xref: /freebsd/sys/vm/vm_object.c (revision f7c32ed617858bcd22f8d1b03199099d50125721)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	from: @(#)vm_object.c	8.5 (Berkeley) 3/22/94
35  *
36  *
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *	Virtual memory object module.
65  */
66 
67 #include <sys/cdefs.h>
68 __FBSDID("$FreeBSD$");
69 
70 #include "opt_vm.h"
71 
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/blockcount.h>
75 #include <sys/cpuset.h>
76 #include <sys/limits.h>
77 #include <sys/lock.h>
78 #include <sys/mman.h>
79 #include <sys/mount.h>
80 #include <sys/kernel.h>
81 #include <sys/pctrie.h>
82 #include <sys/sysctl.h>
83 #include <sys/mutex.h>
84 #include <sys/proc.h>		/* for curproc, pageproc */
85 #include <sys/refcount.h>
86 #include <sys/socket.h>
87 #include <sys/resourcevar.h>
88 #include <sys/refcount.h>
89 #include <sys/rwlock.h>
90 #include <sys/user.h>
91 #include <sys/vnode.h>
92 #include <sys/vmmeter.h>
93 #include <sys/sx.h>
94 
95 #include <vm/vm.h>
96 #include <vm/vm_param.h>
97 #include <vm/pmap.h>
98 #include <vm/vm_map.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_pageout.h>
102 #include <vm/vm_pager.h>
103 #include <vm/vm_phys.h>
104 #include <vm/vm_pagequeue.h>
105 #include <vm/swap_pager.h>
106 #include <vm/vm_kern.h>
107 #include <vm/vm_extern.h>
108 #include <vm/vm_radix.h>
109 #include <vm/vm_reserv.h>
110 #include <vm/uma.h>
111 
112 static int old_msync;
113 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
114     "Use old (insecure) msync behavior");
115 
116 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
117 		    int pagerflags, int flags, boolean_t *allclean,
118 		    boolean_t *eio);
119 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
120 		    boolean_t *allclean);
121 static void	vm_object_backing_remove(vm_object_t object);
122 
123 /*
124  *	Virtual memory objects maintain the actual data
125  *	associated with allocated virtual memory.  A given
126  *	page of memory exists within exactly one object.
127  *
128  *	An object is only deallocated when all "references"
129  *	are given up.  Only one "reference" to a given
130  *	region of an object should be writeable.
131  *
132  *	Associated with each object is a list of all resident
133  *	memory pages belonging to that object; this list is
134  *	maintained by the "vm_page" module, and locked by the object's
135  *	lock.
136  *
137  *	Each object also records a "pager" routine which is
138  *	used to retrieve (and store) pages to the proper backing
139  *	storage.  In addition, objects may be backed by other
140  *	objects from which they were virtual-copied.
141  *
142  *	The only items within the object structure which are
143  *	modified after time of creation are:
144  *		reference count		locked by object's lock
145  *		pager routine		locked by object's lock
146  *
147  */
148 
149 struct object_q vm_object_list;
150 struct mtx vm_object_list_mtx;	/* lock for object list and count */
151 
152 struct vm_object kernel_object_store;
153 
154 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
155     "VM object stats");
156 
157 static COUNTER_U64_DEFINE_EARLY(object_collapses);
158 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
159     &object_collapses,
160     "VM object collapses");
161 
162 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
163 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
164     &object_bypasses,
165     "VM object bypasses");
166 
167 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
168 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
169     &object_collapse_waits,
170     "Number of sleeps for collapse");
171 
172 static uma_zone_t obj_zone;
173 
174 static int vm_object_zinit(void *mem, int size, int flags);
175 
176 #ifdef INVARIANTS
177 static void vm_object_zdtor(void *mem, int size, void *arg);
178 
179 static void
180 vm_object_zdtor(void *mem, int size, void *arg)
181 {
182 	vm_object_t object;
183 
184 	object = (vm_object_t)mem;
185 	KASSERT(object->ref_count == 0,
186 	    ("object %p ref_count = %d", object, object->ref_count));
187 	KASSERT(TAILQ_EMPTY(&object->memq),
188 	    ("object %p has resident pages in its memq", object));
189 	KASSERT(vm_radix_is_empty(&object->rtree),
190 	    ("object %p has resident pages in its trie", object));
191 #if VM_NRESERVLEVEL > 0
192 	KASSERT(LIST_EMPTY(&object->rvq),
193 	    ("object %p has reservations",
194 	    object));
195 #endif
196 	KASSERT(!vm_object_busied(object),
197 	    ("object %p busy = %d", object, blockcount_read(&object->busy)));
198 	KASSERT(object->resident_page_count == 0,
199 	    ("object %p resident_page_count = %d",
200 	    object, object->resident_page_count));
201 	KASSERT(object->shadow_count == 0,
202 	    ("object %p shadow_count = %d",
203 	    object, object->shadow_count));
204 	KASSERT(object->type == OBJT_DEAD,
205 	    ("object %p has non-dead type %d",
206 	    object, object->type));
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, "vm object", 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 	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 
247 	/*
248 	 * Ensure that swap_pager_swapoff() iteration over object_list
249 	 * sees up to date type and pctrie head if it observed
250 	 * non-dead object.
251 	 */
252 	atomic_thread_fence_rel();
253 
254 	object->pg_color = 0;
255 	object->size = size;
256 	object->domain.dr_policy = NULL;
257 	object->generation = 1;
258 	object->cleangeneration = 1;
259 	refcount_init(&object->ref_count, 1);
260 	object->memattr = VM_MEMATTR_DEFAULT;
261 	object->cred = NULL;
262 	object->charge = 0;
263 	object->handle = handle;
264 	object->backing_object = NULL;
265 	object->backing_object_offset = (vm_ooffset_t) 0;
266 #if VM_NRESERVLEVEL > 0
267 	LIST_INIT(&object->rvq);
268 #endif
269 	umtx_shm_object_init(object);
270 }
271 
272 /*
273  *	vm_object_init:
274  *
275  *	Initialize the VM objects module.
276  */
277 void
278 vm_object_init(void)
279 {
280 	TAILQ_INIT(&vm_object_list);
281 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
282 
283 	rw_init(&kernel_object->lock, "kernel vm object");
284 	_vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
285 	    VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
286 #if VM_NRESERVLEVEL > 0
287 	kernel_object->flags |= OBJ_COLORED;
288 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
289 #endif
290 	kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
291 
292 	/*
293 	 * The lock portion of struct vm_object must be type stable due
294 	 * to vm_pageout_fallback_object_lock locking a vm object
295 	 * without holding any references to it.
296 	 *
297 	 * paging_in_progress is valid always.  Lockless references to
298 	 * the objects may acquire pip and then check OBJ_DEAD.
299 	 */
300 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
301 #ifdef INVARIANTS
302 	    vm_object_zdtor,
303 #else
304 	    NULL,
305 #endif
306 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
307 
308 	vm_radix_zinit();
309 }
310 
311 void
312 vm_object_clear_flag(vm_object_t object, u_short bits)
313 {
314 
315 	VM_OBJECT_ASSERT_WLOCKED(object);
316 	object->flags &= ~bits;
317 }
318 
319 /*
320  *	Sets the default memory attribute for the specified object.  Pages
321  *	that are allocated to this object are by default assigned this memory
322  *	attribute.
323  *
324  *	Presently, this function must be called before any pages are allocated
325  *	to the object.  In the future, this requirement may be relaxed for
326  *	"default" and "swap" objects.
327  */
328 int
329 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
330 {
331 
332 	VM_OBJECT_ASSERT_WLOCKED(object);
333 
334 	if (object->type == OBJT_DEAD)
335 		return (KERN_INVALID_ARGUMENT);
336 	if (!TAILQ_EMPTY(&object->memq))
337 		return (KERN_FAILURE);
338 
339 	object->memattr = memattr;
340 	return (KERN_SUCCESS);
341 }
342 
343 void
344 vm_object_pip_add(vm_object_t object, short i)
345 {
346 
347 	if (i > 0)
348 		blockcount_acquire(&object->paging_in_progress, i);
349 }
350 
351 void
352 vm_object_pip_wakeup(vm_object_t object)
353 {
354 
355 	vm_object_pip_wakeupn(object, 1);
356 }
357 
358 void
359 vm_object_pip_wakeupn(vm_object_t object, short i)
360 {
361 
362 	if (i > 0)
363 		blockcount_release(&object->paging_in_progress, i);
364 }
365 
366 /*
367  * Atomically drop the object lock and wait for pip to drain.  This protects
368  * from sleep/wakeup races due to identity changes.  The lock is not re-acquired
369  * on return.
370  */
371 static void
372 vm_object_pip_sleep(vm_object_t object, const char *waitid)
373 {
374 
375 	(void)blockcount_sleep(&object->paging_in_progress, &object->lock,
376 	    waitid, PVM | PDROP);
377 }
378 
379 void
380 vm_object_pip_wait(vm_object_t object, const char *waitid)
381 {
382 
383 	VM_OBJECT_ASSERT_WLOCKED(object);
384 
385 	blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
386 	    PVM);
387 }
388 
389 void
390 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
391 {
392 
393 	VM_OBJECT_ASSERT_UNLOCKED(object);
394 
395 	blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
396 }
397 
398 /*
399  *	vm_object_allocate:
400  *
401  *	Returns a new object with the given size.
402  */
403 vm_object_t
404 vm_object_allocate(objtype_t type, vm_pindex_t size)
405 {
406 	vm_object_t object;
407 	u_short flags;
408 
409 	switch (type) {
410 	case OBJT_DEAD:
411 		panic("vm_object_allocate: can't create OBJT_DEAD");
412 	case OBJT_DEFAULT:
413 		flags = OBJ_COLORED;
414 		break;
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_DEFAULT, size, OBJ_ANON | OBJ_ONEMAPPING,
474 	    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 && backing_object->shadow_count == 1,
578 	    ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
579 	    backing_object->ref_count, backing_object->shadow_count));
580 	KASSERT((object->flags & OBJ_ANON) != 0,
581 	    ("invalid shadow object %p", object));
582 
583 	if (!VM_OBJECT_TRYWLOCK(object)) {
584 		/*
585 		 * Prevent object from disappearing since we do not have a
586 		 * reference.
587 		 */
588 		vm_object_pip_add(object, 1);
589 		VM_OBJECT_WUNLOCK(backing_object);
590 		VM_OBJECT_WLOCK(object);
591 		vm_object_pip_wakeup(object);
592 	} else
593 		VM_OBJECT_WUNLOCK(backing_object);
594 
595 	/*
596 	 * Check for a collapse/terminate race with the last reference holder.
