xref: /freebsd/sys/vm/vm_object.c (revision c99fb5f907eccde7b4c6d53e650e1a32a558f0d8)
1 /*-
2  * Copyright (c) 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * The Mach Operating System project at Carnegie-Mellon University.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 4. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	from: @(#)vm_object.c	8.5 (Berkeley) 3/22/94
33  *
34  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39  *
40  * Permission to use, copy, modify and distribute this software and
41  * its documentation is hereby granted, provided that both the copyright
42  * notice and this permission notice appear in all copies of the
43  * software, derivative works or modified versions, and any portions
44  * thereof, and that both notices appear in supporting documentation.
45  *
46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49  *
50  * Carnegie Mellon requests users of this software to return to
51  *
52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
53  *  School of Computer Science
54  *  Carnegie Mellon University
55  *  Pittsburgh PA 15213-3890
56  *
57  * any improvements or extensions that they make and grant Carnegie the
58  * rights to redistribute these changes.
59  */
60 
61 /*
62  *	Virtual memory object module.
63  */
64 
65 #include <sys/cdefs.h>
66 __FBSDID("$FreeBSD$");
67 
68 #include "opt_vm.h"
69 
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/lock.h>
73 #include <sys/mman.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/mutex.h>
78 #include <sys/proc.h>		/* for curproc, pageproc */
79 #include <sys/socket.h>
80 #include <sys/resourcevar.h>
81 #include <sys/rwlock.h>
82 #include <sys/vnode.h>
83 #include <sys/vmmeter.h>
84 #include <sys/sx.h>
85 
86 #include <vm/vm.h>
87 #include <vm/vm_param.h>
88 #include <vm/pmap.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pageout.h>
93 #include <vm/vm_pager.h>
94 #include <vm/swap_pager.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_extern.h>
97 #include <vm/vm_radix.h>
98 #include <vm/vm_reserv.h>
99 #include <vm/uma.h>
100 
101 static int old_msync;
102 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
103     "Use old (insecure) msync behavior");
104 
105 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
106 		    int pagerflags, int flags, boolean_t *clearobjflags,
107 		    boolean_t *eio);
108 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
109 		    boolean_t *clearobjflags);
110 static void	vm_object_qcollapse(vm_object_t object);
111 static void	vm_object_vndeallocate(vm_object_t object);
112 
113 /*
114  *	Virtual memory objects maintain the actual data
115  *	associated with allocated virtual memory.  A given
116  *	page of memory exists within exactly one object.
117  *
118  *	An object is only deallocated when all "references"
119  *	are given up.  Only one "reference" to a given
120  *	region of an object should be writeable.
121  *
122  *	Associated with each object is a list of all resident
123  *	memory pages belonging to that object; this list is
124  *	maintained by the "vm_page" module, and locked by the object's
125  *	lock.
126  *
127  *	Each object also records a "pager" routine which is
128  *	used to retrieve (and store) pages to the proper backing
129  *	storage.  In addition, objects may be backed by other
130  *	objects from which they were virtual-copied.
131  *
132  *	The only items within the object structure which are
133  *	modified after time of creation are:
134  *		reference count		locked by object's lock
135  *		pager routine		locked by object's lock
136  *
137  */
138 
139 struct object_q vm_object_list;
140 struct mtx vm_object_list_mtx;	/* lock for object list and count */
141 
142 struct vm_object kernel_object_store;
143 struct vm_object kmem_object_store;
144 
145 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
146     "VM object stats");
147 
148 static long object_collapses;
149 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
150     &object_collapses, 0, "VM object collapses");
151 
152 static long object_bypasses;
153 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
154     &object_bypasses, 0, "VM object bypasses");
155 
156 static uma_zone_t obj_zone;
157 
158 static int vm_object_zinit(void *mem, int size, int flags);
159 
160 #ifdef INVARIANTS
161 static void vm_object_zdtor(void *mem, int size, void *arg);
162 
163 static void
164 vm_object_zdtor(void *mem, int size, void *arg)
165 {
166 	vm_object_t object;
167 
168 	object = (vm_object_t)mem;
169 	KASSERT(TAILQ_EMPTY(&object->memq),
170 	    ("object %p has resident pages in its memq", object));
171 	KASSERT(vm_radix_is_empty(&object->rtree),
172 	    ("object %p has resident pages in its trie", object));
173 #if VM_NRESERVLEVEL > 0
174 	KASSERT(LIST_EMPTY(&object->rvq),
175 	    ("object %p has reservations",
176 	    object));
177 #endif
178 	KASSERT(vm_object_cache_is_empty(object),
179 	    ("object %p has cached pages",
180 	    object));
181 	KASSERT(object->paging_in_progress == 0,
182 	    ("object %p paging_in_progress = %d",
183 	    object, object->paging_in_progress));
184 	KASSERT(object->resident_page_count == 0,
185 	    ("object %p resident_page_count = %d",
186 	    object, object->resident_page_count));
187 	KASSERT(object->shadow_count == 0,
188 	    ("object %p shadow_count = %d",
189 	    object, object->shadow_count));
190 }
191 #endif
192 
193 static int
194 vm_object_zinit(void *mem, int size, int flags)
195 {
196 	vm_object_t object;
197 
198 	object = (vm_object_t)mem;
199 	bzero(&object->lock, sizeof(object->lock));
200 	rw_init_flags(&object->lock, "vm object", RW_DUPOK);
201 
202 	/* These are true for any object that has been freed */
203 	object->rtree.rt_root = 0;
204 	object->rtree.rt_flags = 0;
205 	object->paging_in_progress = 0;
206 	object->resident_page_count = 0;
207 	object->shadow_count = 0;
208 	object->cache.rt_root = 0;
209 	object->cache.rt_flags = 0;
210 	return (0);
211 }
212 
213 static void
214 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
215 {
216 
217 	TAILQ_INIT(&object->memq);
218 	LIST_INIT(&object->shadow_head);
219 
220 	object->type = type;
221 	switch (type) {
222 	case OBJT_DEAD:
223 		panic("_vm_object_allocate: can't create OBJT_DEAD");
224 	case OBJT_DEFAULT:
225 	case OBJT_SWAP:
226 		object->flags = OBJ_ONEMAPPING;
227 		break;
228 	case OBJT_DEVICE:
229 	case OBJT_SG:
230 		object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
231 		break;
232 	case OBJT_MGTDEVICE:
233 		object->flags = OBJ_FICTITIOUS;
234 		break;
235 	case OBJT_PHYS:
236 		object->flags = OBJ_UNMANAGED;
237 		break;
238 	case OBJT_VNODE:
239 		object->flags = 0;
240 		break;
241 	default:
242 		panic("_vm_object_allocate: type %d is undefined", type);
243 	}
244 	object->size = size;
245 	object->generation = 1;
246 	object->ref_count = 1;
247 	object->memattr = VM_MEMATTR_DEFAULT;
248 	object->cred = NULL;
249 	object->charge = 0;
250 	object->handle = NULL;
251 	object->backing_object = NULL;
252 	object->backing_object_offset = (vm_ooffset_t) 0;
253 #if VM_NRESERVLEVEL > 0
254 	LIST_INIT(&object->rvq);
255 #endif
256 
257 	mtx_lock(&vm_object_list_mtx);
258 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
259 	mtx_unlock(&vm_object_list_mtx);
260 }
261 
262 /*
263  *	vm_object_init:
264  *
265  *	Initialize the VM objects module.
266  */
267 void
268 vm_object_init(void)
269 {
270 	TAILQ_INIT(&vm_object_list);
271 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
272 
273 	rw_init(&kernel_object->lock, "kernel vm object");
274 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
275 	    kernel_object);
276 #if VM_NRESERVLEVEL > 0
277 	kernel_object->flags |= OBJ_COLORED;
278 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
279 #endif
280 
281 	rw_init(&kmem_object->lock, "kmem vm object");
282 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
283 	    kmem_object);
284 #if VM_NRESERVLEVEL > 0
285 	kmem_object->flags |= OBJ_COLORED;
286 	kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
287 #endif
288 
289 	/*
290 	 * The lock portion of struct vm_object must be type stable due
291 	 * to vm_pageout_fallback_object_lock locking a vm object
292 	 * without holding any references to it.
293 	 */
294 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
295 #ifdef INVARIANTS
296 	    vm_object_zdtor,
297 #else
298 	    NULL,
299 #endif
300 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
301 
302 	vm_radix_init();
303 }
304 
305 void
306 vm_object_clear_flag(vm_object_t object, u_short bits)
307 {
308 
309 	VM_OBJECT_ASSERT_WLOCKED(object);
310 	object->flags &= ~bits;
311 }
312 
313 /*
314  *	Sets the default memory attribute for the specified object.  Pages
315  *	that are allocated to this object are by default assigned this memory
316  *	attribute.
317  *
318  *	Presently, this function must be called before any pages are allocated
319  *	to the object.  In the future, this requirement may be relaxed for
320  *	"default" and "swap" objects.
