xref: /freebsd/sys/vm/vm_object.c (revision c14a5a8800a0f7a007f8cd197b4cad4d26a78f8c)
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	from: @(#)vm_object.c	8.5 (Berkeley) 3/22/94
35  *
36  *
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62 
63 /*
64  *	Virtual memory object module.
65  */
66 
67 #include <sys/cdefs.h>
68 __FBSDID("$FreeBSD$");
69 
70 #include "opt_vm.h"
71 
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/cpuset.h>
75 #include <sys/lock.h>
76 #include <sys/mman.h>
77 #include <sys/mount.h>
78 #include <sys/kernel.h>
79 #include <sys/pctrie.h>
80 #include <sys/sysctl.h>
81 #include <sys/mutex.h>
82 #include <sys/proc.h>		/* for curproc, pageproc */
83 #include <sys/refcount.h>
84 #include <sys/socket.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 #include <sys/sx.h>
92 
93 #include <vm/vm.h>
94 #include <vm/vm_param.h>
95 #include <vm/pmap.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_object.h>
98 #include <vm/vm_page.h>
99 #include <vm/vm_pageout.h>
100 #include <vm/vm_pager.h>
101 #include <vm/vm_phys.h>
102 #include <vm/vm_pagequeue.h>
103 #include <vm/swap_pager.h>
104 #include <vm/vm_kern.h>
105 #include <vm/vm_extern.h>
106 #include <vm/vm_radix.h>
107 #include <vm/vm_reserv.h>
108 #include <vm/uma.h>
109 
110 static int old_msync;
111 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
112     "Use old (insecure) msync behavior");
113 
114 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
115 		    int pagerflags, int flags, boolean_t *allclean,
116 		    boolean_t *eio);
117 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
118 		    boolean_t *allclean);
119 static void	vm_object_qcollapse(vm_object_t object);
120 static void	vm_object_vndeallocate(vm_object_t object);
121 static void	vm_object_backing_remove(vm_object_t object);
122 
123 /*
124  *	Virtual memory objects maintain the actual data
125  *	associated with allocated virtual memory.  A given
126  *	page of memory exists within exactly one object.
127  *
128  *	An object is only deallocated when all "references"
129  *	are given up.  Only one "reference" to a given
130  *	region of an object should be writeable.
131  *
132  *	Associated with each object is a list of all resident
133  *	memory pages belonging to that object; this list is
134  *	maintained by the "vm_page" module, and locked by the object's
135  *	lock.
136  *
137  *	Each object also records a "pager" routine which is
138  *	used to retrieve (and store) pages to the proper backing
139  *	storage.  In addition, objects may be backed by other
140  *	objects from which they were virtual-copied.
141  *
142  *	The only items within the object structure which are
143  *	modified after time of creation are:
144  *		reference count		locked by object's lock
145  *		pager routine		locked by object's lock
146  *
147  */
148 
149 struct object_q vm_object_list;
150 struct mtx vm_object_list_mtx;	/* lock for object list and count */
151 
152 struct vm_object kernel_object_store;
153 
154 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
155     "VM object stats");
156 
157 static counter_u64_t object_collapses = EARLY_COUNTER;
158 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
159     &object_collapses,
160     "VM object collapses");
161 
162 static counter_u64_t object_bypasses = EARLY_COUNTER;
163 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
164     &object_bypasses,
165     "VM object bypasses");
166 
167 static void
168 counter_startup(void)
169 {
170 
171 	object_collapses = counter_u64_alloc(M_WAITOK);
172 	object_bypasses = counter_u64_alloc(M_WAITOK);
173 }
174 SYSINIT(object_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL);
175 
176 static uma_zone_t obj_zone;
177 
178 static int vm_object_zinit(void *mem, int size, int flags);
179 
180 #ifdef INVARIANTS
181 static void vm_object_zdtor(void *mem, int size, void *arg);
182 
183 static void
184 vm_object_zdtor(void *mem, int size, void *arg)
185 {
186 	vm_object_t object;
187 
188 	object = (vm_object_t)mem;
189 	KASSERT(object->ref_count == 0,
190 	    ("object %p ref_count = %d", object, object->ref_count));
191 	KASSERT(TAILQ_EMPTY(&object->memq),
192 	    ("object %p has resident pages in its memq", object));
193 	KASSERT(vm_radix_is_empty(&object->rtree),
194 	    ("object %p has resident pages in its trie", object));
195 #if VM_NRESERVLEVEL > 0
196 	KASSERT(LIST_EMPTY(&object->rvq),
197 	    ("object %p has reservations",
198 	    object));
199 #endif
200 	KASSERT(REFCOUNT_COUNT(object->paging_in_progress) == 0,
201 	    ("object %p paging_in_progress = %d",
202 	    object, REFCOUNT_COUNT(object->paging_in_progress)));
203 	KASSERT(object->busy == 0,
204 	    ("object %p busy = %d",
205 	    object, object->busy));
206 	KASSERT(object->resident_page_count == 0,
207 	    ("object %p resident_page_count = %d",
208 	    object, object->resident_page_count));
209 	KASSERT(object->shadow_count == 0,
210 	    ("object %p shadow_count = %d",
211 	    object, object->shadow_count));
212 	KASSERT(object->type == OBJT_DEAD,
213 	    ("object %p has non-dead type %d",
214 	    object, object->type));
215 }
216 #endif
217 
218 static int
219 vm_object_zinit(void *mem, int size, int flags)
220 {
221 	vm_object_t object;
222 
223 	object = (vm_object_t)mem;
224 	rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
225 
226 	/* These are true for any object that has been freed */
227 	object->type = OBJT_DEAD;
228 	vm_radix_init(&object->rtree);
229 	refcount_init(&object->ref_count, 0);
230 	refcount_init(&object->paging_in_progress, 0);
231 	refcount_init(&object->busy, 0);
232 	object->resident_page_count = 0;
233 	object->shadow_count = 0;
234 	object->flags = OBJ_DEAD;
235 
236 	mtx_lock(&vm_object_list_mtx);
237 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
238 	mtx_unlock(&vm_object_list_mtx);
239 	return (0);
240 }
241 
242 static void
243 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
244     vm_object_t object, void *handle)
245 {
246 
247 	TAILQ_INIT(&object->memq);
248 	LIST_INIT(&object->shadow_head);
249 
250 	object->type = type;
251 	if (type == OBJT_SWAP)
252 		pctrie_init(&object->un_pager.swp.swp_blks);
253 
254 	/*
255 	 * Ensure that swap_pager_swapoff() iteration over object_list
256 	 * sees up to date type and pctrie head if it observed
257 	 * non-dead object.
258 	 */
259 	atomic_thread_fence_rel();
260 
261 	object->pg_color = 0;
262 	object->flags = flags;
263 	object->size = size;
264 	object->domain.dr_policy = NULL;
265 	object->generation = 1;
266 	object->cleangeneration = 1;
267 	refcount_init(&object->ref_count, 1);
268 	object->memattr = VM_MEMATTR_DEFAULT;
269 	object->cred = NULL;
270 	object->charge = 0;
271 	object->handle = handle;
272 	object->backing_object = NULL;
273 	object->backing_object_offset = (vm_ooffset_t) 0;
274 #if VM_NRESERVLEVEL > 0
275 	LIST_INIT(&object->rvq);
276 #endif
277 	umtx_shm_object_init(object);
278 }
279 
280 /*
281  *	vm_object_init:
282  *
283  *	Initialize the VM objects module.
284  */
285 void
286 vm_object_init(void)
287 {
288 	TAILQ_INIT(&vm_object_list);
289 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
290 
291 	rw_init(&kernel_object->lock, "kernel vm object");
292 	_vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
293 	    VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
294 #if VM_NRESERVLEVEL > 0
295 	kernel_object->flags |= OBJ_COLORED;
296 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
297 #endif
298 
299 	/*
300 	 * The lock portion of struct vm_object must be type stable due
301 	 * to vm_pageout_fallback_object_lock locking a vm object
302 	 * without holding any references to it.
303 	 */
304 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
305 #ifdef INVARIANTS
306 	    vm_object_zdtor,
307 #else
308 	    NULL,
309 #endif
310 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
311 
312 	vm_radix_zinit();
313 }
314 
315 void
316 vm_object_clear_flag(vm_object_t object, u_short bits)
317 {
318 
319 	VM_OBJECT_ASSERT_WLOCKED(object);
320 	object->flags &= ~bits;
321 }
322 
323 /*
324  *	Sets the default memory attribute for the specified object.  Pages
325  *	that are allocated to this object are by default assigned this memory
326  *	attribute.
327  *
328  *	Presently, this function must be called before any pages are allocated
329  *	to the object.  In the future, this requirement may be relaxed for
330  *	"default" and "swap" objects.
331  */
332 int
333 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
334 {
335 
336 	VM_OBJECT_ASSERT_WLOCKED(object);
337 	switch (object->type) {
338 	case OBJT_DEFAULT:
339 	case OBJT_DEVICE:
340 	case OBJT_MGTDEVICE:
341 	case OBJT_PHYS:
342 	case OBJT_SG:
343 	case OBJT_SWAP:
344 	case OBJT_VNODE:
345 		if (!TAILQ_EMPTY(&object->memq))
346 			return (KERN_FAILURE);
347 		break;
348 	case OBJT_DEAD:
349 		return (KERN_INVALID_ARGUMENT);
350 	default:
351 		panic("vm_object_set_memattr: object %p is of undefined type",
352 		    object);
353 	}
354 	object->memattr = memattr;
355 	return (KERN_SUCCESS);
356 }
357 
358 void
359 vm_object_pip_add(vm_object_t object, short i)
360 {
361 
362 	refcount_acquiren(&object->paging_in_progress, i);
363 }
364 
365 void
366 vm_object_pip_wakeup(vm_object_t object)
367 {
368 
369 	refcount_release(&object->paging_in_progress);
370 }
371 
372 void
373 vm_object_pip_wakeupn(vm_object_t object, short i)
374 {
375 
376 	refcount_releasen(&object->paging_in_progress, i);
377 }
378 
379 void
380 vm_object_pip_wait(vm_object_t object, char *waitid)
381 {
382 
383 	VM_OBJECT_ASSERT_WLOCKED(object);
384 
385 	while (REFCOUNT_COUNT(object->paging_in_progress) > 0) {
386 		VM_OBJECT_WUNLOCK(object);
387 		refcount_wait(&object->paging_in_progress, waitid, PVM);
388 		VM_OBJECT_WLOCK(object);
389 	}
390 }
391 
392 void
393 vm_object_pip_wait_unlocked(vm_object_t object, char *waitid)
394 {
395 
396 	VM_OBJECT_ASSERT_UNLOCKED(object);
397 
398 	while (REFCOUNT_COUNT(object->paging_in_progress) > 0)
399 		refcount_wait(&object->paging_in_progress, waitid, PVM);
400 }
401 
402 /*
403  *	vm_object_allocate:
404  *
405  *	Returns a new object with the given size.
