object)
737 {
738 	vm_object_t backing_object;
739 
740 	backing_object = object->backing_object;
741 	VM_OBJECT_ASSERT_WLOCKED(object);
742 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
743 
744 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
745 	    ("vm_object_backing_remove: Removing collapsing object."));
746 
747 	vm_object_sub_shadow(backing_object);
748 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
749 		LIST_REMOVE(object, shadow_list);
750 		vm_object_clear_flag(object, OBJ_SHADOWLIST);
751 	}
752 	object->backing_object = NULL;
753 }
754 
755 static void
756 vm_object_backing_remove(vm_object_t object)
757 {
758 	vm_object_t backing_object;
759 
760 	VM_OBJECT_ASSERT_WLOCKED(object);
761 
762 	backing_object = object->backing_object;
763 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
764 		VM_OBJECT_WLOCK(backing_object);
765 		vm_object_backing_remove_locked(object);
766 		VM_OBJECT_WUNLOCK(backing_object);
767 	} else {
768 		object->backing_object = NULL;
769 		vm_object_sub_shadow(backing_object);
770 	}
771 }
772 
773 static void
774 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
775 {
776 
777 	VM_OBJECT_ASSERT_WLOCKED(object);
778 
779 	atomic_add_int(&backing_object->shadow_count, 1);
780 	if ((backing_object->flags & OBJ_ANON) != 0) {
781 		VM_OBJECT_ASSERT_WLOCKED(backing_object);
782 		LIST_INSERT_HEAD(&backing_object->shadow_head, object,
783 		    shadow_list);
784 		vm_object_set_flag(object, OBJ_SHADOWLIST);
785 	}
786 	object->backing_object = backing_object;
787 }
788 
789 static void
790 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
791 {
792 
793 	VM_OBJECT_ASSERT_WLOCKED(object);
794 
795 	if ((backing_object->flags & OBJ_ANON) != 0) {
796 		VM_OBJECT_WLOCK(backing_object);
797 		vm_object_backing_insert_locked(object, backing_object);
798 		VM_OBJECT_WUNLOCK(backing_object);
799 	} else {
800 		object->backing_object = backing_object;
801 		atomic_add_int(&backing_object->shadow_count, 1);
802 	}
803 }
804 
805 /*
806  * Insert an object into a backing_object's shadow list with an additional
807  * reference to the backing_object added.
808  */
809 static void
810 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
811 {
812 
813 	VM_OBJECT_ASSERT_WLOCKED(object);
814 
815 	if ((backing_object->flags & OBJ_ANON) != 0) {
816 		VM_OBJECT_WLOCK(backing_object);
817 		KASSERT((backing_object->flags & OBJ_DEAD) == 0,
818 		    ("shadowing dead anonymous object"));
819 		vm_object_reference_locked(backing_object);
820 		vm_object_backing_insert_locked(object, backing_object);
821 		vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
822 		VM_OBJECT_WUNLOCK(backing_object);
823 	} else {
824 		vm_object_reference(backing_object);
825 		atomic_add_int(&backing_object->shadow_count, 1);
826 		object->backing_object = backing_object;
827 	}
828 }
829 
830 /*
831  * Transfer a backing reference from backing_object to object.
832  */
833 static void
834 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
835 {
836 	vm_object_t new_backing_object;
837 
838 	/*
839 	 * Note that the reference to backing_object->backing_object
840 	 * moves from within backing_object to within object.
841 	 */
842 	vm_object_backing_remove_locked(object);
843 	new_backing_object = backing_object->backing_object;
844 	if (new_backing_object == NULL)
845 		return;
846 	if ((new_backing_object->flags & OBJ_ANON) != 0) {
847 		VM_OBJECT_WLOCK(new_backing_object);
848 		vm_object_backing_remove_locked(backing_object);
849 		vm_object_backing_insert_locked(object, new_backing_object);
850 		VM_OBJECT_WUNLOCK(new_backing_object);
851 	} else {
852 		/*
853 		 * shadow_count for new_backing_object is left
854 		 * unchanged, its reference provided by backing_object
855 		 * is replaced by object.
856 		 */
857 		object->backing_object = new_backing_object;
858 		backing_object->backing_object = NULL;
859 	}
860 }
861 
862 /*
863  * Wait for a concurrent collapse to settle.
864  */
865 static void
866 vm_object_collapse_wait(vm_object_t object)
867 {
868 
869 	VM_OBJECT_ASSERT_WLOCKED(object);
870 
871 	while ((object->flags & OBJ_COLLAPSING) != 0) {
872 		vm_object_pip_wait(object, "vmcolwait");
873 		counter_u64_add(object_collapse_waits, 1);
874 	}
875 }
876 
877 /*
878  * Waits for a backing object to clear a pending collapse and returns
879  * it locked if it is an ANON object.
880  */
881 static vm_object_t
882 vm_object_backing_collapse_wait(vm_object_t object)
883 {
884 	vm_object_t backing_object;
885 
886 	VM_OBJECT_ASSERT_WLOCKED(object);
887 
888 	for (;;) {
889 		backing_object = object->backing_object;
890 		if (backing_object == NULL ||
891 		    (backing_object->flags & OBJ_ANON) == 0)
892 			return (NULL);
893 		VM_OBJECT_WLOCK(backing_object);
894 		if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
895 			break;
896 		VM_OBJECT_WUNLOCK(object);
897 		vm_object_pip_sleep(backing_object, "vmbckwait");
898 		counter_u64_add(object_collapse_waits, 1);
899 		VM_OBJECT_WLOCK(object);
900 	}
901 	return (backing_object);
902 }
903 
904 /*
905  *	vm_object_terminate_pages removes any remaining pageable pages
906  *	from the object and resets the object to an empty state.
907  */
908 static void
909 vm_object_terminate_pages(vm_object_t object)
910 {
911 	vm_page_t p, p_next;
912 
913 	VM_OBJECT_ASSERT_WLOCKED(object);
914 
915 	/*
916 	 * Free any remaining pageable pages.  This also removes them from the
917 	 * paging queues.  However, don't free wired pages, just remove them
918 	 * from the object.  Rather than incrementally removing each page from
919 	 * the object, the page and object are reset to any empty state.
920 	 */
921 	TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
922 		vm_page_assert_unbusied(p);
923 		KASSERT(p->object == object &&
924 		    (p->ref_count & VPRC_OBJREF) != 0,
925 		    ("vm_object_terminate_pages: page %p is inconsistent", p));
926 
927 		p->object = NULL;
928 		if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
929 			VM_CNT_INC(v_pfree);
930 			vm_page_free(p);
931 		}
932 	}
933 
934 	/*
935 	 * If the object contained any pages, then reset it to an empty state.
936 	 * None of the object's fields, including "resident_page_count", were
937 	 * modified by the preceding loop.
938 	 */
939 	if (object->resident_page_count != 0) {
940 		vm_radix_reclaim_allnodes(&object->rtree);
941 		TAILQ_INIT(&object->memq);
942 		object->resident_page_count = 0;
943 		if (object->type == OBJT_VNODE)
944 			vdrop(object->handle);
945 	}
946 }
947 
948 /*
949  *	vm_object_terminate actually destroys the specified object, freeing
950  *	up all previously used resources.
951  *
952  *	The object must be locked.
953  *	This routine may block.
954  */
955 void
956 vm_object_terminate(vm_object_t object)
957 {
958 
959 	VM_OBJECT_ASSERT_WLOCKED(object);
960 	KASSERT((object->flags & OBJ_DEAD) != 0,
961 	    ("terminating non-dead obj %p", object));
962 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
963 	    ("terminating collapsing obj %p", object));
964 	KASSERT(object->backing_object == NULL,
965 	    ("terminating shadow obj %p", object));
966 
967 	/*
968 	 * Wait for the pageout daemon and other current users to be
969 	 * done with the object.  Note that new paging_in_progress
970 	 * users can come after this wait, but they must check
971 	 * OBJ_DEAD flag set (without unlocking the object), and avoid
972 	 * the object being terminated.
973 	 */
974 	vm_object_pip_wait(object, "objtrm");
975 
976 	KASSERT(object->ref_count == 0,
977 	    ("vm_object_terminate: object with references, ref_count=%d",
978 	    object->ref_count));
979 
980 	if ((object->flags & OBJ_PG_DTOR) == 0)
981 		vm_object_terminate_pages(object);
982 
983 #if VM_NRESERVLEVEL > 0
984 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
985 		vm_reserv_break_all(object);
986 #endif
987 
988 	KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
989 	    (object->flags & OBJ_SWAP) != 0,
990 	    ("%s: non-swap obj %p has cred", __func__, object));
991 
992 	/*
993 	 * Let the pager know object is dead.
994 	 */
995 	vm_pager_deallocate(object);
996 	VM_OBJECT_WUNLOCK(object);
997 
998 	vm_object_destroy(object);
999 }
1000 
1001 /*
1002  * Make the page read-only so that we can clear the object flags.  However, if
1003  * this is a nosync mmap then the object is likely to stay dirty so do not
1004  * mess with the page and do not clear the object flags.  Returns TRUE if the
1005  * page should be flushed, and FALSE otherwise.
