xref: /freebsd/sys/vm/vm_object.c (revision 4d65a7c6951cea0333f1a0c1b32c38489cdfa6c5)
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 <sys/cdefs.h>
66 #include "opt_vm.h"
67 
68 #include <sys/param.h>
69 #include <sys/systm.h>
70 #include <sys/blockcount.h>
71 #include <sys/cpuset.h>
72 #include <sys/limits.h>
73 #include <sys/lock.h>
74 #include <sys/mman.h>
75 #include <sys/mount.h>
76 #include <sys/kernel.h>
77 #include <sys/pctrie.h>
78 #include <sys/sysctl.h>
79 #include <sys/mutex.h>
80 #include <sys/proc.h>		/* for curproc, pageproc */
81 #include <sys/refcount.h>
82 #include <sys/socket.h>
83 #include <sys/resourcevar.h>
84 #include <sys/refcount.h>
85 #include <sys/rwlock.h>
86 #include <sys/user.h>
87 #include <sys/vnode.h>
88 #include <sys/vmmeter.h>
89 #include <sys/sx.h>
90 
91 #include <vm/vm.h>
92 #include <vm/vm_param.h>
93 #include <vm/pmap.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_object.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_pageout.h>
98 #include <vm/vm_pager.h>
99 #include <vm/vm_phys.h>
100 #include <vm/vm_pagequeue.h>
101 #include <vm/swap_pager.h>
102 #include <vm/vm_kern.h>
103 #include <vm/vm_extern.h>
104 #include <vm/vm_radix.h>
105 #include <vm/vm_reserv.h>
106 #include <vm/uma.h>
107 
108 static int old_msync;
109 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
110     "Use old (insecure) msync behavior");
111 
112 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
113 		    int pagerflags, int flags, boolean_t *allclean,
114 		    boolean_t *eio);
115 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
116 		    boolean_t *allclean);
117 static void	vm_object_backing_remove(vm_object_t object);
118 
119 /*
120  *	Virtual memory objects maintain the actual data
121  *	associated with allocated virtual memory.  A given
122  *	page of memory exists within exactly one object.
123  *
124  *	An object is only deallocated when all "references"
125  *	are given up.  Only one "reference" to a given
126  *	region of an object should be writeable.
127  *
128  *	Associated with each object is a list of all resident
129  *	memory pages belonging to that object; this list is
130  *	maintained by the "vm_page" module, and locked by the object's
131  *	lock.
132  *
133  *	Each object also records a "pager" routine which is
134  *	used to retrieve (and store) pages to the proper backing
135  *	storage.  In addition, objects may be backed by other
136  *	objects from which they were virtual-copied.
137  *
138  *	The only items within the object structure which are
139  *	modified after time of creation are:
140  *		reference count		locked by object's lock
141  *		pager routine		locked by object's lock
142  *
143  */
144 
145 struct object_q vm_object_list;
146 struct mtx vm_object_list_mtx;	/* lock for object list and count */
147 
148 struct vm_object kernel_object_store;
149 
150 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
151     "VM object stats");
152 
153 static COUNTER_U64_DEFINE_EARLY(object_collapses);
154 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
155     &object_collapses,
156     "VM object collapses");
157 
158 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
159 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
160     &object_bypasses,
161     "VM object bypasses");
162 
163 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
164 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
165     &object_collapse_waits,
166     "Number of sleeps for collapse");
167 
168 static uma_zone_t obj_zone;
169 
170 static int vm_object_zinit(void *mem, int size, int flags);
171 
172 #ifdef INVARIANTS
173 static void vm_object_zdtor(void *mem, int size, void *arg);
174 
175 static void
176 vm_object_zdtor(void *mem, int size, void *arg)
177 {
178 	vm_object_t object;
179 
180 	object = (vm_object_t)mem;
181 	KASSERT(object->ref_count == 0,
182 	    ("object %p ref_count = %d", object, object->ref_count));
183 	KASSERT(TAILQ_EMPTY(&object->memq),
184 	    ("object %p has resident pages in its memq", object));
185 	KASSERT(vm_radix_is_empty(&object->rtree),
186 	    ("object %p has resident pages in its trie", object));
187 #if VM_NRESERVLEVEL > 0
188 	KASSERT(LIST_EMPTY(&object->rvq),
189 	    ("object %p has reservations",
190 	    object));
191 #endif
192 	KASSERT(!vm_object_busied(object),
193 	    ("object %p busy = %d", object, blockcount_read(&object->busy)));
194 	KASSERT(object->resident_page_count == 0,
195 	    ("object %p resident_page_count = %d",
196 	    object, object->resident_page_count));
197 	KASSERT(atomic_load_int(&object->shadow_count) == 0,
198 	    ("object %p shadow_count = %d",
199 	    object, atomic_load_int(&object->shadow_count)));
200 	KASSERT(object->type == OBJT_DEAD,
201 	    ("object %p has non-dead type %d",
202 	    object, object->type));
203 	KASSERT(object->charge == 0 && object->cred == NULL,
204 	    ("object %p has non-zero charge %ju (%p)",
205 	    object, (uintmax_t)object->charge, object->cred));
206 }
207 #endif
208 
209 static int
210 vm_object_zinit(void *mem, int size, int flags)
211 {
212 	vm_object_t object;
213 
214 	object = (vm_object_t)mem;
215 	rw_init_flags(&object->lock, "vmobject", RW_DUPOK | RW_NEW);
216 
217 	/* These are true for any object that has been freed */
218 	object->type = OBJT_DEAD;
219 	vm_radix_init(&object->rtree);
220 	refcount_init(&object->ref_count, 0);
221 	blockcount_init(&object->paging_in_progress);
222 	blockcount_init(&object->busy);
223 	object->resident_page_count = 0;
224 	atomic_store_int(&object->shadow_count, 0);
225 	object->flags = OBJ_DEAD;
226 
227 	mtx_lock(&vm_object_list_mtx);
228 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
229 	mtx_unlock(&vm_object_list_mtx);
230 	return (0);
231 }
232 
233 static void
234 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
235     vm_object_t object, void *handle)
236 {
237 
238 	TAILQ_INIT(&object->memq);
239 	LIST_INIT(&object->shadow_head);
240 
241 	object->type = type;
242 	object->flags = flags;
243 	if ((flags & OBJ_SWAP) != 0) {
244 		pctrie_init(&object->un_pager.swp.swp_blks);
245 		object->un_pager.swp.writemappings = 0;
246 	}
247 
248 	/*
249 	 * Ensure that swap_pager_swapoff() iteration over object_list
250 	 * sees up to date type and pctrie head if it observed
251 	 * non-dead object.
252 	 */
253 	atomic_thread_fence_rel();
254 
255 	object->pg_color = 0;
256 	object->size = size;
257 	object->domain.dr_policy = NULL;
258 	object->generation = 1;
259 	object->cleangeneration = 1;
260 	refcount_init(&object->ref_count, 1);
261 	object->memattr = VM_MEMATTR_DEFAULT;
262 	object->cred = NULL;
263 	object->charge = 0;
264 	object->handle = handle;
265 	object->backing_object = NULL;
266 	object->backing_object_offset = (vm_ooffset_t) 0;
267 #if VM_NRESERVLEVEL > 0
268 	LIST_INIT(&object->rvq);
269 #endif
270 	umtx_shm_object_init(object);
271 }
272 
273 /*
274  *	vm_object_init:
275  *
276  *	Initialize the VM objects module.
277  */
278 void
279 vm_object_init(void)
280 {
281 	TAILQ_INIT(&vm_object_list);
282 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
283 
284 	rw_init(&kernel_object->lock, "kernel vm object");
285 	vm_radix_init(&kernel_object->rtree);
286 	_vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
287 	    VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
288 #if VM_NRESERVLEVEL > 0
289 	kernel_object->flags |= OBJ_COLORED;
290 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
291 #endif
292 	kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
293 
294 	/*
295 	 * The lock portion of struct vm_object must be type stable due
296 	 * to vm_pageout_fallback_object_lock locking a vm object
297 	 * without holding any references to it.
298 	 *
299 	 * paging_in_progress is valid always.  Lockless references to
300 	 * the objects may acquire pip and then check OBJ_DEAD.
301 	 */
302 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
303 #ifdef INVARIANTS
304 	    vm_object_zdtor,
305 #else
306 	    NULL,
307 #endif
308 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
309 
310 	vm_radix_zinit();
311 }
312 
313 void
314 vm_object_clear_flag(vm_object_t object, u_short bits)
315 {
316 
317 	VM_OBJECT_ASSERT_WLOCKED(object);
318 	object->flags &= ~bits;
319 }
320 
321 /*
322  *	Sets the default memory attribute for the specified object.  Pages
323  *	that are allocated to this object are by default assigned this memory
324  *	attribute.
325  *
326  *	Presently, this function must be called before any pages are allocated
327  *	to the object.  In the future, this requirement may be relaxed for
328  *	"default" and "swap" objects.
329  */
330 int
331 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
332 {
333 
334 	VM_OBJECT_ASSERT_WLOCKED(object);
335 
336 	if (object->type == OBJT_DEAD)
337 		return (KERN_INVALID_ARGUMENT);
338 	if (!TAILQ_EMPTY(&object->memq))
339 		return (KERN_FAILURE);
340 
341 	object->memattr = memattr;
342 	return (KERN_SUCCESS);
343 }
344 
345 void
346 vm_object_pip_add(vm_object_t object, short i)
347 {
348 
349 	if (i > 0)
350 		blockcount_acquire(&object->paging_in_progress, i);
351 }
352 
353 void
354 vm_object_pip_wakeup(vm_object_t object)
355 {
356 
357 	vm_object_pip_wakeupn(object, 1);
358 }
359 
360 void
361 vm_object_pip_wakeupn(vm_object_t object, short i)
362 {
363 
364 	if (i > 0)
365 		blockcount_release(&object->paging_in_progress, i);
366 }
367 
368 /*
369  * Atomically drop the object lock and wait for pip to drain.  This protects
370  * from sleep/wakeup races due to identity changes.  The lock is not re-acquired
371  * on return.
372  */
373 static void
374 vm_object_pip_sleep(vm_object_t object, const char *waitid)
375 {
376 
377 	(void)blockcount_sleep(&object->paging_in_progress, &object->lock,
378 	    waitid, PVM | PDROP);
379 }
380 
381 void
382 vm_object_pip_wait(vm_object_t object, const char *waitid)
383 {
384 
385 	VM_OBJECT_ASSERT_WLOCKED(object);
386 
387 	blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
388 	    PVM);
389 }
390 
391 void
392 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
393 {
394 
395 	VM_OBJECT_ASSERT_UNLOCKED(object);
396 
397 	blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
398 }
399 
400 /*
401  *	vm_object_allocate:
402  *
403  *	Returns a new object with the given size.
