xref: /freebsd/sys/vm/vm_object.c (revision 59f5f100b774de8824fb2fc1a8a11a93bbc2dafd)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39  *
40  * Permission to use, copy, modify and distribute this software and
41  * its documentation is hereby granted, provided that both the copyright
42  * notice and this permission notice appear in all copies of the
43  * software, derivative works or modified versions, and any portions
44  * thereof, and that both notices appear in supporting documentation.
45  *
46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49  *
50  * Carnegie Mellon requests users of this software to return to
51  *
52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
53  *  School of Computer Science
54  *  Carnegie Mellon University
55  *  Pittsburgh PA 15213-3890
56  *
57  * any improvements or extensions that they make and grant Carnegie the
58  * rights to redistribute these changes.
59  */
60 
61 /*
62  *	Virtual memory object module.
63  */
64 
65 #include "opt_vm.h"
66 
67 #include <sys/systm.h>
68 #include <sys/blockcount.h>
69 #include <sys/cpuset.h>
70 #include <sys/jail.h>
71 #include <sys/limits.h>
72 #include <sys/lock.h>
73 #include <sys/mman.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/mutex.h>
77 #include <sys/pctrie.h>
78 #include <sys/proc.h>
79 #include <sys/refcount.h>
80 #include <sys/sx.h>
81 #include <sys/sysctl.h>
82 #include <sys/resourcevar.h>
83 #include <sys/refcount.h>
84 #include <sys/rwlock.h>
85 #include <sys/user.h>
86 #include <sys/vnode.h>
87 #include <sys/vmmeter.h>
88 
89 #include <vm/vm.h>
90 #include <vm/vm_param.h>
91 #include <vm/pmap.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_page.h>
95 #include <vm/vm_pageout.h>
96 #include <vm/vm_pager.h>
97 #include <vm/vm_phys.h>
98 #include <vm/vm_pagequeue.h>
99 #include <vm/swap_pager.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_extern.h>
102 #include <vm/vm_radix.h>
103 #include <vm/vm_reserv.h>
104 #include <vm/uma.h>
105 
106 static int old_msync;
107 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
108     "Use old (insecure) msync behavior");
109 
110 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
111 		    int pagerflags, int flags, boolean_t *allclean,
112 		    boolean_t *eio);
113 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
114 		    boolean_t *allclean);
115 static void	vm_object_backing_remove(vm_object_t object);
116 
117 /*
118  *	Virtual memory objects maintain the actual data
119  *	associated with allocated virtual memory.  A given
120  *	page of memory exists within exactly one object.
121  *
122  *	An object is only deallocated when all "references"
123  *	are given up.  Only one "reference" to a given
124  *	region of an object should be writeable.
125  *
126  *	Associated with each object is a list of all resident
127  *	memory pages belonging to that object; this list is
128  *	maintained by the "vm_page" module, and locked by the object's
129  *	lock.
130  *
131  *	Each object also records a "pager" routine which is
132  *	used to retrieve (and store) pages to the proper backing
133  *	storage.  In addition, objects may be backed by other
134  *	objects from which they were virtual-copied.
135  *
136  *	The only items within the object structure which are
137  *	modified after time of creation are:
138  *		reference count		locked by object's lock
139  *		pager routine		locked by object's lock
140  *
141  */
142 
143 struct object_q vm_object_list;
144 struct mtx vm_object_list_mtx;	/* lock for object list and count */
145 
146 struct vm_object kernel_object_store;
147 
148 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
149     "VM object stats");
150 
151 static COUNTER_U64_DEFINE_EARLY(object_collapses);
152 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
153     &object_collapses,
154     "VM object collapses");
155 
156 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
157 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
158     &object_bypasses,
159     "VM object bypasses");
160 
161 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
162 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
163     &object_collapse_waits,
164     "Number of sleeps for collapse");
165 
166 static uma_zone_t obj_zone;
167 
168 static int vm_object_zinit(void *mem, int size, int flags);
169 
170 #ifdef INVARIANTS
171 static void vm_object_zdtor(void *mem, int size, void *arg);
172 
173 static void
174 vm_object_zdtor(void *mem, int size, void *arg)
175 {
176 	vm_object_t object;
177 
178 	object = (vm_object_t)mem;
179 	KASSERT(object->ref_count == 0,
180 	    ("object %p ref_count = %d", object, object->ref_count));
181 	KASSERT(TAILQ_EMPTY(&object->memq),
182 	    ("object %p has resident pages in its memq", object));
183 	KASSERT(vm_radix_is_empty(&object->rtree),
184 	    ("object %p has resident pages in its trie", object));
185 #if VM_NRESERVLEVEL > 0
186 	KASSERT(LIST_EMPTY(&object->rvq),
187 	    ("object %p has reservations",
188 	    object));
189 #endif
190 	KASSERT(!vm_object_busied(object),
191 	    ("object %p busy = %d", object, blockcount_read(&object->busy)));
192 	KASSERT(object->resident_page_count == 0,
193 	    ("object %p resident_page_count = %d",
194 	    object, object->resident_page_count));
195 	KASSERT(atomic_load_int(&object->shadow_count) == 0,
196 	    ("object %p shadow_count = %d",
197 	    object, atomic_load_int(&object->shadow_count)));
198 	KASSERT(object->type == OBJT_DEAD,
199 	    ("object %p has non-dead type %d",
200 	    object, object->type));
201 	KASSERT(object->charge == 0 && object->cred == NULL,
202 	    ("object %p has non-zero charge %ju (%p)",
203 	    object, (uintmax_t)object->charge, object->cred));
204 }
205 #endif
206 
207 static int
208 vm_object_zinit(void *mem, int size, int flags)
209 {
210 	vm_object_t object;
211 
212 	object = (vm_object_t)mem;
213 	rw_init_flags(&object->lock, "vmobject", RW_DUPOK | RW_NEW);
214 
215 	/* These are true for any object that has been freed */
216 	object->type = OBJT_DEAD;
217 	vm_radix_init(&object->rtree);
218 	refcount_init(&object->ref_count, 0);
219 	blockcount_init(&object->paging_in_progress);
220 	blockcount_init(&object->busy);
221 	object->resident_page_count = 0;
222 	atomic_store_int(&object->shadow_count, 0);
223 	object->flags = OBJ_DEAD;
224 
225 	mtx_lock(&vm_object_list_mtx);
226 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
227 	mtx_unlock(&vm_object_list_mtx);
228 	return (0);
229 }
230 
231 static void
232 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
233     vm_object_t object, void *handle)
234 {
235 
236 	TAILQ_INIT(&object->memq);
237 	LIST_INIT(&object->shadow_head);
238 
239 	object->type = type;
240 	object->flags = flags;
241 	if ((flags & OBJ_SWAP) != 0) {
242 		pctrie_init(&object->un_pager.swp.swp_blks);
243 		object->un_pager.swp.writemappings = 0;
244 	}
245 
246 	/*
247 	 * Ensure that swap_pager_swapoff() iteration over object_list
248 	 * sees up to date type and pctrie head if it observed
249 	 * non-dead object.
250 	 */
251 	atomic_thread_fence_rel();
252 
253 	object->pg_color = 0;
254 	object->size = size;
255 	object->domain.dr_policy = NULL;
256 	object->generation = 1;
257 	object->cleangeneration = 1;
258 	refcount_init(&object->ref_count, 1);
259 	object->memattr = VM_MEMATTR_DEFAULT;
260 	object->cred = NULL;
261 	object->charge = 0;
262 	object->handle = handle;
263 	object->backing_object = NULL;
264 	object->backing_object_offset = (vm_ooffset_t) 0;
265 #if VM_NRESERVLEVEL > 0
266 	LIST_INIT(&object->rvq);
267 #endif
268 	umtx_shm_object_init(object);
269 }
270 
271 /*
272  *	vm_object_init:
273  *
274  *	Initialize the VM objects module.
275  */
276 void
277 vm_object_init(void)
278 {
279 	TAILQ_INIT(&vm_object_list);
280 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
281 
282 	rw_init(&kernel_object->lock, "kernel vm object");
283 	vm_radix_init(&kernel_object->rtree);
284 	_vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
285 	    VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
286 #if VM_NRESERVLEVEL > 0
287 	kernel_object->flags |= OBJ_COLORED;
288 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
289 #endif
290 	kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
291 
292 	/*
293 	 * The lock portion of struct vm_object must be type stable due
294 	 * to vm_pageout_fallback_object_lock locking a vm object
295 	 * without holding any references to it.
296 	 *
297 	 * paging_in_progress is valid always.  Lockless references to
298 	 * the objects may acquire pip and then check OBJ_DEAD.
299 	 */
300 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
301 #ifdef INVARIANTS
302 	    vm_object_zdtor,
303 #else
304 	    NULL,
305 #endif
306 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
307 
308 	vm_radix_zinit();
309 }
310 
311 void
312 vm_object_clear_flag(vm_object_t object, u_short bits)
313 {
314 
315 	VM_OBJECT_ASSERT_WLOCKED(object);
316 	object->flags &= ~bits;
317 }
318 
319 /*
320  *	Sets the default memory attribute for the specified object.  Pages
321  *	that are allocated to this object are by default assigned this memory
322  *	attribute.
323  *
324  *	Presently, this function must be called before any pages are allocated
325  *	to the object.  In the future, this requirement may be relaxed for
326  *	"default" and "swap" objects.
327  */
328 int
329 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
330 {
331 
332 	VM_OBJECT_ASSERT_WLOCKED(object);
333 
334 	if (object->type == OBJT_DEAD)
335 		return (KERN_INVALID_ARGUMENT);
336 	if (!TAILQ_EMPTY(&object->memq))
337 		return (KERN_FAILURE);
338 
339 	object->memattr = memattr;
340 	return (KERN_SUCCESS);
341 }
342 
343 void
344 vm_object_pip_add(vm_object_t object, short i)
345 {
346 
347 	if (i > 0)
348 		blockcount_acquire(&object->paging_in_progress, i);
349 }
350 
351 void
352 vm_object_pip_wakeup(vm_object_t object)
353 {
354 
355 	vm_object_pip_wakeupn(object, 1);
356 }
357 
358 void
359 vm_object_pip_wakeupn(vm_object_t object, short i)
360 {
361 
362 	if (i > 0)
363 		blockcount_release(&object->paging_in_progress, i);
364 }
365 
366 /*
367  * Atomically drop the object lock and wait for pip to drain.  This protects
368  * from sleep/wakeup races due to identity changes.  The lock is not re-acquired
369  * on return.
370  */
371 static void
372 vm_object_pip_sleep(vm_object_t object, const char *waitid)
373 {
374 
375 	(void)blockcount_sleep(&object->paging_in_progress, &object->lock,
376 	    waitid, PVM | PDROP);
377 }
378 
379 void
380 vm_object_pip_wait(vm_object_t object, const char *waitid)
381 {
382 
383 	VM_OBJECT_ASSERT_WLOCKED(object);
384 
385 	blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
386 	    PVM);
387 }
388 
389 void
390 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
391 {
392 
393 	VM_OBJECT_ASSERT_UNLOCKED(object);
394 
395 	blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
396 }
397 
398 /*
399  *	vm_object_allocate:
400  *
401  *	Returns a new object with the given size.
