xref: /freebsd/sys/vm/vm_object.c (revision d876124d6ae9d56da5b4ff4c6015efd1d0c9222a)
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/vnode.h>
81 #include <sys/vmmeter.h>
82 #include <sys/sx.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_param.h>
86 #include <vm/pmap.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_pager.h>
92 #include <vm/swap_pager.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
95 #include <vm/vm_reserv.h>
96 #include <vm/uma.h>
97 
98 #define EASY_SCAN_FACTOR       8
99 
100 #define MSYNC_FLUSH_HARDSEQ	0x01
101 #define MSYNC_FLUSH_SOFTSEQ	0x02
102 
103 /*
104  * msync / VM object flushing optimizations
105  */
106 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ;
107 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags,
108         CTLFLAG_RW, &msync_flush_flags, 0, "");
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 void	vm_object_qcollapse(vm_object_t object);
115 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
116 static void	vm_object_vndeallocate(vm_object_t object);
117 
118 /*
119  *	Virtual memory objects maintain the actual data
120  *	associated with allocated virtual memory.  A given
121  *	page of memory exists within exactly one object.
122  *
123  *	An object is only deallocated when all "references"
124  *	are given up.  Only one "reference" to a given
125  *	region of an object should be writeable.
126  *
127  *	Associated with each object is a list of all resident
128  *	memory pages belonging to that object; this list is
129  *	maintained by the "vm_page" module, and locked by the object's
130  *	lock.
131  *
132  *	Each object also records a "pager" routine which is
133  *	used to retrieve (and store) pages to the proper backing
134  *	storage.  In addition, objects may be backed by other
135  *	objects from which they were virtual-copied.
136  *
137  *	The only items within the object structure which are
138  *	modified after time of creation are:
139  *		reference count		locked by object's lock
140  *		pager routine		locked by object's lock
141  *
142  */
143 
144 struct object_q vm_object_list;
145 struct mtx vm_object_list_mtx;	/* lock for object list and count */
146 
147 struct vm_object kernel_object_store;
148 struct vm_object kmem_object_store;
149 
150 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats");
151 
152 static long object_collapses;
153 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
154     &object_collapses, 0, "VM object collapses");
155 
156 static long object_bypasses;
157 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
158     &object_bypasses, 0, "VM object bypasses");
159 
160 static uma_zone_t obj_zone;
161 
162 static int vm_object_zinit(void *mem, int size, int flags);
163 
164 #ifdef INVARIANTS
165 static void vm_object_zdtor(void *mem, int size, void *arg);
166 
167 static void
168 vm_object_zdtor(void *mem, int size, void *arg)
169 {
170 	vm_object_t object;
171 
172 	object = (vm_object_t)mem;
173 	KASSERT(TAILQ_EMPTY(&object->memq),
174 	    ("object %p has resident pages",
175 	    object));
176 #if VM_NRESERVLEVEL > 0
177 	KASSERT(LIST_EMPTY(&object->rvq),
178 	    ("object %p has reservations",
179 	    object));
180 #endif
181 	KASSERT(object->cache == NULL,
182 	    ("object %p has cached pages",
183 	    object));
184 	KASSERT(object->paging_in_progress == 0,
185 	    ("object %p paging_in_progress = %d",
186 	    object, object->paging_in_progress));
187 	KASSERT(object->resident_page_count == 0,
188 	    ("object %p resident_page_count = %d",
189 	    object, object->resident_page_count));
190 	KASSERT(object->shadow_count == 0,
191 	    ("object %p shadow_count = %d",
192 	    object, object->shadow_count));
193 }
194 #endif
195 
196 static int
197 vm_object_zinit(void *mem, int size, int flags)
198 {
199 	vm_object_t object;
200 
201 	object = (vm_object_t)mem;
202 	bzero(&object->mtx, sizeof(object->mtx));
203 	VM_OBJECT_LOCK_INIT(object, "standard object");
204 
205 	/* These are true for any object that has been freed */
206 	object->paging_in_progress = 0;
207 	object->resident_page_count = 0;
208 	object->shadow_count = 0;
209 	return (0);
210 }
211 
212 void
213 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
214 {
215 
216 	TAILQ_INIT(&object->memq);
217 	LIST_INIT(&object->shadow_head);
218 
219 	object->root = NULL;
220 	object->type = type;
221 	object->size = size;
222 	object->generation = 1;
223 	object->ref_count = 1;
224 	object->flags = 0;
225 	if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
226 		object->flags = OBJ_ONEMAPPING;
227 	object->pg_color = 0;
228 	object->handle = NULL;
229 	object->backing_object = NULL;
230 	object->backing_object_offset = (vm_ooffset_t) 0;
231 #if VM_NRESERVLEVEL > 0
232 	LIST_INIT(&object->rvq);
233 #endif
234 	object->cache = NULL;
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 }
240 
241 /*
242  *	vm_object_init:
243  *
244  *	Initialize the VM objects module.
