xref: /freebsd/sys/vm/vm_object.c (revision 2d4e511ca269f1908d27f4e5779c53475527391d)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	from: @(#)vm_object.c	8.5 (Berkeley) 3/22/94
35  *
36  *
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *	Virtual memory object module.
65  */
66 
67 #include <sys/cdefs.h>
68 __FBSDID("$FreeBSD$");
69 
70 #include "opt_vm.h"
71 
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/blockcount.h>
75 #include <sys/cpuset.h>
76 #include <sys/lock.h>
77 #include <sys/mman.h>
78 #include <sys/mount.h>
79 #include <sys/kernel.h>
80 #include <sys/pctrie.h>
81 #include <sys/sysctl.h>
82 #include <sys/mutex.h>
83 #include <sys/proc.h>		/* for curproc, pageproc */
84 #include <sys/refcount.h>
85 #include <sys/socket.h>
86 #include <sys/resourcevar.h>
87 #include <sys/refcount.h>
88 #include <sys/rwlock.h>
89 #include <sys/user.h>
90 #include <sys/vnode.h>
91 #include <sys/vmmeter.h>
92 #include <sys/sx.h>
93 
94 #include <vm/vm.h>
95 #include <vm/vm_param.h>
96 #include <vm/pmap.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_pageout.h>
101 #include <vm/vm_pager.h>
102 #include <vm/vm_phys.h>
103 #include <vm/vm_pagequeue.h>
104 #include <vm/swap_pager.h>
105 #include <vm/vm_kern.h>
106 #include <vm/vm_extern.h>
107 #include <vm/vm_radix.h>
108 #include <vm/vm_reserv.h>
109 #include <vm/uma.h>
110 
111 static int old_msync;
112 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
113     "Use old (insecure) msync behavior");
114 
115 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
116 		    int pagerflags, int flags, boolean_t *allclean,
117 		    boolean_t *eio);
118 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
119 		    boolean_t *allclean);
120 static void	vm_object_backing_remove(vm_object_t object);
121 
122 /*
123  *	Virtual memory objects maintain the actual data
124  *	associated with allocated virtual memory.  A given
125  *	page of memory exists within exactly one object.
126  *
127  *	An object is only deallocated when all "references"
128  *	are given up.  Only one "reference" to a given
129  *	region of an object should be writeable.
130  *
131  *	Associated with each object is a list of all resident
132  *	memory pages belonging to that object; this list is
133  *	maintained by the "vm_page" module, and locked by the object's
134  *	lock.
135  *
136  *	Each object also records a "pager" routine which is
137  *	used to retrieve (and store) pages to the proper backing
138  *	storage.  In addition, objects may be backed by other
139  *	objects from which they were virtual-copied.
140  *
141  *	The only items within the object structure which are
142  *	modified after time of creation are:
143  *		reference count		locked by object's lock
144  *		pager routine		locked by object's lock
145  *
146  */
147 
148 struct object_q vm_object_list;
149 struct mtx vm_object_list_mtx;	/* lock for object list and count */
150 
151 struct vm_object kernel_object_store;
152 
153 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
154     "VM object stats");
155 
156 static counter_u64_t object_collapses = EARLY_COUNTER;
157 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
158     &object_collapses,
159     "VM object collapses");
160 
161 static counter_u64_t object_bypasses = EARLY_COUNTER;
162 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
163     &object_bypasses,
164     "VM object bypasses");
165 
166 static counter_u64_t object_collapse_waits = EARLY_COUNTER;
167 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
168     &object_collapse_waits,
169     "Number of sleeps for collapse");
170 
171 static void
172 counter_startup(void)
173 {
174 
175 	object_collapses = counter_u64_alloc(M_WAITOK);
176 	object_bypasses = counter_u64_alloc(M_WAITOK);
177 	object_collapse_waits = counter_u64_alloc(M_WAITOK);
178 }
179 SYSINIT(object_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL);
180 
181 static uma_zone_t obj_zone;
182 
183 static int vm_object_zinit(void *mem, int size, int flags);
184 
185 #ifdef INVARIANTS
186 static void vm_object_zdtor(void *mem, int size, void *arg);
187 
188 static void
189 vm_object_zdtor(void *mem, int size, void *arg)
190 {
191 	vm_object_t object;
192 
193 	object = (vm_object_t)mem;
194 	KASSERT(object->ref_count == 0,
195 	    ("object %p ref_count = %d", object, object->ref_count));
196 	KASSERT(TAILQ_EMPTY(&object->memq),
197 	    ("object %p has resident pages in its memq", object));
198 	KASSERT(vm_radix_is_empty(&object->rtree),
199 	    ("object %p has resident pages in its trie", object));
200 #if VM_NRESERVLEVEL > 0
201 	KASSERT(LIST_EMPTY(&object->rvq),
202 	    ("object %p has reservations",
203 	    object));
204 #endif
205 	KASSERT(blockcount_read(&object->paging_in_progress) == 0,
206 	    ("object %p paging_in_progress = %d",
207 	    object, blockcount_read(&object->paging_in_progress)));
208 	KASSERT(!vm_object_busied(object),
209 	    ("object %p busy = %d", object, blockcount_read(&object->busy)));
210 	KASSERT(object->resident_page_count == 0,
211 	    ("object %p resident_page_count = %d",
212 	    object, object->resident_page_count));
213 	KASSERT(object->shadow_count == 0,
214 	    ("object %p shadow_count = %d",
215 	    object, object->shadow_count));
216 	KASSERT(object->type == OBJT_DEAD,
217 	    ("object %p has non-dead type %d",
218 	    object, object->type));
219 }
220 #endif
221 
222 static int
223 vm_object_zinit(void *mem, int size, int flags)
224 {
225 	vm_object_t object;
226 
227 	object = (vm_object_t)mem;
228 	rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
229 
230 	/* These are true for any object that has been freed */
231 	object->type = OBJT_DEAD;
232 	vm_radix_init(&object->rtree);
233 	refcount_init(&object->ref_count, 0);
234 	blockcount_init(&object->paging_in_progress);
235 	blockcount_init(&object->busy);
236 	object->resident_page_count = 0;
237 	object->shadow_count = 0;
238 	object->flags = OBJ_DEAD;
239 
240 	mtx_lock(&vm_object_list_mtx);
241 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
242 	mtx_unlock(&vm_object_list_mtx);
243 	return (0);
244 }
245 
246 static void
247 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
248     vm_object_t object, void *handle)
249 {
250 
251 	TAILQ_INIT(&object->memq);
252 	LIST_INIT(&object->shadow_head);
253 
254 	object->type = type;
255 	if (type == OBJT_SWAP)
256 		pctrie_init(&object->un_pager.swp.swp_blks);
257 
258 	/*
259 	 * Ensure that swap_pager_swapoff() iteration over object_list
260 	 * sees up to date type and pctrie head if it observed
261 	 * non-dead object.
262 	 */
263 	atomic_thread_fence_rel();
264 
265 	object->pg_color = 0;
266 	object->flags = flags;
267 	object->size = size;
268 	object->domain.dr_policy = NULL;
269 	object->generation = 1;
270 	object->cleangeneration = 1;
271 	refcount_init(&object->ref_count, 1);
272 	object->memattr = VM_MEMATTR_DEFAULT;
273 	object->cred = NULL;
274 	object->charge = 0;
275 	object->handle = handle;
276 	object->backing_object = NULL;
277 	object->backing_object_offset = (vm_ooffset_t) 0;
278 #if VM_NRESERVLEVEL > 0
279 	LIST_INIT(&object->rvq);
280 #endif
281 	umtx_shm_object_init(object);
282 }
283 
284 /*
285  *	vm_object_init:
286  *
287  *	Initialize the VM objects module.
288  */
289 void
290 vm_object_init(void)
291 {
292 	TAILQ_INIT(&vm_object_list);
293 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
294 
295 	rw_init(&kernel_object->lock, "kernel vm object");
296 	_vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
297 	    VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
298 #if VM_NRESERVLEVEL > 0
299 	kernel_object->flags |= OBJ_COLORED;
300 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
301 #endif
302 
303 	/*
304 	 * The lock portion of struct vm_object must be type stable due
305 	 * to vm_pageout_fallback_object_lock locking a vm object
306 	 * without holding any references to it.
307 	 */
308 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
309 #ifdef INVARIANTS
310 	    vm_object_zdtor,
311 #else
312 	    NULL,
313 #endif
314 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
315 
316 	vm_radix_zinit();
317 }
318 
319 void
320 vm_object_clear_flag(vm_object_t object, u_short bits)
321 {
322 
323 	VM_OBJECT_ASSERT_WLOCKED(object);
324 	object->flags &= ~bits;
325 }
326 
327 /*
328  *	Sets the default memory attribute for the specified object.  Pages
329  *	that are allocated to this object are by default assigned this memory
330  *	attribute.
331  *
332  *	Presently, this function must be called before any pages are allocated
333  *	to the object.  In the future, this requirement may be relaxed for
334  *	"default" and "swap" objects.
335  */
336 int
337 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
338 {
339 
340 	VM_OBJECT_ASSERT_WLOCKED(object);
341 	switch (object->type) {
342 	case OBJT_DEFAULT:
343 	case OBJT_DEVICE:
344 	case OBJT_MGTDEVICE:
345 	case OBJT_PHYS:
346 	case OBJT_SG:
347 	case OBJT_SWAP:
348 	case OBJT_VNODE:
349 		if (!TAILQ_EMPTY(&object->memq))
350 			return (KERN_FAILURE);
351 		break;
352 	case OBJT_DEAD:
353 		return (KERN_INVALID_ARGUMENT);
354 	default:
355 		panic("vm_object_set_memattr: object %p is of undefined type",
356 		    object);
357 	}
358 	object->memattr = memattr;
359 	return (KERN_SUCCESS);
360 }
361 
362 void
363 vm_object_pip_add(vm_object_t object, short i)
364 {
365 
366 	if (i > 0)
367 		blockcount_acquire(&object->paging_in_progress, i);
368 }
369 
370 void
371 vm_object_pip_wakeup(vm_object_t object)
372 {
373 
374 	vm_object_pip_wakeupn(object, 1);
375 }
376 
377 void
378 vm_object_pip_wakeupn(vm_object_t object, short i)
379 {
380 
381 	if (i > 0)
382 		blockcount_release(&object->paging_in_progress, i);
383 }
384 
385 /*
386  * Atomically drop the object lock and wait for pip to drain.  This protects
387  * from sleep/wakeup races due to identity changes.  The lock is not re-acquired
388  * on return.
389  */
390 static void
391 vm_object_pip_sleep(vm_object_t object, const char *waitid)
392 {
393 
394 	(void)blockcount_sleep(&object->paging_in_progress, &object->lock,
395 	    waitid, PVM | PDROP);
396 }
397 
398 void
399 vm_object_pip_wait(vm_object_t object, const char *waitid)
400 {
401 
402 	VM_OBJECT_ASSERT_WLOCKED(object);
403 
404 	blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
405 	    PVM);
406 }
407 
408 void
409 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
410 {
411 
412 	VM_OBJECT_ASSERT_UNLOCKED(object);
413 
414 	blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
415 }
416 
417 /*
418  *	vm_object_allocate:
419  *
420  *	Returns a new object with the given size.