597 	 */
598 	if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
599 	    !refcount_acquire_if_not_zero(&object->ref_count)) {
600 		VM_OBJECT_WUNLOCK(object);
601 		return (NULL);
602 	}
603 	backing_object = object->backing_object;
604 	if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
605 		vm_object_collapse(object);
606 	VM_OBJECT_WUNLOCK(object);
607 
608 	return (object);
609 }
610 
611 /*
612  *	vm_object_deallocate:
613  *
614  *	Release a reference to the specified object,
615  *	gained either through a vm_object_allocate
616  *	or a vm_object_reference call.  When all references
617  *	are gone, storage associated with this object
618  *	may be relinquished.
619  *
620  *	No object may be locked.
621  */
622 void
623 vm_object_deallocate(vm_object_t object)
624 {
625 	vm_object_t temp;
626 	bool released;
627 
628 	while (object != NULL) {
629 		/*
630 		 * If the reference count goes to 0 we start calling
631 		 * vm_object_terminate() on the object chain.  A ref count
632 		 * of 1 may be a special case depending on the shadow count
633 		 * being 0 or 1.  These cases require a write lock on the
634 		 * object.
635 		 */
636 		if ((object->flags & OBJ_ANON) == 0)
637 			released = refcount_release_if_gt(&object->ref_count, 1);
638 		else
639 			released = refcount_release_if_gt(&object->ref_count, 2);
640 		if (released)
641 			return;
642 
643 		if (object->type == OBJT_VNODE) {
644 			VM_OBJECT_RLOCK(object);
645 			if (object->type == OBJT_VNODE) {
646 				vm_object_deallocate_vnode(object);
647 				return;
648 			}
649 			VM_OBJECT_RUNLOCK(object);
650 		}
651 
652 		VM_OBJECT_WLOCK(object);
653 		KASSERT(object->ref_count > 0,
654 		    ("vm_object_deallocate: object deallocated too many times: %d",
655 		    object->type));
656 
657 		/*
658 		 * If this is not the final reference to an anonymous
659 		 * object we may need to collapse the shadow chain.
660 		 */
661 		if (!refcount_release(&object->ref_count)) {
662 			if (object->ref_count > 1 ||
663 			    object->shadow_count == 0) {
664 				if ((object->flags & OBJ_ANON) != 0 &&
665 				    object->ref_count == 1)
666 					vm_object_set_flag(object,
667 					    OBJ_ONEMAPPING);
668 				VM_OBJECT_WUNLOCK(object);
669 				return;
670 			}
671 
672 			/* Handle collapsing last ref on anonymous objects. */
673 			object = vm_object_deallocate_anon(object);
674 			continue;
675 		}
676 
677 		/*
678 		 * Handle the final reference to an object.  We restart
679 		 * the loop with the backing object to avoid recursion.
680 		 */
681 		umtx_shm_object_terminated(object);
682 		temp = object->backing_object;
683 		if (temp != NULL) {
684 			KASSERT(object->type == OBJT_DEFAULT ||
685 			    object->type == OBJT_SWAP,
686 			    ("shadowed tmpfs v_object 2 %p", object));
687 			vm_object_backing_remove(object);
688 		}
689 
690 		KASSERT((object->flags & OBJ_DEAD) == 0,
691 		    ("vm_object_deallocate: Terminating dead object."));
692 		vm_object_set_flag(object, OBJ_DEAD);
693 		vm_object_terminate(object);
694 		object = temp;
695 	}
696 }
697 
698 /*
699  *	vm_object_destroy removes the object from the global object list
700  *      and frees the space for the object.
701  */
702 void
703 vm_object_destroy(vm_object_t object)
704 {
705 
706 	/*
707 	 * Release the allocation charge.
708 	 */
709 	if (object->cred != NULL) {
710 		swap_release_by_cred(object->charge, object->cred);
711 		object->charge = 0;
712 		crfree(object->cred);
713 		object->cred = NULL;
714 	}
715 
716 	/*
717 	 * Free the space for the object.
718 	 */
719 	uma_zfree(obj_zone, object);
720 }
721 
722 static void
723 vm_object_backing_remove_locked(vm_object_t object)
724 {
725 	vm_object_t backing_object;
726 
727 	backing_object = object->backing_object;
728 	VM_OBJECT_ASSERT_WLOCKED(object);
729 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
730 
731 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
732 	    ("vm_object_backing_remove: Removing collapsing object."));
733 
734 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
735 		LIST_REMOVE(object, shadow_list);
736 		backing_object->shadow_count--;
737 		object->flags &= ~OBJ_SHADOWLIST;
738 	}
739 	object->backing_object = NULL;
740 }
741 
742 static void
743 vm_object_backing_remove(vm_object_t object)
744 {
745 	vm_object_t backing_object;
746 
747 	VM_OBJECT_ASSERT_WLOCKED(object);
748 
749 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
750 		backing_object = object->backing_object;
751 		VM_OBJECT_WLOCK(backing_object);
752 		vm_object_backing_remove_locked(object);
753 		VM_OBJECT_WUNLOCK(backing_object);
754 	} else
755 		object->backing_object = NULL;
756 }
757 
758 static void
759 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
760 {
761 
762 	VM_OBJECT_ASSERT_WLOCKED(object);
763 
764 	if ((backing_object->flags & OBJ_ANON) != 0) {
765 		VM_OBJECT_ASSERT_WLOCKED(backing_object);
766 		LIST_INSERT_HEAD(&backing_object->shadow_head, object,
767 		    shadow_list);
768 		backing_object->shadow_count++;
769 		object->flags |= OBJ_SHADOWLIST;
770 	}
771 	object->backing_object = backing_object;
772 }
773 
774 static void
775 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
776 {
777 
778 	VM_OBJECT_ASSERT_WLOCKED(object);
779 
780 	if ((backing_object->flags & OBJ_ANON) != 0) {
781 		VM_OBJECT_WLOCK(backing_object);
782 		vm_object_backing_insert_locked(object, backing_object);
783 		VM_OBJECT_WUNLOCK(backing_object);
784 	} else
785 		object->backing_object = backing_object;
786 }
787 
788 /*
789  * Insert an object into a backing_object's shadow list with an additional
790  * reference to the backing_object added.
791  */
792 static void
793 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
794 {
795 
796 	VM_OBJECT_ASSERT_WLOCKED(object);
797 
798 	if ((backing_object->flags & OBJ_ANON) != 0) {
799 		VM_OBJECT_WLOCK(backing_object);
800 		KASSERT((backing_object->flags & OBJ_DEAD) == 0,
801 		    ("shadowing dead anonymous object"));
802 		vm_object_reference_locked(backing_object);
803 		vm_object_backing_insert_locked(object, backing_object);
804 		vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
805 		VM_OBJECT_WUNLOCK(backing_object);
806 	} else {
807 		vm_object_reference(backing_object);
808 		object->backing_object = backing_object;
809 	}
810 }
811 
812 /*
813  * Transfer a backing reference from backing_object to object.
814  */
815 static void
816 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
817 {
818 	vm_object_t new_backing_object;
819 
820 	/*
821 	 * Note that the reference to backing_object->backing_object
822 	 * moves from within backing_object to within object.
823 	 */
824 	vm_object_backing_remove_locked(object);
825 	new_backing_object = backing_object->backing_object;
826 	if (new_backing_object == NULL)
827 		return;
828 	if ((new_backing_object->flags & OBJ_ANON) != 0) {
829 		VM_OBJECT_WLOCK(new_backing_object);
830 		vm_object_backing_remove_locked(backing_object);
831 		vm_object_backing_insert_locked(object, new_backing_object);
832 		VM_OBJECT_WUNLOCK(new_backing_object);
833 	} else {
834 		object->backing_object = new_backing_object;
835 		backing_object->backing_object = NULL;
836 	}
837 }
838 
839 /*
840  * Wait for a concurrent collapse to settle.
841  */
842 static void
843 vm_object_collapse_wait(vm_object_t object)
844 {
845 
846 	VM_OBJECT_ASSERT_WLOCKED(object);
847 
848 	while ((object->flags & OBJ_COLLAPSING) != 0) {
849 		vm_object_pip_wait(object, "vmcolwait");
850 		counter_u64_add(object_collapse_waits, 1);
851 	}
852 }
853 
854 /*
855  * Waits for a backing object to clear a pending collapse and returns
856  * it locked if it is an ANON object.
857  */
858 static vm_object_t
859 vm_object_backing_collapse_wait(vm_object_t object)
860 {
861 	vm_object_t backing_object;
862 
863 	VM_OBJECT_ASSERT_WLOCKED(object);
864 
865 	for (;;) {
866 		backing_object = object->backing_object;
867 		if (backing_object == NULL ||
868 		    (backing_object->flags & OBJ_ANON) == 0)
869 			return (NULL);
870 		VM_OBJECT_WLOCK(backing_object);
871 		if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
872 			break;
873 		VM_OBJECT_WUNLOCK(object);
874 		vm_object_pip_sleep(backing_object, "vmbckwait");
875 		counter_u64_add(object_collapse_waits, 1);
876 		VM_OBJECT_WLOCK(object);
877 	}
878 	return (backing_object);
879 }
880 
881 /*
882  *	vm_object_terminate_pages removes any remaining pageable pages
883  *	from the object and resets the object to an empty state.
884  */
885 static void
886 vm_object_terminate_pages(vm_object_t object)
887 {
888 	vm_page_t p, p_next;
889 
890 	VM_OBJECT_ASSERT_WLOCKED(object);
891 
892 	/*
893 	 * Free any remaining pageable pages.  This also removes them from the
894 	 * paging queues.  However, don't free wired pages, just remove them
895 	 * from the object.  Rather than incrementally removing each page from
896 	 * the object, the page and object are reset to any empty state.
897 	 */
898 	TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
899 		vm_page_assert_unbusied(p);
900 		KASSERT(p->object == object &&
901 		    (p->ref_count & VPRC_OBJREF) != 0,
902 		    ("vm_object_terminate_pages: page %p is inconsistent", p));
903 
904 		p->object = NULL;
905 		if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
906 			VM_CNT_INC(v_pfree);
907 			vm_page_free(p);
908 		}
909 	}
910 
911 	/*
912 	 * If the object contained any pages, then reset it to an empty state.
913 	 * None of the object's fields, including "resident_page_count", were
914 	 * modified by the preceding loop.
915 	 */
916 	if (object->resident_page_count != 0) {
917 		vm_radix_reclaim_allnodes(&object->rtree);
918 		TAILQ_INIT(&object->memq);
919 		object->resident_page_count = 0;
920 		if (object->type == OBJT_VNODE)
921 			vdrop(object->handle);
922 	}
923 }
924 
925 /*
926  *	vm_object_terminate actually destroys the specified object, freeing
927  *	up all previously used resources.
928  *
929  *	The object must be locked.
930  *	This routine may block.
