xref: /freebsd/sys/vm/vm_object.c (revision 1bb8b1d7e190d75597d31ab87364a058e8ef8c5b)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39  *
40  * Permission to use, copy, modify and distribute this software and
41  * its documentation is hereby granted, provided that both the copyright
42  * notice and this permission notice appear in all copies of the
43  * software, derivative works or modified versions, and any portions
44  * thereof, and that both notices appear in supporting documentation.
45  *
46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49  *
50  * Carnegie Mellon requests users of this software to return to
51  *
52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
53  *  School of Computer Science
54  *  Carnegie Mellon University
55  *  Pittsburgh PA 15213-3890
56  *
57  * any improvements or extensions that they make and grant Carnegie the
58  * rights to redistribute these changes.
59  */
60 
61 /*
62  *	Virtual memory object module.
63  */
64 
65 #include "opt_vm.h"
66 
67 #include <sys/systm.h>
68 #include <sys/blockcount.h>
69 #include <sys/conf.h>
70 #include <sys/cpuset.h>
71 #include <sys/ipc.h>
72 #include <sys/jail.h>
73 #include <sys/limits.h>
74 #include <sys/lock.h>
75 #include <sys/mman.h>
76 #include <sys/mount.h>
77 #include <sys/kernel.h>
78 #include <sys/mutex.h>
79 #include <sys/pctrie.h>
80 #include <sys/proc.h>
81 #include <sys/refcount.h>
82 #include <sys/shm.h>
83 #include <sys/sx.h>
84 #include <sys/sysctl.h>
85 #include <sys/resourcevar.h>
86 #include <sys/refcount.h>
87 #include <sys/rwlock.h>
88 #include <sys/user.h>
89 #include <sys/vnode.h>
90 #include <sys/vmmeter.h>
91 
92 #include <vm/vm.h>
93 #include <vm/vm_param.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_pager.h>
100 #include <vm/vm_phys.h>
101 #include <vm/vm_pagequeue.h>
102 #include <vm/swap_pager.h>
103 #include <vm/vm_kern.h>
104 #include <vm/vm_extern.h>
105 #include <vm/vm_radix.h>
106 #include <vm/vm_reserv.h>
107 #include <vm/uma.h>
108 
109 static int old_msync;
110 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
111     "Use old (insecure) msync behavior");
112 
113 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
114 		    int pagerflags, int flags, boolean_t *allclean,
115 		    boolean_t *eio);
116 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
117 		    boolean_t *allclean);
118 static void	vm_object_backing_remove(vm_object_t object);
119 
120 /*
121  *	Virtual memory objects maintain the actual data
122  *	associated with allocated virtual memory.  A given
123  *	page of memory exists within exactly one object.
124  *
125  *	An object is only deallocated when all "references"
126  *	are given up.  Only one "reference" to a given
127  *	region of an object should be writeable.
128  *
129  *	Associated with each object is a list of all resident
130  *	memory pages belonging to that object; this list is
131  *	maintained by the "vm_page" module, and locked by the object's
132  *	lock.
133  *
134  *	Each object also records a "pager" routine which is
135  *	used to retrieve (and store) pages to the proper backing
136  *	storage.  In addition, objects may be backed by other
137  *	objects from which they were virtual-copied.
138  *
139  *	The only items within the object structure which are
140  *	modified after time of creation are:
141  *		reference count		locked by object's lock
142  *		pager routine		locked by object's lock
143  *
144  */
145 
146 struct object_q vm_object_list;
147 struct mtx vm_object_list_mtx;	/* lock for object list and count */
148 
149 struct vm_object kernel_object_store;
150 
151 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
152     "VM object stats");
153 
154 static COUNTER_U64_DEFINE_EARLY(object_collapses);
155 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
156     &object_collapses,
157     "VM object collapses");
158 
159 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
160 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
161     &object_bypasses,
162     "VM object bypasses");
163 
164 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
165 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
166     &object_collapse_waits,
167     "Number of sleeps for collapse");
168 
169 static uma_zone_t obj_zone;
170 
171 static int vm_object_zinit(void *mem, int size, int flags);
172 
173 #ifdef INVARIANTS
174 static void vm_object_zdtor(void *mem, int size, void *arg);
175 
176 static void
177 vm_object_zdtor(void *mem, int size, void *arg)
178 {
179 	vm_object_t object;
180 
181 	object = (vm_object_t)mem;
182 	KASSERT(object->ref_count == 0,
183 	    ("object %p ref_count = %d", object, object->ref_count));
184 	KASSERT(TAILQ_EMPTY(&object->memq),
185 	    ("object %p has resident pages in its memq", object));
186 	KASSERT(vm_radix_is_empty(&object->rtree),
187 	    ("object %p has resident pages in its trie", object));
188 #if VM_NRESERVLEVEL > 0
189 	KASSERT(LIST_EMPTY(&object->rvq),
190 	    ("object %p has reservations",
191 	    object));
192 #endif
193 	KASSERT(!vm_object_busied(object),
194 	    ("object %p busy = %d", object, blockcount_read(&object->busy)));
195 	KASSERT(object->resident_page_count == 0,
196 	    ("object %p resident_page_count = %d",
197 	    object, object->resident_page_count));
198 	KASSERT(atomic_load_int(&object->shadow_count) == 0,
199 	    ("object %p shadow_count = %d",
200 	    object, atomic_load_int(&object->shadow_count)));
201 	KASSERT(object->type == OBJT_DEAD,
202 	    ("object %p has non-dead type %d",
203 	    object, object->type));
204 	KASSERT(object->charge == 0 && object->cred == NULL,
205 	    ("object %p has non-zero charge %ju (%p)",
206 	    object, (uintmax_t)object->charge, object->cred));
207 }
208 #endif
209 
210 static int
211 vm_object_zinit(void *mem, int size, int flags)
212 {
213 	vm_object_t object;
214 
215 	object = (vm_object_t)mem;
216 	rw_init_flags(&object->lock, "vmobject", RW_DUPOK | RW_NEW);
217 
218 	/* These are true for any object that has been freed */
219 	object->type = OBJT_DEAD;
220 	vm_radix_init(&object->rtree);
221 	refcount_init(&object->ref_count, 0);
222 	blockcount_init(&object->paging_in_progress);
223 	blockcount_init(&object->busy);
224 	object->resident_page_count = 0;
225 	atomic_store_int(&object->shadow_count, 0);
226 	object->flags = OBJ_DEAD;
227 
228 	mtx_lock(&vm_object_list_mtx);
229 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
230 	mtx_unlock(&vm_object_list_mtx);
231 	return (0);
232 }
233 
234 static void
235 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
236     vm_object_t object, void *handle)
237 {
238 
239 	TAILQ_INIT(&object->memq);
240 	LIST_INIT(&object->shadow_head);
241 
242 	object->type = type;
243 	object->flags = flags;
244 	if ((flags & OBJ_SWAP) != 0) {
245 		pctrie_init(&object->un_pager.swp.swp_blks);
246 		object->un_pager.swp.writemappings = 0;
247 	}
248 
249 	/*
250 	 * Ensure that swap_pager_swapoff() iteration over object_list
251 	 * sees up to date type and pctrie head if it observed
252 	 * non-dead object.
253 	 */
254 	atomic_thread_fence_rel();
255 
256 	object->pg_color = 0;
257 	object->size = size;
258 	object->domain.dr_policy = NULL;
259 	object->generation = 1;
260 	object->cleangeneration = 1;
261 	refcount_init(&object->ref_count, 1);
262 	object->memattr = VM_MEMATTR_DEFAULT;
263 	object->cred = NULL;
264 	object->charge = 0;
265 	object->handle = handle;
266 	object->backing_object = NULL;
267 	object->backing_object_offset = (vm_ooffset_t) 0;
268 #if VM_NRESERVLEVEL > 0
269 	LIST_INIT(&object->rvq);
270 #endif
271 	umtx_shm_object_init(object);
272 }
273 
274 /*
275  *	vm_object_init:
276  *
277  *	Initialize the VM objects module.
278  */
279 void
280 vm_object_init(void)
281 {
282 	TAILQ_INIT(&vm_object_list);
283 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
284 
285 	rw_init(&kernel_object->lock, "kernel vm object");
286 	vm_radix_init(&kernel_object->rtree);
287 	_vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
288 	    VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
289 #if VM_NRESERVLEVEL > 0
290 	kernel_object->flags |= OBJ_COLORED;
291 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
292 #endif
293 	kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
294 
295 	/*
296 	 * The lock portion of struct vm_object must be type stable due
297 	 * to vm_pageout_fallback_object_lock locking a vm object
298 	 * without holding any references to it.
299 	 *
300 	 * paging_in_progress is valid always.  Lockless references to
301 	 * the objects may acquire pip and then check OBJ_DEAD.
302 	 */
303 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
304 #ifdef INVARIANTS
305 	    vm_object_zdtor,
306 #else
307 	    NULL,
308 #endif
309 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
310 
311 	vm_radix_zinit();
312 }
313 
314 void
315 vm_object_clear_flag(vm_object_t object, u_short bits)
316 {
317 
318 	VM_OBJECT_ASSERT_WLOCKED(object);
319 	object->flags &= ~bits;
320 }
321 
322 /*
323  *	Sets the default memory attribute for the specified object.  Pages
324  *	that are allocated to this object are by default assigned this memory
325  *	attribute.
326  *
327  *	Presently, this function must be called before any pages are allocated
328  *	to the object.  In the future, this requirement may be relaxed for
329  *	"default" and "swap" objects.
330  */
331 int
332 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
333 {
334 
335 	VM_OBJECT_ASSERT_WLOCKED(object);
336 
337 	if (object->type == OBJT_DEAD)
338 		return (KERN_INVALID_ARGUMENT);
339 	if (!TAILQ_EMPTY(&object->memq))
340 		return (KERN_FAILURE);
341 
342 	object->memattr = memattr;
343 	return (KERN_SUCCESS);
344 }
345 
346 void
347 vm_object_pip_add(vm_object_t object, short i)
348 {
349 
350 	if (i > 0)
351 		blockcount_acquire(&object->paging_in_progress, i);
352 }
353 
354 void
355 vm_object_pip_wakeup(vm_object_t object)
356 {
357 
358 	vm_object_pip_wakeupn(object, 1);
359 }
360 
361 void
362 vm_object_pip_wakeupn(vm_object_t object, short i)
363 {
364 
365 	if (i > 0)
366 		blockcount_release(&object->paging_in_progress, i);
367 }
368 
369 /*
370  * Atomically drop the object lock and wait for pip to drain.  This protects
371  * from sleep/wakeup races due to identity changes.  The lock is not re-acquired
372  * on return.
373  */
374 static void
375 vm_object_pip_sleep(vm_object_t object, const char *waitid)
376 {
377 
378 	(void)blockcount_sleep(&object->paging_in_progress, &object->lock,
379 	    waitid, PVM | PDROP);
380 }
381 
382 void
383 vm_object_pip_wait(vm_object_t object, const char *waitid)
384 {
385 
386 	VM_OBJECT_ASSERT_WLOCKED(object);
387 
388 	blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
389 	    PVM);
390 }
391 
392 void
393 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
394 {
395 
396 	VM_OBJECT_ASSERT_UNLOCKED(object);
397 
398 	blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
399 }
400 
401 /*
402  *	vm_object_allocate:
403  *
404  *	Returns a new object with the given size.
