xref: /freebsd/sys/vm/vm_object.c (revision 7899f917b1c0ea178f1d2be0cfb452086d079d23)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39  *
40  * Permission to use, copy, modify and distribute this software and
41  * its documentation is hereby granted, provided that both the copyright
42  * notice and this permission notice appear in all copies of the
43  * software, derivative works or modified versions, and any portions
44  * thereof, and that both notices appear in supporting documentation.
45  *
46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49  *
50  * Carnegie Mellon requests users of this software to return to
51  *
52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
53  *  School of Computer Science
54  *  Carnegie Mellon University
55  *  Pittsburgh PA 15213-3890
56  *
57  * any improvements or extensions that they make and grant Carnegie the
58  * rights to redistribute these changes.
59  */
60 
61 /*
62  *	Virtual memory object module.
63  */
64 
65 #include "opt_vm.h"
66 
67 #include <sys/systm.h>
68 #include <sys/blockcount.h>
69 #include <sys/cpuset.h>
70 #include <sys/jail.h>
71 #include <sys/limits.h>
72 #include <sys/lock.h>
73 #include <sys/mman.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/mutex.h>
77 #include <sys/pctrie.h>
78 #include <sys/proc.h>
79 #include <sys/refcount.h>
80 #include <sys/sx.h>
81 #include <sys/sysctl.h>
82 #include <sys/resourcevar.h>
83 #include <sys/refcount.h>
84 #include <sys/rwlock.h>
85 #include <sys/user.h>
86 #include <sys/vnode.h>
87 #include <sys/vmmeter.h>
88 
89 #include <vm/vm.h>
90 #include <vm/vm_param.h>
91 #include <vm/pmap.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_page.h>
95 #include <vm/vm_pageout.h>
96 #include <vm/vm_pager.h>
97 #include <vm/vm_phys.h>
98 #include <vm/vm_pagequeue.h>
99 #include <vm/swap_pager.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_extern.h>
102 #include <vm/vm_radix.h>
103 #include <vm/vm_reserv.h>
104 #include <vm/uma.h>
105 
106 static int old_msync;
107 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
108     "Use old (insecure) msync behavior");
109 
110 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
111 		    int pagerflags, int flags, boolean_t *allclean,
112 		    boolean_t *eio);
113 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
114 		    boolean_t *allclean);
115 static void	vm_object_backing_remove(vm_object_t object);
116 
117 /*
118  *	Virtual memory objects maintain the actual data
119  *	associated with allocated virtual memory.  A given
120  *	page of memory exists within exactly one object.
121  *
122  *	An object is only deallocated when all "references"
123  *	are given up.  Only one "reference" to a given
124  *	region of an object should be writeable.
125  *
126  *	Associated with each object is a list of all resident
127  *	memory pages belonging to that object; this list is
128  *	maintained by the "vm_page" module, and locked by the object's
129  *	lock.
130  *
131  *	Each object also records a "pager" routine which is
132  *	used to retrieve (and store) pages to the proper backing
133  *	storage.  In addition, objects may be backed by other
134  *	objects from which they were virtual-copied.
135  *
136  *	The only items within the object structure which are
137  *	modified after time of creation are:
138  *		reference count		locked by object's lock
139  *		pager routine		locked by object's lock
140  *
141  */
142 
143 struct object_q vm_object_list;
144 struct mtx vm_object_list_mtx;	/* lock for object list and count */
145 
146 struct vm_object kernel_object_store;
147 
148 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
149     "VM object stats");
150 
151 static COUNTER_U64_DEFINE_EARLY(object_collapses);
152 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
153     &object_collapses,
154     "VM object collapses");
155 
156 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
157 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
158     &object_bypasses,
159     "VM object bypasses");
160 
161 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
162 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
163     &object_collapse_waits,
164     "Number of sleeps for collapse");
165 
166 static uma_zone_t obj_zone;
167 
168 static int vm_object_zinit(void *mem, int size, int flags);
169 
170 #ifdef INVARIANTS
171 static void vm_object_zdtor(void *mem, int size, void *arg);
172 
173 static void
174 vm_object_zdtor(void *mem, int size, void *arg)
175 {
176 	vm_object_t object;
177 
178 	object = (vm_object_t)mem;
179 	KASSERT(object->ref_count == 0,
180 	    ("object %p ref_count = %d", object, object->ref_count));
181 	KASSERT(TAILQ_EMPTY(&object->memq),
182 	    ("object %p has resident pages in its memq", object));
183 	KASSERT(vm_radix_is_empty(&object->rtree),
184 	    ("object %p has resident pages in its trie", object));
185 #if VM_NRESERVLEVEL > 0
186 	KASSERT(LIST_EMPTY(&object->rvq),
187 	    ("object %p has reservations",
188 	    object));
189 #endif
190 	KASSERT(!vm_object_busied(object),
191 	    ("object %p busy = %d", object, blockcount_read(&object->busy)));
192 	KASSERT(object->resident_page_count == 0,
193 	    ("object %p resident_page_count = %d",
194 	    object, object->resident_page_count));
195 	KASSERT(atomic_load_int(&object->shadow_count) == 0,
196 	    ("object %p shadow_count = %d",
197 	    object, atomic_load_int(&object->shadow_count)));
198 	KASSERT(object->type == OBJT_DEAD,
199 	    ("object %p has non-dead type %d",
200 	    object, object->type));
201 	KASSERT(object->charge == 0 && object->cred == NULL,
202 	    ("object %p has non-zero charge %ju (%p)",
203 	    object, (uintmax_t)object->charge, object->cred));
204 }
205 #endif
206 
207 static int
208 vm_object_zinit(void *mem, int size, int flags)
209 {
210 	vm_object_t object;
211 
212 	object = (vm_object_t)mem;
213 	rw_init_flags(&object->lock, "vmobject", RW_DUPOK | RW_NEW);
214 
215 	/* These are true for any object that has been freed */
216 	object->type = OBJT_DEAD;
217 	vm_radix_init(&object->rtree);
218 	refcount_init(&object->ref_count, 0);
219 	blockcount_init(&object->paging_in_progress);
220 	blockcount_init(&object->busy);
221 	object->resident_page_count = 0;
222 	atomic_store_int(&object->shadow_count, 0);
223 	object->flags = OBJ_DEAD;
224 
225 	mtx_lock(&vm_object_list_mtx);
226 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
227 	mtx_unlock(&vm_object_list_mtx);
228 	return (0);
229 }
230 
231 static void
232 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
233     vm_object_t object, void *handle)
234 {
235 
236 	TAILQ_INIT(&object->memq);
237 	LIST_INIT(&object->shadow_head);
238 
239 	object->type = type;
240 	object->flags = flags;
241 	if ((flags & OBJ_SWAP) != 0) {
242 		pctrie_init(&object->un_pager.swp.swp_blks);
243 		object->un_pager.swp.writemappings = 0;
244 	}
245 
246 	/*
247 	 * Ensure that swap_pager_swapoff() iteration over object_list
248 	 * sees up to date type and pctrie head if it observed
249 	 * non-dead object.
250 	 */
251 	atomic_thread_fence_rel();
252 
253 	object->pg_color = 0;
254 	object->size = size;
255 	object->domain.dr_policy = NULL;
256 	object->generation = 1;
257 	object->cleangeneration = 1;
258 	refcount_init(&object->ref_count, 1);
259 	object->memattr = VM_MEMATTR_DEFAULT;
260 	object->cred = NULL;
261 	object->charge = 0;
262 	object->handle = handle;
263 	object->backing_object = NULL;
264 	object->backing_object_offset = (vm_ooffset_t) 0;
265 #if VM_NRESERVLEVEL > 0
266 	LIST_INIT(&object->rvq);
267 #endif
268 	umtx_shm_object_init(object);
269 }
270 
271 /*
272  *	vm_object_init:
273  *
274  *	Initialize the VM objects module.
275  */
276 void
277 vm_object_init(void)
278 {
279 	TAILQ_INIT(&vm_object_list);
280 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
281 
282 	rw_init(&kernel_object->lock, "kernel vm object");
283 	vm_radix_init(&kernel_object->rtree);
284 	_vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
285 	    VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
286 #if VM_NRESERVLEVEL > 0
287 	kernel_object->flags |= OBJ_COLORED;
288 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
289 #endif
290 	kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
291 
292 	/*
293 	 * The lock portion of struct vm_object must be type stable due
294 	 * to vm_pageout_fallback_object_lock locking a vm object
295 	 * without holding any references to it.
296 	 *
297 	 * paging_in_progress is valid always.  Lockless references to
298 	 * the objects may acquire pip and then check OBJ_DEAD.
299 	 */
300 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
301 #ifdef INVARIANTS
302 	    vm_object_zdtor,
303 #else
304 	    NULL,
305 #endif
306 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
307 
308 	vm_radix_zinit();
309 }
310 
311 void
312 vm_object_clear_flag(vm_object_t object, u_short bits)
313 {
314 
315 	VM_OBJECT_ASSERT_WLOCKED(object);
316 	object->flags &= ~bits;
317 }
318 
319 /*
320  *	Sets the default memory attribute for the specified object.  Pages
321  *	that are allocated to this object are by default assigned this memory
322  *	attribute.
323  *
324  *	Presently, this function must be called before any pages are allocated
325  *	to the object.  In the future, this requirement may be relaxed for
326  *	"default" and "swap" objects.
327  */
328 int
329 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
330 {
331 
332 	VM_OBJECT_ASSERT_WLOCKED(object);
333 
334 	if (object->type == OBJT_DEAD)
335 		return (KERN_INVALID_ARGUMENT);
336 	if (!TAILQ_EMPTY(&object->memq))
337 		return (KERN_FAILURE);
338 
339 	object->memattr = memattr;
340 	return (KERN_SUCCESS);
341 }
342 
343 void
344 vm_object_pip_add(vm_object_t object, short i)
345 {
346 
347 	if (i > 0)
348 		blockcount_acquire(&object->paging_in_progress, i);
349 }
350 
351 void
352 vm_object_pip_wakeup(vm_object_t object)
353 {
354 
355 	vm_object_pip_wakeupn(object, 1);
356 }
357 
358 void
359 vm_object_pip_wakeupn(vm_object_t object, short i)
360 {
361 
362 	if (i > 0)
363 		blockcount_release(&object->paging_in_progress, i);
364 }
365 
366 /*
367  * Atomically drop the object lock and wait for pip to drain.  This protects
368  * from sleep/wakeup races due to identity changes.  The lock is not re-acquired
369  * on return.
370  */
371 static void
372 vm_object_pip_sleep(vm_object_t object, const char *waitid)
373 {
374 
375 	(void)blockcount_sleep(&object->paging_in_progress, &object->lock,
376 	    waitid, PVM | PDROP);
377 }
378 
379 void
380 vm_object_pip_wait(vm_object_t object, const char *waitid)
381 {
382 
383 	VM_OBJECT_ASSERT_WLOCKED(object);
384 
385 	blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
386 	    PVM);
387 }
388 
389 void
390 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
391 {
392 
393 	VM_OBJECT_ASSERT_UNLOCKED(object);
394 
395 	blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
396 }
397 
398 /*
399  *	vm_object_allocate:
400  *
401  *	Returns a new object with the given size.