931  */
932 void
933 vm_object_terminate(vm_object_t object)
934 {
935 
936 	VM_OBJECT_ASSERT_WLOCKED(object);
937 	KASSERT((object->flags & OBJ_DEAD) != 0,
938 	    ("terminating non-dead obj %p", object));
939 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
940 	    ("terminating collapsing obj %p", object));
941 	KASSERT(object->backing_object == NULL,
942 	    ("terminating shadow obj %p", object));
943 
944 	/*
945 	 * Wait for the pageout daemon and other current users to be
946 	 * done with the object.  Note that new paging_in_progress
947 	 * users can come after this wait, but they must check
948 	 * OBJ_DEAD flag set (without unlocking the object), and avoid
949 	 * the object being terminated.
950 	 */
951 	vm_object_pip_wait(object, "objtrm");
952 
953 	KASSERT(object->ref_count == 0,
954 	    ("vm_object_terminate: object with references, ref_count=%d",
955 	    object->ref_count));
956 
957 	if ((object->flags & OBJ_PG_DTOR) == 0)
958 		vm_object_terminate_pages(object);
959 
960 #if VM_NRESERVLEVEL > 0
961 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
962 		vm_reserv_break_all(object);
963 #endif
964 
965 	KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
966 	    (object->flags & OBJ_SWAP) != 0,
967 	    ("%s: non-swap obj %p has cred", __func__, object));
968 
969 	/*
970 	 * Let the pager know object is dead.
971 	 */
972 	vm_pager_deallocate(object);
973 	VM_OBJECT_WUNLOCK(object);
974 
975 	vm_object_destroy(object);
976 }
977 
978 /*
979  * Make the page read-only so that we can clear the object flags.  However, if
980  * this is a nosync mmap then the object is likely to stay dirty so do not
981  * mess with the page and do not clear the object flags.  Returns TRUE if the
982  * page should be flushed, and FALSE otherwise.
983  */
984 static boolean_t
985 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
986 {
987 
988 	vm_page_assert_busied(p);
989 
990 	/*
991 	 * If we have been asked to skip nosync pages and this is a
992 	 * nosync page, skip it.  Note that the object flags were not
993 	 * cleared in this case so we do not have to set them.
994 	 */
995 	if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
996 		*allclean = FALSE;
997 		return (FALSE);
998 	} else {
999 		pmap_remove_write(p);
1000 		return (p->dirty != 0);
1001 	}
1002 }
1003 
1004 /*
1005  *	vm_object_page_clean
1006  *
1007  *	Clean all dirty pages in the specified range of object.  Leaves page
1008  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
1009  *	write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1010  *	leaving the object dirty.
1011  *
1012  *	For swap objects backing tmpfs regular files, do not flush anything,
1013  *	but remove write protection on the mapped pages to update mtime through
1014  *	mmaped writes.
1015  *
1016  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
1017  *	synchronous clustering mode implementation.
1018  *
1019  *	Odd semantics: if start == end, we clean everything.
1020  *
1021  *	The object must be locked.
1022  *
1023  *	Returns FALSE if some page from the range was not written, as
1024  *	reported by the pager, and TRUE otherwise.
1025  */
1026 boolean_t
1027 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1028     int flags)
1029 {
1030 	vm_page_t np, p;
1031 	vm_pindex_t pi, tend, tstart;
1032 	int curgeneration, n, pagerflags;
1033 	boolean_t eio, res, allclean;
1034 
1035 	VM_OBJECT_ASSERT_WLOCKED(object);
1036 
1037 	if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1038 		return (TRUE);
1039 
1040 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1041 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1042 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1043 
1044 	tstart = OFF_TO_IDX(start);
1045 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1046 	allclean = tstart == 0 && tend >= object->size;
1047 	res = TRUE;
1048 
1049 rescan:
1050 	curgeneration = object->generation;
1051 
1052 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1053 		pi = p->pindex;
1054 		if (pi >= tend)
1055 			break;
1056 		np = TAILQ_NEXT(p, listq);
1057 		if (vm_page_none_valid(p))
1058 			continue;
1059 		if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1060 			if (object->generation != curgeneration &&
1061 			    (flags & OBJPC_SYNC) != 0)
1062 				goto rescan;
1063 			np = vm_page_find_least(object, pi);
1064 			continue;
1065 		}
1066 		if (!vm_object_page_remove_write(p, flags, &allclean)) {
1067 			vm_page_xunbusy(p);
1068 			continue;
1069 		}
1070 		if (object->type == OBJT_VNODE) {
1071 			n = vm_object_page_collect_flush(object, p, pagerflags,
1072 			    flags, &allclean, &eio);
1073 			if (eio) {
1074 				res = FALSE;
1075 				allclean = FALSE;
1076 			}
1077 			if (object->generation != curgeneration &&
1078 			    (flags & OBJPC_SYNC) != 0)
1079 				goto rescan;
1080 
1081 			/*
1082 			 * If the VOP_PUTPAGES() did a truncated write, so
1083 			 * that even the first page of the run is not fully
1084 			 * written, vm_pageout_flush() returns 0 as the run
1085 			 * length.  Since the condition that caused truncated
1086 			 * write may be permanent, e.g. exhausted free space,
1087 			 * accepting n == 0 would cause an infinite loop.
1088 			 *
1089 			 * Forwarding the iterator leaves the unwritten page
1090 			 * behind, but there is not much we can do there if
1091 			 * filesystem refuses to write it.
1092 			 */
1093 			if (n == 0) {
1094 				n = 1;
1095 				allclean = FALSE;
1096 			}
1097 		} else {
1098 			n = 1;
1099 			vm_page_xunbusy(p);
1100 		}
1101 		np = vm_page_find_least(object, pi + n);
1102 	}
1103 #if 0
1104 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1105 #endif
1106 
1107 	/*
1108 	 * Leave updating cleangeneration for tmpfs objects to tmpfs
1109 	 * scan.  It needs to update mtime, which happens for other
1110 	 * filesystems during page writeouts.
1111 	 */
1112 	if (allclean && object->type == OBJT_VNODE)
1113 		object->cleangeneration = curgeneration;
1114 	return (res);
1115 }
1116 
1117 static int
1118 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1119     int flags, boolean_t *allclean, boolean_t *eio)
1120 {
1121 	vm_page_t ma[vm_pageout_page_count], p_first, tp;
1122 	int count, i, mreq, runlen;
1123 
1124 	vm_page_lock_assert(p, MA_NOTOWNED);
1125 	vm_page_assert_xbusied(p);
1126 	VM_OBJECT_ASSERT_WLOCKED(object);
1127 
1128 	count = 1;
1129 	mreq = 0;
1130 
1131 	for (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 	}
1140 
1141 	for (p_first = p; count < vm_pageout_page_count; count++) {
1142 		tp = vm_page_prev(p_first);
1143 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1144 			break;
1145 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1146 			vm_page_xunbusy(tp);
1147 			break;
1148 		}
1149 		p_first = tp;
1150 		mreq++;
1151 	}
1152 
1153 	for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1154 		ma[i] = tp;
1155 
1156 	vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1157 	return (runlen);
1158 }
1159 
1160 /*
1161  * Note that there is absolutely no sense in writing out
1162  * anonymous objects, so we track down the vnode object
1163  * to write out.
1164  * We invalidate (remove) all pages from the address space
1165  * for semantic correctness.
1166  *
1167  * If the backing object is a device object with unmanaged pages, then any
1168  * mappings to the specified range of pages must be removed before this
1169  * function is called.
1170  *
1171  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1172  * may start out with a NULL object.
1173  */
1174 boolean_t
1175 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1176     boolean_t syncio, boolean_t invalidate)
1177 {
1178 	vm_object_t backing_object;
1179 	struct vnode *vp;
1180 	struct mount *mp;
1181 	int error, flags, fsync_after;
1182 	boolean_t res;
1183 
1184 	if (object == NULL)
1185 		return (TRUE);
1186 	res = TRUE;
1187 	error = 0;
1188 	VM_OBJECT_WLOCK(object);
1189 	while ((backing_object = object->backing_object) != NULL) {
1190 		VM_OBJECT_WLOCK(backing_object);
1191 		offset += object->backing_object_offset;
1192 		VM_OBJECT_WUNLOCK(object);
1193 		object = backing_object;
1194 		if (object->size < OFF_TO_IDX(offset + size))
1195 			size = IDX_TO_OFF(object->size) - offset;
1196 	}
1197 	/*
1198 	 * Flush pages if writing is allowed, invalidate them
1199 	 * if invalidation requested.  Pages undergoing I/O
1200 	 * will be ignored by vm_object_page_remove().
1201 	 *
1202 	 * We cannot lock the vnode and then wait for paging
1203 	 * to complete without deadlocking against vm_fault.
1204 	 * Instead we simply call vm_object_page_remove() and
1205 	 * allow it to block internally on a page-by-page
1206 	 * basis when it encounters pages undergoing async
1207 	 * I/O.
1208 	 */
1209 	if (object->type == OBJT_VNODE &&
1210 	    vm_object_mightbedirty(object) != 0 &&
1211 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1212 		VM_OBJECT_WUNLOCK(object);
1213 		(void) vn_start_write(vp, &mp, V_WAIT);
1214 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1215 		if (syncio && !invalidate && offset == 0 &&
1216 		    atop(size) == object->size) {
1217 			/*
1218 			 * If syncing the whole mapping of the file,
1219 			 * it is faster to schedule all the writes in
1220 			 * async mode, also allowing the clustering,
1221 			 * and then wait for i/o to complete.
1222 			 */
1223 			flags = 0;
1224 			fsync_after = TRUE;
1225 		} else {
1226 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1227 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1228 			fsync_after = FALSE;
1229 		}
1230 		VM_OBJECT_WLOCK(object);
1231 		res = vm_object_page_clean(object, offset, offset + size,
1232 		    flags);
1233 		VM_OBJECT_WUNLOCK(object);
1234 		if (fsync_after)
1235 			error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1236 		VOP_UNLOCK(vp);
1237 		vn_finished_write(mp);
1238 		if (error != 0)
1239 			res = FALSE;
1240 		VM_OBJECT_WLOCK(object);
1241 	}
1242 	if ((object->type == OBJT_VNODE ||
1243 	     object->type == OBJT_DEVICE) && invalidate) {
1244 		if (object->type == OBJT_DEVICE)
1245 			/*
1246 			 * The option OBJPR_NOTMAPPED must be passed here
1247 			 * because vm_object_page_remove() cannot remove
1248 			 * unmanaged mappings.
1249 			 */
1250 			flags = OBJPR_NOTMAPPED;
1251 		else if (old_msync)
1252 			flags = 0;
1253 		else
1254 			flags = OBJPR_CLEANONLY;
1255 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1256 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1257 	}
1258 	VM_OBJECT_WUNLOCK(object);
1259 	return (res);
1260 }
1261 
1262 /*
1263  * Determine whether the given advice can be applied to the object.  Advice is
1264  * not applied to unmanaged pages since they never belong to page queues, and
1265  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1266  * have been mapped at most once.