421  */
422 vm_object_t
423 vm_object_allocate(objtype_t type, vm_pindex_t size)
424 {
425 	vm_object_t object;
426 	u_short flags;
427 
428 	switch (type) {
429 	case OBJT_DEAD:
430 		panic("vm_object_allocate: can't create OBJT_DEAD");
431 	case OBJT_DEFAULT:
432 	case OBJT_SWAP:
433 		flags = OBJ_COLORED;
434 		break;
435 	case OBJT_DEVICE:
436 	case OBJT_SG:
437 		flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
438 		break;
439 	case OBJT_MGTDEVICE:
440 		flags = OBJ_FICTITIOUS;
441 		break;
442 	case OBJT_PHYS:
443 		flags = OBJ_UNMANAGED;
444 		break;
445 	case OBJT_VNODE:
446 		flags = 0;
447 		break;
448 	default:
449 		panic("vm_object_allocate: type %d is undefined", type);
450 	}
451 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
452 	_vm_object_allocate(type, size, flags, object, NULL);
453 
454 	return (object);
455 }
456 
457 /*
458  *	vm_object_allocate_anon:
459  *
460  *	Returns a new default object of the given size and marked as
461  *	anonymous memory for special split/collapse handling.  Color
462  *	to be initialized by the caller.
463  */
464 vm_object_t
465 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
466     struct ucred *cred, vm_size_t charge)
467 {
468 	vm_object_t handle, object;
469 
470 	if (backing_object == NULL)
471 		handle = NULL;
472 	else if ((backing_object->flags & OBJ_ANON) != 0)
473 		handle = backing_object->handle;
474 	else
475 		handle = backing_object;
476 	object = uma_zalloc(obj_zone, M_WAITOK);
477 	_vm_object_allocate(OBJT_DEFAULT, size, OBJ_ANON | OBJ_ONEMAPPING,
478 	    object, handle);
479 	object->cred = cred;
480 	object->charge = cred != NULL ? charge : 0;
481 	return (object);
482 }
483 
484 static void
485 vm_object_reference_vnode(vm_object_t object)
486 {
487 	u_int old;
488 
489 	/*
490 	 * vnode objects need the lock for the first reference
491 	 * to serialize with vnode_object_deallocate().
492 	 */
493 	if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
494 		VM_OBJECT_RLOCK(object);
495 		old = refcount_acquire(&object->ref_count);
496 		if (object->type == OBJT_VNODE && old == 0)
497 			vref(object->handle);
498 		VM_OBJECT_RUNLOCK(object);
499 	}
500 }
501 
502 /*
503  *	vm_object_reference:
504  *
505  *	Acquires a reference to the given object.
506  */
507 void
508 vm_object_reference(vm_object_t object)
509 {
510 
511 	if (object == NULL)
512 		return;
513 
514 	if (object->type == OBJT_VNODE)
515 		vm_object_reference_vnode(object);
516 	else
517 		refcount_acquire(&object->ref_count);
518 	KASSERT((object->flags & OBJ_DEAD) == 0,
519 	    ("vm_object_reference: Referenced dead object."));
520 }
521 
522 /*
523  *	vm_object_reference_locked:
524  *
525  *	Gets another reference to the given object.
526  *
527  *	The object must be locked.
528  */
529 void
530 vm_object_reference_locked(vm_object_t object)
531 {
532 	u_int old;
533 
534 	VM_OBJECT_ASSERT_LOCKED(object);
535 	old = refcount_acquire(&object->ref_count);
536 	if (object->type == OBJT_VNODE && old == 0)
537 		vref(object->handle);
538 	KASSERT((object->flags & OBJ_DEAD) == 0,
539 	    ("vm_object_reference: Referenced dead object."));
540 }
541 
542 /*
543  * Handle deallocating an object of type OBJT_VNODE.
544  */
545 static void
546 vm_object_deallocate_vnode(vm_object_t object)
547 {
548 	struct vnode *vp = (struct vnode *) object->handle;
549 	bool last;
550 
551 	KASSERT(object->type == OBJT_VNODE,
552 	    ("vm_object_deallocate_vnode: not a vnode object"));
553 	KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
554 
555 	/* Object lock to protect handle lookup. */
556 	last = refcount_release(&object->ref_count);
557 	VM_OBJECT_RUNLOCK(object);
558 
559 	if (!last)
560 		return;
561 
562 	if (!umtx_shm_vnobj_persistent)
563 		umtx_shm_object_terminated(object);
564 
565 	/* vrele may need the vnode lock. */
566 	vrele(vp);
567 }
568 
569 
570 /*
571  * We dropped a reference on an object and discovered that it had a
572  * single remaining shadow.  This is a sibling of the reference we
573  * dropped.  Attempt to collapse the sibling and backing object.
574  */
575 static vm_object_t
576 vm_object_deallocate_anon(vm_object_t backing_object)
577 {
578 	vm_object_t object;
579 
580 	/* Fetch the final shadow.  */
581 	object = LIST_FIRST(&backing_object->shadow_head);
582 	KASSERT(object != NULL && backing_object->shadow_count == 1,
583 	    ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
584 	    backing_object->ref_count, backing_object->shadow_count));
585 	KASSERT((object->flags & (OBJ_TMPFS_NODE | OBJ_ANON)) == OBJ_ANON,
586 	    ("invalid shadow object %p", object));
587 
588 	if (!VM_OBJECT_TRYWLOCK(object)) {
589 		/*
590 		 * Prevent object from disappearing since we do not have a
591 		 * reference.
592 		 */
593 		vm_object_pip_add(object, 1);
594 		VM_OBJECT_WUNLOCK(backing_object);
595 		VM_OBJECT_WLOCK(object);
596 		vm_object_pip_wakeup(object);
597 	} else
598 		VM_OBJECT_WUNLOCK(backing_object);
599 
600 	/*
601 	 * Check for a collapse/terminate race with the last reference holder.
602 	 */
603 	if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
604 	    !refcount_acquire_if_not_zero(&object->ref_count)) {
605 		VM_OBJECT_WUNLOCK(object);
606 		return (NULL);
607 	}
608 	backing_object = object->backing_object;
609 	if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
610 		vm_object_collapse(object);
611 	VM_OBJECT_WUNLOCK(object);
612 
613 	return (object);
614 }
615 
616 /*
617  *	vm_object_deallocate:
618  *
619  *	Release a reference to the specified object,
620  *	gained either through a vm_object_allocate
621  *	or a vm_object_reference call.  When all references
622  *	are gone, storage associated with this object
623  *	may be relinquished.
624  *
625  *	No object may be locked.
626  */
627 void
628 vm_object_deallocate(vm_object_t object)
629 {
630 	vm_object_t temp;
631 	bool released;
632 
633 	while (object != NULL) {
634 		/*
635 		 * If the reference count goes to 0 we start calling
636 		 * vm_object_terminate() on the object chain.  A ref count
637 		 * of 1 may be a special case depending on the shadow count
638 		 * being 0 or 1.  These cases require a write lock on the
639 		 * object.
640 		 */
641 		if ((object->flags & OBJ_ANON) == 0)
642 			released = refcount_release_if_gt(&object->ref_count, 1);
643 		else
644 			released = refcount_release_if_gt(&object->ref_count, 2);
645 		if (released)
646 			return;
647 
648 		if (object->type == OBJT_VNODE) {
649 			VM_OBJECT_RLOCK(object);
650 			if (object->type == OBJT_VNODE) {
651 				vm_object_deallocate_vnode(object);
652 				return;
653 			}
654 			VM_OBJECT_RUNLOCK(object);
655 		}
656 
657 		VM_OBJECT_WLOCK(object);
658 		KASSERT(object->ref_count > 0,
659 		    ("vm_object_deallocate: object deallocated too many times: %d",
660 		    object->type));
661 
662 		/*
663 		 * If this is not the final reference to an anonymous
664 		 * object we may need to collapse the shadow chain.
665 		 */
666 		if (!refcount_release(&object->ref_count)) {
667 			if (object->ref_count > 1 ||
668 			    object->shadow_count == 0) {
669 				if ((object->flags & OBJ_ANON) != 0 &&
670 				    object->ref_count == 1)
671 					vm_object_set_flag(object,
672 					    OBJ_ONEMAPPING);
673 				VM_OBJECT_WUNLOCK(object);
674 				return;
675 			}
676 
677 			/* Handle collapsing last ref on anonymous objects. */
678 			object = vm_object_deallocate_anon(object);
679 			continue;
680 		}
681 
682 		/*
683 		 * Handle the final reference to an object.  We restart
684 		 * the loop with the backing object to avoid recursion.
685 		 */
686 		umtx_shm_object_terminated(object);
687 		temp = object->backing_object;
688 		if (temp != NULL) {
689 			KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
690 			    ("shadowed tmpfs v_object 2 %p", object));
691 			vm_object_backing_remove(object);
692 		}
693 
694 		KASSERT((object->flags & OBJ_DEAD) == 0,
695 		    ("vm_object_deallocate: Terminating dead object."));
696 		vm_object_set_flag(object, OBJ_DEAD);
697 		vm_object_terminate(object);
698 		object = temp;
699 	}
700 }
701 
702 /*
703  *	vm_object_destroy removes the object from the global object list
704  *      and frees the space for the object.
705  */
706 void
707 vm_object_destroy(vm_object_t object)
708 {
709 
710 	/*
711 	 * Release the allocation charge.
712 	 */
713 	if (object->cred != NULL) {
714 		swap_release_by_cred(object->charge, object->cred);
715 		object->charge = 0;
716 		crfree(object->cred);
717 		object->cred = NULL;
718 	}
719 
720 	/*
721 	 * Free the space for the object.
722 	 */
723 	uma_zfree(obj_zone, object);
724 }
725 
726 static void
727 vm_object_backing_remove_locked(vm_object_t object)
728 {
729 	vm_object_t backing_object;
730 
731 	backing_object = object->backing_object;
732 	VM_OBJECT_ASSERT_WLOCKED(object);
733 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
734 
735 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
736 	    ("vm_object_backing_remove: Removing collapsing object."));
737 
738 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
739 		LIST_REMOVE(object, shadow_list);
740 		backing_object->shadow_count--;
741 		object->flags &= ~OBJ_SHADOWLIST;
742 	}
743 	object->backing_object = NULL;
744 }
745 
746 static void
747 vm_object_backing_remove(vm_object_t object)
748 {
749 	vm_object_t backing_object;
750 
751 	VM_OBJECT_ASSERT_WLOCKED(object);
752 
753 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
754 		backing_object = object->backing_object;
755 		VM_OBJECT_WLOCK(backing_object);
756 		vm_object_backing_remove_locked(object);
757 		VM_OBJECT_WUNLOCK(backing_object);
758 	} else
759 		object->backing_object = NULL;
760 }
761 
762 static void
763 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
764 {
765 
766 	VM_OBJECT_ASSERT_WLOCKED(object);
767 
768 	if ((backing_object->flags & OBJ_ANON) != 0) {
769 		VM_OBJECT_ASSERT_WLOCKED(backing_object);
770 		LIST_INSERT_HEAD(&backing_object->shadow_head, object,
771 		    shadow_list);
772 		backing_object->shadow_count++;
773 		object->flags |= OBJ_SHADOWLIST;
774 	}
775 	object->backing_object = backing_object;
776 }
777 
778 static void
779 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
780 {
781 
782 	VM_OBJECT_ASSERT_WLOCKED(object);
783 
784 	if ((backing_object->flags & OBJ_ANON) != 0) {
785 		VM_OBJECT_WLOCK(backing_object);
786 		vm_object_backing_insert_locked(object, backing_object);
787 		VM_OBJECT_WUNLOCK(backing_object);
788 	} else
789 		object->backing_object = backing_object;
790 }
791 
792 /*
793  * Insert an object into a backing_object's shadow list with an additional
794  * reference to the backing_object added.