402  */
403 vm_object_t
404 vm_object_allocate(objtype_t type, vm_pindex_t size)
405 {
406 	vm_object_t object;
407 	u_short flags;
408 
409 	switch (type) {
410 	case OBJT_DEAD:
411 		panic("vm_object_allocate: can't create OBJT_DEAD");
412 	case OBJT_SWAP:
413 		flags = OBJ_COLORED | OBJ_SWAP;
414 		break;
415 	case OBJT_DEVICE:
416 	case OBJT_SG:
417 		flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
418 		break;
419 	case OBJT_MGTDEVICE:
420 		flags = OBJ_FICTITIOUS;
421 		break;
422 	case OBJT_PHYS:
423 		flags = OBJ_UNMANAGED;
424 		break;
425 	case OBJT_VNODE:
426 		flags = 0;
427 		break;
428 	default:
429 		panic("vm_object_allocate: type %d is undefined or dynamic",
430 		    type);
431 	}
432 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
433 	_vm_object_allocate(type, size, flags, object, NULL);
434 
435 	return (object);
436 }
437 
438 vm_object_t
439 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags)
440 {
441 	vm_object_t object;
442 
443 	MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */);
444 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
445 	_vm_object_allocate(dyntype, size, flags, object, NULL);
446 
447 	return (object);
448 }
449 
450 /*
451  *	vm_object_allocate_anon:
452  *
453  *	Returns a new default object of the given size and marked as
454  *	anonymous memory for special split/collapse handling.  Color
455  *	to be initialized by the caller.
456  */
457 vm_object_t
458 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
459     struct ucred *cred, vm_size_t charge)
460 {
461 	vm_object_t handle, object;
462 
463 	if (backing_object == NULL)
464 		handle = NULL;
465 	else if ((backing_object->flags & OBJ_ANON) != 0)
466 		handle = backing_object->handle;
467 	else
468 		handle = backing_object;
469 	object = uma_zalloc(obj_zone, M_WAITOK);
470 	_vm_object_allocate(OBJT_SWAP, size,
471 	    OBJ_ANON | OBJ_ONEMAPPING | OBJ_SWAP, object, handle);
472 	object->cred = cred;
473 	object->charge = cred != NULL ? charge : 0;
474 	return (object);
475 }
476 
477 static void
478 vm_object_reference_vnode(vm_object_t object)
479 {
480 	u_int old;
481 
482 	/*
483 	 * vnode objects need the lock for the first reference
484 	 * to serialize with vnode_object_deallocate().
485 	 */
486 	if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
487 		VM_OBJECT_RLOCK(object);
488 		old = refcount_acquire(&object->ref_count);
489 		if (object->type == OBJT_VNODE && old == 0)
490 			vref(object->handle);
491 		VM_OBJECT_RUNLOCK(object);
492 	}
493 }
494 
495 /*
496  *	vm_object_reference:
497  *
498  *	Acquires a reference to the given object.
499  */
500 void
501 vm_object_reference(vm_object_t object)
502 {
503 
504 	if (object == NULL)
505 		return;
506 
507 	if (object->type == OBJT_VNODE)
508 		vm_object_reference_vnode(object);
509 	else
510 		refcount_acquire(&object->ref_count);
511 	KASSERT((object->flags & OBJ_DEAD) == 0,
512 	    ("vm_object_reference: Referenced dead object."));
513 }
514 
515 /*
516  *	vm_object_reference_locked:
517  *
518  *	Gets another reference to the given object.
519  *
520  *	The object must be locked.
521  */
522 void
523 vm_object_reference_locked(vm_object_t object)
524 {
525 	u_int old;
526 
527 	VM_OBJECT_ASSERT_LOCKED(object);
528 	old = refcount_acquire(&object->ref_count);
529 	if (object->type == OBJT_VNODE && old == 0)
530 		vref(object->handle);
531 	KASSERT((object->flags & OBJ_DEAD) == 0,
532 	    ("vm_object_reference: Referenced dead object."));
533 }
534 
535 /*
536  * Handle deallocating an object of type OBJT_VNODE.
537  */
538 static void
539 vm_object_deallocate_vnode(vm_object_t object)
540 {
541 	struct vnode *vp = (struct vnode *) object->handle;
542 	bool last;
543 
544 	KASSERT(object->type == OBJT_VNODE,
545 	    ("vm_object_deallocate_vnode: not a vnode object"));
546 	KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
547 
548 	/* Object lock to protect handle lookup. */
549 	last = refcount_release(&object->ref_count);
550 	VM_OBJECT_RUNLOCK(object);
551 
552 	if (!last)
553 		return;
554 
555 	if (!umtx_shm_vnobj_persistent)
556 		umtx_shm_object_terminated(object);
557 
558 	/* vrele may need the vnode lock. */
559 	vrele(vp);
560 }
561 
562 /*
563  * We dropped a reference on an object and discovered that it had a
564  * single remaining shadow.  This is a sibling of the reference we
565  * dropped.  Attempt to collapse the sibling and backing object.
566  */
567 static vm_object_t
568 vm_object_deallocate_anon(vm_object_t backing_object)
569 {
570 	vm_object_t object;
571 
572 	/* Fetch the final shadow.  */
573 	object = LIST_FIRST(&backing_object->shadow_head);
574 	KASSERT(object != NULL &&
575 	    atomic_load_int(&backing_object->shadow_count) == 1,
576 	    ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
577 	    backing_object->ref_count,
578 	    atomic_load_int(&backing_object->shadow_count)));
579 	KASSERT((object->flags & OBJ_ANON) != 0,
580 	    ("invalid shadow object %p", object));
581 
582 	if (!VM_OBJECT_TRYWLOCK(object)) {
583 		/*
584 		 * Prevent object from disappearing since we do not have a
585 		 * reference.
586 		 */
587 		vm_object_pip_add(object, 1);
588 		VM_OBJECT_WUNLOCK(backing_object);
589 		VM_OBJECT_WLOCK(object);
590 		vm_object_pip_wakeup(object);
591 	} else
592 		VM_OBJECT_WUNLOCK(backing_object);
593 
594 	/*
595 	 * Check for a collapse/terminate race with the last reference holder.
596 	 */
597 	if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
598 	    !refcount_acquire_if_not_zero(&object->ref_count)) {
599 		VM_OBJECT_WUNLOCK(object);
600 		return (NULL);
601 	}
602 	backing_object = object->backing_object;
603 	if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
604 		vm_object_collapse(object);
605 	VM_OBJECT_WUNLOCK(object);
606 
607 	return (object);
608 }
609 
610 /*
611  *	vm_object_deallocate:
612  *
613  *	Release a reference to the specified object,
614  *	gained either through a vm_object_allocate
615  *	or a vm_object_reference call.  When all references
616  *	are gone, storage associated with this object
617  *	may be relinquished.
618  *
619  *	No object may be locked.
620  */
621 void
622 vm_object_deallocate(vm_object_t object)
623 {
624 	vm_object_t temp;
625 	bool released;
626 
627 	while (object != NULL) {
628 		/*
629 		 * If the reference count goes to 0 we start calling
630 		 * vm_object_terminate() on the object chain.  A ref count
631 		 * of 1 may be a special case depending on the shadow count
632 		 * being 0 or 1.  These cases require a write lock on the
633 		 * object.
634 		 */
635 		if ((object->flags & OBJ_ANON) == 0)
636 			released = refcount_release_if_gt(&object->ref_count, 1);
637 		else
638 			released = refcount_release_if_gt(&object->ref_count, 2);
639 		if (released)
640 			return;
641 
642 		if (object->type == OBJT_VNODE) {
643 			VM_OBJECT_RLOCK(object);
644 			if (object->type == OBJT_VNODE) {
645 				vm_object_deallocate_vnode(object);
646 				return;
647 			}
648 			VM_OBJECT_RUNLOCK(object);
649 		}
650 
651 		VM_OBJECT_WLOCK(object);
652 		KASSERT(object->ref_count > 0,
653 		    ("vm_object_deallocate: object deallocated too many times: %d",
654 		    object->type));
655 
656 		/*
657 		 * If this is not the final reference to an anonymous
658 		 * object we may need to collapse the shadow chain.
659 		 */
660 		if (!refcount_release(&object->ref_count)) {
661 			if (object->ref_count > 1 ||
662 			    atomic_load_int(&object->shadow_count) == 0) {
663 				if ((object->flags & OBJ_ANON) != 0 &&
664 				    object->ref_count == 1)
665 					vm_object_set_flag(object,
666 					    OBJ_ONEMAPPING);
667 				VM_OBJECT_WUNLOCK(object);
668 				return;
669 			}
670 
671 			/* Handle collapsing last ref on anonymous objects. */
672 			object = vm_object_deallocate_anon(object);
673 			continue;
674 		}
675 
676 		/*
677 		 * Handle the final reference to an object.  We restart
678 		 * the loop with the backing object to avoid recursion.
679 		 */
680 		umtx_shm_object_terminated(object);
681 		temp = object->backing_object;
682 		if (temp != NULL) {
683 			KASSERT(object->type == OBJT_SWAP,
684 			    ("shadowed tmpfs v_object 2 %p", object));
685 			vm_object_backing_remove(object);
686 		}
687 
688 		KASSERT((object->flags & OBJ_DEAD) == 0,
689 		    ("vm_object_deallocate: Terminating dead object."));
690 		vm_object_set_flag(object, OBJ_DEAD);
691 		vm_object_terminate(object);
692 		object = temp;
693 	}
694 }
695 
696 void
697 vm_object_destroy(vm_object_t object)
698 {
699 	uma_zfree(obj_zone, object);
700 }
701 
702 static void
703 vm_object_sub_shadow(vm_object_t object)
704 {
705 	KASSERT(object->shadow_count >= 1,
706 	    ("object %p sub_shadow count zero", object));
707 	atomic_subtract_int(&object->shadow_count, 1);
708 }
709 
710 static void
711 vm_object_backing_remove_locked(vm_object_t object)
712 {
713 	vm_object_t backing_object;
714 
715 	backing_object = object->backing_object;
716 	VM_OBJECT_ASSERT_WLOCKED(object);
717 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
718 
719 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
720 	    ("vm_object_backing_remove: Removing collapsing object."));
721 
722 	vm_object_sub_shadow(backing_object);
723 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
724 		LIST_REMOVE(object, shadow_list);
725 		vm_object_clear_flag(object, OBJ_SHADOWLIST);
726 	}
727 	object->backing_object = NULL;
728 }
729 
730 static void
731 vm_object_backing_remove(vm_object_t object)
732 {
733 	vm_object_t backing_object;
734 
735 	VM_OBJECT_ASSERT_WLOCKED(object);
736 
737 	backing_object = object->backing_object;
738 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
739 		VM_OBJECT_WLOCK(backing_object);
740 		vm_object_backing_remove_locked(object);
741 		VM_OBJECT_WUNLOCK(backing_object);
742 	} else {
743 		object->backing_object = NULL;
744 		vm_object_sub_shadow(backing_object);
745 	}
746 }
747 
748 static void
749 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
750 {
751 
752 	VM_OBJECT_ASSERT_WLOCKED(object);
753 
754 	atomic_add_int(&backing_object->shadow_count, 1);
755 	if ((backing_object->flags & OBJ_ANON) != 0) {
756 		VM_OBJECT_ASSERT_WLOCKED(backing_object);
757 		LIST_INSERT_HEAD(&backing_object->shadow_head, object,
758 		    shadow_list);
759 		vm_object_set_flag(object, OBJ_SHADOWLIST);
760 	}
761 	object->backing_object = backing_object;
762 }
763 
764 static void
765 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
766 {
767 
768 	VM_OBJECT_ASSERT_WLOCKED(object);
769 
770 	if ((backing_object->flags & OBJ_ANON) != 0) {
771 		VM_OBJECT_WLOCK(backing_object);
772 		vm_object_backing_insert_locked(object, backing_object);
773 		VM_OBJECT_WUNLOCK(backing_object);
774 	} else {
775 		object->backing_object = backing_object;
776 		atomic_add_int(&backing_object->shadow_count, 1);
777 	}
778 }
779 
780 /*
781  * Insert an object into a backing_object's shadow list with an additional
782  * reference to the backing_object added.