1006  */
1007 static boolean_t
1008 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
1009 {
1010 
1011 	vm_page_assert_busied(p);
1012 
1013 	/*
1014 	 * If we have been asked to skip nosync pages and this is a
1015 	 * nosync page, skip it.  Note that the object flags were not
1016 	 * cleared in this case so we do not have to set them.
1017 	 */
1018 	if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
1019 		*allclean = FALSE;
1020 		return (FALSE);
1021 	} else {
1022 		pmap_remove_write(p);
1023 		return (p->dirty != 0);
1024 	}
1025 }
1026 
1027 /*
1028  *	vm_object_page_clean
1029  *
1030  *	Clean all dirty pages in the specified range of object.  Leaves page
1031  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
1032  *	write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1033  *	leaving the object dirty.
1034  *
1035  *	For swap objects backing tmpfs regular files, do not flush anything,
1036  *	but remove write protection on the mapped pages to update mtime through
1037  *	mmaped writes.
1038  *
1039  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
1040  *	synchronous clustering mode implementation.
1041  *
1042  *	Odd semantics: if start == end, we clean everything.
1043  *
1044  *	The object must be locked.
1045  *
1046  *	Returns FALSE if some page from the range was not written, as
1047  *	reported by the pager, and TRUE otherwise.
1048  */
1049 boolean_t
1050 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1051     int flags)
1052 {
1053 	vm_page_t np, p;
1054 	vm_pindex_t pi, tend, tstart;
1055 	int curgeneration, n, pagerflags;
1056 	boolean_t eio, res, allclean;
1057 
1058 	VM_OBJECT_ASSERT_WLOCKED(object);
1059 
1060 	if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1061 		return (TRUE);
1062 
1063 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1064 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1065 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1066 
1067 	tstart = OFF_TO_IDX(start);
1068 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1069 	allclean = tstart == 0 && tend >= object->size;
1070 	res = TRUE;
1071 
1072 rescan:
1073 	curgeneration = object->generation;
1074 
1075 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1076 		pi = p->pindex;
1077 		if (pi >= tend)
1078 			break;
1079 		np = TAILQ_NEXT(p, listq);
1080 		if (vm_page_none_valid(p))
1081 			continue;
1082 		if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1083 			if (object->generation != curgeneration &&
1084 			    (flags & OBJPC_SYNC) != 0)
1085 				goto rescan;
1086 			np = vm_page_find_least(object, pi);
1087 			continue;
1088 		}
1089 		if (!vm_object_page_remove_write(p, flags, &allclean)) {
1090 			vm_page_xunbusy(p);
1091 			continue;
1092 		}
1093 		if (object->type == OBJT_VNODE) {
1094 			n = vm_object_page_collect_flush(object, p, pagerflags,
1095 			    flags, &allclean, &eio);
1096 			if (eio) {
1097 				res = FALSE;
1098 				allclean = FALSE;
1099 			}
1100 			if (object->generation != curgeneration &&
1101 			    (flags & OBJPC_SYNC) != 0)
1102 				goto rescan;
1103 
1104 			/*
1105 			 * If the VOP_PUTPAGES() did a truncated write, so
1106 			 * that even the first page of the run is not fully
1107 			 * written, vm_pageout_flush() returns 0 as the run
1108 			 * length.  Since the condition that caused truncated
1109 			 * write may be permanent, e.g. exhausted free space,
1110 			 * accepting n == 0 would cause an infinite loop.
1111 			 *
1112 			 * Forwarding the iterator leaves the unwritten page
1113 			 * behind, but there is not much we can do there if
1114 			 * filesystem refuses to write it.
1115 			 */
1116 			if (n == 0) {
1117 				n = 1;
1118 				allclean = FALSE;
1119 			}
1120 		} else {
1121 			n = 1;
1122 			vm_page_xunbusy(p);
1123 		}
1124 		np = vm_page_find_least(object, pi + n);
1125 	}
1126 #if 0
1127 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1128 #endif
1129 
1130 	/*
1131 	 * Leave updating cleangeneration for tmpfs objects to tmpfs
1132 	 * scan.  It needs to update mtime, which happens for other
1133 	 * filesystems during page writeouts.
1134 	 */
1135 	if (allclean && object->type == OBJT_VNODE)
1136 		object->cleangeneration = curgeneration;
1137 	return (res);
1138 }
1139 
1140 static int
1141 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1142     int flags, boolean_t *allclean, boolean_t *eio)
1143 {
1144 	vm_page_t ma[vm_pageout_page_count], p_first, tp;
1145 	int count, i, mreq, runlen;
1146 
1147 	vm_page_lock_assert(p, MA_NOTOWNED);
1148 	vm_page_assert_xbusied(p);
1149 	VM_OBJECT_ASSERT_WLOCKED(object);
1150 
1151 	count = 1;
1152 	mreq = 0;
1153 
1154 	for (tp = p; count < vm_pageout_page_count; count++) {
1155 		tp = vm_page_next(tp);
1156 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1157 			break;
1158 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1159 			vm_page_xunbusy(tp);
1160 			break;
1161 		}
1162 	}
1163 
1164 	for (p_first = p; count < vm_pageout_page_count; count++) {
1165 		tp = vm_page_prev(p_first);
1166 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1167 			break;
1168 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1169 			vm_page_xunbusy(tp);
1170 			break;
1171 		}
1172 		p_first = tp;
1173 		mreq++;
1174 	}
1175 
1176 	for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1177 		ma[i] = tp;
1178 
1179 	vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1180 	return (runlen);
1181 }
1182 
1183 /*
1184  * Note that there is absolutely no sense in writing out
1185  * anonymous objects, so we track down the vnode object
1186  * to write out.
1187  * We invalidate (remove) all pages from the address space
1188  * for semantic correctness.
1189  *
1190  * If the backing object is a device object with unmanaged pages, then any
1191  * mappings to the specified range of pages must be removed before this
1192  * function is called.
1193  *
1194  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1195  * may start out with a NULL object.
1196  */
1197 boolean_t
1198 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1199     boolean_t syncio, boolean_t invalidate)
1200 {
1201 	vm_object_t backing_object;
1202 	struct vnode *vp;
1203 	struct mount *mp;
1204 	int error, flags, fsync_after;
1205 	boolean_t res;
1206 
1207 	if (object == NULL)
1208 		return (TRUE);
1209 	res = TRUE;
1210 	error = 0;
1211 	VM_OBJECT_WLOCK(object);
1212 	while ((backing_object = object->backing_object) != NULL) {
1213 		VM_OBJECT_WLOCK(backing_object);
1214 		offset += object->backing_object_offset;
1215 		VM_OBJECT_WUNLOCK(object);
1216 		object = backing_object;
1217 		if (object->size < OFF_TO_IDX(offset + size))
1218 			size = IDX_TO_OFF(object->size) - offset;
1219 	}
1220 	/*
1221 	 * Flush pages if writing is allowed, invalidate them
1222 	 * if invalidation requested.  Pages undergoing I/O
1223 	 * will be ignored by vm_object_page_remove().
1224 	 *
1225 	 * We cannot lock the vnode and then wait for paging
1226 	 * to complete without deadlocking against vm_fault.
1227 	 * Instead we simply call vm_object_page_remove() and
1228 	 * allow it to block internally on a page-by-page
1229 	 * basis when it encounters pages undergoing async
1230 	 * I/O.
1231 	 */
1232 	if (object->type == OBJT_VNODE &&
1233 	    vm_object_mightbedirty(object) != 0 &&
1234 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1235 		VM_OBJECT_WUNLOCK(object);
1236 		(void) vn_start_write(vp, &mp, V_WAIT);
1237 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1238 		if (syncio && !invalidate && offset == 0 &&
1239 		    atop(size) == object->size) {
1240 			/*
1241 			 * If syncing the whole mapping of the file,
1242 			 * it is faster to schedule all the writes in
1243 			 * async mode, also allowing the clustering,
1244 			 * and then wait for i/o to complete.
1245 			 */
1246 			flags = 0;
1247 			fsync_after = TRUE;
1248 		} else {
1249 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1250 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1251 			fsync_after = FALSE;
1252 		}
1253 		VM_OBJECT_WLOCK(object);
1254 		res = vm_object_page_clean(object, offset, offset + size,
1255 		    flags);
1256 		VM_OBJECT_WUNLOCK(object);
1257 		if (fsync_after)
1258 			error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1259 		VOP_UNLOCK(vp);
1260 		vn_finished_write(mp);
1261 		if (error != 0)
1262 			res = FALSE;
1263 		VM_OBJECT_WLOCK(object);
1264 	}
1265 	if ((object->type == OBJT_VNODE ||
1266 	     object->type == OBJT_DEVICE) && invalidate) {
1267 		if (object->type == OBJT_DEVICE)
1268 			/*
1269 			 * The option OBJPR_NOTMAPPED must be passed here
1270 			 * because vm_object_page_remove() cannot remove
1271 			 * unmanaged mappings.
1272 			 */
1273 			flags = OBJPR_NOTMAPPED;
1274 		else if (old_msync)
1275 			flags = 0;
1276 		else
1277 			flags = OBJPR_CLEANONLY;
1278 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1279 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1280 	}
1281 	VM_OBJECT_WUNLOCK(object);
1282 	return (res);
1283 }
1284 
1285 /*
1286  * Determine whether the given advice can be applied to the object.  Advice is
1287  * not applied to unmanaged pages since they never belong to page queues, and
1288  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1289  * have been mapped at most once.