402  */
403 vm_object_t
404 vm_object_allocate(objtype_t type, vm_pindex_t size)
405 {
406 	vm_object_t object;
407 	u_short flags;
408 
409 	switch (type) {
410 	case OBJT_DEAD:
411 		panic("vm_object_allocate: can't create OBJT_DEAD");
412 	case OBJT_SWAP:
413 		flags = OBJ_COLORED | OBJ_SWAP;
414 		break;
415 	case OBJT_DEVICE:
416 	case OBJT_SG:
417 		flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
418 		break;
419 	case OBJT_MGTDEVICE:
420 		flags = OBJ_FICTITIOUS;
421 		break;
422 	case OBJT_PHYS:
423 		flags = OBJ_UNMANAGED;
424 		break;
425 	case OBJT_VNODE:
426 		flags = 0;
427 		break;
428 	default:
429 		panic("vm_object_allocate: type %d is undefined or dynamic",
430 		    type);
431 	}
432 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
433 	_vm_object_allocate(type, size, flags, object, NULL);
434 
435 	return (object);
436 }
437 
438 vm_object_t
439 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags)
440 {
441 	vm_object_t object;
442 
443 	MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */);
444 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
445 	_vm_object_allocate(dyntype, size, flags, object, NULL);
446 
447 	return (object);
448 }
449 
450 /*
451  *	vm_object_allocate_anon:
452  *
453  *	Returns a new default object of the given size and marked as
454  *	anonymous memory for special split/collapse handling.  Color
455  *	to be initialized by the caller.
456  */
457 vm_object_t
458 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
459     struct ucred *cred, vm_size_t charge)
460 {
461 	vm_object_t handle, object;
462 
463 	if (backing_object == NULL)
464 		handle = NULL;
465 	else if ((backing_object->flags & OBJ_ANON) != 0)
466 		handle = backing_object->handle;
467 	else
468 		handle = backing_object;
469 	object = uma_zalloc(obj_zone, M_WAITOK);
470 	_vm_object_allocate(OBJT_SWAP, size,
471 	    OBJ_ANON | OBJ_ONEMAPPING | OBJ_SWAP, object, handle);
472 	object->cred = cred;
473 	object->charge = cred != NULL ? charge : 0;
474 	return (object);
475 }
476 
477 static void
478 vm_object_reference_vnode(vm_object_t object)
479 {
480 	u_int old;
481 
482 	/*
483 	 * vnode objects need the lock for the first reference
484 	 * to serialize with vnode_object_deallocate().
485 	 */
486 	if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
487 		VM_OBJECT_RLOCK(object);
488 		old = refcount_acquire(&object->ref_count);
489 		if (object->type == OBJT_VNODE && old == 0)
490 			vref(object->handle);
491 		VM_OBJECT_RUNLOCK(object);
492 	}
493 }
494 
495 /*
496  *	vm_object_reference:
497  *
498  *	Acquires a reference to the given object.
499  */
500 void
501 vm_object_reference(vm_object_t object)
502 {
503 
504 	if (object == NULL)
505 		return;
506 
507 	if (object->type == OBJT_VNODE)
508 		vm_object_reference_vnode(object);
509 	else
510 		refcount_acquire(&object->ref_count);
511 	KASSERT((object->flags & OBJ_DEAD) == 0,
512 	    ("vm_object_reference: Referenced dead object."));
513 }
514 
515 /*
516  *	vm_object_reference_locked:
517  *
518  *	Gets another reference to the given object.
519  *
520  *	The object must be locked.
521  */
522 void
523 vm_object_reference_locked(vm_object_t object)
524 {
525 	u_int old;
526 
527 	VM_OBJECT_ASSERT_LOCKED(object);
528 	old = refcount_acquire(&object->ref_count);
529 	if (object->type == OBJT_VNODE && old == 0)
530 		vref(object->handle);
531 	KASSERT((object->flags & OBJ_DEAD) == 0,
532 	    ("vm_object_reference: Referenced dead object."));
533 }
534 
535 /*
536  * Handle deallocating an object of type OBJT_VNODE.
537  */
538 static void
539 vm_object_deallocate_vnode(vm_object_t object)
540 {
541 	struct vnode *vp = (struct vnode *) object->handle;
542 	bool last;
543 
544 	KASSERT(object->type == OBJT_VNODE,
545 	    ("vm_object_deallocate_vnode: not a vnode object"));
546 	KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
547 
548 	/* Object lock to protect handle lookup. */
549 	last = refcount_release(&object->ref_count);
550 	VM_OBJECT_RUNLOCK(object);
551 
552 	if (!last)
553 		return;
554 
555 	if (!umtx_shm_vnobj_persistent)
556 		umtx_shm_object_terminated(object);
557 
558 	/* vrele may need the vnode lock. */
559 	vrele(vp);
560 }
561 
562 /*
563  * We dropped a reference on an object and discovered that it had a
564  * single remaining shadow.  This is a sibling of the reference we
565  * dropped.  Attempt to collapse the sibling and backing object.
566  */
567 static vm_object_t
568 vm_object_deallocate_anon(vm_object_t backing_object)
569 {
570 	vm_object_t object;
571 
572 	/* Fetch the final shadow.  */
573 	object = LIST_FIRST(&backing_object->shadow_head);
574 	KASSERT(object != NULL &&
575 	    atomic_load_int(&backing_object->shadow_count) == 1,
576 	    ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
577 	    backing_object->ref_count,
578 	    atomic_load_int(&backing_object->shadow_count)));
579 	KASSERT((object->flags & OBJ_ANON) != 0,
580 	    ("invalid shadow object %p", object));
581 
582 	if (!VM_OBJECT_TRYWLOCK(object)) {
583 		/*
584 		 * Prevent object from disappearing since we do not have a
585 		 * reference.
586 		 */
587 		vm_object_pip_add(object, 1);
588 		VM_OBJECT_WUNLOCK(backing_object);
589 		VM_OBJECT_WLOCK(object);
590 		vm_object_pip_wakeup(object);
591 	} else
592 		VM_OBJECT_WUNLOCK(backing_object);
593 
594 	/*
595 	 * Check for a collapse/terminate race with the last reference holder.
596 	 */
597 	if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
598 	    !refcount_acquire_if_not_zero(&object->ref_count)) {
599 		VM_OBJECT_WUNLOCK(object);
600 		return (NULL);
601 	}
602 	backing_object = object->backing_object;
603 	if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
604 		vm_object_collapse(object);
605 	VM_OBJECT_WUNLOCK(object);
606 
607 	return (object);
608 }
609 
610 /*
611  *	vm_object_deallocate:
612  *
613  *	Release a reference to the specified object,
614  *	gained either through a vm_object_allocate
615  *	or a vm_object_reference call.  When all references
616  *	are gone, storage associated with this object
617  *	may be relinquished.
618  *
619  *	No object may be locked.
620  */
621 void
622 vm_object_deallocate(vm_object_t object)
623 {
624 	vm_object_t temp;
625 	bool released;
626 
627 	while (object != NULL) {
628 		/*
629 		 * If the reference count goes to 0 we start calling
630 		 * vm_object_terminate() on the object chain.  A ref count
631 		 * of 1 may be a special case depending on the shadow count
632 		 * being 0 or 1.  These cases require a write lock on the
633 		 * object.
634 		 */
635 		if ((object->flags & OBJ_ANON) == 0)
636 			released = refcount_release_if_gt(&object->ref_count, 1);
637 		else
638 			released = refcount_release_if_gt(&object->ref_count, 2);
639 		if (released)
640 			return;
641 
642 		if (object->type == OBJT_VNODE) {
643 			VM_OBJECT_RLOCK(object);
644 			if (object->type == OBJT_VNODE) {
645 				vm_object_deallocate_vnode(object);
646 				return;
647 			}
648 			VM_OBJECT_RUNLOCK(object);
649 		}
650 
651 		VM_OBJECT_WLOCK(object);
652 		KASSERT(object->ref_count > 0,
653 		    ("vm_object_deallocate: object deallocated too many times: %d",
654 		    object->type));
655 
656 		/*
657 		 * If this is not the final reference to an anonymous
658 		 * object we may need to collapse the shadow chain.
659 		 */
660 		if (!refcount_release(&object->ref_count)) {
661 			if (object->ref_count > 1 ||
662 			    atomic_load_int(&object->shadow_count) == 0) {
663 				if ((object->flags & OBJ_ANON) != 0 &&
664 				    object->ref_count == 1)
665 					vm_object_set_flag(object,
666 					    OBJ_ONEMAPPING);
667 				VM_OBJECT_WUNLOCK(object);
668 				return;
669 			}
670 
671 			/* Handle collapsing last ref on anonymous objects. */
672 			object = vm_object_deallocate_anon(object);
673 			continue;
674 		}
675 
676 		/*
677 		 * Handle the final reference to an object.  We restart
678 		 * the loop with the backing object to avoid recursion.
679 		 */
680 		umtx_shm_object_terminated(object);
681 		temp = object->backing_object;
682 		if (temp != NULL) {
683 			KASSERT(object->type == OBJT_SWAP,
684 			    ("shadowed tmpfs v_object 2 %p", object));
685 			vm_object_backing_remove(object);
686 		}
687 
688 		KASSERT((object->flags & OBJ_DEAD) == 0,
689 		    ("vm_object_deallocate: Terminating dead object."));
690 		vm_object_set_flag(object, OBJ_DEAD);
691 		vm_object_terminate(object);
692 		object = temp;
693 	}
694 }
695 
696 void
697 vm_object_destroy(vm_object_t object)
698 {
699 	uma_zfree(obj_zone, object);
700 }
701 
702 static void
703 vm_object_sub_shadow(vm_object_t object)
704 {
705 	KASSERT(object->shadow_count >= 1,
706 	    ("object %p sub_shadow count zero", object));
707 	atomic_subtract_int(&object->shadow_count, 1);
708 }
709 
710 static void
711 vm_object_backing_remove_locked(vm_object_t object)
712 {
713 	vm_object_t backing_object;
714 
715 	backing_object = object->backing_object;
716 	VM_OBJECT_ASSERT_WLOCKED(object);
717 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
718 
719 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
720 	    ("vm_object_backing_remove: Removing collapsing object."));
721 
722 	vm_object_sub_shadow(backing_object);
723 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
724 		LIST_REMOVE(object, shadow_list);
725 		vm_object_clear_flag(object, OBJ_SHADOWLIST);
726 	}
727 	object->backing_object = NULL;
728 }
729 
730 static void
731 vm_object_backing_remove(vm_object_t object)
732 {
733 	vm_object_t backing_object;
734 
735 	VM_OBJECT_ASSERT_WLOCKED(object);
736 
737 	backing_object = object->backing_object;
738 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
739 		VM_OBJECT_WLOCK(backing_object);
740 		vm_object_backing_remove_locked(object);
741 		VM_OBJECT_WUNLOCK(backing_object);
742 	} else {
743 		object->backing_object = NULL;
744 		vm_object_sub_shadow(backing_object);
745 	}
746 }
747 
748 static void
749 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
750 {
751 
752 	VM_OBJECT_ASSERT_WLOCKED(object);
753 
754 	atomic_add_int(&backing_object->shadow_count, 1);
755 	if ((backing_object->flags & OBJ_ANON) != 0) {
756 		VM_OBJECT_ASSERT_WLOCKED(backing_object);
757 		LIST_INSERT_HEAD(&backing_object->shadow_head, object,
758 		    shadow_list);
759 		vm_object_set_flag(object, OBJ_SHADOWLIST);
760 	}
761 	object->backing_object = backing_object;
762 }
763 
764 static void
765 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
766 {
767 
768 	VM_OBJECT_ASSERT_WLOCKED(object);
769 
770 	if ((backing_object->flags & OBJ_ANON) != 0) {
771 		VM_OBJECT_WLOCK(backing_object);
772 		vm_object_backing_insert_locked(object, backing_object);
773 		VM_OBJECT_WUNLOCK(backing_object);
774 	} else {
775 		object->backing_object = backing_object;
776 		atomic_add_int(&backing_object->shadow_count, 1);
777 	}
778 }
779 
780 /*
781  * Insert an object into a backing_object's shadow list with an additional
782  * reference to the backing_object added.