783  */
784 static void
785 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
786 {
787 
788 	VM_OBJECT_ASSERT_WLOCKED(object);
789 
790 	if ((backing_object->flags & OBJ_ANON) != 0) {
791 		VM_OBJECT_WLOCK(backing_object);
792 		KASSERT((backing_object->flags & OBJ_DEAD) == 0,
793 		    ("shadowing dead anonymous object"));
794 		vm_object_reference_locked(backing_object);
795 		vm_object_backing_insert_locked(object, backing_object);
796 		vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
797 		VM_OBJECT_WUNLOCK(backing_object);
798 	} else {
799 		vm_object_reference(backing_object);
800 		atomic_add_int(&backing_object->shadow_count, 1);
801 		object->backing_object = backing_object;
802 	}
803 }
804 
805 /*
806  * Transfer a backing reference from backing_object to object.
807  */
808 static void
809 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
810 {
811 	vm_object_t new_backing_object;
812 
813 	/*
814 	 * Note that the reference to backing_object->backing_object
815 	 * moves from within backing_object to within object.
816 	 */
817 	vm_object_backing_remove_locked(object);
818 	new_backing_object = backing_object->backing_object;
819 	if (new_backing_object == NULL)
820 		return;
821 	if ((new_backing_object->flags & OBJ_ANON) != 0) {
822 		VM_OBJECT_WLOCK(new_backing_object);
823 		vm_object_backing_remove_locked(backing_object);
824 		vm_object_backing_insert_locked(object, new_backing_object);
825 		VM_OBJECT_WUNLOCK(new_backing_object);
826 	} else {
827 		/*
828 		 * shadow_count for new_backing_object is left
829 		 * unchanged, its reference provided by backing_object
830 		 * is replaced by object.
831 		 */
832 		object->backing_object = new_backing_object;
833 		backing_object->backing_object = NULL;
834 	}
835 }
836 
837 /*
838  * Wait for a concurrent collapse to settle.
839  */
840 static void
841 vm_object_collapse_wait(vm_object_t object)
842 {
843 
844 	VM_OBJECT_ASSERT_WLOCKED(object);
845 
846 	while ((object->flags & OBJ_COLLAPSING) != 0) {
847 		vm_object_pip_wait(object, "vmcolwait");
848 		counter_u64_add(object_collapse_waits, 1);
849 	}
850 }
851 
852 /*
853  * Waits for a backing object to clear a pending collapse and returns
854  * it locked if it is an ANON object.
855  */
856 static vm_object_t
857 vm_object_backing_collapse_wait(vm_object_t object)
858 {
859 	vm_object_t backing_object;
860 
861 	VM_OBJECT_ASSERT_WLOCKED(object);
862 
863 	for (;;) {
864 		backing_object = object->backing_object;
865 		if (backing_object == NULL ||
866 		    (backing_object->flags & OBJ_ANON) == 0)
867 			return (NULL);
868 		VM_OBJECT_WLOCK(backing_object);
869 		if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
870 			break;
871 		VM_OBJECT_WUNLOCK(object);
872 		vm_object_pip_sleep(backing_object, "vmbckwait");
873 		counter_u64_add(object_collapse_waits, 1);
874 		VM_OBJECT_WLOCK(object);
875 	}
876 	return (backing_object);
877 }
878 
879 /*
880  *	vm_object_terminate_pages removes any remaining pageable pages
881  *	from the object and resets the object to an empty state.
882  */
883 static void
884 vm_object_terminate_pages(vm_object_t object)
885 {
886 	vm_page_t p, p_next;
887 
888 	VM_OBJECT_ASSERT_WLOCKED(object);
889 
890 	/*
891 	 * Free any remaining pageable pages.  This also removes them from the
892 	 * paging queues.  However, don't free wired pages, just remove them
893 	 * from the object.  Rather than incrementally removing each page from
894 	 * the object, the page and object are reset to any empty state.
895 	 */
896 	TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
897 		vm_page_assert_unbusied(p);
898 		KASSERT(p->object == object &&
899 		    (p->ref_count & VPRC_OBJREF) != 0,
900 		    ("vm_object_terminate_pages: page %p is inconsistent", p));
901 
902 		p->object = NULL;
903 		if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
904 			VM_CNT_INC(v_pfree);
905 			vm_page_free(p);
906 		}
907 	}
908 
909 	/*
910 	 * If the object contained any pages, then reset it to an empty state.
911 	 * None of the object's fields, including "resident_page_count", were
912 	 * modified by the preceding loop.
913 	 */
914 	if (object->resident_page_count != 0) {
915 		vm_radix_reclaim_allnodes(&object->rtree);
916 		TAILQ_INIT(&object->memq);
917 		object->resident_page_count = 0;
918 		if (object->type == OBJT_VNODE)
919 			vdrop(object->handle);
920 	}
921 }
922 
923 /*
924  *	vm_object_terminate actually destroys the specified object, freeing
925  *	up all previously used resources.
926  *
927  *	The object must be locked.
928  *	This routine may block.
929  */
930 void
931 vm_object_terminate(vm_object_t object)
932 {
933 
934 	VM_OBJECT_ASSERT_WLOCKED(object);
935 	KASSERT((object->flags & OBJ_DEAD) != 0,
936 	    ("terminating non-dead obj %p", object));
937 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
938 	    ("terminating collapsing obj %p", object));
939 	KASSERT(object->backing_object == NULL,
940 	    ("terminating shadow obj %p", object));
941 
942 	/*
943 	 * Wait for the pageout daemon and other current users to be
944 	 * done with the object.  Note that new paging_in_progress
945 	 * users can come after this wait, but they must check
946 	 * OBJ_DEAD flag set (without unlocking the object), and avoid
947 	 * the object being terminated.
948 	 */
949 	vm_object_pip_wait(object, "objtrm");
950 
951 	KASSERT(object->ref_count == 0,
952 	    ("vm_object_terminate: object with references, ref_count=%d",
953 	    object->ref_count));
954 
955 	if ((object->flags & OBJ_PG_DTOR) == 0)
956 		vm_object_terminate_pages(object);
957 
958 #if VM_NRESERVLEVEL > 0
959 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
960 		vm_reserv_break_all(object);
961 #endif
962 
963 	KASSERT(object->cred == NULL || (object->flags & OBJ_SWAP) != 0,
964 	    ("%s: non-swap obj %p has cred", __func__, object));
965 
966 	/*
967 	 * Let the pager know object is dead.
968 	 */
969 	vm_pager_deallocate(object);
970 	VM_OBJECT_WUNLOCK(object);
971 
972 	vm_object_destroy(object);
973 }
974 
975 /*
976  * Make the page read-only so that we can clear the object flags.  However, if
977  * this is a nosync mmap then the object is likely to stay dirty so do not
978  * mess with the page and do not clear the object flags.  Returns TRUE if the
979  * page should be flushed, and FALSE otherwise.
980  */
981 static boolean_t
982 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
983 {
984 
985 	vm_page_assert_busied(p);
986 
987 	/*
988 	 * If we have been asked to skip nosync pages and this is a
989 	 * nosync page, skip it.  Note that the object flags were not
990 	 * cleared in this case so we do not have to set them.
991 	 */
992 	if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
993 		*allclean = FALSE;
994 		return (FALSE);
995 	} else {
996 		pmap_remove_write(p);
997 		return (p->dirty != 0);
998 	}
999 }
1000 
1001 /*
1002  *	vm_object_page_clean
1003  *
1004  *	Clean all dirty pages in the specified range of object.  Leaves page
1005  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
1006  *	write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1007  *	leaving the object dirty.
1008  *
1009  *	For swap objects backing tmpfs regular files, do not flush anything,
1010  *	but remove write protection on the mapped pages to update mtime through
1011  *	mmaped writes.
1012  *
1013  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
1014  *	synchronous clustering mode implementation.
1015  *
1016  *	Odd semantics: if start == end, we clean everything.
1017  *
1018  *	The object must be locked.
1019  *
1020  *	Returns FALSE if some page from the range was not written, as
1021  *	reported by the pager, and TRUE otherwise.
1022  */
1023 boolean_t
1024 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1025     int flags)
1026 {
1027 	vm_page_t np, p;
1028 	vm_pindex_t pi, tend, tstart;
1029 	int curgeneration, n, pagerflags;
1030 	boolean_t eio, res, allclean;
1031 
1032 	VM_OBJECT_ASSERT_WLOCKED(object);
1033 
1034 	if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1035 		return (TRUE);
1036 
1037 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1038 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1039 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1040 
1041 	tstart = OFF_TO_IDX(start);
1042 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1043 	allclean = tstart == 0 && tend >= object->size;
1044 	res = TRUE;
1045 
1046 rescan:
1047 	curgeneration = object->generation;
1048 
1049 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1050 		pi = p->pindex;
1051 		if (pi >= tend)
1052 			break;
1053 		np = TAILQ_NEXT(p, listq);
1054 		if (vm_page_none_valid(p))
1055 			continue;
1056 		if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1057 			if (object->generation != curgeneration &&
1058 			    (flags & OBJPC_SYNC) != 0)
1059 				goto rescan;
1060 			np = vm_page_find_least(object, pi);
1061 			continue;
1062 		}
1063 		if (!vm_object_page_remove_write(p, flags, &allclean)) {
1064 			vm_page_xunbusy(p);
1065 			continue;
1066 		}
1067 		if (object->type == OBJT_VNODE) {
1068 			n = vm_object_page_collect_flush(object, p, pagerflags,
1069 			    flags, &allclean, &eio);
1070 			if (eio) {
1071 				res = FALSE;
1072 				allclean = FALSE;
1073 			}
1074 			if (object->generation != curgeneration &&
1075 			    (flags & OBJPC_SYNC) != 0)
1076 				goto rescan;
1077 
1078 			/*
1079 			 * If the VOP_PUTPAGES() did a truncated write, so
1080 			 * that even the first page of the run is not fully
1081 			 * written, vm_pageout_flush() returns 0 as the run
1082 			 * length.  Since the condition that caused truncated
1083 			 * write may be permanent, e.g. exhausted free space,
1084 			 * accepting n == 0 would cause an infinite loop.
1085 			 *
1086 			 * Forwarding the iterator leaves the unwritten page
1087 			 * behind, but there is not much we can do there if
1088 			 * filesystem refuses to write it.
1089 			 */
1090 			if (n == 0) {
1091 				n = 1;
1092 				allclean = FALSE;
1093 			}
1094 		} else {
1095 			n = 1;
1096 			vm_page_xunbusy(p);
1097 		}
1098 		np = vm_page_find_least(object, pi + n);
1099 	}
1100 #if 0
1101 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1102 #endif
1103 
1104 	/*
1105 	 * Leave updating cleangeneration for tmpfs objects to tmpfs
1106 	 * scan.  It needs to update mtime, which happens for other
1107 	 * filesystems during page writeouts.
1108 	 */
1109 	if (allclean && object->type == OBJT_VNODE)
1110 		object->cleangeneration = curgeneration;
1111 	return (res);
1112 }
1113 
1114 static int
1115 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1116     int flags, boolean_t *allclean, boolean_t *eio)
1117 {
1118 	vm_page_t ma[2 * vm_pageout_page_count - 1], tp;
1119 	int base, count, runlen;
1120 
1121 	vm_page_lock_assert(p, MA_NOTOWNED);
1122 	vm_page_assert_xbusied(p);
1123 	VM_OBJECT_ASSERT_WLOCKED(object);
1124 	base = nitems(ma) / 2;
1125 	ma[base] = p;
1126 	for (count = 1, tp = p; count < vm_pageout_page_count; count++) {
1127 		tp = vm_page_next(tp);
1128 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1129 			break;
1130 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1131 			vm_page_xunbusy(tp);
1132 			break;
1133 		}
1134 		ma[base + count] = tp;
1135 	}
1136 
1137 	for (tp = p; count < vm_pageout_page_count; count++) {
1138 		tp = vm_page_prev(tp);
1139 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1140 			break;
1141 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1142 			vm_page_xunbusy(tp);
1143 			break;
1144 		}
1145 		ma[--base] = tp;
1146 	}
1147 
1148 	vm_pageout_flush(&ma[base], count, pagerflags, nitems(ma) / 2 - base,
1149 	    &runlen, eio);
1150 	return (runlen);
1151 }
1152 
1153 /*
1154  * Note that there is absolutely no sense in writing out
1155  * anonymous objects, so we track down the vnode object
1156  * to write out.