245  */
246 void
247 vm_object_init(void)
248 {
249 	TAILQ_INIT(&vm_object_list);
250 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
251 
252 	VM_OBJECT_LOCK_INIT(&kernel_object_store, "kernel object");
253 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
254 	    kernel_object);
255 #if VM_NRESERVLEVEL > 0
256 	kernel_object->flags |= OBJ_COLORED;
257 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
258 #endif
259 
260 	VM_OBJECT_LOCK_INIT(&kmem_object_store, "kmem object");
261 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
262 	    kmem_object);
263 #if VM_NRESERVLEVEL > 0
264 	kmem_object->flags |= OBJ_COLORED;
265 	kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
266 #endif
267 
268 	/*
269 	 * The lock portion of struct vm_object must be type stable due
270 	 * to vm_pageout_fallback_object_lock locking a vm object
271 	 * without holding any references to it.
272 	 */
273 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
274 #ifdef INVARIANTS
275 	    vm_object_zdtor,
276 #else
277 	    NULL,
278 #endif
279 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
280 }
281 
282 void
283 vm_object_clear_flag(vm_object_t object, u_short bits)
284 {
285 
286 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
287 	object->flags &= ~bits;
288 }
289 
290 void
291 vm_object_pip_add(vm_object_t object, short i)
292 {
293 
294 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
295 	object->paging_in_progress += i;
296 }
297 
298 void
299 vm_object_pip_subtract(vm_object_t object, short i)
300 {
301 
302 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
303 	object->paging_in_progress -= i;
304 }
305 
306 void
307 vm_object_pip_wakeup(vm_object_t object)
308 {
309 
310 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
311 	object->paging_in_progress--;
312 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
313 		vm_object_clear_flag(object, OBJ_PIPWNT);
314 		wakeup(object);
315 	}
316 }
317 
318 void
319 vm_object_pip_wakeupn(vm_object_t object, short i)
320 {
321 
322 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
323 	if (i)
324 		object->paging_in_progress -= i;
325 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
326 		vm_object_clear_flag(object, OBJ_PIPWNT);
327 		wakeup(object);
328 	}
329 }
330 
331 void
332 vm_object_pip_wait(vm_object_t object, char *waitid)
333 {
334 
335 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
336 	while (object->paging_in_progress) {
337 		object->flags |= OBJ_PIPWNT;
338 		msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
339 	}
340 }
341 
342 /*
343  *	vm_object_allocate:
344  *
345  *	Returns a new object with the given size.
346  */
347 vm_object_t
348 vm_object_allocate(objtype_t type, vm_pindex_t size)
349 {
350 	vm_object_t object;
351 
352 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
353 	_vm_object_allocate(type, size, object);
354 	return (object);
355 }
356 
357 
358 /*
359  *	vm_object_reference:
360  *
361  *	Gets another reference to the given object.  Note: OBJ_DEAD
362  *	objects can be referenced during final cleaning.
363  */
364 void
365 vm_object_reference(vm_object_t object)
366 {
367 	if (object == NULL)
368 		return;
369 	VM_OBJECT_LOCK(object);
370 	vm_object_reference_locked(object);
371 	VM_OBJECT_UNLOCK(object);
372 }
373 
374 /*
375  *	vm_object_reference_locked:
376  *
377  *	Gets another reference to the given object.
378  *
379  *	The object must be locked.
380  */
381 void
382 vm_object_reference_locked(vm_object_t object)
383 {
384 	struct vnode *vp;
385 
386 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
387 	object->ref_count++;
388 	if (object->type == OBJT_VNODE) {
389 		vp = object->handle;
390 		vref(vp);
391 	}
392 }
393 
394 /*
395  * Handle deallocating an object of type OBJT_VNODE.
396  */
397 static void
398 vm_object_vndeallocate(vm_object_t object)
399 {
400 	struct vnode *vp = (struct vnode *) object->handle;
401 
402 	VFS_ASSERT_GIANT(vp->v_mount);
403 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
404 	KASSERT(object->type == OBJT_VNODE,
405 	    ("vm_object_vndeallocate: not a vnode object"));
406 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
407 #ifdef INVARIANTS
408 	if (object->ref_count == 0) {
409 		vprint("vm_object_vndeallocate", vp);
410 		panic("vm_object_vndeallocate: bad object reference count");
411 	}
412 #endif
413 
414 	object->ref_count--;
415 	if (object->ref_count == 0) {
416 		mp_fixme("Unlocked vflag access.");
417 		vp->v_vflag &= ~VV_TEXT;
418 	}
419 	VM_OBJECT_UNLOCK(object);
420 	/*
421 	 * vrele may need a vop lock
422 	 */
423 	vrele(vp);
424 }
425 
426 /*
427  *	vm_object_deallocate:
428  *
429  *	Release a reference to the specified object,
430  *	gained either through a vm_object_allocate
431  *	or a vm_object_reference call.  When all references
432  *	are gone, storage associated with this object
433  *	may be relinquished.