321  */
322 int
323 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
324 {
325 
326 	VM_OBJECT_ASSERT_WLOCKED(object);
327 	switch (object->type) {
328 	case OBJT_DEFAULT:
329 	case OBJT_DEVICE:
330 	case OBJT_MGTDEVICE:
331 	case OBJT_PHYS:
332 	case OBJT_SG:
333 	case OBJT_SWAP:
334 	case OBJT_VNODE:
335 		if (!TAILQ_EMPTY(&object->memq))
336 			return (KERN_FAILURE);
337 		break;
338 	case OBJT_DEAD:
339 		return (KERN_INVALID_ARGUMENT);
340 	default:
341 		panic("vm_object_set_memattr: object %p is of undefined type",
342 		    object);
343 	}
344 	object->memattr = memattr;
345 	return (KERN_SUCCESS);
346 }
347 
348 void
349 vm_object_pip_add(vm_object_t object, short i)
350 {
351 
352 	VM_OBJECT_ASSERT_WLOCKED(object);
353 	object->paging_in_progress += i;
354 }
355 
356 void
357 vm_object_pip_subtract(vm_object_t object, short i)
358 {
359 
360 	VM_OBJECT_ASSERT_WLOCKED(object);
361 	object->paging_in_progress -= i;
362 }
363 
364 void
365 vm_object_pip_wakeup(vm_object_t object)
366 {
367 
368 	VM_OBJECT_ASSERT_WLOCKED(object);
369 	object->paging_in_progress--;
370 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
371 		vm_object_clear_flag(object, OBJ_PIPWNT);
372 		wakeup(object);
373 	}
374 }
375 
376 void
377 vm_object_pip_wakeupn(vm_object_t object, short i)
378 {
379 
380 	VM_OBJECT_ASSERT_WLOCKED(object);
381 	if (i)
382 		object->paging_in_progress -= i;
383 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
384 		vm_object_clear_flag(object, OBJ_PIPWNT);
385 		wakeup(object);
386 	}
387 }
388 
389 void
390 vm_object_pip_wait(vm_object_t object, char *waitid)
391 {
392 
393 	VM_OBJECT_ASSERT_WLOCKED(object);
394 	while (object->paging_in_progress) {
395 		object->flags |= OBJ_PIPWNT;
396 		VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
397 	}
398 }
399 
400 /*
401  *	vm_object_allocate:
402  *
403  *	Returns a new object with the given size.
404  */
405 vm_object_t
406 vm_object_allocate(objtype_t type, vm_pindex_t size)
407 {
408 	vm_object_t object;
409 
410 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
411 	_vm_object_allocate(type, size, object);
412 	return (object);
413 }
414 
415 
416 /*
417  *	vm_object_reference:
418  *
419  *	Gets another reference to the given object.  Note: OBJ_DEAD
420  *	objects can be referenced during final cleaning.
421  */
422 void
423 vm_object_reference(vm_object_t object)
424 {
425 	if (object == NULL)
426 		return;
427 	VM_OBJECT_WLOCK(object);
428 	vm_object_reference_locked(object);
429 	VM_OBJECT_WUNLOCK(object);
430 }
431 
432 /*
433  *	vm_object_reference_locked:
434  *
435  *	Gets another reference to the given object.
436  *
437  *	The object must be locked.
438  */
439 void
440 vm_object_reference_locked(vm_object_t object)
441 {
442 	struct vnode *vp;
443 
444 	VM_OBJECT_ASSERT_WLOCKED(object);
445 	object->ref_count++;
446 	if (object->type == OBJT_VNODE) {
447 		vp = object->handle;
448 		vref(vp);
449 	}
450 }
451 
452 /*
453  * Handle deallocating an object of type OBJT_VNODE.
454  */
455 static void
456 vm_object_vndeallocate(vm_object_t object)
457 {
458 	struct vnode *vp = (struct vnode *) object->handle;
459 
460 	VM_OBJECT_ASSERT_WLOCKED(object);
461 	KASSERT(object->type == OBJT_VNODE,
462 	    ("vm_object_vndeallocate: not a vnode object"));
463 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
464 #ifdef INVARIANTS
465 	if (object->ref_count == 0) {
466 		vprint("vm_object_vndeallocate", vp);
467 		panic("vm_object_vndeallocate: bad object reference count");
468 	}
469 #endif
470 
471 	if (object->ref_count > 1) {
472 		object->ref_count--;
473 		VM_OBJECT_WUNLOCK(object);
474 		/* vrele may need the vnode lock. */
475 		vrele(vp);
476 	} else {
477 		vhold(vp);
478 		VM_OBJECT_WUNLOCK(object);
479 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
480 		vdrop(vp);
481 		VM_OBJECT_WLOCK(object);
482 		object->ref_count--;
483 		if (object->type == OBJT_DEAD) {
484 			VM_OBJECT_WUNLOCK(object);
485 			VOP_UNLOCK(vp, 0);
486 		} else {
487 			if (object->ref_count == 0)
488 				VOP_UNSET_TEXT(vp);
489 			VM_OBJECT_WUNLOCK(object);
490 			vput(vp);
491 		}
492 	}
493 }
494 
495 /*
496  *	vm_object_deallocate:
497  *
498  *	Release a reference to the specified object,
499  *	gained either through a vm_object_allocate
500  *	or a vm_object_reference call.  When all references
501  *	are gone, storage associated with this object
502  *	may be relinquished.
503  *
504  *	No object may be locked.
505  */
506 void
507 vm_object_deallocate(vm_object_t object)
508 {
509 	vm_object_t temp;
510 	struct vnode *vp;
511 
512 	while (object != NULL) {
513 		VM_OBJECT_WLOCK(object);
514 		if (object->type == OBJT_VNODE) {
515 			vm_object_vndeallocate(object);
516 			return;
517 		}
518 
519 		KASSERT(object->ref_count != 0,
520 			("vm_object_deallocate: object deallocated too many times: %d", object->type));
521 
522 		/*
523 		 * If the reference count goes to 0 we start calling
524 		 * vm_object_terminate() on the object chain.
525 		 * A ref count of 1 may be a special case depending on the
526 		 * shadow count being 0 or 1.
527 		 */
528 		object->ref_count--;
529 		if (object->ref_count > 1) {
530 			VM_OBJECT_WUNLOCK(object);
531 			return;
532 		} else if (object->ref_count == 1) {
533 			if (object->type == OBJT_SWAP &&
534 			    (object->flags & OBJ_TMPFS) != 0) {
535 				vp = object->un_pager.swp.swp_tmpfs;
536 				vhold(vp);
537 				VM_OBJECT_WUNLOCK(object);
538 				vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
539 				vdrop(vp);
540 				VM_OBJECT_WLOCK(object);
541 				if (object->type == OBJT_DEAD ||
542 				    object->ref_count != 1) {
543 					VM_OBJECT_WUNLOCK(object);
544 					VOP_UNLOCK(vp, 0);
545 					return;
546 				}
547 				if ((object->flags & OBJ_TMPFS) != 0)
548 					VOP_UNSET_TEXT(vp);
549 				VOP_UNLOCK(vp, 0);
550 			}
551 			if (object->shadow_count == 0 &&
552 			    object->handle == NULL &&
553 			    (object->type == OBJT_DEFAULT ||
554 			    (object->type == OBJT_SWAP &&
555 			    (object->flags & OBJ_TMPFS) == 0))) {
556 				vm_object_set_flag(object, OBJ_ONEMAPPING);
557 			} else if ((object->shadow_count == 1) &&
558 			    (object->handle == NULL) &&
559 			    (object->type == OBJT_DEFAULT ||
560 			     object->type == OBJT_SWAP)) {
561 				KASSERT((object->flags & OBJ_TMPFS) == 0,
562 				    ("shadowed tmpfs v_object %p", object));
563 				vm_object_t robject;
564 
565 				robject = LIST_FIRST(&object->shadow_head);
566 				KASSERT(robject != NULL,
567 				    ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
568 					 object->ref_count,
569 					 object->shadow_count));
570 				if (!VM_OBJECT_TRYWLOCK(robject)) {
571 					/*
572 					 * Avoid a potential deadlock.
573 					 */
574 					object->ref_count++;
575 					VM_OBJECT_WUNLOCK(object);
576 					/*
577 					 * More likely than not the thread
578 					 * holding robject's lock has lower
579 					 * priority than the current thread.
580 					 * Let the lower priority thread run.
581 					 */
582 					pause("vmo_de", 1);
583 					continue;
584 				}
585 				/*
586 				 * Collapse object into its shadow unless its
587 				 * shadow is dead.  In that case, object will
588 				 * be deallocated by the thread that is
589 				 * deallocating its shadow.
590 				 */
591 				if ((robject->flags & OBJ_DEAD) == 0 &&
592 				    (robject->handle == NULL) &&
593 				    (robject->type == OBJT_DEFAULT ||
594 				     robject->type == OBJT_SWAP)) {
595 
596 					robject->ref_count++;
597 retry:
598 					if (robject->paging_in_progress) {
599 						VM_OBJECT_WUNLOCK(object);
600 						vm_object_pip_wait(robject,
601 						    "objde1");
602 						temp = robject->backing_object;
603 						if (object == temp) {
604 							VM_OBJECT_WLOCK(object);
605 							goto retry;
606 						}
607 					} else if (object->paging_in_progress) {
608 						VM_OBJECT_WUNLOCK(robject);
609 						object->flags |= OBJ_PIPWNT;
610 						VM_OBJECT_SLEEP(object, object,
611 						    PDROP | PVM, "objde2", 0);
612 						VM_OBJECT_WLOCK(robject);
613 						temp = robject->backing_object;
614 						if (object == temp) {
615 							VM_OBJECT_WLOCK(object);
616 							goto retry;
617 						}
618 					} else
619 						VM_OBJECT_WUNLOCK(object);
620 
621 					if (robject->ref_count == 1) {
622 						robject->ref_count--;
623 						object = robject;
624 						goto doterm;
625 					}
626 					object = robject;
627 					vm_object_collapse(object);
628 					VM_OBJECT_WUNLOCK(object);
629 					continue;
630 				}
631 				VM_OBJECT_WUNLOCK(robject);
632 			}
633 			VM_OBJECT_WUNLOCK(object);
634 			return;
635 		}
636 doterm:
637 		temp = object->backing_object;
638 		if (temp != NULL) {
639 			VM_OBJECT_WLOCK(temp);
640 			LIST_REMOVE(object, shadow_list);
641 			temp->shadow_count--;
642 			VM_OBJECT_WUNLOCK(temp);
643 			object->backing_object = NULL;
644 		}
645 		/*
646 		 * Don't double-terminate, we could be in a termination
647 		 * recursion due to the terminate having to sync data
648 		 * to disk.
649 		 */
650 		if ((object->flags & OBJ_DEAD) == 0)
651 			vm_object_terminate(object);
652 		else
653 			VM_OBJECT_WUNLOCK(object);
654 		object = temp;
655 	}
656 }
657 
658 /*
659  *	vm_object_destroy removes the object from the global object list
660  *      and frees the space for the object.
661  */
662 void
663 vm_object_destroy(vm_object_t object)
664 {
665 
666 	/*
667 	 * Remove the object from the global object list.