405  */
406 vm_object_t
407 vm_object_allocate(objtype_t type, vm_pindex_t size)
408 {
409 	vm_object_t object;
410 	u_short flags;
411 
412 	switch (type) {
413 	case OBJT_DEAD:
414 		panic("vm_object_allocate: can't create OBJT_DEAD");
415 	case OBJT_SWAP:
416 		flags = OBJ_COLORED | OBJ_SWAP;
417 		break;
418 	case OBJT_DEVICE:
419 	case OBJT_SG:
420 		flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
421 		break;
422 	case OBJT_MGTDEVICE:
423 		flags = OBJ_FICTITIOUS;
424 		break;
425 	case OBJT_PHYS:
426 		flags = OBJ_UNMANAGED;
427 		break;
428 	case OBJT_VNODE:
429 		flags = 0;
430 		break;
431 	default:
432 		panic("vm_object_allocate: type %d is undefined or dynamic",
433 		    type);
434 	}
435 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
436 	_vm_object_allocate(type, size, flags, object, NULL);
437 
438 	return (object);
439 }
440 
441 vm_object_t
442 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags)
443 {
444 	vm_object_t object;
445 
446 	MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */);
447 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
448 	_vm_object_allocate(dyntype, size, flags, object, NULL);
449 
450 	return (object);
451 }
452 
453 /*
454  *	vm_object_allocate_anon:
455  *
456  *	Returns a new default object of the given size and marked as
457  *	anonymous memory for special split/collapse handling.  Color
458  *	to be initialized by the caller.
459  */
460 vm_object_t
461 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
462     struct ucred *cred, vm_size_t charge)
463 {
464 	vm_object_t handle, object;
465 
466 	if (backing_object == NULL)
467 		handle = NULL;
468 	else if ((backing_object->flags & OBJ_ANON) != 0)
469 		handle = backing_object->handle;
470 	else
471 		handle = backing_object;
472 	object = uma_zalloc(obj_zone, M_WAITOK);
473 	_vm_object_allocate(OBJT_SWAP, size,
474 	    OBJ_ANON | OBJ_ONEMAPPING | OBJ_SWAP, object, handle);
475 	object->cred = cred;
476 	object->charge = cred != NULL ? charge : 0;
477 	return (object);
478 }
479 
480 static void
481 vm_object_reference_vnode(vm_object_t object)
482 {
483 	u_int old;
484 
485 	/*
486 	 * vnode objects need the lock for the first reference
487 	 * to serialize with vnode_object_deallocate().
488 	 */
489 	if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
490 		VM_OBJECT_RLOCK(object);
491 		old = refcount_acquire(&object->ref_count);
492 		if (object->type == OBJT_VNODE && old == 0)
493 			vref(object->handle);
494 		VM_OBJECT_RUNLOCK(object);
495 	}
496 }
497 
498 /*
499  *	vm_object_reference:
500  *
501  *	Acquires a reference to the given object.
502  */
503 void
504 vm_object_reference(vm_object_t object)
505 {
506 
507 	if (object == NULL)
508 		return;
509 
510 	if (object->type == OBJT_VNODE)
511 		vm_object_reference_vnode(object);
512 	else
513 		refcount_acquire(&object->ref_count);
514 	KASSERT((object->flags & OBJ_DEAD) == 0,
515 	    ("vm_object_reference: Referenced dead object."));
516 }
517 
518 /*
519  *	vm_object_reference_locked:
520  *
521  *	Gets another reference to the given object.
522  *
523  *	The object must be locked.
524  */
525 void
526 vm_object_reference_locked(vm_object_t object)
527 {
528 	u_int old;
529 
530 	VM_OBJECT_ASSERT_LOCKED(object);
531 	old = refcount_acquire(&object->ref_count);
532 	if (object->type == OBJT_VNODE && old == 0)
533 		vref(object->handle);
534 	KASSERT((object->flags & OBJ_DEAD) == 0,
535 	    ("vm_object_reference: Referenced dead object."));
536 }
537 
538 /*
539  * Handle deallocating an object of type OBJT_VNODE.
540  */
541 static void
542 vm_object_deallocate_vnode(vm_object_t object)
543 {
544 	struct vnode *vp = (struct vnode *) object->handle;
545 	bool last;
546 
547 	KASSERT(object->type == OBJT_VNODE,
548 	    ("vm_object_deallocate_vnode: not a vnode object"));
549 	KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
550 
551 	/* Object lock to protect handle lookup. */
552 	last = refcount_release(&object->ref_count);
553 	VM_OBJECT_RUNLOCK(object);
554 
555 	if (!last)
556 		return;
557 
558 	if (!umtx_shm_vnobj_persistent)
559 		umtx_shm_object_terminated(object);
560 
561 	/* vrele may need the vnode lock. */
562 	vrele(vp);
563 }
564 
565 /*
566  * We dropped a reference on an object and discovered that it had a
567  * single remaining shadow.  This is a sibling of the reference we
568  * dropped.  Attempt to collapse the sibling and backing object.
569  */
570 static vm_object_t
571 vm_object_deallocate_anon(vm_object_t backing_object)
572 {
573 	vm_object_t object;
574 
575 	/* Fetch the final shadow.  */
576 	object = LIST_FIRST(&backing_object->shadow_head);
577 	KASSERT(object != NULL &&
578 	    atomic_load_int(&backing_object->shadow_count) == 1,
579 	    ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
580 	    backing_object->ref_count,
581 	    atomic_load_int(&backing_object->shadow_count)));
582 	KASSERT((object->flags & OBJ_ANON) != 0,
583 	    ("invalid shadow object %p", object));
584 
585 	if (!VM_OBJECT_TRYWLOCK(object)) {
586 		/*
587 		 * Prevent object from disappearing since we do not have a
588 		 * reference.
589 		 */
590 		vm_object_pip_add(object, 1);
591 		VM_OBJECT_WUNLOCK(backing_object);
592 		VM_OBJECT_WLOCK(object);
593 		vm_object_pip_wakeup(object);
594 	} else
595 		VM_OBJECT_WUNLOCK(backing_object);
596 
597 	/*
598 	 * Check for a collapse/terminate race with the last reference holder.
599 	 */
600 	if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
601 	    !refcount_acquire_if_not_zero(&object->ref_count)) {
602 		VM_OBJECT_WUNLOCK(object);
603 		return (NULL);
604 	}
605 	backing_object = object->backing_object;
606 	if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
607 		vm_object_collapse(object);
608 	VM_OBJECT_WUNLOCK(object);
609 
610 	return (object);
611 }
612 
613 /*
614  *	vm_object_deallocate:
615  *
616  *	Release a reference to the specified object,
617  *	gained either through a vm_object_allocate
618  *	or a vm_object_reference call.  When all references
619  *	are gone, storage associated with this object
620  *	may be relinquished.
621  *
622  *	No object may be locked.
623  */
624 void
625 vm_object_deallocate(vm_object_t object)
626 {
627 	vm_object_t temp;
628 	bool released;
629 
630 	while (object != NULL) {
631 		/*
632 		 * If the reference count goes to 0 we start calling
633 		 * vm_object_terminate() on the object chain.  A ref count
634 		 * of 1 may be a special case depending on the shadow count
635 		 * being 0 or 1.  These cases require a write lock on the
636 		 * object.
637 		 */
638 		if ((object->flags & OBJ_ANON) == 0)
639 			released = refcount_release_if_gt(&object->ref_count, 1);
640 		else
641 			released = refcount_release_if_gt(&object->ref_count, 2);
642 		if (released)
643 			return;
644 
645 		if (object->type == OBJT_VNODE) {
646 			VM_OBJECT_RLOCK(object);
647 			if (object->type == OBJT_VNODE) {
648 				vm_object_deallocate_vnode(object);
649 				return;
650 			}
651 			VM_OBJECT_RUNLOCK(object);
652 		}
653 
654 		VM_OBJECT_WLOCK(object);
655 		KASSERT(object->ref_count > 0,
656 		    ("vm_object_deallocate: object deallocated too many times: %d",
657 		    object->type));
658 
659 		/*
660 		 * If this is not the final reference to an anonymous
661 		 * object we may need to collapse the shadow chain.
662 		 */
663 		if (!refcount_release(&object->ref_count)) {
664 			if (object->ref_count > 1 ||
665 			    atomic_load_int(&object->shadow_count) == 0) {
666 				if ((object->flags & OBJ_ANON) != 0 &&
667 				    object->ref_count == 1)
668 					vm_object_set_flag(object,
669 					    OBJ_ONEMAPPING);
670 				VM_OBJECT_WUNLOCK(object);
671 				return;
672 			}
673 
674 			/* Handle collapsing last ref on anonymous objects. */
675 			object = vm_object_deallocate_anon(object);
676 			continue;
677 		}
678 
679 		/*
680 		 * Handle the final reference to an object.  We restart
681 		 * the loop with the backing object to avoid recursion.
682 		 */
683 		umtx_shm_object_terminated(object);
684 		temp = object->backing_object;
685 		if (temp != NULL) {
686 			KASSERT(object->type == OBJT_SWAP,
687 			    ("shadowed tmpfs v_object 2 %p", object));
688 			vm_object_backing_remove(object);
689 		}
690 
691 		KASSERT((object->flags & OBJ_DEAD) == 0,
692 		    ("vm_object_deallocate: Terminating dead object."));
693 		vm_object_set_flag(object, OBJ_DEAD);
694 		vm_object_terminate(object);
695 		object = temp;
696 	}
697 }
698 
699 void
700 vm_object_destroy(vm_object_t object)
701 {
702 	uma_zfree(obj_zone, object);
703 }
704 
705 static void
706 vm_object_sub_shadow(vm_object_t object)
707 {
708 	KASSERT(object->shadow_count >= 1,
709 	    ("object %p sub_shadow count zero", object));
710 	atomic_subtract_int(&object->shadow_count, 1);
711 }
712 
713 static void
714 vm_object_backing_remove_locked(vm_object_t object)
715 {
716 	vm_object_t backing_object;
717 
718 	backing_object = object->backing_object;
719 	VM_OBJECT_ASSERT_WLOCKED(object);
720 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
721 
722 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
723 	    ("vm_object_backing_remove: Removing collapsing object."));
724 
725 	vm_object_sub_shadow(backing_object);
726 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
727 		LIST_REMOVE(object, shadow_list);
728 		vm_object_clear_flag(object, OBJ_SHADOWLIST);
729 	}
730 	object->backing_object = NULL;
731 }
732 
733 static void
734 vm_object_backing_remove(vm_object_t object)
735 {
736 	vm_object_t backing_object;
737 
738 	VM_OBJECT_ASSERT_WLOCKED(object);
739 
740 	backing_object = object->backing_object;
741 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
742 		VM_OBJECT_WLOCK(backing_object);
743 		vm_object_backing_remove_locked(object);
744 		VM_OBJECT_WUNLOCK(backing_object);
745 	} else {
746 		object->backing_object = NULL;
747 		vm_object_sub_shadow(backing_object);
748 	}
749 }
750 
751 static void
752 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
753 {
754 
755 	VM_OBJECT_ASSERT_WLOCKED(object);
756 
757 	atomic_add_int(&backing_object->shadow_count, 1);
758 	if ((backing_object->flags & OBJ_ANON) != 0) {
759 		VM_OBJECT_ASSERT_WLOCKED(backing_object);
760 		LIST_INSERT_HEAD(&backing_object->shadow_head, object,
761 		    shadow_list);
762 		vm_object_set_flag(object, OBJ_SHADOWLIST);
763 	}
764 	object->backing_object = backing_object;
765 }
766 
767 static void
768 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
769 {
770 
771 	VM_OBJECT_ASSERT_WLOCKED(object);
772 
773 	if ((backing_object->flags & OBJ_ANON) != 0) {
774 		VM_OBJECT_WLOCK(backing_object);
775 		vm_object_backing_insert_locked(object, backing_object);
776 		VM_OBJECT_WUNLOCK(backing_object);
777 	} else {
778 		object->backing_object = backing_object;
779 		atomic_add_int(&backing_object->shadow_count, 1);
780 	}
781 }
782 
783 /*
784  * Insert an object into a backing_object's shadow list with an additional
785  * reference to the backing_object added.