434  *
435  *	No object may be locked.
436  */
437 void
438 vm_object_deallocate(vm_object_t object)
439 {
440 	vm_object_t temp;
441 
442 	while (object != NULL) {
443 		int vfslocked;
444 
445 		vfslocked = 0;
446 	restart:
447 		VM_OBJECT_LOCK(object);
448 		if (object->type == OBJT_VNODE) {
449 			struct vnode *vp = (struct vnode *) object->handle;
450 
451 			/*
452 			 * Conditionally acquire Giant for a vnode-backed
453 			 * object.  We have to be careful since the type of
454 			 * a vnode object can change while the object is
455 			 * unlocked.
456 			 */
457 			if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
458 				vfslocked = 1;
459 				if (!mtx_trylock(&Giant)) {
460 					VM_OBJECT_UNLOCK(object);
461 					mtx_lock(&Giant);
462 					goto restart;
463 				}
464 			}
465 			vm_object_vndeallocate(object);
466 			VFS_UNLOCK_GIANT(vfslocked);
467 			return;
468 		} else
469 			/*
470 			 * This is to handle the case that the object
471 			 * changed type while we dropped its lock to
472 			 * obtain Giant.
473 			 */
474 			VFS_UNLOCK_GIANT(vfslocked);
475 
476 		KASSERT(object->ref_count != 0,
477 			("vm_object_deallocate: object deallocated too many times: %d", object->type));
478 
479 		/*
480 		 * If the reference count goes to 0 we start calling
481 		 * vm_object_terminate() on the object chain.
482 		 * A ref count of 1 may be a special case depending on the
483 		 * shadow count being 0 or 1.
484 		 */
485 		object->ref_count--;
486 		if (object->ref_count > 1) {
487 			VM_OBJECT_UNLOCK(object);
488 			return;
489 		} else if (object->ref_count == 1) {
490 			if (object->shadow_count == 0 &&
491 			    object->handle == NULL &&
492 			    (object->type == OBJT_DEFAULT ||
493 			     object->type == OBJT_SWAP)) {
494 				vm_object_set_flag(object, OBJ_ONEMAPPING);
495 			} else if ((object->shadow_count == 1) &&
496 			    (object->handle == NULL) &&
497 			    (object->type == OBJT_DEFAULT ||
498 			     object->type == OBJT_SWAP)) {
499 				vm_object_t robject;
500 
501 				robject = LIST_FIRST(&object->shadow_head);
502 				KASSERT(robject != NULL,
503 				    ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
504 					 object->ref_count,
505 					 object->shadow_count));
506 				if (!VM_OBJECT_TRYLOCK(robject)) {
507 					/*
508 					 * Avoid a potential deadlock.
509 					 */
510 					object->ref_count++;
511 					VM_OBJECT_UNLOCK(object);
512 					/*
513 					 * More likely than not the thread
514 					 * holding robject's lock has lower
515 					 * priority than the current thread.
516 					 * Let the lower priority thread run.
517 					 */
518 					pause("vmo_de", 1);
519 					continue;
520 				}
521 				/*
522 				 * Collapse object into its shadow unless its
523 				 * shadow is dead.  In that case, object will
524 				 * be deallocated by the thread that is
525 				 * deallocating its shadow.