404  */
405 vm_object_t
406 vm_object_allocate(objtype_t type, vm_pindex_t size)
407 {
408 	vm_object_t object;
409 	u_short flags;
410 
411 	switch (type) {
412 	case OBJT_DEAD:
413 		panic("vm_object_allocate: can't create OBJT_DEAD");
414 	case OBJT_SWAP:
415 		flags = OBJ_COLORED | OBJ_SWAP;
416 		break;
417 	case OBJT_DEVICE:
418 	case OBJT_SG:
419 		flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
420 		break;
421 	case OBJT_MGTDEVICE:
422 		flags = OBJ_FICTITIOUS;
423 		break;
424 	case OBJT_PHYS:
425 		flags = OBJ_UNMANAGED;
426 		break;
427 	case OBJT_VNODE:
428 		flags = 0;
429 		break;
430 	default:
431 		panic("vm_object_allocate: type %d is undefined or dynamic",
432 		    type);
433 	}
434 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
435 	_vm_object_allocate(type, size, flags, object, NULL);
436 
437 	return (object);
438 }
439 
440 vm_object_t
441 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags)
442 {
443 	vm_object_t object;
444 
445 	MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */);
446 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
447 	_vm_object_allocate(dyntype, size, flags, object, NULL);
448 
449 	return (object);
450 }
451 
452 /*
453  *	vm_object_allocate_anon:
454  *
455  *	Returns a new default object of the given size and marked as
456  *	anonymous memory for special split/collapse handling.  Color
457  *	to be initialized by the caller.
458  */
459 vm_object_t
460 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
461     struct ucred *cred, vm_size_t charge)
462 {
463 	vm_object_t handle, object;
464 
465 	if (backing_object == NULL)
466 		handle = NULL;
467 	else if ((backing_object->flags & OBJ_ANON) != 0)
468 		handle = backing_object->handle;
469 	else
470 		handle = backing_object;
471 	object = uma_zalloc(obj_zone, M_WAITOK);
472 	_vm_object_allocate(OBJT_SWAP, size,
473 	    OBJ_ANON | OBJ_ONEMAPPING | OBJ_SWAP, object, handle);
474 	object->cred = cred;
475 	object->charge = cred != NULL ? charge : 0;
476 	return (object);
477 }
478 
479 static void
480 vm_object_reference_vnode(vm_object_t object)
481 {
482 	u_int old;
483 
484 	/*
485 	 * vnode objects need the lock for the first reference
486 	 * to serialize with vnode_object_deallocate().
487 	 */
488 	if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
489 		VM_OBJECT_RLOCK(object);
490 		old = refcount_acquire(&object->ref_count);
491 		if (object->type == OBJT_VNODE && old == 0)
492 			vref(object->handle);
493 		VM_OBJECT_RUNLOCK(object);
494 	}
495 }
496 
497 /*
498  *	vm_object_reference:
499  *
500  *	Acquires a reference to the given object.
501  */
502 void
503 vm_object_reference(vm_object_t object)
504 {
505 
506 	if (object == NULL)
507 		return;
508 
509 	if (object->type == OBJT_VNODE)
510 		vm_object_reference_vnode(object);
511 	else
512 		refcount_acquire(&object->ref_count);
513 	KASSERT((object->flags & OBJ_DEAD) == 0,
514 	    ("vm_object_reference: Referenced dead object."));
515 }
516 
517 /*
518  *	vm_object_reference_locked:
519  *
520  *	Gets another reference to the given object.
521  *
522  *	The object must be locked.
523  */
524 void
525 vm_object_reference_locked(vm_object_t object)
526 {
527 	u_int old;
528 
529 	VM_OBJECT_ASSERT_LOCKED(object);
530 	old = refcount_acquire(&object->ref_count);
531 	if (object->type == OBJT_VNODE && old == 0)
532 		vref(object->handle);
533 	KASSERT((object->flags & OBJ_DEAD) == 0,
534 	    ("vm_object_reference: Referenced dead object."));
535 }
536 
537 /*
538  * Handle deallocating an object of type OBJT_VNODE.
539  */
540 static void
541 vm_object_deallocate_vnode(vm_object_t object)
542 {
543 	struct vnode *vp = (struct vnode *) object->handle;
544 	bool last;
545 
546 	KASSERT(object->type == OBJT_VNODE,
547 	    ("vm_object_deallocate_vnode: not a vnode object"));
548 	KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
549 
550 	/* Object lock to protect handle lookup. */
551 	last = refcount_release(&object->ref_count);
552 	VM_OBJECT_RUNLOCK(object);
553 
554 	if (!last)
555 		return;
556 
557 	if (!umtx_shm_vnobj_persistent)
558 		umtx_shm_object_terminated(object);
559 
560 	/* vrele may need the vnode lock. */
561 	vrele(vp);
562 }
563 
564 /*
565  * We dropped a reference on an object and discovered that it had a
566  * single remaining shadow.  This is a sibling of the reference we
567  * dropped.  Attempt to collapse the sibling and backing object.
568  */
569 static vm_object_t
570 vm_object_deallocate_anon(vm_object_t backing_object)
571 {
572 	vm_object_t object;
573 
574 	/* Fetch the final shadow.  */
575 	object = LIST_FIRST(&backing_object->shadow_head);
576 	KASSERT(object != NULL &&
577 	    atomic_load_int(&backing_object->shadow_count) == 1,
578 	    ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
579 	    backing_object->ref_count,
580 	    atomic_load_int(&backing_object->shadow_count)));
581 	KASSERT((object->flags & OBJ_ANON) != 0,
582 	    ("invalid shadow object %p", object));
583 
584 	if (!VM_OBJECT_TRYWLOCK(object)) {
585 		/*
586 		 * Prevent object from disappearing since we do not have a
587 		 * reference.
588 		 */
589 		vm_object_pip_add(object, 1);
590 		VM_OBJECT_WUNLOCK(backing_object);
591 		VM_OBJECT_WLOCK(object);
592 		vm_object_pip_wakeup(object);
593 	} else
594 		VM_OBJECT_WUNLOCK(backing_object);
595 
596 	/*
597 	 * Check for a collapse/terminate race with the last reference holder.
598 	 */
599 	if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
600 	    !refcount_acquire_if_not_zero(&object->ref_count)) {
601 		VM_OBJECT_WUNLOCK(object);
602 		return (NULL);
603 	}
604 	backing_object = object->backing_object;
605 	if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
606 		vm_object_collapse(object);
607 	VM_OBJECT_WUNLOCK(object);
608 
609 	return (object);
610 }
611 
612 /*
613  *	vm_object_deallocate:
614  *
615  *	Release a reference to the specified object,
616  *	gained either through a vm_object_allocate
617  *	or a vm_object_reference call.  When all references
618  *	are gone, storage associated with this object
619  *	may be relinquished.
620  *
621  *	No object may be locked.
622  */
623 void
624 vm_object_deallocate(vm_object_t object)
625 {
626 	vm_object_t temp;
627 	bool released;
628 
629 	while (object != NULL) {
630 		/*
631 		 * If the reference count goes to 0 we start calling
632 		 * vm_object_terminate() on the object chain.  A ref count
633 		 * of 1 may be a special case depending on the shadow count
634 		 * being 0 or 1.  These cases require a write lock on the
635 		 * object.
636 		 */
637 		if ((object->flags & OBJ_ANON) == 0)
638 			released = refcount_release_if_gt(&object->ref_count, 1);
639 		else
640 			released = refcount_release_if_gt(&object->ref_count, 2);
641 		if (released)
642 			return;
643 
644 		if (object->type == OBJT_VNODE) {
645 			VM_OBJECT_RLOCK(object);
646 			if (object->type == OBJT_VNODE) {
647 				vm_object_deallocate_vnode(object);
648 				return;
649 			}
650 			VM_OBJECT_RUNLOCK(object);
651 		}
652 
653 		VM_OBJECT_WLOCK(object);
654 		KASSERT(object->ref_count > 0,
655 		    ("vm_object_deallocate: object deallocated too many times: %d",
656 		    object->type));
657 
658 		/*
659 		 * If this is not the final reference to an anonymous
660 		 * object we may need to collapse the shadow chain.
661 		 */
662 		if (!refcount_release(&object->ref_count)) {
663 			if (object->ref_count > 1 ||
664 			    atomic_load_int(&object->shadow_count) == 0) {
665 				if ((object->flags & OBJ_ANON) != 0 &&
666 				    object->ref_count == 1)
667 					vm_object_set_flag(object,
668 					    OBJ_ONEMAPPING);
669 				VM_OBJECT_WUNLOCK(object);
670 				return;
671 			}
672 
673 			/* Handle collapsing last ref on anonymous objects. */
674 			object = vm_object_deallocate_anon(object);
675 			continue;
676 		}
677 
678 		/*
679 		 * Handle the final reference to an object.  We restart
680 		 * the loop with the backing object to avoid recursion.
681 		 */
682 		umtx_shm_object_terminated(object);
683 		temp = object->backing_object;
684 		if (temp != NULL) {
685 			KASSERT(object->type == OBJT_SWAP,
686 			    ("shadowed tmpfs v_object 2 %p", object));
687 			vm_object_backing_remove(object);
688 		}
689 
690 		KASSERT((object->flags & OBJ_DEAD) == 0,
691 		    ("vm_object_deallocate: Terminating dead object."));
692 		vm_object_set_flag(object, OBJ_DEAD);
693 		vm_object_terminate(object);
694 		object = temp;
695 	}
696 }
697 
698 void
699 vm_object_destroy(vm_object_t object)
700 {
701 	uma_zfree(obj_zone, object);
702 }
703 
704 static void
705 vm_object_sub_shadow(vm_object_t object)
706 {
707 	KASSERT(object->shadow_count >= 1,
708 	    ("object %p sub_shadow count zero", object));
709 	atomic_subtract_int(&object->shadow_count, 1);
710 }
711 
712 static void
713 vm_object_backing_remove_locked(vm_object_t object)
714 {
715 	vm_object_t backing_object;
716 
717 	backing_object = object->backing_object;
718 	VM_OBJECT_ASSERT_WLOCKED(object);
719 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
720 
721 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
722 	    ("vm_object_backing_remove: Removing collapsing object."));
723 
724 	vm_object_sub_shadow(backing_object);
725 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
726 		LIST_REMOVE(object, shadow_list);
727 		vm_object_clear_flag(object, OBJ_SHADOWLIST);
728 	}
729 	object->backing_object = NULL;
730 }
731 
732 static void
733 vm_object_backing_remove(vm_object_t object)
734 {
735 	vm_object_t backing_object;
736 
737 	VM_OBJECT_ASSERT_WLOCKED(object);
738 
739 	backing_object = object->backing_object;
740 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
741 		VM_OBJECT_WLOCK(backing_object);
742 		vm_object_backing_remove_locked(object);
743 		VM_OBJECT_WUNLOCK(backing_object);
744 	} else {
745 		object->backing_object = NULL;
746 		vm_object_sub_shadow(backing_object);
747 	}
748 }
749 
750 static void
751 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
752 {
753 
754 	VM_OBJECT_ASSERT_WLOCKED(object);
755 
756 	atomic_add_int(&backing_object->shadow_count, 1);
757 	if ((backing_object->flags & OBJ_ANON) != 0) {
758 		VM_OBJECT_ASSERT_WLOCKED(backing_object);
759 		LIST_INSERT_HEAD(&backing_object->shadow_head, object,
760 		    shadow_list);
761 		vm_object_set_flag(object, OBJ_SHADOWLIST);
762 	}
763 	object->backing_object = backing_object;
764 }
765 
766 static void
767 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
768 {
769 
770 	VM_OBJECT_ASSERT_WLOCKED(object);
771 
772 	if ((backing_object->flags & OBJ_ANON) != 0) {
773 		VM_OBJECT_WLOCK(backing_object);
774 		vm_object_backing_insert_locked(object, backing_object);
775 		VM_OBJECT_WUNLOCK(backing_object);
776 	} else {
777 		object->backing_object = backing_object;
778 		atomic_add_int(&backing_object->shadow_count, 1);
779 	}
780 }
781 
782 /*
783  * Insert an object into a backing_object's shadow list with an additional
784  * reference to the backing_object added.