795  */
796 static void
797 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
798 {
799 
800 	VM_OBJECT_ASSERT_WLOCKED(object);
801 
802 	if ((backing_object->flags & OBJ_ANON) != 0) {
803 		VM_OBJECT_WLOCK(backing_object);
804 		KASSERT((backing_object->flags & OBJ_DEAD) == 0,
805 		    ("shadowing dead anonymous object"));
806 		vm_object_reference_locked(backing_object);
807 		vm_object_backing_insert_locked(object, backing_object);
808 		vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
809 		VM_OBJECT_WUNLOCK(backing_object);
810 	} else {
811 		vm_object_reference(backing_object);
812 		object->backing_object = backing_object;
813 	}
814 }
815 
816 /*
817  * Transfer a backing reference from backing_object to object.
818  */
819 static void
820 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
821 {
822 	vm_object_t new_backing_object;
823 
824 	/*
825 	 * Note that the reference to backing_object->backing_object
826 	 * moves from within backing_object to within object.
827 	 */
828 	vm_object_backing_remove_locked(object);
829 	new_backing_object = backing_object->backing_object;
830 	if (new_backing_object == NULL)
831 		return;
832 	if ((new_backing_object->flags & OBJ_ANON) != 0) {
833 		VM_OBJECT_WLOCK(new_backing_object);
834 		vm_object_backing_remove_locked(backing_object);
835 		vm_object_backing_insert_locked(object, new_backing_object);
836 		VM_OBJECT_WUNLOCK(new_backing_object);
837 	} else {
838 		object->backing_object = new_backing_object;
839 		backing_object->backing_object = NULL;
840 	}
841 }
842 
843 /*
844  * Wait for a concurrent collapse to settle.
845  */
846 static void
847 vm_object_collapse_wait(vm_object_t object)
848 {
849 
850 	VM_OBJECT_ASSERT_WLOCKED(object);
851 
852 	while ((object->flags & OBJ_COLLAPSING) != 0) {
853 		vm_object_pip_wait(object, "vmcolwait");
854 		counter_u64_add(object_collapse_waits, 1);
855 	}
856 }
857 
858 /*
859  * Waits for a backing object to clear a pending collapse and returns
860  * it locked if it is an ANON object.
861  */
862 static vm_object_t
863 vm_object_backing_collapse_wait(vm_object_t object)
864 {
865 	vm_object_t backing_object;
866 
867 	VM_OBJECT_ASSERT_WLOCKED(object);
868 
869 	for (;;) {
870 		backing_object = object->backing_object;
871 		if (backing_object == NULL ||
872 		    (backing_object->flags & OBJ_ANON) == 0)
873 			return (NULL);
874 		VM_OBJECT_WLOCK(backing_object);
875 		if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
876 			break;
877 		VM_OBJECT_WUNLOCK(object);
878 		vm_object_pip_sleep(backing_object, "vmbckwait");
879 		counter_u64_add(object_collapse_waits, 1);
880 		VM_OBJECT_WLOCK(object);
881 	}
882 	return (backing_object);
883 }
884 
885 /*
886  *	vm_object_terminate_pages removes any remaining pageable pages
887  *	from the object and resets the object to an empty state.
888  */
889 static void
890 vm_object_terminate_pages(vm_object_t object)
891 {
892 	vm_page_t p, p_next;
893 
894 	VM_OBJECT_ASSERT_WLOCKED(object);
895 
896 	/*
897 	 * Free any remaining pageable pages.  This also removes them from the
898 	 * paging queues.  However, don't free wired pages, just remove them
899 	 * from the object.  Rather than incrementally removing each page from
900 	 * the object, the page and object are reset to any empty state.
901 	 */
902 	TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
903 		vm_page_assert_unbusied(p);
904 		KASSERT(p->object == object &&
905 		    (p->ref_count & VPRC_OBJREF) != 0,
906 		    ("vm_object_terminate_pages: page %p is inconsistent", p));
907 
908 		p->object = NULL;
909 		if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
910 			VM_CNT_INC(v_pfree);
911 			vm_page_free(p);
912 		}
913 	}
914 
915 	/*
916 	 * If the object contained any pages, then reset it to an empty state.
917 	 * None of the object's fields, including "resident_page_count", were
918 	 * modified by the preceding loop.
919 	 */
920 	if (object->resident_page_count != 0) {
921 		vm_radix_reclaim_allnodes(&object->rtree);
922 		TAILQ_INIT(&object->memq);
923 		object->resident_page_count = 0;
924 		if (object->type == OBJT_VNODE)
925 			vdrop(object->handle);
926 	}
927 }
928 
929 /*
930  *	vm_object_terminate actually destroys the specified object, freeing
931  *	up all previously used resources.
932  *
933  *	The object must be locked.
934  *	This routine may block.
935  */
936 void
937 vm_object_terminate(vm_object_t object)
938 {
939 
940 	VM_OBJECT_ASSERT_WLOCKED(object);
941 	KASSERT((object->flags & OBJ_DEAD) != 0,
942 	    ("terminating non-dead obj %p", object));
943 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
944 	    ("terminating collapsing obj %p", object));
945 	KASSERT(object->backing_object == NULL,
946 	    ("terminating shadow obj %p", object));
947 
948 	/*
949 	 * wait for the pageout daemon to be done with the object
950 	 */
951 	vm_object_pip_wait(object, "objtrm");
952 
953 	KASSERT(!blockcount_read(&object->paging_in_progress),
954 	    ("vm_object_terminate: pageout in progress"));
955 
956 	KASSERT(object->ref_count == 0,
957 	    ("vm_object_terminate: object with references, ref_count=%d",
958 	    object->ref_count));
959 
960 	if ((object->flags & OBJ_PG_DTOR) == 0)
961 		vm_object_terminate_pages(object);
962 
963 #if VM_NRESERVLEVEL > 0
964 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
965 		vm_reserv_break_all(object);
966 #endif
967 
968 	KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
969 	    object->type == OBJT_SWAP,
970 	    ("%s: non-swap obj %p has cred", __func__, object));
971 
972 	/*
973 	 * Let the pager know object is dead.
974 	 */
975 	vm_pager_deallocate(object);
976 	VM_OBJECT_WUNLOCK(object);
977 
978 	vm_object_destroy(object);
979 }
980 
981 /*
982  * Make the page read-only so that we can clear the object flags.  However, if
983  * this is a nosync mmap then the object is likely to stay dirty so do not
984  * mess with the page and do not clear the object flags.  Returns TRUE if the
985  * page should be flushed, and FALSE otherwise.
986  */
987 static boolean_t
988 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
989 {
990 
991 	vm_page_assert_busied(p);
992 
993 	/*
994 	 * If we have been asked to skip nosync pages and this is a
995 	 * nosync page, skip it.  Note that the object flags were not
996 	 * cleared in this case so we do not have to set them.
997 	 */
998 	if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
999 		*allclean = FALSE;
1000 		return (FALSE);
1001 	} else {
1002 		pmap_remove_write(p);
1003 		return (p->dirty != 0);
1004 	}
1005 }
1006 
1007 /*
1008  *	vm_object_page_clean
1009  *
1010  *	Clean all dirty pages in the specified range of object.  Leaves page
1011  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
1012  *	write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1013  *	leaving the object dirty.
1014  *
1015  *	For swap objects backing tmpfs regular files, do not flush anything,
1016  *	but remove write protection on the mapped pages to update mtime through
1017  *	mmaped writes.
1018  *
1019  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
1020  *	synchronous clustering mode implementation.
1021  *
1022  *	Odd semantics: if start == end, we clean everything.
1023  *
1024  *	The object must be locked.
1025  *
1026  *	Returns FALSE if some page from the range was not written, as
1027  *	reported by the pager, and TRUE otherwise.
1028  */
1029 boolean_t
1030 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1031     int flags)
1032 {
1033 	vm_page_t np, p;
1034 	vm_pindex_t pi, tend, tstart;
1035 	int curgeneration, n, pagerflags;
1036 	boolean_t eio, res, allclean;
1037 
1038 	VM_OBJECT_ASSERT_WLOCKED(object);
1039 
1040 	if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1041 		return (TRUE);
1042 
1043 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1044 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1045 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1046 
1047 	tstart = OFF_TO_IDX(start);
1048 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1049 	allclean = tstart == 0 && tend >= object->size;
1050 	res = TRUE;
1051 
1052 rescan:
1053 	curgeneration = object->generation;
1054 
1055 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1056 		pi = p->pindex;
1057 		if (pi >= tend)
1058 			break;
1059 		np = TAILQ_NEXT(p, listq);
1060 		if (vm_page_none_valid(p))
1061 			continue;
1062 		if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1063 			if (object->generation != curgeneration &&
1064 			    (flags & OBJPC_SYNC) != 0)
1065 				goto rescan;
1066 			np = vm_page_find_least(object, pi);
1067 			continue;
1068 		}
1069 		if (!vm_object_page_remove_write(p, flags, &allclean)) {
1070 			vm_page_xunbusy(p);
1071 			continue;
1072 		}
1073 		if (object->type == OBJT_VNODE) {
1074 			n = vm_object_page_collect_flush(object, p, pagerflags,
1075 			    flags, &allclean, &eio);
1076 			if (eio) {
1077 				res = FALSE;
1078 				allclean = FALSE;
1079 			}
1080 			if (object->generation != curgeneration &&
1081 			    (flags & OBJPC_SYNC) != 0)
1082 				goto rescan;
1083 
1084 			/*
1085 			 * If the VOP_PUTPAGES() did a truncated write, so
1086 			 * that even the first page of the run is not fully
1087 			 * written, vm_pageout_flush() returns 0 as the run
1088 			 * length.  Since the condition that caused truncated
1089 			 * write may be permanent, e.g. exhausted free space,
1090 			 * accepting n == 0 would cause an infinite loop.
1091 			 *
1092 			 * Forwarding the iterator leaves the unwritten page
1093 			 * behind, but there is not much we can do there if
1094 			 * filesystem refuses to write it.
1095 			 */
1096 			if (n == 0) {
1097 				n = 1;
1098 				allclean = FALSE;
1099 			}
1100 		} else {
1101 			n = 1;
1102 			vm_page_xunbusy(p);
1103 		}
1104 		np = vm_page_find_least(object, pi + n);
1105 	}
1106 #if 0
1107 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1108 #endif
1109 
1110 	/*
1111 	 * Leave updating cleangeneration for tmpfs objects to tmpfs
1112 	 * scan.  It needs to update mtime, which happens for other
1113 	 * filesystems during page writeouts.
1114 	 */
1115 	if (allclean && object->type == OBJT_VNODE)
1116 		object->cleangeneration = curgeneration;
1117 	return (res);
1118 }
1119 
1120 static int
1121 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1122     int flags, boolean_t *allclean, boolean_t *eio)
1123 {
1124 	vm_page_t ma[vm_pageout_page_count], p_first, tp;
1125 	int count, i, mreq, runlen;
1126 
1127 	vm_page_lock_assert(p, MA_NOTOWNED);
1128 	vm_page_assert_xbusied(p);
1129 	VM_OBJECT_ASSERT_WLOCKED(object);
1130 
1131 	count = 1;
1132 	mreq = 0;
1133 
1134 	for (tp = p; count < vm_pageout_page_count; count++) {
1135 		tp = vm_page_next(tp);
1136 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1137 			break;
1138 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1139 			vm_page_xunbusy(tp);
1140 			break;
1141 		}
1142 	}
1143 
1144 	for (p_first = p; count < vm_pageout_page_count; count++) {
1145 		tp = vm_page_prev(p_first);
1146 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1147 			break;
1148 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1149 			vm_page_xunbusy(tp);
1150 			break;
1151 		}
1152 		p_first = tp;
1153 		mreq++;
1154 	}
1155 
1156 	for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1157 		ma[i] = tp;
1158 
1159 	vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1160 	return (runlen);
1161 }
1162 
1163 /*
1164  * Note that there is absolutely no sense in writing out
1165  * anonymous objects, so we track down the vnode object
1166  * to write out.