786  */
787 static void
788 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
789 {
790 
791 	VM_OBJECT_ASSERT_WLOCKED(object);
792 
793 	if ((backing_object->flags & OBJ_ANON) != 0) {
794 		VM_OBJECT_WLOCK(backing_object);
795 		KASSERT((backing_object->flags & OBJ_DEAD) == 0,
796 		    ("shadowing dead anonymous object"));
797 		vm_object_reference_locked(backing_object);
798 		vm_object_backing_insert_locked(object, backing_object);
799 		vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
800 		VM_OBJECT_WUNLOCK(backing_object);
801 	} else {
802 		vm_object_reference(backing_object);
803 		atomic_add_int(&backing_object->shadow_count, 1);
804 		object->backing_object = backing_object;
805 	}
806 }
807 
808 /*
809  * Transfer a backing reference from backing_object to object.
810  */
811 static void
812 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
813 {
814 	vm_object_t new_backing_object;
815 
816 	/*
817 	 * Note that the reference to backing_object->backing_object
818 	 * moves from within backing_object to within object.
819 	 */
820 	vm_object_backing_remove_locked(object);
821 	new_backing_object = backing_object->backing_object;
822 	if (new_backing_object == NULL)
823 		return;
824 	if ((new_backing_object->flags & OBJ_ANON) != 0) {
825 		VM_OBJECT_WLOCK(new_backing_object);
826 		vm_object_backing_remove_locked(backing_object);
827 		vm_object_backing_insert_locked(object, new_backing_object);
828 		VM_OBJECT_WUNLOCK(new_backing_object);
829 	} else {
830 		/*
831 		 * shadow_count for new_backing_object is left
832 		 * unchanged, its reference provided by backing_object
833 		 * is replaced by object.
834 		 */
835 		object->backing_object = new_backing_object;
836 		backing_object->backing_object = NULL;
837 	}
838 }
839 
840 /*
841  * Wait for a concurrent collapse to settle.
842  */
843 static void
844 vm_object_collapse_wait(vm_object_t object)
845 {
846 
847 	VM_OBJECT_ASSERT_WLOCKED(object);
848 
849 	while ((object->flags & OBJ_COLLAPSING) != 0) {
850 		vm_object_pip_wait(object, "vmcolwait");
851 		counter_u64_add(object_collapse_waits, 1);
852 	}
853 }
854 
855 /*
856  * Waits for a backing object to clear a pending collapse and returns
857  * it locked if it is an ANON object.
858  */
859 static vm_object_t
860 vm_object_backing_collapse_wait(vm_object_t object)
861 {
862 	vm_object_t backing_object;
863 
864 	VM_OBJECT_ASSERT_WLOCKED(object);
865 
866 	for (;;) {
867 		backing_object = object->backing_object;
868 		if (backing_object == NULL ||
869 		    (backing_object->flags & OBJ_ANON) == 0)
870 			return (NULL);
871 		VM_OBJECT_WLOCK(backing_object);
872 		if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
873 			break;
874 		VM_OBJECT_WUNLOCK(object);
875 		vm_object_pip_sleep(backing_object, "vmbckwait");
876 		counter_u64_add(object_collapse_waits, 1);
877 		VM_OBJECT_WLOCK(object);
878 	}
879 	return (backing_object);
880 }
881 
882 /*
883  *	vm_object_terminate_single_page removes a pageable page from the object,
884  *	and removes it from the paging queues and frees it, if it is not wired.
885  *	It is invoked via callback from vm_object_terminate_pages.
886  */
887 static void
888 vm_object_terminate_single_page(vm_page_t p, void *objectv)
889 {
890 	vm_object_t object __diagused = objectv;
891 
892 	vm_page_assert_unbusied(p);
893 	KASSERT(p->object == object &&
894 	    (p->ref_count & VPRC_OBJREF) != 0,
895 	    ("%s: page %p is inconsistent", __func__, p));
896 	p->object = NULL;
897 	if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
898 		KASSERT((object->flags & OBJ_UNMANAGED) != 0 ||
899 		    vm_page_astate_load(p).queue != PQ_NONE,
900 		    ("%s: page %p does not belong to a queue", __func__, p));
901 		VM_CNT_INC(v_pfree);
902 		vm_page_free(p);
903 	}
904 }
905 
906 /*
907  *	vm_object_terminate_pages removes any remaining pageable pages
908  *	from the object and resets the object to an empty state.
909  */
910 static void
911 vm_object_terminate_pages(vm_object_t object)
912 {
913 	VM_OBJECT_ASSERT_WLOCKED(object);
914 
915 	/*
916 	 * If the object contained any pages, then reset it to an empty state.
917 	 * Rather than incrementally removing each page from the object, the
918 	 * page and object are reset to any empty state.
919 	 */
920 	if (object->resident_page_count == 0)
921 		return;
922 
923 	vm_radix_reclaim_callback(&object->rtree,
924 	    vm_object_terminate_single_page, object);
925 	TAILQ_INIT(&object->memq);
926 	object->resident_page_count = 0;
927 	if (object->type == OBJT_VNODE)
928 		vdrop(object->handle);
929 }
930 
931 /*
932  *	vm_object_terminate actually destroys the specified object, freeing
933  *	up all previously used resources.
934  *
935  *	The object must be locked.
936  *	This routine may block.
937  */
938 void
939 vm_object_terminate(vm_object_t object)
940 {
941 
942 	VM_OBJECT_ASSERT_WLOCKED(object);
943 	KASSERT((object->flags & OBJ_DEAD) != 0,
944 	    ("terminating non-dead obj %p", object));
945 	KASSERT((object->flags & OBJ_COLLAPSING) == 0,
946 	    ("terminating collapsing obj %p", object));
947 	KASSERT(object->backing_object == NULL,
948 	    ("terminating shadow obj %p", object));
949 
950 	/*
951 	 * Wait for the pageout daemon and other current users to be
952 	 * done with the object.  Note that new paging_in_progress
953 	 * users can come after this wait, but they must check
954 	 * OBJ_DEAD flag set (without unlocking the object), and avoid
955 	 * the object being terminated.
956 	 */
957 	vm_object_pip_wait(object, "objtrm");
958 
959 	KASSERT(object->ref_count == 0,
960 	    ("vm_object_terminate: object with references, ref_count=%d",
961 	    object->ref_count));
962 
963 	if ((object->flags & OBJ_PG_DTOR) == 0)
964 		vm_object_terminate_pages(object);
965 
966 #if VM_NRESERVLEVEL > 0
967 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
968 		vm_reserv_break_all(object);
969 #endif
970 
971 	KASSERT(object->cred == NULL || (object->flags & OBJ_SWAP) != 0,
972 	    ("%s: non-swap obj %p has cred", __func__, object));
973 
974 	/*
975 	 * Let the pager know object is dead.
976 	 */
977 	vm_pager_deallocate(object);
978 	VM_OBJECT_WUNLOCK(object);
979 
980 	vm_object_destroy(object);
981 }
982 
983 /*
984  * Make the page read-only so that we can clear the object flags.  However, if
985  * this is a nosync mmap then the object is likely to stay dirty so do not
986  * mess with the page and do not clear the object flags.  Returns TRUE if the
987  * page should be flushed, and FALSE otherwise.
988  */
989 static boolean_t
990 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
991 {
992 
993 	vm_page_assert_busied(p);
994 
995 	/*
996 	 * If we have been asked to skip nosync pages and this is a
997 	 * nosync page, skip it.  Note that the object flags were not
998 	 * cleared in this case so we do not have to set them.
999 	 */
1000 	if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
1001 		*allclean = FALSE;
1002 		return (FALSE);
1003 	} else {
1004 		pmap_remove_write(p);
1005 		return (p->dirty != 0);
1006 	}
1007 }
1008 
1009 /*
1010  *	vm_object_page_clean
1011  *
1012  *	Clean all dirty pages in the specified range of object.  Leaves page
1013  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
1014  *	write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1015  *	leaving the object dirty.
1016  *
1017  *	For swap objects backing tmpfs regular files, do not flush anything,
1018  *	but remove write protection on the mapped pages to update mtime through
1019  *	mmaped writes.
1020  *
1021  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
1022  *	synchronous clustering mode implementation.
1023  *
1024  *	Odd semantics: if start == end, we clean everything.
1025  *
1026  *	The object must be locked.
1027  *
1028  *	Returns FALSE if some page from the range was not written, as
1029  *	reported by the pager, and TRUE otherwise.
1030  */
1031 boolean_t
1032 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1033     int flags)
1034 {
1035 	vm_page_t np, p;
1036 	vm_pindex_t pi, tend, tstart;
1037 	int curgeneration, n, pagerflags;
1038 	boolean_t eio, res, allclean;
1039 
1040 	VM_OBJECT_ASSERT_WLOCKED(object);
1041 
1042 	if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1043 		return (TRUE);
1044 
1045 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1046 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1047 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1048 
1049 	tstart = OFF_TO_IDX(start);
1050 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1051 	allclean = tstart == 0 && tend >= object->size;
1052 	res = TRUE;
1053 
1054 rescan:
1055 	curgeneration = object->generation;
1056 
1057 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1058 		pi = p->pindex;
1059 		if (pi >= tend)
1060 			break;
1061 		np = TAILQ_NEXT(p, listq);
1062 		if (vm_page_none_valid(p))
1063 			continue;
1064 		if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1065 			if (object->generation != curgeneration &&
1066 			    (flags & OBJPC_SYNC) != 0)
1067 				goto rescan;
1068 			np = vm_page_find_least(object, pi);
1069 			continue;
1070 		}
1071 		if (!vm_object_page_remove_write(p, flags, &allclean)) {
1072 			vm_page_xunbusy(p);
1073 			continue;
1074 		}
1075 		if (object->type == OBJT_VNODE) {
1076 			n = vm_object_page_collect_flush(object, p, pagerflags,
1077 			    flags, &allclean, &eio);
1078 			if (eio) {
1079 				res = FALSE;
1080 				allclean = FALSE;
1081 			}
1082 			if (object->generation != curgeneration &&
1083 			    (flags & OBJPC_SYNC) != 0)
1084 				goto rescan;
1085 
1086 			/*
1087 			 * If the VOP_PUTPAGES() did a truncated write, so
1088 			 * that even the first page of the run is not fully
1089 			 * written, vm_pageout_flush() returns 0 as the run
1090 			 * length.  Since the condition that caused truncated
1091 			 * write may be permanent, e.g. exhausted free space,
1092 			 * accepting n == 0 would cause an infinite loop.
1093 			 *
1094 			 * Forwarding the iterator leaves the unwritten page
1095 			 * behind, but there is not much we can do there if
1096 			 * filesystem refuses to write it.
1097 			 */
1098 			if (n == 0) {
1099 				n = 1;
1100 				allclean = FALSE;
1101 			}
1102 		} else {
1103 			n = 1;
1104 			vm_page_xunbusy(p);
1105 		}
1106 		np = vm_page_find_least(object, pi + n);
1107 	}
1108 #if 0
1109 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1110 #endif
1111 
1112 	/*
1113 	 * Leave updating cleangeneration for tmpfs objects to tmpfs
1114 	 * scan.  It needs to update mtime, which happens for other
1115 	 * filesystems during page writeouts.