1290  */
1291 static bool
1292 vm_object_advice_applies(vm_object_t object, int advice)
1293 {
1294 
1295 	if ((object->flags & OBJ_UNMANAGED) != 0)
1296 		return (false);
1297 	if (advice != MADV_FREE)
1298 		return (true);
1299 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1300 	    (OBJ_ONEMAPPING | OBJ_ANON));
1301 }
1302 
1303 static void
1304 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1305     vm_size_t size)
1306 {
1307 
1308 	if (advice == MADV_FREE)
1309 		vm_pager_freespace(object, pindex, size);
1310 }
1311 
1312 /*
1313  *	vm_object_madvise:
1314  *
1315  *	Implements the madvise function at the object/page level.
1316  *
1317  *	MADV_WILLNEED	(any object)
1318  *
1319  *	    Activate the specified pages if they are resident.
1320  *
1321  *	MADV_DONTNEED	(any object)
1322  *
1323  *	    Deactivate the specified pages if they are resident.
1324  *
1325  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1326  *			 OBJ_ONEMAPPING only)
1327  *
1328  *	    Deactivate and clean the specified pages if they are
1329  *	    resident.  This permits the process to reuse the pages
1330  *	    without faulting or the kernel to reclaim the pages
1331  *	    without I/O.
1332  */
1333 void
1334 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1335     int advice)
1336 {
1337 	vm_pindex_t tpindex;
1338 	vm_object_t backing_object, tobject;
1339 	vm_page_t m, tm;
1340 
1341 	if (object == NULL)
1342 		return;
1343 
1344 relookup:
1345 	VM_OBJECT_WLOCK(object);
1346 	if (!vm_object_advice_applies(object, advice)) {
1347 		VM_OBJECT_WUNLOCK(object);
1348 		return;
1349 	}
1350 	for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1351 		tobject = object;
1352 
1353 		/*
1354 		 * If the next page isn't resident in the top-level object, we
1355 		 * need to search the shadow chain.  When applying MADV_FREE, we
1356 		 * take care to release any swap space used to store
1357 		 * non-resident pages.
1358 		 */
1359 		if (m == NULL || pindex < m->pindex) {
1360 			/*
1361 			 * Optimize a common case: if the top-level object has
1362 			 * no backing object, we can skip over the non-resident
1363 			 * range in constant time.
1364 			 */
1365 			if (object->backing_object == NULL) {
1366 				tpindex = (m != NULL && m->pindex < end) ?
1367 				    m->pindex : end;
1368 				vm_object_madvise_freespace(object, advice,
1369 				    pindex, tpindex - pindex);
1370 				if ((pindex = tpindex) == end)
1371 					break;
1372 				goto next_page;
1373 			}
1374 
1375 			tpindex = pindex;
1376 			do {
1377 				vm_object_madvise_freespace(tobject, advice,
1378 				    tpindex, 1);
1379 				/*
1380 				 * Prepare to search the next object in the
1381 				 * chain.
1382 				 */
1383 				backing_object = tobject->backing_object;
1384 				if (backing_object == NULL)
1385 					goto next_pindex;
1386 				VM_OBJECT_WLOCK(backing_object);
1387 				tpindex +=
1388 				    OFF_TO_IDX(tobject->backing_object_offset);
1389 				if (tobject != object)
1390 					VM_OBJECT_WUNLOCK(tobject);
1391 				tobject = backing_object;
1392 				if (!vm_object_advice_applies(tobject, advice))
1393 					goto next_pindex;
1394 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1395 			    NULL);
1396 		} else {
1397 next_page:
1398 			tm = m;
1399 			m = TAILQ_NEXT(m, listq);
1400 		}
1401 
1402 		/*
1403 		 * If the page is not in a normal state, skip it.  The page
1404 		 * can not be invalidated while the object lock is held.
1405 		 */
1406 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1407 			goto next_pindex;
1408 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1409 		    ("vm_object_madvise: page %p is fictitious", tm));
1410 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1411 		    ("vm_object_madvise: page %p is not managed", tm));
1412 		if (vm_page_tryxbusy(tm) == 0) {
1413 			if (object != tobject)
1414 				VM_OBJECT_WUNLOCK(object);
1415 			if (advice == MADV_WILLNEED) {
1416 				/*
1417 				 * Reference the page before unlocking and
1418 				 * sleeping so that the page daemon is less
1419 				 * likely to reclaim it.
1420 				 */
1421 				vm_page_aflag_set(tm, PGA_REFERENCED);
1422 			}
1423 			if (!vm_page_busy_sleep(tm, "madvpo", 0))
1424 				VM_OBJECT_WUNLOCK(tobject);
1425   			goto relookup;
1426 		}
1427 		vm_page_advise(tm, advice);
1428 		vm_page_xunbusy(tm);
1429 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1430 next_pindex:
1431 		if (tobject != object)
1432 			VM_OBJECT_WUNLOCK(tobject);
1433 	}
1434 	VM_OBJECT_WUNLOCK(object);
1435 }
1436 
1437 /*
1438  *	vm_object_shadow:
1439  *
1440  *	Create a new object which is backed by the
1441  *	specified existing object range.  The source
1442  *	object reference is deallocated.
1443  *
1444  *	The new object and offset into that object
1445  *	are returned in the source parameters.
1446  */
1447 void
1448 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1449     struct ucred *cred, bool shared)
1450 {
1451 	vm_object_t source;
1452 	vm_object_t result;
1453 
1454 	source = *object;
1455 
1456 	/*
1457 	 * Don't create the new object if the old object isn't shared.
1458 	 *
1459 	 * If we hold the only reference we can guarantee that it won't
1460 	 * increase while we have the map locked.  Otherwise the race is
1461 	 * harmless and we will end up with an extra shadow object that
1462 	 * will be collapsed later.
1463 	 */
1464 	if (source != NULL && source->ref_count == 1 &&
1465 	    (source->flags & OBJ_ANON) != 0)
1466 		return;
1467 
1468 	/*
1469 	 * Allocate a new object with the given length.
1470 	 */
1471 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1472 
1473 	/*
1474 	 * Store the offset into the source object, and fix up the offset into
1475 	 * the new object.
1476 	 */
1477 	result->backing_object_offset = *offset;
1478 
1479 	if (shared || source != NULL) {
1480 		VM_OBJECT_WLOCK(result);
1481 
1482 		/*
1483 		 * The new object shadows the source object, adding a
1484 		 * reference to it.  Our caller changes his reference
1485 		 * to point to the new object, removing a reference to
1486 		 * the source object.  Net result: no change of
1487 		 * reference count, unless the caller needs to add one
1488 		 * more reference due to forking a shared map entry.
1489 		 */
1490 		if (shared) {
1491 			vm_object_reference_locked(result);
1492 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1493 		}
1494 
1495 		/*
1496 		 * Try to optimize the result object's page color when
1497 		 * shadowing in order to maintain page coloring
1498 		 * consistency in the combined shadowed object.
1499 		 */
1500 		if (source != NULL) {
1501 			vm_object_backing_insert(result, source);
1502 			result->domain = source->domain;
1503 #if VM_NRESERVLEVEL > 0
1504 			vm_object_set_flag(result,
1505 			    (source->flags & OBJ_COLORED));
1506 			result->pg_color = (source->pg_color +
1507 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1508 			    1)) - 1);
1509 #endif
1510 		}
1511 		VM_OBJECT_WUNLOCK(result);
1512 	}
1513 
1514 	/*
1515 	 * Return the new things
1516 	 */
1517 	*offset = 0;
1518 	*object = result;
1519 }
1520 
1521 /*
1522  *	vm_object_split:
1523  *
1524  * Split the pages in a map entry into a new object.  This affords
1525  * easier removal of unused pages, and keeps object inheritance from
1526  * being a negative impact on memory usage.
1527  */
1528 void
1529 vm_object_split(vm_map_entry_t entry)
1530 {
1531 	vm_page_t m, m_busy, m_next;
1532 	vm_object_t orig_object, new_object, backing_object;
1533 	vm_pindex_t idx, offidxstart;
1534 	vm_size_t size;
1535 
1536 	orig_object = entry->object.vm_object;
1537 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1538 	    ("vm_object_split:  Splitting object with multiple mappings."));
1539 	if ((orig_object->flags & OBJ_ANON) == 0)
1540 		return;
1541 	if (orig_object->ref_count <= 1)
1542 		return;
1543 	VM_OBJECT_WUNLOCK(orig_object);
1544 
1545 	offidxstart = OFF_TO_IDX(entry->offset);
1546 	size = atop(entry->end - entry->start);
1547 
1548 	/*
1549 	 * If swap_pager_copy() is later called, it will convert new_object
1550 	 * into a swap object.
1551 	 */
1552 	new_object = vm_object_allocate_anon(size, orig_object,
1553 	    orig_object->cred, ptoa(size));
1554 
1555 	/*
1556 	 * We must wait for the orig_object to complete any in-progress
1557 	 * collapse so that the swap blocks are stable below.  The
1558 	 * additional reference on backing_object by new object will
1559 	 * prevent further collapse operations until split completes.