406  */
407 vm_object_t
408 vm_object_allocate(objtype_t type, vm_pindex_t size)
409 {
410 	vm_object_t object;
411 	u_short flags;
412 
413 	switch (type) {
414 	case OBJT_DEAD:
415 		panic("vm_object_allocate: can't create OBJT_DEAD");
416 	case OBJT_DEFAULT:
417 	case OBJT_SWAP:
418 		flags = OBJ_COLORED;
419 		break;
420 	case OBJT_DEVICE:
421 	case OBJT_SG:
422 		flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
423 		break;
424 	case OBJT_MGTDEVICE:
425 		flags = OBJ_FICTITIOUS;
426 		break;
427 	case OBJT_PHYS:
428 		flags = OBJ_UNMANAGED;
429 		break;
430 	case OBJT_VNODE:
431 		flags = 0;
432 		break;
433 	default:
434 		panic("vm_object_allocate: type %d is undefined", type);
435 	}
436 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
437 	_vm_object_allocate(type, size, flags, object, NULL);
438 
439 	return (object);
440 }
441 
442 /*
443  *	vm_object_allocate_anon:
444  *
445  *	Returns a new default object of the given size and marked as
446  *	anonymous memory for special split/collapse handling.  Color
447  *	to be initialized by the caller.
448  */
449 vm_object_t
450 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
451     struct ucred *cred, vm_size_t charge)
452 {
453 	vm_object_t handle, object;
454 
455 	if (backing_object == NULL)
456 		handle = NULL;
457 	else if ((backing_object->flags & OBJ_ANON) != 0)
458 		handle = backing_object->handle;
459 	else
460 		handle = backing_object;
461 	object = uma_zalloc(obj_zone, M_WAITOK);
462 	_vm_object_allocate(OBJT_DEFAULT, size, OBJ_ANON | OBJ_ONEMAPPING,
463 	    object, handle);
464 	object->cred = cred;
465 	object->charge = cred != NULL ? charge : 0;
466 	return (object);
467 }
468 
469 
470 /*
471  *	vm_object_reference:
472  *
473  *	Gets another reference to the given object.  Note: OBJ_DEAD
474  *	objects can be referenced during final cleaning.
475  */
476 void
477 vm_object_reference(vm_object_t object)
478 {
479 	struct vnode *vp;
480 	u_int old;
481 
482 	if (object == NULL)
483 		return;
484 
485 	/*
486 	 * Many places assume exclusive access to objects with a single
487 	 * ref. vm_object_collapse() in particular will directly mainpulate
488 	 * references for objects in this state.  vnode objects only need
489 	 * the lock for the first ref to reference the vnode.
490 	 */
491 	if (!refcount_acquire_if_gt(&object->ref_count,
492 	    object->type == OBJT_VNODE ? 0 : 1)) {
493 		VM_OBJECT_RLOCK(object);
494 		old = refcount_acquire(&object->ref_count);
495 		if (object->type == OBJT_VNODE && old == 0) {
496 			vp = object->handle;
497 			vref(vp);
498 		}
499 		VM_OBJECT_RUNLOCK(object);
500 	}
501 }
502 
503 /*
504  *	vm_object_reference_locked:
505  *
506  *	Gets another reference to the given object.
507  *
508  *	The object must be locked.
509  */
510 void
511 vm_object_reference_locked(vm_object_t object)
512 {
513 	struct vnode *vp;
514 	u_int old;
515 
516 	VM_OBJECT_ASSERT_LOCKED(object);
517 	old = refcount_acquire(&object->ref_count);
518 	if (object->type == OBJT_VNODE && old == 0) {
519 		vp = object->handle;
520 		vref(vp);
521 	}
522 }
523 
524 /*
525  * Handle deallocating an object of type OBJT_VNODE.
526  */
527 static void
528 vm_object_vndeallocate(vm_object_t object)
529 {
530 	struct vnode *vp = (struct vnode *) object->handle;
531 	bool last;
532 
533 	KASSERT(object->type == OBJT_VNODE,
534 	    ("vm_object_vndeallocate: not a vnode object"));
535 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
536 
537 	/* Object lock to protect handle lookup. */
538 	last = refcount_release(&object->ref_count);
539 	VM_OBJECT_RUNLOCK(object);
540 
541 	if (!last)
542 		return;
543 
544 	if (!umtx_shm_vnobj_persistent)
545 		umtx_shm_object_terminated(object);
546 
547 	/* vrele may need the vnode lock. */
548 	vrele(vp);
549 }
550 
551 /*
552  *	vm_object_deallocate:
553  *
554  *	Release a reference to the specified object,
555  *	gained either through a vm_object_allocate
556  *	or a vm_object_reference call.  When all references
557  *	are gone, storage associated with this object
558  *	may be relinquished.
559  *
560  *	No object may be locked.
561  */
562 void
563 vm_object_deallocate(vm_object_t object)
564 {
565 	vm_object_t robject, temp;
566 	bool released;
567 
568 	while (object != NULL) {
569 		/*
570 		 * If the reference count goes to 0 we start calling
571 		 * vm_object_terminate() on the object chain.  A ref count
572 		 * of 1 may be a special case depending on the shadow count
573 		 * being 0 or 1.  These cases require a write lock on the
574 		 * object.
575 		 */
576 		if ((object->flags & OBJ_ANON) == 0)
577 			released = refcount_release_if_gt(&object->ref_count, 1);
578 		else
579 			released = refcount_release_if_gt(&object->ref_count, 2);
580 		if (released)
581 			return;
582 
583 		if (object->type == OBJT_VNODE) {
584 			VM_OBJECT_RLOCK(object);
585 			if (object->type == OBJT_VNODE) {
586 				vm_object_vndeallocate(object);
587 				return;
588 			}
589 			VM_OBJECT_RUNLOCK(object);
590 		}
591 
592 		VM_OBJECT_WLOCK(object);
593 		KASSERT(object->ref_count > 0,
594 		    ("vm_object_deallocate: object deallocated too many times: %d",
595 		    object->type));
596 
597 		if (refcount_release(&object->ref_count))
598 			goto doterm;
599 		if (object->ref_count > 1) {
600 			VM_OBJECT_WUNLOCK(object);
601 			return;
602 		} else if (object->ref_count == 1) {
603 			if (object->shadow_count == 0 &&
604 			    (object->flags & OBJ_ANON) != 0) {
605 				vm_object_set_flag(object, OBJ_ONEMAPPING);
606 			} else if (object->shadow_count == 1) {
607 				KASSERT((object->flags & OBJ_ANON) != 0,
608 				    ("obj %p with shadow_count > 0 is not anon",
609 				    object));
610 				robject = LIST_FIRST(&object->shadow_head);
611 				KASSERT(robject != NULL,
612 				    ("vm_object_deallocate: ref_count: %d, "
613 				    "shadow_count: %d", object->ref_count,
614 				    object->shadow_count));
615 				KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0,
616 				    ("shadowed tmpfs v_object %p", object));
617 				if (!VM_OBJECT_TRYWLOCK(robject)) {
618 					/*
619 					 * Avoid a potential deadlock.
620 					 */
621 					refcount_acquire(&object->ref_count);
622 					VM_OBJECT_WUNLOCK(object);
623 					/*
624 					 * More likely than not the thread
625 					 * holding robject's lock has lower
626 					 * priority than the current thread.
627 					 * Let the lower priority thread run.
628 					 */
629 					pause("vmo_de", 1);
630 					continue;
631 				}
632 				/*
633 				 * Collapse object into its shadow unless its
634 				 * shadow is dead.  In that case, object will
635 				 * be deallocated by the thread that is
636 				 * deallocating its shadow.
637 				 */
638 				if ((robject->flags &
639 				    (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON) {
640 
641 					refcount_acquire(&robject->ref_count);
642 retry:
643 					if (REFCOUNT_COUNT(robject->paging_in_progress) > 0) {
644 						VM_OBJECT_WUNLOCK(object);
645 						vm_object_pip_wait(robject,
646 						    "objde1");
647 						temp = robject->backing_object;
648 						if (object == temp) {
649 							VM_OBJECT_WLOCK(object);
650 							goto retry;
651 						}
652 					} else if (REFCOUNT_COUNT(object->paging_in_progress) > 0) {
653 						VM_OBJECT_WUNLOCK(robject);
654 						VM_OBJECT_WUNLOCK(object);
655 						refcount_wait(
656 						    &object->paging_in_progress,
657 						    "objde2", PVM);
658 						VM_OBJECT_WLOCK(robject);
659 						temp = robject->backing_object;
660 						if (object == temp) {
661 							VM_OBJECT_WLOCK(object);
662 							goto retry;
663 						}
664 					} else
665 						VM_OBJECT_WUNLOCK(object);
666 
667 					if (robject->ref_count == 1) {
668 						refcount_release(&robject->ref_count);
669 						object = robject;
670 						goto doterm;
671 					}
672 					object = robject;
673 					vm_object_collapse(object);
674 					VM_OBJECT_WUNLOCK(object);
675 					continue;
676 				}
677 				VM_OBJECT_WUNLOCK(robject);
678 			}
679 			VM_OBJECT_WUNLOCK(object);
680 			return;
681 		}
682 doterm:
683 		umtx_shm_object_terminated(object);
684 		temp = object->backing_object;
685 		if (temp != NULL) {
686 			KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
687 			    ("shadowed tmpfs v_object 2 %p", object));
688 			vm_object_backing_remove(object);
689 		}
690 		/*
691 		 * Don't double-terminate, we could be in a termination
692 		 * recursion due to the terminate having to sync data
693 		 * to disk.
694 		 */
695 		if ((object->flags & OBJ_DEAD) == 0) {
696 			vm_object_set_flag(object, OBJ_DEAD);
697 			vm_object_terminate(object);
698 		} else
699 			VM_OBJECT_WUNLOCK(object);
700 		object = temp;
701 	}
702 }
703 
704 /*
705  *	vm_object_destroy removes the object from the global object list
706  *      and frees the space for the object.
707  */
708 void
709 vm_object_destroy(vm_object_t object)
710 {
711 
712 	/*
713 	 * Release the allocation charge.
714 	 */
715 	if (object->cred != NULL) {
716 		swap_release_by_cred(object->charge, object->cred);
717 		object->charge = 0;
718 		crfree(object->cred);
719 		object->cred = NULL;
720 	}
721 
722 	/*
723 	 * Free the space for the object.
724 	 */
725 	uma_zfree(obj_zone, object);
726 }
727 
728 static void
729 vm_object_backing_remove_locked(vm_object_t object)
730 {
731 	vm_object_t backing_object;
732 
733 	backing_object = object->backing_object;
734 	VM_OBJECT_ASSERT_WLOCKED(object);
735 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
736 
737 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
738 		LIST_REMOVE(object, shadow_list);
739 		backing_object->shadow_count--;
740 		object->flags &= ~OBJ_SHADOWLIST;
741 	}
742 	object->backing_object = NULL;
743 }
744 
745 static void
746 vm_object_backing_remove(vm_object_t object)
747 {
748 	vm_object_t backing_object;
749 
750 	VM_OBJECT_ASSERT_WLOCKED(object);
751 
752 	if ((object->flags & OBJ_SHADOWLIST) != 0) {
753 		backing_object = object->backing_object;
754 		VM_OBJECT_WLOCK(backing_object);
755 		vm_object_backing_remove_locked(object);
756 		VM_OBJECT_WUNLOCK(backing_object);
757 	} else
758 		object->backing_object = NULL;
759 }
760 
761 static void
762 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
763 {
764 
765 	VM_OBJECT_ASSERT_WLOCKED(object);
766 
767 	if ((backing_object->flags & OBJ_ANON) != 0) {
768 		VM_OBJECT_ASSERT_WLOCKED(backing_object);
769 		LIST_INSERT_HEAD(&backing_object->shadow_head, object,
770 		    shadow_list);
771 		backing_object->shadow_count++;
772 		object->flags |= OBJ_SHADOWLIST;
773 	}
774 	object->backing_object = backing_object;
775 }
776 
777 static void
778 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
779 {
780 
781 	VM_OBJECT_ASSERT_WLOCKED(object);
782 
783 	if ((backing_object->flags & OBJ_ANON) != 0) {
784 		VM_OBJECT_WLOCK(backing_object);
785 		vm_object_backing_insert_locked(object, backing_object);
786 		VM_OBJECT_WUNLOCK(backing_object);
787 	} else
788 		object->backing_object = backing_object;
789 }
790 
791 
792 /*
793  *	vm_object_terminate_pages removes any remaining pageable pages
794  *	from the object and resets the object to an empty state.