785  */
786 static void
787 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
788 {
789 
790 	VM_OBJECT_ASSERT_WLOCKED(object);
791 
792 	if ((backing_object->flags & OBJ_ANON) != 0) {
793 		VM_OBJECT_WLOCK(backing_object);
794 		KASSERT((backing_object->flags & OBJ_DEAD) == 0,
795 		    ("shadowing dead anonymous object"));
796 		vm_object_reference_locked(backing_object);
797 		vm_object_backing_insert_locked(object, backing_object);
798 		vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
799 		VM_OBJECT_WUNLOCK(backing_object);
800 	} else {
801 		vm_object_reference(backing_object);
802 		atomic_add_int(&backing_object->shadow_count, 1);
803 		object->backing_object = backing_object;
804 	}
805 }
806 
807 /*
808  * Transfer a backing reference from backing_object to object.
809  */
810 static void
811 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
812 {
813 	vm_object_t new_backing_object;
814 
815 	/*
816 	 * Note that the reference to backing_object->backing_object
817 	 * moves from within backing_object to within object.
818 	 */
819 	vm_object_backing_remove_locked(object);
820 	new_backing_object = backing_object->backing_object;
821 	if (new_backing_object == NULL)
822 		return;
823 	if ((new_backing_object->flags & OBJ_ANON) != 0) {
824 		VM_OBJECT_WLOCK(new_backing_object);
825 		vm_object_backing_remove_locked(backing_object);
826 		vm_object_backing_insert_locked(object, new_backing_object);
827 		VM_OBJECT_WUNLOCK(new_backing_object);
828 	} else {
829 		/*
830 		 * shadow_count for new_backing_object is left
831 		 * unchanged, its reference provided by backing_object
832 		 * is replaced by object.
833 		 */
834 		object->backing_object = new_backing_object;
835 		backing_object->backing_object = NULL;
836 	}
837 }
838 
839 /*
840  * Wait for a concurrent collapse to settle.
841  */
842 static void
843 vm_object_collapse_wait(vm_object_t object)
844 {
845 
846 	VM_OBJECT_ASSERT_WLOCKED(object);
847 
848 	while ((object->flags & OBJ_COLLAPSING) != 0) {
849 		vm_object_pip_wait(object, "vmcolwait");
850 		counter_u64_add(object_collapse_waits, 1);
851 	}
852 }
853 
854 /*
855  * Waits for a backing object to clear a pending collapse and returns
856  * it locked if it is an ANON object.
857  */
858 static vm_object_t
859 vm_object_backing_collapse_wait(vm_object_t object)
860 {
861 	vm_object_t backing_object;
862 
863 	VM_OBJECT_ASSERT_WLOCKED(object);
864 
865 	for (;;) {
866 		backing_object = object->backing_object;
867 		if (backing_object == NULL ||
868 		    (backing_object->flags & OBJ_ANON) == 0)
869 			return (NULL);
870 		VM_OBJECT_WLOCK(backing_object);
871 		if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
872 			break;
873 		VM_OBJECT_WUNLOCK(object);
874 		vm_object_pip_sleep(backing_object, "vmbckwait");
875 		counter_u64_add(object_collapse_waits, 1);
876 		VM_OBJECT_WLOCK(object);
877 	}
878 	return (backing_object);
879 }
880 
881 /*
882  *	vm_object_terminate_pages removes any remaining pageable pages
883  *	from the object and resets the object to an empty state.
884  */
885 static void
886 vm_object_terminate_pages(vm_object_t object)
887 {
888 	vm_page_t p, p_next;
889 
890 	VM_OBJECT_ASSERT_WLOCKED(object);
891 
892 	/*
893 	 * Free any remaining pageable pages.  This also removes them from the
894 	 * paging queues.  However, don't free wired pages, just remove them
895 	 * from the object.  Rather than incrementally removing each page from
896 	 * the object, the page and object are reset to any empty state.
897 	 */
898 	TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
899 		vm_page_assert_unbusied(p);
900 		KASSERT(p->object == object &&
901 		    (p->ref_count & VPRC_OBJREF) != 0,
902 		    ("vm_object_terminate_pages: page %p is inconsistent", p));
903 
904 		p->object = NULL;
905 		if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
906 			VM_CNT_INC(v_pfree);
907 			vm_page_free(p);
908 		}
909 	}
910 
911 	/*
912 	 * If the object contained any pages, then reset it to an empty state.
913 	 * None of the object's fields, including "resident_page_count", were
914 	 * modified by the preceding loop.
915 	 */
916 	if (object->resident_page_count != 0) {
917 		vm_radix_reclaim_allnodes(&object->rtree);
918 		TAILQ_INIT(&object->memq);
919 		object->resident_page_count = 0;
920 		if (object->type == OBJT_VNODE)
921 			vdrop(object->handle);
922 	}
923 }
924 
925 /*
926  *	vm_object_terminate actually destroys the specified object, freeing
927  *	up all previously used resources.
928  *
929  *	The object must be locked.
930  *	This routine may block.
931  */
932 void
933 vm_object_terminate(vm_object_t object)
934 {
935 
936 	VM_OBJECT_ASSERT_WLOCKED(object);
937 	KASSERT((object->flags & OBJ_DEAD) != 0,
938 	    ("terminating non-dead obj %p", object));
939 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
940 	    ("terminating collapsing obj %p", object));
941 	KASSERT(object->backing_object == NULL,
942 	    ("terminating shadow obj %p", object));
943 
944 	/*
945 	 * Wait for the pageout daemon and other current users to be
946 	 * done with the object.  Note that new paging_in_progress
947 	 * users can come after this wait, but they must check
948 	 * OBJ_DEAD flag set (without unlocking the object), and avoid
949 	 * the object being terminated.
950 	 */
951 	vm_object_pip_wait(object, "objtrm");
952 
953 	KASSERT(object->ref_count == 0,
954 	    ("vm_object_terminate: object with references, ref_count=%d",
955 	    object->ref_count));
956 
957 	if ((object->flags & OBJ_PG_DTOR) == 0)
958 		vm_object_terminate_pages(object);
959 
960 #if VM_NRESERVLEVEL > 0
961 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
962 		vm_reserv_break_all(object);
963 #endif
964 
965 	KASSERT(object->cred == NULL || (object->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[vm_pageout_page_count], p_first, tp;
1121 	int count, i, mreq, runlen;
1122 
1123 	vm_page_lock_assert(p, MA_NOTOWNED);
1124 	vm_page_assert_xbusied(p);
1125 	VM_OBJECT_ASSERT_WLOCKED(object);
1126 
1127 	count = 1;
1128 	mreq = 0;
1129 
1130 	for (tp = p; count < vm_pageout_page_count; count++) {
1131 		tp = vm_page_next(tp);
1132 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1133 			break;
1134 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1135 			vm_page_xunbusy(tp);
1136 			break;
1137 		}
1138 	}
1139 
1140 	for (p_first = p; count < vm_pageout_page_count; count++) {
1141 		tp = vm_page_prev(p_first);
1142 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1143 			break;
1144 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1145 			vm_page_xunbusy(tp);
1146 			break;
1147 		}
1148 		p_first = tp;
1149 		mreq++;
1150 	}
1151 
1152 	for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1153 		ma[i] = tp;
1154 
1155 	vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1156 	return (runlen);
1157 }
1158 
1159 /*
1160  * Note that there is absolutely no sense in writing out
1161  * anonymous objects, so we track down the vnode object
1162  * to write out.
1163  * We invalidate (remove) all pages from the address space
1164  * for semantic correctness.
1165  *
1166  * If the backing object is a device object with unmanaged pages, then any
1167  * mappings to the specified range of pages must be removed before this
1168  * function is called.
1169  *
1170  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1171  * may start out with a NULL object.
1172  */
1173 boolean_t
1174 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1175     boolean_t syncio, boolean_t invalidate)
1176 {
1177 	vm_object_t backing_object;
1178 	struct vnode *vp;
1179 	struct mount *mp;
1180 	int error, flags, fsync_after;
1181 	boolean_t res;
1182 
1183 	if (object == NULL)
1184 		return (TRUE);
1185 	res = TRUE;
1186 	error = 0;
1187 	VM_OBJECT_WLOCK(object);
1188 	while ((backing_object = object->backing_object) != NULL) {
1189 		VM_OBJECT_WLOCK(backing_object);
1190 		offset += object->backing_object_offset;
1191 		VM_OBJECT_WUNLOCK(object);
1192 		object = backing_object;
1193 		if (object->size < OFF_TO_IDX(offset + size))
1194 			size = IDX_TO_OFF(object->size) - offset;
1195 	}
1196 	/*
1197 	 * Flush pages if writing is allowed, invalidate them
1198 	 * if invalidation requested.  Pages undergoing I/O
1199 	 * will be ignored by vm_object_page_remove().
1200 	 *
1201 	 * We cannot lock the vnode and then wait for paging
1202 	 * to complete without deadlocking against vm_fault.
1203 	 * Instead we simply call vm_object_page_remove() and
1204 	 * allow it to block internally on a page-by-page
1205 	 * basis when it encounters pages undergoing async
1206 	 * I/O.
1207 	 */
1208 	if (object->type == OBJT_VNODE &&
1209 	    vm_object_mightbedirty(object) != 0 &&
1210 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1211 		VM_OBJECT_WUNLOCK(object);
1212 		(void)vn_start_write(vp, &mp, V_WAIT);
1213 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1214 		if (syncio && !invalidate && offset == 0 &&
1215 		    atop(size) == object->size) {
1216 			/*
1217 			 * If syncing the whole mapping of the file,
1218 			 * it is faster to schedule all the writes in
1219 			 * async mode, also allowing the clustering,
1220 			 * and then wait for i/o to complete.
1221 			 */
1222 			flags = 0;
1223 			fsync_after = TRUE;
1224 		} else {
1225 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1226 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1227 			fsync_after = FALSE;
1228 		}
1229 		VM_OBJECT_WLOCK(object);
1230 		res = vm_object_page_clean(object, offset, offset + size,
1231 		    flags);
1232 		VM_OBJECT_WUNLOCK(object);
1233 		if (fsync_after) {
1234 			for (;;) {
1235 				error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1236 				if (error != ERELOOKUP)
1237 					break;
1238 
1239 				/*
1240 				 * Allow SU/bufdaemon to handle more
1241 				 * dependencies in the meantime.