1560 	 */
1561 	VM_OBJECT_WLOCK(orig_object);
1562 	vm_object_collapse_wait(orig_object);
1563 
1564 	/*
1565 	 * At this point, the new object is still private, so the order in
1566 	 * which the original and new objects are locked does not matter.
1567 	 */
1568 	VM_OBJECT_WLOCK(new_object);
1569 	new_object->domain = orig_object->domain;
1570 	backing_object = orig_object->backing_object;
1571 	if (backing_object != NULL) {
1572 		vm_object_backing_insert_ref(new_object, backing_object);
1573 		new_object->backing_object_offset =
1574 		    orig_object->backing_object_offset + entry->offset;
1575 	}
1576 	if (orig_object->cred != NULL) {
1577 		crhold(orig_object->cred);
1578 		KASSERT(orig_object->charge >= ptoa(size),
1579 		    ("orig_object->charge < 0"));
1580 		orig_object->charge -= ptoa(size);
1581 	}
1582 
1583 	/*
1584 	 * Mark the split operation so that swap_pager_getpages() knows
1585 	 * that the object is in transition.
1586 	 */
1587 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1588 	m_busy = NULL;
1589 #ifdef INVARIANTS
1590 	idx = 0;
1591 #endif
1592 retry:
1593 	m = vm_page_find_least(orig_object, offidxstart);
1594 	KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1595 	    ("%s: object %p was repopulated", __func__, orig_object));
1596 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1597 	    m = m_next) {
1598 		m_next = TAILQ_NEXT(m, listq);
1599 
1600 		/*
1601 		 * We must wait for pending I/O to complete before we can
1602 		 * rename the page.
1603 		 *
1604 		 * We do not have to VM_PROT_NONE the page as mappings should
1605 		 * not be changed by this operation.
1606 		 */
1607 		if (vm_page_tryxbusy(m) == 0) {
1608 			VM_OBJECT_WUNLOCK(new_object);
1609 			if (vm_page_busy_sleep(m, "spltwt", 0))
1610 				VM_OBJECT_WLOCK(orig_object);
1611 			VM_OBJECT_WLOCK(new_object);
1612 			goto retry;
1613 		}
1614 
1615 		/*
1616 		 * The page was left invalid.  Likely placed there by
1617 		 * an incomplete fault.  Just remove and ignore.
1618 		 */
1619 		if (vm_page_none_valid(m)) {
1620 			if (vm_page_remove(m))
1621 				vm_page_free(m);
1622 			continue;
1623 		}
1624 
1625 		/* vm_page_rename() will dirty the page. */
1626 		if (vm_page_rename(m, new_object, idx)) {
1627 			vm_page_xunbusy(m);
1628 			VM_OBJECT_WUNLOCK(new_object);
1629 			VM_OBJECT_WUNLOCK(orig_object);
1630 			vm_radix_wait();
1631 			VM_OBJECT_WLOCK(orig_object);
1632 			VM_OBJECT_WLOCK(new_object);
1633 			goto retry;
1634 		}
1635 
1636 #if VM_NRESERVLEVEL > 0
1637 		/*
1638 		 * If some of the reservation's allocated pages remain with
1639 		 * the original object, then transferring the reservation to
1640 		 * the new object is neither particularly beneficial nor
1641 		 * particularly harmful as compared to leaving the reservation
1642 		 * with the original object.  If, however, all of the
1643 		 * reservation's allocated pages are transferred to the new
1644 		 * object, then transferring the reservation is typically
1645 		 * beneficial.  Determining which of these two cases applies
1646 		 * would be more costly than unconditionally renaming the
1647 		 * reservation.
1648 		 */
1649 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1650 #endif
1651 
1652 		/*
1653 		 * orig_object's type may change while sleeping, so keep track
1654 		 * of the beginning of the busied range.
1655 		 */
1656 		if (orig_object->type != OBJT_SWAP)
1657 			vm_page_xunbusy(m);
1658 		else if (m_busy == NULL)
1659 			m_busy = m;
1660 	}
1661 	if ((orig_object->flags & OBJ_SWAP) != 0) {
1662 		/*
1663 		 * swap_pager_copy() can sleep, in which case the orig_object's
1664 		 * and new_object's locks are released and reacquired.
1665 		 */
1666 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1667 		if (m_busy != NULL)
1668 			TAILQ_FOREACH_FROM(m_busy, &new_object->memq, listq)
1669 				vm_page_xunbusy(m_busy);
1670 	}
1671 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1672 	VM_OBJECT_WUNLOCK(orig_object);
1673 	VM_OBJECT_WUNLOCK(new_object);
1674 	entry->object.vm_object = new_object;
1675 	entry->offset = 0LL;
1676 	vm_object_deallocate(orig_object);
1677 	VM_OBJECT_WLOCK(new_object);
1678 }
1679 
1680 static vm_page_t
1681 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1682 {
1683 	vm_object_t backing_object;
1684 
1685 	VM_OBJECT_ASSERT_WLOCKED(object);
1686 	backing_object = object->backing_object;
1687 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1688 
1689 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1690 	    ("invalid ownership %p %p %p", p, object, backing_object));
1691 	/* The page is only NULL when rename fails. */
1692 	if (p == NULL) {
1693 		VM_OBJECT_WUNLOCK(object);
1694 		VM_OBJECT_WUNLOCK(backing_object);
1695 		vm_radix_wait();
1696 		VM_OBJECT_WLOCK(object);
1697 	} else if (p->object == object) {
1698 		VM_OBJECT_WUNLOCK(backing_object);
1699 		if (vm_page_busy_sleep(p, "vmocol", 0))
1700 			VM_OBJECT_WLOCK(object);
1701 	} else {
1702 		VM_OBJECT_WUNLOCK(object);
1703 		if (!vm_page_busy_sleep(p, "vmocol", 0))
1704 			VM_OBJECT_WUNLOCK(backing_object);
1705 		VM_OBJECT_WLOCK(object);
1706 	}
1707 	VM_OBJECT_WLOCK(backing_object);
1708 	return (TAILQ_FIRST(&backing_object->memq));
1709 }
1710 
1711 static bool
1712 vm_object_scan_all_shadowed(vm_object_t object)
1713 {
1714 	vm_object_t backing_object;
1715 	vm_page_t p, pp;
1716 	vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1717 
1718 	VM_OBJECT_ASSERT_WLOCKED(object);
1719 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1720 
1721 	backing_object = object->backing_object;
1722 
1723 	if ((backing_object->flags & OBJ_ANON) == 0)
1724 		return (false);
1725 
1726 	pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1727 	p = vm_page_find_least(backing_object, pi);
1728 	ps = swap_pager_find_least(backing_object, pi);
1729 
1730 	/*
1731 	 * Only check pages inside the parent object's range and
1732 	 * inside the parent object's mapping of the backing object.
1733 	 */
1734 	for (;; pi++) {
1735 		if (p != NULL && p->pindex < pi)
1736 			p = TAILQ_NEXT(p, listq);
1737 		if (ps < pi)
1738 			ps = swap_pager_find_least(backing_object, pi);
1739 		if (p == NULL && ps >= backing_object->size)
1740 			break;
1741 		else if (p == NULL)
1742 			pi = ps;
1743 		else
1744 			pi = MIN(p->pindex, ps);
1745 
1746 		new_pindex = pi - backing_offset_index;
1747 		if (new_pindex >= object->size)
1748 			break;
1749 
1750 		if (p != NULL) {
1751 			/*
1752 			 * If the backing object page is busy a
1753 			 * grandparent or older page may still be
1754 			 * undergoing CoW.  It is not safe to collapse
1755 			 * the backing object until it is quiesced.
1756 			 */
1757 			if (vm_page_tryxbusy(p) == 0)
1758 				return (false);
1759 
1760 			/*
1761 			 * We raced with the fault handler that left
1762 			 * newly allocated invalid page on the object
1763 			 * queue and retried.
1764 			 */
1765 			if (!vm_page_all_valid(p))
1766 				goto unbusy_ret;
1767 		}
1768 
1769 		/*
1770 		 * See if the parent has the page or if the parent's object
1771 		 * pager has the page.  If the parent has the page but the page
1772 		 * is not valid, the parent's object pager must have the page.
1773 		 *
1774 		 * If this fails, the parent does not completely shadow the
1775 		 * object and we might as well give up now.
1776 		 */
1777 		pp = vm_page_lookup(object, new_pindex);
1778 
1779 		/*
1780 		 * The valid check here is stable due to object lock
1781 		 * being required to clear valid and initiate paging.
1782 		 * Busy of p disallows fault handler to validate pp.