795  */
796 static void
797 vm_object_terminate_pages(vm_object_t object)
798 {
799 	vm_page_t p, p_next;
800 
801 	VM_OBJECT_ASSERT_WLOCKED(object);
802 
803 	/*
804 	 * Free any remaining pageable pages.  This also removes them from the
805 	 * paging queues.  However, don't free wired pages, just remove them
806 	 * from the object.  Rather than incrementally removing each page from
807 	 * the object, the page and object are reset to any empty state.
808 	 */
809 	TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
810 		vm_page_assert_unbusied(p);
811 		KASSERT(p->object == object &&
812 		    (p->ref_count & VPRC_OBJREF) != 0,
813 		    ("vm_object_terminate_pages: page %p is inconsistent", p));
814 
815 		p->object = NULL;
816 		if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
817 			VM_CNT_INC(v_pfree);
818 			vm_page_free(p);
819 		}
820 	}
821 
822 	/*
823 	 * If the object contained any pages, then reset it to an empty state.
824 	 * None of the object's fields, including "resident_page_count", were
825 	 * modified by the preceding loop.
826 	 */
827 	if (object->resident_page_count != 0) {
828 		vm_radix_reclaim_allnodes(&object->rtree);
829 		TAILQ_INIT(&object->memq);
830 		object->resident_page_count = 0;
831 		if (object->type == OBJT_VNODE)
832 			vdrop(object->handle);
833 	}
834 }
835 
836 /*
837  *	vm_object_terminate actually destroys the specified object, freeing
838  *	up all previously used resources.
839  *
840  *	The object must be locked.
841  *	This routine may block.
842  */
843 void
844 vm_object_terminate(vm_object_t object)
845 {
846 	VM_OBJECT_ASSERT_WLOCKED(object);
847 	KASSERT((object->flags & OBJ_DEAD) != 0,
848 	    ("terminating non-dead obj %p", object));
849 
850 	/*
851 	 * wait for the pageout daemon to be done with the object
852 	 */
853 	vm_object_pip_wait(object, "objtrm");
854 
855 	KASSERT(!REFCOUNT_COUNT(object->paging_in_progress),
856 		("vm_object_terminate: pageout in progress"));
857 
858 	KASSERT(object->ref_count == 0,
859 		("vm_object_terminate: object with references, ref_count=%d",
860 		object->ref_count));
861 
862 	if ((object->flags & OBJ_PG_DTOR) == 0)
863 		vm_object_terminate_pages(object);
864 
865 #if VM_NRESERVLEVEL > 0
866 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
867 		vm_reserv_break_all(object);
868 #endif
869 
870 	KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
871 	    object->type == OBJT_SWAP,
872 	    ("%s: non-swap obj %p has cred", __func__, object));
873 
874 	/*
875 	 * Let the pager know object is dead.
876 	 */
877 	vm_pager_deallocate(object);
878 	VM_OBJECT_WUNLOCK(object);
879 
880 	vm_object_destroy(object);
881 }
882 
883 /*
884  * Make the page read-only so that we can clear the object flags.  However, if
885  * this is a nosync mmap then the object is likely to stay dirty so do not
886  * mess with the page and do not clear the object flags.  Returns TRUE if the
887  * page should be flushed, and FALSE otherwise.
888  */
889 static boolean_t
890 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
891 {
892 
893 	vm_page_assert_busied(p);
894 
895 	/*
896 	 * If we have been asked to skip nosync pages and this is a
897 	 * nosync page, skip it.  Note that the object flags were not
898 	 * cleared in this case so we do not have to set them.
899 	 */
900 	if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
901 		*allclean = FALSE;
902 		return (FALSE);
903 	} else {
904 		pmap_remove_write(p);
905 		return (p->dirty != 0);
906 	}
907 }
908 
909 /*
910  *	vm_object_page_clean
911  *
912  *	Clean all dirty pages in the specified range of object.  Leaves page
913  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
914  *	write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
915  *	leaving the object dirty.
916  *
917  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
918  *	synchronous clustering mode implementation.
919  *
920  *	Odd semantics: if start == end, we clean everything.
921  *
922  *	The object must be locked.
923  *
924  *	Returns FALSE if some page from the range was not written, as
925  *	reported by the pager, and TRUE otherwise.
926  */
927 boolean_t
928 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
929     int flags)
930 {
931 	vm_page_t np, p;
932 	vm_pindex_t pi, tend, tstart;
933 	int curgeneration, n, pagerflags;
934 	boolean_t eio, res, allclean;
935 
936 	VM_OBJECT_ASSERT_WLOCKED(object);
937 
938 	if (object->type != OBJT_VNODE || !vm_object_mightbedirty(object) ||
939 	    object->resident_page_count == 0)
940 		return (TRUE);
941 
942 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
943 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
944 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
945 
946 	tstart = OFF_TO_IDX(start);
947 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
948 	allclean = tstart == 0 && tend >= object->size;
949 	res = TRUE;
950 
951 rescan:
952 	curgeneration = object->generation;
953 
954 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
955 		pi = p->pindex;
956 		if (pi >= tend)
957 			break;
958 		np = TAILQ_NEXT(p, listq);
959 		if (vm_page_none_valid(p))
960 			continue;
961 		if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
962 			if (object->generation != curgeneration &&
963 			    (flags & OBJPC_SYNC) != 0)
964 				goto rescan;
965 			np = vm_page_find_least(object, pi);
966 			continue;
967 		}
968 		if (!vm_object_page_remove_write(p, flags, &allclean)) {
969 			vm_page_xunbusy(p);
970 			continue;
971 		}
972 
973 		n = vm_object_page_collect_flush(object, p, pagerflags,
974 		    flags, &allclean, &eio);
975 		if (eio) {
976 			res = FALSE;
977 			allclean = FALSE;
978 		}
979 		if (object->generation != curgeneration &&
980 		    (flags & OBJPC_SYNC) != 0)
981 			goto rescan;
982 
983 		/*
984 		 * If the VOP_PUTPAGES() did a truncated write, so
985 		 * that even the first page of the run is not fully
986 		 * written, vm_pageout_flush() returns 0 as the run
987 		 * length.  Since the condition that caused truncated
988 		 * write may be permanent, e.g. exhausted free space,
989 		 * accepting n == 0 would cause an infinite loop.
990 		 *
991 		 * Forwarding the iterator leaves the unwritten page
992 		 * behind, but there is not much we can do there if
993 		 * filesystem refuses to write it.
994 		 */
995 		if (n == 0) {
996 			n = 1;
997 			allclean = FALSE;
998 		}
999 		np = vm_page_find_least(object, pi + n);
1000 	}
1001 #if 0
1002 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1003 #endif
1004 
1005 	if (allclean)
1006 		object->cleangeneration = curgeneration;
1007 	return (res);
1008 }
1009 
1010 static int
1011 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1012     int flags, boolean_t *allclean, boolean_t *eio)
1013 {
1014 	vm_page_t ma[vm_pageout_page_count], p_first, tp;
1015 	int count, i, mreq, runlen;
1016 
1017 	vm_page_lock_assert(p, MA_NOTOWNED);
1018 	vm_page_assert_xbusied(p);
1019 	VM_OBJECT_ASSERT_WLOCKED(object);
1020 
1021 	count = 1;
1022 	mreq = 0;
1023 
1024 	for (tp = p; count < vm_pageout_page_count; count++) {
1025 		tp = vm_page_next(tp);
1026 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1027 			break;
1028 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1029 			vm_page_xunbusy(tp);
1030 			break;
1031 		}
1032 	}
1033 
1034 	for (p_first = p; count < vm_pageout_page_count; count++) {
1035 		tp = vm_page_prev(p_first);
1036 		if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1037 			break;
1038 		if (!vm_object_page_remove_write(tp, flags, allclean)) {
1039 			vm_page_xunbusy(tp);
1040 			break;
1041 		}
1042 		p_first = tp;
1043 		mreq++;
1044 	}
1045 
1046 	for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1047 		ma[i] = tp;
1048 
1049 	vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1050 	return (runlen);
1051 }
1052 
1053 /*
1054  * Note that there is absolutely no sense in writing out
1055  * anonymous objects, so we track down the vnode object
1056  * to write out.
1057  * We invalidate (remove) all pages from the address space
1058  * for semantic correctness.
1059  *
1060  * If the backing object is a device object with unmanaged pages, then any
1061  * mappings to the specified range of pages must be removed before this
1062  * function is called.
1063  *
1064  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1065  * may start out with a NULL object.
1066  */
1067 boolean_t
1068 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1069     boolean_t syncio, boolean_t invalidate)
1070 {
1071 	vm_object_t backing_object;
1072 	struct vnode *vp;
1073 	struct mount *mp;
1074 	int error, flags, fsync_after;
1075 	boolean_t res;
1076 
1077 	if (object == NULL)
1078 		return (TRUE);
1079 	res = TRUE;
1080 	error = 0;
1081 	VM_OBJECT_WLOCK(object);
1082 	while ((backing_object = object->backing_object) != NULL) {
1083 		VM_OBJECT_WLOCK(backing_object);
1084 		offset += object->backing_object_offset;
1085 		VM_OBJECT_WUNLOCK(object);
1086 		object = backing_object;
1087 		if (object->size < OFF_TO_IDX(offset + size))
1088 			size = IDX_TO_OFF(object->size) - offset;
1089 	}
1090 	/*
1091 	 * Flush pages if writing is allowed, invalidate them
1092 	 * if invalidation requested.  Pages undergoing I/O
1093 	 * will be ignored by vm_object_page_remove().
1094 	 *
1095 	 * We cannot lock the vnode and then wait for paging
1096 	 * to complete without deadlocking against vm_fault.
1097 	 * Instead we simply call vm_object_page_remove() and
1098 	 * allow it to block internally on a page-by-page
1099 	 * basis when it encounters pages undergoing async
1100 	 * I/O.
1101 	 */
1102 	if (object->type == OBJT_VNODE &&
1103 	    vm_object_mightbedirty(object) != 0 &&
1104 	    ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1105 		VM_OBJECT_WUNLOCK(object);
1106 		(void) vn_start_write(vp, &mp, V_WAIT);
1107 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1108 		if (syncio && !invalidate && offset == 0 &&
1109 		    atop(size) == object->size) {
1110 			/*
1111 			 * If syncing the whole mapping of the file,
1112 			 * it is faster to schedule all the writes in
1113 			 * async mode, also allowing the clustering,
1114 			 * and then wait for i/o to complete.