783  */
784 static void
785 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
786 {
787 
788 	VM_OBJECT_ASSERT_WLOCKED(object);
789 
790 	if ((backing_object->flags & OBJ_ANON) != 0) {
791 		VM_OBJECT_WLOCK(backing_object);
792 		KASSERT((backing_object->flags & OBJ_DEAD) == 0,
793 		    ("shadowing dead anonymous object"));
794 		vm_object_reference_locked(backing_object);
795 		vm_object_backing_insert_locked(object, backing_object);
796 		vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
797 		VM_OBJECT_WUNLOCK(backing_object);
798 	} else {
799 		vm_object_reference(backing_object);
800 		atomic_add_int(&backing_object->shadow_count, 1);
801 		object->backing_object = backing_object;
802 	}
803 }
804 
805 /*
806  * Transfer a backing reference from backing_object to object.
807  */
808 static void
809 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
810 {
811 	vm_object_t new_backing_object;
812 
813 	/*
814 	 * Note that the reference to backing_object->backing_object
815 	 * moves from within backing_object to within object.
816 	 */
817 	vm_object_backing_remove_locked(object);
818 	new_backing_object = backing_object->backing_object;
819 	if (new_backing_object == NULL)
820 		return;
821 	if ((new_backing_object->flags & OBJ_ANON) != 0) {
822 		VM_OBJECT_WLOCK(new_backing_object);
823 		vm_object_backing_remove_locked(backing_object);
824 		vm_object_backing_insert_locked(object, new_backing_object);
825 		VM_OBJECT_WUNLOCK(new_backing_object);
826 	} else {
827 		/*
828 		 * shadow_count for new_backing_object is left
829 		 * unchanged, its reference provided by backing_object
830 		 * is replaced by object.
831 		 */
832 		object->backing_object = new_backing_object;
833 		backing_object->backing_object = NULL;
834 	}
835 }
836 
837 /*
838  * Wait for a concurrent collapse to settle.
839  */
840 static void
841 vm_object_collapse_wait(vm_object_t object)
842 {
843 
844 	VM_OBJECT_ASSERT_WLOCKED(object);
845 
846 	while ((object->flags & OBJ_COLLAPSING) != 0) {
847 		vm_object_pip_wait(object, "vmcolwait");
848 		counter_u64_add(object_collapse_waits, 1);
849 	}
850 }
851 
852 /*
853  * Waits for a backing object to clear a pending collapse and returns
854  * it locked if it is an ANON object.
855  */
856 static vm_object_t
857 vm_object_backing_collapse_wait(vm_object_t object)
858 {
859 	vm_object_t backing_object;
860 
861 	VM_OBJECT_ASSERT_WLOCKED(object);
862 
863 	for (;;) {
864 		backing_object = object->backing_object;
865 		if (backing_object == NULL ||
866 		    (backing_object->flags & OBJ_ANON) == 0)
867 			return (NULL);
868 		VM_OBJECT_WLOCK(backing_object);
869 		if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
870 			break;
871 		VM_OBJECT_WUNLOCK(object);
872 		vm_object_pip_sleep(backing_object, "vmbckwait");
873 		counter_u64_add(object_collapse_waits, 1);
874 		VM_OBJECT_WLOCK(object);
875 	}
876 	return (backing_object);
877 }
878 
879 /*
880  *	vm_object_terminate_single_page removes a pageable page from the object,
881  *	and removes it from the paging queues and frees it, if it is not wired.
882  *	It is invoked via callback from vm_object_terminate_pages.
883  */
884 static void
885 vm_object_terminate_single_page(vm_page_t p, void *objectv)
886 {
887 	vm_object_t object __diagused = objectv;
888 
889 	vm_page_assert_unbusied(p);
890 	KASSERT(p->object == object &&
891 	    (p->ref_count & VPRC_OBJREF) != 0,
892 	    ("%s: page %p is inconsistent", __func__, p));
893 	p->object = NULL;
894 	if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
895 		VM_CNT_INC(v_pfree);
896 		vm_page_free(p);
897 	}
898 }
899 
900 /*
901  *	vm_object_terminate_pages removes any remaining pageable pages
902  *	from the object and resets the object to an empty state.
903  */
904 static void
905 vm_object_terminate_pages(vm_object_t object)
906 {
907 	VM_OBJECT_ASSERT_WLOCKED(object);
908 
909 	/*
910 	 * If the object contained any pages, then reset it to an empty state.
911 	 * Rather than incrementally removing each page from the object, the
912 	 * page and object are reset to any empty state.
913 	 */
914 	if (object->resident_page_count == 0)
915 		return;
916 
917 	vm_radix_reclaim_callback(&object->rtree,
918 	    vm_object_terminate_single_page, object);
919 	TAILQ_INIT(&object->memq);
920 	object->resident_page_count = 0;
921 	if (object->type == OBJT_VNODE)
922 		vdrop(object->handle);
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->flags & OBJ_SWAP) != 0,
966 	    ("%s: non-swap obj %p has cred", __func__, object));
967 
968 	/*
969 	 * Let the pager know object is dead.
970 	 */
971 	vm_pager_deallocate(object);
972 	VM_OBJECT_WUNLOCK(object);
973 
974 	vm_object_destroy(object);
975 }
976 
977 /*
978  * Make the page read-only so that we can clear the object flags.  However, if
979  * this is a nosync mmap then the object is likely to stay dirty so do not
980  * mess with the page and do not clear the object flags.  Returns TRUE if the
981  * page should be flushed, and FALSE otherwise.
982  */
983 static boolean_t
984 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
985 {
986 
987 	vm_page_assert_busied(p);
988 
989 	/*
990 	 * If we have been asked to skip nosync pages and this is a
991 	 * nosync page, skip it.  Note that the object flags were not
992 	 * cleared in this case so we do not have to set them.
993 	 */
994 	if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
995 		*allclean = FALSE;
996 		return (FALSE);
997 	} else {
998 		pmap_remove_write(p);
999 		return (p->dirty != 0);
1000 	}
1001 }
1002 
1003 /*
1004  *	vm_object_page_clean
1005  *
1006  *	Clean all dirty pages in the specified range of object.  Leaves page
1007  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
1008  *	write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1009  *	leaving the object dirty.
1010  *
1011  *	For swap objects backing tmpfs regular files, do not flush anything,
1012  *	but remove write protection on the mapped pages to update mtime through
1013  *	mmaped writes.
1014  *
1015  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
1016  *	synchronous clustering mode implementation.
1017  *
1018  *	Odd semantics: if start == end, we clean everything.
1019  *
1020  *	The object must be locked.
1021  *
1022  *	Returns FALSE if some page from the range was not written, as
1023  *	reported by the pager, and TRUE otherwise.
1024  */
1025 boolean_t
1026 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1027     int flags)
1028 {
1029 	vm_page_t np, p;
1030 	vm_pindex_t pi, tend, tstart;
1031 	int curgeneration, n, pagerflags;
1032 	boolean_t eio, res, allclean;
1033 
1034 	VM_OBJECT_ASSERT_WLOCKED(object);
1035 
1036 	if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1037 		return (TRUE);
1038 
1039 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1040 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1041 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1042 
1043 	tstart = OFF_TO_IDX(start);
1044 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1045 	allclean = tstart == 0 && tend >= object->size;
1046 	res = TRUE;
1047 
1048 rescan:
1049 	curgeneration = object->generation;
1050 
1051 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1052 		pi = p->pindex;
1053 		if (pi >= tend)
1054 			break;
1055 		np = TAILQ_NEXT(p, listq);
1056 		if (vm_page_none_valid(p))
1057 			continue;
1058 		if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1059 			if (object->generation != curgeneration &&
1060 			    (flags & OBJPC_SYNC) != 0)
1061 				goto rescan;
1062 			np = vm_page_find_least(object, pi);
1063 			continue;
1064 		}
1065 		if (!vm_object_page_remove_write(p, flags, &allclean)) {
1066 			vm_page_xunbusy(p);
1067 			continue;
1068 		}
1069 		if (object->type == OBJT_VNODE) {
1070 			n = vm_object_page_collect_flush(object, p, pagerflags,
1071 			    flags, &allclean, &eio);
1072 			if (eio) {
1073 				res = FALSE;
1074 				allclean = FALSE;
1075 			}
1076 			if (object->generation != curgeneration &&
1077 			    (flags & OBJPC_SYNC) != 0)
1078 				goto rescan;
1079 
1080 			/*
1081 			 * If the VOP_PUTPAGES() did a truncated write, so
1082 			 * that even the first page of the run is not fully
1083 			 * written, vm_pageout_flush() returns 0 as the run
1084 			 * length.  Since the condition that caused truncated
1085 			 * write may be permanent, e.g. exhausted free space,
1086 			 * accepting n == 0 would cause an infinite loop.
1087 			 *
1088 			 * Forwarding the iterator leaves the unwritten page
1089 			 * behind, but there is not much we can do there if
1090 			 * filesystem refuses to write it.
1091 			 */
1092 			if (n == 0) {
1093 				n = 1;
1094 				allclean = FALSE;
1095 			}
1096 		} else {
1097 			n = 1;
1098 			vm_page_xunbusy(p);
1099 		}
1100 		np = vm_page_find_least(object, pi + n);
1101 	}
1102 #if 0
1103 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1104 #endif
1105 
1106 	/*
1107 	 * Leave updating cleangeneration for tmpfs objects to tmpfs
1108 	 * scan.  It needs to update mtime, which happens for other
1109 	 * filesystems during page writeouts.
1110 	 */
1111 	if (allclean && object->type == OBJT_VNODE)
1112 		object->cleangeneration = curgeneration;
1113 	return (res);
1114 }
1115 
1116 static int
1117 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1118     int flags, boolean_t *allclean, boolean_t *eio)
1119 {
1120 	vm_page_t ma[2 * vm_pageout_page_count - 1], tp;
1121 	int base, count, runlen;
1122 
1123 	vm_page_lock_assert(p, MA_NOTOWNED);
1124 	vm_page_assert_xbusied(p);
1125 	VM_OBJECT_ASSERT_WLOCKED(object);
1126 	base = nitems(ma) / 2;
1127 	ma[base] = p;
1128 	for (count = 1, tp = p; count < vm_pageout_page_count; count++) {
1129 		tp = vm_page_next(tp);
1130 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1131 			break;
1132 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1133 			vm_page_xunbusy(tp);
1134 			break;
1135 		}
1136 		ma[base + count] = tp;
1137 	}
1138 
1139 	for (tp = p; count < vm_pageout_page_count; count++) {
1140 		tp = vm_page_prev(tp);
1141 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1142 			break;
1143 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1144 			vm_page_xunbusy(tp);
1145 			break;
1146 		}
1147 		ma[--base] = tp;
1148 	}
1149 
1150 	vm_pageout_flush(&ma[base], count, pagerflags, nitems(ma) / 2 - base,
1151 	    &runlen, eio);
1152 	return (runlen);
1153 }
1154 
1155 /*
1156  * Note that there is absolutely no sense in writing out
1157  * anonymous objects, so we track down the vnode object
1158  * to write out.
1159  * We invalidate (remove) all pages from the address space
1160  * for semantic correctness.
1161  *
1162  * If the backing object is a device object with unmanaged pages, then any
1163  * mappings to the specified range of pages must be removed before this
1164  * function is called.
1165  *
1166  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1167  * may start out with a NULL object.