1157  * We invalidate (remove) all pages from the address space
1158  * for semantic correctness.
1159  *
1160  * If the backing object is a device object with unmanaged pages, then any
1161  * mappings to the specified range of pages must be removed before this
1162  * function is called.
1163  *
1164  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1165  * may start out with a NULL object.
1166  */
1167 boolean_t
1168 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1169     boolean_t syncio, boolean_t invalidate)
1170 {
1171 	vm_object_t backing_object;
1172 	struct vnode *vp;
1173 	struct mount *mp;
1174 	int error, flags, fsync_after;
1175 	boolean_t res;
1176 
1177 	if (object == NULL)
1178 		return (TRUE);
1179 	res = TRUE;
1180 	error = 0;
1181 	VM_OBJECT_WLOCK(object);
1182 	while ((backing_object = object->backing_object) != NULL) {
1183 		VM_OBJECT_WLOCK(backing_object);
1184 		offset += object->backing_object_offset;
1185 		VM_OBJECT_WUNLOCK(object);
1186 		object = backing_object;
1187 		if (object->size < OFF_TO_IDX(offset + size))
1188 			size = IDX_TO_OFF(object->size) - offset;
1189 	}
1190 	/*
1191 	 * Flush pages if writing is allowed, invalidate them
1192 	 * if invalidation requested.  Pages undergoing I/O
1193 	 * will be ignored by vm_object_page_remove().
1194 	 *
1195 	 * We cannot lock the vnode and then wait for paging
1196 	 * to complete without deadlocking against vm_fault.
1197 	 * Instead we simply call vm_object_page_remove() and
1198 	 * allow it to block internally on a page-by-page
1199 	 * basis when it encounters pages undergoing async
1200 	 * I/O.
1201 	 */
1202 	if (object->type == OBJT_VNODE &&
1203 	    vm_object_mightbedirty(object) != 0 &&
1204 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1205 		VM_OBJECT_WUNLOCK(object);
1206 		(void)vn_start_write(vp, &mp, V_WAIT);
1207 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1208 		if (syncio && !invalidate && offset == 0 &&
1209 		    atop(size) == object->size) {
1210 			/*
1211 			 * If syncing the whole mapping of the file,
1212 			 * it is faster to schedule all the writes in
1213 			 * async mode, also allowing the clustering,
1214 			 * and then wait for i/o to complete.
1215 			 */
1216 			flags = 0;
1217 			fsync_after = TRUE;
1218 		} else {
1219 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1220 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1221 			fsync_after = FALSE;
1222 		}
1223 		VM_OBJECT_WLOCK(object);
1224 		res = vm_object_page_clean(object, offset, offset + size,
1225 		    flags);
1226 		VM_OBJECT_WUNLOCK(object);
1227 		if (fsync_after) {
1228 			for (;;) {
1229 				error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1230 				if (error != ERELOOKUP)
1231 					break;
1232 
1233 				/*
1234 				 * Allow SU/bufdaemon to handle more
1235 				 * dependencies in the meantime.
1236 				 */
1237 				VOP_UNLOCK(vp);
1238 				vn_finished_write(mp);
1239 
1240 				(void)vn_start_write(vp, &mp, V_WAIT);
1241 				vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1242 			}
1243 		}
1244 		VOP_UNLOCK(vp);
1245 		vn_finished_write(mp);
1246 		if (error != 0)
1247 			res = FALSE;
1248 		VM_OBJECT_WLOCK(object);
1249 	}
1250 	if ((object->type == OBJT_VNODE ||
1251 	     object->type == OBJT_DEVICE) && invalidate) {
1252 		if (object->type == OBJT_DEVICE)
1253 			/*
1254 			 * The option OBJPR_NOTMAPPED must be passed here
1255 			 * because vm_object_page_remove() cannot remove
1256 			 * unmanaged mappings.
1257 			 */
1258 			flags = OBJPR_NOTMAPPED;
1259 		else if (old_msync)
1260 			flags = 0;
1261 		else
1262 			flags = OBJPR_CLEANONLY;
1263 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1264 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1265 	}
1266 	VM_OBJECT_WUNLOCK(object);
1267 	return (res);
1268 }
1269 
1270 /*
1271  * Determine whether the given advice can be applied to the object.  Advice is
1272  * not applied to unmanaged pages since they never belong to page queues, and
1273  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1274  * have been mapped at most once.
1275  */
1276 static bool
1277 vm_object_advice_applies(vm_object_t object, int advice)
1278 {
1279 
1280 	if ((object->flags & OBJ_UNMANAGED) != 0)
1281 		return (false);
1282 	if (advice != MADV_FREE)
1283 		return (true);
1284 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1285 	    (OBJ_ONEMAPPING | OBJ_ANON));
1286 }
1287 
1288 static void
1289 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1290     vm_size_t size)
1291 {
1292 
1293 	if (advice == MADV_FREE)
1294 		vm_pager_freespace(object, pindex, size);
1295 }
1296 
1297 /*
1298  *	vm_object_madvise:
1299  *
1300  *	Implements the madvise function at the object/page level.
1301  *
1302  *	MADV_WILLNEED	(any object)
1303  *
1304  *	    Activate the specified pages if they are resident.
1305  *
1306  *	MADV_DONTNEED	(any object)
1307  *
1308  *	    Deactivate the specified pages if they are resident.
1309  *
1310  *	MADV_FREE	(OBJT_SWAP objects, OBJ_ONEMAPPING only)
1311  *
1312  *	    Deactivate and clean the specified pages if they are
1313  *	    resident.  This permits the process to reuse the pages
1314  *	    without faulting or the kernel to reclaim the pages
1315  *	    without I/O.
1316  */
1317 void
1318 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1319     int advice)
1320 {
1321 	vm_pindex_t tpindex;
1322 	vm_object_t backing_object, tobject;
1323 	vm_page_t m, tm;
1324 
1325 	if (object == NULL)
1326 		return;
1327 
1328 relookup:
1329 	VM_OBJECT_WLOCK(object);
1330 	if (!vm_object_advice_applies(object, advice)) {
1331 		VM_OBJECT_WUNLOCK(object);
1332 		return;
1333 	}
1334 	for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1335 		tobject = object;
1336 
1337 		/*
1338 		 * If the next page isn't resident in the top-level object, we
1339 		 * need to search the shadow chain.  When applying MADV_FREE, we
1340 		 * take care to release any swap space used to store
1341 		 * non-resident pages.
1342 		 */
1343 		if (m == NULL || pindex < m->pindex) {
1344 			/*
1345 			 * Optimize a common case: if the top-level object has
1346 			 * no backing object, we can skip over the non-resident
1347 			 * range in constant time.
1348 			 */
1349 			if (object->backing_object == NULL) {
1350 				tpindex = (m != NULL && m->pindex < end) ?
1351 				    m->pindex : end;
1352 				vm_object_madvise_freespace(object, advice,
1353 				    pindex, tpindex - pindex);
1354 				if ((pindex = tpindex) == end)
1355 					break;
1356 				goto next_page;
1357 			}
1358 
1359 			tpindex = pindex;
1360 			do {
1361 				vm_object_madvise_freespace(tobject, advice,
1362 				    tpindex, 1);
1363 				/*
1364 				 * Prepare to search the next object in the
1365 				 * chain.
1366 				 */
1367 				backing_object = tobject->backing_object;
1368 				if (backing_object == NULL)
1369 					goto next_pindex;
1370 				VM_OBJECT_WLOCK(backing_object);
1371 				tpindex +=
1372 				    OFF_TO_IDX(tobject->backing_object_offset);
1373 				if (tobject != object)
1374 					VM_OBJECT_WUNLOCK(tobject);
1375 				tobject = backing_object;
1376 				if (!vm_object_advice_applies(tobject, advice))
1377 					goto next_pindex;
1378 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1379 			    NULL);
1380 		} else {
1381 next_page:
1382 			tm = m;
1383 			m = TAILQ_NEXT(m, listq);
1384 		}
1385 
1386 		/*
1387 		 * If the page is not in a normal state, skip it.  The page
1388 		 * can not be invalidated while the object lock is held.
1389 		 */
1390 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1391 			goto next_pindex;
1392 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1393 		    ("vm_object_madvise: page %p is fictitious", tm));
1394 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1395 		    ("vm_object_madvise: page %p is not managed", tm));
1396 		if (vm_page_tryxbusy(tm) == 0) {
1397 			if (object != tobject)
1398 				VM_OBJECT_WUNLOCK(object);
1399 			if (advice == MADV_WILLNEED) {
1400 				/*
1401 				 * Reference the page before unlocking and
1402 				 * sleeping so that the page daemon is less
1403 				 * likely to reclaim it.
1404 				 */
1405 				vm_page_aflag_set(tm, PGA_REFERENCED);
1406 			}
1407 			if (!vm_page_busy_sleep(tm, "madvpo", 0))
1408 				VM_OBJECT_WUNLOCK(tobject);
1409   			goto relookup;
1410 		}
1411 		vm_page_advise(tm, advice);
1412 		vm_page_xunbusy(tm);
1413 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1414 next_pindex:
1415 		if (tobject != object)
1416 			VM_OBJECT_WUNLOCK(tobject);
1417 	}
1418 	VM_OBJECT_WUNLOCK(object);
1419 }
1420 
1421 /*
1422  *	vm_object_shadow:
1423  *
1424  *	Create a new object which is backed by the
1425  *	specified existing object range.  The source
1426  *	object reference is deallocated.
1427  *
1428  *	The new object and offset into that object
1429  *	are returned in the source parameters.
1430  */
1431 void
1432 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1433     struct ucred *cred, bool shared)
1434 {
1435 	vm_object_t source;
1436 	vm_object_t result;
1437 
1438 	source = *object;
1439 
1440 	/*
1441 	 * Don't create the new object if the old object isn't shared.
1442 	 *
1443 	 * If we hold the only reference we can guarantee that it won't
1444 	 * increase while we have the map locked.  Otherwise the race is
1445 	 * harmless and we will end up with an extra shadow object that
1446 	 * will be collapsed later.
1447 	 */
1448 	if (source != NULL && source->ref_count == 1 &&
1449 	    (source->flags & OBJ_ANON) != 0)
1450 		return;
1451 
1452 	/*
1453 	 * Allocate a new object with the given length.
1454 	 */
1455 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1456 
1457 	/*
1458 	 * Store the offset into the source object, and fix up the offset into
1459 	 * the new object.
1460 	 */
1461 	result->backing_object_offset = *offset;
1462 
1463 	if (shared || source != NULL) {
1464 		VM_OBJECT_WLOCK(result);
1465 
1466 		/*
1467 		 * The new object shadows the source object, adding a
1468 		 * reference to it.  Our caller changes his reference
1469 		 * to point to the new object, removing a reference to
1470 		 * the source object.  Net result: no change of
1471 		 * reference count, unless the caller needs to add one
1472 		 * more reference due to forking a shared map entry.
1473 		 */
1474 		if (shared) {
1475 			vm_object_reference_locked(result);
1476 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1477 		}
1478 
1479 		/*
1480 		 * Try to optimize the result object's page color when
1481 		 * shadowing in order to maintain page coloring
1482 		 * consistency in the combined shadowed object.