668 	 */
669 	mtx_lock(&vm_object_list_mtx);
670 	TAILQ_REMOVE(&vm_object_list, object, object_list);
671 	mtx_unlock(&vm_object_list_mtx);
672 
673 	/*
674 	 * Release the allocation charge.
675 	 */
676 	if (object->cred != NULL) {
677 		KASSERT(object->type == OBJT_DEFAULT ||
678 		    object->type == OBJT_SWAP,
679 		    ("%s: non-swap obj %p has cred", __func__, object));
680 		swap_release_by_cred(object->charge, object->cred);
681 		object->charge = 0;
682 		crfree(object->cred);
683 		object->cred = NULL;
684 	}
685 
686 	/*
687 	 * Free the space for the object.
688 	 */
689 	uma_zfree(obj_zone, object);
690 }
691 
692 /*
693  *	vm_object_terminate actually destroys the specified object, freeing
694  *	up all previously used resources.
695  *
696  *	The object must be locked.
697  *	This routine may block.
698  */
699 void
700 vm_object_terminate(vm_object_t object)
701 {
702 	vm_page_t p, p_next;
703 
704 	VM_OBJECT_ASSERT_WLOCKED(object);
705 
706 	/*
707 	 * Make sure no one uses us.
708 	 */
709 	vm_object_set_flag(object, OBJ_DEAD);
710 
711 	/*
712 	 * wait for the pageout daemon to be done with the object
713 	 */
714 	vm_object_pip_wait(object, "objtrm");
715 
716 	KASSERT(!object->paging_in_progress,
717 		("vm_object_terminate: pageout in progress"));
718 
719 	/*
720 	 * Clean and free the pages, as appropriate. All references to the
721 	 * object are gone, so we don't need to lock it.
722 	 */
723 	if (object->type == OBJT_VNODE) {
724 		struct vnode *vp = (struct vnode *)object->handle;
725 
726 		/*
727 		 * Clean pages and flush buffers.
728 		 */
729 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
730 		VM_OBJECT_WUNLOCK(object);
731 
732 		vinvalbuf(vp, V_SAVE, 0, 0);
733 
734 		VM_OBJECT_WLOCK(object);
735 	}
736 
737 	KASSERT(object->ref_count == 0,
738 		("vm_object_terminate: object with references, ref_count=%d",
739 		object->ref_count));
740 
741 	/*
742 	 * Free any remaining pageable pages.  This also removes them from the
743 	 * paging queues.  However, don't free wired pages, just remove them
744 	 * from the object.  Rather than incrementally removing each page from
745 	 * the object, the page and object are reset to any empty state.
746 	 */
747 	TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
748 		vm_page_assert_unbusied(p);
749 		vm_page_lock(p);
750 		/*
751 		 * Optimize the page's removal from the object by resetting
752 		 * its "object" field.  Specifically, if the page is not
753 		 * wired, then the effect of this assignment is that
754 		 * vm_page_free()'s call to vm_page_remove() will return
755 		 * immediately without modifying the page or the object.
756 		 */
757 		p->object = NULL;
758 		if (p->wire_count == 0) {
759 			vm_page_free(p);
760 			PCPU_INC(cnt.v_pfree);
761 		}
762 		vm_page_unlock(p);
763 	}
764 	/*
765 	 * If the object contained any pages, then reset it to an empty state.
766 	 * None of the object's fields, including "resident_page_count", were
767 	 * modified by the preceding loop.
768 	 */
769 	if (object->resident_page_count != 0) {
770 		vm_radix_reclaim_allnodes(&object->rtree);
771 		TAILQ_INIT(&object->memq);
772 		object->resident_page_count = 0;
773 		if (object->type == OBJT_VNODE)
774 			vdrop(object->handle);
775 	}
776 
777 #if VM_NRESERVLEVEL > 0
778 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
779 		vm_reserv_break_all(object);
780 #endif
781 	if (__predict_false(!vm_object_cache_is_empty(object)))
782 		vm_page_cache_free(object, 0, 0);
783 
784 	/*
785 	 * Let the pager know object is dead.
786 	 */
787 	vm_pager_deallocate(object);
788 	VM_OBJECT_WUNLOCK(object);
789 
790 	vm_object_destroy(object);
791 }
792 
793 /*
794  * Make the page read-only so that we can clear the object flags.  However, if
795  * this is a nosync mmap then the object is likely to stay dirty so do not
796  * mess with the page and do not clear the object flags.  Returns TRUE if the
797  * page should be flushed, and FALSE otherwise.
798  */
799 static boolean_t
800 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
801 {
802 
803 	/*
804 	 * If we have been asked to skip nosync pages and this is a
805 	 * nosync page, skip it.  Note that the object flags were not
806 	 * cleared in this case so we do not have to set them.
807 	 */
808 	if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
809 		*clearobjflags = FALSE;
810 		return (FALSE);
811 	} else {
812 		pmap_remove_write(p);
813 		return (p->dirty != 0);
814 	}
815 }
816 
817 /*
818  *	vm_object_page_clean
819  *
820  *	Clean all dirty pages in the specified range of object.  Leaves page
821  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
822  *	write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
823  *	leaving the object dirty.
824  *
825  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
826  *	synchronous clustering mode implementation.
827  *
828  *	Odd semantics: if start == end, we clean everything.
829  *
830  *	The object must be locked.
831  *
832  *	Returns FALSE if some page from the range was not written, as
833  *	reported by the pager, and TRUE otherwise.
834  */
835 boolean_t
836 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
837     int flags)
838 {
839 	vm_page_t np, p;
840 	vm_pindex_t pi, tend, tstart;
841 	int curgeneration, n, pagerflags;
842 	boolean_t clearobjflags, eio, res;
843 
844 	VM_OBJECT_ASSERT_WLOCKED(object);
845 
846 	/*
847 	 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
848 	 * objects.  The check below prevents the function from
849 	 * operating on non-vnode objects.
850 	 */
851 	if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
852 	    object->resident_page_count == 0)
853 		return (TRUE);
854 
855 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
856 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
857 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
858 
859 	tstart = OFF_TO_IDX(start);
860 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
861 	clearobjflags = tstart == 0 && tend >= object->size;
862 	res = TRUE;
863 
864 rescan:
865 	curgeneration = object->generation;
866 
867 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
868 		pi = p->pindex;
869 		if (pi >= tend)
870 			break;
871 		np = TAILQ_NEXT(p, listq);
872 		if (p->valid == 0)
873 			continue;
874 		if (vm_page_sleep_if_busy(p, "vpcwai")) {
875 			if (object->generation != curgeneration) {
876 				if ((flags & OBJPC_SYNC) != 0)
877 					goto rescan;
878 				else
879 					clearobjflags = FALSE;
880 			}
881 			np = vm_page_find_least(object, pi);
882 			continue;
883 		}
884 		if (!vm_object_page_remove_write(p, flags, &clearobjflags))
885 			continue;
886 
887 		n = vm_object_page_collect_flush(object, p, pagerflags,
888 		    flags, &clearobjflags, &eio);
889 		if (eio) {
890 			res = FALSE;
891 			clearobjflags = FALSE;
892 		}
893 		if (object->generation != curgeneration) {
894 			if ((flags & OBJPC_SYNC) != 0)
895 				goto rescan;
896 			else
897 				clearobjflags = FALSE;
898 		}
899 
900 		/*
901 		 * If the VOP_PUTPAGES() did a truncated write, so
902 		 * that even the first page of the run is not fully
903 		 * written, vm_pageout_flush() returns 0 as the run
904 		 * length.  Since the condition that caused truncated
905 		 * write may be permanent, e.g. exhausted free space,
906 		 * accepting n == 0 would cause an infinite loop.
907 		 *
908 		 * Forwarding the iterator leaves the unwritten page
909 		 * behind, but there is not much we can do there if
910 		 * filesystem refuses to write it.
911 		 */
912 		if (n == 0) {
913 			n = 1;
914 			clearobjflags = FALSE;
915 		}
916 		np = vm_page_find_least(object, pi + n);
917 	}
918 #if 0
919 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
920 #endif
921 
922 	if (clearobjflags)
923 		vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
924 	return (res);
925 }
926 
927 static int
928 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
929     int flags, boolean_t *clearobjflags, boolean_t *eio)
930 {
931 	vm_page_t ma[vm_pageout_page_count], p_first, tp;
932 	int count, i, mreq, runlen;
933 
934 	vm_page_lock_assert(p, MA_NOTOWNED);
935 	VM_OBJECT_ASSERT_WLOCKED(object);
936 
937 	count = 1;
938 	mreq = 0;
939 
940 	for (tp = p; count < vm_pageout_page_count; count++) {
941 		tp = vm_page_next(tp);
942 		if (tp == NULL || vm_page_busied(tp))
943 			break;
944 		if (!vm_object_page_remove_write(tp, flags, clearobjflags))
945 			break;
946 	}
947 
948 	for (p_first = p; count < vm_pageout_page_count; count++) {
949 		tp = vm_page_prev(p_first);
950 		if (tp == NULL || vm_page_busied(tp))
951 			break;
952 		if (!vm_object_page_remove_write(tp, flags, clearobjflags))
953 			break;
954 		p_first = tp;
955 		mreq++;
956 	}
957 
958 	for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
959 		ma[i] = tp;
960 
961 	vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
962 	return (runlen);
963 }
964 
965 /*
966  * Note that there is absolutely no sense in writing out
967  * anonymous objects, so we track down the vnode object
968  * to write out.
969  * We invalidate (remove) all pages from the address space
970  * for semantic correctness.
971  *
972  * If the backing object is a device object with unmanaged pages, then any
973  * mappings to the specified range of pages must be removed before this
974  * function is called.
975  *
976  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
977  * may start out with a NULL object.
