xref: /freebsd/sys/vm/vm_object.c (revision bd66c1b43e33540205dbc1187c2f2a15c58b57ba)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39  *
40  * Permission to use, copy, modify and distribute this software and
41  * its documentation is hereby granted, provided that both the copyright
42  * notice and this permission notice appear in all copies of the
43  * software, derivative works or modified versions, and any portions
44  * thereof, and that both notices appear in supporting documentation.
45  *
46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49  *
50  * Carnegie Mellon requests users of this software to return to
51  *
52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
53  *  School of Computer Science
54  *  Carnegie Mellon University
55  *  Pittsburgh PA 15213-3890
56  *
57  * any improvements or extensions that they make and grant Carnegie the
58  * rights to redistribute these changes.
59  */
60 
61 /*
62  *	Virtual memory object module.
63  */
64 
65 #include "opt_vm.h"
66 
67 #include <sys/systm.h>
68 #include <sys/blockcount.h>
69 #include <sys/conf.h>
70 #include <sys/cpuset.h>
71 #include <sys/ipc.h>
72 #include <sys/jail.h>
73 #include <sys/limits.h>
74 #include <sys/lock.h>
75 #include <sys/mman.h>
76 #include <sys/mount.h>
77 #include <sys/kernel.h>
78 #include <sys/mutex.h>
79 #include <sys/pctrie.h>
80 #include <sys/proc.h>
81 #include <sys/refcount.h>
82 #include <sys/shm.h>
83 #include <sys/sx.h>
84 #include <sys/sysctl.h>
85 #include <sys/resourcevar.h>
86 #include <sys/refcount.h>
87 #include <sys/rwlock.h>
88 #include <sys/user.h>
89 #include <sys/vnode.h>
90 #include <sys/vmmeter.h>
91 
92 #include <vm/vm.h>
93 #include <vm/vm_param.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_pager.h>
100 #include <vm/vm_phys.h>
101 #include <vm/vm_pagequeue.h>
102 #include <vm/swap_pager.h>
103 #include <vm/vm_kern.h>
104 #include <vm/vm_extern.h>
105 #include <vm/vm_radix.h>
106 #include <vm/vm_reserv.h>
107 #include <vm/uma.h>
108 
109 static int old_msync;
110 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
111     "Use old (insecure) msync behavior");
112 
113 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
114 		    int pagerflags, int flags, boolean_t *allclean,
115 		    boolean_t *eio);
116 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
117 		    boolean_t *allclean);
118 static void	vm_object_backing_remove(vm_object_t object);
119 
120 /*
121  *	Virtual memory objects maintain the actual data
122  *	associated with allocated virtual memory.  A given
123  *	page of memory exists within exactly one object.
124  *
125  *	An object is only deallocated when all "references"
126  *	are given up.  Only one "reference" to a given
127  *	region of an object should be writeable.
128  *
129  *	Associated with each object is a list of all resident
130  *	memory pages belonging to that object; this list is
131  *	maintained by the "vm_page" module, and locked by the object's
132  *	lock.
133  *
134  *	Each object also records a "pager" routine which is
135  *	used to retrieve (and store) pages to the proper backing
136  *	storage.  In addition, objects may be backed by other
137  *	objects from which they were virtual-copied.
138  *
139  *	The only items within the object structure which are
140  *	modified after time of creation are:
141  *		reference count		locked by object's lock
142  *		pager routine		locked by object's lock
143  *
144  */
145 
146 struct object_q vm_object_list;
147 struct mtx vm_object_list_mtx;	/* lock for object list and count */
148 
149 struct vm_object kernel_object_store;
150 
151 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
152     "VM object stats");
153 
154 static COUNTER_U64_DEFINE_EARLY(object_collapses);
155 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
156     &object_collapses,
157     "VM object collapses");
158 
159 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
160 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
161     &object_bypasses,
162     "VM object bypasses");
163 
164 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
165 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
166     &object_collapse_waits,
167     "Number of sleeps for collapse");
168 
169 static uma_zone_t obj_zone;
170 
171 static int vm_object_zinit(void *mem, int size, int flags);
172 
173 #ifdef INVARIANTS
174 static void vm_object_zdtor(void *mem, int size, void *arg);
175 
176 static void
177 vm_object_zdtor(void *mem, int size, void *arg)
178 {
179 	vm_object_t object;
180 
181 	object = (vm_object_t)mem;
182 	KASSERT(object->ref_count == 0,
183 	    ("object %p ref_count = %d", object, object->ref_count));
184 	KASSERT(TAILQ_EMPTY(&object->memq),
185 	    ("object %p has resident pages in its memq", object));
186 	KASSERT(vm_radix_is_empty(&object->rtree),
187 	    ("object %p has resident pages in its trie", object));
188 #if VM_NRESERVLEVEL > 0
189 	KASSERT(LIST_EMPTY(&object->rvq),
190 	    ("object %p has reservations",
191 	    object));
192 #endif
193 	KASSERT(!vm_object_busied(object),
194 	    ("object %p busy = %d", object, blockcount_read(&object->busy)));
195 	KASSERT(object->resident_page_count == 0,
196 	    ("object %p resident_page_count = %d",
197 	    object, object->resident_page_count));
198 	KASSERT(atomic_load_int(&object->shadow_count) == 0,
199 	    ("object %p shadow_count = %d",
200 	    object, atomic_load_int(&object->shadow_count)));
201 	KASSERT(object->type == OBJT_DEAD,
202 	    ("object %p has non-dead type %d",
203 	    object, object->type));
204 	KASSERT(object->charge == 0 && object->cred == NULL,
205 	    ("object %p has non-zero charge %ju (%p)",
206 	    object, (uintmax_t)object->charge, object->cred));
207 }
208 #endif
209 
210 static int
211 vm_object_zinit(void *mem, int size, int flags)
212 {
213 	vm_object_t object;
214 
215 	object = (vm_object_t)mem;
216 	rw_init_flags(&object->lock, "vmobject", RW_DUPOK | RW_NEW);
217 
218 	/* These are true for any object that has been freed */
219 	object->type = OBJT_DEAD;
220 	vm_radix_init(&object->rtree);
221 	refcount_init(&object->ref_count, 0);
222 	blockcount_init(&object->paging_in_progress);
223 	blockcount_init(&object->busy);
224 	object->resident_page_count = 0;
225 	atomic_store_int(&object->shadow_count, 0);
226 	object->flags = OBJ_DEAD;
227 
228 	mtx_lock(&vm_object_list_mtx);
229 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
230 	mtx_unlock(&vm_object_list_mtx);
231 	return (0);
232 }
233 
234 static void
235 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
236     vm_object_t object, void *handle)
237 {
238 
239 	TAILQ_INIT(&object->memq);
240 	LIST_INIT(&object->shadow_head);
241 
242 	object->type = type;
243 	object->flags = flags;
244 	if ((flags & OBJ_SWAP) != 0) {
245 		pctrie_init(&object->un_pager.swp.swp_blks);
246 		object->un_pager.swp.writemappings = 0;
247 	}
248 
249 	/*
250 	 * Ensure that swap_pager_swapoff() iteration over object_list
251 	 * sees up to date type and pctrie head if it observed
252 	 * non-dead object.
253 	 */
254 	atomic_thread_fence_rel();
255 
256 	object->pg_color = 0;
257 	object->size = size;
258 	object->domain.dr_policy = NULL;
259 	object->generation = 1;
260 	object->cleangeneration = 1;
261 	refcount_init(&object->ref_count, 1);
262 	object->memattr = VM_MEMATTR_DEFAULT;
263 	object->cred = NULL;
264 	object->charge = 0;
265 	object->handle = handle;
266 	object->backing_object = NULL;
267 	object->backing_object_offset = (vm_ooffset_t) 0;
268 #if VM_NRESERVLEVEL > 0
269 	LIST_INIT(&object->rvq);
270 #endif
271 	umtx_shm_object_init(object);
272 }
273 
274 /*
275  *	vm_object_init:
276  *
277  *	Initialize the VM objects module.
278  */
279 void
280 vm_object_init(void)
281 {
282 	TAILQ_INIT(&vm_object_list);
283 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
284 
285 	rw_init(&kernel_object->lock, "kernel vm object");
286 	vm_radix_init(&kernel_object->rtree);
287 	_vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
288 	    VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
289 #if VM_NRESERVLEVEL > 0
290 	kernel_object->flags |= OBJ_COLORED;
291 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
292 #endif
293 	kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
294 
295 	/*
296 	 * The lock portion of struct vm_object must be type stable due
297 	 * to vm_pageout_fallback_object_lock locking a vm object
298 	 * without holding any references to it.
299 	 *
300 	 * paging_in_progress is valid always.  Lockless references to
301 	 * the objects may acquire pip and then check OBJ_DEAD.
302 	 */
303 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
304 #ifdef INVARIANTS
305 	    vm_object_zdtor,
306 #else
307 	    NULL,
308 #endif
309 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
310 
311 	vm_radix_zinit();
312 }
313 
314 void
315 vm_object_clear_flag(vm_object_t object, u_short bits)
316 {
317 
318 	VM_OBJECT_ASSERT_WLOCKED(object);
319 	object->flags &= ~bits;
320 }
321 
322 /*
323  *	Sets the default memory attribute for the specified object.  Pages
324  *	that are allocated to this object are by default assigned this memory
325  *	attribute.
326  *
327  *	Presently, this function must be called before any pages are allocated
328  *	to the object.  In the future, this requirement may be relaxed for
329  *	"default" and "swap" objects.
330  */
331 int
332 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
333 {
334 
335 	VM_OBJECT_ASSERT_WLOCKED(object);
336 
337 	if (object->type == OBJT_DEAD)
338 		return (KERN_INVALID_ARGUMENT);
339 	if (!TAILQ_EMPTY(&object->memq))
340 		return (KERN_FAILURE);
341 
342 	object->memattr = memattr;
343 	return (KERN_SUCCESS);
344 }
345 
346 void
347 vm_object_pip_add(vm_object_t object, short i)
348 {
349 
350 	if (i > 0)
351 		blockcount_acquire(&object->paging_in_progress, i);
352 }
353 
354 void
355 vm_object_pip_wakeup(vm_object_t object)
356 {
357 
358 	vm_object_pip_wakeupn(object, 1);
359 }
360 
361 void
362 vm_object_pip_wakeupn(vm_object_t object, short i)
363 {
364 
365 	if (i > 0)
366 		blockcount_release(&object->paging_in_progress, i);
367 }
368 
369 /*
370  * Atomically drop the object lock and wait for pip to drain.  This protects
371  * from sleep/wakeup races due to identity changes.  The lock is not re-acquired
372  * on return.
373  */
374 static void
375 vm_object_pip_sleep(vm_object_t object, const char *waitid)
376 {
377 
378 	(void)blockcount_sleep(&object->paging_in_progress, &object->lock,
379 	    waitid, PVM | PDROP);
380 }
381 
382 void
383 vm_object_pip_wait(vm_object_t object, const char *waitid)
384 {
385 
386 	VM_OBJECT_ASSERT_WLOCKED(object);
387 
388 	blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
389 	    PVM);
390 }
391 
392 void
393 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
394 {
395 
396 	VM_OBJECT_ASSERT_UNLOCKED(object);
397 
398 	blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
399 }
400 
401 /*
402  *	vm_object_allocate:
403  *
404  *	Returns a new object with the given size.