1167  * We invalidate (remove) all pages from the address space
1168  * for semantic correctness.
1169  *
1170  * If the backing object is a device object with unmanaged pages, then any
1171  * mappings to the specified range of pages must be removed before this
1172  * function is called.
1173  *
1174  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1175  * may start out with a NULL object.
1176  */
1177 boolean_t
1178 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1179     boolean_t syncio, boolean_t invalidate)
1180 {
1181 	vm_object_t backing_object;
1182 	struct vnode *vp;
1183 	struct mount *mp;
1184 	int error, flags, fsync_after;
1185 	boolean_t res;
1186 
1187 	if (object == NULL)
1188 		return (TRUE);
1189 	res = TRUE;
1190 	error = 0;
1191 	VM_OBJECT_WLOCK(object);
1192 	while ((backing_object = object->backing_object) != NULL) {
1193 		VM_OBJECT_WLOCK(backing_object);
1194 		offset += object->backing_object_offset;
1195 		VM_OBJECT_WUNLOCK(object);
1196 		object = backing_object;
1197 		if (object->size < OFF_TO_IDX(offset + size))
1198 			size = IDX_TO_OFF(object->size) - offset;
1199 	}
1200 	/*
1201 	 * Flush pages if writing is allowed, invalidate them
1202 	 * if invalidation requested.  Pages undergoing I/O
1203 	 * will be ignored by vm_object_page_remove().
1204 	 *
1205 	 * We cannot lock the vnode and then wait for paging
1206 	 * to complete without deadlocking against vm_fault.
1207 	 * Instead we simply call vm_object_page_remove() and
1208 	 * allow it to block internally on a page-by-page
1209 	 * basis when it encounters pages undergoing async
1210 	 * I/O.
1211 	 */
1212 	if (object->type == OBJT_VNODE &&
1213 	    vm_object_mightbedirty(object) != 0 &&
1214 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1215 		VM_OBJECT_WUNLOCK(object);
1216 		(void) vn_start_write(vp, &mp, V_WAIT);
1217 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1218 		if (syncio && !invalidate && offset == 0 &&
1219 		    atop(size) == object->size) {
1220 			/*
1221 			 * If syncing the whole mapping of the file,
1222 			 * it is faster to schedule all the writes in
1223 			 * async mode, also allowing the clustering,
1224 			 * and then wait for i/o to complete.
1225 			 */
1226 			flags = 0;
1227 			fsync_after = TRUE;
1228 		} else {
1229 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1230 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1231 			fsync_after = FALSE;
1232 		}
1233 		VM_OBJECT_WLOCK(object);
1234 		res = vm_object_page_clean(object, offset, offset + size,
1235 		    flags);
1236 		VM_OBJECT_WUNLOCK(object);
1237 		if (fsync_after)
1238 			error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1239 		VOP_UNLOCK(vp);
1240 		vn_finished_write(mp);
1241 		if (error != 0)
1242 			res = FALSE;
1243 		VM_OBJECT_WLOCK(object);
1244 	}
1245 	if ((object->type == OBJT_VNODE ||
1246 	     object->type == OBJT_DEVICE) && invalidate) {
1247 		if (object->type == OBJT_DEVICE)
1248 			/*
1249 			 * The option OBJPR_NOTMAPPED must be passed here
1250 			 * because vm_object_page_remove() cannot remove
1251 			 * unmanaged mappings.
1252 			 */
1253 			flags = OBJPR_NOTMAPPED;
1254 		else if (old_msync)
1255 			flags = 0;
1256 		else
1257 			flags = OBJPR_CLEANONLY;
1258 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1259 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1260 	}
1261 	VM_OBJECT_WUNLOCK(object);
1262 	return (res);
1263 }
1264 
1265 /*
1266  * Determine whether the given advice can be applied to the object.  Advice is
1267  * not applied to unmanaged pages since they never belong to page queues, and
1268  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1269  * have been mapped at most once.
1270  */
1271 static bool
1272 vm_object_advice_applies(vm_object_t object, int advice)
1273 {
1274 
1275 	if ((object->flags & OBJ_UNMANAGED) != 0)
1276 		return (false);
1277 	if (advice != MADV_FREE)
1278 		return (true);
1279 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1280 	    (OBJ_ONEMAPPING | OBJ_ANON));
1281 }
1282 
1283 static void
1284 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1285     vm_size_t size)
1286 {
1287 
1288 	if (advice == MADV_FREE && object->type == OBJT_SWAP)
1289 		swap_pager_freespace(object, pindex, size);
1290 }
1291 
1292 /*
1293  *	vm_object_madvise:
1294  *
1295  *	Implements the madvise function at the object/page level.
1296  *
1297  *	MADV_WILLNEED	(any object)
1298  *
1299  *	    Activate the specified pages if they are resident.
1300  *
1301  *	MADV_DONTNEED	(any object)
1302  *
1303  *	    Deactivate the specified pages if they are resident.
1304  *
1305  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1306  *			 OBJ_ONEMAPPING only)
1307  *
1308  *	    Deactivate and clean the specified pages if they are
1309  *	    resident.  This permits the process to reuse the pages
1310  *	    without faulting or the kernel to reclaim the pages
1311  *	    without I/O.
1312  */
1313 void
1314 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1315     int advice)
1316 {
1317 	vm_pindex_t tpindex;
1318 	vm_object_t backing_object, tobject;
1319 	vm_page_t m, tm;
1320 
1321 	if (object == NULL)
1322 		return;
1323 
1324 relookup:
1325 	VM_OBJECT_WLOCK(object);
1326 	if (!vm_object_advice_applies(object, advice)) {
1327 		VM_OBJECT_WUNLOCK(object);
1328 		return;
1329 	}
1330 	for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1331 		tobject = object;
1332 
1333 		/*
1334 		 * If the next page isn't resident in the top-level object, we
1335 		 * need to search the shadow chain.  When applying MADV_FREE, we
1336 		 * take care to release any swap space used to store
1337 		 * non-resident pages.
1338 		 */
1339 		if (m == NULL || pindex < m->pindex) {
1340 			/*
1341 			 * Optimize a common case: if the top-level object has
1342 			 * no backing object, we can skip over the non-resident
1343 			 * range in constant time.
1344 			 */
1345 			if (object->backing_object == NULL) {
1346 				tpindex = (m != NULL && m->pindex < end) ?
1347 				    m->pindex : end;
1348 				vm_object_madvise_freespace(object, advice,
1349 				    pindex, tpindex - pindex);
1350 				if ((pindex = tpindex) == end)
1351 					break;
1352 				goto next_page;
1353 			}
1354 
1355 			tpindex = pindex;
1356 			do {
1357 				vm_object_madvise_freespace(tobject, advice,
1358 				    tpindex, 1);
1359 				/*
1360 				 * Prepare to search the next object in the
1361 				 * chain.
1362 				 */
1363 				backing_object = tobject->backing_object;
1364 				if (backing_object == NULL)
1365 					goto next_pindex;
1366 				VM_OBJECT_WLOCK(backing_object);
1367 				tpindex +=
1368 				    OFF_TO_IDX(tobject->backing_object_offset);
1369 				if (tobject != object)
1370 					VM_OBJECT_WUNLOCK(tobject);
1371 				tobject = backing_object;
1372 				if (!vm_object_advice_applies(tobject, advice))
1373 					goto next_pindex;
1374 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1375 			    NULL);
1376 		} else {
1377 next_page:
1378 			tm = m;
1379 			m = TAILQ_NEXT(m, listq);
1380 		}
1381 
1382 		/*
1383 		 * If the page is not in a normal state, skip it.  The page
1384 		 * can not be invalidated while the object lock is held.
1385 		 */
1386 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1387 			goto next_pindex;
1388 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1389 		    ("vm_object_madvise: page %p is fictitious", tm));
1390 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1391 		    ("vm_object_madvise: page %p is not managed", tm));
1392 		if (vm_page_tryxbusy(tm) == 0) {
1393 			if (object != tobject)
1394 				VM_OBJECT_WUNLOCK(object);
1395 			if (advice == MADV_WILLNEED) {
1396 				/*
1397 				 * Reference the page before unlocking and
1398 				 * sleeping so that the page daemon is less
1399 				 * likely to reclaim it.
1400 				 */
1401 				vm_page_aflag_set(tm, PGA_REFERENCED);
1402 			}
1403 			vm_page_busy_sleep(tm, "madvpo", false);
1404   			goto relookup;
1405 		}
1406 		vm_page_advise(tm, advice);
1407 		vm_page_xunbusy(tm);
1408 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1409 next_pindex:
1410 		if (tobject != object)
1411 			VM_OBJECT_WUNLOCK(tobject);
1412 	}
1413 	VM_OBJECT_WUNLOCK(object);
1414 }
1415 
1416 /*
1417  *	vm_object_shadow:
1418  *
1419  *	Create a new object which is backed by the
1420  *	specified existing object range.  The source
1421  *	object reference is deallocated.
1422  *
1423  *	The new object and offset into that object
1424  *	are returned in the source parameters.
1425  */
1426 void
1427 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1428     struct ucred *cred, bool shared)
1429 {
1430 	vm_object_t source;
1431 	vm_object_t result;
1432 
1433 	source = *object;
1434 
1435 	/*
1436 	 * Don't create the new object if the old object isn't shared.
1437 	 *
1438 	 * If we hold the only reference we can guarantee that it won't
1439 	 * increase while we have the map locked.  Otherwise the race is
1440 	 * harmless and we will end up with an extra shadow object that
1441 	 * will be collapsed later.
1442 	 */
1443 	if (source != NULL && source->ref_count == 1 &&
1444 	    (source->flags & OBJ_ANON) != 0)
1445 		return;
1446 
1447 	/*
1448 	 * Allocate a new object with the given length.
1449 	 */
1450 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1451 
1452 	/*
1453 	 * Store the offset into the source object, and fix up the offset into
1454 	 * the new object.
1455 	 */
1456 	result->backing_object_offset = *offset;
1457 
1458 	if (shared || source != NULL) {
1459 		VM_OBJECT_WLOCK(result);
1460 
1461 		/*
1462 		 * The new object shadows the source object, adding a
1463 		 * reference to it.  Our caller changes his reference
1464 		 * to point to the new object, removing a reference to
1465 		 * the source object.  Net result: no change of
1466 		 * reference count, unless the caller needs to add one
1467 		 * more reference due to forking a shared map entry.
1468 		 */
1469 		if (shared) {
1470 			vm_object_reference_locked(result);
1471 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1472 		}
1473 
1474 		/*
1475 		 * Try to optimize the result object's page color when
1476 		 * shadowing in order to maintain page coloring
1477 		 * consistency in the combined shadowed object.
1478 		 */
1479 		if (source != NULL) {
1480 			vm_object_backing_insert(result, source);
1481 			result->domain = source->domain;
1482 #if VM_NRESERVLEVEL > 0
1483 			result->flags |= source->flags & OBJ_COLORED;
1484 			result->pg_color = (source->pg_color +
1485 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1486 			    1)) - 1);
1487 #endif
1488 		}
1489 		VM_OBJECT_WUNLOCK(result);
1490 	}
1491 
1492 	/*
1493 	 * Return the new things
1494 	 */
1495 	*offset = 0;
1496 	*object = result;
1497 }
1498 
1499 /*
1500  *	vm_object_split:
1501  *
1502  * Split the pages in a map entry into a new object.  This affords
1503  * easier removal of unused pages, and keeps object inheritance from
1504  * being a negative impact on memory usage.