1783 		 */
1784 		if ((pp == NULL || vm_page_none_valid(pp)) &&
1785 		    !vm_pager_has_page(object, new_pindex, NULL, NULL))
1786 			goto unbusy_ret;
1787 		if (p != NULL)
1788 			vm_page_xunbusy(p);
1789 	}
1790 	return (true);
1791 
1792 unbusy_ret:
1793 	if (p != NULL)
1794 		vm_page_xunbusy(p);
1795 	return (false);
1796 }
1797 
1798 static void
1799 vm_object_collapse_scan(vm_object_t object)
1800 {
1801 	vm_object_t backing_object;
1802 	vm_page_t next, p, pp;
1803 	vm_pindex_t backing_offset_index, new_pindex;
1804 
1805 	VM_OBJECT_ASSERT_WLOCKED(object);
1806 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1807 
1808 	backing_object = object->backing_object;
1809 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1810 
1811 	/*
1812 	 * Our scan
1813 	 */
1814 	for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1815 		next = TAILQ_NEXT(p, listq);
1816 		new_pindex = p->pindex - backing_offset_index;
1817 
1818 		/*
1819 		 * Check for busy page
1820 		 */
1821 		if (vm_page_tryxbusy(p) == 0) {
1822 			next = vm_object_collapse_scan_wait(object, p);
1823 			continue;
1824 		}
1825 
1826 		KASSERT(object->backing_object == backing_object,
1827 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1828 		    object->backing_object, backing_object));
1829 		KASSERT(p->object == backing_object,
1830 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1831 		    p->object, backing_object));
1832 
1833 		if (p->pindex < backing_offset_index ||
1834 		    new_pindex >= object->size) {
1835 			vm_pager_freespace(backing_object, p->pindex, 1);
1836 
1837 			KASSERT(!pmap_page_is_mapped(p),
1838 			    ("freeing mapped page %p", p));
1839 			if (vm_page_remove(p))
1840 				vm_page_free(p);
1841 			continue;
1842 		}
1843 
1844 		if (!vm_page_all_valid(p)) {
1845 			KASSERT(!pmap_page_is_mapped(p),
1846 			    ("freeing mapped page %p", p));
1847 			if (vm_page_remove(p))
1848 				vm_page_free(p);
1849 			continue;
1850 		}
1851 
1852 		pp = vm_page_lookup(object, new_pindex);
1853 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1854 			vm_page_xunbusy(p);
1855 			/*
1856 			 * The page in the parent is busy and possibly not
1857 			 * (yet) valid.  Until its state is finalized by the
1858 			 * busy bit owner, we can't tell whether it shadows the
1859 			 * original page.
1860 			 */
1861 			next = vm_object_collapse_scan_wait(object, pp);
1862 			continue;
1863 		}
1864 
1865 		if (pp != NULL && vm_page_none_valid(pp)) {
1866 			/*
1867 			 * The page was invalid in the parent.  Likely placed
1868 			 * there by an incomplete fault.  Just remove and
1869 			 * ignore.  p can replace it.
1870 			 */
1871 			if (vm_page_remove(pp))
1872 				vm_page_free(pp);
1873 			pp = NULL;
1874 		}
1875 
1876 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1877 			NULL)) {
1878 			/*
1879 			 * The page already exists in the parent OR swap exists
1880 			 * for this location in the parent.  Leave the parent's
1881 			 * page alone.  Destroy the original page from the
1882 			 * backing object.
1883 			 */
1884 			vm_pager_freespace(backing_object, p->pindex, 1);
1885 			KASSERT(!pmap_page_is_mapped(p),
1886 			    ("freeing mapped page %p", p));
1887 			if (vm_page_remove(p))
1888 				vm_page_free(p);
1889 			if (pp != NULL)
1890 				vm_page_xunbusy(pp);
1891 			continue;
1892 		}
1893 
1894 		/*
1895 		 * Page does not exist in parent, rename the page from the
1896 		 * backing object to the main object.
1897 		 *
1898 		 * If the page was mapped to a process, it can remain mapped
1899 		 * through the rename.  vm_page_rename() will dirty the page.
1900 		 */
1901 		if (vm_page_rename(p, object, new_pindex)) {
1902 			vm_page_xunbusy(p);
1903 			next = vm_object_collapse_scan_wait(object, NULL);
1904 			continue;
1905 		}
1906 
1907 		/* Use the old pindex to free the right page. */
1908 		vm_pager_freespace(backing_object, new_pindex +
1909 		    backing_offset_index, 1);
1910 
1911 #if VM_NRESERVLEVEL > 0
1912 		/*
1913 		 * Rename the reservation.
1914 		 */
1915 		vm_reserv_rename(p, object, backing_object,
1916 		    backing_offset_index);
1917 #endif
1918 		vm_page_xunbusy(p);
1919 	}
1920 	return;
1921 }
1922 
1923 /*
1924  *	vm_object_collapse:
1925  *
1926  *	Collapse an object with the object backing it.
1927  *	Pages in the backing object are moved into the
1928  *	parent, and the backing object is deallocated.
1929  */
1930 void
1931 vm_object_collapse(vm_object_t object)
1932 {
1933 	vm_object_t backing_object, new_backing_object;
1934 
1935 	VM_OBJECT_ASSERT_WLOCKED(object);
1936 
1937 	while (TRUE) {
1938 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1939 		    ("collapsing invalid object"));
1940 
1941 		/*
1942 		 * Wait for the backing_object to finish any pending
1943 		 * collapse so that the caller sees the shortest possible
1944 		 * shadow chain.
1945 		 */
1946 		backing_object = vm_object_backing_collapse_wait(object);
1947 		if (backing_object == NULL)
1948 			return;
1949 
1950 		KASSERT(object->ref_count > 0 &&
1951 		    object->ref_count > atomic_load_int(&object->shadow_count),
1952 		    ("collapse with invalid ref %d or shadow %d count.",
1953 		    object->ref_count, atomic_load_int(&object->shadow_count)));
1954 		KASSERT((backing_object->flags &
1955 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1956 		    ("vm_object_collapse: Backing object already collapsing."));
1957 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1958 		    ("vm_object_collapse: object is already collapsing."));
1959 
1960 		/*
1961 		 * We know that we can either collapse the backing object if
1962 		 * the parent is the only reference to it, or (perhaps) have
1963 		 * the parent bypass the object if the parent happens to shadow
1964 		 * all the resident pages in the entire backing object.
1965 		 */
1966 		if (backing_object->ref_count == 1) {
1967 			KASSERT(atomic_load_int(&backing_object->shadow_count)
1968 			    == 1,
1969 			    ("vm_object_collapse: shadow_count: %d",
1970 			    atomic_load_int(&backing_object->shadow_count)));
1971 			vm_object_pip_add(object, 1);
1972 			vm_object_set_flag(object, OBJ_COLLAPSING);
1973 			vm_object_pip_add(backing_object, 1);
1974 			vm_object_set_flag(backing_object, OBJ_DEAD);
1975 
1976 			/*
1977 			 * If there is exactly one reference to the backing
1978 			 * object, we can collapse it into the parent.
1979 			 */
1980 			vm_object_collapse_scan(object);
1981 
1982 #if VM_NRESERVLEVEL > 0
1983 			/*
1984 			 * Break any reservations from backing_object.
1985 			 */
1986 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1987 				vm_reserv_break_all(backing_object);
1988 #endif
1989 
1990 			/*
1991 			 * Move the pager from backing_object to object.
1992 			 */
1993 			if ((backing_object->flags & OBJ_SWAP) != 0) {
1994 				/*
1995 				 * swap_pager_copy() can sleep, in which case
1996 				 * the backing_object's and object's locks are
1997 				 * released and reacquired.
1998 				 * Since swap_pager_copy() is being asked to
1999 				 * destroy backing_object, it will change the
2000 				 * type to OBJT_DEFAULT.
2001 				 */
2002 				swap_pager_copy(
2003 				    backing_object,
2004 				    object,
2005 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
2006 			}
2007 
2008 			/*
2009 			 * Object now shadows whatever backing_object did.
2010 			 */
2011 			vm_object_clear_flag(object, OBJ_COLLAPSING);
2012 			vm_object_backing_transfer(object, backing_object);
2013 			object->backing_object_offset +=
2014 			    backing_object->backing_object_offset;
2015 			VM_OBJECT_WUNLOCK(object);
2016 			vm_object_pip_wakeup(object);
2017 
2018 			/*
2019 			 * Discard backing_object.
2020 			 *
2021 			 * Since the backing object has no pages, no pager left,
2022 			 * and no object references within it, all that is
2023 			 * necessary is to dispose of it.
2024 			 */
2025 			KASSERT(backing_object->ref_count == 1, (
2026 "backing_object %p was somehow re-referenced during collapse!",
2027 			    backing_object));
2028 			vm_object_pip_wakeup(backing_object);
2029 			(void)refcount_release(&backing_object->ref_count);
2030 			vm_object_terminate(backing_object);
2031 			counter_u64_add(object_collapses, 1);
2032 			VM_OBJECT_WLOCK(object);
2033 		} else {
2034 			/*
2035 			 * If we do not entirely shadow the backing object,
2036 			 * there is nothing we can do so we give up.
2037 			 *
2038 			 * The object lock and backing_object lock must not
2039 			 * be dropped during this sequence.
2040 			 */
2041 			if (!vm_object_scan_all_shadowed(object)) {
2042 				VM_OBJECT_WUNLOCK(backing_object);
2043 				break;
2044 			}
2045 
2046 			/*
2047 			 * Make the parent shadow the next object in the
2048 			 * chain.  Deallocating backing_object will not remove
2049 			 * it, since its reference count is at least 2.