1267  */
1268 static bool
1269 vm_object_advice_applies(vm_object_t object, int advice)
1270 {
1271 
1272 	if ((object->flags & OBJ_UNMANAGED) != 0)
1273 		return (false);
1274 	if (advice != MADV_FREE)
1275 		return (true);
1276 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1277 	    (OBJ_ONEMAPPING | OBJ_ANON));
1278 }
1279 
1280 static void
1281 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1282     vm_size_t size)
1283 {
1284 
1285 	if (advice == MADV_FREE)
1286 		vm_pager_freespace(object, pindex, size);
1287 }
1288 
1289 /*
1290  *	vm_object_madvise:
1291  *
1292  *	Implements the madvise function at the object/page level.
1293  *
1294  *	MADV_WILLNEED	(any object)
1295  *
1296  *	    Activate the specified pages if they are resident.
1297  *
1298  *	MADV_DONTNEED	(any object)
1299  *
1300  *	    Deactivate the specified pages if they are resident.
1301  *
1302  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1303  *			 OBJ_ONEMAPPING only)
1304  *
1305  *	    Deactivate and clean the specified pages if they are
1306  *	    resident.  This permits the process to reuse the pages
1307  *	    without faulting or the kernel to reclaim the pages
1308  *	    without I/O.
1309  */
1310 void
1311 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1312     int advice)
1313 {
1314 	vm_pindex_t tpindex;
1315 	vm_object_t backing_object, tobject;
1316 	vm_page_t m, tm;
1317 
1318 	if (object == NULL)
1319 		return;
1320 
1321 relookup:
1322 	VM_OBJECT_WLOCK(object);
1323 	if (!vm_object_advice_applies(object, advice)) {
1324 		VM_OBJECT_WUNLOCK(object);
1325 		return;
1326 	}
1327 	for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1328 		tobject = object;
1329 
1330 		/*
1331 		 * If the next page isn't resident in the top-level object, we
1332 		 * need to search the shadow chain.  When applying MADV_FREE, we
1333 		 * take care to release any swap space used to store
1334 		 * non-resident pages.
1335 		 */
1336 		if (m == NULL || pindex < m->pindex) {
1337 			/*
1338 			 * Optimize a common case: if the top-level object has
1339 			 * no backing object, we can skip over the non-resident
1340 			 * range in constant time.
1341 			 */
1342 			if (object->backing_object == NULL) {
1343 				tpindex = (m != NULL && m->pindex < end) ?
1344 				    m->pindex : end;
1345 				vm_object_madvise_freespace(object, advice,
1346 				    pindex, tpindex - pindex);
1347 				if ((pindex = tpindex) == end)
1348 					break;
1349 				goto next_page;
1350 			}
1351 
1352 			tpindex = pindex;
1353 			do {
1354 				vm_object_madvise_freespace(tobject, advice,
1355 				    tpindex, 1);
1356 				/*
1357 				 * Prepare to search the next object in the
1358 				 * chain.
1359 				 */
1360 				backing_object = tobject->backing_object;
1361 				if (backing_object == NULL)
1362 					goto next_pindex;
1363 				VM_OBJECT_WLOCK(backing_object);
1364 				tpindex +=
1365 				    OFF_TO_IDX(tobject->backing_object_offset);
1366 				if (tobject != object)
1367 					VM_OBJECT_WUNLOCK(tobject);
1368 				tobject = backing_object;
1369 				if (!vm_object_advice_applies(tobject, advice))
1370 					goto next_pindex;
1371 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1372 			    NULL);
1373 		} else {
1374 next_page:
1375 			tm = m;
1376 			m = TAILQ_NEXT(m, listq);
1377 		}
1378 
1379 		/*
1380 		 * If the page is not in a normal state, skip it.  The page
1381 		 * can not be invalidated while the object lock is held.
1382 		 */
1383 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1384 			goto next_pindex;
1385 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1386 		    ("vm_object_madvise: page %p is fictitious", tm));
1387 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1388 		    ("vm_object_madvise: page %p is not managed", tm));
1389 		if (vm_page_tryxbusy(tm) == 0) {
1390 			if (object != tobject)
1391 				VM_OBJECT_WUNLOCK(object);
1392 			if (advice == MADV_WILLNEED) {
1393 				/*
1394 				 * Reference the page before unlocking and
1395 				 * sleeping so that the page daemon is less
1396 				 * likely to reclaim it.
1397 				 */
1398 				vm_page_aflag_set(tm, PGA_REFERENCED);
1399 			}
1400 			if (!vm_page_busy_sleep(tm, "madvpo", 0))
1401 				VM_OBJECT_WUNLOCK(tobject);
1402   			goto relookup;
1403 		}
1404 		vm_page_advise(tm, advice);
1405 		vm_page_xunbusy(tm);
1406 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1407 next_pindex:
1408 		if (tobject != object)
1409 			VM_OBJECT_WUNLOCK(tobject);
1410 	}
1411 	VM_OBJECT_WUNLOCK(object);
1412 }
1413 
1414 /*
1415  *	vm_object_shadow:
1416  *
1417  *	Create a new object which is backed by the
1418  *	specified existing object range.  The source
1419  *	object reference is deallocated.
1420  *
1421  *	The new object and offset into that object
1422  *	are returned in the source parameters.
1423  */
1424 void
1425 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1426     struct ucred *cred, bool shared)
1427 {
1428 	vm_object_t source;
1429 	vm_object_t result;
1430 
1431 	source = *object;
1432 
1433 	/*
1434 	 * Don't create the new object if the old object isn't shared.
1435 	 *
1436 	 * If we hold the only reference we can guarantee that it won't
1437 	 * increase while we have the map locked.  Otherwise the race is
1438 	 * harmless and we will end up with an extra shadow object that
1439 	 * will be collapsed later.
1440 	 */
1441 	if (source != NULL && source->ref_count == 1 &&
1442 	    (source->flags & OBJ_ANON) != 0)
1443 		return;
1444 
1445 	/*
1446 	 * Allocate a new object with the given length.
1447 	 */
1448 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1449 
1450 	/*
1451 	 * Store the offset into the source object, and fix up the offset into
1452 	 * the new object.
1453 	 */
1454 	result->backing_object_offset = *offset;
1455 
1456 	if (shared || source != NULL) {
1457 		VM_OBJECT_WLOCK(result);
1458 
1459 		/*
1460 		 * The new object shadows the source object, adding a
1461 		 * reference to it.  Our caller changes his reference
1462 		 * to point to the new object, removing a reference to
1463 		 * the source object.  Net result: no change of
1464 		 * reference count, unless the caller needs to add one
1465 		 * more reference due to forking a shared map entry.
1466 		 */
1467 		if (shared) {
1468 			vm_object_reference_locked(result);
1469 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1470 		}
1471 
1472 		/*
1473 		 * Try to optimize the result object's page color when
1474 		 * shadowing in order to maintain page coloring
1475 		 * consistency in the combined shadowed object.
1476 		 */
1477 		if (source != NULL) {
1478 			vm_object_backing_insert(result, source);
1479 			result->domain = source->domain;
1480 #if VM_NRESERVLEVEL > 0
1481 			result->flags |= source->flags & OBJ_COLORED;
1482 			result->pg_color = (source->pg_color +
1483 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1484 			    1)) - 1);
1485 #endif
1486 		}
1487 		VM_OBJECT_WUNLOCK(result);
1488 	}
1489 
1490 	/*
1491 	 * Return the new things
1492 	 */
1493 	*offset = 0;
1494 	*object = result;
1495 }
1496 
1497 /*
1498  *	vm_object_split:
1499  *
1500  * Split the pages in a map entry into a new object.  This affords
1501  * easier removal of unused pages, and keeps object inheritance from
1502  * being a negative impact on memory usage.
1503  */
1504 void
1505 vm_object_split(vm_map_entry_t entry)
1506 {
1507 	vm_page_t m, m_busy, m_next;
1508 	vm_object_t orig_object, new_object, backing_object;
1509 	vm_pindex_t idx, offidxstart;
1510 	vm_size_t size;
1511 
1512 	orig_object = entry->object.vm_object;
1513 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1514 	    ("vm_object_split:  Splitting object with multiple mappings."));
1515 	if ((orig_object->flags & OBJ_ANON) == 0)
1516 		return;
1517 	if (orig_object->ref_count <= 1)
1518 		return;
1519 	VM_OBJECT_WUNLOCK(orig_object);
1520 
1521 	offidxstart = OFF_TO_IDX(entry->offset);
1522 	size = atop(entry->end - entry->start);
1523 
1524 	/*
1525 	 * If swap_pager_copy() is later called, it will convert new_object
1526 	 * into a swap object.
1527 	 */
1528 	new_object = vm_object_allocate_anon(size, orig_object,
1529 	    orig_object->cred, ptoa(size));
1530 
1531 	/*
1532 	 * We must wait for the orig_object to complete any in-progress
1533 	 * collapse so that the swap blocks are stable below.  The
1534 	 * additional reference on backing_object by new object will
1535 	 * prevent further collapse operations until split completes.
1536 	 */
1537 	VM_OBJECT_WLOCK(orig_object);
1538 	vm_object_collapse_wait(orig_object);
1539 
1540 	/*
1541 	 * At this point, the new object is still private, so the order in
1542 	 * which the original and new objects are locked does not matter.
1543 	 */
1544 	VM_OBJECT_WLOCK(new_object);
1545 	new_object->domain = orig_object->domain;
1546 	backing_object = orig_object->backing_object;
1547 	if (backing_object != NULL) {
1548 		vm_object_backing_insert_ref(new_object, backing_object);
1549 		new_object->backing_object_offset =
1550 		    orig_object->backing_object_offset + entry->offset;
1551 	}
1552 	if (orig_object->cred != NULL) {
1553 		crhold(orig_object->cred);
1554 		KASSERT(orig_object->charge >= ptoa(size),
1555 		    ("orig_object->charge < 0"));
1556 		orig_object->charge -= ptoa(size);
1557 	}
1558 
1559 	/*
1560 	 * Mark the split operation so that swap_pager_getpages() knows
1561 	 * that the object is in transition.
1562 	 */
1563 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1564 	m_busy = NULL;
1565 #ifdef INVARIANTS
1566 	idx = 0;
1567 #endif
1568 retry:
1569 	m = vm_page_find_least(orig_object, offidxstart);
1570 	KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1571 	    ("%s: object %p was repopulated", __func__, orig_object));
1572 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1573 	    m = m_next) {
1574 		m_next = TAILQ_NEXT(m, listq);
1575 
1576 		/*
1577 		 * We must wait for pending I/O to complete before we can
1578 		 * rename the page.
1579 		 *
1580 		 * We do not have to VM_PROT_NONE the page as mappings should
1581 		 * not be changed by this operation.