978  */
979 boolean_t
980 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
981     boolean_t syncio, boolean_t invalidate)
982 {
983 	vm_object_t backing_object;
984 	struct vnode *vp;
985 	struct mount *mp;
986 	int error, flags, fsync_after;
987 	boolean_t res;
988 
989 	if (object == NULL)
990 		return (TRUE);
991 	res = TRUE;
992 	error = 0;
993 	VM_OBJECT_WLOCK(object);
994 	while ((backing_object = object->backing_object) != NULL) {
995 		VM_OBJECT_WLOCK(backing_object);
996 		offset += object->backing_object_offset;
997 		VM_OBJECT_WUNLOCK(object);
998 		object = backing_object;
999 		if (object->size < OFF_TO_IDX(offset + size))
1000 			size = IDX_TO_OFF(object->size) - offset;
1001 	}
1002 	/*
1003 	 * Flush pages if writing is allowed, invalidate them
1004 	 * if invalidation requested.  Pages undergoing I/O
1005 	 * will be ignored by vm_object_page_remove().
1006 	 *
1007 	 * We cannot lock the vnode and then wait for paging
1008 	 * to complete without deadlocking against vm_fault.
1009 	 * Instead we simply call vm_object_page_remove() and
1010 	 * allow it to block internally on a page-by-page
1011 	 * basis when it encounters pages undergoing async
1012 	 * I/O.
1013 	 */
1014 	if (object->type == OBJT_VNODE &&
1015 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1016 		vp = object->handle;
1017 		VM_OBJECT_WUNLOCK(object);
1018 		(void) vn_start_write(vp, &mp, V_WAIT);
1019 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1020 		if (syncio && !invalidate && offset == 0 &&
1021 		    OFF_TO_IDX(size) == object->size) {
1022 			/*
1023 			 * If syncing the whole mapping of the file,
1024 			 * it is faster to schedule all the writes in
1025 			 * async mode, also allowing the clustering,
1026 			 * and then wait for i/o to complete.
1027 			 */
1028 			flags = 0;
1029 			fsync_after = TRUE;
1030 		} else {
1031 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1032 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1033 			fsync_after = FALSE;
1034 		}
1035 		VM_OBJECT_WLOCK(object);
1036 		res = vm_object_page_clean(object, offset, offset + size,
1037 		    flags);
1038 		VM_OBJECT_WUNLOCK(object);
1039 		if (fsync_after)
1040 			error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1041 		VOP_UNLOCK(vp, 0);
1042 		vn_finished_write(mp);
1043 		if (error != 0)
1044 			res = FALSE;
1045 		VM_OBJECT_WLOCK(object);
1046 	}
1047 	if ((object->type == OBJT_VNODE ||
1048 	     object->type == OBJT_DEVICE) && invalidate) {
1049 		if (object->type == OBJT_DEVICE)
1050 			/*
1051 			 * The option OBJPR_NOTMAPPED must be passed here
1052 			 * because vm_object_page_remove() cannot remove
1053 			 * unmanaged mappings.
1054 			 */
1055 			flags = OBJPR_NOTMAPPED;
1056 		else if (old_msync)
1057 			flags = OBJPR_NOTWIRED;
1058 		else
1059 			flags = OBJPR_CLEANONLY | OBJPR_NOTWIRED;
1060 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1061 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1062 	}
1063 	VM_OBJECT_WUNLOCK(object);
1064 	return (res);
1065 }
1066 
1067 /*
1068  *	vm_object_madvise:
1069  *
1070  *	Implements the madvise function at the object/page level.
1071  *
1072  *	MADV_WILLNEED	(any object)
1073  *
1074  *	    Activate the specified pages if they are resident.
1075  *
1076  *	MADV_DONTNEED	(any object)
1077  *
1078  *	    Deactivate the specified pages if they are resident.
1079  *
1080  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1081  *			 OBJ_ONEMAPPING only)
1082  *
1083  *	    Deactivate and clean the specified pages if they are
1084  *	    resident.  This permits the process to reuse the pages
1085  *	    without faulting or the kernel to reclaim the pages
1086  *	    without I/O.
1087  */
1088 void
1089 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1090     int advise)
1091 {
1092 	vm_pindex_t tpindex;
1093 	vm_object_t backing_object, tobject;
1094 	vm_page_t m;
1095 
1096 	if (object == NULL)
1097 		return;
1098 	VM_OBJECT_WLOCK(object);
1099 	/*
1100 	 * Locate and adjust resident pages
1101 	 */
1102 	for (; pindex < end; pindex += 1) {
1103 relookup:
1104 		tobject = object;
1105 		tpindex = pindex;
1106 shadowlookup:
1107 		/*
1108 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1109 		 * and those pages must be OBJ_ONEMAPPING.
1110 		 */
1111 		if (advise == MADV_FREE) {
1112 			if ((tobject->type != OBJT_DEFAULT &&
1113 			     tobject->type != OBJT_SWAP) ||
1114 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
1115 				goto unlock_tobject;
1116 			}
1117 		} else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1118 			goto unlock_tobject;
1119 		m = vm_page_lookup(tobject, tpindex);
1120 		if (m == NULL && advise == MADV_WILLNEED) {
1121 			/*
1122 			 * If the page is cached, reactivate it.
1123 			 */
1124 			m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1125 			    VM_ALLOC_NOBUSY);
1126 		}
1127 		if (m == NULL) {
1128 			/*
1129 			 * There may be swap even if there is no backing page
1130 			 */
1131 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1132 				swap_pager_freespace(tobject, tpindex, 1);
1133 			/*
1134 			 * next object
1135 			 */
1136 			backing_object = tobject->backing_object;
1137 			if (backing_object == NULL)
1138 				goto unlock_tobject;
1139 			VM_OBJECT_WLOCK(backing_object);
1140 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1141 			if (tobject != object)
1142 				VM_OBJECT_WUNLOCK(tobject);
1143 			tobject = backing_object;
1144 			goto shadowlookup;
1145 		} else if (m->valid != VM_PAGE_BITS_ALL)
1146 			goto unlock_tobject;
1147 		/*
1148 		 * If the page is not in a normal state, skip it.
1149 		 */
1150 		vm_page_lock(m);
1151 		if (m->hold_count != 0 || m->wire_count != 0) {
1152 			vm_page_unlock(m);
1153 			goto unlock_tobject;
1154 		}
1155 		KASSERT((m->flags & PG_FICTITIOUS) == 0,
1156 		    ("vm_object_madvise: page %p is fictitious", m));
1157 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1158 		    ("vm_object_madvise: page %p is not managed", m));
1159 		if (vm_page_busied(m)) {
1160 			if (advise == MADV_WILLNEED) {
1161 				/*
1162 				 * Reference the page before unlocking and
1163 				 * sleeping so that the page daemon is less
1164 				 * likely to reclaim it.
1165 				 */
1166 				vm_page_aflag_set(m, PGA_REFERENCED);
1167 			}
1168 			if (object != tobject)
1169 				VM_OBJECT_WUNLOCK(object);
1170 			VM_OBJECT_WUNLOCK(tobject);
1171 			vm_page_busy_sleep(m, "madvpo");
1172 			VM_OBJECT_WLOCK(object);
1173   			goto relookup;
1174 		}
1175 		if (advise == MADV_WILLNEED) {
1176 			vm_page_activate(m);
1177 		} else {
1178 			vm_page_advise(m, advise);
1179 		}
1180 		vm_page_unlock(m);
1181 		if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1182 			swap_pager_freespace(tobject, tpindex, 1);
1183 unlock_tobject:
1184 		if (tobject != object)
1185 			VM_OBJECT_WUNLOCK(tobject);
1186 	}
1187 	VM_OBJECT_WUNLOCK(object);
1188 }
1189 
1190 /*
1191  *	vm_object_shadow:
1192  *
1193  *	Create a new object which is backed by the
1194  *	specified existing object range.  The source
1195  *	object reference is deallocated.
1196  *
1197  *	The new object and offset into that object
1198  *	are returned in the source parameters.
1199  */
1200 void
1201 vm_object_shadow(
1202 	vm_object_t *object,	/* IN/OUT */
1203 	vm_ooffset_t *offset,	/* IN/OUT */
1204 	vm_size_t length)
1205 {
1206 	vm_object_t source;
1207 	vm_object_t result;
1208 
1209 	source = *object;
1210 
1211 	/*
1212 	 * Don't create the new object if the old object isn't shared.
1213 	 */
1214 	if (source != NULL) {
1215 		VM_OBJECT_WLOCK(source);
1216 		if (source->ref_count == 1 &&
1217 		    source->handle == NULL &&
1218 		    (source->type == OBJT_DEFAULT ||
1219 		     source->type == OBJT_SWAP)) {
1220 			VM_OBJECT_WUNLOCK(source);
1221 			return;
1222 		}
1223 		VM_OBJECT_WUNLOCK(source);
1224 	}
1225 
1226 	/*
1227 	 * Allocate a new object with the given length.
1228 	 */
1229 	result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1230 
1231 	/*
1232 	 * The new object shadows the source object, adding a reference to it.
1233 	 * Our caller changes his reference to point to the new object,
1234 	 * removing a reference to the source object.  Net result: no change
1235 	 * of reference count.
1236 	 *
1237 	 * Try to optimize the result object's page color when shadowing
1238 	 * in order to maintain page coloring consistency in the combined
1239 	 * shadowed object.
1240 	 */
1241 	result->backing_object = source;
1242 	/*
1243 	 * Store the offset into the source object, and fix up the offset into
1244 	 * the new object.
1245 	 */
1246 	result->backing_object_offset = *offset;
1247 	if (source != NULL) {
1248 		VM_OBJECT_WLOCK(source);
1249 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1250 		source->shadow_count++;
1251 #if VM_NRESERVLEVEL > 0
1252 		result->flags |= source->flags & OBJ_COLORED;
1253 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1254 		    ((1 << (VM_NFREEORDER - 1)) - 1);
1255 #endif
1256 		VM_OBJECT_WUNLOCK(source);
1257 	}
1258 
1259 
1260 	/*
1261 	 * Return the new things
1262 	 */
1263 	*offset = 0;
1264 	*object = result;
1265 }
1266 
1267 /*
1268  *	vm_object_split:
1269  *
1270  * Split the pages in a map entry into a new object.  This affords
1271  * easier removal of unused pages, and keeps object inheritance from
1272  * being a negative impact on memory usage.