1116 	 */
1117 	if (allclean && object->type == OBJT_VNODE)
1118 		object->cleangeneration = curgeneration;
1119 	return (res);
1120 }
1121 
1122 static int
1123 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1124     int flags, boolean_t *allclean, boolean_t *eio)
1125 {
1126 	vm_page_t ma[2 * vm_pageout_page_count - 1], tp;
1127 	int base, count, runlen;
1128 
1129 	vm_page_lock_assert(p, MA_NOTOWNED);
1130 	vm_page_assert_xbusied(p);
1131 	VM_OBJECT_ASSERT_WLOCKED(object);
1132 	base = nitems(ma) / 2;
1133 	ma[base] = p;
1134 	for (count = 1, 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 		ma[base + count] = tp;
1143 	}
1144 
1145 	for (tp = p; count < vm_pageout_page_count; count++) {
1146 		tp = vm_page_prev(tp);
1147 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1148 			break;
1149 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1150 			vm_page_xunbusy(tp);
1151 			break;
1152 		}
1153 		ma[--base] = tp;
1154 	}
1155 
1156 	vm_pageout_flush(&ma[base], count, pagerflags, nitems(ma) / 2 - base,
1157 	    &runlen, eio);
1158 	return (runlen);
1159 }
1160 
1161 /*
1162  * Note that there is absolutely no sense in writing out
1163  * anonymous objects, so we track down the vnode object
1164  * to write out.
1165  * We invalidate (remove) all pages from the address space
1166  * for semantic correctness.
1167  *
1168  * If the backing object is a device object with unmanaged pages, then any
1169  * mappings to the specified range of pages must be removed before this
1170  * function is called.
1171  *
1172  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1173  * may start out with a NULL object.
1174  */
1175 boolean_t
1176 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1177     boolean_t syncio, boolean_t invalidate)
1178 {
1179 	vm_object_t backing_object;
1180 	struct vnode *vp;
1181 	struct mount *mp;
1182 	int error, flags, fsync_after;
1183 	boolean_t res;
1184 
1185 	if (object == NULL)
1186 		return (TRUE);
1187 	res = TRUE;
1188 	error = 0;
1189 	VM_OBJECT_WLOCK(object);
1190 	while ((backing_object = object->backing_object) != NULL) {
1191 		VM_OBJECT_WLOCK(backing_object);
1192 		offset += object->backing_object_offset;
1193 		VM_OBJECT_WUNLOCK(object);
1194 		object = backing_object;
1195 		if (object->size < OFF_TO_IDX(offset + size))
1196 			size = IDX_TO_OFF(object->size) - offset;
1197 	}
1198 	/*
1199 	 * Flush pages if writing is allowed, invalidate them
1200 	 * if invalidation requested.  Pages undergoing I/O
1201 	 * will be ignored by vm_object_page_remove().
1202 	 *
1203 	 * We cannot lock the vnode and then wait for paging
1204 	 * to complete without deadlocking against vm_fault.
1205 	 * Instead we simply call vm_object_page_remove() and
1206 	 * allow it to block internally on a page-by-page
1207 	 * basis when it encounters pages undergoing async
1208 	 * I/O.
1209 	 */
1210 	if (object->type == OBJT_VNODE &&
1211 	    vm_object_mightbedirty(object) != 0 &&
1212 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1213 		VM_OBJECT_WUNLOCK(object);
1214 		(void)vn_start_write(vp, &mp, V_WAIT);
1215 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1216 		if (syncio && !invalidate && offset == 0 &&
1217 		    atop(size) == object->size) {
1218 			/*
1219 			 * If syncing the whole mapping of the file,
1220 			 * it is faster to schedule all the writes in
1221 			 * async mode, also allowing the clustering,
1222 			 * and then wait for i/o to complete.
1223 			 */
1224 			flags = 0;
1225 			fsync_after = TRUE;
1226 		} else {
1227 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1228 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1229 			fsync_after = FALSE;
1230 		}
1231 		VM_OBJECT_WLOCK(object);
1232 		res = vm_object_page_clean(object, offset, offset + size,
1233 		    flags);
1234 		VM_OBJECT_WUNLOCK(object);
1235 		if (fsync_after) {
1236 			for (;;) {
1237 				error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1238 				if (error != ERELOOKUP)
1239 					break;
1240 
1241 				/*
1242 				 * Allow SU/bufdaemon to handle more
1243 				 * dependencies in the meantime.
1244 				 */
1245 				VOP_UNLOCK(vp);
1246 				vn_finished_write(mp);
1247 
1248 				(void)vn_start_write(vp, &mp, V_WAIT);
1249 				vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1250 			}
1251 		}
1252 		VOP_UNLOCK(vp);
1253 		vn_finished_write(mp);
1254 		if (error != 0)
1255 			res = FALSE;
1256 		VM_OBJECT_WLOCK(object);
1257 	}
1258 	if ((object->type == OBJT_VNODE ||
1259 	     object->type == OBJT_DEVICE) && invalidate) {
1260 		if (object->type == OBJT_DEVICE)
1261 			/*
1262 			 * The option OBJPR_NOTMAPPED must be passed here
1263 			 * because vm_object_page_remove() cannot remove
1264 			 * unmanaged mappings.
1265 			 */
1266 			flags = OBJPR_NOTMAPPED;
1267 		else if (old_msync)
1268 			flags = 0;
1269 		else
1270 			flags = OBJPR_CLEANONLY;
1271 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1272 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1273 	}
1274 	VM_OBJECT_WUNLOCK(object);
1275 	return (res);
1276 }
1277 
1278 /*
1279  * Determine whether the given advice can be applied to the object.  Advice is
1280  * not applied to unmanaged pages since they never belong to page queues, and
1281  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1282  * have been mapped at most once.
1283  */
1284 static bool
1285 vm_object_advice_applies(vm_object_t object, int advice)
1286 {
1287 
1288 	if ((object->flags & OBJ_UNMANAGED) != 0)
1289 		return (false);
1290 	if (advice != MADV_FREE)
1291 		return (true);
1292 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1293 	    (OBJ_ONEMAPPING | OBJ_ANON));
1294 }
1295 
1296 static void
1297 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1298     vm_size_t size)
1299 {
1300 
1301 	if (advice == MADV_FREE)
1302 		vm_pager_freespace(object, pindex, size);
1303 }
1304 
1305 /*
1306  *	vm_object_madvise:
1307  *
1308  *	Implements the madvise function at the object/page level.
1309  *
1310  *	MADV_WILLNEED	(any object)
1311  *
1312  *	    Activate the specified pages if they are resident.
1313  *
1314  *	MADV_DONTNEED	(any object)
1315  *
1316  *	    Deactivate the specified pages if they are resident.
1317  *
1318  *	MADV_FREE	(OBJT_SWAP objects, OBJ_ONEMAPPING only)
1319  *
1320  *	    Deactivate and clean the specified pages if they are
1321  *	    resident.  This permits the process to reuse the pages
1322  *	    without faulting or the kernel to reclaim the pages
1323  *	    without I/O.
1324  */
1325 void
1326 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1327     int advice)
1328 {
1329 	vm_pindex_t tpindex;
1330 	vm_object_t backing_object, tobject;
1331 	vm_page_t m, tm;
1332 
1333 	if (object == NULL)
1334 		return;
1335 
1336 relookup:
1337 	VM_OBJECT_WLOCK(object);
1338 	if (!vm_object_advice_applies(object, advice)) {
1339 		VM_OBJECT_WUNLOCK(object);
1340 		return;
1341 	}
1342 	for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1343 		tobject = object;
1344 
1345 		/*
1346 		 * If the next page isn't resident in the top-level object, we
1347 		 * need to search the shadow chain.  When applying MADV_FREE, we
1348 		 * take care to release any swap space used to store
1349 		 * non-resident pages.
1350 		 */
1351 		if (m == NULL || pindex < m->pindex) {
1352 			/*
1353 			 * Optimize a common case: if the top-level object has
1354 			 * no backing object, we can skip over the non-resident
1355 			 * range in constant time.
1356 			 */
1357 			if (object->backing_object == NULL) {
1358 				tpindex = (m != NULL && m->pindex < end) ?
1359 				    m->pindex : end;
1360 				vm_object_madvise_freespace(object, advice,
1361 				    pindex, tpindex - pindex);
1362 				if ((pindex = tpindex) == end)
1363 					break;
1364 				goto next_page;
1365 			}
1366 
1367 			tpindex = pindex;
1368 			do {
1369 				vm_object_madvise_freespace(tobject, advice,
1370 				    tpindex, 1);
1371 				/*
1372 				 * Prepare to search the next object in the
1373 				 * chain.
1374 				 */
1375 				backing_object = tobject->backing_object;
1376 				if (backing_object == NULL)
1377 					goto next_pindex;
1378 				VM_OBJECT_WLOCK(backing_object);
1379 				tpindex +=
1380 				    OFF_TO_IDX(tobject->backing_object_offset);
1381 				if (tobject != object)
1382 					VM_OBJECT_WUNLOCK(tobject);
1383 				tobject = backing_object;
1384 				if (!vm_object_advice_applies(tobject, advice))
1385 					goto next_pindex;
1386 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1387 			    NULL);
1388 		} else {
1389 next_page:
1390 			tm = m;
1391 			m = TAILQ_NEXT(m, listq);
1392 		}
1393 
1394 		/*
1395 		 * If the page is not in a normal state, skip it.  The page
1396 		 * can not be invalidated while the object lock is held.
1397 		 */
1398 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1399 			goto next_pindex;
1400 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1401 		    ("vm_object_madvise: page %p is fictitious", tm));
1402 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1403 		    ("vm_object_madvise: page %p is not managed", tm));
1404 		if (vm_page_tryxbusy(tm) == 0) {
1405 			if (object != tobject)
1406 				VM_OBJECT_WUNLOCK(object);
1407 			if (advice == MADV_WILLNEED) {
1408 				/*
1409 				 * Reference the page before unlocking and
1410 				 * sleeping so that the page daemon is less
1411 				 * likely to reclaim it.
1412 				 */
1413 				vm_page_aflag_set(tm, PGA_REFERENCED);
1414 			}
1415 			if (!vm_page_busy_sleep(tm, "madvpo", 0))
1416 				VM_OBJECT_WUNLOCK(tobject);
1417   			goto relookup;
1418 		}
1419 		vm_page_advise(tm, advice);
1420 		vm_page_xunbusy(tm);
1421 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1422 next_pindex:
1423 		if (tobject != object)
1424 			VM_OBJECT_WUNLOCK(tobject);
1425 	}
1426 	VM_OBJECT_WUNLOCK(object);
1427 }
1428 
1429 /*
1430  *	vm_object_shadow:
1431  *
1432  *	Create a new object which is backed by the
1433  *	specified existing object range.  The source
1434  *	object reference is deallocated.
1435  *
1436  *	The new object and offset into that object
1437  *	are returned in the source parameters.
1438  */
1439 void
1440 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1441     struct ucred *cred, bool shared)
1442 {
1443 	vm_object_t source;
1444 	vm_object_t result;
1445 
1446 	source = *object;
1447 
1448 	/*
1449 	 * Don't create the new object if the old object isn't shared.
1450 	 *
1451 	 * If we hold the only reference we can guarantee that it won't
1452 	 * increase while we have the map locked.  Otherwise the race is
1453 	 * harmless and we will end up with an extra shadow object that
1454 	 * will be collapsed later.
1455 	 */
1456 	if (source != NULL && source->ref_count == 1 &&
1457 	    (source->flags & OBJ_ANON) != 0)
1458 		return;
1459 
1460 	/*
1461 	 * Allocate a new object with the given length.
1462 	 */
1463 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1464 
1465 	/*
1466 	 * Store the offset into the source object, and fix up the offset into
1467 	 * the new object.
1468 	 */
1469 	result->backing_object_offset = *offset;
1470 
1471 	if (shared || source != NULL) {
1472 		VM_OBJECT_WLOCK(result);
1473 
1474 		/*
1475 		 * The new object shadows the source object, adding a
1476 		 * reference to it.  Our caller changes his reference
1477 		 * to point to the new object, removing a reference to
1478 		 * the source object.  Net result: no change of
1479 		 * reference count, unless the caller needs to add one
1480 		 * more reference due to forking a shared map entry.