1582 		 */
1583 		if (vm_page_tryxbusy(m) == 0) {
1584 			VM_OBJECT_WUNLOCK(new_object);
1585 			if (vm_page_busy_sleep(m, "spltwt", 0))
1586 				VM_OBJECT_WLOCK(orig_object);
1587 			VM_OBJECT_WLOCK(new_object);
1588 			goto retry;
1589 		}
1590 
1591 		/*
1592 		 * The page was left invalid.  Likely placed there by
1593 		 * an incomplete fault.  Just remove and ignore.
1594 		 */
1595 		if (vm_page_none_valid(m)) {
1596 			if (vm_page_remove(m))
1597 				vm_page_free(m);
1598 			continue;
1599 		}
1600 
1601 		/* vm_page_rename() will dirty the page. */
1602 		if (vm_page_rename(m, new_object, idx)) {
1603 			vm_page_xunbusy(m);
1604 			VM_OBJECT_WUNLOCK(new_object);
1605 			VM_OBJECT_WUNLOCK(orig_object);
1606 			vm_radix_wait();
1607 			VM_OBJECT_WLOCK(orig_object);
1608 			VM_OBJECT_WLOCK(new_object);
1609 			goto retry;
1610 		}
1611 
1612 #if VM_NRESERVLEVEL > 0
1613 		/*
1614 		 * If some of the reservation's allocated pages remain with
1615 		 * the original object, then transferring the reservation to
1616 		 * the new object is neither particularly beneficial nor
1617 		 * particularly harmful as compared to leaving the reservation
1618 		 * with the original object.  If, however, all of the
1619 		 * reservation's allocated pages are transferred to the new
1620 		 * object, then transferring the reservation is typically
1621 		 * beneficial.  Determining which of these two cases applies
1622 		 * would be more costly than unconditionally renaming the
1623 		 * reservation.
1624 		 */
1625 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1626 #endif
1627 
1628 		/*
1629 		 * orig_object's type may change while sleeping, so keep track
1630 		 * of the beginning of the busied range.
1631 		 */
1632 		if (orig_object->type != OBJT_SWAP)
1633 			vm_page_xunbusy(m);
1634 		else if (m_busy == NULL)
1635 			m_busy = m;
1636 	}
1637 	if ((orig_object->flags & OBJ_SWAP) != 0) {
1638 		/*
1639 		 * swap_pager_copy() can sleep, in which case the orig_object's
1640 		 * and new_object's locks are released and reacquired.
1641 		 */
1642 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1643 		if (m_busy != NULL)
1644 			TAILQ_FOREACH_FROM(m_busy, &new_object->memq, listq)
1645 				vm_page_xunbusy(m_busy);
1646 	}
1647 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1648 	VM_OBJECT_WUNLOCK(orig_object);
1649 	VM_OBJECT_WUNLOCK(new_object);
1650 	entry->object.vm_object = new_object;
1651 	entry->offset = 0LL;
1652 	vm_object_deallocate(orig_object);
1653 	VM_OBJECT_WLOCK(new_object);
1654 }
1655 
1656 static vm_page_t
1657 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1658 {
1659 	vm_object_t backing_object;
1660 
1661 	VM_OBJECT_ASSERT_WLOCKED(object);
1662 	backing_object = object->backing_object;
1663 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1664 
1665 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1666 	    ("invalid ownership %p %p %p", p, object, backing_object));
1667 	/* The page is only NULL when rename fails. */
1668 	if (p == NULL) {
1669 		VM_OBJECT_WUNLOCK(object);
1670 		VM_OBJECT_WUNLOCK(backing_object);
1671 		vm_radix_wait();
1672 		VM_OBJECT_WLOCK(object);
1673 	} else if (p->object == object) {
1674 		VM_OBJECT_WUNLOCK(backing_object);
1675 		if (vm_page_busy_sleep(p, "vmocol", 0))
1676 			VM_OBJECT_WLOCK(object);
1677 	} else {
1678 		VM_OBJECT_WUNLOCK(object);
1679 		if (!vm_page_busy_sleep(p, "vmocol", 0))
1680 			VM_OBJECT_WUNLOCK(backing_object);
1681 		VM_OBJECT_WLOCK(object);
1682 	}
1683 	VM_OBJECT_WLOCK(backing_object);
1684 	return (TAILQ_FIRST(&backing_object->memq));
1685 }
1686 
1687 static bool
1688 vm_object_scan_all_shadowed(vm_object_t object)
1689 {
1690 	vm_object_t backing_object;
1691 	vm_page_t p, pp;
1692 	vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1693 
1694 	VM_OBJECT_ASSERT_WLOCKED(object);
1695 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1696 
1697 	backing_object = object->backing_object;
1698 
1699 	if ((backing_object->flags & OBJ_ANON) == 0)
1700 		return (false);
1701 
1702 	pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1703 	p = vm_page_find_least(backing_object, pi);
1704 	ps = swap_pager_find_least(backing_object, pi);
1705 
1706 	/*
1707 	 * Only check pages inside the parent object's range and
1708 	 * inside the parent object's mapping of the backing object.
1709 	 */
1710 	for (;; pi++) {
1711 		if (p != NULL && p->pindex < pi)
1712 			p = TAILQ_NEXT(p, listq);
1713 		if (ps < pi)
1714 			ps = swap_pager_find_least(backing_object, pi);
1715 		if (p == NULL && ps >= backing_object->size)
1716 			break;
1717 		else if (p == NULL)
1718 			pi = ps;
1719 		else
1720 			pi = MIN(p->pindex, ps);
1721 
1722 		new_pindex = pi - backing_offset_index;
1723 		if (new_pindex >= object->size)
1724 			break;
1725 
1726 		if (p != NULL) {
1727 			/*
1728 			 * If the backing object page is busy a
1729 			 * grandparent or older page may still be
1730 			 * undergoing CoW.  It is not safe to collapse
1731 			 * the backing object until it is quiesced.
1732 			 */
1733 			if (vm_page_tryxbusy(p) == 0)
1734 				return (false);
1735 
1736 			/*
1737 			 * We raced with the fault handler that left
1738 			 * newly allocated invalid page on the object
1739 			 * queue and retried.
1740 			 */
1741 			if (!vm_page_all_valid(p))
1742 				goto unbusy_ret;
1743 		}
1744 
1745 		/*
1746 		 * See if the parent has the page or if the parent's object
1747 		 * pager has the page.  If the parent has the page but the page
1748 		 * is not valid, the parent's object pager must have the page.
1749 		 *
1750 		 * If this fails, the parent does not completely shadow the
1751 		 * object and we might as well give up now.
1752 		 */
1753 		pp = vm_page_lookup(object, new_pindex);
1754 
1755 		/*
1756 		 * The valid check here is stable due to object lock
1757 		 * being required to clear valid and initiate paging.
1758 		 * Busy of p disallows fault handler to validate pp.
1759 		 */
1760 		if ((pp == NULL || vm_page_none_valid(pp)) &&
1761 		    !vm_pager_has_page(object, new_pindex, NULL, NULL))
1762 			goto unbusy_ret;
1763 		if (p != NULL)
1764 			vm_page_xunbusy(p);
1765 	}
1766 	return (true);
1767 
1768 unbusy_ret:
1769 	if (p != NULL)
1770 		vm_page_xunbusy(p);
1771 	return (false);
1772 }
1773 
1774 static void
1775 vm_object_collapse_scan(vm_object_t object)
1776 {
1777 	vm_object_t backing_object;
1778 	vm_page_t next, p, pp;
1779 	vm_pindex_t backing_offset_index, new_pindex;
1780 
1781 	VM_OBJECT_ASSERT_WLOCKED(object);
1782 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1783 
1784 	backing_object = object->backing_object;
1785 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1786 
1787 	/*
1788 	 * Our scan
1789 	 */
1790 	for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1791 		next = TAILQ_NEXT(p, listq);
1792 		new_pindex = p->pindex - backing_offset_index;
1793 
1794 		/*
1795 		 * Check for busy page
1796 		 */
1797 		if (vm_page_tryxbusy(p) == 0) {
1798 			next = vm_object_collapse_scan_wait(object, p);
1799 			continue;
1800 		}
1801 
1802 		KASSERT(object->backing_object == backing_object,
1803 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1804 		    object->backing_object, backing_object));
1805 		KASSERT(p->object == backing_object,
1806 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1807 		    p->object, backing_object));
1808 
1809 		if (p->pindex < backing_offset_index ||
1810 		    new_pindex >= object->size) {
1811 			vm_pager_freespace(backing_object, p->pindex, 1);
1812 
1813 			KASSERT(!pmap_page_is_mapped(p),
1814 			    ("freeing mapped page %p", p));
1815 			if (vm_page_remove(p))
1816 				vm_page_free(p);
1817 			continue;
1818 		}
1819 
1820 		if (!vm_page_all_valid(p)) {
1821 			KASSERT(!pmap_page_is_mapped(p),
1822 			    ("freeing mapped page %p", p));
1823 			if (vm_page_remove(p))
1824 				vm_page_free(p);
1825 			continue;
1826 		}
1827 
1828 		pp = vm_page_lookup(object, new_pindex);
1829 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1830 			vm_page_xunbusy(p);
1831 			/*
1832 			 * The page in the parent is busy and possibly not
1833 			 * (yet) valid.  Until its state is finalized by the
1834 			 * busy bit owner, we can't tell whether it shadows the
1835 			 * original page.
1836 			 */
1837 			next = vm_object_collapse_scan_wait(object, pp);
1838 			continue;
1839 		}
1840 
1841 		if (pp != NULL && vm_page_none_valid(pp)) {
1842 			/*
1843 			 * The page was invalid in the parent.  Likely placed
1844 			 * there by an incomplete fault.  Just remove and
1845 			 * ignore.  p can replace it.
1846 			 */
1847 			if (vm_page_remove(pp))
1848 				vm_page_free(pp);
1849 			pp = NULL;
1850 		}
1851 
1852 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1853 			NULL)) {
1854 			/*
1855 			 * The page already exists in the parent OR swap exists
1856 			 * for this location in the parent.  Leave the parent's
1857 			 * page alone.  Destroy the original page from the
1858 			 * backing object.
1859 			 */
1860 			vm_pager_freespace(backing_object, p->pindex, 1);
1861 			KASSERT(!pmap_page_is_mapped(p),
1862 			    ("freeing mapped page %p", p));
1863 			if (vm_page_remove(p))
1864 				vm_page_free(p);
1865 			if (pp != NULL)
1866 				vm_page_xunbusy(pp);
1867 			continue;
1868 		}
1869 
1870 		/*
1871 		 * Page does not exist in parent, rename the page from the
1872 		 * backing object to the main object.
1873 		 *
1874 		 * If the page was mapped to a process, it can remain mapped
1875 		 * through the rename.  vm_page_rename() will dirty the page.