1115 			 */
1116 			flags = 0;
1117 			fsync_after = TRUE;
1118 		} else {
1119 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1120 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1121 			fsync_after = FALSE;
1122 		}
1123 		VM_OBJECT_WLOCK(object);
1124 		res = vm_object_page_clean(object, offset, offset + size,
1125 		    flags);
1126 		VM_OBJECT_WUNLOCK(object);
1127 		if (fsync_after)
1128 			error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1129 		VOP_UNLOCK(vp, 0);
1130 		vn_finished_write(mp);
1131 		if (error != 0)
1132 			res = FALSE;
1133 		VM_OBJECT_WLOCK(object);
1134 	}
1135 	if ((object->type == OBJT_VNODE ||
1136 	     object->type == OBJT_DEVICE) && invalidate) {
1137 		if (object->type == OBJT_DEVICE)
1138 			/*
1139 			 * The option OBJPR_NOTMAPPED must be passed here
1140 			 * because vm_object_page_remove() cannot remove
1141 			 * unmanaged mappings.
1142 			 */
1143 			flags = OBJPR_NOTMAPPED;
1144 		else if (old_msync)
1145 			flags = 0;
1146 		else
1147 			flags = OBJPR_CLEANONLY;
1148 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1149 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1150 	}
1151 	VM_OBJECT_WUNLOCK(object);
1152 	return (res);
1153 }
1154 
1155 /*
1156  * Determine whether the given advice can be applied to the object.  Advice is
1157  * not applied to unmanaged pages since they never belong to page queues, and
1158  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1159  * have been mapped at most once.
1160  */
1161 static bool
1162 vm_object_advice_applies(vm_object_t object, int advice)
1163 {
1164 
1165 	if ((object->flags & OBJ_UNMANAGED) != 0)
1166 		return (false);
1167 	if (advice != MADV_FREE)
1168 		return (true);
1169 	return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1170 	    (OBJ_ONEMAPPING | OBJ_ANON));
1171 }
1172 
1173 static void
1174 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1175     vm_size_t size)
1176 {
1177 
1178 	if (advice == MADV_FREE && object->type == OBJT_SWAP)
1179 		swap_pager_freespace(object, pindex, size);
1180 }
1181 
1182 /*
1183  *	vm_object_madvise:
1184  *
1185  *	Implements the madvise function at the object/page level.
1186  *
1187  *	MADV_WILLNEED	(any object)
1188  *
1189  *	    Activate the specified pages if they are resident.
1190  *
1191  *	MADV_DONTNEED	(any object)
1192  *
1193  *	    Deactivate the specified pages if they are resident.
1194  *
1195  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1196  *			 OBJ_ONEMAPPING only)
1197  *
1198  *	    Deactivate and clean the specified pages if they are
1199  *	    resident.  This permits the process to reuse the pages
1200  *	    without faulting or the kernel to reclaim the pages
1201  *	    without I/O.
1202  */
1203 void
1204 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1205     int advice)
1206 {
1207 	vm_pindex_t tpindex;
1208 	vm_object_t backing_object, tobject;
1209 	vm_page_t m, tm;
1210 
1211 	if (object == NULL)
1212 		return;
1213 
1214 relookup:
1215 	VM_OBJECT_WLOCK(object);
1216 	if (!vm_object_advice_applies(object, advice)) {
1217 		VM_OBJECT_WUNLOCK(object);
1218 		return;
1219 	}
1220 	for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1221 		tobject = object;
1222 
1223 		/*
1224 		 * If the next page isn't resident in the top-level object, we
1225 		 * need to search the shadow chain.  When applying MADV_FREE, we
1226 		 * take care to release any swap space used to store
1227 		 * non-resident pages.
1228 		 */
1229 		if (m == NULL || pindex < m->pindex) {
1230 			/*
1231 			 * Optimize a common case: if the top-level object has
1232 			 * no backing object, we can skip over the non-resident
1233 			 * range in constant time.
1234 			 */
1235 			if (object->backing_object == NULL) {
1236 				tpindex = (m != NULL && m->pindex < end) ?
1237 				    m->pindex : end;
1238 				vm_object_madvise_freespace(object, advice,
1239 				    pindex, tpindex - pindex);
1240 				if ((pindex = tpindex) == end)
1241 					break;
1242 				goto next_page;
1243 			}
1244 
1245 			tpindex = pindex;
1246 			do {
1247 				vm_object_madvise_freespace(tobject, advice,
1248 				    tpindex, 1);
1249 				/*
1250 				 * Prepare to search the next object in the
1251 				 * chain.
1252 				 */
1253 				backing_object = tobject->backing_object;
1254 				if (backing_object == NULL)
1255 					goto next_pindex;
1256 				VM_OBJECT_WLOCK(backing_object);
1257 				tpindex +=
1258 				    OFF_TO_IDX(tobject->backing_object_offset);
1259 				if (tobject != object)
1260 					VM_OBJECT_WUNLOCK(tobject);
1261 				tobject = backing_object;
1262 				if (!vm_object_advice_applies(tobject, advice))
1263 					goto next_pindex;
1264 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
1265 			    NULL);
1266 		} else {
1267 next_page:
1268 			tm = m;
1269 			m = TAILQ_NEXT(m, listq);
1270 		}
1271 
1272 		/*
1273 		 * If the page is not in a normal state, skip it.  The page
1274 		 * can not be invalidated while the object lock is held.
1275 		 */
1276 		if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1277 			goto next_pindex;
1278 		KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1279 		    ("vm_object_madvise: page %p is fictitious", tm));
1280 		KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1281 		    ("vm_object_madvise: page %p is not managed", tm));
1282 		if (vm_page_tryxbusy(tm) == 0) {
1283 			if (object != tobject)
1284 				VM_OBJECT_WUNLOCK(object);
1285 			if (advice == MADV_WILLNEED) {
1286 				/*
1287 				 * Reference the page before unlocking and
1288 				 * sleeping so that the page daemon is less
1289 				 * likely to reclaim it.
1290 				 */
1291 				vm_page_aflag_set(tm, PGA_REFERENCED);
1292 			}
1293 			vm_page_busy_sleep(tm, "madvpo", false);
1294   			goto relookup;
1295 		}
1296 		vm_page_lock(tm);
1297 		vm_page_advise(tm, advice);
1298 		vm_page_unlock(tm);
1299 		vm_page_xunbusy(tm);
1300 		vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1301 next_pindex:
1302 		if (tobject != object)
1303 			VM_OBJECT_WUNLOCK(tobject);
1304 	}
1305 	VM_OBJECT_WUNLOCK(object);
1306 }
1307 
1308 /*
1309  *	vm_object_shadow:
1310  *
1311  *	Create a new object which is backed by the
1312  *	specified existing object range.  The source
1313  *	object reference is deallocated.
1314  *
1315  *	The new object and offset into that object
1316  *	are returned in the source parameters.
1317  */
1318 void
1319 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1320     struct ucred *cred, bool shared)
1321 {
1322 	vm_object_t source;
1323 	vm_object_t result;
1324 
1325 	source = *object;
1326 
1327 	/*
1328 	 * Don't create the new object if the old object isn't shared.
1329 	 *
1330 	 * If we hold the only reference we can guarantee that it won't
1331 	 * increase while we have the map locked.  Otherwise the race is
1332 	 * harmless and we will end up with an extra shadow object that
1333 	 * will be collapsed later.
1334 	 */
1335 	if (source != NULL && source->ref_count == 1 &&
1336 	    (source->flags & OBJ_ANON) != 0)
1337 		return;
1338 
1339 	/*
1340 	 * Allocate a new object with the given length.
1341 	 */
1342 	result = vm_object_allocate_anon(atop(length), source, cred, length);
1343 
1344 	/*
1345 	 * Store the offset into the source object, and fix up the offset into
1346 	 * the new object.
1347 	 */
1348 	result->backing_object_offset = *offset;
1349 
1350 	if (shared || source != NULL) {
1351 		VM_OBJECT_WLOCK(result);
1352 
1353 		/*
1354 		 * The new object shadows the source object, adding a
1355 		 * reference to it.  Our caller changes his reference
1356 		 * to point to the new object, removing a reference to
1357 		 * the source object.  Net result: no change of
1358 		 * reference count, unless the caller needs to add one
1359 		 * more reference due to forking a shared map entry.
1360 		 */
1361 		if (shared) {
1362 			vm_object_reference_locked(result);
1363 			vm_object_clear_flag(result, OBJ_ONEMAPPING);
1364 		}
1365 
1366 		/*
1367 		 * Try to optimize the result object's page color when
1368 		 * shadowing in order to maintain page coloring
1369 		 * consistency in the combined shadowed object.
1370 		 */
1371 		if (source != NULL) {
1372 			vm_object_backing_insert(result, source);
1373 			result->domain = source->domain;
1374 #if VM_NRESERVLEVEL > 0
1375 			result->flags |= source->flags & OBJ_COLORED;
1376 			result->pg_color = (source->pg_color +
1377 			    OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1378 			    1)) - 1);
1379 #endif
1380 		}
1381 		VM_OBJECT_WUNLOCK(result);
1382 	}
1383 
1384 	/*
1385 	 * Return the new things
1386 	 */
1387 	*offset = 0;
1388 	*object = result;
1389 }
1390 
1391 /*
1392  *	vm_object_split:
1393  *
1394  * Split the pages in a map entry into a new object.  This affords
1395  * easier removal of unused pages, and keeps object inheritance from
1396  * being a negative impact on memory usage.
1397  */
1398 void
1399 vm_object_split(vm_map_entry_t entry)
1400 {
1401 	vm_page_t m, m_next;
1402 	vm_object_t orig_object, new_object, source;
1403 	vm_pindex_t idx, offidxstart;
1404 	vm_size_t size;
1405 
1406 	orig_object = entry->object.vm_object;
1407 	if ((orig_object->flags & OBJ_ANON) == 0)
1408 		return;
1409 	if (orig_object->ref_count <= 1)
1410 		return;
1411 	VM_OBJECT_WUNLOCK(orig_object);
1412 
1413 	offidxstart = OFF_TO_IDX(entry->offset);
1414 	size = atop(entry->end - entry->start);
1415 
1416 	/*
1417 	 * If swap_pager_copy() is later called, it will convert new_object
1418 	 * into a swap object.
1419 	 */
1420 	new_object = vm_object_allocate_anon(size, orig_object,
1421 	    orig_object->cred, ptoa(size));
1422 
1423 	/*
1424 	 * At this point, the new object is still private, so the order in
1425 	 * which the original and new objects are locked does not matter.
1426 	 */
1427 	VM_OBJECT_WLOCK(new_object);
1428 	VM_OBJECT_WLOCK(orig_object);
1429 	new_object->domain = orig_object->domain;
1430 	source = orig_object->backing_object;
1431 	if (source != NULL) {
1432 		if ((source->flags & (OBJ_ANON | OBJ_DEAD)) != 0) {
1433 			VM_OBJECT_WLOCK(source);
1434 			if ((source->flags & OBJ_DEAD) != 0) {
1435 				VM_OBJECT_WUNLOCK(source);
1436 				VM_OBJECT_WUNLOCK(orig_object);
1437 				VM_OBJECT_WUNLOCK(new_object);
1438 				new_object->cred = NULL;
1439 				vm_object_deallocate(new_object);
1440 				VM_OBJECT_WLOCK(orig_object);
1441 				return;
1442 			}
1443 			vm_object_backing_insert_locked(new_object, source);
1444 			vm_object_reference_locked(source);	/* for new_object */
1445 			vm_object_clear_flag(source, OBJ_ONEMAPPING);
1446 			VM_OBJECT_WUNLOCK(source);
1447 		} else {
1448 			vm_object_backing_insert(new_object, source);
1449 			vm_object_reference(source);
1450 		}
1451 		new_object->backing_object_offset =
1452 			orig_object->backing_object_offset + entry->offset;
1453 	}
1454 	if (orig_object->cred != NULL) {
1455 		crhold(orig_object->cred);
1456 		KASSERT(orig_object->charge >= ptoa(size),
1457 		    ("orig_object->charge < 0"));
1458 		orig_object->charge -= ptoa(size);
1459 	}
1460 retry:
1461 	m = vm_page_find_least(orig_object, offidxstart);
1462 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1463 	    m = m_next) {
1464 		m_next = TAILQ_NEXT(m, listq);
1465 
1466 		/*
1467 		 * We must wait for pending I/O to complete before we can
1468 		 * rename the page.