1481 		 */
1482 		if (shared) {
1483 			vm_object_reference_locked(result);
1484 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1485 		}
1486 
1487 		/*
1488 		 * Try to optimize the result object's page color when
1489 		 * shadowing in order to maintain page coloring
1490 		 * consistency in the combined shadowed object.
1491 		 */
1492 		if (source != NULL) {
1493 			vm_object_backing_insert(result, source);
1494 			result->domain = source->domain;
1495 #if VM_NRESERVLEVEL > 0
1496 			vm_object_set_flag(result,
1497 			    (source->flags & OBJ_COLORED));
1498 			result->pg_color = (source->pg_color +
1499 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1500 			    1)) - 1);
1501 #endif
1502 		}
1503 		VM_OBJECT_WUNLOCK(result);
1504 	}
1505 
1506 	/*
1507 	 * Return the new things
1508 	 */
1509 	*offset = 0;
1510 	*object = result;
1511 }
1512 
1513 /*
1514  *	vm_object_split:
1515  *
1516  * Split the pages in a map entry into a new object.  This affords
1517  * easier removal of unused pages, and keeps object inheritance from
1518  * being a negative impact on memory usage.
1519  */
1520 void
1521 vm_object_split(vm_map_entry_t entry)
1522 {
1523 	struct pctrie_iter pages;
1524 	vm_page_t m;
1525 	vm_object_t orig_object, new_object, backing_object;
1526 	vm_pindex_t offidxstart;
1527 	vm_size_t size;
1528 
1529 	orig_object = entry->object.vm_object;
1530 	KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1531 	    ("vm_object_split:  Splitting object with multiple mappings."));
1532 	if ((orig_object->flags & OBJ_ANON) == 0)
1533 		return;
1534 	if (orig_object->ref_count <= 1)
1535 		return;
1536 	VM_OBJECT_WUNLOCK(orig_object);
1537 
1538 	offidxstart = OFF_TO_IDX(entry->offset);
1539 	size = atop(entry->end - entry->start);
1540 
1541 	new_object = vm_object_allocate_anon(size, orig_object,
1542 	    orig_object->cred, ptoa(size));
1543 
1544 	/*
1545 	 * We must wait for the orig_object to complete any in-progress
1546 	 * collapse so that the swap blocks are stable below.  The
1547 	 * additional reference on backing_object by new object will
1548 	 * prevent further collapse operations until split completes.
1549 	 */
1550 	VM_OBJECT_WLOCK(orig_object);
1551 	vm_object_collapse_wait(orig_object);
1552 
1553 	/*
1554 	 * At this point, the new object is still private, so the order in
1555 	 * which the original and new objects are locked does not matter.
1556 	 */
1557 	VM_OBJECT_WLOCK(new_object);
1558 	new_object->domain = orig_object->domain;
1559 	backing_object = orig_object->backing_object;
1560 	if (backing_object != NULL) {
1561 		vm_object_backing_insert_ref(new_object, backing_object);
1562 		new_object->backing_object_offset =
1563 		    orig_object->backing_object_offset + entry->offset;
1564 	}
1565 	if (orig_object->cred != NULL) {
1566 		crhold(orig_object->cred);
1567 		KASSERT(orig_object->charge >= ptoa(size),
1568 		    ("orig_object->charge < 0"));
1569 		orig_object->charge -= ptoa(size);
1570 	}
1571 
1572 	/*
1573 	 * Mark the split operation so that swap_pager_getpages() knows
1574 	 * that the object is in transition.
1575 	 */
1576 	vm_object_set_flag(orig_object, OBJ_SPLIT);
1577 	vm_page_iter_limit_init(&pages, orig_object, offidxstart + size);
1578 retry:
1579 	KASSERT(pctrie_iter_is_reset(&pages),
1580 	    ("%s: pctrie_iter not reset for retry", __func__));
1581 	for (m = vm_radix_iter_lookup_ge(&pages, offidxstart); m != NULL;
1582 	    m = vm_radix_iter_step(&pages)) {
1583 		/*
1584 		 * We must wait for pending I/O to complete before we can
1585 		 * rename the page.
1586 		 *
1587 		 * We do not have to VM_PROT_NONE the page as mappings should
1588 		 * not be changed by this operation.
1589 		 */
1590 		if (vm_page_tryxbusy(m) == 0) {
1591 			VM_OBJECT_WUNLOCK(new_object);
1592 			if (vm_page_busy_sleep(m, "spltwt", 0))
1593 				VM_OBJECT_WLOCK(orig_object);
1594 			pctrie_iter_reset(&pages);
1595 			VM_OBJECT_WLOCK(new_object);
1596 			goto retry;
1597 		}
1598 
1599 		/*
1600 		 * If the page was left invalid, it was likely placed there by
1601 		 * an incomplete fault.  Just remove and ignore.
1602 		 *
1603 		 * One other possibility is that the map entry is wired, in
1604 		 * which case we must hang on to the page to avoid leaking it,
1605 		 * as the map entry owns the wiring.  This case can arise if the
1606 		 * backing object is truncated by the pager.
1607 		 */
1608 		if (vm_page_none_valid(m) && entry->wired_count == 0) {
1609 			if (vm_page_iter_remove(&pages, m))
1610 				vm_page_free(m);
1611 			continue;
1612 		}
1613 
1614 		/* vm_page_iter_rename() will dirty the page if it is valid. */
1615 		if (!vm_page_iter_rename(&pages, m, new_object, m->pindex -
1616 		    offidxstart)) {
1617 			vm_page_xunbusy(m);
1618 			VM_OBJECT_WUNLOCK(new_object);
1619 			VM_OBJECT_WUNLOCK(orig_object);
1620 			vm_radix_wait();
1621 			pctrie_iter_reset(&pages);
1622 			VM_OBJECT_WLOCK(orig_object);
1623 			VM_OBJECT_WLOCK(new_object);
1624 			goto retry;
1625 		}
1626 
1627 #if VM_NRESERVLEVEL > 0
1628 		/*
1629 		 * If some of the reservation's allocated pages remain with
1630 		 * the original object, then transferring the reservation to
1631 		 * the new object is neither particularly beneficial nor
1632 		 * particularly harmful as compared to leaving the reservation
1633 		 * with the original object.  If, however, all of the
1634 		 * reservation's allocated pages are transferred to the new
1635 		 * object, then transferring the reservation is typically
1636 		 * beneficial.  Determining which of these two cases applies
1637 		 * would be more costly than unconditionally renaming the
1638 		 * reservation.
1639 		 */
1640 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1641 #endif
1642 	}
1643 
1644 	/*
1645 	 * swap_pager_copy() can sleep, in which case the orig_object's
1646 	 * and new_object's locks are released and reacquired.
1647 	 */
1648 	swap_pager_copy(orig_object, new_object, offidxstart, 0);
1649 
1650 	TAILQ_FOREACH(m, &new_object->memq, listq)
1651 		vm_page_xunbusy(m);
1652 
1653 	vm_object_clear_flag(orig_object, OBJ_SPLIT);
1654 	VM_OBJECT_WUNLOCK(orig_object);
1655 	VM_OBJECT_WUNLOCK(new_object);
1656 	entry->object.vm_object = new_object;
1657 	entry->offset = 0LL;
1658 	vm_object_deallocate(orig_object);
1659 	VM_OBJECT_WLOCK(new_object);
1660 }
1661 
1662 static vm_page_t
1663 vm_object_collapse_scan_wait(struct pctrie_iter *pages, vm_object_t object,
1664     vm_page_t p)
1665 {
1666 	vm_object_t backing_object;
1667 
1668 	VM_OBJECT_ASSERT_WLOCKED(object);
1669 	backing_object = object->backing_object;
1670 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1671 
1672 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1673 	    ("invalid ownership %p %p %p", p, object, backing_object));
1674 	/* The page is only NULL when rename fails. */
1675 	if (p == NULL) {
1676 		VM_OBJECT_WUNLOCK(object);
1677 		VM_OBJECT_WUNLOCK(backing_object);
1678 		vm_radix_wait();
1679 		VM_OBJECT_WLOCK(object);
1680 	} else if (p->object == object) {
1681 		VM_OBJECT_WUNLOCK(backing_object);
1682 		if (vm_page_busy_sleep(p, "vmocol", 0))
1683 			VM_OBJECT_WLOCK(object);
1684 	} else {
1685 		VM_OBJECT_WUNLOCK(object);
1686 		if (!vm_page_busy_sleep(p, "vmocol", 0))
1687 			VM_OBJECT_WUNLOCK(backing_object);
1688 		VM_OBJECT_WLOCK(object);
1689 	}
1690 	VM_OBJECT_WLOCK(backing_object);
1691 	vm_page_iter_init(pages, backing_object);
1692 	return (vm_radix_iter_lookup_ge(pages, 0));
1693 }
1694 
1695 static void
1696 vm_object_collapse_scan(vm_object_t object)
1697 {
1698 	struct pctrie_iter pages;
1699 	vm_object_t backing_object;
1700 	vm_page_t next, p, pp;
1701 	vm_pindex_t backing_offset_index, new_pindex;
1702 
1703 	VM_OBJECT_ASSERT_WLOCKED(object);
1704 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1705 
1706 	backing_object = object->backing_object;
1707 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1708 
1709 	/*
1710 	 * Our scan
1711 	 */
1712 	vm_page_iter_init(&pages, backing_object);
1713 	for (p = vm_radix_iter_lookup_ge(&pages, 0); p != NULL; p = next) {
1714 		/*
1715 		 * Check for busy page
1716 		 */
1717 		if (vm_page_tryxbusy(p) == 0) {
1718 			next = vm_object_collapse_scan_wait(&pages, object, p);
1719 			continue;
1720 		}
1721 
1722 		KASSERT(object->backing_object == backing_object,
1723 		    ("vm_object_collapse_scan: backing object mismatch %p != %p",
1724 		    object->backing_object, backing_object));
1725 		KASSERT(p->object == backing_object,
1726 		    ("vm_object_collapse_scan: object mismatch %p != %p",
1727 		    p->object, backing_object));
1728 
1729 		if (p->pindex < backing_offset_index || object->size <=
1730 		    (new_pindex = p->pindex - backing_offset_index)) {
1731 			vm_pager_freespace(backing_object, p->pindex, 1);
1732 
1733 			KASSERT(!pmap_page_is_mapped(p),
1734 			    ("freeing mapped page %p", p));
1735 			if (vm_page_iter_remove(&pages, p))
1736 				vm_page_free(p);
1737 			next = vm_radix_iter_step(&pages);
1738 			continue;
1739 		}
1740 
1741 		if (!vm_page_all_valid(p)) {
1742 			KASSERT(!pmap_page_is_mapped(p),
1743 			    ("freeing mapped page %p", p));
1744 			if (vm_page_iter_remove(&pages, p))
1745 				vm_page_free(p);
1746 			next = vm_radix_iter_step(&pages);
1747 			continue;
1748 		}
1749 
1750 		pp = vm_page_lookup(object, new_pindex);
1751 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1752 			vm_page_xunbusy(p);
1753 			/*
1754 			 * The page in the parent is busy and possibly not
1755 			 * (yet) valid.  Until its state is finalized by the
1756 			 * busy bit owner, we can't tell whether it shadows the
1757 			 * original page.
1758 			 */
1759 			next = vm_object_collapse_scan_wait(&pages, object, pp);
1760 			continue;
1761 		}
1762 
1763 		if (pp != NULL && vm_page_none_valid(pp)) {
1764 			/*
1765 			 * The page was invalid in the parent.  Likely placed
1766 			 * there by an incomplete fault.  Just remove and
1767 			 * ignore.  p can replace it.