1273  */
1274 void
1275 vm_object_split(vm_map_entry_t entry)
1276 {
1277 	vm_page_t m, m_next;
1278 	vm_object_t orig_object, new_object, source;
1279 	vm_pindex_t idx, offidxstart;
1280 	vm_size_t size;
1281 
1282 	orig_object = entry->object.vm_object;
1283 	if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1284 		return;
1285 	if (orig_object->ref_count <= 1)
1286 		return;
1287 	VM_OBJECT_WUNLOCK(orig_object);
1288 
1289 	offidxstart = OFF_TO_IDX(entry->offset);
1290 	size = atop(entry->end - entry->start);
1291 
1292 	/*
1293 	 * If swap_pager_copy() is later called, it will convert new_object
1294 	 * into a swap object.
1295 	 */
1296 	new_object = vm_object_allocate(OBJT_DEFAULT, size);
1297 
1298 	/*
1299 	 * At this point, the new object is still private, so the order in
1300 	 * which the original and new objects are locked does not matter.
1301 	 */
1302 	VM_OBJECT_WLOCK(new_object);
1303 	VM_OBJECT_WLOCK(orig_object);
1304 	source = orig_object->backing_object;
1305 	if (source != NULL) {
1306 		VM_OBJECT_WLOCK(source);
1307 		if ((source->flags & OBJ_DEAD) != 0) {
1308 			VM_OBJECT_WUNLOCK(source);
1309 			VM_OBJECT_WUNLOCK(orig_object);
1310 			VM_OBJECT_WUNLOCK(new_object);
1311 			vm_object_deallocate(new_object);
1312 			VM_OBJECT_WLOCK(orig_object);
1313 			return;
1314 		}
1315 		LIST_INSERT_HEAD(&source->shadow_head,
1316 				  new_object, shadow_list);
1317 		source->shadow_count++;
1318 		vm_object_reference_locked(source);	/* for new_object */
1319 		vm_object_clear_flag(source, OBJ_ONEMAPPING);
1320 		VM_OBJECT_WUNLOCK(source);
1321 		new_object->backing_object_offset =
1322 			orig_object->backing_object_offset + entry->offset;
1323 		new_object->backing_object = source;
1324 	}
1325 	if (orig_object->cred != NULL) {
1326 		new_object->cred = orig_object->cred;
1327 		crhold(orig_object->cred);
1328 		new_object->charge = ptoa(size);
1329 		KASSERT(orig_object->charge >= ptoa(size),
1330 		    ("orig_object->charge < 0"));
1331 		orig_object->charge -= ptoa(size);
1332 	}
1333 retry:
1334 	m = vm_page_find_least(orig_object, offidxstart);
1335 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1336 	    m = m_next) {
1337 		m_next = TAILQ_NEXT(m, listq);
1338 
1339 		/*
1340 		 * We must wait for pending I/O to complete before we can
1341 		 * rename the page.
1342 		 *
1343 		 * We do not have to VM_PROT_NONE the page as mappings should
1344 		 * not be changed by this operation.
1345 		 */
1346 		if (vm_page_busied(m)) {
1347 			VM_OBJECT_WUNLOCK(new_object);
1348 			vm_page_lock(m);
1349 			VM_OBJECT_WUNLOCK(orig_object);
1350 			vm_page_busy_sleep(m, "spltwt");
1351 			VM_OBJECT_WLOCK(orig_object);
1352 			VM_OBJECT_WLOCK(new_object);
1353 			goto retry;
1354 		}
1355 
1356 		/* vm_page_rename() will handle dirty and cache. */
1357 		if (vm_page_rename(m, new_object, idx)) {
1358 			VM_OBJECT_WUNLOCK(new_object);
1359 			VM_OBJECT_WUNLOCK(orig_object);
1360 			VM_WAIT;
1361 			VM_OBJECT_WLOCK(orig_object);
1362 			VM_OBJECT_WLOCK(new_object);
1363 			goto retry;
1364 		}
1365 #if VM_NRESERVLEVEL > 0
1366 		/*
1367 		 * If some of the reservation's allocated pages remain with
1368 		 * the original object, then transferring the reservation to
1369 		 * the new object is neither particularly beneficial nor
1370 		 * particularly harmful as compared to leaving the reservation
1371 		 * with the original object.  If, however, all of the
1372 		 * reservation's allocated pages are transferred to the new
1373 		 * object, then transferring the reservation is typically
1374 		 * beneficial.  Determining which of these two cases applies
1375 		 * would be more costly than unconditionally renaming the
1376 		 * reservation.
1377 		 */
1378 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1379 #endif
1380 		if (orig_object->type == OBJT_SWAP)
1381 			vm_page_xbusy(m);
1382 	}
1383 	if (orig_object->type == OBJT_SWAP) {
1384 		/*
1385 		 * swap_pager_copy() can sleep, in which case the orig_object's
1386 		 * and new_object's locks are released and reacquired.
1387 		 */
1388 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1389 		TAILQ_FOREACH(m, &new_object->memq, listq)
1390 			vm_page_xunbusy(m);
1391 
1392 		/*
1393 		 * Transfer any cached pages from orig_object to new_object.
1394 		 * If swap_pager_copy() found swapped out pages within the
1395 		 * specified range of orig_object, then it changed
1396 		 * new_object's type to OBJT_SWAP when it transferred those
1397 		 * pages to new_object.  Otherwise, new_object's type
1398 		 * should still be OBJT_DEFAULT and orig_object should not
1399 		 * contain any cached pages within the specified range.
1400 		 */
1401 		if (__predict_false(!vm_object_cache_is_empty(orig_object)))
1402 			vm_page_cache_transfer(orig_object, offidxstart,
1403 			    new_object);
1404 	}
1405 	VM_OBJECT_WUNLOCK(orig_object);
1406 	VM_OBJECT_WUNLOCK(new_object);
1407 	entry->object.vm_object = new_object;
1408 	entry->offset = 0LL;
1409 	vm_object_deallocate(orig_object);
1410 	VM_OBJECT_WLOCK(new_object);
1411 }
1412 
1413 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1414 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1415 #define	OBSC_COLLAPSE_WAIT	0x0004
1416 
1417 static int
1418 vm_object_backing_scan(vm_object_t object, int op)
1419 {
1420 	int r = 1;
1421 	vm_page_t p;
1422 	vm_object_t backing_object;
1423 	vm_pindex_t backing_offset_index;
1424 
1425 	VM_OBJECT_ASSERT_WLOCKED(object);
1426 	VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1427 
1428 	backing_object = object->backing_object;
1429 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1430 
1431 	/*
1432 	 * Initial conditions
1433 	 */
1434 	if (op & OBSC_TEST_ALL_SHADOWED) {
1435 		/*
1436 		 * We do not want to have to test for the existence of cache
1437 		 * or swap pages in the backing object.  XXX but with the
1438 		 * new swapper this would be pretty easy to do.
1439 		 *
1440 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1441 		 * been ZFOD faulted yet?  If we do not test for this, the
1442 		 * shadow test may succeed! XXX
1443 		 */
1444 		if (backing_object->type != OBJT_DEFAULT) {
1445 			return (0);
1446 		}
1447 	}
1448 	if (op & OBSC_COLLAPSE_WAIT) {
1449 		vm_object_set_flag(backing_object, OBJ_DEAD);
1450 	}
1451 
1452 	/*
1453 	 * Our scan
1454 	 */
1455 	p = TAILQ_FIRST(&backing_object->memq);
1456 	while (p) {
1457 		vm_page_t next = TAILQ_NEXT(p, listq);
1458 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1459 
1460 		if (op & OBSC_TEST_ALL_SHADOWED) {
1461 			vm_page_t pp;
1462 
1463 			/*
1464 			 * Ignore pages outside the parent object's range
1465 			 * and outside the parent object's mapping of the
1466 			 * backing object.
1467 			 *
1468 			 * note that we do not busy the backing object's
1469 			 * page.
1470 			 */
1471 			if (
1472 			    p->pindex < backing_offset_index ||
1473 			    new_pindex >= object->size
1474 			) {
1475 				p = next;
1476 				continue;
1477 			}
1478 
1479 			/*
1480 			 * See if the parent has the page or if the parent's
1481 			 * object pager has the page.  If the parent has the
1482 			 * page but the page is not valid, the parent's
1483 			 * object pager must have the page.
1484 			 *
1485 			 * If this fails, the parent does not completely shadow
1486 			 * the object and we might as well give up now.
1487 			 */
1488 
1489 			pp = vm_page_lookup(object, new_pindex);
1490 			if (
1491 			    (pp == NULL || pp->valid == 0) &&
1492 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1493 			) {
1494 				r = 0;
1495 				break;
1496 			}
1497 		}
1498 
1499 		/*
1500 		 * Check for busy page
1501 		 */
1502 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1503 			vm_page_t pp;
1504 
1505 			if (op & OBSC_COLLAPSE_NOWAIT) {
1506 				if (!p->valid || vm_page_busied(p)) {
1507 					p = next;
1508 					continue;
1509 				}
1510 			} else if (op & OBSC_COLLAPSE_WAIT) {
1511 				if (vm_page_busied(p)) {
1512 					VM_OBJECT_WUNLOCK(object);
1513 					vm_page_lock(p);
1514 					VM_OBJECT_WUNLOCK(backing_object);
1515 					vm_page_busy_sleep(p, "vmocol");
1516 					VM_OBJECT_WLOCK(object);
1517 					VM_OBJECT_WLOCK(backing_object);
1518 					/*
1519 					 * If we slept, anything could have
1520 					 * happened.  Since the object is
1521 					 * marked dead, the backing offset
1522 					 * should not have changed so we
1523 					 * just restart our scan.
1524 					 */
1525 					p = TAILQ_FIRST(&backing_object->memq);
1526 					continue;
1527 				}
1528 			}
1529 
1530 			KASSERT(
1531 			    p->object == backing_object,
1532 			    ("vm_object_backing_scan: object mismatch")
1533 			);
1534 
1535 			if (
1536 			    p->pindex < backing_offset_index ||
1537 			    new_pindex >= object->size
1538 			) {
1539 				if (backing_object->type == OBJT_SWAP)
1540 					swap_pager_freespace(backing_object,
1541 					    p->pindex, 1);
1542 
1543 				/*
1544 				 * Page is out of the parent object's range, we
1545 				 * can simply destroy it.