1469 		 *
1470 		 * We do not have to VM_PROT_NONE the page as mappings should
1471 		 * not be changed by this operation.
1472 		 */
1473 		if (vm_page_tryxbusy(m) == 0) {
1474 			VM_OBJECT_WUNLOCK(new_object);
1475 			vm_page_sleep_if_busy(m, "spltwt");
1476 			VM_OBJECT_WLOCK(new_object);
1477 			goto retry;
1478 		}
1479 
1480 		/*
1481 		 * The page was left invalid.  Likely placed there by
1482 		 * an incomplete fault.  Just remove and ignore.
1483 		 */
1484 		if (vm_page_none_valid(m)) {
1485 			if (vm_page_remove(m))
1486 				vm_page_free(m);
1487 			continue;
1488 		}
1489 
1490 		/* vm_page_rename() will dirty the page. */
1491 		if (vm_page_rename(m, new_object, idx)) {
1492 			vm_page_xunbusy(m);
1493 			VM_OBJECT_WUNLOCK(new_object);
1494 			VM_OBJECT_WUNLOCK(orig_object);
1495 			vm_radix_wait();
1496 			VM_OBJECT_WLOCK(orig_object);
1497 			VM_OBJECT_WLOCK(new_object);
1498 			goto retry;
1499 		}
1500 
1501 #if VM_NRESERVLEVEL > 0
1502 		/*
1503 		 * If some of the reservation's allocated pages remain with
1504 		 * the original object, then transferring the reservation to
1505 		 * the new object is neither particularly beneficial nor
1506 		 * particularly harmful as compared to leaving the reservation
1507 		 * with the original object.  If, however, all of the
1508 		 * reservation's allocated pages are transferred to the new
1509 		 * object, then transferring the reservation is typically
1510 		 * beneficial.  Determining which of these two cases applies
1511 		 * would be more costly than unconditionally renaming the
1512 		 * reservation.
1513 		 */
1514 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1515 #endif
1516 		if (orig_object->type != OBJT_SWAP)
1517 			vm_page_xunbusy(m);
1518 	}
1519 	if (orig_object->type == OBJT_SWAP) {
1520 		/*
1521 		 * swap_pager_copy() can sleep, in which case the orig_object's
1522 		 * and new_object's locks are released and reacquired.
1523 		 */
1524 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1525 		TAILQ_FOREACH(m, &new_object->memq, listq)
1526 			vm_page_xunbusy(m);
1527 	}
1528 	VM_OBJECT_WUNLOCK(orig_object);
1529 	VM_OBJECT_WUNLOCK(new_object);
1530 	entry->object.vm_object = new_object;
1531 	entry->offset = 0LL;
1532 	vm_object_deallocate(orig_object);
1533 	VM_OBJECT_WLOCK(new_object);
1534 }
1535 
1536 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1537 #define	OBSC_COLLAPSE_WAIT	0x0004
1538 
1539 static vm_page_t
1540 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next,
1541     int op)
1542 {
1543 	vm_object_t backing_object;
1544 
1545 	VM_OBJECT_ASSERT_WLOCKED(object);
1546 	backing_object = object->backing_object;
1547 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1548 
1549 	KASSERT(p == NULL || p->object == object || p->object == backing_object,
1550 	    ("invalid ownership %p %p %p", p, object, backing_object));
1551 	if ((op & OBSC_COLLAPSE_NOWAIT) != 0)
1552 		return (next);
1553 	/* The page is only NULL when rename fails. */
1554 	if (p == NULL) {
1555 		VM_OBJECT_WUNLOCK(object);
1556 		VM_OBJECT_WUNLOCK(backing_object);
1557 		vm_radix_wait();
1558 	} else {
1559 		if (p->object == object)
1560 			VM_OBJECT_WUNLOCK(backing_object);
1561 		else
1562 			VM_OBJECT_WUNLOCK(object);
1563 		vm_page_busy_sleep(p, "vmocol", false);
1564 	}
1565 	VM_OBJECT_WLOCK(object);
1566 	VM_OBJECT_WLOCK(backing_object);
1567 	return (TAILQ_FIRST(&backing_object->memq));
1568 }
1569 
1570 static bool
1571 vm_object_scan_all_shadowed(vm_object_t object)
1572 {
1573 	vm_object_t backing_object;
1574 	vm_page_t p, pp;
1575 	vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1576 
1577 	VM_OBJECT_ASSERT_WLOCKED(object);
1578 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1579 
1580 	backing_object = object->backing_object;
1581 
1582 	if ((backing_object->flags & OBJ_ANON) == 0)
1583 		return (false);
1584 
1585 	pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1586 	p = vm_page_find_least(backing_object, pi);
1587 	ps = swap_pager_find_least(backing_object, pi);
1588 
1589 	/*
1590 	 * Only check pages inside the parent object's range and
1591 	 * inside the parent object's mapping of the backing object.
1592 	 */
1593 	for (;; pi++) {
1594 		if (p != NULL && p->pindex < pi)
1595 			p = TAILQ_NEXT(p, listq);
1596 		if (ps < pi)
1597 			ps = swap_pager_find_least(backing_object, pi);
1598 		if (p == NULL && ps >= backing_object->size)
1599 			break;
1600 		else if (p == NULL)
1601 			pi = ps;
1602 		else
1603 			pi = MIN(p->pindex, ps);
1604 
1605 		new_pindex = pi - backing_offset_index;
1606 		if (new_pindex >= object->size)
1607 			break;
1608 
1609 		/*
1610 		 * See if the parent has the page or if the parent's object
1611 		 * pager has the page.  If the parent has the page but the page
1612 		 * is not valid, the parent's object pager must have the page.
1613 		 *
1614 		 * If this fails, the parent does not completely shadow the
1615 		 * object and we might as well give up now.
1616 		 */
1617 		pp = vm_page_lookup(object, new_pindex);
1618 		/*
1619 		 * The valid check here is stable due to object lock being
1620 		 * required to clear valid and initiate paging.
1621 		 */
1622 		if ((pp == NULL || vm_page_none_valid(pp)) &&
1623 		    !vm_pager_has_page(object, new_pindex, NULL, NULL))
1624 			return (false);
1625 	}
1626 	return (true);
1627 }
1628 
1629 static bool
1630 vm_object_collapse_scan(vm_object_t object, int op)
1631 {
1632 	vm_object_t backing_object;
1633 	vm_page_t next, p, pp;
1634 	vm_pindex_t backing_offset_index, new_pindex;
1635 
1636 	VM_OBJECT_ASSERT_WLOCKED(object);
1637 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1638 
1639 	backing_object = object->backing_object;
1640 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1641 
1642 	/*
1643 	 * Initial conditions
1644 	 */
1645 	if ((op & OBSC_COLLAPSE_WAIT) != 0)
1646 		vm_object_set_flag(backing_object, OBJ_DEAD);
1647 
1648 	/*
1649 	 * Our scan
1650 	 */
1651 	for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1652 		next = TAILQ_NEXT(p, listq);
1653 		new_pindex = p->pindex - backing_offset_index;
1654 
1655 		/*
1656 		 * Check for busy page
1657 		 */
1658 		if (vm_page_tryxbusy(p) == 0) {
1659 			next = vm_object_collapse_scan_wait(object, p, next, op);
1660 			continue;
1661 		}
1662 
1663 		KASSERT(p->object == backing_object,
1664 		    ("vm_object_collapse_scan: object mismatch"));
1665 
1666 		if (p->pindex < backing_offset_index ||
1667 		    new_pindex >= object->size) {
1668 			if (backing_object->type == OBJT_SWAP)
1669 				swap_pager_freespace(backing_object, p->pindex,
1670 				    1);
1671 
1672 			KASSERT(!pmap_page_is_mapped(p),
1673 			    ("freeing mapped page %p", p));
1674 			if (vm_page_remove(p))
1675 				vm_page_free(p);
1676 			continue;
1677 		}
1678 
1679 		pp = vm_page_lookup(object, new_pindex);
1680 		if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1681 			vm_page_xunbusy(p);
1682 			/*
1683 			 * The page in the parent is busy and possibly not
1684 			 * (yet) valid.  Until its state is finalized by the
1685 			 * busy bit owner, we can't tell whether it shadows the
1686 			 * original page.  Therefore, we must either skip it
1687 			 * and the original (backing_object) page or wait for
1688 			 * its state to be finalized.
1689 			 *
1690 			 * This is due to a race with vm_fault() where we must
1691 			 * unbusy the original (backing_obj) page before we can
1692 			 * (re)lock the parent.  Hence we can get here.
1693 			 */
1694 			next = vm_object_collapse_scan_wait(object, pp, next,
1695 			    op);
1696 			continue;
1697 		}
1698 
1699 		if (pp != NULL && vm_page_none_valid(pp)) {
1700 			/*
1701 			 * The page was invalid in the parent.  Likely placed
1702 			 * there by an incomplete fault.  Just remove and
1703 			 * ignore.  p can replace it.
1704 			 */
1705 			if (vm_page_remove(pp))
1706 				vm_page_free(pp);
1707 			pp = NULL;
1708 		}
1709 
1710 		if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1711 			NULL)) {
1712 			/*
1713 			 * The page already exists in the parent OR swap exists
1714 			 * for this location in the parent.  Leave the parent's
1715 			 * page alone.  Destroy the original page from the
1716 			 * backing object.
1717 			 */
1718 			if (backing_object->type == OBJT_SWAP)
1719 				swap_pager_freespace(backing_object, p->pindex,
1720 				    1);
1721 			KASSERT(!pmap_page_is_mapped(p),
1722 			    ("freeing mapped page %p", p));
1723 			if (vm_page_remove(p))
1724 				vm_page_free(p);
1725 			if (pp != NULL)
1726 				vm_page_xunbusy(pp);
1727 			continue;
1728 		}
1729 
1730 		/*
1731 		 * Page does not exist in parent, rename the page from the
1732 		 * backing object to the main object.
1733 		 *
1734 		 * If the page was mapped to a process, it can remain mapped
1735 		 * through the rename.  vm_page_rename() will dirty the page.
1736 		 */
1737 		if (vm_page_rename(p, object, new_pindex)) {
1738 			vm_page_xunbusy(p);
1739 			if (pp != NULL)
1740 				vm_page_xunbusy(pp);
1741 			next = vm_object_collapse_scan_wait(object, NULL, next,
1742 			    op);
1743 			continue;
1744 		}
1745 
1746 		/* Use the old pindex to free the right page. */
1747 		if (backing_object->type == OBJT_SWAP)
1748 			swap_pager_freespace(backing_object,
1749 			    new_pindex + backing_offset_index, 1);
1750 
1751 #if VM_NRESERVLEVEL > 0
1752 		/*
1753 		 * Rename the reservation.