1768 			 */
1769 			if (vm_page_remove(pp))
1770 				vm_page_free(pp);
1771 			pp = NULL;
1772 		}
1773 
1774 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1775 			NULL)) {
1776 			/*
1777 			 * The page already exists in the parent OR swap exists
1778 			 * for this location in the parent.  Leave the parent's
1779 			 * page alone.  Destroy the original page from the
1780 			 * backing object.
1781 			 */
1782 			vm_pager_freespace(backing_object, p->pindex, 1);
1783 			KASSERT(!pmap_page_is_mapped(p),
1784 			    ("freeing mapped page %p", p));
1785 			if (pp != NULL)
1786 				vm_page_xunbusy(pp);
1787 			if (vm_page_iter_remove(&pages, p))
1788 				vm_page_free(p);
1789 			next = vm_radix_iter_step(&pages);
1790 			continue;
1791 		}
1792 
1793 		/*
1794 		 * Page does not exist in parent, rename the page from the
1795 		 * backing object to the main object.
1796 		 *
1797 		 * If the page was mapped to a process, it can remain mapped
1798 		 * through the rename.  vm_page_iter_rename() will dirty the
1799 		 * page.
1800 		 */
1801 		if (!vm_page_iter_rename(&pages, p, object, new_pindex)) {
1802 			vm_page_xunbusy(p);
1803 			next = vm_object_collapse_scan_wait(&pages, object,
1804 			    NULL);
1805 			continue;
1806 		}
1807 
1808 		/* Use the old pindex to free the right page. */
1809 		vm_pager_freespace(backing_object, new_pindex +
1810 		    backing_offset_index, 1);
1811 
1812 #if VM_NRESERVLEVEL > 0
1813 		/*
1814 		 * Rename the reservation.
1815 		 */
1816 		vm_reserv_rename(p, object, backing_object,
1817 		    backing_offset_index);
1818 #endif
1819 		vm_page_xunbusy(p);
1820 		next = vm_radix_iter_step(&pages);
1821 	}
1822 	return;
1823 }
1824 
1825 /*
1826  *	vm_object_collapse:
1827  *
1828  *	Collapse an object with the object backing it.
1829  *	Pages in the backing object are moved into the
1830  *	parent, and the backing object is deallocated.
1831  */
1832 void
1833 vm_object_collapse(vm_object_t object)
1834 {
1835 	vm_object_t backing_object, new_backing_object;
1836 
1837 	VM_OBJECT_ASSERT_WLOCKED(object);
1838 
1839 	while (TRUE) {
1840 		KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1841 		    ("collapsing invalid object"));
1842 
1843 		/*
1844 		 * Wait for the backing_object to finish any pending
1845 		 * collapse so that the caller sees the shortest possible
1846 		 * shadow chain.
1847 		 */
1848 		backing_object = vm_object_backing_collapse_wait(object);
1849 		if (backing_object == NULL)
1850 			return;
1851 
1852 		KASSERT(object->ref_count > 0 &&
1853 		    object->ref_count > atomic_load_int(&object->shadow_count),
1854 		    ("collapse with invalid ref %d or shadow %d count.",
1855 		    object->ref_count, atomic_load_int(&object->shadow_count)));
1856 		KASSERT((backing_object->flags &
1857 		    (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1858 		    ("vm_object_collapse: Backing object already collapsing."));
1859 		KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1860 		    ("vm_object_collapse: object is already collapsing."));
1861 
1862 		/*
1863 		 * We know that we can either collapse the backing object if
1864 		 * the parent is the only reference to it, or (perhaps) have
1865 		 * the parent bypass the object if the parent happens to shadow
1866 		 * all the resident pages in the entire backing object.
1867 		 */
1868 		if (backing_object->ref_count == 1) {
1869 			KASSERT(atomic_load_int(&backing_object->shadow_count)
1870 			    == 1,
1871 			    ("vm_object_collapse: shadow_count: %d",
1872 			    atomic_load_int(&backing_object->shadow_count)));
1873 			vm_object_pip_add(object, 1);
1874 			vm_object_set_flag(object, OBJ_COLLAPSING);
1875 			vm_object_pip_add(backing_object, 1);
1876 			vm_object_set_flag(backing_object, OBJ_DEAD);
1877 
1878 			/*
1879 			 * If there is exactly one reference to the backing
1880 			 * object, we can collapse it into the parent.
1881 			 */
1882 			vm_object_collapse_scan(object);
1883 
1884 			/*
1885 			 * Move the pager from backing_object to object.
1886 			 *
1887 			 * swap_pager_copy() can sleep, in which case the
1888 			 * backing_object's and object's locks are released and
1889 			 * reacquired.
1890 			 */
1891 			swap_pager_copy(backing_object, object,
1892 			    OFF_TO_IDX(object->backing_object_offset), TRUE);
1893 
1894 			/*
1895 			 * Object now shadows whatever backing_object did.
1896 			 */
1897 			vm_object_clear_flag(object, OBJ_COLLAPSING);
1898 			vm_object_backing_transfer(object, backing_object);
1899 			object->backing_object_offset +=
1900 			    backing_object->backing_object_offset;
1901 			VM_OBJECT_WUNLOCK(object);
1902 			vm_object_pip_wakeup(object);
1903 
1904 			/*
1905 			 * Discard backing_object.
1906 			 *
1907 			 * Since the backing object has no pages, no pager left,
1908 			 * and no object references within it, all that is
1909 			 * necessary is to dispose of it.
1910 			 */
1911 			KASSERT(backing_object->ref_count == 1, (
1912 "backing_object %p was somehow re-referenced during collapse!",
1913 			    backing_object));
1914 			vm_object_pip_wakeup(backing_object);
1915 			(void)refcount_release(&backing_object->ref_count);
1916 			umtx_shm_object_terminated(backing_object);
1917 			vm_object_terminate(backing_object);
1918 			counter_u64_add(object_collapses, 1);
1919 			VM_OBJECT_WLOCK(object);
1920 		} else {
1921 			/*
1922 			 * If we do not entirely shadow the backing object,
1923 			 * there is nothing we can do so we give up.
1924 			 *
1925 			 * The object lock and backing_object lock must not
1926 			 * be dropped during this sequence.
1927 			 */
1928 			if (!swap_pager_scan_all_shadowed(object)) {
1929 				VM_OBJECT_WUNLOCK(backing_object);
1930 				break;
1931 			}
1932 
1933 			/*
1934 			 * Make the parent shadow the next object in the
1935 			 * chain.  Deallocating backing_object will not remove
1936 			 * it, since its reference count is at least 2.
1937 			 */
1938 			vm_object_backing_remove_locked(object);
1939 			new_backing_object = backing_object->backing_object;
1940 			if (new_backing_object != NULL) {
1941 				vm_object_backing_insert_ref(object,
1942 				    new_backing_object);
1943 				object->backing_object_offset +=
1944 				    backing_object->backing_object_offset;
1945 			}
1946 
1947 			/*
1948 			 * Drop the reference count on backing_object. Since
1949 			 * its ref_count was at least 2, it will not vanish.
1950 			 */
1951 			(void)refcount_release(&backing_object->ref_count);
1952 			KASSERT(backing_object->ref_count >= 1, (
1953 "backing_object %p was somehow dereferenced during collapse!",
1954 			    backing_object));
1955 			VM_OBJECT_WUNLOCK(backing_object);
1956 			counter_u64_add(object_bypasses, 1);
1957 		}
1958 
1959 		/*
1960 		 * Try again with this object's new backing object.
1961 		 */
1962 	}
1963 }
1964 
1965 /*
1966  *	vm_object_page_remove:
1967  *
1968  *	For the given object, either frees or invalidates each of the
1969  *	specified pages.  In general, a page is freed.  However, if a page is
1970  *	wired for any reason other than the existence of a managed, wired
1971  *	mapping, then it may be invalidated but not removed from the object.
1972  *	Pages are specified by the given range ["start", "end") and the option
1973  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
1974  *	extends from "start" to the end of the object.  If the option
1975  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1976  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
1977  *	specified, then the pages within the specified range must have no
1978  *	mappings.  Otherwise, if this option is not specified, any mappings to
1979  *	the specified pages are removed before the pages are freed or
1980  *	invalidated.
1981  *
1982  *	In general, this operation should only be performed on objects that
1983  *	contain managed pages.  There are, however, two exceptions.  First, it
1984  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
1985  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1986  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
1987  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
1988  *
1989  *	The object must be locked.
1990  */
1991 void
1992 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1993     int options)
1994 {
1995 	struct pctrie_iter pages;
1996 	vm_page_t p;
1997 
1998 	VM_OBJECT_ASSERT_WLOCKED(object);
1999 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2000 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2001 	    ("vm_object_page_remove: illegal options for object %p", object));
2002 	if (object->resident_page_count == 0)
2003 		return;
2004 	vm_object_pip_add(object, 1);
2005 	vm_page_iter_limit_init(&pages, object, end);
2006 again:
2007 	KASSERT(pctrie_iter_is_reset(&pages),
2008 	    ("%s: pctrie_iter not reset for retry", __func__));
2009 	for (p = vm_radix_iter_lookup_ge(&pages, start); p != NULL;
2010 	     p = vm_radix_iter_step(&pages)) {
2011 		/*
2012 		 * Skip invalid pages if asked to do so.  Try to avoid acquiring
2013 		 * the busy lock, as some consumers rely on this to avoid
2014 		 * deadlocks.
2015 		 *
2016 		 * A thread may concurrently transition the page from invalid to
2017 		 * valid using only the busy lock, so the result of this check
2018 		 * is immediately stale.  It is up to consumers to handle this,
2019 		 * for instance by ensuring that all invalid->valid transitions
2020 		 * happen with a mutex held, as may be possible for a
2021 		 * filesystem.
2022 		 */
2023 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2024 			continue;
2025 
2026 		/*
2027 		 * If the page is wired for any reason besides the existence
2028 		 * of managed, wired mappings, then it cannot be freed.  For
2029 		 * example, fictitious pages, which represent device memory,
2030 		 * are inherently wired and cannot be freed.  They can,
2031 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2032 		 * not specified.
2033 		 */
2034 		if (vm_page_tryxbusy(p) == 0) {
2035 			if (vm_page_busy_sleep(p, "vmopar", 0))
2036 				VM_OBJECT_WLOCK(object);
2037 			pctrie_iter_reset(&pages);
2038 			goto again;
2039 		}
2040 		if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2041 			vm_page_xunbusy(p);
2042 			continue;
2043 		}
2044 		if (vm_page_wired(p)) {
2045 wired:
2046 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2047 			    object->ref_count != 0)
2048 				pmap_remove_all(p);
2049 			if ((options & OBJPR_CLEANONLY) == 0) {
2050 				vm_page_invalid(p);
2051 				vm_page_undirty(p);
2052 			}
2053 			vm_page_xunbusy(p);
2054 			continue;
2055 		}
2056 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2057 		    ("vm_object_page_remove: page %p is fictitious", p));
2058 		if ((options & OBJPR_CLEANONLY) != 0 &&
2059 		    !vm_page_none_valid(p)) {
2060 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2061 			    object->ref_count != 0 &&
2062 			    !vm_page_try_remove_write(p))
2063 				goto wired;
2064 			if (p->dirty != 0) {
2065 				vm_page_xunbusy(p);
2066 				continue;
2067 			}
2068 		}
2069 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2070 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2071 			goto wired;
2072 		vm_page_iter_free(&pages, p);
2073 	}
2074 	vm_object_pip_wakeup(object);
2075 
2076 	vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2077 	    start);
2078 }
2079 
2080 /*
2081  *	vm_object_page_noreuse:
2082  *
2083  *	For the given object, attempt to move the specified pages to
2084  *	the head of the inactive queue.  This bypasses regular LRU
2085  *	operation and allows the pages to be reused quickly under memory
2086  *	pressure.  If a page is wired for any reason, then it will not
2087  *	be queued.  Pages are specified by the range ["start", "end").