1483 		 */
1484 		if (source != NULL) {
1485 			vm_object_backing_insert(result, source);
1486 			result->domain = source->domain;
1487 #if VM_NRESERVLEVEL > 0
1488 			vm_object_set_flag(result,
1489 			    (source->flags & OBJ_COLORED));
1490 			result->pg_color = (source->pg_color +
1491 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1492 			    1)) - 1);
1493 #endif
1494 		}
1495 		VM_OBJECT_WUNLOCK(result);
1496 	}
1497 
1498 	/*
1499 	 * Return the new things
1500 	 */
1501 	*offset = 0;
1502 	*object = result;
1503 }
1504 
1505 /*
1506  *	vm_object_split:
1507  *
1508  * Split the pages in a map entry into a new object.  This affords
1509  * easier removal of unused pages, and keeps object inheritance from
1510  * being a negative impact on memory usage.
1511  */
1512 void
1513 vm_object_split(vm_map_entry_t entry)
1514 {
1515 	vm_page_t m, m_next;
1516 	vm_object_t orig_object, new_object, backing_object;
1517 	vm_pindex_t idx, offidxstart;
1518 	vm_size_t size;
1519 
1520 	orig_object = entry->object.vm_object;
1521 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1522 	    ("vm_object_split:  Splitting object with multiple mappings."));
1523 	if ((orig_object->flags & OBJ_ANON) == 0)
1524 		return;
1525 	if (orig_object->ref_count <= 1)
1526 		return;
1527 	VM_OBJECT_WUNLOCK(orig_object);
1528 
1529 	offidxstart = OFF_TO_IDX(entry->offset);
1530 	size = atop(entry->end - entry->start);
1531 
1532 	new_object = vm_object_allocate_anon(size, orig_object,
1533 	    orig_object->cred, ptoa(size));
1534 
1535 	/*
1536 	 * We must wait for the orig_object to complete any in-progress
1537 	 * collapse so that the swap blocks are stable below.  The
1538 	 * additional reference on backing_object by new object will
1539 	 * prevent further collapse operations until split completes.
1540 	 */
1541 	VM_OBJECT_WLOCK(orig_object);
1542 	vm_object_collapse_wait(orig_object);
1543 
1544 	/*
1545 	 * At this point, the new object is still private, so the order in
1546 	 * which the original and new objects are locked does not matter.
1547 	 */
1548 	VM_OBJECT_WLOCK(new_object);
1549 	new_object->domain = orig_object->domain;
1550 	backing_object = orig_object->backing_object;
1551 	if (backing_object != NULL) {
1552 		vm_object_backing_insert_ref(new_object, backing_object);
1553 		new_object->backing_object_offset =
1554 		    orig_object->backing_object_offset + entry->offset;
1555 	}
1556 	if (orig_object->cred != NULL) {
1557 		crhold(orig_object->cred);
1558 		KASSERT(orig_object->charge >= ptoa(size),
1559 		    ("orig_object->charge < 0"));
1560 		orig_object->charge -= ptoa(size);
1561 	}
1562 
1563 	/*
1564 	 * Mark the split operation so that swap_pager_getpages() knows
1565 	 * that the object is in transition.
1566 	 */
1567 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1568 #ifdef INVARIANTS
1569 	idx = 0;
1570 #endif
1571 retry:
1572 	m = vm_page_find_least(orig_object, offidxstart);
1573 	KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1574 	    ("%s: object %p was repopulated", __func__, orig_object));
1575 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1576 	    m = m_next) {
1577 		m_next = TAILQ_NEXT(m, listq);
1578 
1579 		/*
1580 		 * We must wait for pending I/O to complete before we can
1581 		 * rename the page.
1582 		 *
1583 		 * We do not have to VM_PROT_NONE the page as mappings should
1584 		 * not be changed by this operation.
1585 		 */
1586 		if (vm_page_tryxbusy(m) == 0) {
1587 			VM_OBJECT_WUNLOCK(new_object);
1588 			if (vm_page_busy_sleep(m, "spltwt", 0))
1589 				VM_OBJECT_WLOCK(orig_object);
1590 			VM_OBJECT_WLOCK(new_object);
1591 			goto retry;
1592 		}
1593 
1594 		/*
1595 		 * The page was left invalid.  Likely placed there by
1596 		 * an incomplete fault.  Just remove and ignore.
1597 		 */
1598 		if (vm_page_none_valid(m)) {
1599 			if (vm_page_remove(m))
1600 				vm_page_free(m);
1601 			continue;
1602 		}
1603 
1604 		/* vm_page_rename() will dirty the page. */
1605 		if (vm_page_rename(m, new_object, idx)) {
1606 			vm_page_xunbusy(m);
1607 			VM_OBJECT_WUNLOCK(new_object);
1608 			VM_OBJECT_WUNLOCK(orig_object);
1609 			vm_radix_wait();
1610 			VM_OBJECT_WLOCK(orig_object);
1611 			VM_OBJECT_WLOCK(new_object);
1612 			goto retry;
1613 		}
1614 
1615 #if VM_NRESERVLEVEL > 0
1616 		/*
1617 		 * If some of the reservation's allocated pages remain with
1618 		 * the original object, then transferring the reservation to
1619 		 * the new object is neither particularly beneficial nor
1620 		 * particularly harmful as compared to leaving the reservation
1621 		 * with the original object.  If, however, all of the
1622 		 * reservation's allocated pages are transferred to the new
1623 		 * object, then transferring the reservation is typically
1624 		 * beneficial.  Determining which of these two cases applies
1625 		 * would be more costly than unconditionally renaming the
1626 		 * reservation.
1627 		 */
1628 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1629 #endif
1630 	}
1631 
1632 	/*
1633 	 * swap_pager_copy() can sleep, in which case the orig_object's
1634 	 * and new_object's locks are released and reacquired.
1635 	 */
1636 	swap_pager_copy(orig_object, new_object, offidxstart, 0);
1637 
1638 	TAILQ_FOREACH(m, &new_object->memq, listq)
1639 		vm_page_xunbusy(m);
1640 
1641 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1642 	VM_OBJECT_WUNLOCK(orig_object);
1643 	VM_OBJECT_WUNLOCK(new_object);
1644 	entry->object.vm_object = new_object;
1645 	entry->offset = 0LL;
1646 	vm_object_deallocate(orig_object);
1647 	VM_OBJECT_WLOCK(new_object);
1648 }
1649 
1650 static vm_page_t
1651 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1652 {
1653 	vm_object_t backing_object;
1654 
1655 	VM_OBJECT_ASSERT_WLOCKED(object);
1656 	backing_object = object->backing_object;
1657 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1658 
1659 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1660 	    ("invalid ownership %p %p %p", p, object, backing_object));
1661 	/* The page is only NULL when rename fails. */
1662 	if (p == NULL) {
1663 		VM_OBJECT_WUNLOCK(object);
1664 		VM_OBJECT_WUNLOCK(backing_object);
1665 		vm_radix_wait();
1666 		VM_OBJECT_WLOCK(object);
1667 	} else if (p->object == object) {
1668 		VM_OBJECT_WUNLOCK(backing_object);
1669 		if (vm_page_busy_sleep(p, "vmocol", 0))
1670 			VM_OBJECT_WLOCK(object);
1671 	} else {
1672 		VM_OBJECT_WUNLOCK(object);
1673 		if (!vm_page_busy_sleep(p, "vmocol", 0))
1674 			VM_OBJECT_WUNLOCK(backing_object);
1675 		VM_OBJECT_WLOCK(object);
1676 	}
1677 	VM_OBJECT_WLOCK(backing_object);
1678 	return (TAILQ_FIRST(&backing_object->memq));
1679 }
1680 
1681 static bool
1682 vm_object_scan_all_shadowed(vm_object_t object)
1683 {
1684 	vm_object_t backing_object;
1685 	vm_page_t p, pp;
1686 	vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1687 
1688 	VM_OBJECT_ASSERT_WLOCKED(object);
1689 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1690 
1691 	backing_object = object->backing_object;
1692 
1693 	if ((backing_object->flags & OBJ_ANON) == 0)
1694 		return (false);
1695 
1696 	pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1697 	p = vm_page_find_least(backing_object, pi);
1698 	ps = swap_pager_find_least(backing_object, pi);
1699 
1700 	/*
1701 	 * Only check pages inside the parent object's range and
1702 	 * inside the parent object's mapping of the backing object.
1703 	 */
1704 	for (;; pi++) {
1705 		if (p != NULL && p->pindex < pi)
1706 			p = TAILQ_NEXT(p, listq);
1707 		if (ps < pi)
1708 			ps = swap_pager_find_least(backing_object, pi);
1709 		if (p == NULL && ps >= backing_object->size)
1710 			break;
1711 		else if (p == NULL)
1712 			pi = ps;
1713 		else
1714 			pi = MIN(p->pindex, ps);
1715 
1716 		new_pindex = pi - backing_offset_index;
1717 		if (new_pindex >= object->size)
1718 			break;
1719 
1720 		if (p != NULL) {
1721 			/*
1722 			 * If the backing object page is busy a
1723 			 * grandparent or older page may still be
1724 			 * undergoing CoW.  It is not safe to collapse
1725 			 * the backing object until it is quiesced.
1726 			 */
1727 			if (vm_page_tryxbusy(p) == 0)
1728 				return (false);
1729 
1730 			/*
1731 			 * We raced with the fault handler that left
1732 			 * newly allocated invalid page on the object
1733 			 * queue and retried.
1734 			 */
1735 			if (!vm_page_all_valid(p))
1736 				goto unbusy_ret;
1737 		}
1738 
1739 		/*
1740 		 * See if the parent has the page or if the parent's object
1741 		 * pager has the page.  If the parent has the page but the page
1742 		 * is not valid, the parent's object pager must have the page.
1743 		 *
1744 		 * If this fails, the parent does not completely shadow the
1745 		 * object and we might as well give up now.
1746 		 */
1747 		pp = vm_page_lookup(object, new_pindex);
1748 
1749 		/*
1750 		 * The valid check here is stable due to object lock
1751 		 * being required to clear valid and initiate paging.
1752 		 * Busy of p disallows fault handler to validate pp.
1753 		 */
1754 		if ((pp == NULL || vm_page_none_valid(pp)) &&
1755 		    !vm_pager_has_page(object, new_pindex, NULL, NULL))
1756 			goto unbusy_ret;
1757 		if (p != NULL)
1758 			vm_page_xunbusy(p);
1759 	}
1760 	return (true);
1761 
1762 unbusy_ret:
1763 	if (p != NULL)
1764 		vm_page_xunbusy(p);
1765 	return (false);
1766 }
1767 
1768 static void
1769 vm_object_collapse_scan(vm_object_t object)
1770 {
1771 	vm_object_t backing_object;
1772 	vm_page_t next, p, pp;
1773 	vm_pindex_t backing_offset_index, new_pindex;
1774 
1775 	VM_OBJECT_ASSERT_WLOCKED(object);
1776 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1777 
1778 	backing_object = object->backing_object;
1779 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1780 
1781 	/*
1782 	 * Our scan
1783 	 */
1784 	for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1785 		next = TAILQ_NEXT(p, listq);
1786 		new_pindex = p->pindex - backing_offset_index;
1787 
1788 		/*
1789 		 * Check for busy page
1790 		 */
1791 		if (vm_page_tryxbusy(p) == 0) {
1792 			next = vm_object_collapse_scan_wait(object, p);
1793 			continue;
1794 		}
1795 
1796 		KASSERT(object->backing_object == backing_object,
1797 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1798 		    object->backing_object, backing_object));
1799 		KASSERT(p->object == backing_object,
1800 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1801 		    p->object, backing_object));
1802 
1803 		if (p->pindex < backing_offset_index ||
1804 		    new_pindex >= object->size) {
1805 			vm_pager_freespace(backing_object, p->pindex, 1);
1806 
1807 			KASSERT(!pmap_page_is_mapped(p),
1808 			    ("freeing mapped page %p", p));
1809 			if (vm_page_remove(p))
1810 				vm_page_free(p);
1811 			continue;
1812 		}
1813 
1814 		if (!vm_page_all_valid(p)) {
1815 			KASSERT(!pmap_page_is_mapped(p),
1816 			    ("freeing mapped page %p", p));
1817 			if (vm_page_remove(p))
1818 				vm_page_free(p);
1819 			continue;
1820 		}
1821 
1822 		pp = vm_page_lookup(object, new_pindex);
1823 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1824 			vm_page_xunbusy(p);
1825 			/*
1826 			 * The page in the parent is busy and possibly not
1827 			 * (yet) valid.  Until its state is finalized by the
1828 			 * busy bit owner, we can't tell whether it shadows the
1829 			 * original page.