526 				 */
527 				if ((robject->flags & OBJ_DEAD) == 0 &&
528 				    (robject->handle == NULL) &&
529 				    (robject->type == OBJT_DEFAULT ||
530 				     robject->type == OBJT_SWAP)) {
531 
532 					robject->ref_count++;
533 retry:
534 					if (robject->paging_in_progress) {
535 						VM_OBJECT_UNLOCK(object);
536 						vm_object_pip_wait(robject,
537 						    "objde1");
538 						temp = robject->backing_object;
539 						if (object == temp) {
540 							VM_OBJECT_LOCK(object);
541 							goto retry;
542 						}
543 					} else if (object->paging_in_progress) {
544 						VM_OBJECT_UNLOCK(robject);
545 						object->flags |= OBJ_PIPWNT;
546 						msleep(object,
547 						    VM_OBJECT_MTX(object),
548 						    PDROP | PVM, "objde2", 0);
549 						VM_OBJECT_LOCK(robject);
550 						temp = robject->backing_object;
551 						if (object == temp) {
552 							VM_OBJECT_LOCK(object);
553 							goto retry;
554 						}
555 					} else
556 						VM_OBJECT_UNLOCK(object);
557 
558 					if (robject->ref_count == 1) {
559 						robject->ref_count--;
560 						object = robject;
561 						goto doterm;
562 					}
563 					object = robject;
564 					vm_object_collapse(object);
565 					VM_OBJECT_UNLOCK(object);
566 					continue;
567 				}
568 				VM_OBJECT_UNLOCK(robject);
569 			}
570 			VM_OBJECT_UNLOCK(object);
571 			return;
572 		}
573 doterm:
574 		temp = object->backing_object;
575 		if (temp != NULL) {
576 			VM_OBJECT_LOCK(temp);
577 			LIST_REMOVE(object, shadow_list);
578 			temp->shadow_count--;
579 			temp->generation++;
580 			VM_OBJECT_UNLOCK(temp);
581 			object->backing_object = NULL;
582 		}
583 		/*
584 		 * Don't double-terminate, we could be in a termination
585 		 * recursion due to the terminate having to sync data
586 		 * to disk.
587 		 */
588 		if ((object->flags & OBJ_DEAD) == 0)
589 			vm_object_terminate(object);
590 		else
591 			VM_OBJECT_UNLOCK(object);
592 		object = temp;
593 	}
594 }
595 
596 /*
597  *	vm_object_terminate actually destroys the specified object, freeing
598  *	up all previously used resources.
599  *
600  *	The object must be locked.
601  *	This routine may block.
602  */
603 void
604 vm_object_terminate(vm_object_t object)
605 {
606 	vm_page_t p;
607 
608 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
609 
610 	/*
611 	 * Make sure no one uses us.
612 	 */
613 	vm_object_set_flag(object, OBJ_DEAD);
614 
615 	/*
616 	 * wait for the pageout daemon to be done with the object
617 	 */
618 	vm_object_pip_wait(object, "objtrm");
619 
620 	KASSERT(!object->paging_in_progress,
621 		("vm_object_terminate: pageout in progress"));
622 
623 	/*
624 	 * Clean and free the pages, as appropriate. All references to the
625 	 * object are gone, so we don't need to lock it.
626 	 */
627 	if (object->type == OBJT_VNODE) {
628 		struct vnode *vp = (struct vnode *)object->handle;
629 
630 		/*
631 		 * Clean pages and flush buffers.
632 		 */
633 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
634 		VM_OBJECT_UNLOCK(object);
635 
636 		vinvalbuf(vp, V_SAVE, NULL, 0, 0);
637 
638 		VM_OBJECT_LOCK(object);
639 	}
640 
641 	KASSERT(object->ref_count == 0,
642 		("vm_object_terminate: object with references, ref_count=%d",
643 		object->ref_count));
644 
645 	/*
646 	 * Now free any remaining pages. For internal objects, this also
647 	 * removes them from paging queues. Don't free wired pages, just
648 	 * remove them from the object.
649 	 */
650 	vm_page_lock_queues();
651 	while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
652 		KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
653 			("vm_object_terminate: freeing busy page %p "
654 			"p->busy = %d, p->flags %x\n", p, p->busy, p->flags));
655 		if (p->wire_count == 0) {
656 			vm_page_free(p);
657 			cnt.v_pfree++;
658 		} else {
659 			vm_page_remove(p);
660 		}
661 	}
662 	vm_page_unlock_queues();
663 
664 #if VM_NRESERVLEVEL > 0
665 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
666 		vm_reserv_break_all(object);
667 #endif
668 	if (__predict_false(object->cache != NULL))
669 		vm_page_cache_free(object, 0, 0);
670 
671 	/*
672 	 * Let the pager know object is dead.
673 	 */
674 	vm_pager_deallocate(object);
675 	VM_OBJECT_UNLOCK(object);
676 
677 	/*
678 	 * Remove the object from the global object list.
679 	 */
680 	mtx_lock(&vm_object_list_mtx);
681 	TAILQ_REMOVE(&vm_object_list, object, object_list);
682 	mtx_unlock(&vm_object_list_mtx);
683 
684 	/*
685 	 * Free the space for the object.
686 	 */
687 	uma_zfree(obj_zone, object);
688 }
689 
690 /*
691  *	vm_object_page_clean
692  *
693  *	Clean all dirty pages in the specified range of object.  Leaves page
694  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
695  *	write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
696  *	leaving the object dirty.
697  *
698  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
699  *	synchronous clustering mode implementation.