1168  */
1169 boolean_t
1170 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1171     boolean_t syncio, boolean_t invalidate)
1172 {
1173 	vm_object_t backing_object;
1174 	struct vnode *vp;
1175 	struct mount *mp;
1176 	int error, flags, fsync_after;
1177 	boolean_t res;
1178 
1179 	if (object == NULL)
1180 		return (TRUE);
1181 	res = TRUE;
1182 	error = 0;
1183 	VM_OBJECT_WLOCK(object);
1184 	while ((backing_object = object->backing_object) != NULL) {
1185 		VM_OBJECT_WLOCK(backing_object);
1186 		offset += object->backing_object_offset;
1187 		VM_OBJECT_WUNLOCK(object);
1188 		object = backing_object;
1189 		if (object->size < OFF_TO_IDX(offset + size))
1190 			size = IDX_TO_OFF(object->size) - offset;
1191 	}
1192 	/*
1193 	 * Flush pages if writing is allowed, invalidate them
1194 	 * if invalidation requested.  Pages undergoing I/O
1195 	 * will be ignored by vm_object_page_remove().
1196 	 *
1197 	 * We cannot lock the vnode and then wait for paging
1198 	 * to complete without deadlocking against vm_fault.
1199 	 * Instead we simply call vm_object_page_remove() and
1200 	 * allow it to block internally on a page-by-page
1201 	 * basis when it encounters pages undergoing async
1202 	 * I/O.
1203 	 */
1204 	if (object->type == OBJT_VNODE &&
1205 	    vm_object_mightbedirty(object) != 0 &&
1206 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1207 		VM_OBJECT_WUNLOCK(object);
1208 		(void)vn_start_write(vp, &mp, V_WAIT);
1209 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1210 		if (syncio && !invalidate && offset == 0 &&
1211 		    atop(size) == object->size) {
1212 			/*
1213 			 * If syncing the whole mapping of the file,
1214 			 * it is faster to schedule all the writes in
1215 			 * async mode, also allowing the clustering,
1216 			 * and then wait for i/o to complete.
1217 			 */
1218 			flags = 0;
1219 			fsync_after = TRUE;
1220 		} else {
1221 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1222 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1223 			fsync_after = FALSE;
1224 		}
1225 		VM_OBJECT_WLOCK(object);
1226 		res = vm_object_page_clean(object, offset, offset + size,
1227 		    flags);
1228 		VM_OBJECT_WUNLOCK(object);
1229 		if (fsync_after) {
1230 			for (;;) {
1231 				error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1232 				if (error != ERELOOKUP)
1233 					break;
1234 
1235 				/*
1236 				 * Allow SU/bufdaemon to handle more
1237 				 * dependencies in the meantime.
1238 				 */
1239 				VOP_UNLOCK(vp);
1240 				vn_finished_write(mp);
1241 
1242 				(void)vn_start_write(vp, &mp, V_WAIT);
1243 				vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1244 			}
1245 		}
1246 		VOP_UNLOCK(vp);
1247 		vn_finished_write(mp);
1248 		if (error != 0)
1249 			res = FALSE;
1250 		VM_OBJECT_WLOCK(object);
1251 	}
1252 	if ((object->type == OBJT_VNODE ||
1253 	     object->type == OBJT_DEVICE) && invalidate) {
1254 		if (object->type == OBJT_DEVICE)
1255 			/*
1256 			 * The option OBJPR_NOTMAPPED must be passed here
1257 			 * because vm_object_page_remove() cannot remove
1258 			 * unmanaged mappings.
1259 			 */
1260 			flags = OBJPR_NOTMAPPED;
1261 		else if (old_msync)
1262 			flags = 0;
1263 		else
1264 			flags = OBJPR_CLEANONLY;
1265 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1266 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1267 	}
1268 	VM_OBJECT_WUNLOCK(object);
1269 	return (res);
1270 }
1271 
1272 /*
1273  * Determine whether the given advice can be applied to the object.  Advice is
1274  * not applied to unmanaged pages since they never belong to page queues, and
1275  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1276  * have been mapped at most once.
1277  */
1278 static bool
1279 vm_object_advice_applies(vm_object_t object, int advice)
1280 {
1281 
1282 	if ((object->flags & OBJ_UNMANAGED) != 0)
1283 		return (false);
1284 	if (advice != MADV_FREE)
1285 		return (true);
1286 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1287 	    (OBJ_ONEMAPPING | OBJ_ANON));
1288 }
1289 
1290 static void
1291 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1292     vm_size_t size)
1293 {
1294 
1295 	if (advice == MADV_FREE)
1296 		vm_pager_freespace(object, pindex, size);
1297 }
1298 
1299 /*
1300  *	vm_object_madvise:
1301  *
1302  *	Implements the madvise function at the object/page level.
1303  *
1304  *	MADV_WILLNEED	(any object)
1305  *
1306  *	    Activate the specified pages if they are resident.
1307  *
1308  *	MADV_DONTNEED	(any object)
1309  *
1310  *	    Deactivate the specified pages if they are resident.
1311  *
1312  *	MADV_FREE	(OBJT_SWAP objects, OBJ_ONEMAPPING only)
1313  *
1314  *	    Deactivate and clean the specified pages if they are
1315  *	    resident.  This permits the process to reuse the pages
1316  *	    without faulting or the kernel to reclaim the pages
1317  *	    without I/O.
1318  */
1319 void
1320 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1321     int advice)
1322 {
1323 	vm_pindex_t tpindex;
1324 	vm_object_t backing_object, tobject;
1325 	vm_page_t m, tm;
1326 
1327 	if (object == NULL)
1328 		return;
1329 
1330 relookup:
1331 	VM_OBJECT_WLOCK(object);
1332 	if (!vm_object_advice_applies(object, advice)) {
1333 		VM_OBJECT_WUNLOCK(object);
1334 		return;
1335 	}
1336 	for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1337 		tobject = object;
1338 
1339 		/*
1340 		 * If the next page isn't resident in the top-level object, we
1341 		 * need to search the shadow chain.  When applying MADV_FREE, we
1342 		 * take care to release any swap space used to store
1343 		 * non-resident pages.
1344 		 */
1345 		if (m == NULL || pindex < m->pindex) {
1346 			/*
1347 			 * Optimize a common case: if the top-level object has
1348 			 * no backing object, we can skip over the non-resident
1349 			 * range in constant time.
1350 			 */
1351 			if (object->backing_object == NULL) {
1352 				tpindex = (m != NULL && m->pindex < end) ?
1353 				    m->pindex : end;
1354 				vm_object_madvise_freespace(object, advice,
1355 				    pindex, tpindex - pindex);
1356 				if ((pindex = tpindex) == end)
1357 					break;
1358 				goto next_page;
1359 			}
1360 
1361 			tpindex = pindex;
1362 			do {
1363 				vm_object_madvise_freespace(tobject, advice,
1364 				    tpindex, 1);
1365 				/*
1366 				 * Prepare to search the next object in the
1367 				 * chain.
1368 				 */
1369 				backing_object = tobject->backing_object;
1370 				if (backing_object == NULL)
1371 					goto next_pindex;
1372 				VM_OBJECT_WLOCK(backing_object);
1373 				tpindex +=
1374 				    OFF_TO_IDX(tobject->backing_object_offset);
1375 				if (tobject != object)
1376 					VM_OBJECT_WUNLOCK(tobject);
1377 				tobject = backing_object;
1378 				if (!vm_object_advice_applies(tobject, advice))
1379 					goto next_pindex;
1380 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1381 			    NULL);
1382 		} else {
1383 next_page:
1384 			tm = m;
1385 			m = TAILQ_NEXT(m, listq);
1386 		}
1387 
1388 		/*
1389 		 * If the page is not in a normal state, skip it.  The page
1390 		 * can not be invalidated while the object lock is held.
1391 		 */
1392 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1393 			goto next_pindex;
1394 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1395 		    ("vm_object_madvise: page %p is fictitious", tm));
1396 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1397 		    ("vm_object_madvise: page %p is not managed", tm));
1398 		if (vm_page_tryxbusy(tm) == 0) {
1399 			if (object != tobject)
1400 				VM_OBJECT_WUNLOCK(object);
1401 			if (advice == MADV_WILLNEED) {
1402 				/*
1403 				 * Reference the page before unlocking and
1404 				 * sleeping so that the page daemon is less
1405 				 * likely to reclaim it.
1406 				 */
1407 				vm_page_aflag_set(tm, PGA_REFERENCED);
1408 			}
1409 			if (!vm_page_busy_sleep(tm, "madvpo", 0))
1410 				VM_OBJECT_WUNLOCK(tobject);
1411   			goto relookup;
1412 		}
1413 		vm_page_advise(tm, advice);
1414 		vm_page_xunbusy(tm);
1415 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1416 next_pindex:
1417 		if (tobject != object)
1418 			VM_OBJECT_WUNLOCK(tobject);
1419 	}
1420 	VM_OBJECT_WUNLOCK(object);
1421 }
1422 
1423 /*
1424  *	vm_object_shadow:
1425  *
1426  *	Create a new object which is backed by the
1427  *	specified existing object range.  The source
1428  *	object reference is deallocated.
1429  *
1430  *	The new object and offset into that object
1431  *	are returned in the source parameters.
1432  */
1433 void
1434 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1435     struct ucred *cred, bool shared)
1436 {
1437 	vm_object_t source;
1438 	vm_object_t result;
1439 
1440 	source = *object;
1441 
1442 	/*
1443 	 * Don't create the new object if the old object isn't shared.
1444 	 *
1445 	 * If we hold the only reference we can guarantee that it won't
1446 	 * increase while we have the map locked.  Otherwise the race is
1447 	 * harmless and we will end up with an extra shadow object that
1448 	 * will be collapsed later.
1449 	 */
1450 	if (source != NULL && source->ref_count == 1 &&
1451 	    (source->flags & OBJ_ANON) != 0)
1452 		return;
1453 
1454 	/*
1455 	 * Allocate a new object with the given length.
1456 	 */
1457 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1458 
1459 	/*
1460 	 * Store the offset into the source object, and fix up the offset into
1461 	 * the new object.
1462 	 */
1463 	result->backing_object_offset = *offset;
1464 
1465 	if (shared || source != NULL) {
1466 		VM_OBJECT_WLOCK(result);
1467 
1468 		/*
1469 		 * The new object shadows the source object, adding a
1470 		 * reference to it.  Our caller changes his reference
1471 		 * to point to the new object, removing a reference to
1472 		 * the source object.  Net result: no change of
1473 		 * reference count, unless the caller needs to add one
1474 		 * more reference due to forking a shared map entry.
1475 		 */
1476 		if (shared) {
1477 			vm_object_reference_locked(result);
1478 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1479 		}
1480 
1481 		/*
1482 		 * Try to optimize the result object's page color when
1483 		 * shadowing in order to maintain page coloring
1484 		 * consistency in the combined shadowed object.
1485 		 */
1486 		if (source != NULL) {
1487 			vm_object_backing_insert(result, source);
1488 			result->domain = source->domain;
1489 #if VM_NRESERVLEVEL > 0
1490 			vm_object_set_flag(result,
1491 			    (source->flags & OBJ_COLORED));
1492 			result->pg_color = (source->pg_color +
1493 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1494 			    1)) - 1);
1495 #endif
1496 		}
1497 		VM_OBJECT_WUNLOCK(result);
1498 	}
1499 
1500 	/*
1501 	 * Return the new things
1502 	 */
1503 	*offset = 0;
1504 	*object = result;
1505 }
1506 
1507 /*
1508  *	vm_object_split:
1509  *
1510  * Split the pages in a map entry into a new object.  This affords
1511  * easier removal of unused pages, and keeps object inheritance from
1512  * being a negative impact on memory usage.