1505  */
1506 void
1507 vm_object_split(vm_map_entry_t entry)
1508 {
1509 	vm_page_t m, m_next;
1510 	vm_object_t orig_object, new_object, backing_object;
1511 	vm_pindex_t idx, offidxstart;
1512 	vm_size_t size;
1513 
1514 	orig_object = entry->object.vm_object;
1515 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1516 	    ("vm_object_split:  Splitting object with multiple mappings."));
1517 	if ((orig_object->flags & OBJ_ANON) == 0)
1518 		return;
1519 	if (orig_object->ref_count <= 1)
1520 		return;
1521 	VM_OBJECT_WUNLOCK(orig_object);
1522 
1523 	offidxstart = OFF_TO_IDX(entry->offset);
1524 	size = atop(entry->end - entry->start);
1525 
1526 	/*
1527 	 * If swap_pager_copy() is later called, it will convert new_object
1528 	 * into a swap object.
1529 	 */
1530 	new_object = vm_object_allocate_anon(size, orig_object,
1531 	    orig_object->cred, ptoa(size));
1532 
1533 	/*
1534 	 * We must wait for the orig_object to complete any in-progress
1535 	 * collapse so that the swap blocks are stable below.  The
1536 	 * additional reference on backing_object by new object will
1537 	 * prevent further collapse operations until split completes.
1538 	 */
1539 	VM_OBJECT_WLOCK(orig_object);
1540 	vm_object_collapse_wait(orig_object);
1541 
1542 	/*
1543 	 * At this point, the new object is still private, so the order in
1544 	 * which the original and new objects are locked does not matter.
1545 	 */
1546 	VM_OBJECT_WLOCK(new_object);
1547 	new_object->domain = orig_object->domain;
1548 	backing_object = orig_object->backing_object;
1549 	if (backing_object != NULL) {
1550 		vm_object_backing_insert_ref(new_object, backing_object);
1551 		new_object->backing_object_offset =
1552 		    orig_object->backing_object_offset + entry->offset;
1553 	}
1554 	if (orig_object->cred != NULL) {
1555 		crhold(orig_object->cred);
1556 		KASSERT(orig_object->charge >= ptoa(size),
1557 		    ("orig_object->charge < 0"));
1558 		orig_object->charge -= ptoa(size);
1559 	}
1560 
1561 	/*
1562 	 * Mark the split operation so that swap_pager_getpages() knows
1563 	 * that the object is in transition.
1564 	 */
1565 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1566 retry:
1567 	m = vm_page_find_least(orig_object, offidxstart);
1568 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1569 	    m = m_next) {
1570 		m_next = TAILQ_NEXT(m, listq);
1571 
1572 		/*
1573 		 * We must wait for pending I/O to complete before we can
1574 		 * rename the page.
1575 		 *
1576 		 * We do not have to VM_PROT_NONE the page as mappings should
1577 		 * not be changed by this operation.
1578 		 */
1579 		if (vm_page_tryxbusy(m) == 0) {
1580 			VM_OBJECT_WUNLOCK(new_object);
1581 			vm_page_sleep_if_busy(m, "spltwt");
1582 			VM_OBJECT_WLOCK(new_object);
1583 			goto retry;
1584 		}
1585 
1586 		/*
1587 		 * The page was left invalid.  Likely placed there by
1588 		 * an incomplete fault.  Just remove and ignore.
1589 		 */
1590 		if (vm_page_none_valid(m)) {
1591 			if (vm_page_remove(m))
1592 				vm_page_free(m);
1593 			continue;
1594 		}
1595 
1596 		/* vm_page_rename() will dirty the page. */
1597 		if (vm_page_rename(m, new_object, idx)) {
1598 			vm_page_xunbusy(m);
1599 			VM_OBJECT_WUNLOCK(new_object);
1600 			VM_OBJECT_WUNLOCK(orig_object);
1601 			vm_radix_wait();
1602 			VM_OBJECT_WLOCK(orig_object);
1603 			VM_OBJECT_WLOCK(new_object);
1604 			goto retry;
1605 		}
1606 
1607 #if VM_NRESERVLEVEL > 0
1608 		/*
1609 		 * If some of the reservation's allocated pages remain with
1610 		 * the original object, then transferring the reservation to
1611 		 * the new object is neither particularly beneficial nor
1612 		 * particularly harmful as compared to leaving the reservation
1613 		 * with the original object.  If, however, all of the
1614 		 * reservation's allocated pages are transferred to the new
1615 		 * object, then transferring the reservation is typically
1616 		 * beneficial.  Determining which of these two cases applies
1617 		 * would be more costly than unconditionally renaming the
1618 		 * reservation.
1619 		 */
1620 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1621 #endif
1622 		if (orig_object->type != OBJT_SWAP)
1623 			vm_page_xunbusy(m);
1624 	}
1625 	if (orig_object->type == OBJT_SWAP) {
1626 		/*
1627 		 * swap_pager_copy() can sleep, in which case the orig_object's
1628 		 * and new_object's locks are released and reacquired.
1629 		 */
1630 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1631 		TAILQ_FOREACH(m, &new_object->memq, listq)
1632 			vm_page_xunbusy(m);
1633 	}
1634 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1635 	VM_OBJECT_WUNLOCK(orig_object);
1636 	VM_OBJECT_WUNLOCK(new_object);
1637 	entry->object.vm_object = new_object;
1638 	entry->offset = 0LL;
1639 	vm_object_deallocate(orig_object);
1640 	VM_OBJECT_WLOCK(new_object);
1641 }
1642 
1643 static vm_page_t
1644 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1645 {
1646 	vm_object_t backing_object;
1647 
1648 	VM_OBJECT_ASSERT_WLOCKED(object);
1649 	backing_object = object->backing_object;
1650 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1651 
1652 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1653 	    ("invalid ownership %p %p %p", p, object, backing_object));
1654 	/* The page is only NULL when rename fails. */
1655 	if (p == NULL) {
1656 		VM_OBJECT_WUNLOCK(object);
1657 		VM_OBJECT_WUNLOCK(backing_object);
1658 		vm_radix_wait();
1659 	} else {
1660 		if (p->object == object)
1661 			VM_OBJECT_WUNLOCK(backing_object);
1662 		else
1663 			VM_OBJECT_WUNLOCK(object);
1664 		vm_page_busy_sleep(p, "vmocol", false);
1665 	}
1666 	VM_OBJECT_WLOCK(object);
1667 	VM_OBJECT_WLOCK(backing_object);
1668 	return (TAILQ_FIRST(&backing_object->memq));
1669 }
1670 
1671 static bool
1672 vm_object_scan_all_shadowed(vm_object_t object)
1673 {
1674 	vm_object_t backing_object;
1675 	vm_page_t p, pp;
1676 	vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1677 
1678 	VM_OBJECT_ASSERT_WLOCKED(object);
1679 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1680 
1681 	backing_object = object->backing_object;
1682 
1683 	if ((backing_object->flags & OBJ_ANON) == 0)
1684 		return (false);
1685 
1686 	pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1687 	p = vm_page_find_least(backing_object, pi);
1688 	ps = swap_pager_find_least(backing_object, pi);
1689 
1690 	/*
1691 	 * Only check pages inside the parent object's range and
1692 	 * inside the parent object's mapping of the backing object.
1693 	 */
1694 	for (;; pi++) {
1695 		if (p != NULL && p->pindex < pi)
1696 			p = TAILQ_NEXT(p, listq);
1697 		if (ps < pi)
1698 			ps = swap_pager_find_least(backing_object, pi);
1699 		if (p == NULL && ps >= backing_object->size)
1700 			break;
1701 		else if (p == NULL)
1702 			pi = ps;
1703 		else
1704 			pi = MIN(p->pindex, ps);
1705 
1706 		new_pindex = pi - backing_offset_index;
1707 		if (new_pindex >= object->size)
1708 			break;
1709 
1710 		if (p != NULL) {
1711 			/*
1712 			 * If the backing object page is busy a
1713 			 * grandparent or older page may still be
1714 			 * undergoing CoW.  It is not safe to collapse
1715 			 * the backing object until it is quiesced.
1716 			 */
1717 			if (vm_page_tryxbusy(p) == 0)
1718 				return (false);
1719 
1720 			/*
1721 			 * We raced with the fault handler that left
1722 			 * newly allocated invalid page on the object
1723 			 * queue and retried.
1724 			 */
1725 			if (!vm_page_all_valid(p))
1726 				goto unbusy_ret;
1727 		}
1728 
1729 		/*
1730 		 * See if the parent has the page or if the parent's object
1731 		 * pager has the page.  If the parent has the page but the page
1732 		 * is not valid, the parent's object pager must have the page.
1733 		 *
1734 		 * If this fails, the parent does not completely shadow the
1735 		 * object and we might as well give up now.
1736 		 */
1737 		pp = vm_page_lookup(object, new_pindex);
1738 
1739 		/*
1740 		 * The valid check here is stable due to object lock
1741 		 * being required to clear valid and initiate paging.
1742 		 * Busy of p disallows fault handler to validate pp.
1743 		 */
1744 		if ((pp == NULL || vm_page_none_valid(pp)) &&
1745 		    !vm_pager_has_page(object, new_pindex, NULL, NULL))
1746 			goto unbusy_ret;
1747 		if (p != NULL)
1748 			vm_page_xunbusy(p);
1749 	}
1750 	return (true);
1751 
1752 unbusy_ret:
1753 	if (p != NULL)
1754 		vm_page_xunbusy(p);
1755 	return (false);
1756 }
1757 
1758 static void
1759 vm_object_collapse_scan(vm_object_t object)
1760 {
1761 	vm_object_t backing_object;
1762 	vm_page_t next, p, pp;
1763 	vm_pindex_t backing_offset_index, new_pindex;
1764 
1765 	VM_OBJECT_ASSERT_WLOCKED(object);
1766 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1767 
1768 	backing_object = object->backing_object;
1769 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1770 
1771 	/*
1772 	 * Our scan
1773 	 */
1774 	for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1775 		next = TAILQ_NEXT(p, listq);
1776 		new_pindex = p->pindex - backing_offset_index;
1777 
1778 		/*
1779 		 * Check for busy page
1780 		 */
1781 		if (vm_page_tryxbusy(p) == 0) {
1782 			next = vm_object_collapse_scan_wait(object, p);
1783 			continue;
1784 		}
1785 
1786 		KASSERT(object->backing_object == backing_object,
1787 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1788 		    object->backing_object, backing_object));
1789 		KASSERT(p->object == backing_object,
1790 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1791 		    p->object, backing_object));
1792 
1793 		if (p->pindex < backing_offset_index ||
1794 		    new_pindex >= object->size) {
1795 			if (backing_object->type == OBJT_SWAP)
1796 				swap_pager_freespace(backing_object, p->pindex,
1797 				    1);
1798 
1799 			KASSERT(!pmap_page_is_mapped(p),
1800 			    ("freeing mapped page %p", p));
1801 			if (vm_page_remove(p))
1802 				vm_page_free(p);
1803 			continue;
1804 		}
1805 
1806 		if (!vm_page_all_valid(p)) {
1807 			KASSERT(!pmap_page_is_mapped(p),
1808 			    ("freeing mapped page %p", p));
1809 			if (vm_page_remove(p))
1810 				vm_page_free(p);
1811 			continue;
1812 		}
1813 
1814 		pp = vm_page_lookup(object, new_pindex);
1815 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1816 			vm_page_xunbusy(p);
1817 			/*
1818 			 * The page in the parent is busy and possibly not
1819 			 * (yet) valid.  Until its state is finalized by the
1820 			 * busy bit owner, we can't tell whether it shadows the
1821 			 * original page.