1830 			 */
1831 			next = vm_object_collapse_scan_wait(object, pp);
1832 			continue;
1833 		}
1834 
1835 		if (pp != NULL && vm_page_none_valid(pp)) {
1836 			/*
1837 			 * The page was invalid in the parent.  Likely placed
1838 			 * there by an incomplete fault.  Just remove and
1839 			 * ignore.  p can replace it.
1840 			 */
1841 			if (vm_page_remove(pp))
1842 				vm_page_free(pp);
1843 			pp = NULL;
1844 		}
1845 
1846 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1847 			NULL)) {
1848 			/*
1849 			 * The page already exists in the parent OR swap exists
1850 			 * for this location in the parent.  Leave the parent's
1851 			 * page alone.  Destroy the original page from the
1852 			 * backing object.
1853 			 */
1854 			vm_pager_freespace(backing_object, p->pindex, 1);
1855 			KASSERT(!pmap_page_is_mapped(p),
1856 			    ("freeing mapped page %p", p));
1857 			if (vm_page_remove(p))
1858 				vm_page_free(p);
1859 			if (pp != NULL)
1860 				vm_page_xunbusy(pp);
1861 			continue;
1862 		}
1863 
1864 		/*
1865 		 * Page does not exist in parent, rename the page from the
1866 		 * backing object to the main object.
1867 		 *
1868 		 * If the page was mapped to a process, it can remain mapped
1869 		 * through the rename.  vm_page_rename() will dirty the page.
1870 		 */
1871 		if (vm_page_rename(p, object, new_pindex)) {
1872 			vm_page_xunbusy(p);
1873 			next = vm_object_collapse_scan_wait(object, NULL);
1874 			continue;
1875 		}
1876 
1877 		/* Use the old pindex to free the right page. */
1878 		vm_pager_freespace(backing_object, new_pindex +
1879 		    backing_offset_index, 1);
1880 
1881 #if VM_NRESERVLEVEL > 0
1882 		/*
1883 		 * Rename the reservation.
1884 		 */
1885 		vm_reserv_rename(p, object, backing_object,
1886 		    backing_offset_index);
1887 #endif
1888 		vm_page_xunbusy(p);
1889 	}
1890 	return;
1891 }
1892 
1893 /*
1894  *	vm_object_collapse:
1895  *
1896  *	Collapse an object with the object backing it.
1897  *	Pages in the backing object are moved into the
1898  *	parent, and the backing object is deallocated.
1899  */
1900 void
1901 vm_object_collapse(vm_object_t object)
1902 {
1903 	vm_object_t backing_object, new_backing_object;
1904 
1905 	VM_OBJECT_ASSERT_WLOCKED(object);
1906 
1907 	while (TRUE) {
1908 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1909 		    ("collapsing invalid object"));
1910 
1911 		/*
1912 		 * Wait for the backing_object to finish any pending
1913 		 * collapse so that the caller sees the shortest possible
1914 		 * shadow chain.
1915 		 */
1916 		backing_object = vm_object_backing_collapse_wait(object);
1917 		if (backing_object == NULL)
1918 			return;
1919 
1920 		KASSERT(object->ref_count > 0 &&
1921 		    object->ref_count > atomic_load_int(&object->shadow_count),
1922 		    ("collapse with invalid ref %d or shadow %d count.",
1923 		    object->ref_count, atomic_load_int(&object->shadow_count)));
1924 		KASSERT((backing_object->flags &
1925 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1926 		    ("vm_object_collapse: Backing object already collapsing."));
1927 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1928 		    ("vm_object_collapse: object is already collapsing."));
1929 
1930 		/*
1931 		 * We know that we can either collapse the backing object if
1932 		 * the parent is the only reference to it, or (perhaps) have
1933 		 * the parent bypass the object if the parent happens to shadow
1934 		 * all the resident pages in the entire backing object.
1935 		 */
1936 		if (backing_object->ref_count == 1) {
1937 			KASSERT(atomic_load_int(&backing_object->shadow_count)
1938 			    == 1,
1939 			    ("vm_object_collapse: shadow_count: %d",
1940 			    atomic_load_int(&backing_object->shadow_count)));
1941 			vm_object_pip_add(object, 1);
1942 			vm_object_set_flag(object, OBJ_COLLAPSING);
1943 			vm_object_pip_add(backing_object, 1);
1944 			vm_object_set_flag(backing_object, OBJ_DEAD);
1945 
1946 			/*
1947 			 * If there is exactly one reference to the backing
1948 			 * object, we can collapse it into the parent.
1949 			 */
1950 			vm_object_collapse_scan(object);
1951 
1952 			/*
1953 			 * Move the pager from backing_object to object.
1954 			 *
1955 			 * swap_pager_copy() can sleep, in which case the
1956 			 * backing_object's and object's locks are released and
1957 			 * reacquired.
1958 			 */
1959 			swap_pager_copy(backing_object, object,
1960 			    OFF_TO_IDX(object->backing_object_offset), TRUE);
1961 
1962 			/*
1963 			 * Object now shadows whatever backing_object did.
1964 			 */
1965 			vm_object_clear_flag(object, OBJ_COLLAPSING);
1966 			vm_object_backing_transfer(object, backing_object);
1967 			object->backing_object_offset +=
1968 			    backing_object->backing_object_offset;
1969 			VM_OBJECT_WUNLOCK(object);
1970 			vm_object_pip_wakeup(object);
1971 
1972 			/*
1973 			 * Discard backing_object.
1974 			 *
1975 			 * Since the backing object has no pages, no pager left,
1976 			 * and no object references within it, all that is
1977 			 * necessary is to dispose of it.
1978 			 */
1979 			KASSERT(backing_object->ref_count == 1, (
1980 "backing_object %p was somehow re-referenced during collapse!",
1981 			    backing_object));
1982 			vm_object_pip_wakeup(backing_object);
1983 			(void)refcount_release(&backing_object->ref_count);
1984 			umtx_shm_object_terminated(backing_object);
1985 			vm_object_terminate(backing_object);
1986 			counter_u64_add(object_collapses, 1);
1987 			VM_OBJECT_WLOCK(object);
1988 		} else {
1989 			/*
1990 			 * If we do not entirely shadow the backing object,
1991 			 * there is nothing we can do so we give up.
1992 			 *
1993 			 * The object lock and backing_object lock must not
1994 			 * be dropped during this sequence.
1995 			 */
1996 			if (!vm_object_scan_all_shadowed(object)) {
1997 				VM_OBJECT_WUNLOCK(backing_object);
1998 				break;
1999 			}
2000 
2001 			/*
2002 			 * Make the parent shadow the next object in the
2003 			 * chain.  Deallocating backing_object will not remove
2004 			 * it, since its reference count is at least 2.
2005 			 */
2006 			vm_object_backing_remove_locked(object);
2007 			new_backing_object = backing_object->backing_object;
2008 			if (new_backing_object != NULL) {
2009 				vm_object_backing_insert_ref(object,
2010 				    new_backing_object);
2011 				object->backing_object_offset +=
2012 				    backing_object->backing_object_offset;
2013 			}
2014 
2015 			/*
2016 			 * Drop the reference count on backing_object. Since
2017 			 * its ref_count was at least 2, it will not vanish.
2018 			 */
2019 			(void)refcount_release(&backing_object->ref_count);
2020 			KASSERT(backing_object->ref_count >= 1, (
2021 "backing_object %p was somehow dereferenced during collapse!",
2022 			    backing_object));
2023 			VM_OBJECT_WUNLOCK(backing_object);
2024 			counter_u64_add(object_bypasses, 1);
2025 		}
2026 
2027 		/*
2028 		 * Try again with this object's new backing object.
2029 		 */
2030 	}
2031 }
2032 
2033 /*
2034  *	vm_object_page_remove:
2035  *
2036  *	For the given object, either frees or invalidates each of the
2037  *	specified pages.  In general, a page is freed.  However, if a page is
2038  *	wired for any reason other than the existence of a managed, wired
2039  *	mapping, then it may be invalidated but not removed from the object.
2040  *	Pages are specified by the given range ["start", "end") and the option
2041  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
2042  *	extends from "start" to the end of the object.  If the option
2043  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2044  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
2045  *	specified, then the pages within the specified range must have no
2046  *	mappings.  Otherwise, if this option is not specified, any mappings to
2047  *	the specified pages are removed before the pages are freed or
2048  *	invalidated.
2049  *
2050  *	In general, this operation should only be performed on objects that
2051  *	contain managed pages.  There are, however, two exceptions.  First, it
2052  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
2053  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2054  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
2055  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
2056  *
2057  *	The object must be locked.
2058  */
2059 void
2060 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2061     int options)
2062 {
2063 	vm_page_t p, next;
2064 
2065 	VM_OBJECT_ASSERT_WLOCKED(object);
2066 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2067 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2068 	    ("vm_object_page_remove: illegal options for object %p", object));
2069 	if (object->resident_page_count == 0)
2070 		return;
2071 	vm_object_pip_add(object, 1);
2072 again:
2073 	p = vm_page_find_least(object, start);
2074 
2075 	/*
2076 	 * Here, the variable "p" is either (1) the page with the least pindex
2077 	 * greater than or equal to the parameter "start" or (2) NULL.
2078 	 */
2079 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2080 		next = TAILQ_NEXT(p, listq);
2081 
2082 		/*
2083 		 * Skip invalid pages if asked to do so.  Try to avoid acquiring
2084 		 * the busy lock, as some consumers rely on this to avoid
2085 		 * deadlocks.
2086 		 *
2087 		 * A thread may concurrently transition the page from invalid to
2088 		 * valid using only the busy lock, so the result of this check
2089 		 * is immediately stale.  It is up to consumers to handle this,
2090 		 * for instance by ensuring that all invalid->valid transitions
2091 		 * happen with a mutex held, as may be possible for a
2092 		 * filesystem.
2093 		 */
2094 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2095 			continue;
2096 
2097 		/*
2098 		 * If the page is wired for any reason besides the existence
2099 		 * of managed, wired mappings, then it cannot be freed.  For
2100 		 * example, fictitious pages, which represent device memory,
2101 		 * are inherently wired and cannot be freed.  They can,
2102 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2103 		 * not specified.