405  */
406 vm_object_t
407 vm_object_allocate(objtype_t type, vm_pindex_t size)
408 {
409 	vm_object_t object;
410 	u_short flags;
411 
412 	switch (type) {
413 	case OBJT_DEAD:
414 		panic("vm_object_allocate: can't create OBJT_DEAD");
415 	case OBJT_SWAP:
416 		flags = OBJ_COLORED | OBJ_SWAP;
417 		break;
418 	case OBJT_DEVICE:
419 	case OBJT_SG:
420 		flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
421 		break;
422 	case OBJT_MGTDEVICE:
423 		flags = OBJ_FICTITIOUS;
424 		break;
425 	case OBJT_PHYS:
426 		flags = OBJ_UNMANAGED;
427 		break;
428 	case OBJT_VNODE:
429 		flags = 0;
430 		break;
431 	default:
432 		panic("vm_object_allocate: type %d is undefined or dynamic",
433 		    type);
434 	}
435 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
436 	_vm_object_allocate(type, size, flags, object, NULL);
437 
438 	return (object);
439 }
440 
441 vm_object_t
442 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags)
443 {
444 	vm_object_t object;
445 
446 	MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */);
447 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
448 	_vm_object_allocate(dyntype, size, flags, object, NULL);
449 
450 	return (object);
451 }
452 
453 /*
454  *	vm_object_allocate_anon:
455  *
456  *	Returns a new default object of the given size and marked as
457  *	anonymous memory for special split/collapse handling.  Color
458  *	to be initialized by the caller.
459  */
460 vm_object_t
461 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
462     struct ucred *cred, vm_size_t charge)
463 {
464 	vm_object_t handle, object;
465 
466 	if (backing_object == NULL)
467 		handle = NULL;
468 	else if ((backing_object->flags & OBJ_ANON) != 0)
469 		handle = backing_object->handle;
470 	else
471 		handle = backing_object;
472 	object = uma_zalloc(obj_zone, M_WAITOK);
473 	_vm_object_allocate(OBJT_SWAP, size,
474 	    OBJ_ANON | OBJ_ONEMAPPING | OBJ_SWAP, object, handle);
475 	object->cred = cred;
476 	object->charge = cred != NULL ? charge : 0;
477 	return (object);
478 }
479 
480 static void
481 vm_object_reference_vnode(vm_object_t object)
482 {
483 	u_int old;
484 
485 	/*
486 	 * vnode objects need the lock for the first reference
487 	 * to serialize with vnode_object_deallocate().
488 	 */
489 	if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
490 		VM_OBJECT_RLOCK(object);
491 		old = refcount_acquire(&object->ref_count);
492 		if (object->type == OBJT_VNODE && old == 0)
493 			vref(object->handle);
494 		VM_OBJECT_RUNLOCK(object);
495 	}
496 }
497 
498 /*
499  *	vm_object_reference:
500  *
501  *	Acquires a reference to the given object.
502  */
503 void
504 vm_object_reference(vm_object_t object)
505 {
506 
507 	if (object == NULL)
508 		return;
509 
510 	if (object->type == OBJT_VNODE)
511 		vm_object_reference_vnode(object);
512 	else
513 		refcount_acquire(&object->ref_count);
514 	KASSERT((object->flags & OBJ_DEAD) == 0,
515 	    ("vm_object_reference: Referenced dead object."));
516 }
517 
518 /*
519  *	vm_object_reference_locked:
520  *
521  *	Gets another reference to the given object.
522  *
523  *	The object must be locked.
524  */
525 void
526 vm_object_reference_locked(vm_object_t object)
527 {
528 	u_int old;
529 
530 	VM_OBJECT_ASSERT_LOCKED(object);
531 	old = refcount_acquire(&object->ref_count);
532 	if (object->type == OBJT_VNODE && old == 0)
533 		vref(object->handle);
534 	KASSERT((object->flags & OBJ_DEAD) == 0,
535 	    ("vm_object_reference: Referenced dead object."));
536 }
537 
538 /*
539  * Handle deallocating an object of type OBJT_VNODE.
540  */
541 static void
542 vm_object_deallocate_vnode(vm_object_t object)
543 {
544 	struct vnode *vp = (struct vnode *) object->handle;
545 	bool last;
546 
547 	KASSERT(object->type == OBJT_VNODE,
548 	    ("vm_object_deallocate_vnode: not a vnode object"));
549 	KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
550 
551 	/* Object lock to protect handle lookup. */
552 	last = refcount_release(&object->ref_count);
553 	VM_OBJECT_RUNLOCK(object);
554 
555 	if (!last)
556 		return;
557 
558 	if (!umtx_shm_vnobj_persistent)
559 		umtx_shm_object_terminated(object);
560 
561 	/* vrele may need the vnode lock. */
562 	vrele(vp);
563 }
564 
565 /*
566  * We dropped a reference on an object and discovered that it had a
567  * single remaining shadow.  This is a sibling of the reference we
568  * dropped.  Attempt to collapse the sibling and backing object.
569  */
570 static vm_object_t
571 vm_object_deallocate_anon(vm_object_t backing_object)
572 {
573 	vm_object_t object;
574 
575 	/* Fetch the final shadow.  */
576 	object = LIST_FIRST(&backing_object->shadow_head);
577 	KASSERT(object != NULL &&
578 	    atomic_load_int(&backing_object->shadow_count) == 1,
579 	    ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
580 	    backing_object->ref_count,
581 	    atomic_load_int(&backing_object->shadow_count)));
582 	KASSERT((object->flags & OBJ_ANON) != 0,
583 	    ("invalid shadow object %p", object));
584 
585 	if (!VM_OBJECT_TRYWLOCK(object)) {
586 		/*
587 		 * Prevent object from disappearing since we do not have a
588 		 * reference.
589 		 */
590 		vm_object_pip_add(object, 1);
591 		VM_OBJECT_WUNLOCK(backing_object);
592 		VM_OBJECT_WLOCK(object);
593 		vm_object_pip_wakeup(object);
594 	} else
595 		VM_OBJECT_WUNLOCK(backing_object);
596 
597 	/*
598 	 * Check for a collapse/terminate race with the last reference holder.
599 	 */
600 	if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
601 	    !refcount_acquire_if_not_zero(&object->ref_count)) {
602 		VM_OBJECT_WUNLOCK(object);
603 		return (NULL);
604 	}
605 	backing_object = object->backing_object;
606 	if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
607 		vm_object_collapse(object);
608 	VM_OBJECT_WUNLOCK(object);
609 
610 	return (object);
611 }
612 
613 /*
614  *	vm_object_deallocate:
615  *
616  *	Release a reference to the specified object,
617  *	gained either through a vm_object_allocate
618  *	or a vm_object_reference call.  When all references
619  *	are gone, storage associated with this object
620  *	may be relinquished.
621  *
622  *	No object may be locked.
623  */
624 void
625 vm_object_deallocate(vm_object_t object)
626 {
627 	vm_object_t temp;
628 	bool released;
629 
630 	while (object != NULL) {
631 		/*
632 		 * If the reference count goes to 0 we start calling
633 		 * vm_object_terminate() on the object chain.  A ref count
634 		 * of 1 may be a special case depending on the shadow count
635 		 * being 0 or 1.  These cases require a write lock on the
636 		 * object.
637 		 */
638 		if ((object->flags & OBJ_ANON) == 0)
639 			released = refcount_release_if_gt(&object->ref_count, 1);
640 		else
641 			released = refcount_release_if_gt(&object->ref_count, 2);
642 		if (released)
643 			return;
644 
645 		if (object->type == OBJT_VNODE) {
646 			VM_OBJECT_RLOCK(object);
647 			if (object->type == OBJT_VNODE) {
648 				vm_object_deallocate_vnode(object);
649 				return;
650 			}
651 			VM_OBJECT_RUNLOCK(object);
652 		}
653 
654 		VM_OBJECT_WLOCK(object);
655 		KASSERT(object->ref_count > 0,
656 		    ("vm_object_deallocate: object deallocated too many times: %d",
657 		    object->type));
658 
659 		/*
660 		 * If this is not the final reference to an anonymous
661 		 * object we may need to collapse the shadow chain.
662 		 */
663 		if (!refcount_release(&object->ref_count)) {
664 			if (object->ref_count > 1 ||
665 			    atomic_load_int(&object->shadow_count) == 0) {
666 				if ((object->flags & OBJ_ANON) != 0 &&
667 				    object->ref_count == 1)
668 					vm_object_set_flag(object,
669 					    OBJ_ONEMAPPING);
670 				VM_OBJECT_WUNLOCK(object);
671 				return;
672 			}
673 
674 			/* Handle collapsing last ref on anonymous objects. */
675 			object = vm_object_deallocate_anon(object);
676 			continue;
677 		}
678 
679 		/*
680 		 * Handle the final reference to an object.  We restart
681 		 * the loop with the backing object to avoid recursion.
682 		 */
683 		umtx_shm_object_terminated(object);
684 		temp = object->backing_object;
685 		if (temp != NULL) {
686 			KASSERT(object->type == OBJT_SWAP,
687 			    ("shadowed tmpfs v_object 2 %p", object));
688 			vm_object_backing_remove(object);
689 		}
690 
691 		KASSERT((object->flags & OBJ_DEAD) == 0,
692 		    ("vm_object_deallocate: Terminating dead object."));
693 		vm_object_set_flag(object, OBJ_DEAD);
694 		vm_object_terminate(object);
695 		object = temp;
696 	}
697 }
698 
699 void
700 vm_object_destroy(vm_object_t object)
701 {
702 	uma_zfree(obj_zone, object);
703 }
704 
705 static void
706 vm_object_sub_shadow(vm_object_t object)
707 {
708 	KASSERT(object->shadow_count >= 1,
709 	    ("object %p sub_shadow count zero", object));
710 	atomic_subtract_int(&object->shadow_count, 1);
711 }
712 
713 static void
714 vm_object_backing_remove_locked(vm_object_t object)
715 {
716 	vm_object_t backing_object;
717 
718 	backing_object = object->backing_object;
719 	VM_OBJECT_ASSERT_WLOCKED(object);
720 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
721 
722 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
723 	    ("vm_object_backing_remove: Removing collapsing object."));
724 
725 	vm_object_sub_shadow(backing_object);
726 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
727 		LIST_REMOVE(object, shadow_list);
728 		vm_object_clear_flag(object, OBJ_SHADOWLIST);
729 	}
730 	object->backing_object = NULL;
731 }
732 
733 static void
734 vm_object_backing_remove(vm_object_t object)
735 {
736 	vm_object_t backing_object;
737 
738 	VM_OBJECT_ASSERT_WLOCKED(object);
739 
740 	backing_object = object->backing_object;
741 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
742 		VM_OBJECT_WLOCK(backing_object);
743 		vm_object_backing_remove_locked(object);
744 		VM_OBJECT_WUNLOCK(backing_object);
745 	} else {
746 		object->backing_object = NULL;
747 		vm_object_sub_shadow(backing_object);
748 	}
749 }
750 
751 static void
752 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
753 {
754 
755 	VM_OBJECT_ASSERT_WLOCKED(object);
756 
757 	atomic_add_int(&backing_object->shadow_count, 1);
758 	if ((backing_object->flags & OBJ_ANON) != 0) {
759 		VM_OBJECT_ASSERT_WLOCKED(backing_object);
760 		LIST_INSERT_HEAD(&backing_object->shadow_head, object,
761 		    shadow_list);
762 		vm_object_set_flag(object, OBJ_SHADOWLIST);
763 	}
764 	object->backing_object = backing_object;
765 }
766 
767 static void
768 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
769 {
770 
771 	VM_OBJECT_ASSERT_WLOCKED(object);
772 
773 	if ((backing_object->flags & OBJ_ANON) != 0) {
774 		VM_OBJECT_WLOCK(backing_object);
775 		vm_object_backing_insert_locked(object, backing_object);
776 		VM_OBJECT_WUNLOCK(backing_object);
777 	} else {
778 		object->backing_object = backing_object;
779 		atomic_add_int(&backing_object->shadow_count, 1);
780 	}
781 }
782 
783 /*
784  * Insert an object into a backing_object's shadow list with an additional
785  * reference to the backing_object added.