1242 				 */
1243 				VOP_UNLOCK(vp);
1244 				vn_finished_write(mp);
1245 
1246 				(void)vn_start_write(vp, &mp, V_WAIT);
1247 				vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1248 			}
1249 		}
1250 		VOP_UNLOCK(vp);
1251 		vn_finished_write(mp);
1252 		if (error != 0)
1253 			res = FALSE;
1254 		VM_OBJECT_WLOCK(object);
1255 	}
1256 	if ((object->type == OBJT_VNODE ||
1257 	     object->type == OBJT_DEVICE) && invalidate) {
1258 		if (object->type == OBJT_DEVICE)
1259 			/*
1260 			 * The option OBJPR_NOTMAPPED must be passed here
1261 			 * because vm_object_page_remove() cannot remove
1262 			 * unmanaged mappings.
1263 			 */
1264 			flags = OBJPR_NOTMAPPED;
1265 		else if (old_msync)
1266 			flags = 0;
1267 		else
1268 			flags = OBJPR_CLEANONLY;
1269 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1270 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1271 	}
1272 	VM_OBJECT_WUNLOCK(object);
1273 	return (res);
1274 }
1275 
1276 /*
1277  * Determine whether the given advice can be applied to the object.  Advice is
1278  * not applied to unmanaged pages since they never belong to page queues, and
1279  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1280  * have been mapped at most once.
1281  */
1282 static bool
1283 vm_object_advice_applies(vm_object_t object, int advice)
1284 {
1285 
1286 	if ((object->flags & OBJ_UNMANAGED) != 0)
1287 		return (false);
1288 	if (advice != MADV_FREE)
1289 		return (true);
1290 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1291 	    (OBJ_ONEMAPPING | OBJ_ANON));
1292 }
1293 
1294 static void
1295 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1296     vm_size_t size)
1297 {
1298 
1299 	if (advice == MADV_FREE)
1300 		vm_pager_freespace(object, pindex, size);
1301 }
1302 
1303 /*
1304  *	vm_object_madvise:
1305  *
1306  *	Implements the madvise function at the object/page level.
1307  *
1308  *	MADV_WILLNEED	(any object)
1309  *
1310  *	    Activate the specified pages if they are resident.
1311  *
1312  *	MADV_DONTNEED	(any object)
1313  *
1314  *	    Deactivate the specified pages if they are resident.
1315  *
1316  *	MADV_FREE	(OBJT_SWAP objects, OBJ_ONEMAPPING only)
1317  *
1318  *	    Deactivate and clean the specified pages if they are
1319  *	    resident.  This permits the process to reuse the pages
1320  *	    without faulting or the kernel to reclaim the pages
1321  *	    without I/O.
1322  */
1323 void
1324 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1325     int advice)
1326 {
1327 	vm_pindex_t tpindex;
1328 	vm_object_t backing_object, tobject;
1329 	vm_page_t m, tm;
1330 
1331 	if (object == NULL)
1332 		return;
1333 
1334 relookup:
1335 	VM_OBJECT_WLOCK(object);
1336 	if (!vm_object_advice_applies(object, advice)) {
1337 		VM_OBJECT_WUNLOCK(object);
1338 		return;
1339 	}
1340 	for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1341 		tobject = object;
1342 
1343 		/*
1344 		 * If the next page isn't resident in the top-level object, we
1345 		 * need to search the shadow chain.  When applying MADV_FREE, we
1346 		 * take care to release any swap space used to store
1347 		 * non-resident pages.
1348 		 */
1349 		if (m == NULL || pindex < m->pindex) {
1350 			/*
1351 			 * Optimize a common case: if the top-level object has
1352 			 * no backing object, we can skip over the non-resident
1353 			 * range in constant time.
1354 			 */
1355 			if (object->backing_object == NULL) {
1356 				tpindex = (m != NULL && m->pindex < end) ?
1357 				    m->pindex : end;
1358 				vm_object_madvise_freespace(object, advice,
1359 				    pindex, tpindex - pindex);
1360 				if ((pindex = tpindex) == end)
1361 					break;
1362 				goto next_page;
1363 			}
1364 
1365 			tpindex = pindex;
1366 			do {
1367 				vm_object_madvise_freespace(tobject, advice,
1368 				    tpindex, 1);
1369 				/*
1370 				 * Prepare to search the next object in the
1371 				 * chain.
1372 				 */
1373 				backing_object = tobject->backing_object;
1374 				if (backing_object == NULL)
1375 					goto next_pindex;
1376 				VM_OBJECT_WLOCK(backing_object);
1377 				tpindex +=
1378 				    OFF_TO_IDX(tobject->backing_object_offset);
1379 				if (tobject != object)
1380 					VM_OBJECT_WUNLOCK(tobject);
1381 				tobject = backing_object;
1382 				if (!vm_object_advice_applies(tobject, advice))
1383 					goto next_pindex;
1384 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1385 			    NULL);
1386 		} else {
1387 next_page:
1388 			tm = m;
1389 			m = TAILQ_NEXT(m, listq);
1390 		}
1391 
1392 		/*
1393 		 * If the page is not in a normal state, skip it.  The page
1394 		 * can not be invalidated while the object lock is held.
1395 		 */
1396 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1397 			goto next_pindex;
1398 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1399 		    ("vm_object_madvise: page %p is fictitious", tm));
1400 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1401 		    ("vm_object_madvise: page %p is not managed", tm));
1402 		if (vm_page_tryxbusy(tm) == 0) {
1403 			if (object != tobject)
1404 				VM_OBJECT_WUNLOCK(object);
1405 			if (advice == MADV_WILLNEED) {
1406 				/*
1407 				 * Reference the page before unlocking and
1408 				 * sleeping so that the page daemon is less
1409 				 * likely to reclaim it.
1410 				 */
1411 				vm_page_aflag_set(tm, PGA_REFERENCED);
1412 			}
1413 			if (!vm_page_busy_sleep(tm, "madvpo", 0))
1414 				VM_OBJECT_WUNLOCK(tobject);
1415   			goto relookup;
1416 		}
1417 		vm_page_advise(tm, advice);
1418 		vm_page_xunbusy(tm);
1419 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1420 next_pindex:
1421 		if (tobject != object)
1422 			VM_OBJECT_WUNLOCK(tobject);
1423 	}
1424 	VM_OBJECT_WUNLOCK(object);
1425 }
1426 
1427 /*
1428  *	vm_object_shadow:
1429  *
1430  *	Create a new object which is backed by the
1431  *	specified existing object range.  The source
1432  *	object reference is deallocated.
1433  *
1434  *	The new object and offset into that object
1435  *	are returned in the source parameters.
1436  */
1437 void
1438 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1439     struct ucred *cred, bool shared)
1440 {
1441 	vm_object_t source;
1442 	vm_object_t result;
1443 
1444 	source = *object;
1445 
1446 	/*
1447 	 * Don't create the new object if the old object isn't shared.
1448 	 *
1449 	 * If we hold the only reference we can guarantee that it won't
1450 	 * increase while we have the map locked.  Otherwise the race is
1451 	 * harmless and we will end up with an extra shadow object that
1452 	 * will be collapsed later.
1453 	 */
1454 	if (source != NULL && source->ref_count == 1 &&
1455 	    (source->flags & OBJ_ANON) != 0)
1456 		return;
1457 
1458 	/*
1459 	 * Allocate a new object with the given length.
1460 	 */
1461 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1462 
1463 	/*
1464 	 * Store the offset into the source object, and fix up the offset into
1465 	 * the new object.
1466 	 */
1467 	result->backing_object_offset = *offset;
1468 
1469 	if (shared || source != NULL) {
1470 		VM_OBJECT_WLOCK(result);
1471 
1472 		/*
1473 		 * The new object shadows the source object, adding a
1474 		 * reference to it.  Our caller changes his reference
1475 		 * to point to the new object, removing a reference to
1476 		 * the source object.  Net result: no change of
1477 		 * reference count, unless the caller needs to add one
1478 		 * more reference due to forking a shared map entry.
1479 		 */
1480 		if (shared) {
1481 			vm_object_reference_locked(result);
1482 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1483 		}
1484 
1485 		/*
1486 		 * Try to optimize the result object's page color when
1487 		 * shadowing in order to maintain page coloring
1488 		 * consistency in the combined shadowed object.
1489 		 */
1490 		if (source != NULL) {
1491 			vm_object_backing_insert(result, source);
1492 			result->domain = source->domain;
1493 #if VM_NRESERVLEVEL > 0
1494 			vm_object_set_flag(result,
1495 			    (source->flags & OBJ_COLORED));
1496 			result->pg_color = (source->pg_color +
1497 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1498 			    1)) - 1);
1499 #endif
1500 		}
1501 		VM_OBJECT_WUNLOCK(result);
1502 	}
1503 
1504 	/*
1505 	 * Return the new things
1506 	 */
1507 	*offset = 0;
1508 	*object = result;
1509 }
1510 
1511 /*
1512  *	vm_object_split:
1513  *
1514  * Split the pages in a map entry into a new object.  This affords
1515  * easier removal of unused pages, and keeps object inheritance from
1516  * being a negative impact on memory usage.
1517  */
1518 void
1519 vm_object_split(vm_map_entry_t entry)
1520 {
1521 	vm_page_t m, m_next;
1522 	vm_object_t orig_object, new_object, backing_object;
1523 	vm_pindex_t idx, offidxstart;
1524 	vm_size_t size;
1525 
1526 	orig_object = entry->object.vm_object;
1527 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1528 	    ("vm_object_split:  Splitting object with multiple mappings."));
1529 	if ((orig_object->flags & OBJ_ANON) == 0)
1530 		return;
1531 	if (orig_object->ref_count <= 1)
1532 		return;
1533 	VM_OBJECT_WUNLOCK(orig_object);
1534 
1535 	offidxstart = OFF_TO_IDX(entry->offset);
1536 	size = atop(entry->end - entry->start);
1537 
1538 	new_object = vm_object_allocate_anon(size, orig_object,
1539 	    orig_object->cred, ptoa(size));
1540 
1541 	/*
1542 	 * We must wait for the orig_object to complete any in-progress
1543 	 * collapse so that the swap blocks are stable below.  The
1544 	 * additional reference on backing_object by new object will
1545 	 * prevent further collapse operations until split completes.
1546 	 */
1547 	VM_OBJECT_WLOCK(orig_object);
1548 	vm_object_collapse_wait(orig_object);
1549 
1550 	/*
1551 	 * At this point, the new object is still private, so the order in
1552 	 * which the original and new objects are locked does not matter.
1553 	 */
1554 	VM_OBJECT_WLOCK(new_object);
1555 	new_object->domain = orig_object->domain;
1556 	backing_object = orig_object->backing_object;
1557 	if (backing_object != NULL) {
1558 		vm_object_backing_insert_ref(new_object, backing_object);
1559 		new_object->backing_object_offset =
1560 		    orig_object->backing_object_offset + entry->offset;
1561 	}
1562 	if (orig_object->cred != NULL) {
1563 		crhold(orig_object->cred);
1564 		KASSERT(orig_object->charge >= ptoa(size),
1565 		    ("orig_object->charge < 0"));
1566 		orig_object->charge -= ptoa(size);
1567 	}
1568 
1569 	/*
1570 	 * Mark the split operation so that swap_pager_getpages() knows
1571 	 * that the object is in transition.