2104 		 */
2105 		if (vm_page_tryxbusy(p) == 0) {
2106 			if (vm_page_busy_sleep(p, "vmopar", 0))
2107 				VM_OBJECT_WLOCK(object);
2108 			goto again;
2109 		}
2110 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2111 			vm_page_xunbusy(p);
2112 			continue;
2113 		}
2114 		if (vm_page_wired(p)) {
2115 wired:
2116 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2117 			    object->ref_count != 0)
2118 				pmap_remove_all(p);
2119 			if ((options & OBJPR_CLEANONLY) == 0) {
2120 				vm_page_invalid(p);
2121 				vm_page_undirty(p);
2122 			}
2123 			vm_page_xunbusy(p);
2124 			continue;
2125 		}
2126 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2127 		    ("vm_object_page_remove: page %p is fictitious", p));
2128 		if ((options & OBJPR_CLEANONLY) != 0 &&
2129 		    !vm_page_none_valid(p)) {
2130 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2131 			    object->ref_count != 0 &&
2132 			    !vm_page_try_remove_write(p))
2133 				goto wired;
2134 			if (p->dirty != 0) {
2135 				vm_page_xunbusy(p);
2136 				continue;
2137 			}
2138 		}
2139 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2140 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2141 			goto wired;
2142 		vm_page_free(p);
2143 	}
2144 	vm_object_pip_wakeup(object);
2145 
2146 	vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2147 	    start);
2148 }
2149 
2150 /*
2151  *	vm_object_page_noreuse:
2152  *
2153  *	For the given object, attempt to move the specified pages to
2154  *	the head of the inactive queue.  This bypasses regular LRU
2155  *	operation and allows the pages to be reused quickly under memory
2156  *	pressure.  If a page is wired for any reason, then it will not
2157  *	be queued.  Pages are specified by the range ["start", "end").
2158  *	As a special case, if "end" is zero, then the range extends from
2159  *	"start" to the end of the object.
2160  *
2161  *	This operation should only be performed on objects that
2162  *	contain non-fictitious, managed pages.
2163  *
2164  *	The object must be locked.
2165  */
2166 void
2167 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2168 {
2169 	vm_page_t p, next;
2170 
2171 	VM_OBJECT_ASSERT_LOCKED(object);
2172 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2173 	    ("vm_object_page_noreuse: illegal object %p", object));
2174 	if (object->resident_page_count == 0)
2175 		return;
2176 	p = vm_page_find_least(object, start);
2177 
2178 	/*
2179 	 * Here, the variable "p" is either (1) the page with the least pindex
2180 	 * greater than or equal to the parameter "start" or (2) NULL.
2181 	 */
2182 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2183 		next = TAILQ_NEXT(p, listq);
2184 		vm_page_deactivate_noreuse(p);
2185 	}
2186 }
2187 
2188 /*
2189  *	Populate the specified range of the object with valid pages.  Returns
2190  *	TRUE if the range is successfully populated and FALSE otherwise.
2191  *
2192  *	Note: This function should be optimized to pass a larger array of
2193  *	pages to vm_pager_get_pages() before it is applied to a non-
2194  *	OBJT_DEVICE object.
2195  *
2196  *	The object must be locked.
2197  */
2198 boolean_t
2199 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2200 {
2201 	vm_page_t m;
2202 	vm_pindex_t pindex;
2203 	int rv;
2204 
2205 	VM_OBJECT_ASSERT_WLOCKED(object);
2206 	for (pindex = start; pindex < end; pindex++) {
2207 		rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2208 		if (rv != VM_PAGER_OK)
2209 			break;
2210 
2211 		/*
2212 		 * Keep "m" busy because a subsequent iteration may unlock
2213 		 * the object.
2214 		 */
2215 	}
2216 	if (pindex > start) {
2217 		m = vm_page_lookup(object, start);
2218 		while (m != NULL && m->pindex < pindex) {
2219 			vm_page_xunbusy(m);
2220 			m = TAILQ_NEXT(m, listq);
2221 		}
2222 	}
2223 	return (pindex == end);
2224 }
2225 
2226 /*
2227  *	Routine:	vm_object_coalesce
2228  *	Function:	Coalesces two objects backing up adjoining
2229  *			regions of memory into a single object.
2230  *
2231  *	returns TRUE if objects were combined.
2232  *
2233  *	NOTE:	Only works at the moment if the second object is NULL -
2234  *		if it's not, which object do we lock first?
2235  *
2236  *	Parameters:
2237  *		prev_object	First object to coalesce
2238  *		prev_offset	Offset into prev_object
2239  *		prev_size	Size of reference to prev_object
2240  *		next_size	Size of reference to the second object
2241  *		reserved	Indicator that extension region has
2242  *				swap accounted for
2243  *
2244  *	Conditions:
2245  *	The object must *not* be locked.
2246  */
2247 boolean_t
2248 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2249     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2250 {
2251 	vm_pindex_t next_pindex;
2252 
2253 	if (prev_object == NULL)
2254 		return (TRUE);
2255 	if ((prev_object->flags & OBJ_ANON) == 0)
2256 		return (FALSE);
2257 
2258 	VM_OBJECT_WLOCK(prev_object);
2259 	/*
2260 	 * Try to collapse the object first.
2261 	 */
2262 	vm_object_collapse(prev_object);
2263 
2264 	/*
2265 	 * Can't coalesce if: . more than one reference . paged out . shadows
2266 	 * another object . has a copy elsewhere (any of which mean that the
2267 	 * pages not mapped to prev_entry may be in use anyway)
2268 	 */
2269 	if (prev_object->backing_object != NULL) {
2270 		VM_OBJECT_WUNLOCK(prev_object);
2271 		return (FALSE);
2272 	}
2273 
2274 	prev_size >>= PAGE_SHIFT;
2275 	next_size >>= PAGE_SHIFT;
2276 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2277 
2278 	if (prev_object->ref_count > 1 &&
2279 	    prev_object->size != next_pindex &&
2280 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2281 		VM_OBJECT_WUNLOCK(prev_object);
2282 		return (FALSE);
2283 	}
2284 
2285 	/*
2286 	 * Account for the charge.
2287 	 */
2288 	if (prev_object->cred != NULL) {
2289 		/*
2290 		 * If prev_object was charged, then this mapping,
2291 		 * although not charged now, may become writable
2292 		 * later. Non-NULL cred in the object would prevent
2293 		 * swap reservation during enabling of the write
2294 		 * access, so reserve swap now. Failed reservation
2295 		 * cause allocation of the separate object for the map
2296 		 * entry, and swap reservation for this entry is
2297 		 * managed in appropriate time.
2298 		 */
2299 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2300 		    prev_object->cred)) {
2301 			VM_OBJECT_WUNLOCK(prev_object);
2302 			return (FALSE);
2303 		}
2304 		prev_object->charge += ptoa(next_size);
2305 	}
2306 
2307 	/*
2308 	 * Remove any pages that may still be in the object from a previous
2309 	 * deallocation.
2310 	 */
2311 	if (next_pindex < prev_object->size) {
2312 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2313 		    next_size, 0);
2314 #if 0
2315 		if (prev_object->cred != NULL) {
2316 			KASSERT(prev_object->charge >=
2317 			    ptoa(prev_object->size - next_pindex),
2318 			    ("object %p overcharged 1 %jx %jx", prev_object,
2319 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2320 			prev_object->charge -= ptoa(prev_object->size -
2321 			    next_pindex);
2322 		}
2323 #endif
2324 	}
2325 
2326 	/*
2327 	 * Extend the object if necessary.
2328 	 */
2329 	if (next_pindex + next_size > prev_object->size)
2330 		prev_object->size = next_pindex + next_size;
2331 
2332 	VM_OBJECT_WUNLOCK(prev_object);
2333 	return (TRUE);
2334 }
2335 
2336 void
2337 vm_object_set_writeable_dirty_(vm_object_t object)
2338 {
2339 	atomic_add_int(&object->generation, 1);
2340 }
2341 
2342 bool
2343 vm_object_mightbedirty_(vm_object_t object)
2344 {
2345 	return (object->generation != object->cleangeneration);
2346 }
2347 
2348 /*
2349  *	vm_object_unwire:
2350  *
2351  *	For each page offset within the specified range of the given object,
2352  *	find the highest-level page in the shadow chain and unwire it.  A page
2353  *	must exist at every page offset, and the highest-level page must be
2354  *	wired.
2355  */
2356 void
2357 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2358     uint8_t queue)
2359 {
2360 	vm_object_t tobject, t1object;
2361 	vm_page_t m, tm;
2362 	vm_pindex_t end_pindex, pindex, tpindex;
2363 	int depth, locked_depth;
2364 
2365 	KASSERT((offset & PAGE_MASK) == 0,
2366 	    ("vm_object_unwire: offset is not page aligned"));
2367 	KASSERT((length & PAGE_MASK) == 0,
2368 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2369 	/* The wired count of a fictitious page never changes. */
2370 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2371 		return;
2372 	pindex = OFF_TO_IDX(offset);
2373 	end_pindex = pindex + atop(length);
2374 again:
2375 	locked_depth = 1;
2376 	VM_OBJECT_RLOCK(object);
2377 	m = vm_page_find_least(object, pindex);
2378 	while (pindex < end_pindex) {
2379 		if (m == NULL || pindex < m->pindex) {
2380 			/*
2381 			 * The first object in the shadow chain doesn't
2382 			 * contain a page at the current index.  Therefore,
2383 			 * the page must exist in a backing object.
2384 			 */
2385 			tobject = object;
2386 			tpindex = pindex;
2387 			depth = 0;
2388 			do {
2389 				tpindex +=
2390 				    OFF_TO_IDX(tobject->backing_object_offset);
2391 				tobject = tobject->backing_object;
2392 				KASSERT(tobject != NULL,
2393 				    ("vm_object_unwire: missing page"));
2394 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2395 					goto next_page;
2396 				depth++;
2397 				if (depth == locked_depth) {
2398 					locked_depth++;
2399 					VM_OBJECT_RLOCK(tobject);
2400 				}
2401 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2402 			    NULL);
2403 		} else {
2404 			tm = m;
2405 			m = TAILQ_NEXT(m, listq);
2406 		}
2407 		if (vm_page_trysbusy(tm) == 0) {
2408 			for (tobject = object; locked_depth >= 1;
2409 			    locked_depth--) {
2410 				t1object = tobject->backing_object;
2411 				if (tm->object != tobject)
2412 					VM_OBJECT_RUNLOCK(tobject);
2413 				tobject = t1object;
2414 			}
2415 			tobject = tm->object;
2416 			if (!vm_page_busy_sleep(tm, "unwbo",
2417 			    VM_ALLOC_IGN_SBUSY))
2418 				VM_OBJECT_RUNLOCK(tobject);
2419 			goto again;
2420 		}
2421 		vm_page_unwire(tm, queue);
2422 		vm_page_sunbusy(tm);
2423 next_page:
2424 		pindex++;
2425 	}
2426 	/* Release the accumulated object locks. */
2427 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2428 		t1object = tobject->backing_object;
2429 		VM_OBJECT_RUNLOCK(tobject);
2430 		tobject = t1object;
2431 	}
2432 }
2433 
2434 /*
2435  * Return the vnode for the given object, or NULL if none exists.
2436  * For tmpfs objects, the function may return NULL if there is
2437  * no vnode allocated at the time of the call.
2438  */
2439 struct vnode *
2440 vm_object_vnode(vm_object_t object)
2441 {
2442 	struct vnode *vp;
2443 
2444 	VM_OBJECT_ASSERT_LOCKED(object);
2445 	vm_pager_getvp(object, &vp, NULL);
2446 	return (vp);
2447 }
2448 
2449 /*
2450  * Busy the vm object.  This prevents new pages belonging to the object from
2451  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2452  * for checking page state before proceeding.
2453  */
2454 void
2455 vm_object_busy(vm_object_t obj)
2456 {
2457 
2458 	VM_OBJECT_ASSERT_LOCKED(obj);
2459 
2460 	blockcount_acquire(&obj->busy, 1);
2461 	/* The fence is required to order loads of page busy. */
2462 	atomic_thread_fence_acq_rel();
2463 }
2464 
2465 void
2466 vm_object_unbusy(vm_object_t obj)
2467 {
2468 
2469 	blockcount_release(&obj->busy, 1);
2470 }
2471 
2472 void
2473 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2474 {
2475 
2476 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2477 
2478 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2479 }
2480 
2481 /*
2482  * This function aims to determine if the object is mapped,
2483  * specifically, if it is referenced by a vm_map_entry.  Because
2484  * objects occasionally acquire transient references that do not
2485  * represent a mapping, the method used here is inexact.  However, it
2486  * has very low overhead and is good enough for the advisory
2487  * vm.vmtotal sysctl.