786  */
787 static void
788 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
789 {
790 
791 	VM_OBJECT_ASSERT_WLOCKED(object);
792 
793 	if ((backing_object->flags & OBJ_ANON) != 0) {
794 		VM_OBJECT_WLOCK(backing_object);
795 		KASSERT((backing_object->flags & OBJ_DEAD) == 0,
796 		    ("shadowing dead anonymous object"));
797 		vm_object_reference_locked(backing_object);
798 		vm_object_backing_insert_locked(object, backing_object);
799 		vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
800 		VM_OBJECT_WUNLOCK(backing_object);
801 	} else {
802 		vm_object_reference(backing_object);
803 		atomic_add_int(&backing_object->shadow_count, 1);
804 		object->backing_object = backing_object;
805 	}
806 }
807 
808 /*
809  * Transfer a backing reference from backing_object to object.
810  */
811 static void
812 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
813 {
814 	vm_object_t new_backing_object;
815 
816 	/*
817 	 * Note that the reference to backing_object->backing_object
818 	 * moves from within backing_object to within object.
819 	 */
820 	vm_object_backing_remove_locked(object);
821 	new_backing_object = backing_object->backing_object;
822 	if (new_backing_object == NULL)
823 		return;
824 	if ((new_backing_object->flags & OBJ_ANON) != 0) {
825 		VM_OBJECT_WLOCK(new_backing_object);
826 		vm_object_backing_remove_locked(backing_object);
827 		vm_object_backing_insert_locked(object, new_backing_object);
828 		VM_OBJECT_WUNLOCK(new_backing_object);
829 	} else {
830 		/*
831 		 * shadow_count for new_backing_object is left
832 		 * unchanged, its reference provided by backing_object
833 		 * is replaced by object.
834 		 */
835 		object->backing_object = new_backing_object;
836 		backing_object->backing_object = NULL;
837 	}
838 }
839 
840 /*
841  * Wait for a concurrent collapse to settle.
842  */
843 static void
844 vm_object_collapse_wait(vm_object_t object)
845 {
846 
847 	VM_OBJECT_ASSERT_WLOCKED(object);
848 
849 	while ((object->flags & OBJ_COLLAPSING) != 0) {
850 		vm_object_pip_wait(object, "vmcolwait");
851 		counter_u64_add(object_collapse_waits, 1);
852 	}
853 }
854 
855 /*
856  * Waits for a backing object to clear a pending collapse and returns
857  * it locked if it is an ANON object.
858  */
859 static vm_object_t
860 vm_object_backing_collapse_wait(vm_object_t object)
861 {
862 	vm_object_t backing_object;
863 
864 	VM_OBJECT_ASSERT_WLOCKED(object);
865 
866 	for (;;) {
867 		backing_object = object->backing_object;
868 		if (backing_object == NULL ||
869 		    (backing_object->flags & OBJ_ANON) == 0)
870 			return (NULL);
871 		VM_OBJECT_WLOCK(backing_object);
872 		if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
873 			break;
874 		VM_OBJECT_WUNLOCK(object);
875 		vm_object_pip_sleep(backing_object, "vmbckwait");
876 		counter_u64_add(object_collapse_waits, 1);
877 		VM_OBJECT_WLOCK(object);
878 	}
879 	return (backing_object);
880 }
881 
882 /*
883  *	vm_object_terminate_single_page removes a pageable page from the object,
884  *	and removes it from the paging queues and frees it, if it is not wired.
885  *	It is invoked via callback from vm_object_terminate_pages.
886  */
887 static void
888 vm_object_terminate_single_page(vm_page_t p, void *objectv)
889 {
890 	vm_object_t object __diagused = objectv;
891 
892 	vm_page_assert_unbusied(p);
893 	KASSERT(p->object == object &&
894 	    (p->ref_count & VPRC_OBJREF) != 0,
895 	    ("%s: page %p is inconsistent", __func__, p));
896 	p->object = NULL;
897 	if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
898 		VM_CNT_INC(v_pfree);
899 		vm_page_free(p);
900 	}
901 }
902 
903 /*
904  *	vm_object_terminate_pages removes any remaining pageable pages
905  *	from the object and resets the object to an empty state.
906  */
907 static void
908 vm_object_terminate_pages(vm_object_t object)
909 {
910 	VM_OBJECT_ASSERT_WLOCKED(object);
911 
912 	/*
913 	 * If the object contained any pages, then reset it to an empty state.
914 	 * Rather than incrementally removing each page from the object, the
915 	 * page and object are reset to any empty state.
916 	 */
917 	if (object->resident_page_count == 0)
918 		return;
919 
920 	vm_radix_reclaim_callback(&object->rtree,
921 	    vm_object_terminate_single_page, object);
922 	TAILQ_INIT(&object->memq);
923 	object->resident_page_count = 0;
924 	if (object->type == OBJT_VNODE)
925 		vdrop(object->handle);
926 }
927 
928 /*
929  *	vm_object_terminate actually destroys the specified object, freeing
930  *	up all previously used resources.
931  *
932  *	The object must be locked.
933  *	This routine may block.
934  */
935 void
936 vm_object_terminate(vm_object_t object)
937 {
938 
939 	VM_OBJECT_ASSERT_WLOCKED(object);
940 	KASSERT((object->flags & OBJ_DEAD) != 0,
941 	    ("terminating non-dead obj %p", object));
942 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
943 	    ("terminating collapsing obj %p", object));
944 	KASSERT(object->backing_object == NULL,
945 	    ("terminating shadow obj %p", object));
946 
947 	/*
948 	 * Wait for the pageout daemon and other current users to be
949 	 * done with the object.  Note that new paging_in_progress
950 	 * users can come after this wait, but they must check
951 	 * OBJ_DEAD flag set (without unlocking the object), and avoid
952 	 * the object being terminated.
953 	 */
954 	vm_object_pip_wait(object, "objtrm");
955 
956 	KASSERT(object->ref_count == 0,
957 	    ("vm_object_terminate: object with references, ref_count=%d",
958 	    object->ref_count));
959 
960 	if ((object->flags & OBJ_PG_DTOR) == 0)
961 		vm_object_terminate_pages(object);
962 
963 #if VM_NRESERVLEVEL > 0
964 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
965 		vm_reserv_break_all(object);
966 #endif
967 
968 	KASSERT(object->cred == NULL || (object->flags & OBJ_SWAP) != 0,
969 	    ("%s: non-swap obj %p has cred", __func__, object));
970 
971 	/*
972 	 * Let the pager know object is dead.
973 	 */
974 	vm_pager_deallocate(object);
975 	VM_OBJECT_WUNLOCK(object);
976 
977 	vm_object_destroy(object);
978 }
979 
980 /*
981  * Make the page read-only so that we can clear the object flags.  However, if
982  * this is a nosync mmap then the object is likely to stay dirty so do not
983  * mess with the page and do not clear the object flags.  Returns TRUE if the
984  * page should be flushed, and FALSE otherwise.
985  */
986 static boolean_t
987 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
988 {
989 
990 	vm_page_assert_busied(p);
991 
992 	/*
993 	 * If we have been asked to skip nosync pages and this is a
994 	 * nosync page, skip it.  Note that the object flags were not
995 	 * cleared in this case so we do not have to set them.
996 	 */
997 	if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
998 		*allclean = FALSE;
999 		return (FALSE);
1000 	} else {
1001 		pmap_remove_write(p);
1002 		return (p->dirty != 0);
1003 	}
1004 }
1005 
1006 /*
1007  *	vm_object_page_clean
1008  *
1009  *	Clean all dirty pages in the specified range of object.  Leaves page
1010  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
1011  *	write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1012  *	leaving the object dirty.
1013  *
1014  *	For swap objects backing tmpfs regular files, do not flush anything,
1015  *	but remove write protection on the mapped pages to update mtime through
1016  *	mmaped writes.
1017  *
1018  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
1019  *	synchronous clustering mode implementation.
1020  *
1021  *	Odd semantics: if start == end, we clean everything.
1022  *
1023  *	The object must be locked.
1024  *
1025  *	Returns FALSE if some page from the range was not written, as
1026  *	reported by the pager, and TRUE otherwise.
1027  */
1028 boolean_t
1029 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1030     int flags)
1031 {
1032 	vm_page_t np, p;
1033 	vm_pindex_t pi, tend, tstart;
1034 	int curgeneration, n, pagerflags;
1035 	boolean_t eio, res, allclean;
1036 
1037 	VM_OBJECT_ASSERT_WLOCKED(object);
1038 
1039 	if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1040 		return (TRUE);
1041 
1042 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1043 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1044 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1045 
1046 	tstart = OFF_TO_IDX(start);
1047 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1048 	allclean = tstart == 0 && tend >= object->size;
1049 	res = TRUE;
1050 
1051 rescan:
1052 	curgeneration = object->generation;
1053 
1054 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1055 		pi = p->pindex;
1056 		if (pi >= tend)
1057 			break;
1058 		np = TAILQ_NEXT(p, listq);
1059 		if (vm_page_none_valid(p))
1060 			continue;
1061 		if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1062 			if (object->generation != curgeneration &&
1063 			    (flags & OBJPC_SYNC) != 0)
1064 				goto rescan;
1065 			np = vm_page_find_least(object, pi);
1066 			continue;
1067 		}
1068 		if (!vm_object_page_remove_write(p, flags, &allclean)) {
1069 			vm_page_xunbusy(p);
1070 			continue;
1071 		}
1072 		if (object->type == OBJT_VNODE) {
1073 			n = vm_object_page_collect_flush(object, p, pagerflags,
1074 			    flags, &allclean, &eio);
1075 			if (eio) {
1076 				res = FALSE;
1077 				allclean = FALSE;
1078 			}
1079 			if (object->generation != curgeneration &&
1080 			    (flags & OBJPC_SYNC) != 0)
1081 				goto rescan;
1082 
1083 			/*
1084 			 * If the VOP_PUTPAGES() did a truncated write, so
1085 			 * that even the first page of the run is not fully
1086 			 * written, vm_pageout_flush() returns 0 as the run
1087 			 * length.  Since the condition that caused truncated
1088 			 * write may be permanent, e.g. exhausted free space,
1089 			 * accepting n == 0 would cause an infinite loop.
1090 			 *
1091 			 * Forwarding the iterator leaves the unwritten page
1092 			 * behind, but there is not much we can do there if
1093 			 * filesystem refuses to write it.
1094 			 */
1095 			if (n == 0) {
1096 				n = 1;
1097 				allclean = FALSE;
1098 			}
1099 		} else {
1100 			n = 1;
1101 			vm_page_xunbusy(p);
1102 		}
1103 		np = vm_page_find_least(object, pi + n);
1104 	}
1105 #if 0
1106 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1107 #endif
1108 
1109 	/*
1110 	 * Leave updating cleangeneration for tmpfs objects to tmpfs
1111 	 * scan.  It needs to update mtime, which happens for other
1112 	 * filesystems during page writeouts.
1113 	 */
1114 	if (allclean && object->type == OBJT_VNODE)
1115 		object->cleangeneration = curgeneration;
1116 	return (res);
1117 }
1118 
1119 static int
1120 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1121     int flags, boolean_t *allclean, boolean_t *eio)
1122 {
1123 	vm_page_t ma[2 * vm_pageout_page_count - 1], tp;
1124 	int base, count, runlen;
1125 
1126 	vm_page_lock_assert(p, MA_NOTOWNED);
1127 	vm_page_assert_xbusied(p);
1128 	VM_OBJECT_ASSERT_WLOCKED(object);
1129 	base = nitems(ma) / 2;
1130 	ma[base] = p;
1131 	for (count = 1, tp = p; count < vm_pageout_page_count; count++) {
1132 		tp = vm_page_next(tp);
1133 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1134 			break;
1135 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1136 			vm_page_xunbusy(tp);
1137 			break;
1138 		}
1139 		ma[base + count] = tp;
1140 	}
1141 
1142 	for (tp = p; count < vm_pageout_page_count; count++) {
1143 		tp = vm_page_prev(tp);
1144 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1145 			break;
1146 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1147 			vm_page_xunbusy(tp);
1148 			break;
1149 		}
1150 		ma[--base] = tp;
1151 	}
1152 
1153 	vm_pageout_flush(&ma[base], count, pagerflags, nitems(ma) / 2 - base,
1154 	    &runlen, eio);
1155 	return (runlen);
1156 }
1157 
1158 /*
1159  * Note that there is absolutely no sense in writing out
1160  * anonymous objects, so we track down the vnode object
1161  * to write out.