1572 	 */
1573 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1574 #ifdef INVARIANTS
1575 	idx = 0;
1576 #endif
1577 retry:
1578 	m = vm_page_find_least(orig_object, offidxstart);
1579 	KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1580 	    ("%s: object %p was repopulated", __func__, orig_object));
1581 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1582 	    m = m_next) {
1583 		m_next = TAILQ_NEXT(m, listq);
1584 
1585 		/*
1586 		 * We must wait for pending I/O to complete before we can
1587 		 * rename the page.
1588 		 *
1589 		 * We do not have to VM_PROT_NONE the page as mappings should
1590 		 * not be changed by this operation.
1591 		 */
1592 		if (vm_page_tryxbusy(m) == 0) {
1593 			VM_OBJECT_WUNLOCK(new_object);
1594 			if (vm_page_busy_sleep(m, "spltwt", 0))
1595 				VM_OBJECT_WLOCK(orig_object);
1596 			VM_OBJECT_WLOCK(new_object);
1597 			goto retry;
1598 		}
1599 
1600 		/*
1601 		 * The page was left invalid.  Likely placed there by
1602 		 * an incomplete fault.  Just remove and ignore.
1603 		 */
1604 		if (vm_page_none_valid(m)) {
1605 			if (vm_page_remove(m))
1606 				vm_page_free(m);
1607 			continue;
1608 		}
1609 
1610 		/* vm_page_rename() will dirty the page. */
1611 		if (vm_page_rename(m, new_object, idx)) {
1612 			vm_page_xunbusy(m);
1613 			VM_OBJECT_WUNLOCK(new_object);
1614 			VM_OBJECT_WUNLOCK(orig_object);
1615 			vm_radix_wait();
1616 			VM_OBJECT_WLOCK(orig_object);
1617 			VM_OBJECT_WLOCK(new_object);
1618 			goto retry;
1619 		}
1620 
1621 #if VM_NRESERVLEVEL > 0
1622 		/*
1623 		 * If some of the reservation's allocated pages remain with
1624 		 * the original object, then transferring the reservation to
1625 		 * the new object is neither particularly beneficial nor
1626 		 * particularly harmful as compared to leaving the reservation
1627 		 * with the original object.  If, however, all of the
1628 		 * reservation's allocated pages are transferred to the new
1629 		 * object, then transferring the reservation is typically
1630 		 * beneficial.  Determining which of these two cases applies
1631 		 * would be more costly than unconditionally renaming the
1632 		 * reservation.
1633 		 */
1634 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1635 #endif
1636 	}
1637 
1638 	/*
1639 	 * swap_pager_copy() can sleep, in which case the orig_object's
1640 	 * and new_object's locks are released and reacquired.
1641 	 */
1642 	swap_pager_copy(orig_object, new_object, offidxstart, 0);
1643 
1644 	TAILQ_FOREACH(m, &new_object->memq, listq)
1645 		vm_page_xunbusy(m);
1646 
1647 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1648 	VM_OBJECT_WUNLOCK(orig_object);
1649 	VM_OBJECT_WUNLOCK(new_object);
1650 	entry->object.vm_object = new_object;
1651 	entry->offset = 0LL;
1652 	vm_object_deallocate(orig_object);
1653 	VM_OBJECT_WLOCK(new_object);
1654 }
1655 
1656 static vm_page_t
1657 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1658 {
1659 	vm_object_t backing_object;
1660 
1661 	VM_OBJECT_ASSERT_WLOCKED(object);
1662 	backing_object = object->backing_object;
1663 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1664 
1665 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1666 	    ("invalid ownership %p %p %p", p, object, backing_object));
1667 	/* The page is only NULL when rename fails. */
1668 	if (p == NULL) {
1669 		VM_OBJECT_WUNLOCK(object);
1670 		VM_OBJECT_WUNLOCK(backing_object);
1671 		vm_radix_wait();
1672 		VM_OBJECT_WLOCK(object);
1673 	} else if (p->object == object) {
1674 		VM_OBJECT_WUNLOCK(backing_object);
1675 		if (vm_page_busy_sleep(p, "vmocol", 0))
1676 			VM_OBJECT_WLOCK(object);
1677 	} else {
1678 		VM_OBJECT_WUNLOCK(object);
1679 		if (!vm_page_busy_sleep(p, "vmocol", 0))
1680 			VM_OBJECT_WUNLOCK(backing_object);
1681 		VM_OBJECT_WLOCK(object);
1682 	}
1683 	VM_OBJECT_WLOCK(backing_object);
1684 	return (TAILQ_FIRST(&backing_object->memq));
1685 }
1686 
1687 static bool
1688 vm_object_scan_all_shadowed(vm_object_t object)
1689 {
1690 	vm_object_t backing_object;
1691 	vm_page_t p, pp;
1692 	vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1693 
1694 	VM_OBJECT_ASSERT_WLOCKED(object);
1695 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1696 
1697 	backing_object = object->backing_object;
1698 
1699 	if ((backing_object->flags & OBJ_ANON) == 0)
1700 		return (false);
1701 
1702 	pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1703 	p = vm_page_find_least(backing_object, pi);
1704 	ps = swap_pager_find_least(backing_object, pi);
1705 
1706 	/*
1707 	 * Only check pages inside the parent object's range and
1708 	 * inside the parent object's mapping of the backing object.
1709 	 */
1710 	for (;; pi++) {
1711 		if (p != NULL && p->pindex < pi)
1712 			p = TAILQ_NEXT(p, listq);
1713 		if (ps < pi)
1714 			ps = swap_pager_find_least(backing_object, pi);
1715 		if (p == NULL && ps >= backing_object->size)
1716 			break;
1717 		else if (p == NULL)
1718 			pi = ps;
1719 		else
1720 			pi = MIN(p->pindex, ps);
1721 
1722 		new_pindex = pi - backing_offset_index;
1723 		if (new_pindex >= object->size)
1724 			break;
1725 
1726 		if (p != NULL) {
1727 			/*
1728 			 * If the backing object page is busy a
1729 			 * grandparent or older page may still be
1730 			 * undergoing CoW.  It is not safe to collapse
1731 			 * the backing object until it is quiesced.
1732 			 */
1733 			if (vm_page_tryxbusy(p) == 0)
1734 				return (false);
1735 
1736 			/*
1737 			 * We raced with the fault handler that left
1738 			 * newly allocated invalid page on the object
1739 			 * queue and retried.
1740 			 */
1741 			if (!vm_page_all_valid(p))
1742 				goto unbusy_ret;
1743 		}
1744 
1745 		/*
1746 		 * See if the parent has the page or if the parent's object
1747 		 * pager has the page.  If the parent has the page but the page
1748 		 * is not valid, the parent's object pager must have the page.
1749 		 *
1750 		 * If this fails, the parent does not completely shadow the
1751 		 * object and we might as well give up now.
1752 		 */
1753 		pp = vm_page_lookup(object, new_pindex);
1754 
1755 		/*
1756 		 * The valid check here is stable due to object lock
1757 		 * being required to clear valid and initiate paging.
1758 		 * Busy of p disallows fault handler to validate pp.
1759 		 */
1760 		if ((pp == NULL || vm_page_none_valid(pp)) &&
1761 		    !vm_pager_has_page(object, new_pindex, NULL, NULL))
1762 			goto unbusy_ret;
1763 		if (p != NULL)
1764 			vm_page_xunbusy(p);
1765 	}
1766 	return (true);
1767 
1768 unbusy_ret:
1769 	if (p != NULL)
1770 		vm_page_xunbusy(p);
1771 	return (false);
1772 }
1773 
1774 static void
1775 vm_object_collapse_scan(vm_object_t object)
1776 {
1777 	vm_object_t backing_object;
1778 	vm_page_t next, p, pp;
1779 	vm_pindex_t backing_offset_index, new_pindex;
1780 
1781 	VM_OBJECT_ASSERT_WLOCKED(object);
1782 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1783 
1784 	backing_object = object->backing_object;
1785 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1786 
1787 	/*
1788 	 * Our scan
1789 	 */
1790 	for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1791 		next = TAILQ_NEXT(p, listq);
1792 		new_pindex = p->pindex - backing_offset_index;
1793 
1794 		/*
1795 		 * Check for busy page
1796 		 */
1797 		if (vm_page_tryxbusy(p) == 0) {
1798 			next = vm_object_collapse_scan_wait(object, p);
1799 			continue;
1800 		}
1801 
1802 		KASSERT(object->backing_object == backing_object,
1803 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1804 		    object->backing_object, backing_object));
1805 		KASSERT(p->object == backing_object,
1806 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1807 		    p->object, backing_object));
1808 
1809 		if (p->pindex < backing_offset_index ||
1810 		    new_pindex >= object->size) {
1811 			vm_pager_freespace(backing_object, p->pindex, 1);
1812 
1813 			KASSERT(!pmap_page_is_mapped(p),
1814 			    ("freeing mapped page %p", p));
1815 			if (vm_page_remove(p))
1816 				vm_page_free(p);
1817 			continue;
1818 		}
1819 
1820 		if (!vm_page_all_valid(p)) {
1821 			KASSERT(!pmap_page_is_mapped(p),
1822 			    ("freeing mapped page %p", p));
1823 			if (vm_page_remove(p))
1824 				vm_page_free(p);
1825 			continue;
1826 		}
1827 
1828 		pp = vm_page_lookup(object, new_pindex);
1829 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1830 			vm_page_xunbusy(p);
1831 			/*
1832 			 * The page in the parent is busy and possibly not
1833 			 * (yet) valid.  Until its state is finalized by the
1834 			 * busy bit owner, we can't tell whether it shadows the
1835 			 * original page.
1836 			 */
1837 			next = vm_object_collapse_scan_wait(object, pp);
1838 			continue;
1839 		}
1840 
1841 		if (pp != NULL && vm_page_none_valid(pp)) {
1842 			/*
1843 			 * The page was invalid in the parent.  Likely placed
1844 			 * there by an incomplete fault.  Just remove and
1845 			 * ignore.  p can replace it.
1846 			 */
1847 			if (vm_page_remove(pp))
1848 				vm_page_free(pp);
1849 			pp = NULL;
1850 		}
1851 
1852 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1853 			NULL)) {
1854 			/*
1855 			 * The page already exists in the parent OR swap exists
1856 			 * for this location in the parent.  Leave the parent's
1857 			 * page alone.  Destroy the original page from the
1858 			 * backing object.
1859 			 */
1860 			vm_pager_freespace(backing_object, p->pindex, 1);
1861 			KASSERT(!pmap_page_is_mapped(p),
1862 			    ("freeing mapped page %p", p));
1863 			if (vm_page_remove(p))
1864 				vm_page_free(p);
1865 			if (pp != NULL)
1866 				vm_page_xunbusy(pp);
1867 			continue;
1868 		}
1869 
1870 		/*
1871 		 * Page does not exist in parent, rename the page from the
1872 		 * backing object to the main object.