2050 			 */
2051 			vm_object_backing_remove_locked(object);
2052 			new_backing_object = backing_object->backing_object;
2053 			if (new_backing_object != NULL) {
2054 				vm_object_backing_insert_ref(object,
2055 				    new_backing_object);
2056 				object->backing_object_offset +=
2057 				    backing_object->backing_object_offset;
2058 			}
2059 
2060 			/*
2061 			 * Drop the reference count on backing_object. Since
2062 			 * its ref_count was at least 2, it will not vanish.
2063 			 */
2064 			(void)refcount_release(&backing_object->ref_count);
2065 			KASSERT(backing_object->ref_count >= 1, (
2066 "backing_object %p was somehow dereferenced during collapse!",
2067 			    backing_object));
2068 			VM_OBJECT_WUNLOCK(backing_object);
2069 			counter_u64_add(object_bypasses, 1);
2070 		}
2071 
2072 		/*
2073 		 * Try again with this object's new backing object.
2074 		 */
2075 	}
2076 }
2077 
2078 /*
2079  *	vm_object_page_remove:
2080  *
2081  *	For the given object, either frees or invalidates each of the
2082  *	specified pages.  In general, a page is freed.  However, if a page is
2083  *	wired for any reason other than the existence of a managed, wired
2084  *	mapping, then it may be invalidated but not removed from the object.
2085  *	Pages are specified by the given range ["start", "end") and the option
2086  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
2087  *	extends from "start" to the end of the object.  If the option
2088  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2089  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
2090  *	specified, then the pages within the specified range must have no
2091  *	mappings.  Otherwise, if this option is not specified, any mappings to
2092  *	the specified pages are removed before the pages are freed or
2093  *	invalidated.
2094  *
2095  *	In general, this operation should only be performed on objects that
2096  *	contain managed pages.  There are, however, two exceptions.  First, it
2097  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
2098  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2099  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
2100  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
2101  *
2102  *	The object must be locked.
2103  */
2104 void
2105 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2106     int options)
2107 {
2108 	vm_page_t p, next;
2109 
2110 	VM_OBJECT_ASSERT_WLOCKED(object);
2111 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2112 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2113 	    ("vm_object_page_remove: illegal options for object %p", object));
2114 	if (object->resident_page_count == 0)
2115 		return;
2116 	vm_object_pip_add(object, 1);
2117 again:
2118 	p = vm_page_find_least(object, start);
2119 
2120 	/*
2121 	 * Here, the variable "p" is either (1) the page with the least pindex
2122 	 * greater than or equal to the parameter "start" or (2) NULL.
2123 	 */
2124 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2125 		next = TAILQ_NEXT(p, listq);
2126 
2127 		/*
2128 		 * Skip invalid pages if asked to do so.  Try to avoid acquiring
2129 		 * the busy lock, as some consumers rely on this to avoid
2130 		 * deadlocks.
2131 		 *
2132 		 * A thread may concurrently transition the page from invalid to
2133 		 * valid using only the busy lock, so the result of this check
2134 		 * is immediately stale.  It is up to consumers to handle this,
2135 		 * for instance by ensuring that all invalid->valid transitions
2136 		 * happen with a mutex held, as may be possible for a
2137 		 * filesystem.
2138 		 */
2139 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2140 			continue;
2141 
2142 		/*
2143 		 * If the page is wired for any reason besides the existence
2144 		 * of managed, wired mappings, then it cannot be freed.  For
2145 		 * example, fictitious pages, which represent device memory,
2146 		 * are inherently wired and cannot be freed.  They can,
2147 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2148 		 * not specified.
2149 		 */
2150 		if (vm_page_tryxbusy(p) == 0) {
2151 			if (vm_page_busy_sleep(p, "vmopar", 0))
2152 				VM_OBJECT_WLOCK(object);
2153 			goto again;
2154 		}
2155 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2156 			vm_page_xunbusy(p);
2157 			continue;
2158 		}
2159 		if (vm_page_wired(p)) {
2160 wired:
2161 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2162 			    object->ref_count != 0)
2163 				pmap_remove_all(p);
2164 			if ((options & OBJPR_CLEANONLY) == 0) {
2165 				vm_page_invalid(p);
2166 				vm_page_undirty(p);
2167 			}
2168 			vm_page_xunbusy(p);
2169 			continue;
2170 		}
2171 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2172 		    ("vm_object_page_remove: page %p is fictitious", p));
2173 		if ((options & OBJPR_CLEANONLY) != 0 &&
2174 		    !vm_page_none_valid(p)) {
2175 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2176 			    object->ref_count != 0 &&
2177 			    !vm_page_try_remove_write(p))
2178 				goto wired;
2179 			if (p->dirty != 0) {
2180 				vm_page_xunbusy(p);
2181 				continue;
2182 			}
2183 		}
2184 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2185 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2186 			goto wired;
2187 		vm_page_free(p);
2188 	}
2189 	vm_object_pip_wakeup(object);
2190 
2191 	vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2192 	    start);
2193 }
2194 
2195 /*
2196  *	vm_object_page_noreuse:
2197  *
2198  *	For the given object, attempt to move the specified pages to
2199  *	the head of the inactive queue.  This bypasses regular LRU
2200  *	operation and allows the pages to be reused quickly under memory
2201  *	pressure.  If a page is wired for any reason, then it will not
2202  *	be queued.  Pages are specified by the range ["start", "end").
2203  *	As a special case, if "end" is zero, then the range extends from
2204  *	"start" to the end of the object.
2205  *
2206  *	This operation should only be performed on objects that
2207  *	contain non-fictitious, managed pages.
2208  *
2209  *	The object must be locked.
2210  */
2211 void
2212 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2213 {
2214 	vm_page_t p, next;
2215 
2216 	VM_OBJECT_ASSERT_LOCKED(object);
2217 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2218 	    ("vm_object_page_noreuse: illegal object %p", object));
2219 	if (object->resident_page_count == 0)
2220 		return;
2221 	p = vm_page_find_least(object, start);
2222 
2223 	/*
2224 	 * Here, the variable "p" is either (1) the page with the least pindex
2225 	 * greater than or equal to the parameter "start" or (2) NULL.
2226 	 */
2227 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2228 		next = TAILQ_NEXT(p, listq);
2229 		vm_page_deactivate_noreuse(p);
2230 	}
2231 }
2232 
2233 /*
2234  *	Populate the specified range of the object with valid pages.  Returns
2235  *	TRUE if the range is successfully populated and FALSE otherwise.
2236  *
2237  *	Note: This function should be optimized to pass a larger array of
2238  *	pages to vm_pager_get_pages() before it is applied to a non-
2239  *	OBJT_DEVICE object.
2240  *
2241  *	The object must be locked.
2242  */
2243 boolean_t
2244 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2245 {
2246 	vm_page_t m;
2247 	vm_pindex_t pindex;
2248 	int rv;
2249 
2250 	VM_OBJECT_ASSERT_WLOCKED(object);
2251 	for (pindex = start; pindex < end; pindex++) {
2252 		rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2253 		if (rv != VM_PAGER_OK)
2254 			break;
2255 
2256 		/*
2257 		 * Keep "m" busy because a subsequent iteration may unlock
2258 		 * the object.
2259 		 */
2260 	}
2261 	if (pindex > start) {
2262 		m = vm_page_lookup(object, start);
2263 		while (m != NULL && m->pindex < pindex) {
2264 			vm_page_xunbusy(m);
2265 			m = TAILQ_NEXT(m, listq);
2266 		}
2267 	}
2268 	return (pindex == end);
2269 }
2270 
2271 /*
2272  *	Routine:	vm_object_coalesce
2273  *	Function:	Coalesces two objects backing up adjoining
2274  *			regions of memory into a single object.
2275  *
2276  *	returns TRUE if objects were combined.
2277  *
2278  *	NOTE:	Only works at the moment if the second object is NULL -
2279  *		if it's not, which object do we lock first?
2280  *
2281  *	Parameters:
2282  *		prev_object	First object to coalesce
2283  *		prev_offset	Offset into prev_object
2284  *		prev_size	Size of reference to prev_object
2285  *		next_size	Size of reference to the second object
2286  *		reserved	Indicator that extension region has
2287  *				swap accounted for
2288  *
2289  *	Conditions:
2290  *	The object must *not* be locked.
2291  */
2292 boolean_t
2293 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2294     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2295 {
2296 	vm_pindex_t next_pindex;
2297 
2298 	if (prev_object == NULL)
2299 		return (TRUE);
2300 	if ((prev_object->flags & OBJ_ANON) == 0)
2301 		return (FALSE);
2302 
2303 	VM_OBJECT_WLOCK(prev_object);
2304 	/*
2305 	 * Try to collapse the object first.
2306 	 */
2307 	vm_object_collapse(prev_object);
2308 
2309 	/*
2310 	 * Can't coalesce if: . more than one reference . paged out . shadows
2311 	 * another object . has a copy elsewhere (any of which mean that the
2312 	 * pages not mapped to prev_entry may be in use anyway)
2313 	 */
2314 	if (prev_object->backing_object != NULL) {
2315 		VM_OBJECT_WUNLOCK(prev_object);
2316 		return (FALSE);
2317 	}
2318 
2319 	prev_size >>= PAGE_SHIFT;
2320 	next_size >>= PAGE_SHIFT;
2321 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2322 
2323 	if (prev_object->ref_count > 1 &&
2324 	    prev_object->size != next_pindex &&
2325 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2326 		VM_OBJECT_WUNLOCK(prev_object);
2327 		return (FALSE);
2328 	}
2329 
2330 	/*
2331 	 * Account for the charge.