1162  * We invalidate (remove) all pages from the address space
1163  * for semantic correctness.
1164  *
1165  * If the backing object is a device object with unmanaged pages, then any
1166  * mappings to the specified range of pages must be removed before this
1167  * function is called.
1168  *
1169  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1170  * may start out with a NULL object.
1171  */
1172 boolean_t
1173 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1174     boolean_t syncio, boolean_t invalidate)
1175 {
1176 	vm_object_t backing_object;
1177 	struct vnode *vp;
1178 	struct mount *mp;
1179 	int error, flags, fsync_after;
1180 	boolean_t res;
1181 
1182 	if (object == NULL)
1183 		return (TRUE);
1184 	res = TRUE;
1185 	error = 0;
1186 	VM_OBJECT_WLOCK(object);
1187 	while ((backing_object = object->backing_object) != NULL) {
1188 		VM_OBJECT_WLOCK(backing_object);
1189 		offset += object->backing_object_offset;
1190 		VM_OBJECT_WUNLOCK(object);
1191 		object = backing_object;
1192 		if (object->size < OFF_TO_IDX(offset + size))
1193 			size = IDX_TO_OFF(object->size) - offset;
1194 	}
1195 	/*
1196 	 * Flush pages if writing is allowed, invalidate them
1197 	 * if invalidation requested.  Pages undergoing I/O
1198 	 * will be ignored by vm_object_page_remove().
1199 	 *
1200 	 * We cannot lock the vnode and then wait for paging
1201 	 * to complete without deadlocking against vm_fault.
1202 	 * Instead we simply call vm_object_page_remove() and
1203 	 * allow it to block internally on a page-by-page
1204 	 * basis when it encounters pages undergoing async
1205 	 * I/O.
1206 	 */
1207 	if (object->type == OBJT_VNODE &&
1208 	    vm_object_mightbedirty(object) != 0 &&
1209 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1210 		VM_OBJECT_WUNLOCK(object);
1211 		(void)vn_start_write(vp, &mp, V_WAIT);
1212 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1213 		if (syncio && !invalidate && offset == 0 &&
1214 		    atop(size) == object->size) {
1215 			/*
1216 			 * If syncing the whole mapping of the file,
1217 			 * it is faster to schedule all the writes in
1218 			 * async mode, also allowing the clustering,
1219 			 * and then wait for i/o to complete.
1220 			 */
1221 			flags = 0;
1222 			fsync_after = TRUE;
1223 		} else {
1224 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1225 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1226 			fsync_after = FALSE;
1227 		}
1228 		VM_OBJECT_WLOCK(object);
1229 		res = vm_object_page_clean(object, offset, offset + size,
1230 		    flags);
1231 		VM_OBJECT_WUNLOCK(object);
1232 		if (fsync_after) {
1233 			for (;;) {
1234 				error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1235 				if (error != ERELOOKUP)
1236 					break;
1237 
1238 				/*
1239 				 * Allow SU/bufdaemon to handle more
1240 				 * dependencies in the meantime.
1241 				 */
1242 				VOP_UNLOCK(vp);
1243 				vn_finished_write(mp);
1244 
1245 				(void)vn_start_write(vp, &mp, V_WAIT);
1246 				vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1247 			}
1248 		}
1249 		VOP_UNLOCK(vp);
1250 		vn_finished_write(mp);
1251 		if (error != 0)
1252 			res = FALSE;
1253 		VM_OBJECT_WLOCK(object);
1254 	}
1255 	if ((object->type == OBJT_VNODE ||
1256 	     object->type == OBJT_DEVICE) && invalidate) {
1257 		if (object->type == OBJT_DEVICE)
1258 			/*
1259 			 * The option OBJPR_NOTMAPPED must be passed here
1260 			 * because vm_object_page_remove() cannot remove
1261 			 * unmanaged mappings.
1262 			 */
1263 			flags = OBJPR_NOTMAPPED;
1264 		else if (old_msync)
1265 			flags = 0;
1266 		else
1267 			flags = OBJPR_CLEANONLY;
1268 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1269 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1270 	}
1271 	VM_OBJECT_WUNLOCK(object);
1272 	return (res);
1273 }
1274 
1275 /*
1276  * Determine whether the given advice can be applied to the object.  Advice is
1277  * not applied to unmanaged pages since they never belong to page queues, and
1278  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1279  * have been mapped at most once.
1280  */
1281 static bool
1282 vm_object_advice_applies(vm_object_t object, int advice)
1283 {
1284 
1285 	if ((object->flags & OBJ_UNMANAGED) != 0)
1286 		return (false);
1287 	if (advice != MADV_FREE)
1288 		return (true);
1289 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1290 	    (OBJ_ONEMAPPING | OBJ_ANON));
1291 }
1292 
1293 static void
1294 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1295     vm_size_t size)
1296 {
1297 
1298 	if (advice == MADV_FREE)
1299 		vm_pager_freespace(object, pindex, size);
1300 }
1301 
1302 /*
1303  *	vm_object_madvise:
1304  *
1305  *	Implements the madvise function at the object/page level.
1306  *
1307  *	MADV_WILLNEED	(any object)
1308  *
1309  *	    Activate the specified pages if they are resident.
1310  *
1311  *	MADV_DONTNEED	(any object)
1312  *
1313  *	    Deactivate the specified pages if they are resident.
1314  *
1315  *	MADV_FREE	(OBJT_SWAP objects, OBJ_ONEMAPPING only)
1316  *
1317  *	    Deactivate and clean the specified pages if they are
1318  *	    resident.  This permits the process to reuse the pages
1319  *	    without faulting or the kernel to reclaim the pages
1320  *	    without I/O.
1321  */
1322 void
1323 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1324     int advice)
1325 {
1326 	vm_pindex_t tpindex;
1327 	vm_object_t backing_object, tobject;
1328 	vm_page_t m, tm;
1329 
1330 	if (object == NULL)
1331 		return;
1332 
1333 relookup:
1334 	VM_OBJECT_WLOCK(object);
1335 	if (!vm_object_advice_applies(object, advice)) {
1336 		VM_OBJECT_WUNLOCK(object);
1337 		return;
1338 	}
1339 	for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1340 		tobject = object;
1341 
1342 		/*
1343 		 * If the next page isn't resident in the top-level object, we
1344 		 * need to search the shadow chain.  When applying MADV_FREE, we
1345 		 * take care to release any swap space used to store
1346 		 * non-resident pages.
1347 		 */
1348 		if (m == NULL || pindex < m->pindex) {
1349 			/*
1350 			 * Optimize a common case: if the top-level object has
1351 			 * no backing object, we can skip over the non-resident
1352 			 * range in constant time.
1353 			 */
1354 			if (object->backing_object == NULL) {
1355 				tpindex = (m != NULL && m->pindex < end) ?
1356 				    m->pindex : end;
1357 				vm_object_madvise_freespace(object, advice,
1358 				    pindex, tpindex - pindex);
1359 				if ((pindex = tpindex) == end)
1360 					break;
1361 				goto next_page;
1362 			}
1363 
1364 			tpindex = pindex;
1365 			do {
1366 				vm_object_madvise_freespace(tobject, advice,
1367 				    tpindex, 1);
1368 				/*
1369 				 * Prepare to search the next object in the
1370 				 * chain.
1371 				 */
1372 				backing_object = tobject->backing_object;
1373 				if (backing_object == NULL)
1374 					goto next_pindex;
1375 				VM_OBJECT_WLOCK(backing_object);
1376 				tpindex +=
1377 				    OFF_TO_IDX(tobject->backing_object_offset);
1378 				if (tobject != object)
1379 					VM_OBJECT_WUNLOCK(tobject);
1380 				tobject = backing_object;
1381 				if (!vm_object_advice_applies(tobject, advice))
1382 					goto next_pindex;
1383 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1384 			    NULL);
1385 		} else {
1386 next_page:
1387 			tm = m;
1388 			m = TAILQ_NEXT(m, listq);
1389 		}
1390 
1391 		/*
1392 		 * If the page is not in a normal state, skip it.  The page
1393 		 * can not be invalidated while the object lock is held.
1394 		 */
1395 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1396 			goto next_pindex;
1397 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1398 		    ("vm_object_madvise: page %p is fictitious", tm));
1399 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1400 		    ("vm_object_madvise: page %p is not managed", tm));
1401 		if (vm_page_tryxbusy(tm) == 0) {
1402 			if (object != tobject)
1403 				VM_OBJECT_WUNLOCK(object);
1404 			if (advice == MADV_WILLNEED) {
1405 				/*
1406 				 * Reference the page before unlocking and
1407 				 * sleeping so that the page daemon is less
1408 				 * likely to reclaim it.
1409 				 */
1410 				vm_page_aflag_set(tm, PGA_REFERENCED);
1411 			}
1412 			if (!vm_page_busy_sleep(tm, "madvpo", 0))
1413 				VM_OBJECT_WUNLOCK(tobject);
1414   			goto relookup;
1415 		}
1416 		vm_page_advise(tm, advice);
1417 		vm_page_xunbusy(tm);
1418 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1419 next_pindex:
1420 		if (tobject != object)
1421 			VM_OBJECT_WUNLOCK(tobject);
1422 	}
1423 	VM_OBJECT_WUNLOCK(object);
1424 }
1425 
1426 /*
1427  *	vm_object_shadow:
1428  *
1429  *	Create a new object which is backed by the
1430  *	specified existing object range.  The source
1431  *	object reference is deallocated.
1432  *
1433  *	The new object and offset into that object
1434  *	are returned in the source parameters.
1435  */
1436 void
1437 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1438     struct ucred *cred, bool shared)
1439 {
1440 	vm_object_t source;
1441 	vm_object_t result;
1442 
1443 	source = *object;
1444 
1445 	/*
1446 	 * Don't create the new object if the old object isn't shared.
1447 	 *
1448 	 * If we hold the only reference we can guarantee that it won't
1449 	 * increase while we have the map locked.  Otherwise the race is
1450 	 * harmless and we will end up with an extra shadow object that
1451 	 * will be collapsed later.
1452 	 */
1453 	if (source != NULL && source->ref_count == 1 &&
1454 	    (source->flags & OBJ_ANON) != 0)
1455 		return;
1456 
1457 	/*
1458 	 * Allocate a new object with the given length.
1459 	 */
1460 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1461 
1462 	/*
1463 	 * Store the offset into the source object, and fix up the offset into
1464 	 * the new object.
1465 	 */
1466 	result->backing_object_offset = *offset;
1467 
1468 	if (shared || source != NULL) {
1469 		VM_OBJECT_WLOCK(result);
1470 
1471 		/*
1472 		 * The new object shadows the source object, adding a
1473 		 * reference to it.  Our caller changes his reference
1474 		 * to point to the new object, removing a reference to
1475 		 * the source object.  Net result: no change of
1476 		 * reference count, unless the caller needs to add one
1477 		 * more reference due to forking a shared map entry.
1478 		 */
1479 		if (shared) {
1480 			vm_object_reference_locked(result);
1481 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1482 		}
1483 
1484 		/*
1485 		 * Try to optimize the result object's page color when
1486 		 * shadowing in order to maintain page coloring
1487 		 * consistency in the combined shadowed object.
1488 		 */
1489 		if (source != NULL) {
1490 			vm_object_backing_insert(result, source);
1491 			result->domain = source->domain;
1492 #if VM_NRESERVLEVEL > 0
1493 			vm_object_set_flag(result,
1494 			    (source->flags & OBJ_COLORED));
1495 			result->pg_color = (source->pg_color +
1496 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1497 			    1)) - 1);
1498 #endif
1499 		}
1500 		VM_OBJECT_WUNLOCK(result);
1501 	}
1502 
1503 	/*
1504 	 * Return the new things
1505 	 */
1506 	*offset = 0;
1507 	*object = result;
1508 }
1509 
1510 /*
1511  *	vm_object_split:
1512  *
1513  * Split the pages in a map entry into a new object.  This affords
1514  * easier removal of unused pages, and keeps object inheritance from
1515  * being a negative impact on memory usage.