1873 		 *
1874 		 * If the page was mapped to a process, it can remain mapped
1875 		 * through the rename.  vm_page_rename() will dirty the page.
1876 		 */
1877 		if (vm_page_rename(p, object, new_pindex)) {
1878 			vm_page_xunbusy(p);
1879 			next = vm_object_collapse_scan_wait(object, NULL);
1880 			continue;
1881 		}
1882 
1883 		/* Use the old pindex to free the right page. */
1884 		vm_pager_freespace(backing_object, new_pindex +
1885 		    backing_offset_index, 1);
1886 
1887 #if VM_NRESERVLEVEL > 0
1888 		/*
1889 		 * Rename the reservation.
1890 		 */
1891 		vm_reserv_rename(p, object, backing_object,
1892 		    backing_offset_index);
1893 #endif
1894 		vm_page_xunbusy(p);
1895 	}
1896 	return;
1897 }
1898 
1899 /*
1900  *	vm_object_collapse:
1901  *
1902  *	Collapse an object with the object backing it.
1903  *	Pages in the backing object are moved into the
1904  *	parent, and the backing object is deallocated.
1905  */
1906 void
1907 vm_object_collapse(vm_object_t object)
1908 {
1909 	vm_object_t backing_object, new_backing_object;
1910 
1911 	VM_OBJECT_ASSERT_WLOCKED(object);
1912 
1913 	while (TRUE) {
1914 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1915 		    ("collapsing invalid object"));
1916 
1917 		/*
1918 		 * Wait for the backing_object to finish any pending
1919 		 * collapse so that the caller sees the shortest possible
1920 		 * shadow chain.
1921 		 */
1922 		backing_object = vm_object_backing_collapse_wait(object);
1923 		if (backing_object == NULL)
1924 			return;
1925 
1926 		KASSERT(object->ref_count > 0 &&
1927 		    object->ref_count > atomic_load_int(&object->shadow_count),
1928 		    ("collapse with invalid ref %d or shadow %d count.",
1929 		    object->ref_count, atomic_load_int(&object->shadow_count)));
1930 		KASSERT((backing_object->flags &
1931 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1932 		    ("vm_object_collapse: Backing object already collapsing."));
1933 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1934 		    ("vm_object_collapse: object is already collapsing."));
1935 
1936 		/*
1937 		 * We know that we can either collapse the backing object if
1938 		 * the parent is the only reference to it, or (perhaps) have
1939 		 * the parent bypass the object if the parent happens to shadow
1940 		 * all the resident pages in the entire backing object.
1941 		 */
1942 		if (backing_object->ref_count == 1) {
1943 			KASSERT(atomic_load_int(&backing_object->shadow_count)
1944 			    == 1,
1945 			    ("vm_object_collapse: shadow_count: %d",
1946 			    atomic_load_int(&backing_object->shadow_count)));
1947 			vm_object_pip_add(object, 1);
1948 			vm_object_set_flag(object, OBJ_COLLAPSING);
1949 			vm_object_pip_add(backing_object, 1);
1950 			vm_object_set_flag(backing_object, OBJ_DEAD);
1951 
1952 			/*
1953 			 * If there is exactly one reference to the backing
1954 			 * object, we can collapse it into the parent.
1955 			 */
1956 			vm_object_collapse_scan(object);
1957 
1958 #if VM_NRESERVLEVEL > 0
1959 			/*
1960 			 * Break any reservations from backing_object.
1961 			 */
1962 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1963 				vm_reserv_break_all(backing_object);
1964 #endif
1965 
1966 			/*
1967 			 * Move the pager from backing_object to object.
1968 			 *
1969 			 * swap_pager_copy() can sleep, in which case the
1970 			 * backing_object's and object's locks are released and
1971 			 * reacquired.
1972 			 */
1973 			swap_pager_copy(backing_object, object,
1974 			    OFF_TO_IDX(object->backing_object_offset), TRUE);
1975 
1976 			/*
1977 			 * Object now shadows whatever backing_object did.
1978 			 */
1979 			vm_object_clear_flag(object, OBJ_COLLAPSING);
1980 			vm_object_backing_transfer(object, backing_object);
1981 			object->backing_object_offset +=
1982 			    backing_object->backing_object_offset;
1983 			VM_OBJECT_WUNLOCK(object);
1984 			vm_object_pip_wakeup(object);
1985 
1986 			/*
1987 			 * Discard backing_object.
1988 			 *
1989 			 * Since the backing object has no pages, no pager left,
1990 			 * and no object references within it, all that is
1991 			 * necessary is to dispose of it.
1992 			 */
1993 			KASSERT(backing_object->ref_count == 1, (
1994 "backing_object %p was somehow re-referenced during collapse!",
1995 			    backing_object));
1996 			vm_object_pip_wakeup(backing_object);
1997 			(void)refcount_release(&backing_object->ref_count);
1998 			vm_object_terminate(backing_object);
1999 			counter_u64_add(object_collapses, 1);
2000 			VM_OBJECT_WLOCK(object);
2001 		} else {
2002 			/*
2003 			 * If we do not entirely shadow the backing object,
2004 			 * there is nothing we can do so we give up.
2005 			 *
2006 			 * The object lock and backing_object lock must not
2007 			 * be dropped during this sequence.
2008 			 */
2009 			if (!vm_object_scan_all_shadowed(object)) {
2010 				VM_OBJECT_WUNLOCK(backing_object);
2011 				break;
2012 			}
2013 
2014 			/*
2015 			 * Make the parent shadow the next object in the
2016 			 * chain.  Deallocating backing_object will not remove
2017 			 * it, since its reference count is at least 2.
2018 			 */
2019 			vm_object_backing_remove_locked(object);
2020 			new_backing_object = backing_object->backing_object;
2021 			if (new_backing_object != NULL) {
2022 				vm_object_backing_insert_ref(object,
2023 				    new_backing_object);
2024 				object->backing_object_offset +=
2025 				    backing_object->backing_object_offset;
2026 			}
2027 
2028 			/*
2029 			 * Drop the reference count on backing_object. Since
2030 			 * its ref_count was at least 2, it will not vanish.
2031 			 */
2032 			(void)refcount_release(&backing_object->ref_count);
2033 			KASSERT(backing_object->ref_count >= 1, (
2034 "backing_object %p was somehow dereferenced during collapse!",
2035 			    backing_object));
2036 			VM_OBJECT_WUNLOCK(backing_object);
2037 			counter_u64_add(object_bypasses, 1);
2038 		}
2039 
2040 		/*
2041 		 * Try again with this object's new backing object.
2042 		 */
2043 	}
2044 }
2045 
2046 /*
2047  *	vm_object_page_remove:
2048  *
2049  *	For the given object, either frees or invalidates each of the
2050  *	specified pages.  In general, a page is freed.  However, if a page is
2051  *	wired for any reason other than the existence of a managed, wired
2052  *	mapping, then it may be invalidated but not removed from the object.
2053  *	Pages are specified by the given range ["start", "end") and the option
2054  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
2055  *	extends from "start" to the end of the object.  If the option
2056  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2057  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
2058  *	specified, then the pages within the specified range must have no
2059  *	mappings.  Otherwise, if this option is not specified, any mappings to
2060  *	the specified pages are removed before the pages are freed or
2061  *	invalidated.
2062  *
2063  *	In general, this operation should only be performed on objects that
2064  *	contain managed pages.  There are, however, two exceptions.  First, it
2065  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
2066  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2067  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
2068  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
2069  *
2070  *	The object must be locked.
2071  */
2072 void
2073 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2074     int options)
2075 {
2076 	vm_page_t p, next;
2077 
2078 	VM_OBJECT_ASSERT_WLOCKED(object);
2079 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2080 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2081 	    ("vm_object_page_remove: illegal options for object %p", object));
2082 	if (object->resident_page_count == 0)
2083 		return;
2084 	vm_object_pip_add(object, 1);
2085 again:
2086 	p = vm_page_find_least(object, start);
2087 
2088 	/*
2089 	 * Here, the variable "p" is either (1) the page with the least pindex
2090 	 * greater than or equal to the parameter "start" or (2) NULL.
2091 	 */
2092 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2093 		next = TAILQ_NEXT(p, listq);
2094 
2095 		/*
2096 		 * Skip invalid pages if asked to do so.  Try to avoid acquiring
2097 		 * the busy lock, as some consumers rely on this to avoid
2098 		 * deadlocks.
2099 		 *
2100 		 * A thread may concurrently transition the page from invalid to
2101 		 * valid using only the busy lock, so the result of this check
2102 		 * is immediately stale.  It is up to consumers to handle this,
2103 		 * for instance by ensuring that all invalid->valid transitions
2104 		 * happen with a mutex held, as may be possible for a
2105 		 * filesystem.
2106 		 */
2107 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2108 			continue;
2109 
2110 		/*
2111 		 * If the page is wired for any reason besides the existence
2112 		 * of managed, wired mappings, then it cannot be freed.  For
2113 		 * example, fictitious pages, which represent device memory,
2114 		 * are inherently wired and cannot be freed.  They can,
2115 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2116 		 * not specified.
2117 		 */
2118 		if (vm_page_tryxbusy(p) == 0) {
2119 			if (vm_page_busy_sleep(p, "vmopar", 0))
2120 				VM_OBJECT_WLOCK(object);
2121 			goto again;
2122 		}
2123 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2124 			vm_page_xunbusy(p);
2125 			continue;
2126 		}
2127 		if (vm_page_wired(p)) {
2128 wired:
2129 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2130 			    object->ref_count != 0)
2131 				pmap_remove_all(p);
2132 			if ((options & OBJPR_CLEANONLY) == 0) {
2133 				vm_page_invalid(p);
2134 				vm_page_undirty(p);
2135 			}
2136 			vm_page_xunbusy(p);
2137 			continue;
2138 		}
2139 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2140 		    ("vm_object_page_remove: page %p is fictitious", p));
2141 		if ((options & OBJPR_CLEANONLY) != 0 &&
2142 		    !vm_page_none_valid(p)) {
2143 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2144 			    object->ref_count != 0 &&
2145 			    !vm_page_try_remove_write(p))
2146 				goto wired;
2147 			if (p->dirty != 0) {
2148 				vm_page_xunbusy(p);
2149 				continue;
2150 			}
2151 		}
2152 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2153 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2154 			goto wired;
2155 		vm_page_free(p);
2156 	}
2157 	vm_object_pip_wakeup(object);
2158 
2159 	vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2160 	    start);
2161 }
2162 
2163 /*
2164  *	vm_object_page_noreuse:
2165  *
2166  *	For the given object, attempt to move the specified pages to
2167  *	the head of the inactive queue.  This bypasses regular LRU
2168  *	operation and allows the pages to be reused quickly under memory
2169  *	pressure.  If a page is wired for any reason, then it will not
2170  *	be queued.  Pages are specified by the range ["start", "end").
2171  *	As a special case, if "end" is zero, then the range extends from
2172  *	"start" to the end of the object.