700  *
701  *	Odd semantics: if start == end, we clean everything.
702  *
703  *	The object must be locked.
704  */
705 void
706 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags)
707 {
708 	vm_page_t p, np;
709 	vm_pindex_t tstart, tend;
710 	vm_pindex_t pi;
711 	int clearobjflags;
712 	int pagerflags;
713 	int curgeneration;
714 
715 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
716 	if (object->type != OBJT_VNODE ||
717 		(object->flags & OBJ_MIGHTBEDIRTY) == 0)
718 		return;
719 
720 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
721 	pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
722 
723 	vm_object_set_flag(object, OBJ_CLEANING);
724 
725 	tstart = start;
726 	if (end == 0) {
727 		tend = object->size;
728 	} else {
729 		tend = end;
730 	}
731 
732 	vm_page_lock_queues();
733 	/*
734 	 * If the caller is smart and only msync()s a range he knows is
735 	 * dirty, we may be able to avoid an object scan.  This results in
736 	 * a phenominal improvement in performance.  We cannot do this
737 	 * as a matter of course because the object may be huge - e.g.
738 	 * the size might be in the gigabytes or terrabytes.
739 	 */
740 	if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
741 		vm_pindex_t tscan;
742 		int scanlimit;
743 		int scanreset;
744 
745 		scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
746 		if (scanreset < 16)
747 			scanreset = 16;
748 		pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
749 
750 		scanlimit = scanreset;
751 		tscan = tstart;
752 		while (tscan < tend) {
753 			curgeneration = object->generation;
754 			p = vm_page_lookup(object, tscan);
755 			if (p == NULL || p->valid == 0) {
756 				if (--scanlimit == 0)
757 					break;
758 				++tscan;
759 				continue;
760 			}
761 			vm_page_test_dirty(p);
762 			if ((p->dirty & p->valid) == 0) {
763 				if (--scanlimit == 0)
764 					break;
765 				++tscan;
766 				continue;
767 			}
768 			/*
769 			 * If we have been asked to skip nosync pages and
770 			 * this is a nosync page, we can't continue.
771 			 */
772 			if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
773 				if (--scanlimit == 0)
774 					break;
775 				++tscan;
776 				continue;
777 			}
778 			scanlimit = scanreset;
779 
780 			/*
781 			 * This returns 0 if it was unable to busy the first
782 			 * page (i.e. had to sleep).
783 			 */
784 			tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
785 		}
786 
787 		/*
788 		 * If everything was dirty and we flushed it successfully,
789 		 * and the requested range is not the entire object, we
790 		 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
791 		 * return immediately.
792 		 */
793 		if (tscan >= tend && (tstart || tend < object->size)) {
794 			vm_page_unlock_queues();
795 			vm_object_clear_flag(object, OBJ_CLEANING);
796 			return;
797 		}
798 		pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
799 	}
800 
801 	/*
802 	 * Generally set CLEANCHK interlock and make the page read-only so
803 	 * we can then clear the object flags.
804 	 *
805 	 * However, if this is a nosync mmap then the object is likely to
806 	 * stay dirty so do not mess with the page and do not clear the
807 	 * object flags.
808 	 */
809 	clearobjflags = 1;
810 	TAILQ_FOREACH(p, &object->memq, listq) {
811 		p->oflags |= VPO_CLEANCHK;
812 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC))
813 			clearobjflags = 0;
814 		else
815 			pmap_remove_write(p);
816 	}
817 
818 	if (clearobjflags && (tstart == 0) && (tend == object->size)) {
819 		struct vnode *vp;
820 
821 		vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
822 		if (object->type == OBJT_VNODE &&
823 		    (vp = (struct vnode *)object->handle) != NULL) {
824 			VI_LOCK(vp);
825 			if (vp->v_iflag & VI_OBJDIRTY)
826 				vp->v_iflag &= ~VI_OBJDIRTY;
827 			VI_UNLOCK(vp);
828 		}
829 	}
830 
831 rescan:
832 	curgeneration = object->generation;
833 
834 	for (p = TAILQ_FIRST(&object->memq); p; p = np) {
835 		int n;
836 
837 		np = TAILQ_NEXT(p, listq);
838 
839 again:
840 		pi = p->pindex;
841 		if ((p->oflags & VPO_CLEANCHK) == 0 ||
842 			(pi < tstart) || (pi >= tend) ||
843 		    p->valid == 0) {
844 			p->oflags &= ~VPO_CLEANCHK;
845 			continue;
846 		}
847 
848 		vm_page_test_dirty(p);
849 		if ((p->dirty & p->valid) == 0) {
850 			p->oflags &= ~VPO_CLEANCHK;
851 			continue;
852 		}
853 
854 		/*
855 		 * If we have been asked to skip nosync pages and this is a
856 		 * nosync page, skip it.  Note that the object flags were
857 		 * not cleared in this case so we do not have to set them.