1754 		 */
1755 		vm_reserv_rename(p, object, backing_object,
1756 		    backing_offset_index);
1757 #endif
1758 		vm_page_xunbusy(p);
1759 	}
1760 	return (true);
1761 }
1762 
1763 
1764 /*
1765  * this version of collapse allows the operation to occur earlier and
1766  * when paging_in_progress is true for an object...  This is not a complete
1767  * operation, but should plug 99.9% of the rest of the leaks.
1768  */
1769 static void
1770 vm_object_qcollapse(vm_object_t object)
1771 {
1772 	vm_object_t backing_object = object->backing_object;
1773 
1774 	VM_OBJECT_ASSERT_WLOCKED(object);
1775 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1776 
1777 	if (backing_object->ref_count != 1)
1778 		return;
1779 
1780 	vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT);
1781 }
1782 
1783 /*
1784  *	vm_object_collapse:
1785  *
1786  *	Collapse an object with the object backing it.
1787  *	Pages in the backing object are moved into the
1788  *	parent, and the backing object is deallocated.
1789  */
1790 void
1791 vm_object_collapse(vm_object_t object)
1792 {
1793 	vm_object_t backing_object, new_backing_object;
1794 
1795 	VM_OBJECT_ASSERT_WLOCKED(object);
1796 
1797 	while (TRUE) {
1798 		/*
1799 		 * Verify that the conditions are right for collapse:
1800 		 *
1801 		 * The object exists and the backing object exists.
1802 		 */
1803 		if ((backing_object = object->backing_object) == NULL)
1804 			break;
1805 
1806 		/*
1807 		 * we check the backing object first, because it is most likely
1808 		 * not collapsable.
1809 		 */
1810 		if ((backing_object->flags & OBJ_ANON) == 0)
1811 			break;
1812 		VM_OBJECT_WLOCK(backing_object);
1813 		if ((backing_object->flags & OBJ_DEAD) != 0 ||
1814 		    (object->flags & (OBJ_DEAD | OBJ_ANON)) != OBJ_ANON) {
1815 			VM_OBJECT_WUNLOCK(backing_object);
1816 			break;
1817 		}
1818 
1819 		if (REFCOUNT_COUNT(object->paging_in_progress) > 0 ||
1820 		    REFCOUNT_COUNT(backing_object->paging_in_progress) > 0) {
1821 			vm_object_qcollapse(object);
1822 			VM_OBJECT_WUNLOCK(backing_object);
1823 			break;
1824 		}
1825 
1826 		/*
1827 		 * We know that we can either collapse the backing object (if
1828 		 * the parent is the only reference to it) or (perhaps) have
1829 		 * the parent bypass the object if the parent happens to shadow
1830 		 * all the resident pages in the entire backing object.
1831 		 *
1832 		 * This is ignoring pager-backed pages such as swap pages.
1833 		 * vm_object_collapse_scan fails the shadowing test in this
1834 		 * case.
1835 		 */
1836 		if (backing_object->ref_count == 1) {
1837 			vm_object_pip_add(object, 1);
1838 			vm_object_pip_add(backing_object, 1);
1839 
1840 			/*
1841 			 * If there is exactly one reference to the backing
1842 			 * object, we can collapse it into the parent.
1843 			 */
1844 			vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT);
1845 
1846 #if VM_NRESERVLEVEL > 0
1847 			/*
1848 			 * Break any reservations from backing_object.
1849 			 */
1850 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1851 				vm_reserv_break_all(backing_object);
1852 #endif
1853 
1854 			/*
1855 			 * Move the pager from backing_object to object.
1856 			 */
1857 			if (backing_object->type == OBJT_SWAP) {
1858 				/*
1859 				 * swap_pager_copy() can sleep, in which case
1860 				 * the backing_object's and object's locks are
1861 				 * released and reacquired.
1862 				 * Since swap_pager_copy() is being asked to
1863 				 * destroy the source, it will change the
1864 				 * backing_object's type to OBJT_DEFAULT.
1865 				 */
1866 				swap_pager_copy(
1867 				    backing_object,
1868 				    object,
1869 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1870 			}
1871 			/*
1872 			 * Object now shadows whatever backing_object did.
1873 			 * Note that the reference to
1874 			 * backing_object->backing_object moves from within
1875 			 * backing_object to within object.
1876 			 */
1877 			vm_object_backing_remove_locked(object);
1878 			new_backing_object = backing_object->backing_object;
1879 			if (new_backing_object != NULL) {
1880 				VM_OBJECT_WLOCK(new_backing_object);
1881 				vm_object_backing_remove_locked(backing_object);
1882 				vm_object_backing_insert_locked(object,
1883 				    new_backing_object);
1884 				VM_OBJECT_WUNLOCK(new_backing_object);
1885 			}
1886 			object->backing_object_offset +=
1887 			    backing_object->backing_object_offset;
1888 
1889 			/*
1890 			 * Discard backing_object.
1891 			 *
1892 			 * Since the backing object has no pages, no pager left,
1893 			 * and no object references within it, all that is
1894 			 * necessary is to dispose of it.
1895 			 */
1896 			KASSERT(backing_object->ref_count == 1, (
1897 "backing_object %p was somehow re-referenced during collapse!",
1898 			    backing_object));
1899 			vm_object_pip_wakeup(backing_object);
1900 			backing_object->type = OBJT_DEAD;
1901 			refcount_release(&backing_object->ref_count);
1902 			VM_OBJECT_WUNLOCK(backing_object);
1903 			vm_object_destroy(backing_object);
1904 
1905 			vm_object_pip_wakeup(object);
1906 			counter_u64_add(object_collapses, 1);
1907 		} else {
1908 			/*
1909 			 * If we do not entirely shadow the backing object,
1910 			 * there is nothing we can do so we give up.
1911 			 */
1912 			if (object->resident_page_count != object->size &&
1913 			    !vm_object_scan_all_shadowed(object)) {
1914 				VM_OBJECT_WUNLOCK(backing_object);
1915 				break;
1916 			}
1917 
1918 			/*
1919 			 * Make the parent shadow the next object in the
1920 			 * chain.  Deallocating backing_object will not remove
1921 			 * it, since its reference count is at least 2.
1922 			 */
1923 			vm_object_backing_remove_locked(object);
1924 
1925 			new_backing_object = backing_object->backing_object;
1926 			if (new_backing_object != NULL) {
1927 				vm_object_backing_insert(object,
1928 				    new_backing_object);
1929 				vm_object_reference(new_backing_object);
1930 				object->backing_object_offset +=
1931 					backing_object->backing_object_offset;
1932 			}
1933 
1934 			/*
1935 			 * Drop the reference count on backing_object. Since
1936 			 * its ref_count was at least 2, it will not vanish.
1937 			 */
1938 			refcount_release(&backing_object->ref_count);
1939 			VM_OBJECT_WUNLOCK(backing_object);
1940 			counter_u64_add(object_bypasses, 1);
1941 		}
1942 
1943 		/*
1944 		 * Try again with this object's new backing object.
1945 		 */
1946 	}
1947 }
1948 
1949 /*
1950  *	vm_object_page_remove:
1951  *
1952  *	For the given object, either frees or invalidates each of the
1953  *	specified pages.  In general, a page is freed.  However, if a page is
1954  *	wired for any reason other than the existence of a managed, wired
1955  *	mapping, then it may be invalidated but not removed from the object.
1956  *	Pages are specified by the given range ["start", "end") and the option
1957  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
1958  *	extends from "start" to the end of the object.  If the option
1959  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1960  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
1961  *	specified, then the pages within the specified range must have no
1962  *	mappings.  Otherwise, if this option is not specified, any mappings to
1963  *	the specified pages are removed before the pages are freed or
1964  *	invalidated.
1965  *
1966  *	In general, this operation should only be performed on objects that
1967  *	contain managed pages.  There are, however, two exceptions.  First, it
1968  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
1969  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1970  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
1971  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
1972  *
1973  *	The object must be locked.
1974  */
1975 void
1976 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1977     int options)
1978 {
1979 	vm_page_t p, next;
1980 
1981 	VM_OBJECT_ASSERT_WLOCKED(object);
1982 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1983 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1984 	    ("vm_object_page_remove: illegal options for object %p", object));
1985 	if (object->resident_page_count == 0)
1986 		return;
1987 	vm_object_pip_add(object, 1);
1988 again:
1989 	p = vm_page_find_least(object, start);
1990 
1991 	/*
1992 	 * Here, the variable "p" is either (1) the page with the least pindex
1993 	 * greater than or equal to the parameter "start" or (2) NULL.
1994 	 */
1995 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1996 		next = TAILQ_NEXT(p, listq);
1997 
1998 		/*
1999 		 * If the page is wired for any reason besides the existence
2000 		 * of managed, wired mappings, then it cannot be freed.  For
2001 		 * example, fictitious pages, which represent device memory,
2002 		 * are inherently wired and cannot be freed.  They can,
2003 		 * however, be invalidated if the option OBJPR_CLEANONLY is
2004 		 * not specified.
2005 		 */
2006 		if (vm_page_tryxbusy(p) == 0) {
2007 			vm_page_sleep_if_busy(p, "vmopar");
2008 			goto again;
2009 		}
2010 		if (vm_page_wired(p)) {
2011 wired:
2012 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2013 			    object->ref_count != 0)
2014 				pmap_remove_all(p);
2015 			if ((options & OBJPR_CLEANONLY) == 0) {
2016 				vm_page_invalid(p);
2017 				vm_page_undirty(p);
2018 			}
2019 			vm_page_xunbusy(p);
2020 			continue;
2021 		}
2022 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2023 		    ("vm_object_page_remove: page %p is fictitious", p));
2024 		if ((options & OBJPR_CLEANONLY) != 0 &&
2025 		    !vm_page_none_valid(p)) {
2026 			if ((options & OBJPR_NOTMAPPED) == 0 &&
2027 			    object->ref_count != 0 &&
2028 			    !vm_page_try_remove_write(p))
2029 				goto wired;
2030 			if (p->dirty != 0) {
2031 				vm_page_xunbusy(p);
2032 				continue;
2033 			}
2034 		}
2035 		if ((options & OBJPR_NOTMAPPED) == 0 &&
2036 		    object->ref_count != 0 && !vm_page_try_remove_all(p))
2037 			goto wired;
2038 		vm_page_free(p);
2039 	}
2040 	vm_object_pip_wakeup(object);
2041 }
2042 
2043 /*
2044  *	vm_object_page_noreuse:
2045  *
2046  *	For the given object, attempt to move the specified pages to
2047  *	the head of the inactive queue.  This bypasses regular LRU
2048  *	operation and allows the pages to be reused quickly under memory
2049  *	pressure.  If a page is wired for any reason, then it will not
2050  *	be queued.  Pages are specified by the range ["start", "end").
2051  *	As a special case, if "end" is zero, then the range extends from
2052  *	"start" to the end of the object.
2053  *
2054  *	This operation should only be performed on objects that
2055  *	contain non-fictitious, managed pages.
2056  *
2057  *	The object must be locked.
2058  */
2059 void
2060 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2061 {
2062 	struct mtx *mtx;
2063 	vm_page_t p, next;
2064 
2065 	VM_OBJECT_ASSERT_LOCKED(object);
2066 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2067 	    ("vm_object_page_noreuse: illegal object %p", object));
2068 	if (object->resident_page_count == 0)
2069 		return;
2070 	p = vm_page_find_least(object, start);
2071 
2072 	/*
2073 	 * Here, the variable "p" is either (1) the page with the least pindex
2074 	 * greater than or equal to the parameter "start" or (2) NULL.