2488  */
2489 bool
2490 vm_object_is_active(vm_object_t obj)
2491 {
2492 
2493 	return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2494 }
2495 
2496 static int
2497 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2498 {
2499 	struct kinfo_vmobject *kvo;
2500 	char *fullpath, *freepath;
2501 	struct vnode *vp;
2502 	struct vattr va;
2503 	vm_object_t obj;
2504 	vm_page_t m;
2505 	u_long sp;
2506 	int count, error;
2507 	bool want_path;
2508 
2509 	if (req->oldptr == NULL) {
2510 		/*
2511 		 * If an old buffer has not been provided, generate an
2512 		 * estimate of the space needed for a subsequent call.
2513 		 */
2514 		mtx_lock(&vm_object_list_mtx);
2515 		count = 0;
2516 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2517 			if (obj->type == OBJT_DEAD)
2518 				continue;
2519 			count++;
2520 		}
2521 		mtx_unlock(&vm_object_list_mtx);
2522 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2523 		    count * 11 / 10));
2524 	}
2525 
2526 	want_path = !(swap_only || jailed(curthread->td_ucred));
2527 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO);
2528 	error = 0;
2529 
2530 	/*
2531 	 * VM objects are type stable and are never removed from the
2532 	 * list once added.  This allows us to safely read obj->object_list
2533 	 * after reacquiring the VM object lock.
2534 	 */
2535 	mtx_lock(&vm_object_list_mtx);
2536 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2537 		if (obj->type == OBJT_DEAD ||
2538 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2539 			continue;
2540 		VM_OBJECT_RLOCK(obj);
2541 		if (obj->type == OBJT_DEAD ||
2542 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2543 			VM_OBJECT_RUNLOCK(obj);
2544 			continue;
2545 		}
2546 		mtx_unlock(&vm_object_list_mtx);
2547 		kvo->kvo_size = ptoa(obj->size);
2548 		kvo->kvo_resident = obj->resident_page_count;
2549 		kvo->kvo_ref_count = obj->ref_count;
2550 		kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2551 		kvo->kvo_memattr = obj->memattr;
2552 		kvo->kvo_active = 0;
2553 		kvo->kvo_inactive = 0;
2554 		if (!swap_only) {
2555 			TAILQ_FOREACH(m, &obj->memq, listq) {
2556 				/*
2557 				 * A page may belong to the object but be
2558 				 * dequeued and set to PQ_NONE while the
2559 				 * object lock is not held.  This makes the
2560 				 * reads of m->queue below racy, and we do not
2561 				 * count pages set to PQ_NONE.  However, this
2562 				 * sysctl is only meant to give an
2563 				 * approximation of the system anyway.
2564 				 */
2565 				if (m->a.queue == PQ_ACTIVE)
2566 					kvo->kvo_active++;
2567 				else if (m->a.queue == PQ_INACTIVE)
2568 					kvo->kvo_inactive++;
2569 			}
2570 		}
2571 
2572 		kvo->kvo_vn_fileid = 0;
2573 		kvo->kvo_vn_fsid = 0;
2574 		kvo->kvo_vn_fsid_freebsd11 = 0;
2575 		freepath = NULL;
2576 		fullpath = "";
2577 		vp = NULL;
2578 		kvo->kvo_type = vm_object_kvme_type(obj, want_path ? &vp :
2579 		    NULL);
2580 		if (vp != NULL) {
2581 			vref(vp);
2582 		} else if ((obj->flags & OBJ_ANON) != 0) {
2583 			MPASS(kvo->kvo_type == KVME_TYPE_SWAP);
2584 			kvo->kvo_me = (uintptr_t)obj;
2585 			/* tmpfs objs are reported as vnodes */
2586 			kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2587 			sp = swap_pager_swapped_pages(obj);
2588 			kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2589 		}
2590 		VM_OBJECT_RUNLOCK(obj);
2591 		if (vp != NULL) {
2592 			vn_fullpath(vp, &fullpath, &freepath);
2593 			vn_lock(vp, LK_SHARED | LK_RETRY);
2594 			if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2595 				kvo->kvo_vn_fileid = va.va_fileid;
2596 				kvo->kvo_vn_fsid = va.va_fsid;
2597 				kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2598 								/* truncate */
2599 			}
2600 			vput(vp);
2601 		}
2602 
2603 		strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2604 		free(freepath, M_TEMP);
2605 
2606 		/* Pack record size down */
2607 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2608 		    + strlen(kvo->kvo_path) + 1;
2609 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2610 		    sizeof(uint64_t));
2611 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2612 		maybe_yield();
2613 		mtx_lock(&vm_object_list_mtx);
2614 		if (error)
2615 			break;
2616 	}
2617 	mtx_unlock(&vm_object_list_mtx);
2618 	free(kvo, M_TEMP);
2619 	return (error);
2620 }
2621 
2622 static int
2623 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2624 {
2625 	return (vm_object_list_handler(req, false));
2626 }
2627 
2628 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2629     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2630     "List of VM objects");
2631 
2632 static int
2633 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2634 {
2635 	return (vm_object_list_handler(req, true));
2636 }
2637 
2638 /*
2639  * This sysctl returns list of the anonymous or swap objects. Intent
2640  * is to provide stripped optimized list useful to analyze swap use.
2641  * Since technically non-swap (default) objects participate in the
2642  * shadow chains, and are converted to swap type as needed by swap
2643  * pager, we must report them.
2644  */
2645 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2646     CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2647     sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2648     "List of swap VM objects");
2649 
2650 #include "opt_ddb.h"
2651 #ifdef DDB
2652 #include <sys/kernel.h>
2653 
2654 #include <sys/cons.h>
2655 
2656 #include <ddb/ddb.h>
2657 
2658 static int
2659 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2660 {
2661 	vm_map_t tmpm;
2662 	vm_map_entry_t tmpe;
2663 	vm_object_t obj;
2664 
2665 	if (map == 0)
2666 		return 0;
2667 
2668 	if (entry == 0) {
2669 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2670 			if (_vm_object_in_map(map, object, tmpe)) {
2671 				return 1;
2672 			}
2673 		}
2674 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2675 		tmpm = entry->object.sub_map;
2676 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2677 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2678 				return 1;
2679 			}
2680 		}
2681 	} else if ((obj = entry->object.vm_object) != NULL) {
2682 		for (; obj; obj = obj->backing_object)
2683 			if (obj == object) {
2684 				return 1;
2685 			}
2686 	}
2687 	return 0;
2688 }
2689 
2690 static int
2691 vm_object_in_map(vm_object_t object)
2692 {
2693 	struct proc *p;
2694 
2695 	/* sx_slock(&allproc_lock); */
2696 	FOREACH_PROC_IN_SYSTEM(p) {
2697 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2698 			continue;
2699 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2700 			/* sx_sunlock(&allproc_lock); */
2701 			return 1;
2702 		}
2703 	}
2704 	/* sx_sunlock(&allproc_lock); */
2705 	if (_vm_object_in_map(kernel_map, object, 0))
2706 		return 1;
2707 	return 0;
2708 }
2709 
2710 DB_SHOW_COMMAND_FLAGS(vmochk, vm_object_check, DB_CMD_MEMSAFE)
2711 {
2712 	vm_object_t object;
2713 
2714 	/*
2715 	 * make sure that internal objs are in a map somewhere
2716 	 * and none have zero ref counts.
2717 	 */
2718 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2719 		if ((object->flags & OBJ_ANON) != 0) {
2720 			if (object->ref_count == 0) {
2721 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2722 					(long)object->size);
2723 			}
2724 			if (!vm_object_in_map(object)) {
2725 				db_printf(
2726 			"vmochk: internal obj is not in a map: "
2727 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2728 				    object->ref_count, (u_long)object->size,
2729 				    (u_long)object->size,
2730 				    (void *)object->backing_object);
2731 			}
2732 		}
2733 		if (db_pager_quit)
2734 			return;
2735 	}
2736 }
2737 
2738 /*
2739  *	vm_object_print:	[ debug ]
2740  */
2741 DB_SHOW_COMMAND(object, vm_object_print_static)
2742 {
2743 	/* XXX convert args. */
2744 	vm_object_t object = (vm_object_t)addr;
2745 	boolean_t full = have_addr;
2746 
2747 	vm_page_t p;
2748 
2749 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2750 #define	count	was_count
2751 
2752 	int count;
2753 
2754 	if (object == NULL)
2755 		return;
2756 
2757 	db_iprintf(
2758 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2759 	    object, (int)object->type, (uintmax_t)object->size,
2760 	    object->resident_page_count, object->ref_count, object->flags,
2761 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2762 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2763 	    atomic_load_int(&object->shadow_count),
2764 	    object->backing_object ? object->backing_object->ref_count : 0,
2765 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2766 
2767 	if (!full)
2768 		return;
2769 
2770 	db_indent += 2;
2771 	count = 0;
2772 	TAILQ_FOREACH(p, &object->memq, listq) {
2773 		if (count == 0)
2774 			db_iprintf("memory:=");
2775 		else if (count == 6) {
2776 			db_printf("\n");
2777 			db_iprintf(" ...");
2778 			count = 0;
2779 		} else
2780 			db_printf(",");
2781 		count++;
2782 
2783 		db_printf("(off=0x%jx,page=0x%jx)",
2784 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2785 
2786 		if (db_pager_quit)
2787 			break;
2788 	}
2789 	if (count != 0)
2790 		db_printf("\n");
2791 	db_indent -= 2;
2792 }
2793 
2794 /* XXX. */
2795 #undef count
2796 
2797 /* XXX need this non-static entry for calling from vm_map_print. */
2798 void
2799 vm_object_print(
2800         /* db_expr_t */ long addr,
2801 	boolean_t have_addr,
2802 	/* db_expr_t */ long count,
2803 	char *modif)
2804 {
2805 	vm_object_print_static(addr, have_addr, count, modif);
2806 }
2807 
2808 DB_SHOW_COMMAND_FLAGS(vmopag, vm_object_print_pages, DB_CMD_MEMSAFE)
2809 {
2810 	vm_object_t object;
2811 	vm_pindex_t fidx;
2812 	vm_paddr_t pa;
2813 	vm_page_t m, prev_m;
2814 	int rcount;
2815 
2816 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2817 		db_printf("new object: %p\n", (void *)object);
2818 		if (db_pager_quit)
2819 			return;
2820 
2821 		rcount = 0;
2822 		fidx = 0;
2823 		pa = -1;
2824 		TAILQ_FOREACH(m, &object->memq, listq) {
2825 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2826 			    prev_m->pindex + 1 != m->pindex) {
2827 				if (rcount) {
2828 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2829 						(long)fidx, rcount, (long)pa);
2830 					if (db_pager_quit)
2831 						return;
2832 					rcount = 0;
2833 				}
2834 			}
2835 			if (rcount &&
2836 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2837 				++rcount;
2838 				continue;
2839 			}
2840 			if (rcount) {
2841 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2842 					(long)fidx, rcount, (long)pa);
2843 				if (db_pager_quit)
2844 					return;
2845 			}
2846 			fidx = m->pindex;
2847 			pa = VM_PAGE_TO_PHYS(m);
2848 			rcount = 1;
2849 		}
2850 		if (rcount) {
2851 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2852 				(long)fidx, rcount, (long)pa);
2853 			if (db_pager_quit)
2854 				return;
2855 		}
2856 	}
2857 }
2858 #endif /* DDB */
2859