858 		 */
859 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
860 			p->oflags &= ~VPO_CLEANCHK;
861 			continue;
862 		}
863 
864 		n = vm_object_page_collect_flush(object, p,
865 			curgeneration, pagerflags);
866 		if (n == 0)
867 			goto rescan;
868 
869 		if (object->generation != curgeneration)
870 			goto rescan;
871 
872 		/*
873 		 * Try to optimize the next page.  If we can't we pick up
874 		 * our (random) scan where we left off.
875 		 */
876 		if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
877 			if ((p = vm_page_lookup(object, pi + n)) != NULL)
878 				goto again;
879 		}
880 	}
881 	vm_page_unlock_queues();
882 #if 0
883 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
884 #endif
885 
886 	vm_object_clear_flag(object, OBJ_CLEANING);
887 	return;
888 }
889 
890 static int
891 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
892 {
893 	int runlen;
894 	int maxf;
895 	int chkb;
896 	int maxb;
897 	int i;
898 	vm_pindex_t pi;
899 	vm_page_t maf[vm_pageout_page_count];
900 	vm_page_t mab[vm_pageout_page_count];
901 	vm_page_t ma[vm_pageout_page_count];
902 
903 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
904 	pi = p->pindex;
905 	while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
906 		vm_page_lock_queues();
907 		if (object->generation != curgeneration) {
908 			return(0);
909 		}
910 	}
911 	maxf = 0;
912 	for(i = 1; i < vm_pageout_page_count; i++) {
913 		vm_page_t tp;
914 
915 		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
916 			if ((tp->oflags & VPO_BUSY) ||
917 				((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
918 				 (tp->oflags & VPO_CLEANCHK) == 0) ||
919 				(tp->busy != 0))
920 				break;
921 			vm_page_test_dirty(tp);
922 			if ((tp->dirty & tp->valid) == 0) {
923 				tp->oflags &= ~VPO_CLEANCHK;
924 				break;
925 			}
926 			maf[ i - 1 ] = tp;
927 			maxf++;
928 			continue;
929 		}
930 		break;
931 	}
932 
933 	maxb = 0;
934 	chkb = vm_pageout_page_count -  maxf;
935 	if (chkb) {
936 		for(i = 1; i < chkb;i++) {
937 			vm_page_t tp;
938 
939 			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
940 				if ((tp->oflags & VPO_BUSY) ||
941 					((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
942 					 (tp->oflags & VPO_CLEANCHK) == 0) ||
943 					(tp->busy != 0))
944 					break;
945 				vm_page_test_dirty(tp);
946 				if ((tp->dirty & tp->valid) == 0) {
947 					tp->oflags &= ~VPO_CLEANCHK;
948 					break;
949 				}
950 				mab[ i - 1 ] = tp;
951 				maxb++;
952 				continue;
953 			}
954 			break;
955 		}
956 	}
957 
958 	for(i = 0; i < maxb; i++) {
959 		int index = (maxb - i) - 1;
960 		ma[index] = mab[i];
961 		ma[index]->oflags &= ~VPO_CLEANCHK;
962 	}
963 	p->oflags &= ~VPO_CLEANCHK;
964 	ma[maxb] = p;
965 	for(i = 0; i < maxf; i++) {
966 		int index = (maxb + i) + 1;
967 		ma[index] = maf[i];
968 		ma[index]->oflags &= ~VPO_CLEANCHK;
969 	}
970 	runlen = maxb + maxf + 1;
971 
972 	vm_pageout_flush(ma, runlen, pagerflags);
973 	for (i = 0; i < runlen; i++) {
974 		if (ma[i]->valid & ma[i]->dirty) {
975 			pmap_remove_write(ma[i]);
976 			ma[i]->oflags |= VPO_CLEANCHK;
977 
978 			/*
979 			 * maxf will end up being the actual number of pages
980 			 * we wrote out contiguously, non-inclusive of the
981 			 * first page.  We do not count look-behind pages.
982 			 */
983 			if (i >= maxb + 1 && (maxf > i - maxb - 1))
984 				maxf = i - maxb - 1;
985 		}
986 	}
987 	return(maxf + 1);
988 }
989 
990 /*
991  * Note that there is absolutely no sense in writing out
992  * anonymous objects, so we track down the vnode object
993  * to write out.
994  * We invalidate (remove) all pages from the address space
995  * for semantic correctness.
996  *
997  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
998  * may start out with a NULL object.