1516  */
1517 void
1518 vm_object_split(vm_map_entry_t entry)
1519 {
1520 	vm_page_t m, m_next;
1521 	vm_object_t orig_object, new_object, backing_object;
1522 	vm_pindex_t idx, offidxstart;
1523 	vm_size_t size;
1524 
1525 	orig_object = entry->object.vm_object;
1526 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1527 	    ("vm_object_split:  Splitting object with multiple mappings."));
1528 	if ((orig_object->flags & OBJ_ANON) == 0)
1529 		return;
1530 	if (orig_object->ref_count <= 1)
1531 		return;
1532 	VM_OBJECT_WUNLOCK(orig_object);
1533 
1534 	offidxstart = OFF_TO_IDX(entry->offset);
1535 	size = atop(entry->end - entry->start);
1536 
1537 	new_object = vm_object_allocate_anon(size, orig_object,
1538 	    orig_object->cred, ptoa(size));
1539 
1540 	/*
1541 	 * We must wait for the orig_object to complete any in-progress
1542 	 * collapse so that the swap blocks are stable below.  The
1543 	 * additional reference on backing_object by new object will
1544 	 * prevent further collapse operations until split completes.
1545 	 */
1546 	VM_OBJECT_WLOCK(orig_object);
1547 	vm_object_collapse_wait(orig_object);
1548 
1549 	/*
1550 	 * At this point, the new object is still private, so the order in
1551 	 * which the original and new objects are locked does not matter.
1552 	 */
1553 	VM_OBJECT_WLOCK(new_object);
1554 	new_object->domain = orig_object->domain;
1555 	backing_object = orig_object->backing_object;
1556 	if (backing_object != NULL) {
1557 		vm_object_backing_insert_ref(new_object, backing_object);
1558 		new_object->backing_object_offset =
1559 		    orig_object->backing_object_offset + entry->offset;
1560 	}
1561 	if (orig_object->cred != NULL) {
1562 		crhold(orig_object->cred);
1563 		KASSERT(orig_object->charge >= ptoa(size),
1564 		    ("orig_object->charge < 0"));
1565 		orig_object->charge -= ptoa(size);
1566 	}
1567 
1568 	/*
1569 	 * Mark the split operation so that swap_pager_getpages() knows
1570 	 * that the object is in transition.
1571 	 */
1572 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1573 #ifdef INVARIANTS
1574 	idx = 0;
1575 #endif
1576 retry:
1577 	m = vm_page_find_least(orig_object, offidxstart);
1578 	KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1579 	    ("%s: object %p was repopulated", __func__, orig_object));
1580 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1581 	    m = m_next) {
1582 		m_next = TAILQ_NEXT(m, listq);
1583 
1584 		/*
1585 		 * We must wait for pending I/O to complete before we can
1586 		 * rename the page.
1587 		 *
1588 		 * We do not have to VM_PROT_NONE the page as mappings should
1589 		 * not be changed by this operation.
1590 		 */
1591 		if (vm_page_tryxbusy(m) == 0) {
1592 			VM_OBJECT_WUNLOCK(new_object);
1593 			if (vm_page_busy_sleep(m, "spltwt", 0))
1594 				VM_OBJECT_WLOCK(orig_object);
1595 			VM_OBJECT_WLOCK(new_object);
1596 			goto retry;
1597 		}
1598 
1599 		/*
1600 		 * The page was left invalid.  Likely placed there by
1601 		 * an incomplete fault.  Just remove and ignore.
1602 		 */
1603 		if (vm_page_none_valid(m)) {
1604 			if (vm_page_remove(m))
1605 				vm_page_free(m);
1606 			continue;
1607 		}
1608 
1609 		/* vm_page_rename() will dirty the page. */
1610 		if (vm_page_rename(m, new_object, idx)) {
1611 			vm_page_xunbusy(m);
1612 			VM_OBJECT_WUNLOCK(new_object);
1613 			VM_OBJECT_WUNLOCK(orig_object);
1614 			vm_radix_wait();
1615 			VM_OBJECT_WLOCK(orig_object);
1616 			VM_OBJECT_WLOCK(new_object);
1617 			goto retry;
1618 		}
1619 
1620 #if VM_NRESERVLEVEL > 0
1621 		/*
1622 		 * If some of the reservation's allocated pages remain with
1623 		 * the original object, then transferring the reservation to
1624 		 * the new object is neither particularly beneficial nor
1625 		 * particularly harmful as compared to leaving the reservation
1626 		 * with the original object.  If, however, all of the
1627 		 * reservation's allocated pages are transferred to the new
1628 		 * object, then transferring the reservation is typically
1629 		 * beneficial.  Determining which of these two cases applies
1630 		 * would be more costly than unconditionally renaming the
1631 		 * reservation.
1632 		 */
1633 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1634 #endif
1635 	}
1636 
1637 	/*
1638 	 * swap_pager_copy() can sleep, in which case the orig_object's
1639 	 * and new_object's locks are released and reacquired.
1640 	 */
1641 	swap_pager_copy(orig_object, new_object, offidxstart, 0);
1642 
1643 	TAILQ_FOREACH(m, &new_object->memq, listq)
1644 		vm_page_xunbusy(m);
1645 
1646 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1647 	VM_OBJECT_WUNLOCK(orig_object);
1648 	VM_OBJECT_WUNLOCK(new_object);
1649 	entry->object.vm_object = new_object;
1650 	entry->offset = 0LL;
1651 	vm_object_deallocate(orig_object);
1652 	VM_OBJECT_WLOCK(new_object);
1653 }
1654 
1655 static vm_page_t
1656 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1657 {
1658 	vm_object_t backing_object;
1659 
1660 	VM_OBJECT_ASSERT_WLOCKED(object);
1661 	backing_object = object->backing_object;
1662 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1663 
1664 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1665 	    ("invalid ownership %p %p %p", p, object, backing_object));
1666 	/* The page is only NULL when rename fails. */
1667 	if (p == NULL) {
1668 		VM_OBJECT_WUNLOCK(object);
1669 		VM_OBJECT_WUNLOCK(backing_object);
1670 		vm_radix_wait();
1671 		VM_OBJECT_WLOCK(object);
1672 	} else if (p->object == object) {
1673 		VM_OBJECT_WUNLOCK(backing_object);
1674 		if (vm_page_busy_sleep(p, "vmocol", 0))
1675 			VM_OBJECT_WLOCK(object);
1676 	} else {
1677 		VM_OBJECT_WUNLOCK(object);
1678 		if (!vm_page_busy_sleep(p, "vmocol", 0))
1679 			VM_OBJECT_WUNLOCK(backing_object);
1680 		VM_OBJECT_WLOCK(object);
1681 	}
1682 	VM_OBJECT_WLOCK(backing_object);
1683 	return (TAILQ_FIRST(&backing_object->memq));
1684 }
1685 
1686 static bool
1687 vm_object_scan_all_shadowed(vm_object_t object)
1688 {
1689 	vm_object_t backing_object;
1690 	vm_page_t p, pp;
1691 	vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1692 
1693 	VM_OBJECT_ASSERT_WLOCKED(object);
1694 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1695 
1696 	backing_object = object->backing_object;
1697 
1698 	if ((backing_object->flags & OBJ_ANON) == 0)
1699 		return (false);
1700 
1701 	pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1702 	p = vm_page_find_least(backing_object, pi);
1703 	ps = swap_pager_find_least(backing_object, pi);
1704 
1705 	/*
1706 	 * Only check pages inside the parent object's range and
1707 	 * inside the parent object's mapping of the backing object.
1708 	 */
1709 	for (;; pi++) {
1710 		if (p != NULL && p->pindex < pi)
1711 			p = TAILQ_NEXT(p, listq);
1712 		if (ps < pi)
1713 			ps = swap_pager_find_least(backing_object, pi);
1714 		if (p == NULL && ps >= backing_object->size)
1715 			break;
1716 		else if (p == NULL)
1717 			pi = ps;
1718 		else
1719 			pi = MIN(p->pindex, ps);
1720 
1721 		new_pindex = pi - backing_offset_index;
1722 		if (new_pindex >= object->size)
1723 			break;
1724 
1725 		if (p != NULL) {
1726 			/*
1727 			 * If the backing object page is busy a
1728 			 * grandparent or older page may still be
1729 			 * undergoing CoW.  It is not safe to collapse
1730 			 * the backing object until it is quiesced.
1731 			 */
1732 			if (vm_page_tryxbusy(p) == 0)
1733 				return (false);
1734 
1735 			/*
1736 			 * We raced with the fault handler that left
1737 			 * newly allocated invalid page on the object
1738 			 * queue and retried.
1739 			 */
1740 			if (!vm_page_all_valid(p))
1741 				goto unbusy_ret;
1742 		}
1743 
1744 		/*
1745 		 * See if the parent has the page or if the parent's object
1746 		 * pager has the page.  If the parent has the page but the page
1747 		 * is not valid, the parent's object pager must have the page.
1748 		 *
1749 		 * If this fails, the parent does not completely shadow the
1750 		 * object and we might as well give up now.
1751 		 */
1752 		pp = vm_page_lookup(object, new_pindex);
1753 
1754 		/*
1755 		 * The valid check here is stable due to object lock
1756 		 * being required to clear valid and initiate paging.
1757 		 * Busy of p disallows fault handler to validate pp.
1758 		 */
1759 		if ((pp == NULL || vm_page_none_valid(pp)) &&
1760 		    !vm_pager_has_page(object, new_pindex, NULL, NULL))
1761 			goto unbusy_ret;
1762 		if (p != NULL)
1763 			vm_page_xunbusy(p);
1764 	}
1765 	return (true);
1766 
1767 unbusy_ret:
1768 	if (p != NULL)
1769 		vm_page_xunbusy(p);
1770 	return (false);
1771 }
1772 
1773 static void
1774 vm_object_collapse_scan(vm_object_t object)
1775 {
1776 	vm_object_t backing_object;
1777 	vm_page_t next, p, pp;
1778 	vm_pindex_t backing_offset_index, new_pindex;
1779 
1780 	VM_OBJECT_ASSERT_WLOCKED(object);
1781 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1782 
1783 	backing_object = object->backing_object;
1784 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1785 
1786 	/*
1787 	 * Our scan
1788 	 */
1789 	for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1790 		next = TAILQ_NEXT(p, listq);
1791 		new_pindex = p->pindex - backing_offset_index;
1792 
1793 		/*
1794 		 * Check for busy page
1795 		 */
1796 		if (vm_page_tryxbusy(p) == 0) {
1797 			next = vm_object_collapse_scan_wait(object, p);
1798 			continue;
1799 		}
1800 
1801 		KASSERT(object->backing_object == backing_object,
1802 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1803 		    object->backing_object, backing_object));
1804 		KASSERT(p->object == backing_object,
1805 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1806 		    p->object, backing_object));
1807 
1808 		if (p->pindex < backing_offset_index ||
1809 		    new_pindex >= object->size) {
1810 			vm_pager_freespace(backing_object, p->pindex, 1);
1811 
1812 			KASSERT(!pmap_page_is_mapped(p),
1813 			    ("freeing mapped page %p", p));
1814 			if (vm_page_remove(p))
1815 				vm_page_free(p);
1816 			continue;
1817 		}
1818 
1819 		if (!vm_page_all_valid(p)) {
1820 			KASSERT(!pmap_page_is_mapped(p),
1821 			    ("freeing mapped page %p", p));
1822 			if (vm_page_remove(p))
1823 				vm_page_free(p);
1824 			continue;
1825 		}
1826 
1827 		pp = vm_page_lookup(object, new_pindex);
1828 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1829 			vm_page_xunbusy(p);
1830 			/*
1831 			 * The page in the parent is busy and possibly not
1832 			 * (yet) valid.  Until its state is finalized by the
1833 			 * busy bit owner, we can't tell whether it shadows the
1834 			 * original page.