1513  */
1514 void
1515 vm_object_split(vm_map_entry_t entry)
1516 {
1517 	vm_page_t m, m_next;
1518 	vm_object_t orig_object, new_object, backing_object;
1519 	vm_pindex_t idx, offidxstart;
1520 	vm_size_t size;
1521 
1522 	orig_object = entry->object.vm_object;
1523 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1524 	    ("vm_object_split:  Splitting object with multiple mappings."));
1525 	if ((orig_object->flags & OBJ_ANON) == 0)
1526 		return;
1527 	if (orig_object->ref_count <= 1)
1528 		return;
1529 	VM_OBJECT_WUNLOCK(orig_object);
1530 
1531 	offidxstart = OFF_TO_IDX(entry->offset);
1532 	size = atop(entry->end - entry->start);
1533 
1534 	new_object = vm_object_allocate_anon(size, orig_object,
1535 	    orig_object->cred, ptoa(size));
1536 
1537 	/*
1538 	 * We must wait for the orig_object to complete any in-progress
1539 	 * collapse so that the swap blocks are stable below.  The
1540 	 * additional reference on backing_object by new object will
1541 	 * prevent further collapse operations until split completes.
1542 	 */
1543 	VM_OBJECT_WLOCK(orig_object);
1544 	vm_object_collapse_wait(orig_object);
1545 
1546 	/*
1547 	 * At this point, the new object is still private, so the order in
1548 	 * which the original and new objects are locked does not matter.
1549 	 */
1550 	VM_OBJECT_WLOCK(new_object);
1551 	new_object->domain = orig_object->domain;
1552 	backing_object = orig_object->backing_object;
1553 	if (backing_object != NULL) {
1554 		vm_object_backing_insert_ref(new_object, backing_object);
1555 		new_object->backing_object_offset =
1556 		    orig_object->backing_object_offset + entry->offset;
1557 	}
1558 	if (orig_object->cred != NULL) {
1559 		crhold(orig_object->cred);
1560 		KASSERT(orig_object->charge >= ptoa(size),
1561 		    ("orig_object->charge < 0"));
1562 		orig_object->charge -= ptoa(size);
1563 	}
1564 
1565 	/*
1566 	 * Mark the split operation so that swap_pager_getpages() knows
1567 	 * that the object is in transition.
1568 	 */
1569 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1570 #ifdef INVARIANTS
1571 	idx = 0;
1572 #endif
1573 retry:
1574 	m = vm_page_find_least(orig_object, offidxstart);
1575 	KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1576 	    ("%s: object %p was repopulated", __func__, orig_object));
1577 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1578 	    m = m_next) {
1579 		m_next = TAILQ_NEXT(m, listq);
1580 
1581 		/*
1582 		 * We must wait for pending I/O to complete before we can
1583 		 * rename the page.
1584 		 *
1585 		 * We do not have to VM_PROT_NONE the page as mappings should
1586 		 * not be changed by this operation.
1587 		 */
1588 		if (vm_page_tryxbusy(m) == 0) {
1589 			VM_OBJECT_WUNLOCK(new_object);
1590 			if (vm_page_busy_sleep(m, "spltwt", 0))
1591 				VM_OBJECT_WLOCK(orig_object);
1592 			VM_OBJECT_WLOCK(new_object);
1593 			goto retry;
1594 		}
1595 
1596 		/*
1597 		 * The page was left invalid.  Likely placed there by
1598 		 * an incomplete fault.  Just remove and ignore.
1599 		 */
1600 		if (vm_page_none_valid(m)) {
1601 			if (vm_page_remove(m))
1602 				vm_page_free(m);
1603 			continue;
1604 		}
1605 
1606 		/* vm_page_rename() will dirty the page. */
1607 		if (vm_page_rename(m, new_object, idx)) {
1608 			vm_page_xunbusy(m);
1609 			VM_OBJECT_WUNLOCK(new_object);
1610 			VM_OBJECT_WUNLOCK(orig_object);
1611 			vm_radix_wait();
1612 			VM_OBJECT_WLOCK(orig_object);
1613 			VM_OBJECT_WLOCK(new_object);
1614 			goto retry;
1615 		}
1616 
1617 #if VM_NRESERVLEVEL > 0
1618 		/*
1619 		 * If some of the reservation's allocated pages remain with
1620 		 * the original object, then transferring the reservation to
1621 		 * the new object is neither particularly beneficial nor
1622 		 * particularly harmful as compared to leaving the reservation
1623 		 * with the original object.  If, however, all of the
1624 		 * reservation's allocated pages are transferred to the new
1625 		 * object, then transferring the reservation is typically
1626 		 * beneficial.  Determining which of these two cases applies
1627 		 * would be more costly than unconditionally renaming the
1628 		 * reservation.
1629 		 */
1630 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1631 #endif
1632 	}
1633 
1634 	/*
1635 	 * swap_pager_copy() can sleep, in which case the orig_object's
1636 	 * and new_object's locks are released and reacquired.
1637 	 */
1638 	swap_pager_copy(orig_object, new_object, offidxstart, 0);
1639 
1640 	TAILQ_FOREACH(m, &new_object->memq, listq)
1641 		vm_page_xunbusy(m);
1642 
1643 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1644 	VM_OBJECT_WUNLOCK(orig_object);
1645 	VM_OBJECT_WUNLOCK(new_object);
1646 	entry->object.vm_object = new_object;
1647 	entry->offset = 0LL;
1648 	vm_object_deallocate(orig_object);
1649 	VM_OBJECT_WLOCK(new_object);
1650 }
1651 
1652 static vm_page_t
1653 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1654 {
1655 	vm_object_t backing_object;
1656 
1657 	VM_OBJECT_ASSERT_WLOCKED(object);
1658 	backing_object = object->backing_object;
1659 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1660 
1661 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1662 	    ("invalid ownership %p %p %p", p, object, backing_object));
1663 	/* The page is only NULL when rename fails. */
1664 	if (p == NULL) {
1665 		VM_OBJECT_WUNLOCK(object);
1666 		VM_OBJECT_WUNLOCK(backing_object);
1667 		vm_radix_wait();
1668 		VM_OBJECT_WLOCK(object);
1669 	} else if (p->object == object) {
1670 		VM_OBJECT_WUNLOCK(backing_object);
1671 		if (vm_page_busy_sleep(p, "vmocol", 0))
1672 			VM_OBJECT_WLOCK(object);
1673 	} else {
1674 		VM_OBJECT_WUNLOCK(object);
1675 		if (!vm_page_busy_sleep(p, "vmocol", 0))
1676 			VM_OBJECT_WUNLOCK(backing_object);
1677 		VM_OBJECT_WLOCK(object);
1678 	}
1679 	VM_OBJECT_WLOCK(backing_object);
1680 	return (TAILQ_FIRST(&backing_object->memq));
1681 }
1682 
1683 static bool
1684 vm_object_scan_all_shadowed(vm_object_t object)
1685 {
1686 	vm_object_t backing_object;
1687 	vm_page_t p, pp;
1688 	vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1689 
1690 	VM_OBJECT_ASSERT_WLOCKED(object);
1691 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1692 
1693 	backing_object = object->backing_object;
1694 
1695 	if ((backing_object->flags & OBJ_ANON) == 0)
1696 		return (false);
1697 
1698 	pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1699 	p = vm_page_find_least(backing_object, pi);
1700 	ps = swap_pager_find_least(backing_object, pi);
1701 
1702 	/*
1703 	 * Only check pages inside the parent object's range and
1704 	 * inside the parent object's mapping of the backing object.
1705 	 */
1706 	for (;; pi++) {
1707 		if (p != NULL && p->pindex < pi)
1708 			p = TAILQ_NEXT(p, listq);
1709 		if (ps < pi)
1710 			ps = swap_pager_find_least(backing_object, pi);
1711 		if (p == NULL && ps >= backing_object->size)
1712 			break;
1713 		else if (p == NULL)
1714 			pi = ps;
1715 		else
1716 			pi = MIN(p->pindex, ps);
1717 
1718 		new_pindex = pi - backing_offset_index;
1719 		if (new_pindex >= object->size)
1720 			break;
1721 
1722 		if (p != NULL) {
1723 			/*
1724 			 * If the backing object page is busy a
1725 			 * grandparent or older page may still be
1726 			 * undergoing CoW.  It is not safe to collapse
1727 			 * the backing object until it is quiesced.
1728 			 */
1729 			if (vm_page_tryxbusy(p) == 0)
1730 				return (false);
1731 
1732 			/*
1733 			 * We raced with the fault handler that left
1734 			 * newly allocated invalid page on the object
1735 			 * queue and retried.
1736 			 */
1737 			if (!vm_page_all_valid(p))
1738 				goto unbusy_ret;
1739 		}
1740 
1741 		/*
1742 		 * See if the parent has the page or if the parent's object
1743 		 * pager has the page.  If the parent has the page but the page
1744 		 * is not valid, the parent's object pager must have the page.
1745 		 *
1746 		 * If this fails, the parent does not completely shadow the
1747 		 * object and we might as well give up now.
1748 		 */
1749 		pp = vm_page_lookup(object, new_pindex);
1750 
1751 		/*
1752 		 * The valid check here is stable due to object lock
1753 		 * being required to clear valid and initiate paging.
1754 		 * Busy of p disallows fault handler to validate pp.
1755 		 */
1756 		if ((pp == NULL || vm_page_none_valid(pp)) &&
1757 		    !vm_pager_has_page(object, new_pindex, NULL, NULL))
1758 			goto unbusy_ret;
1759 		if (p != NULL)
1760 			vm_page_xunbusy(p);
1761 	}
1762 	return (true);
1763 
1764 unbusy_ret:
1765 	if (p != NULL)
1766 		vm_page_xunbusy(p);
1767 	return (false);
1768 }
1769 
1770 static void
1771 vm_object_collapse_scan(vm_object_t object)
1772 {
1773 	vm_object_t backing_object;
1774 	vm_page_t next, p, pp;
1775 	vm_pindex_t backing_offset_index, new_pindex;
1776 
1777 	VM_OBJECT_ASSERT_WLOCKED(object);
1778 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1779 
1780 	backing_object = object->backing_object;
1781 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1782 
1783 	/*
1784 	 * Our scan
1785 	 */
1786 	for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1787 		next = TAILQ_NEXT(p, listq);
1788 		new_pindex = p->pindex - backing_offset_index;
1789 
1790 		/*
1791 		 * Check for busy page
1792 		 */
1793 		if (vm_page_tryxbusy(p) == 0) {
1794 			next = vm_object_collapse_scan_wait(object, p);
1795 			continue;
1796 		}
1797 
1798 		KASSERT(object->backing_object == backing_object,
1799 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1800 		    object->backing_object, backing_object));
1801 		KASSERT(p->object == backing_object,
1802 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1803 		    p->object, backing_object));
1804 
1805 		if (p->pindex < backing_offset_index ||
1806 		    new_pindex >= object->size) {
1807 			vm_pager_freespace(backing_object, p->pindex, 1);
1808 
1809 			KASSERT(!pmap_page_is_mapped(p),
1810 			    ("freeing mapped page %p", p));
1811 			if (vm_page_remove(p))
1812 				vm_page_free(p);
1813 			continue;
1814 		}
1815 
1816 		if (!vm_page_all_valid(p)) {
1817 			KASSERT(!pmap_page_is_mapped(p),
1818 			    ("freeing mapped page %p", p));
1819 			if (vm_page_remove(p))
1820 				vm_page_free(p);
1821 			continue;
1822 		}
1823 
1824 		pp = vm_page_lookup(object, new_pindex);
1825 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1826 			vm_page_xunbusy(p);
1827 			/*
1828 			 * The page in the parent is busy and possibly not
1829 			 * (yet) valid.  Until its state is finalized by the
1830 			 * busy bit owner, we can't tell whether it shadows the
1831 			 * original page.