999  */
1000 void
1001 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1002     boolean_t syncio, boolean_t invalidate)
1003 {
1004 	vm_object_t backing_object;
1005 	struct vnode *vp;
1006 	struct mount *mp;
1007 	int flags;
1008 
1009 	if (object == NULL)
1010 		return;
1011 	VM_OBJECT_LOCK(object);
1012 	while ((backing_object = object->backing_object) != NULL) {
1013 		VM_OBJECT_LOCK(backing_object);
1014 		offset += object->backing_object_offset;
1015 		VM_OBJECT_UNLOCK(object);
1016 		object = backing_object;
1017 		if (object->size < OFF_TO_IDX(offset + size))
1018 			size = IDX_TO_OFF(object->size) - offset;
1019 	}
1020 	/*
1021 	 * Flush pages if writing is allowed, invalidate them
1022 	 * if invalidation requested.  Pages undergoing I/O
1023 	 * will be ignored by vm_object_page_remove().
1024 	 *
1025 	 * We cannot lock the vnode and then wait for paging
1026 	 * to complete without deadlocking against vm_fault.
1027 	 * Instead we simply call vm_object_page_remove() and
1028 	 * allow it to block internally on a page-by-page
1029 	 * basis when it encounters pages undergoing async
1030 	 * I/O.
1031 	 */
1032 	if (object->type == OBJT_VNODE &&
1033 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1034 		int vfslocked;
1035 		vp = object->handle;
1036 		VM_OBJECT_UNLOCK(object);
1037 		(void) vn_start_write(vp, &mp, V_WAIT);
1038 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1039 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1040 		flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1041 		flags |= invalidate ? OBJPC_INVAL : 0;
1042 		VM_OBJECT_LOCK(object);
1043 		vm_object_page_clean(object,
1044 		    OFF_TO_IDX(offset),
1045 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1046 		    flags);
1047 		VM_OBJECT_UNLOCK(object);
1048 		VOP_UNLOCK(vp, 0);
1049 		VFS_UNLOCK_GIANT(vfslocked);
1050 		vn_finished_write(mp);
1051 		VM_OBJECT_LOCK(object);
1052 	}
1053 	if ((object->type == OBJT_VNODE ||
1054 	     object->type == OBJT_DEVICE) && invalidate) {
1055 		boolean_t purge;
1056 		purge = old_msync || (object->type == OBJT_DEVICE);
1057 		vm_object_page_remove(object,
1058 		    OFF_TO_IDX(offset),
1059 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1060 		    purge ? FALSE : TRUE);
1061 	}
1062 	VM_OBJECT_UNLOCK(object);
1063 }
1064 
1065 /*
1066  *	vm_object_madvise:
1067  *
1068  *	Implements the madvise function at the object/page level.
1069  *
1070  *	MADV_WILLNEED	(any object)
1071  *
1072  *	    Activate the specified pages if they are resident.
1073  *
1074  *	MADV_DONTNEED	(any object)
1075  *
1076  *	    Deactivate the specified pages if they are resident.
1077  *
1078  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1079  *			 OBJ_ONEMAPPING only)
1080  *
1081  *	    Deactivate and clean the specified pages if they are
1082  *	    resident.  This permits the process to reuse the pages
1083  *	    without faulting or the kernel to reclaim the pages
1084  *	    without I/O.
1085  */
1086 void
1087 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1088 {
1089 	vm_pindex_t end, tpindex;
1090 	vm_object_t backing_object, tobject;
1091 	vm_page_t m;
1092 
1093 	if (object == NULL)
1094 		return;
1095 	VM_OBJECT_LOCK(object);
1096 	end = pindex + count;
1097 	/*
1098 	 * Locate and adjust resident pages
1099 	 */
1100 	for (; pindex < end; pindex += 1) {
1101 relookup:
1102 		tobject = object;
1103 		tpindex = pindex;
1104 shadowlookup:
1105 		/*
1106 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1107 		 * and those pages must be OBJ_ONEMAPPING.
1108 		 */
1109 		if (advise == MADV_FREE) {
1110 			if ((tobject->type != OBJT_DEFAULT &&
1111 			     tobject->type != OBJT_SWAP) ||
1112 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
1113 				goto unlock_tobject;
1114 			}
1115 		}
1116 		m = vm_page_lookup(tobject, tpindex);
1117 		if (m == NULL && advise == MADV_WILLNEED) {
1118 			/*
1119 			 * If the page is cached, reactivate it.