1546 				 */
1547 				vm_page_lock(p);
1548 				KASSERT(!pmap_page_is_mapped(p),
1549 				    ("freeing mapped page %p", p));
1550 				if (p->wire_count == 0)
1551 					vm_page_free(p);
1552 				else
1553 					vm_page_remove(p);
1554 				vm_page_unlock(p);
1555 				p = next;
1556 				continue;
1557 			}
1558 
1559 			pp = vm_page_lookup(object, new_pindex);
1560 			if (
1561 			    (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1562 			    (pp != NULL && pp->valid == 0)
1563 			) {
1564 				if (backing_object->type == OBJT_SWAP)
1565 					swap_pager_freespace(backing_object,
1566 					    p->pindex, 1);
1567 
1568 				/*
1569 				 * The page in the parent is not (yet) valid.
1570 				 * We don't know anything about the state of
1571 				 * the original page.  It might be mapped,
1572 				 * so we must avoid the next if here.
1573 				 *
1574 				 * This is due to a race in vm_fault() where
1575 				 * we must unbusy the original (backing_obj)
1576 				 * page before we can (re)lock the parent.
1577 				 * Hence we can get here.
1578 				 */
1579 				p = next;
1580 				continue;
1581 			}
1582 			if (
1583 			    pp != NULL ||
1584 			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1585 			) {
1586 				if (backing_object->type == OBJT_SWAP)
1587 					swap_pager_freespace(backing_object,
1588 					    p->pindex, 1);
1589 
1590 				/*
1591 				 * page already exists in parent OR swap exists
1592 				 * for this location in the parent.  Destroy
1593 				 * the original page from the backing object.
1594 				 *
1595 				 * Leave the parent's page alone
1596 				 */
1597 				vm_page_lock(p);
1598 				KASSERT(!pmap_page_is_mapped(p),
1599 				    ("freeing mapped page %p", p));
1600 				if (p->wire_count == 0)
1601 					vm_page_free(p);
1602 				else
1603 					vm_page_remove(p);
1604 				vm_page_unlock(p);
1605 				p = next;
1606 				continue;
1607 			}
1608 
1609 			/*
1610 			 * Page does not exist in parent, rename the
1611 			 * page from the backing object to the main object.
1612 			 *
1613 			 * If the page was mapped to a process, it can remain
1614 			 * mapped through the rename.
1615 			 * vm_page_rename() will handle dirty and cache.
1616 			 */
1617 			if (vm_page_rename(p, object, new_pindex)) {
1618 				if (op & OBSC_COLLAPSE_NOWAIT) {
1619 					p = next;
1620 					continue;
1621 				}
1622 				VM_OBJECT_WLOCK(backing_object);
1623 				VM_OBJECT_WUNLOCK(object);
1624 				VM_WAIT;
1625 				VM_OBJECT_WLOCK(object);
1626 				VM_OBJECT_WLOCK(backing_object);
1627 				p = TAILQ_FIRST(&backing_object->memq);
1628 				continue;
1629 			}
1630 
1631 			/* Use the old pindex to free the right page. */
1632 			if (backing_object->type == OBJT_SWAP)
1633 				swap_pager_freespace(backing_object,
1634 				    new_pindex + backing_offset_index, 1);
1635 
1636 #if VM_NRESERVLEVEL > 0
1637 			/*
1638 			 * Rename the reservation.
1639 			 */
1640 			vm_reserv_rename(p, object, backing_object,
1641 			    backing_offset_index);
1642 #endif
1643 		}
1644 		p = next;
1645 	}
1646 	return (r);
1647 }
1648 
1649 
1650 /*
1651  * this version of collapse allows the operation to occur earlier and
1652  * when paging_in_progress is true for an object...  This is not a complete
1653  * operation, but should plug 99.9% of the rest of the leaks.
1654  */
1655 static void
1656 vm_object_qcollapse(vm_object_t object)
1657 {
1658 	vm_object_t backing_object = object->backing_object;
1659 
1660 	VM_OBJECT_ASSERT_WLOCKED(object);
1661 	VM_OBJECT_ASSERT_WLOCKED(backing_object);
1662 
1663 	if (backing_object->ref_count != 1)
1664 		return;
1665 
1666 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1667 }
1668 
1669 /*
1670  *	vm_object_collapse:
1671  *
1672  *	Collapse an object with the object backing it.
1673  *	Pages in the backing object are moved into the
1674  *	parent, and the backing object is deallocated.
1675  */
1676 void
1677 vm_object_collapse(vm_object_t object)
1678 {
1679 	VM_OBJECT_ASSERT_WLOCKED(object);
1680 
1681 	while (TRUE) {
1682 		vm_object_t backing_object;
1683 
1684 		/*
1685 		 * Verify that the conditions are right for collapse:
1686 		 *
1687 		 * The object exists and the backing object exists.
1688 		 */
1689 		if ((backing_object = object->backing_object) == NULL)
1690 			break;
1691 
1692 		/*
1693 		 * we check the backing object first, because it is most likely
1694 		 * not collapsable.
1695 		 */
1696 		VM_OBJECT_WLOCK(backing_object);
1697 		if (backing_object->handle != NULL ||
1698 		    (backing_object->type != OBJT_DEFAULT &&
1699 		     backing_object->type != OBJT_SWAP) ||
1700 		    (backing_object->flags & OBJ_DEAD) ||
1701 		    object->handle != NULL ||
1702 		    (object->type != OBJT_DEFAULT &&
1703 		     object->type != OBJT_SWAP) ||
1704 		    (object->flags & OBJ_DEAD)) {
1705 			VM_OBJECT_WUNLOCK(backing_object);
1706 			break;
1707 		}
1708 
1709 		if (
1710 		    object->paging_in_progress != 0 ||
1711 		    backing_object->paging_in_progress != 0
1712 		) {
1713 			vm_object_qcollapse(object);
1714 			VM_OBJECT_WUNLOCK(backing_object);
1715 			break;
1716 		}
1717 		/*
1718 		 * We know that we can either collapse the backing object (if
1719 		 * the parent is the only reference to it) or (perhaps) have
1720 		 * the parent bypass the object if the parent happens to shadow
1721 		 * all the resident pages in the entire backing object.
1722 		 *
1723 		 * This is ignoring pager-backed pages such as swap pages.
1724 		 * vm_object_backing_scan fails the shadowing test in this
1725 		 * case.
1726 		 */
1727 		if (backing_object->ref_count == 1) {
1728 			/*
1729 			 * If there is exactly one reference to the backing
1730 			 * object, we can collapse it into the parent.
1731 			 */
1732 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1733 
1734 #if VM_NRESERVLEVEL > 0
1735 			/*
1736 			 * Break any reservations from backing_object.
1737 			 */
1738 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1739 				vm_reserv_break_all(backing_object);
1740 #endif
1741 
1742 			/*
1743 			 * Move the pager from backing_object to object.
1744 			 */
1745 			if (backing_object->type == OBJT_SWAP) {
1746 				/*
1747 				 * swap_pager_copy() can sleep, in which case
1748 				 * the backing_object's and object's locks are
1749 				 * released and reacquired.
1750 				 * Since swap_pager_copy() is being asked to
1751 				 * destroy the source, it will change the
1752 				 * backing_object's type to OBJT_DEFAULT.
1753 				 */
1754 				swap_pager_copy(
1755 				    backing_object,
1756 				    object,
1757 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1758 
1759 				/*
1760 				 * Free any cached pages from backing_object.
1761 				 */
1762 				if (__predict_false(
1763 				    !vm_object_cache_is_empty(backing_object)))
1764 					vm_page_cache_free(backing_object, 0, 0);
1765 			}
1766 			/*
1767 			 * Object now shadows whatever backing_object did.
1768 			 * Note that the reference to
1769 			 * backing_object->backing_object moves from within
1770 			 * backing_object to within object.
1771 			 */
1772 			LIST_REMOVE(object, shadow_list);
1773 			backing_object->shadow_count--;
1774 			if (backing_object->backing_object) {
1775 				VM_OBJECT_WLOCK(backing_object->backing_object);
1776 				LIST_REMOVE(backing_object, shadow_list);
1777 				LIST_INSERT_HEAD(
1778 				    &backing_object->backing_object->shadow_head,
1779 				    object, shadow_list);
1780 				/*
1781 				 * The shadow_count has not changed.
1782 				 */
1783 				VM_OBJECT_WUNLOCK(backing_object->backing_object);
1784 			}
1785 			object->backing_object = backing_object->backing_object;
1786 			object->backing_object_offset +=
1787 			    backing_object->backing_object_offset;
1788 
1789 			/*
1790 			 * Discard backing_object.
1791 			 *
1792 			 * Since the backing object has no pages, no pager left,
1793 			 * and no object references within it, all that is
1794 			 * necessary is to dispose of it.
1795 			 */
1796 			KASSERT(backing_object->ref_count == 1, (
1797 "backing_object %p was somehow re-referenced during collapse!",
1798 			    backing_object));
1799 			VM_OBJECT_WUNLOCK(backing_object);
1800 			vm_object_destroy(backing_object);
1801 
1802 			object_collapses++;
1803 		} else {
1804 			vm_object_t new_backing_object;
1805 
1806 			/*
1807 			 * If we do not entirely shadow the backing object,
1808 			 * there is nothing we can do so we give up.
1809 			 */
1810 			if (object->resident_page_count != object->size &&
1811 			    vm_object_backing_scan(object,
1812 			    OBSC_TEST_ALL_SHADOWED) == 0) {
1813 				VM_OBJECT_WUNLOCK(backing_object);
1814 				break;
1815 			}
1816 
1817 			/*
1818 			 * Make the parent shadow the next object in the
1819 			 * chain.  Deallocating backing_object will not remove
1820 			 * it, since its reference count is at least 2.