2332 	 */
2333 	if (prev_object->cred != NULL) {
2334 		/*
2335 		 * If prev_object was charged, then this mapping,
2336 		 * although not charged now, may become writable
2337 		 * later. Non-NULL cred in the object would prevent
2338 		 * swap reservation during enabling of the write
2339 		 * access, so reserve swap now. Failed reservation
2340 		 * cause allocation of the separate object for the map
2341 		 * entry, and swap reservation for this entry is
2342 		 * managed in appropriate time.
2343 		 */
2344 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2345 		    prev_object->cred)) {
2346 			VM_OBJECT_WUNLOCK(prev_object);
2347 			return (FALSE);
2348 		}
2349 		prev_object->charge += ptoa(next_size);
2350 	}
2351 
2352 	/*
2353 	 * Remove any pages that may still be in the object from a previous
2354 	 * deallocation.
2355 	 */
2356 	if (next_pindex < prev_object->size) {
2357 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2358 		    next_size, 0);
2359 #if 0
2360 		if (prev_object->cred != NULL) {
2361 			KASSERT(prev_object->charge >=
2362 			    ptoa(prev_object->size - next_pindex),
2363 			    ("object %p overcharged 1 %jx %jx", prev_object,
2364 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2365 			prev_object->charge -= ptoa(prev_object->size -
2366 			    next_pindex);
2367 		}
2368 #endif
2369 	}
2370 
2371 	/*
2372 	 * Extend the object if necessary.
2373 	 */
2374 	if (next_pindex + next_size > prev_object->size)
2375 		prev_object->size = next_pindex + next_size;
2376 
2377 	VM_OBJECT_WUNLOCK(prev_object);
2378 	return (TRUE);
2379 }
2380 
2381 void
2382 vm_object_set_writeable_dirty_(vm_object_t object)
2383 {
2384 	atomic_add_int(&object->generation, 1);
2385 }
2386 
2387 bool
2388 vm_object_mightbedirty_(vm_object_t object)
2389 {
2390 	return (object->generation != object->cleangeneration);
2391 }
2392 
2393 /*
2394  *	vm_object_unwire:
2395  *
2396  *	For each page offset within the specified range of the given object,
2397  *	find the highest-level page in the shadow chain and unwire it.  A page
2398  *	must exist at every page offset, and the highest-level page must be
2399  *	wired.
2400  */
2401 void
2402 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2403     uint8_t queue)
2404 {
2405 	vm_object_t tobject, t1object;
2406 	vm_page_t m, tm;
2407 	vm_pindex_t end_pindex, pindex, tpindex;
2408 	int depth, locked_depth;
2409 
2410 	KASSERT((offset & PAGE_MASK) == 0,
2411 	    ("vm_object_unwire: offset is not page aligned"));
2412 	KASSERT((length & PAGE_MASK) == 0,
2413 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2414 	/* The wired count of a fictitious page never changes. */
2415 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2416 		return;
2417 	pindex = OFF_TO_IDX(offset);
2418 	end_pindex = pindex + atop(length);
2419 again:
2420 	locked_depth = 1;
2421 	VM_OBJECT_RLOCK(object);
2422 	m = vm_page_find_least(object, pindex);
2423 	while (pindex < end_pindex) {
2424 		if (m == NULL || pindex < m->pindex) {
2425 			/*
2426 			 * The first object in the shadow chain doesn't
2427 			 * contain a page at the current index.  Therefore,
2428 			 * the page must exist in a backing object.
2429 			 */
2430 			tobject = object;
2431 			tpindex = pindex;
2432 			depth = 0;
2433 			do {
2434 				tpindex +=
2435 				    OFF_TO_IDX(tobject->backing_object_offset);
2436 				tobject = tobject->backing_object;
2437 				KASSERT(tobject != NULL,
2438 				    ("vm_object_unwire: missing page"));
2439 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2440 					goto next_page;
2441 				depth++;
2442 				if (depth == locked_depth) {
2443 					locked_depth++;
2444 					VM_OBJECT_RLOCK(tobject);
2445 				}
2446 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2447 			    NULL);
2448 		} else {
2449 			tm = m;
2450 			m = TAILQ_NEXT(m, listq);
2451 		}
2452 		if (vm_page_trysbusy(tm) == 0) {
2453 			for (tobject = object; locked_depth >= 1;
2454 			    locked_depth--) {
2455 				t1object = tobject->backing_object;
2456 				if (tm->object != tobject)
2457 					VM_OBJECT_RUNLOCK(tobject);
2458 				tobject = t1object;
2459 			}
2460 			tobject = tm->object;
2461 			if (!vm_page_busy_sleep(tm, "unwbo",
2462 			    VM_ALLOC_IGN_SBUSY))
2463 				VM_OBJECT_RUNLOCK(tobject);
2464 			goto again;
2465 		}
2466 		vm_page_unwire(tm, queue);
2467 		vm_page_sunbusy(tm);
2468 next_page:
2469 		pindex++;
2470 	}
2471 	/* Release the accumulated object locks. */
2472 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2473 		t1object = tobject->backing_object;
2474 		VM_OBJECT_RUNLOCK(tobject);
2475 		tobject = t1object;
2476 	}
2477 }
2478 
2479 /*
2480  * Return the vnode for the given object, or NULL if none exists.
2481  * For tmpfs objects, the function may return NULL if there is
2482  * no vnode allocated at the time of the call.
2483  */
2484 struct vnode *
2485 vm_object_vnode(vm_object_t object)
2486 {
2487 	struct vnode *vp;
2488 
2489 	VM_OBJECT_ASSERT_LOCKED(object);
2490 	vm_pager_getvp(object, &vp, NULL);
2491 	return (vp);
2492 }
2493 
2494 /*
2495  * Busy the vm object.  This prevents new pages belonging to the object from
2496  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2497  * for checking page state before proceeding.
2498  */
2499 void
2500 vm_object_busy(vm_object_t obj)
2501 {
2502 
2503 	VM_OBJECT_ASSERT_LOCKED(obj);
2504 
2505 	blockcount_acquire(&obj->busy, 1);
2506 	/* The fence is required to order loads of page busy. */
2507 	atomic_thread_fence_acq_rel();
2508 }
2509 
2510 void
2511 vm_object_unbusy(vm_object_t obj)
2512 {
2513 
2514 	blockcount_release(&obj->busy, 1);
2515 }
2516 
2517 void
2518 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2519 {
2520 
2521 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2522 
2523 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2524 }
2525 
2526 /*
2527  * This function aims to determine if the object is mapped,
2528  * specifically, if it is referenced by a vm_map_entry.  Because
2529  * objects occasionally acquire transient references that do not
2530  * represent a mapping, the method used here is inexact.  However, it
2531  * has very low overhead and is good enough for the advisory
2532  * vm.vmtotal sysctl.
2533  */
2534 bool
2535 vm_object_is_active(vm_object_t obj)
2536 {
2537 
2538 	return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2539 }
2540 
2541 static int
2542 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2543 {
2544 	struct kinfo_vmobject *kvo;
2545 	char *fullpath, *freepath;
2546 	struct vnode *vp;
2547 	struct vattr va;
2548 	vm_object_t obj;
2549 	vm_page_t m;
2550 	u_long sp;
2551 	int count, error;
2552 
2553 	if (req->oldptr == NULL) {
2554 		/*
2555 		 * If an old buffer has not been provided, generate an
2556 		 * estimate of the space needed for a subsequent call.
2557 		 */
2558 		mtx_lock(&vm_object_list_mtx);
2559 		count = 0;
2560 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2561 			if (obj->type == OBJT_DEAD)
2562 				continue;
2563 			count++;
2564 		}
2565 		mtx_unlock(&vm_object_list_mtx);
2566 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2567 		    count * 11 / 10));
2568 	}
2569 
2570 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK);
2571 	error = 0;
2572 
2573 	/*
2574 	 * VM objects are type stable and are never removed from the
2575 	 * list once added.  This allows us to safely read obj->object_list
2576 	 * after reacquiring the VM object lock.
2577 	 */
2578 	mtx_lock(&vm_object_list_mtx);
2579 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2580 		if (obj->type == OBJT_DEAD ||
2581 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2582 			continue;
2583 		VM_OBJECT_RLOCK(obj);
2584 		if (obj->type == OBJT_DEAD ||
2585 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2586 			VM_OBJECT_RUNLOCK(obj);
2587 			continue;
2588 		}
2589 		mtx_unlock(&vm_object_list_mtx);
2590 		kvo->kvo_size = ptoa(obj->size);
2591 		kvo->kvo_resident = obj->resident_page_count;
2592 		kvo->kvo_ref_count = obj->ref_count;
2593 		kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2594 		kvo->kvo_memattr = obj->memattr;
2595 		kvo->kvo_active = 0;
2596 		kvo->kvo_inactive = 0;
2597 		if (!swap_only) {
2598 			TAILQ_FOREACH(m, &obj->memq, listq) {
2599 				/*
2600 				 * A page may belong to the object but be
2601 				 * dequeued and set to PQ_NONE while the
2602 				 * object lock is not held.  This makes the
2603 				 * reads of m->queue below racy, and we do not
2604 				 * count pages set to PQ_NONE.  However, this
2605 				 * sysctl is only meant to give an
2606 				 * approximation of the system anyway.