1876 		 */
1877 		if (vm_page_rename(p, object, new_pindex)) {
1878 			vm_page_xunbusy(p);
1879 			next = vm_object_collapse_scan_wait(object, NULL);
1880 			continue;
1881 		}
1882 
1883 		/* Use the old pindex to free the right page. */
1884 		vm_pager_freespace(backing_object, new_pindex +
1885 		    backing_offset_index, 1);
1886 
1887 #if VM_NRESERVLEVEL > 0
1888 		/*
1889 		 * Rename the reservation.
1890 		 */
1891 		vm_reserv_rename(p, object, backing_object,
1892 		    backing_offset_index);
1893 #endif
1894 		vm_page_xunbusy(p);
1895 	}
1896 	return;
1897 }
1898 
1899 /*
1900  *	vm_object_collapse:
1901  *
1902  *	Collapse an object with the object backing it.
1903  *	Pages in the backing object are moved into the
1904  *	parent, and the backing object is deallocated.
1905  */
1906 void
1907 vm_object_collapse(vm_object_t object)
1908 {
1909 	vm_object_t backing_object, new_backing_object;
1910 
1911 	VM_OBJECT_ASSERT_WLOCKED(object);
1912 
1913 	while (TRUE) {
1914 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1915 		    ("collapsing invalid object"));
1916 
1917 		/*
1918 		 * Wait for the backing_object to finish any pending
1919 		 * collapse so that the caller sees the shortest possible
1920 		 * shadow chain.
1921 		 */
1922 		backing_object = vm_object_backing_collapse_wait(object);
1923 		if (backing_object == NULL)
1924 			return;
1925 
1926 		KASSERT(object->ref_count > 0 &&
1927 		    object->ref_count > object->shadow_count,
1928 		    ("collapse with invalid ref %d or shadow %d count.",
1929 		    object->ref_count, object->shadow_count));
1930 		KASSERT((backing_object->flags &
1931 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1932 		    ("vm_object_collapse: Backing object already collapsing."));
1933 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1934 		    ("vm_object_collapse: object is already collapsing."));
1935 
1936 		/*
1937 		 * We know that we can either collapse the backing object if
1938 		 * the parent is the only reference to it, or (perhaps) have
1939 		 * the parent bypass the object if the parent happens to shadow
1940 		 * all the resident pages in the entire backing object.
1941 		 */
1942 		if (backing_object->ref_count == 1) {
1943 			KASSERT(backing_object->shadow_count == 1,
1944 			    ("vm_object_collapse: shadow_count: %d",
1945 			    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 #if VM_NRESERVLEVEL > 0
1958 			/*
1959 			 * Break any reservations from backing_object.
1960 			 */
1961 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1962 				vm_reserv_break_all(backing_object);
1963 #endif
1964 
1965 			/*
1966 			 * Move the pager from backing_object to object.
1967 			 */
1968 			if ((backing_object->flags & OBJ_SWAP) != 0) {
1969 				/*
1970 				 * swap_pager_copy() can sleep, in which case
1971 				 * the backing_object's and object's locks are
1972 				 * released and reacquired.
1973 				 * Since swap_pager_copy() is being asked to
1974 				 * destroy backing_object, it will change the
1975 				 * type to OBJT_DEFAULT.
1976 				 */
1977 				swap_pager_copy(
1978 				    backing_object,
1979 				    object,
1980 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1981 			}
1982 
1983 			/*
1984 			 * Object now shadows whatever backing_object did.
1985 			 */
1986 			vm_object_clear_flag(object, OBJ_COLLAPSING);
1987 			vm_object_backing_transfer(object, backing_object);
1988 			object->backing_object_offset +=
1989 			    backing_object->backing_object_offset;
1990 			VM_OBJECT_WUNLOCK(object);
1991 			vm_object_pip_wakeup(object);
1992 
1993 			/*
1994 			 * Discard backing_object.
1995 			 *
1996 			 * Since the backing object has no pages, no pager left,
1997 			 * and no object references within it, all that is
1998 			 * necessary is to dispose of it.
1999 			 */
2000 			KASSERT(backing_object->ref_count == 1, (
2001 "backing_object %p was somehow re-referenced during collapse!",
2002 			    backing_object));
2003 			vm_object_pip_wakeup(backing_object);
2004 			(void)refcount_release(&backing_object->ref_count);
2005 			vm_object_terminate(backing_object);
2006 			counter_u64_add(object_collapses, 1);
2007 			VM_OBJECT_WLOCK(object);
2008 		} else {
2009 			/*
2010 			 * If we do not entirely shadow the backing object,
2011 			 * there is nothing we can do so we give up.
2012 			 *
2013 			 * The object lock and backing_object lock must not
2014 			 * be dropped during this sequence.
2015 			 */
2016 			if (!vm_object_scan_all_shadowed(object)) {
2017 				VM_OBJECT_WUNLOCK(backing_object);
2018 				break;
2019 			}
2020 
2021 			/*
2022 			 * Make the parent shadow the next object in the
2023 			 * chain.  Deallocating backing_object will not remove
2024 			 * it, since its reference count is at least 2.
2025 			 */
2026 			vm_object_backing_remove_locked(object);
2027 			new_backing_object = backing_object->backing_object;
2028 			if (new_backing_object != NULL) {
2029 				vm_object_backing_insert_ref(object,
2030 				    new_backing_object);
2031 				object->backing_object_offset +=
2032 				    backing_object->backing_object_offset;
2033 			}
2034 
2035 			/*
2036 			 * Drop the reference count on backing_object. Since
2037 			 * its ref_count was at least 2, it will not vanish.
2038 			 */
2039 			(void)refcount_release(&backing_object->ref_count);
2040 			KASSERT(backing_object->ref_count >= 1, (
2041 "backing_object %p was somehow dereferenced during collapse!",
2042 			    backing_object));
2043 			VM_OBJECT_WUNLOCK(backing_object);
2044 			counter_u64_add(object_bypasses, 1);
2045 		}
2046 
2047 		/*
2048 		 * Try again with this object's new backing object.
2049 		 */
2050 	}
2051 }
2052 
2053 /*
2054  *	vm_object_page_remove:
2055  *
2056  *	For the given object, either frees or invalidates each of the
2057  *	specified pages.  In general, a page is freed.  However, if a page is
2058  *	wired for any reason other than the existence of a managed, wired
2059  *	mapping, then it may be invalidated but not removed from the object.
2060  *	Pages are specified by the given range ["start", "end") and the option
2061  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
2062  *	extends from "start" to the end of the object.  If the option
2063  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2064  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
2065  *	specified, then the pages within the specified range must have no
2066  *	mappings.  Otherwise, if this option is not specified, any mappings to
2067  *	the specified pages are removed before the pages are freed or
2068  *	invalidated.
2069  *
2070  *	In general, this operation should only be performed on objects that
2071  *	contain managed pages.  There are, however, two exceptions.  First, it
2072  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
2073  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2074  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
2075  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
2076  *
2077  *	The object must be locked.
2078  */
2079 void
2080 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2081     int options)
2082 {
2083 	vm_page_t p, next;
2084 
2085 	VM_OBJECT_ASSERT_WLOCKED(object);
2086 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2087 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2088 	    ("vm_object_page_remove: illegal options for object %p", object));
2089 	if (object->resident_page_count == 0)
2090 		return;
2091 	vm_object_pip_add(object, 1);
2092 again:
2093 	p = vm_page_find_least(object, start);
2094 
2095 	/*
2096 	 * Here, the variable "p" is either (1) the page with the least pindex
2097 	 * greater than or equal to the parameter "start" or (2) NULL.
2098 	 */
2099 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2100 		next = TAILQ_NEXT(p, listq);
2101 
2102 		/*
2103 		 * Skip invalid pages if asked to do so.  Try to avoid acquiring
2104 		 * the busy lock, as some consumers rely on this to avoid
2105 		 * deadlocks.
2106 		 *
2107 		 * A thread may concurrently transition the page from invalid to
2108 		 * valid using only the busy lock, so the result of this check
2109 		 * is immediately stale.  It is up to consumers to handle this,
2110 		 * for instance by ensuring that all invalid->valid transitions
2111 		 * happen with a mutex held, as may be possible for a
2112 		 * filesystem.
2113 		 */
2114 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2115 			continue;
2116 
2117 		/*
2118 		 * If the page is wired for any reason besides the existence
2119 		 * of managed, wired mappings, then it cannot be freed.  For
2120 		 * example, fictitious pages, which represent device memory,
2121 		 * are inherently wired and cannot be freed.  They can,
2122 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2123 		 * not specified.
2124 		 */
2125 		if (vm_page_tryxbusy(p) == 0) {
2126 			if (vm_page_busy_sleep(p, "vmopar", 0))
2127 				VM_OBJECT_WLOCK(object);
2128 			goto again;
2129 		}
2130 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2131 			vm_page_xunbusy(p);
2132 			continue;
2133 		}
2134 		if (vm_page_wired(p)) {
2135 wired:
2136 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2137 			    object->ref_count != 0)
2138 				pmap_remove_all(p);
2139 			if ((options & OBJPR_CLEANONLY) == 0) {
2140 				vm_page_invalid(p);
2141 				vm_page_undirty(p);
2142 			}
2143 			vm_page_xunbusy(p);
2144 			continue;
2145 		}
2146 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2147 		    ("vm_object_page_remove: page %p is fictitious", p));
2148 		if ((options & OBJPR_CLEANONLY) != 0 &&
2149 		    !vm_page_none_valid(p)) {
2150 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2151 			    object->ref_count != 0 &&
2152 			    !vm_page_try_remove_write(p))
2153 				goto wired;
2154 			if (p->dirty != 0) {
2155 				vm_page_xunbusy(p);
2156 				continue;
2157 			}
2158 		}
2159 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2160 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2161 			goto wired;
2162 		vm_page_free(p);
2163 	}
2164 	vm_object_pip_wakeup(object);
2165 
2166 	vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2167 	    start);
2168 }
2169 
2170 /*
2171  *	vm_object_page_noreuse:
2172  *
2173  *	For the given object, attempt to move the specified pages to
2174  *	the head of the inactive queue.  This bypasses regular LRU
2175  *	operation and allows the pages to be reused quickly under memory
2176  *	pressure.  If a page is wired for any reason, then it will not
2177  *	be queued.  Pages are specified by the range ["start", "end").
2178  *	As a special case, if "end" is zero, then the range extends from
2179  *	"start" to the end of the object.
2180  *
2181  *	This operation should only be performed on objects that
2182  *	contain non-fictitious, managed pages.
2183  *
2184  *	The object must be locked.