2075 	 */
2076 	mtx = NULL;
2077 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2078 		next = TAILQ_NEXT(p, listq);
2079 		vm_page_change_lock(p, &mtx);
2080 		vm_page_deactivate_noreuse(p);
2081 	}
2082 	if (mtx != NULL)
2083 		mtx_unlock(mtx);
2084 }
2085 
2086 /*
2087  *	Populate the specified range of the object with valid pages.  Returns
2088  *	TRUE if the range is successfully populated and FALSE otherwise.
2089  *
2090  *	Note: This function should be optimized to pass a larger array of
2091  *	pages to vm_pager_get_pages() before it is applied to a non-
2092  *	OBJT_DEVICE object.
2093  *
2094  *	The object must be locked.
2095  */
2096 boolean_t
2097 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2098 {
2099 	vm_page_t m;
2100 	vm_pindex_t pindex;
2101 	int rv;
2102 
2103 	VM_OBJECT_ASSERT_WLOCKED(object);
2104 	for (pindex = start; pindex < end; pindex++) {
2105 		rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2106 		if (rv != VM_PAGER_OK)
2107 			break;
2108 
2109 		/*
2110 		 * Keep "m" busy because a subsequent iteration may unlock
2111 		 * the object.
2112 		 */
2113 	}
2114 	if (pindex > start) {
2115 		m = vm_page_lookup(object, start);
2116 		while (m != NULL && m->pindex < pindex) {
2117 			vm_page_xunbusy(m);
2118 			m = TAILQ_NEXT(m, listq);
2119 		}
2120 	}
2121 	return (pindex == end);
2122 }
2123 
2124 /*
2125  *	Routine:	vm_object_coalesce
2126  *	Function:	Coalesces two objects backing up adjoining
2127  *			regions of memory into a single object.
2128  *
2129  *	returns TRUE if objects were combined.
2130  *
2131  *	NOTE:	Only works at the moment if the second object is NULL -
2132  *		if it's not, which object do we lock first?
2133  *
2134  *	Parameters:
2135  *		prev_object	First object to coalesce
2136  *		prev_offset	Offset into prev_object
2137  *		prev_size	Size of reference to prev_object
2138  *		next_size	Size of reference to the second object
2139  *		reserved	Indicator that extension region has
2140  *				swap accounted for
2141  *
2142  *	Conditions:
2143  *	The object must *not* be locked.
2144  */
2145 boolean_t
2146 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2147     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2148 {
2149 	vm_pindex_t next_pindex;
2150 
2151 	if (prev_object == NULL)
2152 		return (TRUE);
2153 	if ((prev_object->flags & OBJ_ANON) == 0)
2154 		return (FALSE);
2155 
2156 	VM_OBJECT_WLOCK(prev_object);
2157 	/*
2158 	 * Try to collapse the object first
2159 	 */
2160 	vm_object_collapse(prev_object);
2161 
2162 	/*
2163 	 * Can't coalesce if: . more than one reference . paged out . shadows
2164 	 * another object . has a copy elsewhere (any of which mean that the
2165 	 * pages not mapped to prev_entry may be in use anyway)
2166 	 */
2167 	if (prev_object->backing_object != NULL) {
2168 		VM_OBJECT_WUNLOCK(prev_object);
2169 		return (FALSE);
2170 	}
2171 
2172 	prev_size >>= PAGE_SHIFT;
2173 	next_size >>= PAGE_SHIFT;
2174 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2175 
2176 	if (prev_object->ref_count > 1 &&
2177 	    prev_object->size != next_pindex &&
2178 	    (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2179 		VM_OBJECT_WUNLOCK(prev_object);
2180 		return (FALSE);
2181 	}
2182 
2183 	/*
2184 	 * Account for the charge.
2185 	 */
2186 	if (prev_object->cred != NULL) {
2187 
2188 		/*
2189 		 * If prev_object was charged, then this mapping,
2190 		 * although not charged now, may become writable
2191 		 * later. Non-NULL cred in the object would prevent
2192 		 * swap reservation during enabling of the write
2193 		 * access, so reserve swap now. Failed reservation
2194 		 * cause allocation of the separate object for the map
2195 		 * entry, and swap reservation for this entry is
2196 		 * managed in appropriate time.
2197 		 */
2198 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2199 		    prev_object->cred)) {
2200 			VM_OBJECT_WUNLOCK(prev_object);
2201 			return (FALSE);
2202 		}
2203 		prev_object->charge += ptoa(next_size);
2204 	}
2205 
2206 	/*
2207 	 * Remove any pages that may still be in the object from a previous
2208 	 * deallocation.
2209 	 */
2210 	if (next_pindex < prev_object->size) {
2211 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2212 		    next_size, 0);
2213 		if (prev_object->type == OBJT_SWAP)
2214 			swap_pager_freespace(prev_object,
2215 					     next_pindex, next_size);
2216 #if 0
2217 		if (prev_object->cred != NULL) {
2218 			KASSERT(prev_object->charge >=
2219 			    ptoa(prev_object->size - next_pindex),
2220 			    ("object %p overcharged 1 %jx %jx", prev_object,
2221 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2222 			prev_object->charge -= ptoa(prev_object->size -
2223 			    next_pindex);
2224 		}
2225 #endif
2226 	}
2227 
2228 	/*
2229 	 * Extend the object if necessary.
2230 	 */
2231 	if (next_pindex + next_size > prev_object->size)
2232 		prev_object->size = next_pindex + next_size;
2233 
2234 	VM_OBJECT_WUNLOCK(prev_object);
2235 	return (TRUE);
2236 }
2237 
2238 void
2239 vm_object_set_writeable_dirty(vm_object_t object)
2240 {
2241 
2242 	/* Only set for vnodes & tmpfs */
2243 	if (object->type != OBJT_VNODE &&
2244 	    (object->flags & OBJ_TMPFS_NODE) == 0)
2245 		return;
2246 	atomic_add_int(&object->generation, 1);
2247 }
2248 
2249 /*
2250  *	vm_object_unwire:
2251  *
2252  *	For each page offset within the specified range of the given object,
2253  *	find the highest-level page in the shadow chain and unwire it.  A page
2254  *	must exist at every page offset, and the highest-level page must be
2255  *	wired.
2256  */
2257 void
2258 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2259     uint8_t queue)
2260 {
2261 	vm_object_t tobject, t1object;
2262 	vm_page_t m, tm;
2263 	vm_pindex_t end_pindex, pindex, tpindex;
2264 	int depth, locked_depth;
2265 
2266 	KASSERT((offset & PAGE_MASK) == 0,
2267 	    ("vm_object_unwire: offset is not page aligned"));
2268 	KASSERT((length & PAGE_MASK) == 0,
2269 	    ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2270 	/* The wired count of a fictitious page never changes. */
2271 	if ((object->flags & OBJ_FICTITIOUS) != 0)
2272 		return;
2273 	pindex = OFF_TO_IDX(offset);
2274 	end_pindex = pindex + atop(length);
2275 again:
2276 	locked_depth = 1;
2277 	VM_OBJECT_RLOCK(object);
2278 	m = vm_page_find_least(object, pindex);
2279 	while (pindex < end_pindex) {
2280 		if (m == NULL || pindex < m->pindex) {
2281 			/*
2282 			 * The first object in the shadow chain doesn't
2283 			 * contain a page at the current index.  Therefore,
2284 			 * the page must exist in a backing object.
2285 			 */
2286 			tobject = object;
2287 			tpindex = pindex;
2288 			depth = 0;
2289 			do {
2290 				tpindex +=
2291 				    OFF_TO_IDX(tobject->backing_object_offset);
2292 				tobject = tobject->backing_object;
2293 				KASSERT(tobject != NULL,
2294 				    ("vm_object_unwire: missing page"));
2295 				if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2296 					goto next_page;
2297 				depth++;
2298 				if (depth == locked_depth) {
2299 					locked_depth++;
2300 					VM_OBJECT_RLOCK(tobject);
2301 				}
2302 			} while ((tm = vm_page_lookup(tobject, tpindex)) ==
2303 			    NULL);
2304 		} else {
2305 			tm = m;
2306 			m = TAILQ_NEXT(m, listq);
2307 		}
2308 		if (vm_page_trysbusy(tm) == 0) {
2309 			for (tobject = object; locked_depth >= 1;
2310 			    locked_depth--) {
2311 				t1object = tobject->backing_object;
2312 				if (tm->object != tobject)
2313 					VM_OBJECT_RUNLOCK(tobject);
2314 				tobject = t1object;
2315 			}
2316 			vm_page_busy_sleep(tm, "unwbo", true);
2317 			goto again;
2318 		}
2319 		vm_page_unwire(tm, queue);
2320 		vm_page_sunbusy(tm);
2321 next_page:
2322 		pindex++;
2323 	}
2324 	/* Release the accumulated object locks. */
2325 	for (tobject = object; locked_depth >= 1; locked_depth--) {
2326 		t1object = tobject->backing_object;
2327 		VM_OBJECT_RUNLOCK(tobject);
2328 		tobject = t1object;
2329 	}
2330 }
2331 
2332 /*
2333  * Return the vnode for the given object, or NULL if none exists.
2334  * For tmpfs objects, the function may return NULL if there is
2335  * no vnode allocated at the time of the call.
2336  */
2337 struct vnode *
2338 vm_object_vnode(vm_object_t object)
2339 {
2340 	struct vnode *vp;
2341 
2342 	VM_OBJECT_ASSERT_LOCKED(object);
2343 	if (object->type == OBJT_VNODE) {
2344 		vp = object->handle;
2345 		KASSERT(vp != NULL, ("%s: OBJT_VNODE has no vnode", __func__));
2346 	} else if (object->type == OBJT_SWAP &&
2347 	    (object->flags & OBJ_TMPFS) != 0) {
2348 		vp = object->un_pager.swp.swp_tmpfs;
2349 		KASSERT(vp != NULL, ("%s: OBJT_TMPFS has no vnode", __func__));
2350 	} else {
2351 		vp = NULL;
2352 	}
2353 	return (vp);
2354 }
2355 
2356 
2357 /*
2358  * Busy the vm object.  This prevents new pages belonging to the object from
2359  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2360  * for checking page state before proceeding.
2361  */
2362 void
2363 vm_object_busy(vm_object_t obj)
2364 {
2365 
2366 	VM_OBJECT_ASSERT_LOCKED(obj);
2367 
2368 	refcount_acquire(&obj->busy);
2369 	/* The fence is required to order loads of page busy. */
2370 	atomic_thread_fence_acq_rel();
2371 }
2372 
2373 void
2374 vm_object_unbusy(vm_object_t obj)
2375 {
2376 
2377 
2378 	refcount_release(&obj->busy);
2379 }
2380 
2381 void
2382 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2383 {
2384 
2385 	VM_OBJECT_ASSERT_UNLOCKED(obj);
2386 
2387 	if (obj->busy)
2388 		refcount_sleep(&obj->busy, wmesg, PVM);
2389 }
2390 
2391 /*
2392  * Return the kvme type of the given object.
2393  * If vpp is not NULL, set it to the object's vm_object_vnode() or NULL.