1832 			 */
1833 			next = vm_object_collapse_scan_wait(object, pp);
1834 			continue;
1835 		}
1836 
1837 		if (pp != NULL && vm_page_none_valid(pp)) {
1838 			/*
1839 			 * The page was invalid in the parent.  Likely placed
1840 			 * there by an incomplete fault.  Just remove and
1841 			 * ignore.  p can replace it.
1842 			 */
1843 			if (vm_page_remove(pp))
1844 				vm_page_free(pp);
1845 			pp = NULL;
1846 		}
1847 
1848 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1849 			NULL)) {
1850 			/*
1851 			 * The page already exists in the parent OR swap exists
1852 			 * for this location in the parent.  Leave the parent's
1853 			 * page alone.  Destroy the original page from the
1854 			 * backing object.
1855 			 */
1856 			vm_pager_freespace(backing_object, p->pindex, 1);
1857 			KASSERT(!pmap_page_is_mapped(p),
1858 			    ("freeing mapped page %p", p));
1859 			if (vm_page_remove(p))
1860 				vm_page_free(p);
1861 			if (pp != NULL)
1862 				vm_page_xunbusy(pp);
1863 			continue;
1864 		}
1865 
1866 		/*
1867 		 * Page does not exist in parent, rename the page from the
1868 		 * backing object to the main object.
1869 		 *
1870 		 * If the page was mapped to a process, it can remain mapped
1871 		 * through the rename.  vm_page_rename() will dirty the page.
1872 		 */
1873 		if (vm_page_rename(p, object, new_pindex)) {
1874 			vm_page_xunbusy(p);
1875 			next = vm_object_collapse_scan_wait(object, NULL);
1876 			continue;
1877 		}
1878 
1879 		/* Use the old pindex to free the right page. */
1880 		vm_pager_freespace(backing_object, new_pindex +
1881 		    backing_offset_index, 1);
1882 
1883 #if VM_NRESERVLEVEL > 0
1884 		/*
1885 		 * Rename the reservation.
1886 		 */
1887 		vm_reserv_rename(p, object, backing_object,
1888 		    backing_offset_index);
1889 #endif
1890 		vm_page_xunbusy(p);
1891 	}
1892 	return;
1893 }
1894 
1895 /*
1896  *	vm_object_collapse:
1897  *
1898  *	Collapse an object with the object backing it.
1899  *	Pages in the backing object are moved into the
1900  *	parent, and the backing object is deallocated.
1901  */
1902 void
1903 vm_object_collapse(vm_object_t object)
1904 {
1905 	vm_object_t backing_object, new_backing_object;
1906 
1907 	VM_OBJECT_ASSERT_WLOCKED(object);
1908 
1909 	while (TRUE) {
1910 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1911 		    ("collapsing invalid object"));
1912 
1913 		/*
1914 		 * Wait for the backing_object to finish any pending
1915 		 * collapse so that the caller sees the shortest possible
1916 		 * shadow chain.
1917 		 */
1918 		backing_object = vm_object_backing_collapse_wait(object);
1919 		if (backing_object == NULL)
1920 			return;
1921 
1922 		KASSERT(object->ref_count > 0 &&
1923 		    object->ref_count > atomic_load_int(&object->shadow_count),
1924 		    ("collapse with invalid ref %d or shadow %d count.",
1925 		    object->ref_count, atomic_load_int(&object->shadow_count)));
1926 		KASSERT((backing_object->flags &
1927 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1928 		    ("vm_object_collapse: Backing object already collapsing."));
1929 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1930 		    ("vm_object_collapse: object is already collapsing."));
1931 
1932 		/*
1933 		 * We know that we can either collapse the backing object if
1934 		 * the parent is the only reference to it, or (perhaps) have
1935 		 * the parent bypass the object if the parent happens to shadow
1936 		 * all the resident pages in the entire backing object.
1937 		 */
1938 		if (backing_object->ref_count == 1) {
1939 			KASSERT(atomic_load_int(&backing_object->shadow_count)
1940 			    == 1,
1941 			    ("vm_object_collapse: shadow_count: %d",
1942 			    atomic_load_int(&backing_object->shadow_count)));
1943 			vm_object_pip_add(object, 1);
1944 			vm_object_set_flag(object, OBJ_COLLAPSING);
1945 			vm_object_pip_add(backing_object, 1);
1946 			vm_object_set_flag(backing_object, OBJ_DEAD);
1947 
1948 			/*
1949 			 * If there is exactly one reference to the backing
1950 			 * object, we can collapse it into the parent.
1951 			 */
1952 			vm_object_collapse_scan(object);
1953 
1954 			/*
1955 			 * Move the pager from backing_object to object.
1956 			 *
1957 			 * swap_pager_copy() can sleep, in which case the
1958 			 * backing_object's and object's locks are released and
1959 			 * reacquired.
1960 			 */
1961 			swap_pager_copy(backing_object, object,
1962 			    OFF_TO_IDX(object->backing_object_offset), TRUE);
1963 
1964 			/*
1965 			 * Object now shadows whatever backing_object did.
1966 			 */
1967 			vm_object_clear_flag(object, OBJ_COLLAPSING);
1968 			vm_object_backing_transfer(object, backing_object);
1969 			object->backing_object_offset +=
1970 			    backing_object->backing_object_offset;
1971 			VM_OBJECT_WUNLOCK(object);
1972 			vm_object_pip_wakeup(object);
1973 
1974 			/*
1975 			 * Discard backing_object.
1976 			 *
1977 			 * Since the backing object has no pages, no pager left,
1978 			 * and no object references within it, all that is
1979 			 * necessary is to dispose of it.
1980 			 */
1981 			KASSERT(backing_object->ref_count == 1, (
1982 "backing_object %p was somehow re-referenced during collapse!",
1983 			    backing_object));
1984 			vm_object_pip_wakeup(backing_object);
1985 			(void)refcount_release(&backing_object->ref_count);
1986 			umtx_shm_object_terminated(backing_object);
1987 			vm_object_terminate(backing_object);
1988 			counter_u64_add(object_collapses, 1);
1989 			VM_OBJECT_WLOCK(object);
1990 		} else {
1991 			/*
1992 			 * If we do not entirely shadow the backing object,
1993 			 * there is nothing we can do so we give up.
1994 			 *
1995 			 * The object lock and backing_object lock must not
1996 			 * be dropped during this sequence.
1997 			 */
1998 			if (!vm_object_scan_all_shadowed(object)) {
1999 				VM_OBJECT_WUNLOCK(backing_object);
2000 				break;
2001 			}
2002 
2003 			/*
2004 			 * Make the parent shadow the next object in the
2005 			 * chain.  Deallocating backing_object will not remove
2006 			 * it, since its reference count is at least 2.
2007 			 */
2008 			vm_object_backing_remove_locked(object);
2009 			new_backing_object = backing_object->backing_object;
2010 			if (new_backing_object != NULL) {
2011 				vm_object_backing_insert_ref(object,
2012 				    new_backing_object);
2013 				object->backing_object_offset +=
2014 				    backing_object->backing_object_offset;
2015 			}
2016 
2017 			/*
2018 			 * Drop the reference count on backing_object. Since
2019 			 * its ref_count was at least 2, it will not vanish.
2020 			 */
2021 			(void)refcount_release(&backing_object->ref_count);
2022 			KASSERT(backing_object->ref_count >= 1, (
2023 "backing_object %p was somehow dereferenced during collapse!",
2024 			    backing_object));
2025 			VM_OBJECT_WUNLOCK(backing_object);
2026 			counter_u64_add(object_bypasses, 1);
2027 		}
2028 
2029 		/*
2030 		 * Try again with this object's new backing object.
2031 		 */
2032 	}
2033 }
2034 
2035 /*
2036  *	vm_object_page_remove:
2037  *
2038  *	For the given object, either frees or invalidates each of the
2039  *	specified pages.  In general, a page is freed.  However, if a page is
2040  *	wired for any reason other than the existence of a managed, wired
2041  *	mapping, then it may be invalidated but not removed from the object.
2042  *	Pages are specified by the given range ["start", "end") and the option
2043  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
2044  *	extends from "start" to the end of the object.  If the option
2045  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2046  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
2047  *	specified, then the pages within the specified range must have no
2048  *	mappings.  Otherwise, if this option is not specified, any mappings to
2049  *	the specified pages are removed before the pages are freed or
2050  *	invalidated.
2051  *
2052  *	In general, this operation should only be performed on objects that
2053  *	contain managed pages.  There are, however, two exceptions.  First, it
2054  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
2055  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2056  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
2057  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
2058  *
2059  *	The object must be locked.
2060  */
2061 void
2062 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2063     int options)
2064 {
2065 	vm_page_t p, next;
2066 
2067 	VM_OBJECT_ASSERT_WLOCKED(object);
2068 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2069 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2070 	    ("vm_object_page_remove: illegal options for object %p", object));
2071 	if (object->resident_page_count == 0)
2072 		return;
2073 	vm_object_pip_add(object, 1);
2074 again:
2075 	p = vm_page_find_least(object, start);
2076 
2077 	/*
2078 	 * Here, the variable "p" is either (1) the page with the least pindex
2079 	 * greater than or equal to the parameter "start" or (2) NULL.
2080 	 */
2081 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2082 		next = TAILQ_NEXT(p, listq);
2083 
2084 		/*
2085 		 * Skip invalid pages if asked to do so.  Try to avoid acquiring
2086 		 * the busy lock, as some consumers rely on this to avoid
2087 		 * deadlocks.
2088 		 *
2089 		 * A thread may concurrently transition the page from invalid to
2090 		 * valid using only the busy lock, so the result of this check
2091 		 * is immediately stale.  It is up to consumers to handle this,
2092 		 * for instance by ensuring that all invalid->valid transitions
2093 		 * happen with a mutex held, as may be possible for a
2094 		 * filesystem.
2095 		 */
2096 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2097 			continue;
2098 
2099 		/*
2100 		 * If the page is wired for any reason besides the existence
2101 		 * of managed, wired mappings, then it cannot be freed.  For
2102 		 * example, fictitious pages, which represent device memory,
2103 		 * are inherently wired and cannot be freed.  They can,
2104 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2105 		 * not specified.
2106 		 */
2107 		if (vm_page_tryxbusy(p) == 0) {
2108 			if (vm_page_busy_sleep(p, "vmopar", 0))
2109 				VM_OBJECT_WLOCK(object);
2110 			goto again;
2111 		}
2112 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2113 			vm_page_xunbusy(p);
2114 			continue;
2115 		}
2116 		if (vm_page_wired(p)) {
2117 wired:
2118 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2119 			    object->ref_count != 0)
2120 				pmap_remove_all(p);
2121 			if ((options & OBJPR_CLEANONLY) == 0) {
2122 				vm_page_invalid(p);
2123 				vm_page_undirty(p);
2124 			}
2125 			vm_page_xunbusy(p);
2126 			continue;
2127 		}
2128 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2129 		    ("vm_object_page_remove: page %p is fictitious", p));
2130 		if ((options & OBJPR_CLEANONLY) != 0 &&
2131 		    !vm_page_none_valid(p)) {
2132 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2133 			    object->ref_count != 0 &&
2134 			    !vm_page_try_remove_write(p))
2135 				goto wired;
2136 			if (p->dirty != 0) {
2137 				vm_page_xunbusy(p);
2138 				continue;
2139 			}
2140 		}
2141 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2142 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2143 			goto wired;
2144 		vm_page_free(p);
2145 	}
2146 	vm_object_pip_wakeup(object);
2147 
2148 	vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2149 	    start);
2150 }
2151 
2152 /*
2153  *	vm_object_page_noreuse:
2154  *
2155  *	For the given object, attempt to move the specified pages to
2156  *	the head of the inactive queue.  This bypasses regular LRU
2157  *	operation and allows the pages to be reused quickly under memory
2158  *	pressure.  If a page is wired for any reason, then it will not
2159  *	be queued.  Pages are specified by the range ["start", "end").