2088  *	As a special case, if "end" is zero, then the range extends from
2089  *	"start" to the end of the object.
2090  *
2091  *	This operation should only be performed on objects that
2092  *	contain non-fictitious, managed pages.
2093  *
2094  *	The object must be locked.
2095  */
2096 void
2097 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2098 {
2099 	vm_page_t p, next;
2100 
2101 	VM_OBJECT_ASSERT_LOCKED(object);
2102 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2103 	    ("vm_object_page_noreuse: illegal object %p", object));
2104 	if (object->resident_page_count == 0)
2105 		return;
2106 	p = vm_page_find_least(object, start);
2107 
2108 	/*
2109 	 * Here, the variable "p" is either (1) the page with the least pindex
2110 	 * greater than or equal to the parameter "start" or (2) NULL.
2111 	 */
2112 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2113 		next = TAILQ_NEXT(p, listq);
2114 		vm_page_deactivate_noreuse(p);
2115 	}
2116 }
2117 
2118 /*
2119  *	Populate the specified range of the object with valid pages.  Returns
2120  *	TRUE if the range is successfully populated and FALSE otherwise.
2121  *
2122  *	Note: This function should be optimized to pass a larger array of
2123  *	pages to vm_pager_get_pages() before it is applied to a non-
2124  *	OBJT_DEVICE object.
2125  *
2126  *	The object must be locked.
2127  */
2128 boolean_t
2129 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2130 {
2131 	vm_page_t m;
2132 	vm_pindex_t pindex;
2133 	int rv;
2134 
2135 	VM_OBJECT_ASSERT_WLOCKED(object);
2136 	for (pindex = start; pindex < end; pindex++) {
2137 		rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2138 		if (rv != VM_PAGER_OK)
2139 			break;
2140 
2141 		/*
2142 		 * Keep "m" busy because a subsequent iteration may unlock
2143 		 * the object.
2144 		 */
2145 	}
2146 	if (pindex > start) {
2147 		m = vm_page_lookup(object, start);
2148 		while (m != NULL && m->pindex < pindex) {
2149 			vm_page_xunbusy(m);
2150 			m = TAILQ_NEXT(m, listq);
2151 		}
2152 	}
2153 	return (pindex == end);
2154 }
2155 
2156 /*
2157  *	Routine:	vm_object_coalesce
2158  *	Function:	Coalesces two objects backing up adjoining
2159  *			regions of memory into a single object.
2160  *
2161  *	returns TRUE if objects were combined.
2162  *
2163  *	NOTE:	Only works at the moment if the second object is NULL -
2164  *		if it's not, which object do we lock first?
2165  *
2166  *	Parameters:
2167  *		prev_object	First object to coalesce
2168  *		prev_offset	Offset into prev_object
2169  *		prev_size	Size of reference to prev_object
2170  *		next_size	Size of reference to the second object
2171  *		reserved	Indicator that extension region has
2172  *				swap accounted for
2173  *
2174  *	Conditions:
2175  *	The object must *not* be locked.
2176  */
2177 boolean_t
2178 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2179     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2180 {
2181 	vm_pindex_t next_pindex;
2182 
2183 	if (prev_object == NULL)
2184 		return (TRUE);
2185 	if ((prev_object->flags & OBJ_ANON) == 0)
2186 		return (FALSE);
2187 
2188 	VM_OBJECT_WLOCK(prev_object);
2189 	/*
2190 	 * Try to collapse the object first.
2191 	 */
2192 	vm_object_collapse(prev_object);
2193 
2194 	/*
2195 	 * Can't coalesce if: . more than one reference . paged out . shadows
2196 	 * another object . has a copy elsewhere (any of which mean that the
2197 	 * pages not mapped to prev_entry may be in use anyway)
2198 	 */
2199 	if (prev_object->backing_object != NULL) {
2200 		VM_OBJECT_WUNLOCK(prev_object);
2201 		return (FALSE);
2202 	}
2203 
2204 	prev_size >>= PAGE_SHIFT;
2205 	next_size >>= PAGE_SHIFT;
2206 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2207 
2208 	if (prev_object->ref_count > 1 &&
2209 	    prev_object->size != next_pindex &&
2210 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2211 		VM_OBJECT_WUNLOCK(prev_object);
2212 		return (FALSE);
2213 	}
2214 
2215 	/*
2216 	 * Account for the charge.
2217 	 */
2218 	if (prev_object->cred != NULL) {
2219 		/*
2220 		 * If prev_object was charged, then this mapping,
2221 		 * although not charged now, may become writable
2222 		 * later. Non-NULL cred in the object would prevent
2223 		 * swap reservation during enabling of the write
2224 		 * access, so reserve swap now. Failed reservation
2225 		 * cause allocation of the separate object for the map
2226 		 * entry, and swap reservation for this entry is
2227 		 * managed in appropriate time.
2228 		 */
2229 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2230 		    prev_object->cred)) {
2231 			VM_OBJECT_WUNLOCK(prev_object);
2232 			return (FALSE);
2233 		}
2234 		prev_object->charge += ptoa(next_size);
2235 	}
2236 
2237 	/*
2238 	 * Remove any pages that may still be in the object from a previous
2239 	 * deallocation.
2240 	 */
2241 	if (next_pindex < prev_object->size) {
2242 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2243 		    next_size, 0);
2244 #if 0
2245 		if (prev_object->cred != NULL) {
2246 			KASSERT(prev_object->charge >=
2247 			    ptoa(prev_object->size - next_pindex),
2248 			    ("object %p overcharged 1 %jx %jx", prev_object,
2249 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2250 			prev_object->charge -= ptoa(prev_object->size -
2251 			    next_pindex);
2252 		}
2253 #endif
2254 	}
2255 
2256 	/*
2257 	 * Extend the object if necessary.
2258 	 */
2259 	if (next_pindex + next_size > prev_object->size)
2260 		prev_object->size = next_pindex + next_size;
2261 
2262 	VM_OBJECT_WUNLOCK(prev_object);
2263 	return (TRUE);
2264 }
2265 
2266 /*
2267  * Fill in the m_dst array with up to *rbehind optional pages before m_src[0]
2268  * and up to *rahead optional pages after m_src[count - 1].  In both cases, stop
2269  * the filling-in short on encountering a cached page, an object boundary limit,
2270  * or an allocation error.  Update *rbehind and *rahead to indicate the number
2271  * of pages allocated.  Copy elements of m_src into array elements from
2272  * m_dst[*rbehind] to m_dst[*rbehind + count -1].
2273  */
2274 void
2275 vm_object_prepare_buf_pages(vm_object_t object, vm_page_t *ma_dst, int count,
2276     int *rbehind, int *rahead, vm_page_t *ma_src)
2277 {
2278 	struct pctrie_iter pages;
2279 	vm_pindex_t pindex;
2280 	vm_page_t m, mpred, msucc;
2281 
2282 	vm_page_iter_init(&pages, object);
2283 	VM_OBJECT_ASSERT_LOCKED(object);
2284 	if (*rbehind != 0) {
2285 		m = ma_src[0];
2286 		pindex = m->pindex;
2287 		mpred = vm_radix_iter_lookup_lt(&pages, pindex);
2288 		*rbehind = MIN(*rbehind,
2289 		    pindex - (mpred != NULL ? mpred->pindex + 1 : 0));
2290 		/* Stepping backward from pindex, mpred doesn't change. */
2291 		for (int i = 0; i < *rbehind; i++) {
2292 			m = vm_page_alloc_after(object, &pages, pindex - i - 1,
2293 			    VM_ALLOC_NORMAL, mpred);
2294 			if (m == NULL) {
2295 				/* Shift the array. */
2296 				for (int j = 0; j < i; j++)
2297 					ma_dst[j] = ma_dst[j + *rbehind - i];
2298 				*rbehind = i;
2299 				*rahead = 0;
2300 				break;
2301 			}
2302 			ma_dst[*rbehind - i - 1] = m;
2303 		}
2304 	}
2305 	for (int i = 0; i < count; i++)
2306 		ma_dst[*rbehind + i] = ma_src[i];
2307 	if (*rahead != 0) {
2308 		m = ma_src[count - 1];
2309 		pindex = m->pindex + 1;
2310 		msucc = vm_radix_iter_lookup_ge(&pages, pindex);
2311 		*rahead = MIN(*rahead,
2312 		    (msucc != NULL ? msucc->pindex : object->size) - pindex);
2313 		mpred = m;
2314 		for (int i = 0; i < *rahead; i++) {
2315 			m = vm_page_alloc_after(object, &pages, pindex + i,
2316 			    VM_ALLOC_NORMAL, mpred);
2317 			if (m == NULL) {
2318 				*rahead = i;
2319 				break;
2320 			}
2321 			ma_dst[*rbehind + count + i] = mpred = m;
2322 		}
2323 	}
2324 }
2325 
2326 void
2327 vm_object_set_writeable_dirty_(vm_object_t object)
2328 {
2329 	atomic_add_int(&object->generation, 1);
2330 }
2331 
2332 bool
2333 vm_object_mightbedirty_(vm_object_t object)
2334 {
2335 	return (object->generation != object->cleangeneration);
2336 }
2337 
2338 /*
2339  *	vm_object_unwire:
2340  *
2341  *	For each page offset within the specified range of the given object,
2342  *	find the highest-level page in the shadow chain and unwire it.  A page
2343  *	must exist at every page offset, and the highest-level page must be
2344  *	wired.
2345  */
2346 void
2347 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2348     uint8_t queue)
2349 {
2350 	vm_object_t tobject, t1object;
2351 	vm_page_t m, tm;
2352 	vm_pindex_t end_pindex, pindex, tpindex;
2353 	int depth, locked_depth;
2354 
2355 	KASSERT((offset & PAGE_MASK) == 0,
2356 	    ("vm_object_unwire: offset is not page aligned"));
2357 	KASSERT((length & PAGE_MASK) == 0,
2358 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2359 	/* The wired count of a fictitious page never changes. */
2360 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2361 		return;
2362 	pindex = OFF_TO_IDX(offset);
2363 	end_pindex = pindex + atop(length);
2364 again:
2365 	locked_depth = 1;
2366 	VM_OBJECT_RLOCK(object);
2367 	m = vm_page_find_least(object, pindex);
2368 	while (pindex < end_pindex) {
2369 		if (m == NULL || pindex < m->pindex) {
2370 			/*
2371 			 * The first object in the shadow chain doesn't
2372 			 * contain a page at the current index.  Therefore,
2373 			 * the page must exist in a backing object.
2374 			 */
2375 			tobject = object;
2376 			tpindex = pindex;
2377 			depth = 0;
2378 			do {
2379 				tpindex +=
2380 				    OFF_TO_IDX(tobject->backing_object_offset);
2381 				tobject = tobject->backing_object;
2382 				KASSERT(tobject != NULL,
2383 				    ("vm_object_unwire: missing page"));
2384 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2385 					goto next_page;
2386 				depth++;
2387 				if (depth == locked_depth) {
2388 					locked_depth++;
2389 					VM_OBJECT_RLOCK(tobject);
2390 				}
2391 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2392 			    NULL);
2393 		} else {
2394 			tm = m;
2395 			m = TAILQ_NEXT(m, listq);
2396 		}
2397 		if (vm_page_trysbusy(tm) == 0) {
2398 			for (tobject = object; locked_depth >= 1;
2399 			    locked_depth--) {
2400 				t1object = tobject->backing_object;
2401 				if (tm->object != tobject)
2402 					VM_OBJECT_RUNLOCK(tobject);
2403 				tobject = t1object;
2404 			}
2405 			tobject = tm->object;
2406 			if (!vm_page_busy_sleep(tm, "unwbo",
2407 			    VM_ALLOC_IGN_SBUSY))
2408 				VM_OBJECT_RUNLOCK(tobject);
2409 			goto again;
2410 		}
2411 		vm_page_unwire(tm, queue);
2412 		vm_page_sunbusy(tm);
2413 next_page:
2414 		pindex++;
2415 	}
2416 	/* Release the accumulated object locks. */
2417 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2418 		t1object = tobject->backing_object;
2419 		VM_OBJECT_RUNLOCK(tobject);
2420 		tobject = t1object;
2421 	}
2422 }
2423 
2424 /*
2425  * Return the vnode for the given object, or NULL if none exists.