2185  */
2186 void
2187 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2188 {
2189 	vm_page_t p, next;
2190 
2191 	VM_OBJECT_ASSERT_LOCKED(object);
2192 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2193 	    ("vm_object_page_noreuse: illegal object %p", object));
2194 	if (object->resident_page_count == 0)
2195 		return;
2196 	p = vm_page_find_least(object, start);
2197 
2198 	/*
2199 	 * Here, the variable "p" is either (1) the page with the least pindex
2200 	 * greater than or equal to the parameter "start" or (2) NULL.
2201 	 */
2202 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2203 		next = TAILQ_NEXT(p, listq);
2204 		vm_page_deactivate_noreuse(p);
2205 	}
2206 }
2207 
2208 /*
2209  *	Populate the specified range of the object with valid pages.  Returns
2210  *	TRUE if the range is successfully populated and FALSE otherwise.
2211  *
2212  *	Note: This function should be optimized to pass a larger array of
2213  *	pages to vm_pager_get_pages() before it is applied to a non-
2214  *	OBJT_DEVICE object.
2215  *
2216  *	The object must be locked.
2217  */
2218 boolean_t
2219 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2220 {
2221 	vm_page_t m;
2222 	vm_pindex_t pindex;
2223 	int rv;
2224 
2225 	VM_OBJECT_ASSERT_WLOCKED(object);
2226 	for (pindex = start; pindex < end; pindex++) {
2227 		rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2228 		if (rv != VM_PAGER_OK)
2229 			break;
2230 
2231 		/*
2232 		 * Keep "m" busy because a subsequent iteration may unlock
2233 		 * the object.
2234 		 */
2235 	}
2236 	if (pindex > start) {
2237 		m = vm_page_lookup(object, start);
2238 		while (m != NULL && m->pindex < pindex) {
2239 			vm_page_xunbusy(m);
2240 			m = TAILQ_NEXT(m, listq);
2241 		}
2242 	}
2243 	return (pindex == end);
2244 }
2245 
2246 /*
2247  *	Routine:	vm_object_coalesce
2248  *	Function:	Coalesces two objects backing up adjoining
2249  *			regions of memory into a single object.
2250  *
2251  *	returns TRUE if objects were combined.
2252  *
2253  *	NOTE:	Only works at the moment if the second object is NULL -
2254  *		if it's not, which object do we lock first?
2255  *
2256  *	Parameters:
2257  *		prev_object	First object to coalesce
2258  *		prev_offset	Offset into prev_object
2259  *		prev_size	Size of reference to prev_object
2260  *		next_size	Size of reference to the second object
2261  *		reserved	Indicator that extension region has
2262  *				swap accounted for
2263  *
2264  *	Conditions:
2265  *	The object must *not* be locked.
2266  */
2267 boolean_t
2268 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2269     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2270 {
2271 	vm_pindex_t next_pindex;
2272 
2273 	if (prev_object == NULL)
2274 		return (TRUE);
2275 	if ((prev_object->flags & OBJ_ANON) == 0)
2276 		return (FALSE);
2277 
2278 	VM_OBJECT_WLOCK(prev_object);
2279 	/*
2280 	 * Try to collapse the object first.
2281 	 */
2282 	vm_object_collapse(prev_object);
2283 
2284 	/*
2285 	 * Can't coalesce if: . more than one reference . paged out . shadows
2286 	 * another object . has a copy elsewhere (any of which mean that the
2287 	 * pages not mapped to prev_entry may be in use anyway)
2288 	 */
2289 	if (prev_object->backing_object != NULL) {
2290 		VM_OBJECT_WUNLOCK(prev_object);
2291 		return (FALSE);
2292 	}
2293 
2294 	prev_size >>= PAGE_SHIFT;
2295 	next_size >>= PAGE_SHIFT;
2296 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2297 
2298 	if (prev_object->ref_count > 1 &&
2299 	    prev_object->size != next_pindex &&
2300 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2301 		VM_OBJECT_WUNLOCK(prev_object);
2302 		return (FALSE);
2303 	}
2304 
2305 	/*
2306 	 * Account for the charge.
2307 	 */
2308 	if (prev_object->cred != NULL) {
2309 		/*
2310 		 * If prev_object was charged, then this mapping,
2311 		 * although not charged now, may become writable
2312 		 * later. Non-NULL cred in the object would prevent
2313 		 * swap reservation during enabling of the write
2314 		 * access, so reserve swap now. Failed reservation
2315 		 * cause allocation of the separate object for the map
2316 		 * entry, and swap reservation for this entry is
2317 		 * managed in appropriate time.
2318 		 */
2319 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2320 		    prev_object->cred)) {
2321 			VM_OBJECT_WUNLOCK(prev_object);
2322 			return (FALSE);
2323 		}
2324 		prev_object->charge += ptoa(next_size);
2325 	}
2326 
2327 	/*
2328 	 * Remove any pages that may still be in the object from a previous
2329 	 * deallocation.
2330 	 */
2331 	if (next_pindex < prev_object->size) {
2332 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2333 		    next_size, 0);
2334 #if 0
2335 		if (prev_object->cred != NULL) {
2336 			KASSERT(prev_object->charge >=
2337 			    ptoa(prev_object->size - next_pindex),
2338 			    ("object %p overcharged 1 %jx %jx", prev_object,
2339 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2340 			prev_object->charge -= ptoa(prev_object->size -
2341 			    next_pindex);
2342 		}
2343 #endif
2344 	}
2345 
2346 	/*
2347 	 * Extend the object if necessary.
2348 	 */
2349 	if (next_pindex + next_size > prev_object->size)
2350 		prev_object->size = next_pindex + next_size;
2351 
2352 	VM_OBJECT_WUNLOCK(prev_object);
2353 	return (TRUE);
2354 }
2355 
2356 void
2357 vm_object_set_writeable_dirty_(vm_object_t object)
2358 {
2359 	atomic_add_int(&object->generation, 1);
2360 }
2361 
2362 bool
2363 vm_object_mightbedirty_(vm_object_t object)
2364 {
2365 	return (object->generation != object->cleangeneration);
2366 }
2367 
2368 /*
2369  *	vm_object_unwire:
2370  *
2371  *	For each page offset within the specified range of the given object,
2372  *	find the highest-level page in the shadow chain and unwire it.  A page
2373  *	must exist at every page offset, and the highest-level page must be
2374  *	wired.
2375  */
2376 void
2377 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2378     uint8_t queue)
2379 {
2380 	vm_object_t tobject, t1object;
2381 	vm_page_t m, tm;
2382 	vm_pindex_t end_pindex, pindex, tpindex;
2383 	int depth, locked_depth;
2384 
2385 	KASSERT((offset & PAGE_MASK) == 0,
2386 	    ("vm_object_unwire: offset is not page aligned"));
2387 	KASSERT((length & PAGE_MASK) == 0,
2388 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2389 	/* The wired count of a fictitious page never changes. */
2390 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2391 		return;
2392 	pindex = OFF_TO_IDX(offset);
2393 	end_pindex = pindex + atop(length);
2394 again:
2395 	locked_depth = 1;
2396 	VM_OBJECT_RLOCK(object);
2397 	m = vm_page_find_least(object, pindex);
2398 	while (pindex < end_pindex) {
2399 		if (m == NULL || pindex < m->pindex) {
2400 			/*
2401 			 * The first object in the shadow chain doesn't
2402 			 * contain a page at the current index.  Therefore,
2403 			 * the page must exist in a backing object.
2404 			 */
2405 			tobject = object;
2406 			tpindex = pindex;
2407 			depth = 0;
2408 			do {
2409 				tpindex +=
2410 				    OFF_TO_IDX(tobject->backing_object_offset);
2411 				tobject = tobject->backing_object;
2412 				KASSERT(tobject != NULL,
2413 				    ("vm_object_unwire: missing page"));
2414 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2415 					goto next_page;
2416 				depth++;
2417 				if (depth == locked_depth) {
2418 					locked_depth++;
2419 					VM_OBJECT_RLOCK(tobject);
2420 				}
2421 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2422 			    NULL);
2423 		} else {
2424 			tm = m;
2425 			m = TAILQ_NEXT(m, listq);
2426 		}
2427 		if (vm_page_trysbusy(tm) == 0) {
2428 			for (tobject = object; locked_depth >= 1;
2429 			    locked_depth--) {
2430 				t1object = tobject->backing_object;
2431 				if (tm->object != tobject)
2432 					VM_OBJECT_RUNLOCK(tobject);
2433 				tobject = t1object;
2434 			}
2435 			tobject = tm->object;
2436 			if (!vm_page_busy_sleep(tm, "unwbo",
2437 			    VM_ALLOC_IGN_SBUSY))
2438 				VM_OBJECT_RUNLOCK(tobject);
2439 			goto again;
2440 		}
2441 		vm_page_unwire(tm, queue);
2442 		vm_page_sunbusy(tm);
2443 next_page:
2444 		pindex++;
2445 	}
2446 	/* Release the accumulated object locks. */
2447 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2448 		t1object = tobject->backing_object;
2449 		VM_OBJECT_RUNLOCK(tobject);
2450 		tobject = t1object;
2451 	}
2452 }
2453 
2454 /*
2455  * Return the vnode for the given object, or NULL if none exists.
2456  * For tmpfs objects, the function may return NULL if there is
2457  * no vnode allocated at the time of the call.
2458  */
2459 struct vnode *
2460 vm_object_vnode(vm_object_t object)
2461 {
2462 	struct vnode *vp;
2463 
2464 	VM_OBJECT_ASSERT_LOCKED(object);
2465 	vm_pager_getvp(object, &vp, NULL);
2466 	return (vp);
2467 }
2468 
2469 /*
2470  * Busy the vm object.  This prevents new pages belonging to the object from
2471  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2472  * for checking page state before proceeding.
2473  */
2474 void
2475 vm_object_busy(vm_object_t obj)
2476 {
2477 
2478 	VM_OBJECT_ASSERT_LOCKED(obj);
2479 
2480 	blockcount_acquire(&obj->busy, 1);
2481 	/* The fence is required to order loads of page busy. */
2482 	atomic_thread_fence_acq_rel();
2483 }
2484 
2485 void
2486 vm_object_unbusy(vm_object_t obj)
2487 {
2488 
2489 	blockcount_release(&obj->busy, 1);
2490 }
2491 
2492 void
2493 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2494 {
2495 
2496 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2497 
2498 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
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 	u_long sp;
2511 	int count, error;
2512 
2513 	if (req->oldptr == NULL) {
2514 		/*
2515 		 * If an old buffer has not been provided, generate an
2516 		 * estimate of the space needed for a subsequent call.
2517 		 */
2518 		mtx_lock(&vm_object_list_mtx);
2519 		count = 0;
2520 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2521 			if (obj->type == OBJT_DEAD)
2522 				continue;
2523 			count++;
2524 		}
2525 		mtx_unlock(&vm_object_list_mtx);
2526 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2527 		    count * 11 / 10));
2528 	}
2529 
2530 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK);
2531 	error = 0;
2532 
2533 	/*
2534 	 * VM objects are type stable and are never removed from the
2535 	 * list once added.  This allows us to safely read obj->object_list
2536 	 * after reacquiring the VM object lock.