2173  *
2174  *	This operation should only be performed on objects that
2175  *	contain non-fictitious, managed pages.
2176  *
2177  *	The object must be locked.
2178  */
2179 void
2180 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2181 {
2182 	vm_page_t p, next;
2183 
2184 	VM_OBJECT_ASSERT_LOCKED(object);
2185 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2186 	    ("vm_object_page_noreuse: illegal object %p", object));
2187 	if (object->resident_page_count == 0)
2188 		return;
2189 	p = vm_page_find_least(object, start);
2190 
2191 	/*
2192 	 * Here, the variable "p" is either (1) the page with the least pindex
2193 	 * greater than or equal to the parameter "start" or (2) NULL.
2194 	 */
2195 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2196 		next = TAILQ_NEXT(p, listq);
2197 		vm_page_deactivate_noreuse(p);
2198 	}
2199 }
2200 
2201 /*
2202  *	Populate the specified range of the object with valid pages.  Returns
2203  *	TRUE if the range is successfully populated and FALSE otherwise.
2204  *
2205  *	Note: This function should be optimized to pass a larger array of
2206  *	pages to vm_pager_get_pages() before it is applied to a non-
2207  *	OBJT_DEVICE object.
2208  *
2209  *	The object must be locked.
2210  */
2211 boolean_t
2212 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2213 {
2214 	vm_page_t m;
2215 	vm_pindex_t pindex;
2216 	int rv;
2217 
2218 	VM_OBJECT_ASSERT_WLOCKED(object);
2219 	for (pindex = start; pindex < end; pindex++) {
2220 		rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2221 		if (rv != VM_PAGER_OK)
2222 			break;
2223 
2224 		/*
2225 		 * Keep "m" busy because a subsequent iteration may unlock
2226 		 * the object.
2227 		 */
2228 	}
2229 	if (pindex > start) {
2230 		m = vm_page_lookup(object, start);
2231 		while (m != NULL && m->pindex < pindex) {
2232 			vm_page_xunbusy(m);
2233 			m = TAILQ_NEXT(m, listq);
2234 		}
2235 	}
2236 	return (pindex == end);
2237 }
2238 
2239 /*
2240  *	Routine:	vm_object_coalesce
2241  *	Function:	Coalesces two objects backing up adjoining
2242  *			regions of memory into a single object.
2243  *
2244  *	returns TRUE if objects were combined.
2245  *
2246  *	NOTE:	Only works at the moment if the second object is NULL -
2247  *		if it's not, which object do we lock first?
2248  *
2249  *	Parameters:
2250  *		prev_object	First object to coalesce
2251  *		prev_offset	Offset into prev_object
2252  *		prev_size	Size of reference to prev_object
2253  *		next_size	Size of reference to the second object
2254  *		reserved	Indicator that extension region has
2255  *				swap accounted for
2256  *
2257  *	Conditions:
2258  *	The object must *not* be locked.
2259  */
2260 boolean_t
2261 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2262     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2263 {
2264 	vm_pindex_t next_pindex;
2265 
2266 	if (prev_object == NULL)
2267 		return (TRUE);
2268 	if ((prev_object->flags & OBJ_ANON) == 0)
2269 		return (FALSE);
2270 
2271 	VM_OBJECT_WLOCK(prev_object);
2272 	/*
2273 	 * Try to collapse the object first.
2274 	 */
2275 	vm_object_collapse(prev_object);
2276 
2277 	/*
2278 	 * Can't coalesce if: . more than one reference . paged out . shadows
2279 	 * another object . has a copy elsewhere (any of which mean that the
2280 	 * pages not mapped to prev_entry may be in use anyway)
2281 	 */
2282 	if (prev_object->backing_object != NULL) {
2283 		VM_OBJECT_WUNLOCK(prev_object);
2284 		return (FALSE);
2285 	}
2286 
2287 	prev_size >>= PAGE_SHIFT;
2288 	next_size >>= PAGE_SHIFT;
2289 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2290 
2291 	if (prev_object->ref_count > 1 &&
2292 	    prev_object->size != next_pindex &&
2293 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2294 		VM_OBJECT_WUNLOCK(prev_object);
2295 		return (FALSE);
2296 	}
2297 
2298 	/*
2299 	 * Account for the charge.
2300 	 */
2301 	if (prev_object->cred != NULL) {
2302 		/*
2303 		 * If prev_object was charged, then this mapping,
2304 		 * although not charged now, may become writable
2305 		 * later. Non-NULL cred in the object would prevent
2306 		 * swap reservation during enabling of the write
2307 		 * access, so reserve swap now. Failed reservation
2308 		 * cause allocation of the separate object for the map
2309 		 * entry, and swap reservation for this entry is
2310 		 * managed in appropriate time.
2311 		 */
2312 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2313 		    prev_object->cred)) {
2314 			VM_OBJECT_WUNLOCK(prev_object);
2315 			return (FALSE);
2316 		}
2317 		prev_object->charge += ptoa(next_size);
2318 	}
2319 
2320 	/*
2321 	 * Remove any pages that may still be in the object from a previous
2322 	 * deallocation.
2323 	 */
2324 	if (next_pindex < prev_object->size) {
2325 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2326 		    next_size, 0);
2327 #if 0
2328 		if (prev_object->cred != NULL) {
2329 			KASSERT(prev_object->charge >=
2330 			    ptoa(prev_object->size - next_pindex),
2331 			    ("object %p overcharged 1 %jx %jx", prev_object,
2332 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2333 			prev_object->charge -= ptoa(prev_object->size -
2334 			    next_pindex);
2335 		}
2336 #endif
2337 	}
2338 
2339 	/*
2340 	 * Extend the object if necessary.
2341 	 */
2342 	if (next_pindex + next_size > prev_object->size)
2343 		prev_object->size = next_pindex + next_size;
2344 
2345 	VM_OBJECT_WUNLOCK(prev_object);
2346 	return (TRUE);
2347 }
2348 
2349 void
2350 vm_object_set_writeable_dirty_(vm_object_t object)
2351 {
2352 	atomic_add_int(&object->generation, 1);
2353 }
2354 
2355 bool
2356 vm_object_mightbedirty_(vm_object_t object)
2357 {
2358 	return (object->generation != object->cleangeneration);
2359 }
2360 
2361 /*
2362  *	vm_object_unwire:
2363  *
2364  *	For each page offset within the specified range of the given object,
2365  *	find the highest-level page in the shadow chain and unwire it.  A page
2366  *	must exist at every page offset, and the highest-level page must be
2367  *	wired.
2368  */
2369 void
2370 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2371     uint8_t queue)
2372 {
2373 	vm_object_t tobject, t1object;
2374 	vm_page_t m, tm;
2375 	vm_pindex_t end_pindex, pindex, tpindex;
2376 	int depth, locked_depth;
2377 
2378 	KASSERT((offset & PAGE_MASK) == 0,
2379 	    ("vm_object_unwire: offset is not page aligned"));
2380 	KASSERT((length & PAGE_MASK) == 0,
2381 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2382 	/* The wired count of a fictitious page never changes. */
2383 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2384 		return;
2385 	pindex = OFF_TO_IDX(offset);
2386 	end_pindex = pindex + atop(length);
2387 again:
2388 	locked_depth = 1;
2389 	VM_OBJECT_RLOCK(object);
2390 	m = vm_page_find_least(object, pindex);
2391 	while (pindex < end_pindex) {
2392 		if (m == NULL || pindex < m->pindex) {
2393 			/*
2394 			 * The first object in the shadow chain doesn't
2395 			 * contain a page at the current index.  Therefore,
2396 			 * the page must exist in a backing object.
2397 			 */
2398 			tobject = object;
2399 			tpindex = pindex;
2400 			depth = 0;
2401 			do {
2402 				tpindex +=
2403 				    OFF_TO_IDX(tobject->backing_object_offset);
2404 				tobject = tobject->backing_object;
2405 				KASSERT(tobject != NULL,
2406 				    ("vm_object_unwire: missing page"));
2407 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2408 					goto next_page;
2409 				depth++;
2410 				if (depth == locked_depth) {
2411 					locked_depth++;
2412 					VM_OBJECT_RLOCK(tobject);
2413 				}
2414 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2415 			    NULL);
2416 		} else {
2417 			tm = m;
2418 			m = TAILQ_NEXT(m, listq);
2419 		}
2420 		if (vm_page_trysbusy(tm) == 0) {
2421 			for (tobject = object; locked_depth >= 1;
2422 			    locked_depth--) {
2423 				t1object = tobject->backing_object;
2424 				if (tm->object != tobject)
2425 					VM_OBJECT_RUNLOCK(tobject);
2426 				tobject = t1object;
2427 			}
2428 			tobject = tm->object;
2429 			if (!vm_page_busy_sleep(tm, "unwbo",
2430 			    VM_ALLOC_IGN_SBUSY))
2431 				VM_OBJECT_RUNLOCK(tobject);
2432 			goto again;
2433 		}
2434 		vm_page_unwire(tm, queue);
2435 		vm_page_sunbusy(tm);
2436 next_page:
2437 		pindex++;
2438 	}
2439 	/* Release the accumulated object locks. */
2440 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2441 		t1object = tobject->backing_object;
2442 		VM_OBJECT_RUNLOCK(tobject);
2443 		tobject = t1object;
2444 	}
2445 }
2446 
2447 /*
2448  * Return the vnode for the given object, or NULL if none exists.
2449  * For tmpfs objects, the function may return NULL if there is
2450  * no vnode allocated at the time of the call.
2451  */
2452 struct vnode *
2453 vm_object_vnode(vm_object_t object)
2454 {
2455 	struct vnode *vp;
2456 
2457 	VM_OBJECT_ASSERT_LOCKED(object);
2458 	vm_pager_getvp(object, &vp, NULL);
2459 	return (vp);
2460 }
2461 
2462 /*
2463  * Busy the vm object.  This prevents new pages belonging to the object from
2464  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2465  * for checking page state before proceeding.
2466  */
2467 void
2468 vm_object_busy(vm_object_t obj)
2469 {
2470 
2471 	VM_OBJECT_ASSERT_LOCKED(obj);
2472 
2473 	blockcount_acquire(&obj->busy, 1);
2474 	/* The fence is required to order loads of page busy. */
2475 	atomic_thread_fence_acq_rel();
2476 }
2477 
2478 void
2479 vm_object_unbusy(vm_object_t obj)
2480 {
2481 
2482 	blockcount_release(&obj->busy, 1);
2483 }
2484 
2485 void
2486 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2487 {
2488 
2489 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2490 
2491 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2492 }
2493 
2494 /*
2495  * This function aims to determine if the object is mapped,
2496  * specifically, if it is referenced by a vm_map_entry.  Because
2497  * objects occasionally acquire transient references that do not
2498  * represent a mapping, the method used here is inexact.  However, it
2499  * has very low overhead and is good enough for the advisory
2500  * vm.vmtotal sysctl.