2426  * For tmpfs objects, the function may return NULL if there is
2427  * no vnode allocated at the time of the call.
2428  */
2429 struct vnode *
2430 vm_object_vnode(vm_object_t object)
2431 {
2432 	struct vnode *vp;
2433 
2434 	VM_OBJECT_ASSERT_LOCKED(object);
2435 	vm_pager_getvp(object, &vp, NULL);
2436 	return (vp);
2437 }
2438 
2439 /*
2440  * Busy the vm object.  This prevents new pages belonging to the object from
2441  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2442  * for checking page state before proceeding.
2443  */
2444 void
2445 vm_object_busy(vm_object_t obj)
2446 {
2447 
2448 	VM_OBJECT_ASSERT_LOCKED(obj);
2449 
2450 	blockcount_acquire(&obj->busy, 1);
2451 	/* The fence is required to order loads of page busy. */
2452 	atomic_thread_fence_acq_rel();
2453 }
2454 
2455 void
2456 vm_object_unbusy(vm_object_t obj)
2457 {
2458 
2459 	blockcount_release(&obj->busy, 1);
2460 }
2461 
2462 void
2463 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2464 {
2465 
2466 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2467 
2468 	(void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2469 }
2470 
2471 /*
2472  * This function aims to determine if the object is mapped,
2473  * specifically, if it is referenced by a vm_map_entry.  Because
2474  * objects occasionally acquire transient references that do not
2475  * represent a mapping, the method used here is inexact.  However, it
2476  * has very low overhead and is good enough for the advisory
2477  * vm.vmtotal sysctl.
2478  */
2479 bool
2480 vm_object_is_active(vm_object_t obj)
2481 {
2482 
2483 	return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2484 }
2485 
2486 static int
2487 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2488 {
2489 	struct kinfo_vmobject *kvo;
2490 	char *fullpath, *freepath;
2491 	struct vnode *vp;
2492 	struct vattr va;
2493 	vm_object_t obj;
2494 	vm_page_t m;
2495 	u_long sp;
2496 	int count, error;
2497 	key_t key;
2498 	unsigned short seq;
2499 	bool want_path;
2500 
2501 	if (req->oldptr == NULL) {
2502 		/*
2503 		 * If an old buffer has not been provided, generate an
2504 		 * estimate of the space needed for a subsequent call.
2505 		 */
2506 		mtx_lock(&vm_object_list_mtx);
2507 		count = 0;
2508 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2509 			if (obj->type == OBJT_DEAD)
2510 				continue;
2511 			count++;
2512 		}
2513 		mtx_unlock(&vm_object_list_mtx);
2514 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2515 		    count * 11 / 10));
2516 	}
2517 
2518 	want_path = !(swap_only || jailed(curthread->td_ucred));
2519 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO);
2520 	error = 0;
2521 
2522 	/*
2523 	 * VM objects are type stable and are never removed from the
2524 	 * list once added.  This allows us to safely read obj->object_list
2525 	 * after reacquiring the VM object lock.
2526 	 */
2527 	mtx_lock(&vm_object_list_mtx);
2528 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2529 		if (obj->type == OBJT_DEAD ||
2530 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2531 			continue;
2532 		VM_OBJECT_RLOCK(obj);
2533 		if (obj->type == OBJT_DEAD ||
2534 		    (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2535 			VM_OBJECT_RUNLOCK(obj);
2536 			continue;
2537 		}
2538 		mtx_unlock(&vm_object_list_mtx);
2539 		kvo->kvo_size = ptoa(obj->size);
2540 		kvo->kvo_resident = obj->resident_page_count;
2541 		kvo->kvo_ref_count = obj->ref_count;
2542 		kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2543 		kvo->kvo_memattr = obj->memattr;
2544 		kvo->kvo_active = 0;
2545 		kvo->kvo_inactive = 0;
2546 		kvo->kvo_flags = 0;
2547 		if (!swap_only) {
2548 			TAILQ_FOREACH(m, &obj->memq, listq) {
2549 				/*
2550 				 * A page may belong to the object but be
2551 				 * dequeued and set to PQ_NONE while the
2552 				 * object lock is not held.  This makes the
2553 				 * reads of m->queue below racy, and we do not
2554 				 * count pages set to PQ_NONE.  However, this
2555 				 * sysctl is only meant to give an
2556 				 * approximation of the system anyway.
2557 				 */
2558 				if (vm_page_active(m))
2559 					kvo->kvo_active++;
2560 				else if (vm_page_inactive(m))
2561 					kvo->kvo_inactive++;
2562 				else if (vm_page_in_laundry(m))
2563 					kvo->kvo_laundry++;
2564 			}
2565 		}
2566 
2567 		kvo->kvo_vn_fileid = 0;
2568 		kvo->kvo_vn_fsid = 0;
2569 		kvo->kvo_vn_fsid_freebsd11 = 0;
2570 		freepath = NULL;
2571 		fullpath = "";
2572 		vp = NULL;
2573 		kvo->kvo_type = vm_object_kvme_type(obj, want_path ? &vp :
2574 		    NULL);
2575 		if (vp != NULL) {
2576 			vref(vp);
2577 		} else if ((obj->flags & OBJ_ANON) != 0) {
2578 			MPASS(kvo->kvo_type == KVME_TYPE_SWAP);
2579 			kvo->kvo_me = (uintptr_t)obj;
2580 			/* tmpfs objs are reported as vnodes */
2581 			kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2582 			sp = swap_pager_swapped_pages(obj);
2583 			kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2584 		}
2585 		if (obj->type == OBJT_DEVICE || obj->type == OBJT_MGTDEVICE) {
2586 			cdev_pager_get_path(obj, kvo->kvo_path,
2587 			    sizeof(kvo->kvo_path));
2588 		}
2589 		VM_OBJECT_RUNLOCK(obj);
2590 		if ((obj->flags & OBJ_SYSVSHM) != 0) {
2591 			kvo->kvo_flags |= KVMO_FLAG_SYSVSHM;
2592 			shmobjinfo(obj, &key, &seq);
2593 			kvo->kvo_vn_fileid = key;
2594 			kvo->kvo_vn_fsid_freebsd11 = seq;
2595 		}
2596 		if ((obj->flags & OBJ_POSIXSHM) != 0) {
2597 			kvo->kvo_flags |= KVMO_FLAG_POSIXSHM;
2598 			shm_get_path(obj, kvo->kvo_path,
2599 			    sizeof(kvo->kvo_path));
2600 		}
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 			strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2612 			free(freepath, M_TEMP);
2613 		}
2614 
2615 		/* Pack record size down */
2616 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2617 		    + strlen(kvo->kvo_path) + 1;
2618 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2619 		    sizeof(uint64_t));
2620 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2621 		maybe_yield();
2622 		mtx_lock(&vm_object_list_mtx);
2623 		if (error)
2624 			break;
2625 	}
2626 	mtx_unlock(&vm_object_list_mtx);
2627 	free(kvo, M_TEMP);
2628 	return (error);
2629 }
2630 
2631 static int
2632 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2633 {
2634 	return (vm_object_list_handler(req, false));
2635 }
2636 
2637 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2638     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2639     "List of VM objects");
2640 
2641 static int
2642 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2643 {
2644 	return (vm_object_list_handler(req, true));
2645 }
2646 
2647 /*
2648  * This sysctl returns list of the anonymous or swap objects. Intent
2649  * is to provide stripped optimized list useful to analyze swap use.
2650  * Since technically non-swap (default) objects participate in the
2651  * shadow chains, and are converted to swap type as needed by swap
2652  * pager, we must report them.
2653  */
2654 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2655     CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2656     sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2657     "List of swap VM objects");
2658 
2659 #include "opt_ddb.h"
2660 #ifdef DDB
2661 #include <sys/kernel.h>
2662 
2663 #include <sys/cons.h>
2664 
2665 #include <ddb/ddb.h>
2666 
2667 static int
2668 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2669 {
2670 	vm_map_t tmpm;
2671 	vm_map_entry_t tmpe;
2672 	vm_object_t obj;
2673 
2674 	if (map == 0)
2675 		return 0;
2676 
2677 	if (entry == 0) {
2678 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2679 			if (_vm_object_in_map(map, object, tmpe)) {
2680 				return 1;
2681 			}
2682 		}
2683 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2684 		tmpm = entry->object.sub_map;
2685 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2686 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2687 				return 1;
2688 			}
2689 		}
2690 	} else if ((obj = entry->object.vm_object) != NULL) {
2691 		for (; obj; obj = obj->backing_object)
2692 			if (obj == object) {
2693 				return 1;
2694 			}
2695 	}
2696 	return 0;
2697 }
2698 
2699 static int
2700 vm_object_in_map(vm_object_t object)
2701 {
2702 	struct proc *p;
2703 
2704 	/* sx_slock(&allproc_lock); */
2705 	FOREACH_PROC_IN_SYSTEM(p) {
2706 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2707 			continue;
2708 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2709 			/* sx_sunlock(&allproc_lock); */
2710 			return 1;
2711 		}
2712 	}
2713 	/* sx_sunlock(&allproc_lock); */
2714 	if (_vm_object_in_map(kernel_map, object, 0))
2715 		return 1;
2716 	return 0;
2717 }
2718 
2719 DB_SHOW_COMMAND_FLAGS(vmochk, vm_object_check, DB_CMD_MEMSAFE)
2720 {
2721 	vm_object_t object;
2722 
2723 	/*
2724 	 * make sure that internal objs are in a map somewhere
2725 	 * and none have zero ref counts.
2726 	 */
2727 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2728 		if ((object->flags & OBJ_ANON) != 0) {
2729 			if (object->ref_count == 0) {
2730 				db_printf(
2731 			"vmochk: internal obj has zero ref count: %lu\n",
2732 				    (u_long)object->size);
2733 			}
2734 			if (!vm_object_in_map(object)) {
2735 				db_printf(
2736 			"vmochk: internal obj is not in a map: "
2737 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2738 				    object->ref_count, (u_long)object->size,
2739 				    (u_long)object->size,
2740 				    (void *)object->backing_object);
2741 			}
2742 		}
2743 		if (db_pager_quit)
2744 			return;
2745 	}
2746 }
2747 
2748 /*
2749  *	vm_object_print:	[ debug ]
2750  */
2751 DB_SHOW_COMMAND(object, vm_object_print_static)
2752 {
2753 	/* XXX convert args. */
2754 	vm_object_t object = (vm_object_t)addr;
2755 	boolean_t full = have_addr;
2756 
2757 	vm_page_t p;
2758 
2759 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2760 #define	count	was_count
2761 
2762 	int count;
2763 
2764 	if (object == NULL)
2765 		return;
2766 
2767 	db_iprintf("Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x",
2768 	    object, (int)object->type, (uintmax_t)object->size,
2769 	    object->resident_page_count, object->ref_count, object->flags);
2770 	db_iprintf(" ruid %d charge %jx\n",
2771 	    object->cred ? object->cred->cr_ruid : -1,
2772 	    (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