2160  *	As a special case, if "end" is zero, then the range extends from
2161  *	"start" to the end of the object.
2162  *
2163  *	This operation should only be performed on objects that
2164  *	contain non-fictitious, managed pages.
2165  *
2166  *	The object must be locked.
2167  */
2168 void
2169 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2170 {
2171 	vm_page_t p, next;
2172 
2173 	VM_OBJECT_ASSERT_LOCKED(object);
2174 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2175 	    ("vm_object_page_noreuse: illegal object %p", object));
2176 	if (object->resident_page_count == 0)
2177 		return;
2178 	p = vm_page_find_least(object, start);
2179 
2180 	/*
2181 	 * Here, the variable "p" is either (1) the page with the least pindex
2182 	 * greater than or equal to the parameter "start" or (2) NULL.
2183 	 */
2184 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2185 		next = TAILQ_NEXT(p, listq);
2186 		vm_page_deactivate_noreuse(p);
2187 	}
2188 }
2189 
2190 /*
2191  *	Populate the specified range of the object with valid pages.  Returns
2192  *	TRUE if the range is successfully populated and FALSE otherwise.
2193  *
2194  *	Note: This function should be optimized to pass a larger array of
2195  *	pages to vm_pager_get_pages() before it is applied to a non-
2196  *	OBJT_DEVICE object.
2197  *
2198  *	The object must be locked.
2199  */
2200 boolean_t
2201 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2202 {
2203 	vm_page_t m;
2204 	vm_pindex_t pindex;
2205 	int rv;
2206 
2207 	VM_OBJECT_ASSERT_WLOCKED(object);
2208 	for (pindex = start; pindex < end; pindex++) {
2209 		rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2210 		if (rv != VM_PAGER_OK)
2211 			break;
2212 
2213 		/*
2214 		 * Keep "m" busy because a subsequent iteration may unlock
2215 		 * the object.
2216 		 */
2217 	}
2218 	if (pindex > start) {
2219 		m = vm_page_lookup(object, start);
2220 		while (m != NULL && m->pindex < pindex) {
2221 			vm_page_xunbusy(m);
2222 			m = TAILQ_NEXT(m, listq);
2223 		}
2224 	}
2225 	return (pindex == end);
2226 }
2227 
2228 /*
2229  *	Routine:	vm_object_coalesce
2230  *	Function:	Coalesces two objects backing up adjoining
2231  *			regions of memory into a single object.
2232  *
2233  *	returns TRUE if objects were combined.
2234  *
2235  *	NOTE:	Only works at the moment if the second object is NULL -
2236  *		if it's not, which object do we lock first?
2237  *
2238  *	Parameters:
2239  *		prev_object	First object to coalesce
2240  *		prev_offset	Offset into prev_object
2241  *		prev_size	Size of reference to prev_object
2242  *		next_size	Size of reference to the second object
2243  *		reserved	Indicator that extension region has
2244  *				swap accounted for
2245  *
2246  *	Conditions:
2247  *	The object must *not* be locked.
2248  */
2249 boolean_t
2250 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2251     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2252 {
2253 	vm_pindex_t next_pindex;
2254 
2255 	if (prev_object == NULL)
2256 		return (TRUE);
2257 	if ((prev_object->flags & OBJ_ANON) == 0)
2258 		return (FALSE);
2259 
2260 	VM_OBJECT_WLOCK(prev_object);
2261 	/*
2262 	 * Try to collapse the object first.
2263 	 */
2264 	vm_object_collapse(prev_object);
2265 
2266 	/*
2267 	 * Can't coalesce if: . more than one reference . paged out . shadows
2268 	 * another object . has a copy elsewhere (any of which mean that the
2269 	 * pages not mapped to prev_entry may be in use anyway)
2270 	 */
2271 	if (prev_object->backing_object != NULL) {
2272 		VM_OBJECT_WUNLOCK(prev_object);
2273 		return (FALSE);
2274 	}
2275 
2276 	prev_size >>= PAGE_SHIFT;
2277 	next_size >>= PAGE_SHIFT;
2278 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2279 
2280 	if (prev_object->ref_count > 1 &&
2281 	    prev_object->size != next_pindex &&
2282 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2283 		VM_OBJECT_WUNLOCK(prev_object);
2284 		return (FALSE);
2285 	}
2286 
2287 	/*
2288 	 * Account for the charge.
2289 	 */
2290 	if (prev_object->cred != NULL) {
2291 		/*
2292 		 * If prev_object was charged, then this mapping,
2293 		 * although not charged now, may become writable
2294 		 * later. Non-NULL cred in the object would prevent
2295 		 * swap reservation during enabling of the write
2296 		 * access, so reserve swap now. Failed reservation
2297 		 * cause allocation of the separate object for the map
2298 		 * entry, and swap reservation for this entry is
2299 		 * managed in appropriate time.
2300 		 */
2301 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2302 		    prev_object->cred)) {
2303 			VM_OBJECT_WUNLOCK(prev_object);
2304 			return (FALSE);
2305 		}
2306 		prev_object->charge += ptoa(next_size);
2307 	}
2308 
2309 	/*
2310 	 * Remove any pages that may still be in the object from a previous
2311 	 * deallocation.
2312 	 */
2313 	if (next_pindex < prev_object->size) {
2314 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2315 		    next_size, 0);
2316 #if 0
2317 		if (prev_object->cred != NULL) {
2318 			KASSERT(prev_object->charge >=
2319 			    ptoa(prev_object->size - next_pindex),
2320 			    ("object %p overcharged 1 %jx %jx", prev_object,
2321 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2322 			prev_object->charge -= ptoa(prev_object->size -
2323 			    next_pindex);
2324 		}
2325 #endif
2326 	}
2327 
2328 	/*
2329 	 * Extend the object if necessary.
2330 	 */
2331 	if (next_pindex + next_size > prev_object->size)
2332 		prev_object->size = next_pindex + next_size;
2333 
2334 	VM_OBJECT_WUNLOCK(prev_object);
2335 	return (TRUE);
2336 }
2337 
2338 void
2339 vm_object_set_writeable_dirty_(vm_object_t object)
2340 {
2341 	atomic_add_int(&object->generation, 1);
2342 }
2343 
2344 bool
2345 vm_object_mightbedirty_(vm_object_t object)
2346 {
2347 	return (object->generation != object->cleangeneration);
2348 }
2349 
2350 /*
2351  *	vm_object_unwire:
2352  *
2353  *	For each page offset within the specified range of the given object,
2354  *	find the highest-level page in the shadow chain and unwire it.  A page
2355  *	must exist at every page offset, and the highest-level page must be
2356  *	wired.
2357  */
2358 void
2359 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2360     uint8_t queue)
2361 {
2362 	vm_object_t tobject, t1object;
2363 	vm_page_t m, tm;
2364 	vm_pindex_t end_pindex, pindex, tpindex;
2365 	int depth, locked_depth;
2366 
2367 	KASSERT((offset & PAGE_MASK) == 0,
2368 	    ("vm_object_unwire: offset is not page aligned"));
2369 	KASSERT((length & PAGE_MASK) == 0,
2370 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2371 	/* The wired count of a fictitious page never changes. */
2372 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2373 		return;
2374 	pindex = OFF_TO_IDX(offset);
2375 	end_pindex = pindex + atop(length);
2376 again:
2377 	locked_depth = 1;
2378 	VM_OBJECT_RLOCK(object);
2379 	m = vm_page_find_least(object, pindex);
2380 	while (pindex < end_pindex) {
2381 		if (m == NULL || pindex < m->pindex) {
2382 			/*
2383 			 * The first object in the shadow chain doesn't
2384 			 * contain a page at the current index.  Therefore,
2385 			 * the page must exist in a backing object.
2386 			 */
2387 			tobject = object;
2388 			tpindex = pindex;
2389 			depth = 0;
2390 			do {
2391 				tpindex +=
2392 				    OFF_TO_IDX(tobject->backing_object_offset);
2393 				tobject = tobject->backing_object;
2394 				KASSERT(tobject != NULL,
2395 				    ("vm_object_unwire: missing page"));
2396 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2397 					goto next_page;
2398 				depth++;
2399 				if (depth == locked_depth) {
2400 					locked_depth++;
2401 					VM_OBJECT_RLOCK(tobject);
2402 				}
2403 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2404 			    NULL);
2405 		} else {
2406 			tm = m;
2407 			m = TAILQ_NEXT(m, listq);
2408 		}
2409 		if (vm_page_trysbusy(tm) == 0) {
2410 			for (tobject = object; locked_depth >= 1;
2411 			    locked_depth--) {
2412 				t1object = tobject->backing_object;
2413 				if (tm->object != tobject)
2414 					VM_OBJECT_RUNLOCK(tobject);
2415 				tobject = t1object;
2416 			}
2417 			tobject = tm->object;
2418 			if (!vm_page_busy_sleep(tm, "unwbo",
2419 			    VM_ALLOC_IGN_SBUSY))
2420 				VM_OBJECT_RUNLOCK(tobject);
2421 			goto again;
2422 		}
2423 		vm_page_unwire(tm, queue);
2424 		vm_page_sunbusy(tm);
2425 next_page:
2426 		pindex++;
2427 	}
2428 	/* Release the accumulated object locks. */
2429 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2430 		t1object = tobject->backing_object;
2431 		VM_OBJECT_RUNLOCK(tobject);
2432 		tobject = t1object;
2433 	}
2434 }
2435 
2436 /*
2437  * Return the vnode for the given object, or NULL if none exists.
2438  * For tmpfs objects, the function may return NULL if there is
2439  * no vnode allocated at the time of the call.
2440  */
2441 struct vnode *
2442 vm_object_vnode(vm_object_t object)
2443 {
2444 	struct vnode *vp;
2445 
2446 	VM_OBJECT_ASSERT_LOCKED(object);
2447 	vm_pager_getvp(object, &vp, NULL);
2448 	return (vp);
2449 }
2450 
2451 /*
2452  * Busy the vm object.  This prevents new pages belonging to the object from
2453  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2454  * for checking page state before proceeding.
2455  */
2456 void
2457 vm_object_busy(vm_object_t obj)
2458 {
2459 
2460 	VM_OBJECT_ASSERT_LOCKED(obj);
2461 
2462 	blockcount_acquire(&obj->busy, 1);
2463 	/* The fence is required to order loads of page busy. */
2464 	atomic_thread_fence_acq_rel();
2465 }
2466 
2467 void
2468 vm_object_unbusy(vm_object_t obj)
2469 {
2470 
2471 	blockcount_release(&obj->busy, 1);
2472 }
2473 
2474 void
2475 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2476 {
2477 
2478 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2479 
2480 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2481 }
2482 
2483 /*
2484  * This function aims to determine if the object is mapped,
2485  * specifically, if it is referenced by a vm_map_entry.  Because
2486  * objects occasionally acquire transient references that do not
2487  * represent a mapping, the method used here is inexact.  However, it
2488  * has very low overhead and is good enough for the advisory
2489  * vm.vmtotal sysctl.