1821 			 */
1822 			LIST_REMOVE(object, shadow_list);
1823 			backing_object->shadow_count--;
1824 
1825 			new_backing_object = backing_object->backing_object;
1826 			if ((object->backing_object = new_backing_object) != NULL) {
1827 				VM_OBJECT_WLOCK(new_backing_object);
1828 				LIST_INSERT_HEAD(
1829 				    &new_backing_object->shadow_head,
1830 				    object,
1831 				    shadow_list
1832 				);
1833 				new_backing_object->shadow_count++;
1834 				vm_object_reference_locked(new_backing_object);
1835 				VM_OBJECT_WUNLOCK(new_backing_object);
1836 				object->backing_object_offset +=
1837 					backing_object->backing_object_offset;
1838 			}
1839 
1840 			/*
1841 			 * Drop the reference count on backing_object. Since
1842 			 * its ref_count was at least 2, it will not vanish.
1843 			 */
1844 			backing_object->ref_count--;
1845 			VM_OBJECT_WUNLOCK(backing_object);
1846 			object_bypasses++;
1847 		}
1848 
1849 		/*
1850 		 * Try again with this object's new backing object.
1851 		 */
1852 	}
1853 }
1854 
1855 /*
1856  *	vm_object_page_remove:
1857  *
1858  *	For the given object, either frees or invalidates each of the
1859  *	specified pages.  In general, a page is freed.  However, if a page is
1860  *	wired for any reason other than the existence of a managed, wired
1861  *	mapping, then it may be invalidated but not removed from the object.
1862  *	Pages are specified by the given range ["start", "end") and the option
1863  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
1864  *	extends from "start" to the end of the object.  If the option
1865  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1866  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
1867  *	specified, then the pages within the specified range must have no
1868  *	mappings.  Otherwise, if this option is not specified, any mappings to
1869  *	the specified pages are removed before the pages are freed or
1870  *	invalidated.
1871  *
1872  *	In general, this operation should only be performed on objects that
1873  *	contain managed pages.  There are, however, two exceptions.  First, it
1874  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
1875  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1876  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
1877  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
1878  *
1879  *	The object must be locked.
1880  */
1881 void
1882 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1883     int options)
1884 {
1885 	vm_page_t p, next;
1886 	int wirings;
1887 
1888 	VM_OBJECT_ASSERT_WLOCKED(object);
1889 	KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1890 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1891 	    ("vm_object_page_remove: illegal options for object %p", object));
1892 	if (object->resident_page_count == 0)
1893 		goto skipmemq;
1894 	vm_object_pip_add(object, 1);
1895 again:
1896 	p = vm_page_find_least(object, start);
1897 
1898 	/*
1899 	 * Here, the variable "p" is either (1) the page with the least pindex
1900 	 * greater than or equal to the parameter "start" or (2) NULL.
1901 	 */
1902 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1903 		next = TAILQ_NEXT(p, listq);
1904 
1905 		/*
1906 		 * If the page is wired for any reason besides the existence
1907 		 * of managed, wired mappings, then it cannot be freed.  For
1908 		 * example, fictitious pages, which represent device memory,
1909 		 * are inherently wired and cannot be freed.  They can,
1910 		 * however, be invalidated if the option OBJPR_CLEANONLY is
1911 		 * not specified.
1912 		 */
1913 		vm_page_lock(p);
1914 		if (vm_page_xbusied(p)) {
1915 			VM_OBJECT_WUNLOCK(object);
1916 			vm_page_busy_sleep(p, "vmopax");
1917 			VM_OBJECT_WLOCK(object);
1918 			goto again;
1919 		}
1920 		if ((wirings = p->wire_count) != 0 &&
1921 		    (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1922 			if ((options & (OBJPR_NOTWIRED | OBJPR_NOTMAPPED)) ==
1923 			    0) {
1924 				pmap_remove_all(p);
1925 				/* Account for removal of wired mappings. */
1926 				if (wirings != 0)
1927 					p->wire_count -= wirings;
1928 			}
1929 			if ((options & OBJPR_CLEANONLY) == 0) {
1930 				p->valid = 0;
1931 				vm_page_undirty(p);
1932 			}
1933 			goto next;
1934 		}
1935 		if (vm_page_busied(p)) {
1936 			VM_OBJECT_WUNLOCK(object);
1937 			vm_page_busy_sleep(p, "vmopar");
1938 			VM_OBJECT_WLOCK(object);
1939 			goto again;
1940 		}
1941 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
1942 		    ("vm_object_page_remove: page %p is fictitious", p));
1943 		if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1944 			if ((options & OBJPR_NOTMAPPED) == 0)
1945 				pmap_remove_write(p);
1946 			if (p->dirty)
1947 				goto next;
1948 		}
1949 		if ((options & OBJPR_NOTMAPPED) == 0) {
1950 			if ((options & OBJPR_NOTWIRED) != 0 && wirings != 0)
1951 				goto next;
1952 			pmap_remove_all(p);
1953 			/* Account for removal of wired mappings. */
1954 			if (wirings != 0) {
1955 				KASSERT(p->wire_count == wirings,
1956 				    ("inconsistent wire count %d %d %p",
1957 				    p->wire_count, wirings, p));
1958 				p->wire_count = 0;
1959 				atomic_subtract_int(&cnt.v_wire_count, 1);
1960 			}
1961 		}
1962 		vm_page_free(p);
1963 next:
1964 		vm_page_unlock(p);
1965 	}
1966 	vm_object_pip_wakeup(object);
1967 skipmemq:
1968 	if (__predict_false(!vm_object_cache_is_empty(object)))
1969 		vm_page_cache_free(object, start, end);
1970 }
1971 
1972 /*
1973  *	vm_object_page_cache:
1974  *
1975  *	For the given object, attempt to move the specified clean
1976  *	pages to the cache queue.  If a page is wired for any reason,
1977  *	then it will not be changed.  Pages are specified by the given
1978  *	range ["start", "end").  As a special case, if "end" is zero,
1979  *	then the range extends from "start" to the end of the object.
1980  *	Any mappings to the specified pages are removed before the
1981  *	pages are moved to the cache queue.
1982  *
1983  *	This operation should only be performed on objects that
1984  *	contain non-fictitious, managed pages.
1985  *
1986  *	The object must be locked.
1987  */
1988 void
1989 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1990 {
1991 	struct mtx *mtx, *new_mtx;
1992 	vm_page_t p, next;
1993 
1994 	VM_OBJECT_ASSERT_WLOCKED(object);
1995 	KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1996 	    ("vm_object_page_cache: illegal object %p", object));
1997 	if (object->resident_page_count == 0)
1998 		return;
1999 	p = vm_page_find_least(object, start);
2000 
2001 	/*
2002 	 * Here, the variable "p" is either (1) the page with the least pindex
2003 	 * greater than or equal to the parameter "start" or (2) NULL.
2004 	 */
2005 	mtx = NULL;
2006 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2007 		next = TAILQ_NEXT(p, listq);
2008 
2009 		/*
2010 		 * Avoid releasing and reacquiring the same page lock.
2011 		 */
2012 		new_mtx = vm_page_lockptr(p);
2013 		if (mtx != new_mtx) {
2014 			if (mtx != NULL)
2015 				mtx_unlock(mtx);
2016 			mtx = new_mtx;
2017 			mtx_lock(mtx);
2018 		}
2019 		vm_page_try_to_cache(p);
2020 	}
2021 	if (mtx != NULL)
2022 		mtx_unlock(mtx);
2023 }
2024 
2025 /*
2026  *	Populate the specified range of the object with valid pages.  Returns
2027  *	TRUE if the range is successfully populated and FALSE otherwise.
2028  *
2029  *	Note: This function should be optimized to pass a larger array of
2030  *	pages to vm_pager_get_pages() before it is applied to a non-
2031  *	OBJT_DEVICE object.
2032  *
2033  *	The object must be locked.
2034  */
2035 boolean_t
2036 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2037 {
2038 	vm_page_t m, ma[1];
2039 	vm_pindex_t pindex;
2040 	int rv;
2041 
2042 	VM_OBJECT_ASSERT_WLOCKED(object);
2043 	for (pindex = start; pindex < end; pindex++) {
2044 		m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
2045 		if (m->valid != VM_PAGE_BITS_ALL) {
2046 			ma[0] = m;
2047 			rv = vm_pager_get_pages(object, ma, 1, 0);
2048 			m = vm_page_lookup(object, pindex);
2049 			if (m == NULL)
2050 				break;
2051 			if (rv != VM_PAGER_OK) {
2052 				vm_page_lock(m);
2053 				vm_page_free(m);
2054 				vm_page_unlock(m);
2055 				break;
2056 			}
2057 		}
2058 		/*
2059 		 * Keep "m" busy because a subsequent iteration may unlock
2060 		 * the object.
2061 		 */
2062 	}
2063 	if (pindex > start) {
2064 		m = vm_page_lookup(object, start);
2065 		while (m != NULL && m->pindex < pindex) {
2066 			vm_page_xunbusy(m);
2067 			m = TAILQ_NEXT(m, listq);
2068 		}
2069 	}
2070 	return (pindex == end);
2071 }
2072 
2073 /*
2074  *	Routine:	vm_object_coalesce
2075  *	Function:	Coalesces two objects backing up adjoining
2076  *			regions of memory into a single object.
2077  *
2078  *	returns TRUE if objects were combined.
2079  *
2080  *	NOTE:	Only works at the moment if the second object is NULL -
2081  *		if it's not, which object do we lock first?
2082  *
2083  *	Parameters:
2084  *		prev_object	First object to coalesce
2085  *		prev_offset	Offset into prev_object
2086  *		prev_size	Size of reference to prev_object
2087  *		next_size	Size of reference to the second object
2088  *		reserved	Indicator that extension region has
2089  *				swap accounted for
2090  *
2091  *	Conditions:
2092  *	The object must *not* be locked.