2537 	 */
2538 	mtx_lock(&vm_object_list_mtx);
2539 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2540 		if (obj->type == OBJT_DEAD ||
2541 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2542 			continue;
2543 		VM_OBJECT_RLOCK(obj);
2544 		if (obj->type == OBJT_DEAD ||
2545 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2546 			VM_OBJECT_RUNLOCK(obj);
2547 			continue;
2548 		}
2549 		mtx_unlock(&vm_object_list_mtx);
2550 		kvo->kvo_size = ptoa(obj->size);
2551 		kvo->kvo_resident = obj->resident_page_count;
2552 		kvo->kvo_ref_count = obj->ref_count;
2553 		kvo->kvo_shadow_count = obj->shadow_count;
2554 		kvo->kvo_memattr = obj->memattr;
2555 		kvo->kvo_active = 0;
2556 		kvo->kvo_inactive = 0;
2557 		if (!swap_only) {
2558 			TAILQ_FOREACH(m, &obj->memq, listq) {
2559 				/*
2560 				 * A page may belong to the object but be
2561 				 * dequeued and set to PQ_NONE while the
2562 				 * object lock is not held.  This makes the
2563 				 * reads of m->queue below racy, and we do not
2564 				 * count pages set to PQ_NONE.  However, this
2565 				 * sysctl is only meant to give an
2566 				 * approximation of the system anyway.
2567 				 */
2568 				if (m->a.queue == PQ_ACTIVE)
2569 					kvo->kvo_active++;
2570 				else if (m->a.queue == PQ_INACTIVE)
2571 					kvo->kvo_inactive++;
2572 			}
2573 		}
2574 
2575 		kvo->kvo_vn_fileid = 0;
2576 		kvo->kvo_vn_fsid = 0;
2577 		kvo->kvo_vn_fsid_freebsd11 = 0;
2578 		freepath = NULL;
2579 		fullpath = "";
2580 		vp = NULL;
2581 		kvo->kvo_type = vm_object_kvme_type(obj, swap_only ? NULL : &vp);
2582 		if (vp != NULL) {
2583 			vref(vp);
2584 		} else if ((obj->flags & OBJ_ANON) != 0) {
2585 			MPASS(kvo->kvo_type == KVME_TYPE_DEFAULT ||
2586 			    kvo->kvo_type == KVME_TYPE_SWAP);
2587 			kvo->kvo_me = (uintptr_t)obj;
2588 			/* tmpfs objs are reported as vnodes */
2589 			kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2590 			sp = swap_pager_swapped_pages(obj);
2591 			kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2592 		}
2593 		VM_OBJECT_RUNLOCK(obj);
2594 		if (vp != NULL) {
2595 			vn_fullpath(vp, &fullpath, &freepath);
2596 			vn_lock(vp, LK_SHARED | LK_RETRY);
2597 			if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2598 				kvo->kvo_vn_fileid = va.va_fileid;
2599 				kvo->kvo_vn_fsid = va.va_fsid;
2600 				kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2601 								/* truncate */
2602 			}
2603 			vput(vp);
2604 		}
2605 
2606 		strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2607 		if (freepath != NULL)
2608 			free(freepath, M_TEMP);
2609 
2610 		/* Pack record size down */
2611 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2612 		    + strlen(kvo->kvo_path) + 1;
2613 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2614 		    sizeof(uint64_t));
2615 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2616 		maybe_yield();
2617 		mtx_lock(&vm_object_list_mtx);
2618 		if (error)
2619 			break;
2620 	}
2621 	mtx_unlock(&vm_object_list_mtx);
2622 	free(kvo, M_TEMP);
2623 	return (error);
2624 }
2625 
2626 static int
2627 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2628 {
2629 	return (vm_object_list_handler(req, false));
2630 }
2631 
2632 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2633     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2634     "List of VM objects");
2635 
2636 static int
2637 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2638 {
2639 	return (vm_object_list_handler(req, true));
2640 }
2641 
2642 /*
2643  * This sysctl returns list of the anonymous or swap objects. Intent
2644  * is to provide stripped optimized list useful to analyze swap use.
2645  * Since technically non-swap (default) objects participate in the
2646  * shadow chains, and are converted to swap type as needed by swap
2647  * pager, we must report them.
2648  */
2649 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2650     CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2651     sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2652     "List of swap VM objects");
2653 
2654 #include "opt_ddb.h"
2655 #ifdef DDB
2656 #include <sys/kernel.h>
2657 
2658 #include <sys/cons.h>
2659 
2660 #include <ddb/ddb.h>
2661 
2662 static int
2663 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2664 {
2665 	vm_map_t tmpm;
2666 	vm_map_entry_t tmpe;
2667 	vm_object_t obj;
2668 
2669 	if (map == 0)
2670 		return 0;
2671 
2672 	if (entry == 0) {
2673 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2674 			if (_vm_object_in_map(map, object, tmpe)) {
2675 				return 1;
2676 			}
2677 		}
2678 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2679 		tmpm = entry->object.sub_map;
2680 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2681 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2682 				return 1;
2683 			}
2684 		}
2685 	} else if ((obj = entry->object.vm_object) != NULL) {
2686 		for (; obj; obj = obj->backing_object)
2687 			if (obj == object) {
2688 				return 1;
2689 			}
2690 	}
2691 	return 0;
2692 }
2693 
2694 static int
2695 vm_object_in_map(vm_object_t object)
2696 {
2697 	struct proc *p;
2698 
2699 	/* sx_slock(&allproc_lock); */
2700 	FOREACH_PROC_IN_SYSTEM(p) {
2701 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2702 			continue;
2703 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2704 			/* sx_sunlock(&allproc_lock); */
2705 			return 1;
2706 		}
2707 	}
2708 	/* sx_sunlock(&allproc_lock); */
2709 	if (_vm_object_in_map(kernel_map, object, 0))
2710 		return 1;
2711 	return 0;
2712 }
2713 
2714 DB_SHOW_COMMAND(vmochk, vm_object_check)
2715 {
2716 	vm_object_t object;
2717 
2718 	/*
2719 	 * make sure that internal objs are in a map somewhere
2720 	 * and none have zero ref counts.
2721 	 */
2722 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2723 		if ((object->flags & OBJ_ANON) != 0) {
2724 			if (object->ref_count == 0) {
2725 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2726 					(long)object->size);
2727 			}
2728 			if (!vm_object_in_map(object)) {
2729 				db_printf(
2730 			"vmochk: internal obj is not in a map: "
2731 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2732 				    object->ref_count, (u_long)object->size,
2733 				    (u_long)object->size,
2734 				    (void *)object->backing_object);
2735 			}
2736 		}
2737 		if (db_pager_quit)
2738 			return;
2739 	}
2740 }
2741 
2742 /*
2743  *	vm_object_print:	[ debug ]
2744  */
2745 DB_SHOW_COMMAND(object, vm_object_print_static)
2746 {
2747 	/* XXX convert args. */
2748 	vm_object_t object = (vm_object_t)addr;
2749 	boolean_t full = have_addr;
2750 
2751 	vm_page_t p;
2752 
2753 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2754 #define	count	was_count
2755 
2756 	int count;
2757 
2758 	if (object == NULL)
2759 		return;
2760 
2761 	db_iprintf(
2762 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2763 	    object, (int)object->type, (uintmax_t)object->size,
2764 	    object->resident_page_count, object->ref_count, object->flags,
2765 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2766 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2767 	    object->shadow_count,
2768 	    object->backing_object ? object->backing_object->ref_count : 0,
2769 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2770 
2771 	if (!full)
2772 		return;
2773 
2774 	db_indent += 2;
2775 	count = 0;
2776 	TAILQ_FOREACH(p, &object->memq, listq) {
2777 		if (count == 0)
2778 			db_iprintf("memory:=");
2779 		else if (count == 6) {
2780 			db_printf("\n");
2781 			db_iprintf(" ...");
2782 			count = 0;
2783 		} else
2784 			db_printf(",");
2785 		count++;
2786 
2787 		db_printf("(off=0x%jx,page=0x%jx)",
2788 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2789 
2790 		if (db_pager_quit)
2791 			break;
2792 	}
2793 	if (count != 0)
2794 		db_printf("\n");
2795 	db_indent -= 2;
2796 }
2797 
2798 /* XXX. */
2799 #undef count
2800 
2801 /* XXX need this non-static entry for calling from vm_map_print. */
2802 void
2803 vm_object_print(
2804         /* db_expr_t */ long addr,
2805 	boolean_t have_addr,
2806 	/* db_expr_t */ long count,
2807 	char *modif)
2808 {
2809 	vm_object_print_static(addr, have_addr, count, modif);
2810 }
2811 
2812 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2813 {
2814 	vm_object_t object;
2815 	vm_pindex_t fidx;
2816 	vm_paddr_t pa;
2817 	vm_page_t m, prev_m;
2818 	int rcount, nl, c;
2819 
2820 	nl = 0;
2821 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2822 		db_printf("new object: %p\n", (void *)object);
2823 		if (nl > 18) {
2824 			c = cngetc();
2825 			if (c != ' ')
2826 				return;
2827 			nl = 0;
2828 		}
2829 		nl++;
2830 		rcount = 0;
2831 		fidx = 0;
2832 		pa = -1;
2833 		TAILQ_FOREACH(m, &object->memq, listq) {
2834 			if (m->pindex > 128)
2835 				break;
2836 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2837 			    prev_m->pindex + 1 != m->pindex) {
2838 				if (rcount) {
2839 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2840 						(long)fidx, rcount, (long)pa);
2841 					if (nl > 18) {
2842 						c = cngetc();
2843 						if (c != ' ')
2844 							return;
2845 						nl = 0;
2846 					}
2847 					nl++;
2848 					rcount = 0;
2849 				}
2850 			}
2851 			if (rcount &&
2852 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2853 				++rcount;
2854 				continue;
2855 			}
2856 			if (rcount) {
2857 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2858 					(long)fidx, rcount, (long)pa);
2859 				if (nl > 18) {
2860 					c = cngetc();
2861 					if (c != ' ')
2862 						return;
2863 					nl = 0;
2864 				}
2865 				nl++;
2866 			}
2867 			fidx = m->pindex;
2868 			pa = VM_PAGE_TO_PHYS(m);
2869 			rcount = 1;
2870 		}
2871 		if (rcount) {
2872 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2873 				(long)fidx, rcount, (long)pa);
2874 			if (nl > 18) {
2875 				c = cngetc();
2876 				if (c != ' ')
2877 					return;
2878 				nl = 0;
2879 			}
2880 			nl++;
2881 		}
2882 	}
2883 }
2884 #endif /* DDB */
2885