2501  */
2502 bool
2503 vm_object_is_active(vm_object_t obj)
2504 {
2505 
2506 	return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2507 }
2508 
2509 static int
2510 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2511 {
2512 	struct kinfo_vmobject *kvo;
2513 	char *fullpath, *freepath;
2514 	struct vnode *vp;
2515 	struct vattr va;
2516 	vm_object_t obj;
2517 	vm_page_t m;
2518 	u_long sp;
2519 	int count, error;
2520 
2521 	if (req->oldptr == NULL) {
2522 		/*
2523 		 * If an old buffer has not been provided, generate an
2524 		 * estimate of the space needed for a subsequent call.
2525 		 */
2526 		mtx_lock(&vm_object_list_mtx);
2527 		count = 0;
2528 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2529 			if (obj->type == OBJT_DEAD)
2530 				continue;
2531 			count++;
2532 		}
2533 		mtx_unlock(&vm_object_list_mtx);
2534 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2535 		    count * 11 / 10));
2536 	}
2537 
2538 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO);
2539 	error = 0;
2540 
2541 	/*
2542 	 * VM objects are type stable and are never removed from the
2543 	 * list once added.  This allows us to safely read obj->object_list
2544 	 * after reacquiring the VM object lock.
2545 	 */
2546 	mtx_lock(&vm_object_list_mtx);
2547 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2548 		if (obj->type == OBJT_DEAD ||
2549 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2550 			continue;
2551 		VM_OBJECT_RLOCK(obj);
2552 		if (obj->type == OBJT_DEAD ||
2553 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2554 			VM_OBJECT_RUNLOCK(obj);
2555 			continue;
2556 		}
2557 		mtx_unlock(&vm_object_list_mtx);
2558 		kvo->kvo_size = ptoa(obj->size);
2559 		kvo->kvo_resident = obj->resident_page_count;
2560 		kvo->kvo_ref_count = obj->ref_count;
2561 		kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2562 		kvo->kvo_memattr = obj->memattr;
2563 		kvo->kvo_active = 0;
2564 		kvo->kvo_inactive = 0;
2565 		if (!swap_only) {
2566 			TAILQ_FOREACH(m, &obj->memq, listq) {
2567 				/*
2568 				 * A page may belong to the object but be
2569 				 * dequeued and set to PQ_NONE while the
2570 				 * object lock is not held.  This makes the
2571 				 * reads of m->queue below racy, and we do not
2572 				 * count pages set to PQ_NONE.  However, this
2573 				 * sysctl is only meant to give an
2574 				 * approximation of the system anyway.
2575 				 */
2576 				if (m->a.queue == PQ_ACTIVE)
2577 					kvo->kvo_active++;
2578 				else if (m->a.queue == PQ_INACTIVE)
2579 					kvo->kvo_inactive++;
2580 			}
2581 		}
2582 
2583 		kvo->kvo_vn_fileid = 0;
2584 		kvo->kvo_vn_fsid = 0;
2585 		kvo->kvo_vn_fsid_freebsd11 = 0;
2586 		freepath = NULL;
2587 		fullpath = "";
2588 		vp = NULL;
2589 		kvo->kvo_type = vm_object_kvme_type(obj, swap_only ? NULL : &vp);
2590 		if (vp != NULL) {
2591 			vref(vp);
2592 		} else if ((obj->flags & OBJ_ANON) != 0) {
2593 			MPASS(kvo->kvo_type == KVME_TYPE_SWAP);
2594 			kvo->kvo_me = (uintptr_t)obj;
2595 			/* tmpfs objs are reported as vnodes */
2596 			kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2597 			sp = swap_pager_swapped_pages(obj);
2598 			kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2599 		}
2600 		VM_OBJECT_RUNLOCK(obj);
2601 		if (vp != NULL) {
2602 			vn_fullpath(vp, &fullpath, &freepath);
2603 			vn_lock(vp, LK_SHARED | LK_RETRY);
2604 			if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2605 				kvo->kvo_vn_fileid = va.va_fileid;
2606 				kvo->kvo_vn_fsid = va.va_fsid;
2607 				kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2608 								/* truncate */
2609 			}
2610 			vput(vp);
2611 		}
2612 
2613 		strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2614 		if (freepath != NULL)
2615 			free(freepath, M_TEMP);
2616 
2617 		/* Pack record size down */
2618 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2619 		    + strlen(kvo->kvo_path) + 1;
2620 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2621 		    sizeof(uint64_t));
2622 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2623 		maybe_yield();
2624 		mtx_lock(&vm_object_list_mtx);
2625 		if (error)
2626 			break;
2627 	}
2628 	mtx_unlock(&vm_object_list_mtx);
2629 	free(kvo, M_TEMP);
2630 	return (error);
2631 }
2632 
2633 static int
2634 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2635 {
2636 	return (vm_object_list_handler(req, false));
2637 }
2638 
2639 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2640     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2641     "List of VM objects");
2642 
2643 static int
2644 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2645 {
2646 	return (vm_object_list_handler(req, true));
2647 }
2648 
2649 /*
2650  * This sysctl returns list of the anonymous or swap objects. Intent
2651  * is to provide stripped optimized list useful to analyze swap use.
2652  * Since technically non-swap (default) objects participate in the
2653  * shadow chains, and are converted to swap type as needed by swap
2654  * pager, we must report them.
2655  */
2656 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2657     CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2658     sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2659     "List of swap VM objects");
2660 
2661 #include "opt_ddb.h"
2662 #ifdef DDB
2663 #include <sys/kernel.h>
2664 
2665 #include <sys/cons.h>
2666 
2667 #include <ddb/ddb.h>
2668 
2669 static int
2670 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2671 {
2672 	vm_map_t tmpm;
2673 	vm_map_entry_t tmpe;
2674 	vm_object_t obj;
2675 
2676 	if (map == 0)
2677 		return 0;
2678 
2679 	if (entry == 0) {
2680 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2681 			if (_vm_object_in_map(map, object, tmpe)) {
2682 				return 1;
2683 			}
2684 		}
2685 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2686 		tmpm = entry->object.sub_map;
2687 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2688 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2689 				return 1;
2690 			}
2691 		}
2692 	} else if ((obj = entry->object.vm_object) != NULL) {
2693 		for (; obj; obj = obj->backing_object)
2694 			if (obj == object) {
2695 				return 1;
2696 			}
2697 	}
2698 	return 0;
2699 }
2700 
2701 static int
2702 vm_object_in_map(vm_object_t object)
2703 {
2704 	struct proc *p;
2705 
2706 	/* sx_slock(&allproc_lock); */
2707 	FOREACH_PROC_IN_SYSTEM(p) {
2708 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2709 			continue;
2710 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2711 			/* sx_sunlock(&allproc_lock); */
2712 			return 1;
2713 		}
2714 	}
2715 	/* sx_sunlock(&allproc_lock); */
2716 	if (_vm_object_in_map(kernel_map, object, 0))
2717 		return 1;
2718 	return 0;
2719 }
2720 
2721 DB_SHOW_COMMAND_FLAGS(vmochk, vm_object_check, DB_CMD_MEMSAFE)
2722 {
2723 	vm_object_t object;
2724 
2725 	/*
2726 	 * make sure that internal objs are in a map somewhere
2727 	 * and none have zero ref counts.
2728 	 */
2729 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2730 		if ((object->flags & OBJ_ANON) != 0) {
2731 			if (object->ref_count == 0) {
2732 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2733 					(long)object->size);
2734 			}
2735 			if (!vm_object_in_map(object)) {
2736 				db_printf(
2737 			"vmochk: internal obj is not in a map: "
2738 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2739 				    object->ref_count, (u_long)object->size,
2740 				    (u_long)object->size,
2741 				    (void *)object->backing_object);
2742 			}
2743 		}
2744 		if (db_pager_quit)
2745 			return;
2746 	}
2747 }
2748 
2749 /*
2750  *	vm_object_print:	[ debug ]
2751  */
2752 DB_SHOW_COMMAND(object, vm_object_print_static)
2753 {
2754 	/* XXX convert args. */
2755 	vm_object_t object = (vm_object_t)addr;
2756 	boolean_t full = have_addr;
2757 
2758 	vm_page_t p;
2759 
2760 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2761 #define	count	was_count
2762 
2763 	int count;
2764 
2765 	if (object == NULL)
2766 		return;
2767 
2768 	db_iprintf(
2769 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2770 	    object, (int)object->type, (uintmax_t)object->size,
2771 	    object->resident_page_count, object->ref_count, object->flags,
2772 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2773 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2774 	    atomic_load_int(&object->shadow_count),
2775 	    object->backing_object ? object->backing_object->ref_count : 0,
2776 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2777 
2778 	if (!full)
2779 		return;
2780 
2781 	db_indent += 2;
2782 	count = 0;
2783 	TAILQ_FOREACH(p, &object->memq, listq) {
2784 		if (count == 0)
2785 			db_iprintf("memory:=");
2786 		else if (count == 6) {
2787 			db_printf("\n");
2788 			db_iprintf(" ...");
2789 			count = 0;
2790 		} else
2791 			db_printf(",");
2792 		count++;
2793 
2794 		db_printf("(off=0x%jx,page=0x%jx)",
2795 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2796 
2797 		if (db_pager_quit)
2798 			break;
2799 	}
2800 	if (count != 0)
2801 		db_printf("\n");
2802 	db_indent -= 2;
2803 }
2804 
2805 /* XXX. */
2806 #undef count
2807 
2808 /* XXX need this non-static entry for calling from vm_map_print. */
2809 void
2810 vm_object_print(
2811         /* db_expr_t */ long addr,
2812 	boolean_t have_addr,
2813 	/* db_expr_t */ long count,
2814 	char *modif)
2815 {
2816 	vm_object_print_static(addr, have_addr, count, modif);
2817 }
2818 
2819 DB_SHOW_COMMAND_FLAGS(vmopag, vm_object_print_pages, DB_CMD_MEMSAFE)
2820 {
2821 	vm_object_t object;
2822 	vm_pindex_t fidx;
2823 	vm_paddr_t pa;
2824 	vm_page_t m, prev_m;
2825 	int rcount;
2826 
2827 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2828 		db_printf("new object: %p\n", (void *)object);
2829 		if (db_pager_quit)
2830 			return;
2831 
2832 		rcount = 0;
2833 		fidx = 0;
2834 		pa = -1;
2835 		TAILQ_FOREACH(m, &object->memq, listq) {
2836 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2837 			    prev_m->pindex + 1 != m->pindex) {
2838 				if (rcount) {
2839 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2840 						(long)fidx, rcount, (long)pa);
2841 					if (db_pager_quit)
2842 						return;
2843 					rcount = 0;
2844 				}
2845 			}
2846 			if (rcount &&
2847 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2848 				++rcount;
2849 				continue;
2850 			}
2851 			if (rcount) {
2852 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2853 					(long)fidx, rcount, (long)pa);
2854 				if (db_pager_quit)
2855 					return;
2856 			}
2857 			fidx = m->pindex;
2858 			pa = VM_PAGE_TO_PHYS(m);
2859 			rcount = 1;
2860 		}
2861 		if (rcount) {
2862 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2863 				(long)fidx, rcount, (long)pa);
2864 			if (db_pager_quit)
2865 				return;
2866 		}
2867 	}
2868 }
2869 #endif /* DDB */
2870