2093  */
2094 boolean_t
2095 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2096     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2097 {
2098 	vm_pindex_t next_pindex;
2099 
2100 	if (prev_object == NULL)
2101 		return (TRUE);
2102 	VM_OBJECT_WLOCK(prev_object);
2103 	if ((prev_object->type != OBJT_DEFAULT &&
2104 	    prev_object->type != OBJT_SWAP) ||
2105 	    (prev_object->flags & OBJ_TMPFS) != 0) {
2106 		VM_OBJECT_WUNLOCK(prev_object);
2107 		return (FALSE);
2108 	}
2109 
2110 	/*
2111 	 * Try to collapse the object first
2112 	 */
2113 	vm_object_collapse(prev_object);
2114 
2115 	/*
2116 	 * Can't coalesce if: . more than one reference . paged out . shadows
2117 	 * another object . has a copy elsewhere (any of which mean that the
2118 	 * pages not mapped to prev_entry may be in use anyway)
2119 	 */
2120 	if (prev_object->backing_object != NULL) {
2121 		VM_OBJECT_WUNLOCK(prev_object);
2122 		return (FALSE);
2123 	}
2124 
2125 	prev_size >>= PAGE_SHIFT;
2126 	next_size >>= PAGE_SHIFT;
2127 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2128 
2129 	if ((prev_object->ref_count > 1) &&
2130 	    (prev_object->size != next_pindex)) {
2131 		VM_OBJECT_WUNLOCK(prev_object);
2132 		return (FALSE);
2133 	}
2134 
2135 	/*
2136 	 * Account for the charge.
2137 	 */
2138 	if (prev_object->cred != NULL) {
2139 
2140 		/*
2141 		 * If prev_object was charged, then this mapping,
2142 		 * althought not charged now, may become writable
2143 		 * later. Non-NULL cred in the object would prevent
2144 		 * swap reservation during enabling of the write
2145 		 * access, so reserve swap now. Failed reservation
2146 		 * cause allocation of the separate object for the map
2147 		 * entry, and swap reservation for this entry is
2148 		 * managed in appropriate time.
2149 		 */
2150 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2151 		    prev_object->cred)) {
2152 			return (FALSE);
2153 		}
2154 		prev_object->charge += ptoa(next_size);
2155 	}
2156 
2157 	/*
2158 	 * Remove any pages that may still be in the object from a previous
2159 	 * deallocation.
2160 	 */
2161 	if (next_pindex < prev_object->size) {
2162 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2163 		    next_size, 0);
2164 		if (prev_object->type == OBJT_SWAP)
2165 			swap_pager_freespace(prev_object,
2166 					     next_pindex, next_size);
2167 #if 0
2168 		if (prev_object->cred != NULL) {
2169 			KASSERT(prev_object->charge >=
2170 			    ptoa(prev_object->size - next_pindex),
2171 			    ("object %p overcharged 1 %jx %jx", prev_object,
2172 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2173 			prev_object->charge -= ptoa(prev_object->size -
2174 			    next_pindex);
2175 		}
2176 #endif
2177 	}
2178 
2179 	/*
2180 	 * Extend the object if necessary.
2181 	 */
2182 	if (next_pindex + next_size > prev_object->size)
2183 		prev_object->size = next_pindex + next_size;
2184 
2185 	VM_OBJECT_WUNLOCK(prev_object);
2186 	return (TRUE);
2187 }
2188 
2189 void
2190 vm_object_set_writeable_dirty(vm_object_t object)
2191 {
2192 
2193 	VM_OBJECT_ASSERT_WLOCKED(object);
2194 	if (object->type != OBJT_VNODE)
2195 		return;
2196 	object->generation++;
2197 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2198 		return;
2199 	vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2200 }
2201 
2202 #include "opt_ddb.h"
2203 #ifdef DDB
2204 #include <sys/kernel.h>
2205 
2206 #include <sys/cons.h>
2207 
2208 #include <ddb/ddb.h>
2209 
2210 static int
2211 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2212 {
2213 	vm_map_t tmpm;
2214 	vm_map_entry_t tmpe;
2215 	vm_object_t obj;
2216 	int entcount;
2217 
2218 	if (map == 0)
2219 		return 0;
2220 
2221 	if (entry == 0) {
2222 		tmpe = map->header.next;
2223 		entcount = map->nentries;
2224 		while (entcount-- && (tmpe != &map->header)) {
2225 			if (_vm_object_in_map(map, object, tmpe)) {
2226 				return 1;
2227 			}
2228 			tmpe = tmpe->next;
2229 		}
2230 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2231 		tmpm = entry->object.sub_map;
2232 		tmpe = tmpm->header.next;
2233 		entcount = tmpm->nentries;
2234 		while (entcount-- && tmpe != &tmpm->header) {
2235 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2236 				return 1;
2237 			}
2238 			tmpe = tmpe->next;
2239 		}
2240 	} else if ((obj = entry->object.vm_object) != NULL) {
2241 		for (; obj; obj = obj->backing_object)
2242 			if (obj == object) {
2243 				return 1;
2244 			}
2245 	}
2246 	return 0;
2247 }
2248 
2249 static int
2250 vm_object_in_map(vm_object_t object)
2251 {
2252 	struct proc *p;
2253 
2254 	/* sx_slock(&allproc_lock); */
2255 	FOREACH_PROC_IN_SYSTEM(p) {
2256 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2257 			continue;
2258 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2259 			/* sx_sunlock(&allproc_lock); */
2260 			return 1;
2261 		}
2262 	}
2263 	/* sx_sunlock(&allproc_lock); */
2264 	if (_vm_object_in_map(kernel_map, object, 0))
2265 		return 1;
2266 	return 0;
2267 }
2268 
2269 DB_SHOW_COMMAND(vmochk, vm_object_check)
2270 {
2271 	vm_object_t object;
2272 
2273 	/*
2274 	 * make sure that internal objs are in a map somewhere
2275 	 * and none have zero ref counts.
2276 	 */
2277 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2278 		if (object->handle == NULL &&
2279 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2280 			if (object->ref_count == 0) {
2281 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2282 					(long)object->size);
2283 			}
2284 			if (!vm_object_in_map(object)) {
2285 				db_printf(
2286 			"vmochk: internal obj is not in a map: "
2287 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2288 				    object->ref_count, (u_long)object->size,
2289 				    (u_long)object->size,
2290 				    (void *)object->backing_object);
2291 			}
2292 		}
2293 	}
2294 }
2295 
2296 /*
2297  *	vm_object_print:	[ debug ]
2298  */
2299 DB_SHOW_COMMAND(object, vm_object_print_static)
2300 {
2301 	/* XXX convert args. */
2302 	vm_object_t object = (vm_object_t)addr;
2303 	boolean_t full = have_addr;
2304 
2305 	vm_page_t p;
2306 
2307 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2308 #define	count	was_count
2309 
2310 	int count;
2311 
2312 	if (object == NULL)
2313 		return;
2314 
2315 	db_iprintf(
2316 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2317 	    object, (int)object->type, (uintmax_t)object->size,
2318 	    object->resident_page_count, object->ref_count, object->flags,
2319 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2320 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2321 	    object->shadow_count,
2322 	    object->backing_object ? object->backing_object->ref_count : 0,
2323 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2324 
2325 	if (!full)
2326 		return;
2327 
2328 	db_indent += 2;
2329 	count = 0;
2330 	TAILQ_FOREACH(p, &object->memq, listq) {
2331 		if (count == 0)
2332 			db_iprintf("memory:=");
2333 		else if (count == 6) {
2334 			db_printf("\n");
2335 			db_iprintf(" ...");
2336 			count = 0;
2337 		} else
2338 			db_printf(",");
2339 		count++;
2340 
2341 		db_printf("(off=0x%jx,page=0x%jx)",
2342 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2343 	}
2344 	if (count != 0)
2345 		db_printf("\n");
2346 	db_indent -= 2;
2347 }
2348 
2349 /* XXX. */
2350 #undef count
2351 
2352 /* XXX need this non-static entry for calling from vm_map_print. */
2353 void
2354 vm_object_print(
2355         /* db_expr_t */ long addr,
2356 	boolean_t have_addr,
2357 	/* db_expr_t */ long count,
2358 	char *modif)
2359 {
2360 	vm_object_print_static(addr, have_addr, count, modif);
2361 }
2362 
2363 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2364 {
2365 	vm_object_t object;
2366 	vm_pindex_t fidx;
2367 	vm_paddr_t pa;
2368 	vm_page_t m, prev_m;
2369 	int rcount, nl, c;
2370 
2371 	nl = 0;
2372 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2373 		db_printf("new object: %p\n", (void *)object);
2374 		if (nl > 18) {
2375 			c = cngetc();
2376 			if (c != ' ')
2377 				return;
2378 			nl = 0;
2379 		}
2380 		nl++;
2381 		rcount = 0;
2382 		fidx = 0;
2383 		pa = -1;
2384 		TAILQ_FOREACH(m, &object->memq, listq) {
2385 			if (m->pindex > 128)
2386 				break;
2387 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2388 			    prev_m->pindex + 1 != m->pindex) {
2389 				if (rcount) {
2390 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2391 						(long)fidx, rcount, (long)pa);
2392 					if (nl > 18) {
2393 						c = cngetc();
2394 						if (c != ' ')
2395 							return;
2396 						nl = 0;
2397 					}
2398 					nl++;
2399 					rcount = 0;
2400 				}
2401 			}
2402 			if (rcount &&
2403 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2404 				++rcount;
2405 				continue;
2406 			}
2407 			if (rcount) {
2408 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2409 					(long)fidx, rcount, (long)pa);
2410 				if (nl > 18) {
2411 					c = cngetc();
2412 					if (c != ' ')
2413 						return;
2414 					nl = 0;
2415 				}
2416 				nl++;
2417 			}
2418 			fidx = m->pindex;
2419 			pa = VM_PAGE_TO_PHYS(m);
2420 			rcount = 1;
2421 		}
2422 		if (rcount) {
2423 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2424 				(long)fidx, rcount, (long)pa);
2425 			if (nl > 18) {
2426 				c = cngetc();
2427 				if (c != ' ')
2428 					return;
2429 				nl = 0;
2430 			}
2431 			nl++;
2432 		}
2433 	}
2434 }
2435 #endif /* DDB */
2436