1835 			 */
1836 			next = vm_object_collapse_scan_wait(object, pp);
1837 			continue;
1838 		}
1839 
1840 		if (pp != NULL && vm_page_none_valid(pp)) {
1841 			/*
1842 			 * The page was invalid in the parent.  Likely placed
1843 			 * there by an incomplete fault.  Just remove and
1844 			 * ignore.  p can replace it.
1845 			 */
1846 			if (vm_page_remove(pp))
1847 				vm_page_free(pp);
1848 			pp = NULL;
1849 		}
1850 
1851 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1852 			NULL)) {
1853 			/*
1854 			 * The page already exists in the parent OR swap exists
1855 			 * for this location in the parent.  Leave the parent's
1856 			 * page alone.  Destroy the original page from the
1857 			 * backing object.
1858 			 */
1859 			vm_pager_freespace(backing_object, p->pindex, 1);
1860 			KASSERT(!pmap_page_is_mapped(p),
1861 			    ("freeing mapped page %p", p));
1862 			if (vm_page_remove(p))
1863 				vm_page_free(p);
1864 			if (pp != NULL)
1865 				vm_page_xunbusy(pp);
1866 			continue;
1867 		}
1868 
1869 		/*
1870 		 * Page does not exist in parent, rename the page from the
1871 		 * backing object to the main object.
1872 		 *
1873 		 * If the page was mapped to a process, it can remain mapped
1874 		 * through the rename.  vm_page_rename() will dirty the page.
1875 		 */
1876 		if (vm_page_rename(p, object, new_pindex)) {
1877 			vm_page_xunbusy(p);
1878 			next = vm_object_collapse_scan_wait(object, NULL);
1879 			continue;
1880 		}
1881 
1882 		/* Use the old pindex to free the right page. */
1883 		vm_pager_freespace(backing_object, new_pindex +
1884 		    backing_offset_index, 1);
1885 
1886 #if VM_NRESERVLEVEL > 0
1887 		/*
1888 		 * Rename the reservation.
1889 		 */
1890 		vm_reserv_rename(p, object, backing_object,
1891 		    backing_offset_index);
1892 #endif
1893 		vm_page_xunbusy(p);
1894 	}
1895 	return;
1896 }
1897 
1898 /*
1899  *	vm_object_collapse:
1900  *
1901  *	Collapse an object with the object backing it.
1902  *	Pages in the backing object are moved into the
1903  *	parent, and the backing object is deallocated.
1904  */
1905 void
1906 vm_object_collapse(vm_object_t object)
1907 {
1908 	vm_object_t backing_object, new_backing_object;
1909 
1910 	VM_OBJECT_ASSERT_WLOCKED(object);
1911 
1912 	while (TRUE) {
1913 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1914 		    ("collapsing invalid object"));
1915 
1916 		/*
1917 		 * Wait for the backing_object to finish any pending
1918 		 * collapse so that the caller sees the shortest possible
1919 		 * shadow chain.
1920 		 */
1921 		backing_object = vm_object_backing_collapse_wait(object);
1922 		if (backing_object == NULL)
1923 			return;
1924 
1925 		KASSERT(object->ref_count > 0 &&
1926 		    object->ref_count > atomic_load_int(&object->shadow_count),
1927 		    ("collapse with invalid ref %d or shadow %d count.",
1928 		    object->ref_count, atomic_load_int(&object->shadow_count)));
1929 		KASSERT((backing_object->flags &
1930 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1931 		    ("vm_object_collapse: Backing object already collapsing."));
1932 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1933 		    ("vm_object_collapse: object is already collapsing."));
1934 
1935 		/*
1936 		 * We know that we can either collapse the backing object if
1937 		 * the parent is the only reference to it, or (perhaps) have
1938 		 * the parent bypass the object if the parent happens to shadow
1939 		 * all the resident pages in the entire backing object.
1940 		 */
1941 		if (backing_object->ref_count == 1) {
1942 			KASSERT(atomic_load_int(&backing_object->shadow_count)
1943 			    == 1,
1944 			    ("vm_object_collapse: shadow_count: %d",
1945 			    atomic_load_int(&backing_object->shadow_count)));
1946 			vm_object_pip_add(object, 1);
1947 			vm_object_set_flag(object, OBJ_COLLAPSING);
1948 			vm_object_pip_add(backing_object, 1);
1949 			vm_object_set_flag(backing_object, OBJ_DEAD);
1950 
1951 			/*
1952 			 * If there is exactly one reference to the backing
1953 			 * object, we can collapse it into the parent.
1954 			 */
1955 			vm_object_collapse_scan(object);
1956 
1957 			/*
1958 			 * Move the pager from backing_object to object.
1959 			 *
1960 			 * swap_pager_copy() can sleep, in which case the
1961 			 * backing_object's and object's locks are released and
1962 			 * reacquired.
1963 			 */
1964 			swap_pager_copy(backing_object, object,
1965 			    OFF_TO_IDX(object->backing_object_offset), TRUE);
1966 
1967 			/*
1968 			 * Object now shadows whatever backing_object did.
1969 			 */
1970 			vm_object_clear_flag(object, OBJ_COLLAPSING);
1971 			vm_object_backing_transfer(object, backing_object);
1972 			object->backing_object_offset +=
1973 			    backing_object->backing_object_offset;
1974 			VM_OBJECT_WUNLOCK(object);
1975 			vm_object_pip_wakeup(object);
1976 
1977 			/*
1978 			 * Discard backing_object.
1979 			 *
1980 			 * Since the backing object has no pages, no pager left,
1981 			 * and no object references within it, all that is
1982 			 * necessary is to dispose of it.
1983 			 */
1984 			KASSERT(backing_object->ref_count == 1, (
1985 "backing_object %p was somehow re-referenced during collapse!",
1986 			    backing_object));
1987 			vm_object_pip_wakeup(backing_object);
1988 			(void)refcount_release(&backing_object->ref_count);
1989 			umtx_shm_object_terminated(backing_object);
1990 			vm_object_terminate(backing_object);
1991 			counter_u64_add(object_collapses, 1);
1992 			VM_OBJECT_WLOCK(object);
1993 		} else {
1994 			/*
1995 			 * If we do not entirely shadow the backing object,
1996 			 * there is nothing we can do so we give up.
1997 			 *
1998 			 * The object lock and backing_object lock must not
1999 			 * be dropped during this sequence.
2000 			 */
2001 			if (!vm_object_scan_all_shadowed(object)) {
2002 				VM_OBJECT_WUNLOCK(backing_object);
2003 				break;
2004 			}
2005 
2006 			/*
2007 			 * Make the parent shadow the next object in the
2008 			 * chain.  Deallocating backing_object will not remove
2009 			 * it, since its reference count is at least 2.
2010 			 */
2011 			vm_object_backing_remove_locked(object);
2012 			new_backing_object = backing_object->backing_object;
2013 			if (new_backing_object != NULL) {
2014 				vm_object_backing_insert_ref(object,
2015 				    new_backing_object);
2016 				object->backing_object_offset +=
2017 				    backing_object->backing_object_offset;
2018 			}
2019 
2020 			/*
2021 			 * Drop the reference count on backing_object. Since
2022 			 * its ref_count was at least 2, it will not vanish.
2023 			 */
2024 			(void)refcount_release(&backing_object->ref_count);
2025 			KASSERT(backing_object->ref_count >= 1, (
2026 "backing_object %p was somehow dereferenced during collapse!",
2027 			    backing_object));
2028 			VM_OBJECT_WUNLOCK(backing_object);
2029 			counter_u64_add(object_bypasses, 1);
2030 		}
2031 
2032 		/*
2033 		 * Try again with this object's new backing object.
2034 		 */
2035 	}
2036 }
2037 
2038 /*
2039  *	vm_object_page_remove:
2040  *
2041  *	For the given object, either frees or invalidates each of the
2042  *	specified pages.  In general, a page is freed.  However, if a page is
2043  *	wired for any reason other than the existence of a managed, wired
2044  *	mapping, then it may be invalidated but not removed from the object.
2045  *	Pages are specified by the given range ["start", "end") and the option
2046  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
2047  *	extends from "start" to the end of the object.  If the option
2048  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2049  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
2050  *	specified, then the pages within the specified range must have no
2051  *	mappings.  Otherwise, if this option is not specified, any mappings to
2052  *	the specified pages are removed before the pages are freed or
2053  *	invalidated.
2054  *
2055  *	In general, this operation should only be performed on objects that
2056  *	contain managed pages.  There are, however, two exceptions.  First, it
2057  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
2058  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2059  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
2060  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
2061  *
2062  *	The object must be locked.
2063  */
2064 void
2065 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2066     int options)
2067 {
2068 	vm_page_t p, next;
2069 
2070 	VM_OBJECT_ASSERT_WLOCKED(object);
2071 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2072 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2073 	    ("vm_object_page_remove: illegal options for object %p", object));
2074 	if (object->resident_page_count == 0)
2075 		return;
2076 	vm_object_pip_add(object, 1);
2077 again:
2078 	p = vm_page_find_least(object, start);
2079 
2080 	/*
2081 	 * Here, the variable "p" is either (1) the page with the least pindex
2082 	 * greater than or equal to the parameter "start" or (2) NULL.
2083 	 */
2084 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2085 		next = TAILQ_NEXT(p, listq);
2086 
2087 		/*
2088 		 * Skip invalid pages if asked to do so.  Try to avoid acquiring
2089 		 * the busy lock, as some consumers rely on this to avoid
2090 		 * deadlocks.
2091 		 *
2092 		 * A thread may concurrently transition the page from invalid to
2093 		 * valid using only the busy lock, so the result of this check
2094 		 * is immediately stale.  It is up to consumers to handle this,
2095 		 * for instance by ensuring that all invalid->valid transitions
2096 		 * happen with a mutex held, as may be possible for a
2097 		 * filesystem.
2098 		 */
2099 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2100 			continue;
2101 
2102 		/*
2103 		 * If the page is wired for any reason besides the existence
2104 		 * of managed, wired mappings, then it cannot be freed.  For
2105 		 * example, fictitious pages, which represent device memory,
2106 		 * are inherently wired and cannot be freed.  They can,
2107 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2108 		 * not specified.
2109 		 */
2110 		if (vm_page_tryxbusy(p) == 0) {
2111 			if (vm_page_busy_sleep(p, "vmopar", 0))
2112 				VM_OBJECT_WLOCK(object);
2113 			goto again;
2114 		}
2115 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2116 			vm_page_xunbusy(p);
2117 			continue;
2118 		}
2119 		if (vm_page_wired(p)) {
2120 wired:
2121 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2122 			    object->ref_count != 0)
2123 				pmap_remove_all(p);
2124 			if ((options & OBJPR_CLEANONLY) == 0) {
2125 				vm_page_invalid(p);
2126 				vm_page_undirty(p);
2127 			}
2128 			vm_page_xunbusy(p);
2129 			continue;
2130 		}
2131 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2132 		    ("vm_object_page_remove: page %p is fictitious", p));
2133 		if ((options & OBJPR_CLEANONLY) != 0 &&
2134 		    !vm_page_none_valid(p)) {
2135 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2136 			    object->ref_count != 0 &&
2137 			    !vm_page_try_remove_write(p))
2138 				goto wired;
2139 			if (p->dirty != 0) {
2140 				vm_page_xunbusy(p);
2141 				continue;
2142 			}
2143 		}
2144 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2145 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2146 			goto wired;
2147 		vm_page_free(p);
2148 	}
2149 	vm_object_pip_wakeup(object);
2150 
2151 	vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2152 	    start);
2153 }
2154 
2155 /*
2156  *	vm_object_page_noreuse:
2157  *
2158  *	For the given object, attempt to move the specified pages to
2159  *	the head of the inactive queue.  This bypasses regular LRU
2160  *	operation and allows the pages to be reused quickly under memory
2161  *	pressure.  If a page is wired for any reason, then it will not
2162  *	be queued.  Pages are specified by the range ["start", "end").