1822 			 */
1823 			next = vm_object_collapse_scan_wait(object, pp);
1824 			continue;
1825 		}
1826 
1827 		if (pp != NULL && vm_page_none_valid(pp)) {
1828 			/*
1829 			 * The page was invalid in the parent.  Likely placed
1830 			 * there by an incomplete fault.  Just remove and
1831 			 * ignore.  p can replace it.
1832 			 */
1833 			if (vm_page_remove(pp))
1834 				vm_page_free(pp);
1835 			pp = NULL;
1836 		}
1837 
1838 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1839 			NULL)) {
1840 			/*
1841 			 * The page already exists in the parent OR swap exists
1842 			 * for this location in the parent.  Leave the parent's
1843 			 * page alone.  Destroy the original page from the
1844 			 * backing object.
1845 			 */
1846 			if (backing_object->type == OBJT_SWAP)
1847 				swap_pager_freespace(backing_object, p->pindex,
1848 				    1);
1849 			KASSERT(!pmap_page_is_mapped(p),
1850 			    ("freeing mapped page %p", p));
1851 			if (vm_page_remove(p))
1852 				vm_page_free(p);
1853 			if (pp != NULL)
1854 				vm_page_xunbusy(pp);
1855 			continue;
1856 		}
1857 
1858 		/*
1859 		 * Page does not exist in parent, rename the page from the
1860 		 * backing object to the main object.
1861 		 *
1862 		 * If the page was mapped to a process, it can remain mapped
1863 		 * through the rename.  vm_page_rename() will dirty the page.
1864 		 */
1865 		if (vm_page_rename(p, object, new_pindex)) {
1866 			vm_page_xunbusy(p);
1867 			next = vm_object_collapse_scan_wait(object, NULL);
1868 			continue;
1869 		}
1870 
1871 		/* Use the old pindex to free the right page. */
1872 		if (backing_object->type == OBJT_SWAP)
1873 			swap_pager_freespace(backing_object,
1874 			    new_pindex + backing_offset_index, 1);
1875 
1876 #if VM_NRESERVLEVEL > 0
1877 		/*
1878 		 * Rename the reservation.
1879 		 */
1880 		vm_reserv_rename(p, object, backing_object,
1881 		    backing_offset_index);
1882 #endif
1883 		vm_page_xunbusy(p);
1884 	}
1885 	return;
1886 }
1887 
1888 /*
1889  *	vm_object_collapse:
1890  *
1891  *	Collapse an object with the object backing it.
1892  *	Pages in the backing object are moved into the
1893  *	parent, and the backing object is deallocated.
1894  */
1895 void
1896 vm_object_collapse(vm_object_t object)
1897 {
1898 	vm_object_t backing_object, new_backing_object;
1899 
1900 	VM_OBJECT_ASSERT_WLOCKED(object);
1901 
1902 	while (TRUE) {
1903 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1904 		    ("collapsing invalid object"));
1905 
1906 		/*
1907 		 * Wait for the backing_object to finish any pending
1908 		 * collapse so that the caller sees the shortest possible
1909 		 * shadow chain.
1910 		 */
1911 		backing_object = vm_object_backing_collapse_wait(object);
1912 		if (backing_object == NULL)
1913 			return;
1914 
1915 		KASSERT(object->ref_count > 0 &&
1916 		    object->ref_count > object->shadow_count,
1917 		    ("collapse with invalid ref %d or shadow %d count.",
1918 		    object->ref_count, object->shadow_count));
1919 		KASSERT((backing_object->flags &
1920 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1921 		    ("vm_object_collapse: Backing object already collapsing."));
1922 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1923 		    ("vm_object_collapse: object is already collapsing."));
1924 
1925 		/*
1926 		 * We know that we can either collapse the backing object if
1927 		 * the parent is the only reference to it, or (perhaps) have
1928 		 * the parent bypass the object if the parent happens to shadow
1929 		 * all the resident pages in the entire backing object.
1930 		 */
1931 		if (backing_object->ref_count == 1) {
1932 			KASSERT(backing_object->shadow_count == 1,
1933 			    ("vm_object_collapse: shadow_count: %d",
1934 			    backing_object->shadow_count));
1935 			vm_object_pip_add(object, 1);
1936 			vm_object_set_flag(object, OBJ_COLLAPSING);
1937 			vm_object_pip_add(backing_object, 1);
1938 			vm_object_set_flag(backing_object, OBJ_DEAD);
1939 
1940 			/*
1941 			 * If there is exactly one reference to the backing
1942 			 * object, we can collapse it into the parent.
1943 			 */
1944 			vm_object_collapse_scan(object);
1945 
1946 #if VM_NRESERVLEVEL > 0
1947 			/*
1948 			 * Break any reservations from backing_object.
1949 			 */
1950 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1951 				vm_reserv_break_all(backing_object);
1952 #endif
1953 
1954 			/*
1955 			 * Move the pager from backing_object to object.
1956 			 */
1957 			if (backing_object->type == OBJT_SWAP) {
1958 				/*
1959 				 * swap_pager_copy() can sleep, in which case
1960 				 * the backing_object's and object's locks are
1961 				 * released and reacquired.
1962 				 * Since swap_pager_copy() is being asked to
1963 				 * destroy backing_object, it will change the
1964 				 * type to OBJT_DEFAULT.
1965 				 */
1966 				swap_pager_copy(
1967 				    backing_object,
1968 				    object,
1969 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1970 			}
1971 
1972 			/*
1973 			 * Object now shadows whatever backing_object did.
1974 			 */
1975 			vm_object_clear_flag(object, OBJ_COLLAPSING);
1976 			vm_object_backing_transfer(object, backing_object);
1977 			object->backing_object_offset +=
1978 			    backing_object->backing_object_offset;
1979 			VM_OBJECT_WUNLOCK(object);
1980 			vm_object_pip_wakeup(object);
1981 
1982 			/*
1983 			 * Discard backing_object.
1984 			 *
1985 			 * Since the backing object has no pages, no pager left,
1986 			 * and no object references within it, all that is
1987 			 * necessary is to dispose of it.
1988 			 */
1989 			KASSERT(backing_object->ref_count == 1, (
1990 "backing_object %p was somehow re-referenced during collapse!",
1991 			    backing_object));
1992 			vm_object_pip_wakeup(backing_object);
1993 			(void)refcount_release(&backing_object->ref_count);
1994 			vm_object_terminate(backing_object);
1995 			counter_u64_add(object_collapses, 1);
1996 			VM_OBJECT_WLOCK(object);
1997 		} else {
1998 			/*
1999 			 * If we do not entirely shadow the backing object,
2000 			 * there is nothing we can do so we give up.
2001 			 *
2002 			 * The object lock and backing_object lock must not
2003 			 * be dropped during this sequence.
2004 			 */
2005 			if (!vm_object_scan_all_shadowed(object)) {
2006 				VM_OBJECT_WUNLOCK(backing_object);
2007 				break;
2008 			}
2009 
2010 			/*
2011 			 * Make the parent shadow the next object in the
2012 			 * chain.  Deallocating backing_object will not remove
2013 			 * it, since its reference count is at least 2.
2014 			 */
2015 			vm_object_backing_remove_locked(object);
2016 			new_backing_object = backing_object->backing_object;
2017 			if (new_backing_object != NULL) {
2018 				vm_object_backing_insert_ref(object,
2019 				    new_backing_object);
2020 				object->backing_object_offset +=
2021 				    backing_object->backing_object_offset;
2022 			}
2023 
2024 			/*
2025 			 * Drop the reference count on backing_object. Since
2026 			 * its ref_count was at least 2, it will not vanish.
2027 			 */
2028 			(void)refcount_release(&backing_object->ref_count);
2029 			KASSERT(backing_object->ref_count >= 1, (
2030 "backing_object %p was somehow dereferenced during collapse!",
2031 			    backing_object));
2032 			VM_OBJECT_WUNLOCK(backing_object);
2033 			counter_u64_add(object_bypasses, 1);
2034 		}
2035 
2036 		/*
2037 		 * Try again with this object's new backing object.
2038 		 */
2039 	}
2040 }
2041 
2042 /*
2043  *	vm_object_page_remove:
2044  *
2045  *	For the given object, either frees or invalidates each of the
2046  *	specified pages.  In general, a page is freed.  However, if a page is
2047  *	wired for any reason other than the existence of a managed, wired
2048  *	mapping, then it may be invalidated but not removed from the object.
2049  *	Pages are specified by the given range ["start", "end") and the option
2050  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
2051  *	extends from "start" to the end of the object.  If the option
2052  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2053  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
2054  *	specified, then the pages within the specified range must have no
2055  *	mappings.  Otherwise, if this option is not specified, any mappings to
2056  *	the specified pages are removed before the pages are freed or
2057  *	invalidated.
2058  *
2059  *	In general, this operation should only be performed on objects that
2060  *	contain managed pages.  There are, however, two exceptions.  First, it
2061  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
2062  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2063  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
2064  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
2065  *
2066  *	The object must be locked.
2067  */
2068 void
2069 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2070     int options)
2071 {
2072 	vm_page_t p, next;
2073 
2074 	VM_OBJECT_ASSERT_WLOCKED(object);
2075 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2076 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2077 	    ("vm_object_page_remove: illegal options for object %p", object));
2078 	if (object->resident_page_count == 0)
2079 		return;
2080 	vm_object_pip_add(object, 1);
2081 again:
2082 	p = vm_page_find_least(object, start);
2083 
2084 	/*
2085 	 * Here, the variable "p" is either (1) the page with the least pindex
2086 	 * greater than or equal to the parameter "start" or (2) NULL.
2087 	 */
2088 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2089 		next = TAILQ_NEXT(p, listq);
2090 
2091 		/*
2092 		 * If the page is wired for any reason besides the existence
2093 		 * of managed, wired mappings, then it cannot be freed.  For
2094 		 * example, fictitious pages, which represent device memory,
2095 		 * are inherently wired and cannot be freed.  They can,
2096 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2097 		 * not specified.
2098 		 */
2099 		if (vm_page_tryxbusy(p) == 0) {
2100 			vm_page_sleep_if_busy(p, "vmopar");
2101 			goto again;
2102 		}
2103 		if (vm_page_wired(p)) {
2104 wired:
2105 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2106 			    object->ref_count != 0)
2107 				pmap_remove_all(p);
2108 			if ((options & OBJPR_CLEANONLY) == 0) {
2109 				vm_page_invalid(p);
2110 				vm_page_undirty(p);
2111 			}
2112 			vm_page_xunbusy(p);
2113 			continue;
2114 		}
2115 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2116 		    ("vm_object_page_remove: page %p is fictitious", p));
2117 		if ((options & OBJPR_CLEANONLY) != 0 &&
2118 		    !vm_page_none_valid(p)) {
2119 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2120 			    object->ref_count != 0 &&
2121 			    !vm_page_try_remove_write(p))
2122 				goto wired;
2123 			if (p->dirty != 0) {
2124 				vm_page_xunbusy(p);
2125 				continue;
2126 			}
2127 		}
2128 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2129 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2130 			goto wired;
2131 		vm_page_free(p);
2132 	}
2133 	vm_object_pip_wakeup(object);
2134 }
2135 
2136 /*
2137  *	vm_object_page_noreuse:
2138  *
2139  *	For the given object, attempt to move the specified pages to
2140  *	the head of the inactive queue.  This bypasses regular LRU
2141  *	operation and allows the pages to be reused quickly under memory
2142  *	pressure.  If a page is wired for any reason, then it will not
2143  *	be queued.  Pages are specified by the range ["start", "end").