2490  */
2491 bool
2492 vm_object_is_active(vm_object_t obj)
2493 {
2494 
2495 	return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2496 }
2497 
2498 static int
2499 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2500 {
2501 	struct kinfo_vmobject *kvo;
2502 	char *fullpath, *freepath;
2503 	struct vnode *vp;
2504 	struct vattr va;
2505 	vm_object_t obj;
2506 	vm_page_t m;
2507 	u_long sp;
2508 	int count, error;
2509 	bool want_path;
2510 
2511 	if (req->oldptr == NULL) {
2512 		/*
2513 		 * If an old buffer has not been provided, generate an
2514 		 * estimate of the space needed for a subsequent call.
2515 		 */
2516 		mtx_lock(&vm_object_list_mtx);
2517 		count = 0;
2518 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2519 			if (obj->type == OBJT_DEAD)
2520 				continue;
2521 			count++;
2522 		}
2523 		mtx_unlock(&vm_object_list_mtx);
2524 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2525 		    count * 11 / 10));
2526 	}
2527 
2528 	want_path = !(swap_only || jailed(curthread->td_ucred));
2529 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO);
2530 	error = 0;
2531 
2532 	/*
2533 	 * VM objects are type stable and are never removed from the
2534 	 * list once added.  This allows us to safely read obj->object_list
2535 	 * after reacquiring the VM object lock.
2536 	 */
2537 	mtx_lock(&vm_object_list_mtx);
2538 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2539 		if (obj->type == OBJT_DEAD ||
2540 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2541 			continue;
2542 		VM_OBJECT_RLOCK(obj);
2543 		if (obj->type == OBJT_DEAD ||
2544 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2545 			VM_OBJECT_RUNLOCK(obj);
2546 			continue;
2547 		}
2548 		mtx_unlock(&vm_object_list_mtx);
2549 		kvo->kvo_size = ptoa(obj->size);
2550 		kvo->kvo_resident = obj->resident_page_count;
2551 		kvo->kvo_ref_count = obj->ref_count;
2552 		kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2553 		kvo->kvo_memattr = obj->memattr;
2554 		kvo->kvo_active = 0;
2555 		kvo->kvo_inactive = 0;
2556 		if (!swap_only) {
2557 			TAILQ_FOREACH(m, &obj->memq, listq) {
2558 				/*
2559 				 * A page may belong to the object but be
2560 				 * dequeued and set to PQ_NONE while the
2561 				 * object lock is not held.  This makes the
2562 				 * reads of m->queue below racy, and we do not
2563 				 * count pages set to PQ_NONE.  However, this
2564 				 * sysctl is only meant to give an
2565 				 * approximation of the system anyway.
2566 				 */
2567 				if (m->a.queue == PQ_ACTIVE)
2568 					kvo->kvo_active++;
2569 				else if (m->a.queue == PQ_INACTIVE)
2570 					kvo->kvo_inactive++;
2571 			}
2572 		}
2573 
2574 		kvo->kvo_vn_fileid = 0;
2575 		kvo->kvo_vn_fsid = 0;
2576 		kvo->kvo_vn_fsid_freebsd11 = 0;
2577 		freepath = NULL;
2578 		fullpath = "";
2579 		vp = NULL;
2580 		kvo->kvo_type = vm_object_kvme_type(obj, want_path ? &vp :
2581 		    NULL);
2582 		if (vp != NULL) {
2583 			vref(vp);
2584 		} else if ((obj->flags & OBJ_ANON) != 0) {
2585 			MPASS(kvo->kvo_type == KVME_TYPE_SWAP);
2586 			kvo->kvo_me = (uintptr_t)obj;
2587 			/* tmpfs objs are reported as vnodes */
2588 			kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2589 			sp = swap_pager_swapped_pages(obj);
2590 			kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2591 		}
2592 		VM_OBJECT_RUNLOCK(obj);
2593 		if (vp != NULL) {
2594 			vn_fullpath(vp, &fullpath, &freepath);
2595 			vn_lock(vp, LK_SHARED | LK_RETRY);
2596 			if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2597 				kvo->kvo_vn_fileid = va.va_fileid;
2598 				kvo->kvo_vn_fsid = va.va_fsid;
2599 				kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2600 								/* truncate */
2601 			}
2602 			vput(vp);
2603 		}
2604 
2605 		strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2606 		free(freepath, M_TEMP);
2607 
2608 		/* Pack record size down */
2609 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2610 		    + strlen(kvo->kvo_path) + 1;
2611 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2612 		    sizeof(uint64_t));
2613 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2614 		maybe_yield();
2615 		mtx_lock(&vm_object_list_mtx);
2616 		if (error)
2617 			break;
2618 	}
2619 	mtx_unlock(&vm_object_list_mtx);
2620 	free(kvo, M_TEMP);
2621 	return (error);
2622 }
2623 
2624 static int
2625 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2626 {
2627 	return (vm_object_list_handler(req, false));
2628 }
2629 
2630 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2631     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2632     "List of VM objects");
2633 
2634 static int
2635 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2636 {
2637 	return (vm_object_list_handler(req, true));
2638 }
2639 
2640 /*
2641  * This sysctl returns list of the anonymous or swap objects. Intent
2642  * is to provide stripped optimized list useful to analyze swap use.
2643  * Since technically non-swap (default) objects participate in the
2644  * shadow chains, and are converted to swap type as needed by swap
2645  * pager, we must report them.
2646  */
2647 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2648     CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2649     sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2650     "List of swap VM objects");
2651 
2652 #include "opt_ddb.h"
2653 #ifdef DDB
2654 #include <sys/kernel.h>
2655 
2656 #include <sys/cons.h>
2657 
2658 #include <ddb/ddb.h>
2659 
2660 static int
2661 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2662 {
2663 	vm_map_t tmpm;
2664 	vm_map_entry_t tmpe;
2665 	vm_object_t obj;
2666 
2667 	if (map == 0)
2668 		return 0;
2669 
2670 	if (entry == 0) {
2671 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2672 			if (_vm_object_in_map(map, object, tmpe)) {
2673 				return 1;
2674 			}
2675 		}
2676 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2677 		tmpm = entry->object.sub_map;
2678 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2679 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2680 				return 1;
2681 			}
2682 		}
2683 	} else if ((obj = entry->object.vm_object) != NULL) {
2684 		for (; obj; obj = obj->backing_object)
2685 			if (obj == object) {
2686 				return 1;
2687 			}
2688 	}
2689 	return 0;
2690 }
2691 
2692 static int
2693 vm_object_in_map(vm_object_t object)
2694 {
2695 	struct proc *p;
2696 
2697 	/* sx_slock(&allproc_lock); */
2698 	FOREACH_PROC_IN_SYSTEM(p) {
2699 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2700 			continue;
2701 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2702 			/* sx_sunlock(&allproc_lock); */
2703 			return 1;
2704 		}
2705 	}
2706 	/* sx_sunlock(&allproc_lock); */
2707 	if (_vm_object_in_map(kernel_map, object, 0))
2708 		return 1;
2709 	return 0;
2710 }
2711 
2712 DB_SHOW_COMMAND_FLAGS(vmochk, vm_object_check, DB_CMD_MEMSAFE)
2713 {
2714 	vm_object_t object;
2715 
2716 	/*
2717 	 * make sure that internal objs are in a map somewhere
2718 	 * and none have zero ref counts.
2719 	 */
2720 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2721 		if ((object->flags & OBJ_ANON) != 0) {
2722 			if (object->ref_count == 0) {
2723 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2724 					(long)object->size);
2725 			}
2726 			if (!vm_object_in_map(object)) {
2727 				db_printf(
2728 			"vmochk: internal obj is not in a map: "
2729 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2730 				    object->ref_count, (u_long)object->size,
2731 				    (u_long)object->size,
2732 				    (void *)object->backing_object);
2733 			}
2734 		}
2735 		if (db_pager_quit)
2736 			return;
2737 	}
2738 }
2739 
2740 /*
2741  *	vm_object_print:	[ debug ]
2742  */
2743 DB_SHOW_COMMAND(object, vm_object_print_static)
2744 {
2745 	/* XXX convert args. */
2746 	vm_object_t object = (vm_object_t)addr;
2747 	boolean_t full = have_addr;
2748 
2749 	vm_page_t p;
2750 
2751 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2752 #define	count	was_count
2753 
2754 	int count;
2755 
2756 	if (object == NULL)
2757 		return;
2758 
2759 	db_iprintf(
2760 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2761 	    object, (int)object->type, (uintmax_t)object->size,
2762 	    object->resident_page_count, object->ref_count, object->flags,
2763 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2764 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2765 	    atomic_load_int(&object->shadow_count),
2766 	    object->backing_object ? object->backing_object->ref_count : 0,
2767 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2768 
2769 	if (!full)
2770 		return;
2771 
2772 	db_indent += 2;
2773 	count = 0;
2774 	TAILQ_FOREACH(p, &object->memq, listq) {
2775 		if (count == 0)
2776 			db_iprintf("memory:=");
2777 		else if (count == 6) {
2778 			db_printf("\n");
2779 			db_iprintf(" ...");
2780 			count = 0;
2781 		} else
2782 			db_printf(",");
2783 		count++;
2784 
2785 		db_printf("(off=0x%jx,page=0x%jx)",
2786 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2787 
2788 		if (db_pager_quit)
2789 			break;
2790 	}
2791 	if (count != 0)
2792 		db_printf("\n");
2793 	db_indent -= 2;
2794 }
2795 
2796 /* XXX. */
2797 #undef count
2798 
2799 /* XXX need this non-static entry for calling from vm_map_print. */
2800 void
2801 vm_object_print(
2802         /* db_expr_t */ long addr,
2803 	boolean_t have_addr,
2804 	/* db_expr_t */ long count,
2805 	char *modif)
2806 {
2807 	vm_object_print_static(addr, have_addr, count, modif);
2808 }
2809 
2810 DB_SHOW_COMMAND_FLAGS(vmopag, vm_object_print_pages, DB_CMD_MEMSAFE)
2811 {
2812 	vm_object_t object;
2813 	vm_pindex_t fidx;
2814 	vm_paddr_t pa;
2815 	vm_page_t m, prev_m;
2816 	int rcount;
2817 
2818 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2819 		db_printf("new object: %p\n", (void *)object);
2820 		if (db_pager_quit)
2821 			return;
2822 
2823 		rcount = 0;
2824 		fidx = 0;
2825 		pa = -1;
2826 		TAILQ_FOREACH(m, &object->memq, listq) {
2827 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2828 			    prev_m->pindex + 1 != m->pindex) {
2829 				if (rcount) {
2830 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2831 						(long)fidx, rcount, (long)pa);
2832 					if (db_pager_quit)
2833 						return;
2834 					rcount = 0;
2835 				}
2836 			}
2837 			if (rcount &&
2838 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2839 				++rcount;
2840 				continue;
2841 			}
2842 			if (rcount) {
2843 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2844 					(long)fidx, rcount, (long)pa);
2845 				if (db_pager_quit)
2846 					return;
2847 			}
2848 			fidx = m->pindex;
2849 			pa = VM_PAGE_TO_PHYS(m);
2850 			rcount = 1;
2851 		}
2852 		if (rcount) {
2853 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2854 				(long)fidx, rcount, (long)pa);
2855 			if (db_pager_quit)
2856 				return;
2857 		}
2858 	}
2859 }
2860 #endif /* DDB */
2861