2163  *	As a special case, if "end" is zero, then the range extends from
2164  *	"start" to the end of the object.
2165  *
2166  *	This operation should only be performed on objects that
2167  *	contain non-fictitious, managed pages.
2168  *
2169  *	The object must be locked.
2170  */
2171 void
2172 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2173 {
2174 	vm_page_t p, next;
2175 
2176 	VM_OBJECT_ASSERT_LOCKED(object);
2177 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2178 	    ("vm_object_page_noreuse: illegal object %p", object));
2179 	if (object->resident_page_count == 0)
2180 		return;
2181 	p = vm_page_find_least(object, start);
2182 
2183 	/*
2184 	 * Here, the variable "p" is either (1) the page with the least pindex
2185 	 * greater than or equal to the parameter "start" or (2) NULL.
2186 	 */
2187 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2188 		next = TAILQ_NEXT(p, listq);
2189 		vm_page_deactivate_noreuse(p);
2190 	}
2191 }
2192 
2193 /*
2194  *	Populate the specified range of the object with valid pages.  Returns
2195  *	TRUE if the range is successfully populated and FALSE otherwise.
2196  *
2197  *	Note: This function should be optimized to pass a larger array of
2198  *	pages to vm_pager_get_pages() before it is applied to a non-
2199  *	OBJT_DEVICE object.
2200  *
2201  *	The object must be locked.
2202  */
2203 boolean_t
2204 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2205 {
2206 	vm_page_t m;
2207 	vm_pindex_t pindex;
2208 	int rv;
2209 
2210 	VM_OBJECT_ASSERT_WLOCKED(object);
2211 	for (pindex = start; pindex < end; pindex++) {
2212 		rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2213 		if (rv != VM_PAGER_OK)
2214 			break;
2215 
2216 		/*
2217 		 * Keep "m" busy because a subsequent iteration may unlock
2218 		 * the object.
2219 		 */
2220 	}
2221 	if (pindex > start) {
2222 		m = vm_page_lookup(object, start);
2223 		while (m != NULL && m->pindex < pindex) {
2224 			vm_page_xunbusy(m);
2225 			m = TAILQ_NEXT(m, listq);
2226 		}
2227 	}
2228 	return (pindex == end);
2229 }
2230 
2231 /*
2232  *	Routine:	vm_object_coalesce
2233  *	Function:	Coalesces two objects backing up adjoining
2234  *			regions of memory into a single object.
2235  *
2236  *	returns TRUE if objects were combined.
2237  *
2238  *	NOTE:	Only works at the moment if the second object is NULL -
2239  *		if it's not, which object do we lock first?
2240  *
2241  *	Parameters:
2242  *		prev_object	First object to coalesce
2243  *		prev_offset	Offset into prev_object
2244  *		prev_size	Size of reference to prev_object
2245  *		next_size	Size of reference to the second object
2246  *		reserved	Indicator that extension region has
2247  *				swap accounted for
2248  *
2249  *	Conditions:
2250  *	The object must *not* be locked.
2251  */
2252 boolean_t
2253 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2254     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2255 {
2256 	vm_pindex_t next_pindex;
2257 
2258 	if (prev_object == NULL)
2259 		return (TRUE);
2260 	if ((prev_object->flags & OBJ_ANON) == 0)
2261 		return (FALSE);
2262 
2263 	VM_OBJECT_WLOCK(prev_object);
2264 	/*
2265 	 * Try to collapse the object first.
2266 	 */
2267 	vm_object_collapse(prev_object);
2268 
2269 	/*
2270 	 * Can't coalesce if: . more than one reference . paged out . shadows
2271 	 * another object . has a copy elsewhere (any of which mean that the
2272 	 * pages not mapped to prev_entry may be in use anyway)
2273 	 */
2274 	if (prev_object->backing_object != NULL) {
2275 		VM_OBJECT_WUNLOCK(prev_object);
2276 		return (FALSE);
2277 	}
2278 
2279 	prev_size >>= PAGE_SHIFT;
2280 	next_size >>= PAGE_SHIFT;
2281 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2282 
2283 	if (prev_object->ref_count > 1 &&
2284 	    prev_object->size != next_pindex &&
2285 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2286 		VM_OBJECT_WUNLOCK(prev_object);
2287 		return (FALSE);
2288 	}
2289 
2290 	/*
2291 	 * Account for the charge.
2292 	 */
2293 	if (prev_object->cred != NULL) {
2294 		/*
2295 		 * If prev_object was charged, then this mapping,
2296 		 * although not charged now, may become writable
2297 		 * later. Non-NULL cred in the object would prevent
2298 		 * swap reservation during enabling of the write
2299 		 * access, so reserve swap now. Failed reservation
2300 		 * cause allocation of the separate object for the map
2301 		 * entry, and swap reservation for this entry is
2302 		 * managed in appropriate time.
2303 		 */
2304 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2305 		    prev_object->cred)) {
2306 			VM_OBJECT_WUNLOCK(prev_object);
2307 			return (FALSE);
2308 		}
2309 		prev_object->charge += ptoa(next_size);
2310 	}
2311 
2312 	/*
2313 	 * Remove any pages that may still be in the object from a previous
2314 	 * deallocation.
2315 	 */
2316 	if (next_pindex < prev_object->size) {
2317 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2318 		    next_size, 0);
2319 #if 0
2320 		if (prev_object->cred != NULL) {
2321 			KASSERT(prev_object->charge >=
2322 			    ptoa(prev_object->size - next_pindex),
2323 			    ("object %p overcharged 1 %jx %jx", prev_object,
2324 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2325 			prev_object->charge -= ptoa(prev_object->size -
2326 			    next_pindex);
2327 		}
2328 #endif
2329 	}
2330 
2331 	/*
2332 	 * Extend the object if necessary.
2333 	 */
2334 	if (next_pindex + next_size > prev_object->size)
2335 		prev_object->size = next_pindex + next_size;
2336 
2337 	VM_OBJECT_WUNLOCK(prev_object);
2338 	return (TRUE);
2339 }
2340 
2341 void
2342 vm_object_set_writeable_dirty_(vm_object_t object)
2343 {
2344 	atomic_add_int(&object->generation, 1);
2345 }
2346 
2347 bool
2348 vm_object_mightbedirty_(vm_object_t object)
2349 {
2350 	return (object->generation != object->cleangeneration);
2351 }
2352 
2353 /*
2354  *	vm_object_unwire:
2355  *
2356  *	For each page offset within the specified range of the given object,
2357  *	find the highest-level page in the shadow chain and unwire it.  A page
2358  *	must exist at every page offset, and the highest-level page must be
2359  *	wired.
2360  */
2361 void
2362 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2363     uint8_t queue)
2364 {
2365 	vm_object_t tobject, t1object;
2366 	vm_page_t m, tm;
2367 	vm_pindex_t end_pindex, pindex, tpindex;
2368 	int depth, locked_depth;
2369 
2370 	KASSERT((offset & PAGE_MASK) == 0,
2371 	    ("vm_object_unwire: offset is not page aligned"));
2372 	KASSERT((length & PAGE_MASK) == 0,
2373 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2374 	/* The wired count of a fictitious page never changes. */
2375 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2376 		return;
2377 	pindex = OFF_TO_IDX(offset);
2378 	end_pindex = pindex + atop(length);
2379 again:
2380 	locked_depth = 1;
2381 	VM_OBJECT_RLOCK(object);
2382 	m = vm_page_find_least(object, pindex);
2383 	while (pindex < end_pindex) {
2384 		if (m == NULL || pindex < m->pindex) {
2385 			/*
2386 			 * The first object in the shadow chain doesn't
2387 			 * contain a page at the current index.  Therefore,
2388 			 * the page must exist in a backing object.
2389 			 */
2390 			tobject = object;
2391 			tpindex = pindex;
2392 			depth = 0;
2393 			do {
2394 				tpindex +=
2395 				    OFF_TO_IDX(tobject->backing_object_offset);
2396 				tobject = tobject->backing_object;
2397 				KASSERT(tobject != NULL,
2398 				    ("vm_object_unwire: missing page"));
2399 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2400 					goto next_page;
2401 				depth++;
2402 				if (depth == locked_depth) {
2403 					locked_depth++;
2404 					VM_OBJECT_RLOCK(tobject);
2405 				}
2406 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2407 			    NULL);
2408 		} else {
2409 			tm = m;
2410 			m = TAILQ_NEXT(m, listq);
2411 		}
2412 		if (vm_page_trysbusy(tm) == 0) {
2413 			for (tobject = object; locked_depth >= 1;
2414 			    locked_depth--) {
2415 				t1object = tobject->backing_object;
2416 				if (tm->object != tobject)
2417 					VM_OBJECT_RUNLOCK(tobject);
2418 				tobject = t1object;
2419 			}
2420 			tobject = tm->object;
2421 			if (!vm_page_busy_sleep(tm, "unwbo",
2422 			    VM_ALLOC_IGN_SBUSY))
2423 				VM_OBJECT_RUNLOCK(tobject);
2424 			goto again;
2425 		}
2426 		vm_page_unwire(tm, queue);
2427 		vm_page_sunbusy(tm);
2428 next_page:
2429 		pindex++;
2430 	}
2431 	/* Release the accumulated object locks. */
2432 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2433 		t1object = tobject->backing_object;
2434 		VM_OBJECT_RUNLOCK(tobject);
2435 		tobject = t1object;
2436 	}
2437 }
2438 
2439 /*
2440  * Return the vnode for the given object, or NULL if none exists.
2441  * For tmpfs objects, the function may return NULL if there is
2442  * no vnode allocated at the time of the call.
2443  */
2444 struct vnode *
2445 vm_object_vnode(vm_object_t object)
2446 {
2447 	struct vnode *vp;
2448 
2449 	VM_OBJECT_ASSERT_LOCKED(object);
2450 	vm_pager_getvp(object, &vp, NULL);
2451 	return (vp);
2452 }
2453 
2454 /*
2455  * Busy the vm object.  This prevents new pages belonging to the object from
2456  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2457  * for checking page state before proceeding.
2458  */
2459 void
2460 vm_object_busy(vm_object_t obj)
2461 {
2462 
2463 	VM_OBJECT_ASSERT_LOCKED(obj);
2464 
2465 	blockcount_acquire(&obj->busy, 1);
2466 	/* The fence is required to order loads of page busy. */
2467 	atomic_thread_fence_acq_rel();
2468 }
2469 
2470 void
2471 vm_object_unbusy(vm_object_t obj)
2472 {
2473 
2474 	blockcount_release(&obj->busy, 1);
2475 }
2476 
2477 void
2478 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2479 {
2480 
2481 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2482 
2483 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2484 }
2485 
2486 /*
2487  * This function aims to determine if the object is mapped,
2488  * specifically, if it is referenced by a vm_map_entry.  Because
2489  * objects occasionally acquire transient references that do not
2490  * represent a mapping, the method used here is inexact.  However, it
2491  * has very low overhead and is good enough for the advisory
2492  * vm.vmtotal sysctl.
2493  */
2494 bool
2495 vm_object_is_active(vm_object_t obj)
2496 {
2497 
2498 	return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2499 }
2500 
2501 static int
2502 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2503 {
2504 	struct kinfo_vmobject *kvo;
2505 	char *fullpath, *freepath;
2506 	struct vnode *vp;
2507 	struct vattr va;
2508 	vm_object_t obj;
2509 	vm_page_t m;
2510 	struct cdev *cdev;
2511 	struct cdevsw *csw;
2512 	u_long sp;
2513 	int count, error, ref;
2514 	key_t key;
2515 	unsigned short seq;
2516 	bool want_path;
2517 
2518 	if (req->oldptr == NULL) {
2519 		/*
2520 		 * If an old buffer has not been provided, generate an
2521 		 * estimate of the space needed for a subsequent call.