2144  *	As a special case, if "end" is zero, then the range extends from
2145  *	"start" to the end of the object.
2146  *
2147  *	This operation should only be performed on objects that
2148  *	contain non-fictitious, managed pages.
2149  *
2150  *	The object must be locked.
2151  */
2152 void
2153 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2154 {
2155 	vm_page_t p, next;
2156 
2157 	VM_OBJECT_ASSERT_LOCKED(object);
2158 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2159 	    ("vm_object_page_noreuse: illegal object %p", object));
2160 	if (object->resident_page_count == 0)
2161 		return;
2162 	p = vm_page_find_least(object, start);
2163 
2164 	/*
2165 	 * Here, the variable "p" is either (1) the page with the least pindex
2166 	 * greater than or equal to the parameter "start" or (2) NULL.
2167 	 */
2168 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2169 		next = TAILQ_NEXT(p, listq);
2170 		vm_page_deactivate_noreuse(p);
2171 	}
2172 }
2173 
2174 /*
2175  *	Populate the specified range of the object with valid pages.  Returns
2176  *	TRUE if the range is successfully populated and FALSE otherwise.
2177  *
2178  *	Note: This function should be optimized to pass a larger array of
2179  *	pages to vm_pager_get_pages() before it is applied to a non-
2180  *	OBJT_DEVICE object.
2181  *
2182  *	The object must be locked.
2183  */
2184 boolean_t
2185 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2186 {
2187 	vm_page_t m;
2188 	vm_pindex_t pindex;
2189 	int rv;
2190 
2191 	VM_OBJECT_ASSERT_WLOCKED(object);
2192 	for (pindex = start; pindex < end; pindex++) {
2193 		rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2194 		if (rv != VM_PAGER_OK)
2195 			break;
2196 
2197 		/*
2198 		 * Keep "m" busy because a subsequent iteration may unlock
2199 		 * the object.
2200 		 */
2201 	}
2202 	if (pindex > start) {
2203 		m = vm_page_lookup(object, start);
2204 		while (m != NULL && m->pindex < pindex) {
2205 			vm_page_xunbusy(m);
2206 			m = TAILQ_NEXT(m, listq);
2207 		}
2208 	}
2209 	return (pindex == end);
2210 }
2211 
2212 /*
2213  *	Routine:	vm_object_coalesce
2214  *	Function:	Coalesces two objects backing up adjoining
2215  *			regions of memory into a single object.
2216  *
2217  *	returns TRUE if objects were combined.
2218  *
2219  *	NOTE:	Only works at the moment if the second object is NULL -
2220  *		if it's not, which object do we lock first?
2221  *
2222  *	Parameters:
2223  *		prev_object	First object to coalesce
2224  *		prev_offset	Offset into prev_object
2225  *		prev_size	Size of reference to prev_object
2226  *		next_size	Size of reference to the second object
2227  *		reserved	Indicator that extension region has
2228  *				swap accounted for
2229  *
2230  *	Conditions:
2231  *	The object must *not* be locked.
2232  */
2233 boolean_t
2234 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2235     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2236 {
2237 	vm_pindex_t next_pindex;
2238 
2239 	if (prev_object == NULL)
2240 		return (TRUE);
2241 	if ((prev_object->flags & OBJ_ANON) == 0)
2242 		return (FALSE);
2243 
2244 	VM_OBJECT_WLOCK(prev_object);
2245 	/*
2246 	 * Try to collapse the object first.
2247 	 */
2248 	vm_object_collapse(prev_object);
2249 
2250 	/*
2251 	 * Can't coalesce if: . more than one reference . paged out . shadows
2252 	 * another object . has a copy elsewhere (any of which mean that the
2253 	 * pages not mapped to prev_entry may be in use anyway)
2254 	 */
2255 	if (prev_object->backing_object != NULL) {
2256 		VM_OBJECT_WUNLOCK(prev_object);
2257 		return (FALSE);
2258 	}
2259 
2260 	prev_size >>= PAGE_SHIFT;
2261 	next_size >>= PAGE_SHIFT;
2262 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2263 
2264 	if (prev_object->ref_count > 1 &&
2265 	    prev_object->size != next_pindex &&
2266 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2267 		VM_OBJECT_WUNLOCK(prev_object);
2268 		return (FALSE);
2269 	}
2270 
2271 	/*
2272 	 * Account for the charge.
2273 	 */
2274 	if (prev_object->cred != NULL) {
2275 
2276 		/*
2277 		 * If prev_object was charged, then this mapping,
2278 		 * although not charged now, may become writable
2279 		 * later. Non-NULL cred in the object would prevent
2280 		 * swap reservation during enabling of the write
2281 		 * access, so reserve swap now. Failed reservation
2282 		 * cause allocation of the separate object for the map
2283 		 * entry, and swap reservation for this entry is
2284 		 * managed in appropriate time.
2285 		 */
2286 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2287 		    prev_object->cred)) {
2288 			VM_OBJECT_WUNLOCK(prev_object);
2289 			return (FALSE);
2290 		}
2291 		prev_object->charge += ptoa(next_size);
2292 	}
2293 
2294 	/*
2295 	 * Remove any pages that may still be in the object from a previous
2296 	 * deallocation.
2297 	 */
2298 	if (next_pindex < prev_object->size) {
2299 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2300 		    next_size, 0);
2301 		if (prev_object->type == OBJT_SWAP)
2302 			swap_pager_freespace(prev_object,
2303 					     next_pindex, next_size);
2304 #if 0
2305 		if (prev_object->cred != NULL) {
2306 			KASSERT(prev_object->charge >=
2307 			    ptoa(prev_object->size - next_pindex),
2308 			    ("object %p overcharged 1 %jx %jx", prev_object,
2309 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2310 			prev_object->charge -= ptoa(prev_object->size -
2311 			    next_pindex);
2312 		}
2313 #endif
2314 	}
2315 
2316 	/*
2317 	 * Extend the object if necessary.
2318 	 */
2319 	if (next_pindex + next_size > prev_object->size)
2320 		prev_object->size = next_pindex + next_size;
2321 
2322 	VM_OBJECT_WUNLOCK(prev_object);
2323 	return (TRUE);
2324 }
2325 
2326 void
2327 vm_object_set_writeable_dirty(vm_object_t object)
2328 {
2329 
2330 	/* Only set for vnodes & tmpfs */
2331 	if (object->type != OBJT_VNODE &&
2332 	    (object->flags & OBJ_TMPFS_NODE) == 0)
2333 		return;
2334 	atomic_add_int(&object->generation, 1);
2335 }
2336 
2337 /*
2338  *	vm_object_unwire:
2339  *
2340  *	For each page offset within the specified range of the given object,
2341  *	find the highest-level page in the shadow chain and unwire it.  A page
2342  *	must exist at every page offset, and the highest-level page must be
2343  *	wired.
2344  */
2345 void
2346 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2347     uint8_t queue)
2348 {
2349 	vm_object_t tobject, t1object;
2350 	vm_page_t m, tm;
2351 	vm_pindex_t end_pindex, pindex, tpindex;
2352 	int depth, locked_depth;
2353 
2354 	KASSERT((offset & PAGE_MASK) == 0,
2355 	    ("vm_object_unwire: offset is not page aligned"));
2356 	KASSERT((length & PAGE_MASK) == 0,
2357 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2358 	/* The wired count of a fictitious page never changes. */
2359 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2360 		return;
2361 	pindex = OFF_TO_IDX(offset);
2362 	end_pindex = pindex + atop(length);
2363 again:
2364 	locked_depth = 1;
2365 	VM_OBJECT_RLOCK(object);
2366 	m = vm_page_find_least(object, pindex);
2367 	while (pindex < end_pindex) {
2368 		if (m == NULL || pindex < m->pindex) {
2369 			/*
2370 			 * The first object in the shadow chain doesn't
2371 			 * contain a page at the current index.  Therefore,
2372 			 * the page must exist in a backing object.
2373 			 */
2374 			tobject = object;
2375 			tpindex = pindex;
2376 			depth = 0;
2377 			do {
2378 				tpindex +=
2379 				    OFF_TO_IDX(tobject->backing_object_offset);
2380 				tobject = tobject->backing_object;
2381 				KASSERT(tobject != NULL,
2382 				    ("vm_object_unwire: missing page"));
2383 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2384 					goto next_page;
2385 				depth++;
2386 				if (depth == locked_depth) {
2387 					locked_depth++;
2388 					VM_OBJECT_RLOCK(tobject);
2389 				}
2390 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2391 			    NULL);
2392 		} else {
2393 			tm = m;
2394 			m = TAILQ_NEXT(m, listq);
2395 		}
2396 		if (vm_page_trysbusy(tm) == 0) {
2397 			for (tobject = object; locked_depth >= 1;
2398 			    locked_depth--) {
2399 				t1object = tobject->backing_object;
2400 				if (tm->object != tobject)
2401 					VM_OBJECT_RUNLOCK(tobject);
2402 				tobject = t1object;
2403 			}
2404 			vm_page_busy_sleep(tm, "unwbo", true);
2405 			goto again;
2406 		}
2407 		vm_page_unwire(tm, queue);
2408 		vm_page_sunbusy(tm);
2409 next_page:
2410 		pindex++;
2411 	}
2412 	/* Release the accumulated object locks. */
2413 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2414 		t1object = tobject->backing_object;
2415 		VM_OBJECT_RUNLOCK(tobject);
2416 		tobject = t1object;
2417 	}
2418 }
2419 
2420 /*
2421  * Return the vnode for the given object, or NULL if none exists.
2422  * For tmpfs objects, the function may return NULL if there is
2423  * no vnode allocated at the time of the call.
2424  */
2425 struct vnode *
2426 vm_object_vnode(vm_object_t object)
2427 {
2428 	struct vnode *vp;
2429 
2430 	VM_OBJECT_ASSERT_LOCKED(object);
2431 	if (object->type == OBJT_VNODE) {
2432 		vp = object->handle;
2433 		KASSERT(vp != NULL, ("%s: OBJT_VNODE has no vnode", __func__));
2434 	} else if (object->type == OBJT_SWAP &&
2435 	    (object->flags & OBJ_TMPFS) != 0) {
2436 		vp = object->un_pager.swp.swp_tmpfs;
2437 		KASSERT(vp != NULL, ("%s: OBJT_TMPFS has no vnode", __func__));
2438 	} else {
2439 		vp = NULL;
2440 	}
2441 	return (vp);
2442 }
2443 
2444 
2445 /*
2446  * Busy the vm object.  This prevents new pages belonging to the object from
2447  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2448  * for checking page state before proceeding.
2449  */
2450 void
2451 vm_object_busy(vm_object_t obj)
2452 {
2453 
2454 	VM_OBJECT_ASSERT_LOCKED(obj);
2455 
2456 	blockcount_acquire(&obj->busy, 1);
2457 	/* The fence is required to order loads of page busy. */
2458 	atomic_thread_fence_acq_rel();
2459 }
2460 
2461 void
2462 vm_object_unbusy(vm_object_t obj)
2463 {
2464 
2465 	blockcount_release(&obj->busy, 1);
2466 }
2467 
2468 void
2469 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2470 {
2471 
2472 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2473 
2474 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2475 }
2476 
2477 /*
2478  * Return the kvme type of the given object.