2607 				 */
2608 				if (m->a.queue == PQ_ACTIVE)
2609 					kvo->kvo_active++;
2610 				else if (m->a.queue == PQ_INACTIVE)
2611 					kvo->kvo_inactive++;
2612 			}
2613 		}
2614 
2615 		kvo->kvo_vn_fileid = 0;
2616 		kvo->kvo_vn_fsid = 0;
2617 		kvo->kvo_vn_fsid_freebsd11 = 0;
2618 		freepath = NULL;
2619 		fullpath = "";
2620 		vp = NULL;
2621 		kvo->kvo_type = vm_object_kvme_type(obj, swap_only ? NULL : &vp);
2622 		if (vp != NULL) {
2623 			vref(vp);
2624 		} else if ((obj->flags & OBJ_ANON) != 0) {
2625 			MPASS(kvo->kvo_type == KVME_TYPE_DEFAULT ||
2626 			    kvo->kvo_type == KVME_TYPE_SWAP);
2627 			kvo->kvo_me = (uintptr_t)obj;
2628 			/* tmpfs objs are reported as vnodes */
2629 			kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2630 			sp = swap_pager_swapped_pages(obj);
2631 			kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2632 		}
2633 		VM_OBJECT_RUNLOCK(obj);
2634 		if (vp != NULL) {
2635 			vn_fullpath(vp, &fullpath, &freepath);
2636 			vn_lock(vp, LK_SHARED | LK_RETRY);
2637 			if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2638 				kvo->kvo_vn_fileid = va.va_fileid;
2639 				kvo->kvo_vn_fsid = va.va_fsid;
2640 				kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2641 								/* truncate */
2642 			}
2643 			vput(vp);
2644 		}
2645 
2646 		strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2647 		if (freepath != NULL)
2648 			free(freepath, M_TEMP);
2649 
2650 		/* Pack record size down */
2651 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2652 		    + strlen(kvo->kvo_path) + 1;
2653 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2654 		    sizeof(uint64_t));
2655 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2656 		maybe_yield();
2657 		mtx_lock(&vm_object_list_mtx);
2658 		if (error)
2659 			break;
2660 	}
2661 	mtx_unlock(&vm_object_list_mtx);
2662 	free(kvo, M_TEMP);
2663 	return (error);
2664 }
2665 
2666 static int
2667 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2668 {
2669 	return (vm_object_list_handler(req, false));
2670 }
2671 
2672 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2673     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2674     "List of VM objects");
2675 
2676 static int
2677 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2678 {
2679 	return (vm_object_list_handler(req, true));
2680 }
2681 
2682 /*
2683  * This sysctl returns list of the anonymous or swap objects. Intent
2684  * is to provide stripped optimized list useful to analyze swap use.
2685  * Since technically non-swap (default) objects participate in the
2686  * shadow chains, and are converted to swap type as needed by swap
2687  * pager, we must report them.
2688  */
2689 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2690     CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2691     sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2692     "List of swap VM objects");
2693 
2694 #include "opt_ddb.h"
2695 #ifdef DDB
2696 #include <sys/kernel.h>
2697 
2698 #include <sys/cons.h>
2699 
2700 #include <ddb/ddb.h>
2701 
2702 static int
2703 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2704 {
2705 	vm_map_t tmpm;
2706 	vm_map_entry_t tmpe;
2707 	vm_object_t obj;
2708 
2709 	if (map == 0)
2710 		return 0;
2711 
2712 	if (entry == 0) {
2713 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2714 			if (_vm_object_in_map(map, object, tmpe)) {
2715 				return 1;
2716 			}
2717 		}
2718 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2719 		tmpm = entry->object.sub_map;
2720 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2721 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2722 				return 1;
2723 			}
2724 		}
2725 	} else if ((obj = entry->object.vm_object) != NULL) {
2726 		for (; obj; obj = obj->backing_object)
2727 			if (obj == object) {
2728 				return 1;
2729 			}
2730 	}
2731 	return 0;
2732 }
2733 
2734 static int
2735 vm_object_in_map(vm_object_t object)
2736 {
2737 	struct proc *p;
2738 
2739 	/* sx_slock(&allproc_lock); */
2740 	FOREACH_PROC_IN_SYSTEM(p) {
2741 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2742 			continue;
2743 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2744 			/* sx_sunlock(&allproc_lock); */
2745 			return 1;
2746 		}
2747 	}
2748 	/* sx_sunlock(&allproc_lock); */
2749 	if (_vm_object_in_map(kernel_map, object, 0))
2750 		return 1;
2751 	return 0;
2752 }
2753 
2754 DB_SHOW_COMMAND(vmochk, vm_object_check)
2755 {
2756 	vm_object_t object;
2757 
2758 	/*
2759 	 * make sure that internal objs are in a map somewhere
2760 	 * and none have zero ref counts.
2761 	 */
2762 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2763 		if ((object->flags & OBJ_ANON) != 0) {
2764 			if (object->ref_count == 0) {
2765 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2766 					(long)object->size);
2767 			}
2768 			if (!vm_object_in_map(object)) {
2769 				db_printf(
2770 			"vmochk: internal obj is not in a map: "
2771 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2772 				    object->ref_count, (u_long)object->size,
2773 				    (u_long)object->size,
2774 				    (void *)object->backing_object);
2775 			}
2776 		}
2777 		if (db_pager_quit)
2778 			return;
2779 	}
2780 }
2781 
2782 /*
2783  *	vm_object_print:	[ debug ]
2784  */
2785 DB_SHOW_COMMAND(object, vm_object_print_static)
2786 {
2787 	/* XXX convert args. */
2788 	vm_object_t object = (vm_object_t)addr;
2789 	boolean_t full = have_addr;
2790 
2791 	vm_page_t p;
2792 
2793 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2794 #define	count	was_count
2795 
2796 	int count;
2797 
2798 	if (object == NULL)
2799 		return;
2800 
2801 	db_iprintf(
2802 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2803 	    object, (int)object->type, (uintmax_t)object->size,
2804 	    object->resident_page_count, object->ref_count, object->flags,
2805 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2806 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2807 	    atomic_load_int(&object->shadow_count),
2808 	    object->backing_object ? object->backing_object->ref_count : 0,
2809 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2810 
2811 	if (!full)
2812 		return;
2813 
2814 	db_indent += 2;
2815 	count = 0;
2816 	TAILQ_FOREACH(p, &object->memq, listq) {
2817 		if (count == 0)
2818 			db_iprintf("memory:=");
2819 		else if (count == 6) {
2820 			db_printf("\n");
2821 			db_iprintf(" ...");
2822 			count = 0;
2823 		} else
2824 			db_printf(",");
2825 		count++;
2826 
2827 		db_printf("(off=0x%jx,page=0x%jx)",
2828 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2829 
2830 		if (db_pager_quit)
2831 			break;
2832 	}
2833 	if (count != 0)
2834 		db_printf("\n");
2835 	db_indent -= 2;
2836 }
2837 
2838 /* XXX. */
2839 #undef count
2840 
2841 /* XXX need this non-static entry for calling from vm_map_print. */
2842 void
2843 vm_object_print(
2844         /* db_expr_t */ long addr,
2845 	boolean_t have_addr,
2846 	/* db_expr_t */ long count,
2847 	char *modif)
2848 {
2849 	vm_object_print_static(addr, have_addr, count, modif);
2850 }
2851 
2852 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2853 {
2854 	vm_object_t object;
2855 	vm_pindex_t fidx;
2856 	vm_paddr_t pa;
2857 	vm_page_t m, prev_m;
2858 	int rcount;
2859 
2860 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2861 		db_printf("new object: %p\n", (void *)object);
2862 		if (db_pager_quit)
2863 			return;
2864 
2865 		rcount = 0;
2866 		fidx = 0;
2867 		pa = -1;
2868 		TAILQ_FOREACH(m, &object->memq, listq) {
2869 			if (m->pindex > 128)
2870 				break;
2871 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2872 			    prev_m->pindex + 1 != m->pindex) {
2873 				if (rcount) {
2874 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2875 						(long)fidx, rcount, (long)pa);
2876 					if (db_pager_quit)
2877 						return;
2878 					rcount = 0;
2879 				}
2880 			}
2881 			if (rcount &&
2882 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2883 				++rcount;
2884 				continue;
2885 			}
2886 			if (rcount) {
2887 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2888 					(long)fidx, rcount, (long)pa);
2889 				if (db_pager_quit)
2890 					return;
2891 			}
2892 			fidx = m->pindex;
2893 			pa = VM_PAGE_TO_PHYS(m);
2894 			rcount = 1;
2895 		}
2896 		if (rcount) {
2897 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2898 				(long)fidx, rcount, (long)pa);
2899 			if (db_pager_quit)
2900 				return;
2901 		}
2902 	}
2903 }
2904 #endif /* DDB */
2905