2522 		 */
2523 		mtx_lock(&vm_object_list_mtx);
2524 		count = 0;
2525 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2526 			if (obj->type == OBJT_DEAD)
2527 				continue;
2528 			count++;
2529 		}
2530 		mtx_unlock(&vm_object_list_mtx);
2531 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2532 		    count * 11 / 10));
2533 	}
2534 
2535 	want_path = !(swap_only || jailed(curthread->td_ucred));
2536 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO);
2537 	error = 0;
2538 
2539 	/*
2540 	 * VM objects are type stable and are never removed from the
2541 	 * list once added.  This allows us to safely read obj->object_list
2542 	 * after reacquiring the VM object lock.
2543 	 */
2544 	mtx_lock(&vm_object_list_mtx);
2545 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2546 		if (obj->type == OBJT_DEAD ||
2547 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2548 			continue;
2549 		VM_OBJECT_RLOCK(obj);
2550 		if (obj->type == OBJT_DEAD ||
2551 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2552 			VM_OBJECT_RUNLOCK(obj);
2553 			continue;
2554 		}
2555 		mtx_unlock(&vm_object_list_mtx);
2556 		kvo->kvo_size = ptoa(obj->size);
2557 		kvo->kvo_resident = obj->resident_page_count;
2558 		kvo->kvo_ref_count = obj->ref_count;
2559 		kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2560 		kvo->kvo_memattr = obj->memattr;
2561 		kvo->kvo_active = 0;
2562 		kvo->kvo_inactive = 0;
2563 		kvo->kvo_flags = 0;
2564 		if (!swap_only) {
2565 			TAILQ_FOREACH(m, &obj->memq, listq) {
2566 				/*
2567 				 * A page may belong to the object but be
2568 				 * dequeued and set to PQ_NONE while the
2569 				 * object lock is not held.  This makes the
2570 				 * reads of m->queue below racy, and we do not
2571 				 * count pages set to PQ_NONE.  However, this
2572 				 * sysctl is only meant to give an
2573 				 * approximation of the system anyway.
2574 				 */
2575 				if (m->a.queue == PQ_ACTIVE)
2576 					kvo->kvo_active++;
2577 				else if (m->a.queue == PQ_INACTIVE)
2578 					kvo->kvo_inactive++;
2579 			}
2580 		}
2581 
2582 		kvo->kvo_vn_fileid = 0;
2583 		kvo->kvo_vn_fsid = 0;
2584 		kvo->kvo_vn_fsid_freebsd11 = 0;
2585 		freepath = NULL;
2586 		fullpath = "";
2587 		vp = NULL;
2588 		kvo->kvo_type = vm_object_kvme_type(obj, want_path ? &vp :
2589 		    NULL);
2590 		if (vp != NULL) {
2591 			vref(vp);
2592 		} else if ((obj->flags & OBJ_ANON) != 0) {
2593 			MPASS(kvo->kvo_type == KVME_TYPE_SWAP);
2594 			kvo->kvo_me = (uintptr_t)obj;
2595 			/* tmpfs objs are reported as vnodes */
2596 			kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2597 			sp = swap_pager_swapped_pages(obj);
2598 			kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2599 		}
2600 		if (obj->type == OBJT_DEVICE || obj->type == OBJT_MGTDEVICE) {
2601 			cdev = obj->un_pager.devp.dev;
2602 			if (cdev != NULL) {
2603 				csw = dev_refthread(cdev, &ref);
2604 				if (csw != NULL) {
2605 					strlcpy(kvo->kvo_path, cdev->si_name,
2606 					    sizeof(kvo->kvo_path));
2607 					dev_relthread(cdev, ref);
2608 				}
2609 			}
2610 		}
2611 		VM_OBJECT_RUNLOCK(obj);
2612 		if ((obj->flags & OBJ_SYSVSHM) != 0) {
2613 			kvo->kvo_flags |= KVMO_FLAG_SYSVSHM;
2614 			shmobjinfo(obj, &key, &seq);
2615 			kvo->kvo_vn_fileid = key;
2616 			kvo->kvo_vn_fsid_freebsd11 = seq;
2617 		}
2618 		if ((obj->flags & OBJ_POSIXSHM) != 0) {
2619 			kvo->kvo_flags |= KVMO_FLAG_POSIXSHM;
2620 			shm_get_path(obj, kvo->kvo_path,
2621 			    sizeof(kvo->kvo_path));
2622 		}
2623 		if (vp != NULL) {
2624 			vn_fullpath(vp, &fullpath, &freepath);
2625 			vn_lock(vp, LK_SHARED | LK_RETRY);
2626 			if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2627 				kvo->kvo_vn_fileid = va.va_fileid;
2628 				kvo->kvo_vn_fsid = va.va_fsid;
2629 				kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2630 								/* truncate */
2631 			}
2632 			vput(vp);
2633 			strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2634 			free(freepath, M_TEMP);
2635 		}
2636 
2637 		/* Pack record size down */
2638 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2639 		    + strlen(kvo->kvo_path) + 1;
2640 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2641 		    sizeof(uint64_t));
2642 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2643 		maybe_yield();
2644 		mtx_lock(&vm_object_list_mtx);
2645 		if (error)
2646 			break;
2647 	}
2648 	mtx_unlock(&vm_object_list_mtx);
2649 	free(kvo, M_TEMP);
2650 	return (error);
2651 }
2652 
2653 static int
2654 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2655 {
2656 	return (vm_object_list_handler(req, false));
2657 }
2658 
2659 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2660     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2661     "List of VM objects");
2662 
2663 static int
2664 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2665 {
2666 	return (vm_object_list_handler(req, true));
2667 }
2668 
2669 /*
2670  * This sysctl returns list of the anonymous or swap objects. Intent
2671  * is to provide stripped optimized list useful to analyze swap use.
2672  * Since technically non-swap (default) objects participate in the
2673  * shadow chains, and are converted to swap type as needed by swap
2674  * pager, we must report them.
2675  */
2676 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2677     CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2678     sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2679     "List of swap VM objects");
2680 
2681 #include "opt_ddb.h"
2682 #ifdef DDB
2683 #include <sys/kernel.h>
2684 
2685 #include <sys/cons.h>
2686 
2687 #include <ddb/ddb.h>
2688 
2689 static int
2690 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2691 {
2692 	vm_map_t tmpm;
2693 	vm_map_entry_t tmpe;
2694 	vm_object_t obj;
2695 
2696 	if (map == 0)
2697 		return 0;
2698 
2699 	if (entry == 0) {
2700 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2701 			if (_vm_object_in_map(map, object, tmpe)) {
2702 				return 1;
2703 			}
2704 		}
2705 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2706 		tmpm = entry->object.sub_map;
2707 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2708 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2709 				return 1;
2710 			}
2711 		}
2712 	} else if ((obj = entry->object.vm_object) != NULL) {
2713 		for (; obj; obj = obj->backing_object)
2714 			if (obj == object) {
2715 				return 1;
2716 			}
2717 	}
2718 	return 0;
2719 }
2720 
2721 static int
2722 vm_object_in_map(vm_object_t object)
2723 {
2724 	struct proc *p;
2725 
2726 	/* sx_slock(&allproc_lock); */
2727 	FOREACH_PROC_IN_SYSTEM(p) {
2728 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2729 			continue;
2730 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2731 			/* sx_sunlock(&allproc_lock); */
2732 			return 1;
2733 		}
2734 	}
2735 	/* sx_sunlock(&allproc_lock); */
2736 	if (_vm_object_in_map(kernel_map, object, 0))
2737 		return 1;
2738 	return 0;
2739 }
2740 
2741 DB_SHOW_COMMAND_FLAGS(vmochk, vm_object_check, DB_CMD_MEMSAFE)
2742 {
2743 	vm_object_t object;
2744 
2745 	/*
2746 	 * make sure that internal objs are in a map somewhere
2747 	 * and none have zero ref counts.
2748 	 */
2749 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2750 		if ((object->flags & OBJ_ANON) != 0) {
2751 			if (object->ref_count == 0) {
2752 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2753 					(long)object->size);
2754 			}
2755 			if (!vm_object_in_map(object)) {
2756 				db_printf(
2757 			"vmochk: internal obj is not in a map: "
2758 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2759 				    object->ref_count, (u_long)object->size,
2760 				    (u_long)object->size,
2761 				    (void *)object->backing_object);
2762 			}
2763 		}
2764 		if (db_pager_quit)
2765 			return;
2766 	}
2767 }
2768 
2769 /*
2770  *	vm_object_print:	[ debug ]
2771  */
2772 DB_SHOW_COMMAND(object, vm_object_print_static)
2773 {
2774 	/* XXX convert args. */
2775 	vm_object_t object = (vm_object_t)addr;
2776 	boolean_t full = have_addr;
2777 
2778 	vm_page_t p;
2779 
2780 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2781 #define	count	was_count
2782 
2783 	int count;
2784 
2785 	if (object == NULL)
2786 		return;
2787 
2788 	db_iprintf(
2789 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2790 	    object, (int)object->type, (uintmax_t)object->size,
2791 	    object->resident_page_count, object->ref_count, object->flags,
2792 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2793 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2794 	    atomic_load_int(&object->shadow_count),
2795 	    object->backing_object ? object->backing_object->ref_count : 0,
2796 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2797 
2798 	if (!full)
2799 		return;
2800 
2801 	db_indent += 2;
2802 	count = 0;
2803 	TAILQ_FOREACH(p, &object->memq, listq) {
2804 		if (count == 0)
2805 			db_iprintf("memory:=");
2806 		else if (count == 6) {
2807 			db_printf("\n");
2808 			db_iprintf(" ...");
2809 			count = 0;
2810 		} else
2811 			db_printf(",");
2812 		count++;
2813 
2814 		db_printf("(off=0x%jx,page=0x%jx)",
2815 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2816 
2817 		if (db_pager_quit)
2818 			break;
2819 	}
2820 	if (count != 0)
2821 		db_printf("\n");
2822 	db_indent -= 2;
2823 }
2824 
2825 /* XXX. */
2826 #undef count
2827 
2828 /* XXX need this non-static entry for calling from vm_map_print. */
2829 void
2830 vm_object_print(
2831         /* db_expr_t */ long addr,
2832 	boolean_t have_addr,
2833 	/* db_expr_t */ long count,
2834 	char *modif)
2835 {
2836 	vm_object_print_static(addr, have_addr, count, modif);
2837 }
2838 
2839 DB_SHOW_COMMAND_FLAGS(vmopag, vm_object_print_pages, DB_CMD_MEMSAFE)
2840 {
2841 	vm_object_t object;
2842 	vm_pindex_t fidx;
2843 	vm_paddr_t pa;
2844 	vm_page_t m, prev_m;
2845 	int rcount;
2846 
2847 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2848 		db_printf("new object: %p\n", (void *)object);
2849 		if (db_pager_quit)
2850 			return;
2851 
2852 		rcount = 0;
2853 		fidx = 0;
2854 		pa = -1;
2855 		TAILQ_FOREACH(m, &object->memq, listq) {
2856 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2857 			    prev_m->pindex + 1 != m->pindex) {
2858 				if (rcount) {
2859 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2860 						(long)fidx, rcount, (long)pa);
2861 					if (db_pager_quit)
2862 						return;
2863 					rcount = 0;
2864 				}
2865 			}
2866 			if (rcount &&
2867 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2868 				++rcount;
2869 				continue;
2870 			}
2871 			if (rcount) {
2872 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2873 					(long)fidx, rcount, (long)pa);
2874 				if (db_pager_quit)
2875 					return;
2876 			}
2877 			fidx = m->pindex;
2878 			pa = VM_PAGE_TO_PHYS(m);
2879 			rcount = 1;
2880 		}
2881 		if (rcount) {
2882 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2883 				(long)fidx, rcount, (long)pa);
2884 			if (db_pager_quit)
2885 				return;
2886 		}
2887 	}
2888 }
2889 #endif /* DDB */
2890