2479  * If vpp is not NULL, set it to the object's vm_object_vnode() or NULL.
2480  */
2481 int
2482 vm_object_kvme_type(vm_object_t object, struct vnode **vpp)
2483 {
2484 
2485 	VM_OBJECT_ASSERT_LOCKED(object);
2486 	if (vpp != NULL)
2487 		*vpp = vm_object_vnode(object);
2488 	switch (object->type) {
2489 	case OBJT_DEFAULT:
2490 		return (KVME_TYPE_DEFAULT);
2491 	case OBJT_VNODE:
2492 		return (KVME_TYPE_VNODE);
2493 	case OBJT_SWAP:
2494 		if ((object->flags & OBJ_TMPFS_NODE) != 0)
2495 			return (KVME_TYPE_VNODE);
2496 		return (KVME_TYPE_SWAP);
2497 	case OBJT_DEVICE:
2498 		return (KVME_TYPE_DEVICE);
2499 	case OBJT_PHYS:
2500 		return (KVME_TYPE_PHYS);
2501 	case OBJT_DEAD:
2502 		return (KVME_TYPE_DEAD);
2503 	case OBJT_SG:
2504 		return (KVME_TYPE_SG);
2505 	case OBJT_MGTDEVICE:
2506 		return (KVME_TYPE_MGTDEVICE);
2507 	default:
2508 		return (KVME_TYPE_UNKNOWN);
2509 	}
2510 }
2511 
2512 static int
2513 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2514 {
2515 	struct kinfo_vmobject *kvo;
2516 	char *fullpath, *freepath;
2517 	struct vnode *vp;
2518 	struct vattr va;
2519 	vm_object_t obj;
2520 	vm_page_t m;
2521 	int count, error;
2522 
2523 	if (req->oldptr == NULL) {
2524 		/*
2525 		 * If an old buffer has not been provided, generate an
2526 		 * estimate of the space needed for a subsequent call.
2527 		 */
2528 		mtx_lock(&vm_object_list_mtx);
2529 		count = 0;
2530 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2531 			if (obj->type == OBJT_DEAD)
2532 				continue;
2533 			count++;
2534 		}
2535 		mtx_unlock(&vm_object_list_mtx);
2536 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2537 		    count * 11 / 10));
2538 	}
2539 
2540 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK);
2541 	error = 0;
2542 
2543 	/*
2544 	 * VM objects are type stable and are never removed from the
2545 	 * list once added.  This allows us to safely read obj->object_list
2546 	 * after reacquiring the VM object lock.
2547 	 */
2548 	mtx_lock(&vm_object_list_mtx);
2549 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2550 		if (obj->type == OBJT_DEAD)
2551 			continue;
2552 		VM_OBJECT_RLOCK(obj);
2553 		if (obj->type == OBJT_DEAD) {
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 = obj->shadow_count;
2562 		kvo->kvo_memattr = obj->memattr;
2563 		kvo->kvo_active = 0;
2564 		kvo->kvo_inactive = 0;
2565 		TAILQ_FOREACH(m, &obj->memq, listq) {
2566 			/*
2567 			 * A page may belong to the object but be
2568 			 * dequeued and set to PQ_NONE while the
2569 			 * object lock is not held.  This makes the
2570 			 * reads of m->queue below racy, and we do not
2571 			 * count pages set to PQ_NONE.  However, this
2572 			 * sysctl is only meant to give an
2573 			 * approximation of the system anyway.
2574 			 */
2575 			if (m->a.queue == PQ_ACTIVE)
2576 				kvo->kvo_active++;
2577 			else if (m->a.queue == PQ_INACTIVE)
2578 				kvo->kvo_inactive++;
2579 		}
2580 
2581 		kvo->kvo_vn_fileid = 0;
2582 		kvo->kvo_vn_fsid = 0;
2583 		kvo->kvo_vn_fsid_freebsd11 = 0;
2584 		freepath = NULL;
2585 		fullpath = "";
2586 		kvo->kvo_type = vm_object_kvme_type(obj, &vp);
2587 		if (vp != NULL)
2588 			vref(vp);
2589 		VM_OBJECT_RUNLOCK(obj);
2590 		if (vp != NULL) {
2591 			vn_fullpath(curthread, vp, &fullpath, &freepath);
2592 			vn_lock(vp, LK_SHARED | LK_RETRY);
2593 			if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2594 				kvo->kvo_vn_fileid = va.va_fileid;
2595 				kvo->kvo_vn_fsid = va.va_fsid;
2596 				kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2597 								/* truncate */
2598 			}
2599 			vput(vp);
2600 		}
2601 
2602 		strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2603 		if (freepath != NULL)
2604 			free(freepath, M_TEMP);
2605 
2606 		/* Pack record size down */
2607 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2608 		    + strlen(kvo->kvo_path) + 1;
2609 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2610 		    sizeof(uint64_t));
2611 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2612 		mtx_lock(&vm_object_list_mtx);
2613 		if (error)
2614 			break;
2615 	}
2616 	mtx_unlock(&vm_object_list_mtx);
2617 	free(kvo, M_TEMP);
2618 	return (error);
2619 }
2620 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2621     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2622     "List of VM objects");
2623 
2624 #include "opt_ddb.h"
2625 #ifdef DDB
2626 #include <sys/kernel.h>
2627 
2628 #include <sys/cons.h>
2629 
2630 #include <ddb/ddb.h>
2631 
2632 static int
2633 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2634 {
2635 	vm_map_t tmpm;
2636 	vm_map_entry_t tmpe;
2637 	vm_object_t obj;
2638 
2639 	if (map == 0)
2640 		return 0;
2641 
2642 	if (entry == 0) {
2643 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2644 			if (_vm_object_in_map(map, object, tmpe)) {
2645 				return 1;
2646 			}
2647 		}
2648 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2649 		tmpm = entry->object.sub_map;
2650 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2651 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2652 				return 1;
2653 			}
2654 		}
2655 	} else if ((obj = entry->object.vm_object) != NULL) {
2656 		for (; obj; obj = obj->backing_object)
2657 			if (obj == object) {
2658 				return 1;
2659 			}
2660 	}
2661 	return 0;
2662 }
2663 
2664 static int
2665 vm_object_in_map(vm_object_t object)
2666 {
2667 	struct proc *p;
2668 
2669 	/* sx_slock(&allproc_lock); */
2670 	FOREACH_PROC_IN_SYSTEM(p) {
2671 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2672 			continue;
2673 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2674 			/* sx_sunlock(&allproc_lock); */
2675 			return 1;
2676 		}
2677 	}
2678 	/* sx_sunlock(&allproc_lock); */
2679 	if (_vm_object_in_map(kernel_map, object, 0))
2680 		return 1;
2681 	return 0;
2682 }
2683 
2684 DB_SHOW_COMMAND(vmochk, vm_object_check)
2685 {
2686 	vm_object_t object;
2687 
2688 	/*
2689 	 * make sure that internal objs are in a map somewhere
2690 	 * and none have zero ref counts.
2691 	 */
2692 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2693 		if ((object->flags & OBJ_ANON) != 0) {
2694 			if (object->ref_count == 0) {
2695 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2696 					(long)object->size);
2697 			}
2698 			if (!vm_object_in_map(object)) {
2699 				db_printf(
2700 			"vmochk: internal obj is not in a map: "
2701 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2702 				    object->ref_count, (u_long)object->size,
2703 				    (u_long)object->size,
2704 				    (void *)object->backing_object);
2705 			}
2706 		}
2707 	}
2708 }
2709 
2710 /*
2711  *	vm_object_print:	[ debug ]
2712  */
2713 DB_SHOW_COMMAND(object, vm_object_print_static)
2714 {
2715 	/* XXX convert args. */
2716 	vm_object_t object = (vm_object_t)addr;
2717 	boolean_t full = have_addr;
2718 
2719 	vm_page_t p;
2720 
2721 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2722 #define	count	was_count
2723 
2724 	int count;
2725 
2726 	if (object == NULL)
2727 		return;
2728 
2729 	db_iprintf(
2730 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2731 	    object, (int)object->type, (uintmax_t)object->size,
2732 	    object->resident_page_count, object->ref_count, object->flags,
2733 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2734 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2735 	    object->shadow_count,
2736 	    object->backing_object ? object->backing_object->ref_count : 0,
2737 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2738 
2739 	if (!full)
2740 		return;
2741 
2742 	db_indent += 2;
2743 	count = 0;
2744 	TAILQ_FOREACH(p, &object->memq, listq) {
2745 		if (count == 0)
2746 			db_iprintf("memory:=");
2747 		else if (count == 6) {
2748 			db_printf("\n");
2749 			db_iprintf(" ...");
2750 			count = 0;
2751 		} else
2752 			db_printf(",");
2753 		count++;
2754 
2755 		db_printf("(off=0x%jx,page=0x%jx)",
2756 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2757 	}
2758 	if (count != 0)
2759 		db_printf("\n");
2760 	db_indent -= 2;
2761 }
2762 
2763 /* XXX. */
2764 #undef count
2765 
2766 /* XXX need this non-static entry for calling from vm_map_print. */
2767 void
2768 vm_object_print(
2769         /* db_expr_t */ long addr,
2770 	boolean_t have_addr,
2771 	/* db_expr_t */ long count,
2772 	char *modif)
2773 {
2774 	vm_object_print_static(addr, have_addr, count, modif);
2775 }
2776 
2777 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2778 {
2779 	vm_object_t object;
2780 	vm_pindex_t fidx;
2781 	vm_paddr_t pa;
2782 	vm_page_t m, prev_m;
2783 	int rcount, nl, c;
2784 
2785 	nl = 0;
2786 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2787 		db_printf("new object: %p\n", (void *)object);
2788 		if (nl > 18) {
2789 			c = cngetc();
2790 			if (c != ' ')
2791 				return;
2792 			nl = 0;
2793 		}
2794 		nl++;
2795 		rcount = 0;
2796 		fidx = 0;
2797 		pa = -1;
2798 		TAILQ_FOREACH(m, &object->memq, listq) {
2799 			if (m->pindex > 128)
2800 				break;
2801 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2802 			    prev_m->pindex + 1 != m->pindex) {
2803 				if (rcount) {
2804 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2805 						(long)fidx, rcount, (long)pa);
2806 					if (nl > 18) {
2807 						c = cngetc();
2808 						if (c != ' ')
2809 							return;
2810 						nl = 0;
2811 					}
2812 					nl++;
2813 					rcount = 0;
2814 				}
2815 			}
2816 			if (rcount &&
2817 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2818 				++rcount;
2819 				continue;
2820 			}
2821 			if (rcount) {
2822 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2823 					(long)fidx, rcount, (long)pa);
2824 				if (nl > 18) {
2825 					c = cngetc();
2826 					if (c != ' ')
2827 						return;
2828 					nl = 0;
2829 				}
2830 				nl++;
2831 			}
2832 			fidx = m->pindex;
2833 			pa = VM_PAGE_TO_PHYS(m);
2834 			rcount = 1;
2835 		}
2836 		if (rcount) {
2837 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2838 				(long)fidx, rcount, (long)pa);
2839 			if (nl > 18) {
2840 				c = cngetc();
2841 				if (c != ' ')
2842 					return;
2843 				nl = 0;
2844 			}
2845 			nl++;
2846 		}
2847 	}
2848 }
2849 #endif /* DDB */
2850