2394  */
2395 int
2396 vm_object_kvme_type(vm_object_t object, struct vnode **vpp)
2397 {
2398 
2399 	VM_OBJECT_ASSERT_LOCKED(object);
2400 	if (vpp != NULL)
2401 		*vpp = vm_object_vnode(object);
2402 	switch (object->type) {
2403 	case OBJT_DEFAULT:
2404 		return (KVME_TYPE_DEFAULT);
2405 	case OBJT_VNODE:
2406 		return (KVME_TYPE_VNODE);
2407 	case OBJT_SWAP:
2408 		if ((object->flags & OBJ_TMPFS_NODE) != 0)
2409 			return (KVME_TYPE_VNODE);
2410 		return (KVME_TYPE_SWAP);
2411 	case OBJT_DEVICE:
2412 		return (KVME_TYPE_DEVICE);
2413 	case OBJT_PHYS:
2414 		return (KVME_TYPE_PHYS);
2415 	case OBJT_DEAD:
2416 		return (KVME_TYPE_DEAD);
2417 	case OBJT_SG:
2418 		return (KVME_TYPE_SG);
2419 	case OBJT_MGTDEVICE:
2420 		return (KVME_TYPE_MGTDEVICE);
2421 	default:
2422 		return (KVME_TYPE_UNKNOWN);
2423 	}
2424 }
2425 
2426 static int
2427 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2428 {
2429 	struct kinfo_vmobject *kvo;
2430 	char *fullpath, *freepath;
2431 	struct vnode *vp;
2432 	struct vattr va;
2433 	vm_object_t obj;
2434 	vm_page_t m;
2435 	int count, error;
2436 
2437 	if (req->oldptr == NULL) {
2438 		/*
2439 		 * If an old buffer has not been provided, generate an
2440 		 * estimate of the space needed for a subsequent call.
2441 		 */
2442 		mtx_lock(&vm_object_list_mtx);
2443 		count = 0;
2444 		TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2445 			if (obj->type == OBJT_DEAD)
2446 				continue;
2447 			count++;
2448 		}
2449 		mtx_unlock(&vm_object_list_mtx);
2450 		return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2451 		    count * 11 / 10));
2452 	}
2453 
2454 	kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK);
2455 	error = 0;
2456 
2457 	/*
2458 	 * VM objects are type stable and are never removed from the
2459 	 * list once added.  This allows us to safely read obj->object_list
2460 	 * after reacquiring the VM object lock.
2461 	 */
2462 	mtx_lock(&vm_object_list_mtx);
2463 	TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2464 		if (obj->type == OBJT_DEAD)
2465 			continue;
2466 		VM_OBJECT_RLOCK(obj);
2467 		if (obj->type == OBJT_DEAD) {
2468 			VM_OBJECT_RUNLOCK(obj);
2469 			continue;
2470 		}
2471 		mtx_unlock(&vm_object_list_mtx);
2472 		kvo->kvo_size = ptoa(obj->size);
2473 		kvo->kvo_resident = obj->resident_page_count;
2474 		kvo->kvo_ref_count = obj->ref_count;
2475 		kvo->kvo_shadow_count = obj->shadow_count;
2476 		kvo->kvo_memattr = obj->memattr;
2477 		kvo->kvo_active = 0;
2478 		kvo->kvo_inactive = 0;
2479 		TAILQ_FOREACH(m, &obj->memq, listq) {
2480 			/*
2481 			 * A page may belong to the object but be
2482 			 * dequeued and set to PQ_NONE while the
2483 			 * object lock is not held.  This makes the
2484 			 * reads of m->queue below racy, and we do not
2485 			 * count pages set to PQ_NONE.  However, this
2486 			 * sysctl is only meant to give an
2487 			 * approximation of the system anyway.
2488 			 */
2489 			if (m->a.queue == PQ_ACTIVE)
2490 				kvo->kvo_active++;
2491 			else if (m->a.queue == PQ_INACTIVE)
2492 				kvo->kvo_inactive++;
2493 		}
2494 
2495 		kvo->kvo_vn_fileid = 0;
2496 		kvo->kvo_vn_fsid = 0;
2497 		kvo->kvo_vn_fsid_freebsd11 = 0;
2498 		freepath = NULL;
2499 		fullpath = "";
2500 		kvo->kvo_type = vm_object_kvme_type(obj, &vp);
2501 		if (vp != NULL)
2502 			vref(vp);
2503 		VM_OBJECT_RUNLOCK(obj);
2504 		if (vp != NULL) {
2505 			vn_fullpath(curthread, vp, &fullpath, &freepath);
2506 			vn_lock(vp, LK_SHARED | LK_RETRY);
2507 			if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2508 				kvo->kvo_vn_fileid = va.va_fileid;
2509 				kvo->kvo_vn_fsid = va.va_fsid;
2510 				kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2511 								/* truncate */
2512 			}
2513 			vput(vp);
2514 		}
2515 
2516 		strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2517 		if (freepath != NULL)
2518 			free(freepath, M_TEMP);
2519 
2520 		/* Pack record size down */
2521 		kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2522 		    + strlen(kvo->kvo_path) + 1;
2523 		kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2524 		    sizeof(uint64_t));
2525 		error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2526 		mtx_lock(&vm_object_list_mtx);
2527 		if (error)
2528 			break;
2529 	}
2530 	mtx_unlock(&vm_object_list_mtx);
2531 	free(kvo, M_TEMP);
2532 	return (error);
2533 }
2534 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2535     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2536     "List of VM objects");
2537 
2538 #include "opt_ddb.h"
2539 #ifdef DDB
2540 #include <sys/kernel.h>
2541 
2542 #include <sys/cons.h>
2543 
2544 #include <ddb/ddb.h>
2545 
2546 static int
2547 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2548 {
2549 	vm_map_t tmpm;
2550 	vm_map_entry_t tmpe;
2551 	vm_object_t obj;
2552 
2553 	if (map == 0)
2554 		return 0;
2555 
2556 	if (entry == 0) {
2557 		VM_MAP_ENTRY_FOREACH(tmpe, map) {
2558 			if (_vm_object_in_map(map, object, tmpe)) {
2559 				return 1;
2560 			}
2561 		}
2562 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2563 		tmpm = entry->object.sub_map;
2564 		VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2565 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2566 				return 1;
2567 			}
2568 		}
2569 	} else if ((obj = entry->object.vm_object) != NULL) {
2570 		for (; obj; obj = obj->backing_object)
2571 			if (obj == object) {
2572 				return 1;
2573 			}
2574 	}
2575 	return 0;
2576 }
2577 
2578 static int
2579 vm_object_in_map(vm_object_t object)
2580 {
2581 	struct proc *p;
2582 
2583 	/* sx_slock(&allproc_lock); */
2584 	FOREACH_PROC_IN_SYSTEM(p) {
2585 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2586 			continue;
2587 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2588 			/* sx_sunlock(&allproc_lock); */
2589 			return 1;
2590 		}
2591 	}
2592 	/* sx_sunlock(&allproc_lock); */
2593 	if (_vm_object_in_map(kernel_map, object, 0))
2594 		return 1;
2595 	return 0;
2596 }
2597 
2598 DB_SHOW_COMMAND(vmochk, vm_object_check)
2599 {
2600 	vm_object_t object;
2601 
2602 	/*
2603 	 * make sure that internal objs are in a map somewhere
2604 	 * and none have zero ref counts.
2605 	 */
2606 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2607 		if ((object->flags & OBJ_ANON) != 0) {
2608 			if (object->ref_count == 0) {
2609 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2610 					(long)object->size);
2611 			}
2612 			if (!vm_object_in_map(object)) {
2613 				db_printf(
2614 			"vmochk: internal obj is not in a map: "
2615 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2616 				    object->ref_count, (u_long)object->size,
2617 				    (u_long)object->size,
2618 				    (void *)object->backing_object);
2619 			}
2620 		}
2621 	}
2622 }
2623 
2624 /*
2625  *	vm_object_print:	[ debug ]
2626  */
2627 DB_SHOW_COMMAND(object, vm_object_print_static)
2628 {
2629 	/* XXX convert args. */
2630 	vm_object_t object = (vm_object_t)addr;
2631 	boolean_t full = have_addr;
2632 
2633 	vm_page_t p;
2634 
2635 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2636 #define	count	was_count
2637 
2638 	int count;
2639 
2640 	if (object == NULL)
2641 		return;
2642 
2643 	db_iprintf(
2644 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2645 	    object, (int)object->type, (uintmax_t)object->size,
2646 	    object->resident_page_count, object->ref_count, object->flags,
2647 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2648 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2649 	    object->shadow_count,
2650 	    object->backing_object ? object->backing_object->ref_count : 0,
2651 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2652 
2653 	if (!full)
2654 		return;
2655 
2656 	db_indent += 2;
2657 	count = 0;
2658 	TAILQ_FOREACH(p, &object->memq, listq) {
2659 		if (count == 0)
2660 			db_iprintf("memory:=");
2661 		else if (count == 6) {
2662 			db_printf("\n");
2663 			db_iprintf(" ...");
2664 			count = 0;
2665 		} else
2666 			db_printf(",");
2667 		count++;
2668 
2669 		db_printf("(off=0x%jx,page=0x%jx)",
2670 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2671 	}
2672 	if (count != 0)
2673 		db_printf("\n");
2674 	db_indent -= 2;
2675 }
2676 
2677 /* XXX. */
2678 #undef count
2679 
2680 /* XXX need this non-static entry for calling from vm_map_print. */
2681 void
2682 vm_object_print(
2683         /* db_expr_t */ long addr,
2684 	boolean_t have_addr,
2685 	/* db_expr_t */ long count,
2686 	char *modif)
2687 {
2688 	vm_object_print_static(addr, have_addr, count, modif);
2689 }
2690 
2691 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2692 {
2693 	vm_object_t object;
2694 	vm_pindex_t fidx;
2695 	vm_paddr_t pa;
2696 	vm_page_t m, prev_m;
2697 	int rcount, nl, c;
2698 
2699 	nl = 0;
2700 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2701 		db_printf("new object: %p\n", (void *)object);
2702 		if (nl > 18) {
2703 			c = cngetc();
2704 			if (c != ' ')
2705 				return;
2706 			nl = 0;
2707 		}
2708 		nl++;
2709 		rcount = 0;
2710 		fidx = 0;
2711 		pa = -1;
2712 		TAILQ_FOREACH(m, &object->memq, listq) {
2713 			if (m->pindex > 128)
2714 				break;
2715 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2716 			    prev_m->pindex + 1 != m->pindex) {
2717 				if (rcount) {
2718 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2719 						(long)fidx, rcount, (long)pa);
2720 					if (nl > 18) {
2721 						c = cngetc();
2722 						if (c != ' ')
2723 							return;
2724 						nl = 0;
2725 					}
2726 					nl++;
2727 					rcount = 0;
2728 				}
2729 			}
2730 			if (rcount &&
2731 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2732 				++rcount;
2733 				continue;
2734 			}
2735 			if (rcount) {
2736 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2737 					(long)fidx, rcount, (long)pa);
2738 				if (nl > 18) {
2739 					c = cngetc();
2740 					if (c != ' ')
2741 						return;
2742 					nl = 0;
2743 				}
2744 				nl++;
2745 			}
2746 			fidx = m->pindex;
2747 			pa = VM_PAGE_TO_PHYS(m);
2748 			rcount = 1;
2749 		}
2750 		if (rcount) {
2751 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2752 				(long)fidx, rcount, (long)pa);
2753 			if (nl > 18) {
2754 				c = cngetc();
2755 				if (c != ' ')
2756 					return;
2757 				nl = 0;
2758 			}
2759 			nl++;
2760 		}
2761 	}
2762 }
2763 #endif /* DDB */
2764