1120 			 */
1121 			m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1122 			    VM_ALLOC_NOBUSY);
1123 		}
1124 		if (m == NULL) {
1125 			/*
1126 			 * There may be swap even if there is no backing page
1127 			 */
1128 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1129 				swap_pager_freespace(tobject, tpindex, 1);
1130 			/*
1131 			 * next object
1132 			 */
1133 			backing_object = tobject->backing_object;
1134 			if (backing_object == NULL)
1135 				goto unlock_tobject;
1136 			VM_OBJECT_LOCK(backing_object);
1137 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1138 			if (tobject != object)
1139 				VM_OBJECT_UNLOCK(tobject);
1140 			tobject = backing_object;
1141 			goto shadowlookup;
1142 		}
1143 		/*
1144 		 * If the page is busy or not in a normal active state,
1145 		 * we skip it.  If the page is not managed there are no
1146 		 * page queues to mess with.  Things can break if we mess
1147 		 * with pages in any of the below states.
1148 		 */
1149 		vm_page_lock_queues();
1150 		if (m->hold_count ||
1151 		    m->wire_count ||
1152 		    (m->flags & PG_UNMANAGED) ||
1153 		    m->valid != VM_PAGE_BITS_ALL) {
1154 			vm_page_unlock_queues();
1155 			goto unlock_tobject;
1156 		}
1157 		if ((m->oflags & VPO_BUSY) || m->busy) {
1158 			vm_page_flag_set(m, PG_REFERENCED);
1159 			vm_page_unlock_queues();
1160 			if (object != tobject)
1161 				VM_OBJECT_UNLOCK(object);
1162 			m->oflags |= VPO_WANTED;
1163 			msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo", 0);
1164 			VM_OBJECT_LOCK(object);
1165   			goto relookup;
1166 		}
1167 		if (advise == MADV_WILLNEED) {
1168 			vm_page_activate(m);
1169 		} else if (advise == MADV_DONTNEED) {
1170 			vm_page_dontneed(m);
1171 		} else if (advise == MADV_FREE) {
1172 			/*
1173 			 * Mark the page clean.  This will allow the page
1174 			 * to be freed up by the system.  However, such pages
1175 			 * are often reused quickly by malloc()/free()
1176 			 * so we do not do anything that would cause
1177 			 * a page fault if we can help it.
1178 			 *
1179 			 * Specifically, we do not try to actually free
1180 			 * the page now nor do we try to put it in the
1181 			 * cache (which would cause a page fault on reuse).
1182 			 *
1183 			 * But we do make the page is freeable as we
1184 			 * can without actually taking the step of unmapping
1185 			 * it.
1186 			 */
1187 			pmap_clear_modify(m);
1188 			m->dirty = 0;
1189 			m->act_count = 0;
1190 			vm_page_dontneed(m);
1191 		}
1192 		vm_page_unlock_queues();
1193 		if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1194 			swap_pager_freespace(tobject, tpindex, 1);
1195 unlock_tobject:
1196 		if (tobject != object)
1197 			VM_OBJECT_UNLOCK(tobject);
1198 	}
1199 	VM_OBJECT_UNLOCK(object);
1200 }
1201 
1202 /*
1203  *	vm_object_shadow:
1204  *
1205  *	Create a new object which is backed by the
1206  *	specified existing object range.  The source
1207  *	object reference is deallocated.
1208  *
1209  *	The new object and offset into that object
1210  *	are returned in the source parameters.
1211  */
1212 void
1213 vm_object_shadow(
1214 	vm_object_t *object,	/* IN/OUT */
1215 	vm_ooffset_t *offset,	/* IN/OUT */
1216 	vm_size_t length)
1217 {
1218 	vm_object_t source;
1219 	vm_object_t result;
1220 
1221 	source = *object;
1222 
1223 	/*
1224 	 * Don't create the new object if the old object isn't shared.
1225 	 */
1226 	if (source != NULL) {
1227 		VM_OBJECT_LOCK(source);
1228 		if (source->ref_count == 1 &&
1229 		    source->handle == NULL &&
1230 		    (source->type == OBJT_DEFAULT ||
1231 		     source->type == OBJT_SWAP)) {
1232 			VM_OBJECT_UNLOCK(source);
1233 			return;
1234 		}
1235 		VM_OBJECT_UNLOCK(source);
1236 	}
1237 
1238 	/*
1239 	 * Allocate a new object with the given length.
1240 	 */
1241 	result = vm_object_allocate(OBJT_DEFAULT, length);
1242 
1243 	/*
1244 	 * The new object shadows the source object, adding a reference to it.
1245 	 * Our caller changes his reference to point to the new object,
1246 	 * removing a reference to the source object.  Net result: no change
1247 	 * of reference count.
1248 	 *
1249 	 * Try to optimize the result object's page color when shadowing
1250 	 * in order to maintain page coloring consistency in the combined
1251 	 * shadowed object.
1252 	 */
1253 	result->backing_object = source;
1254 	/*
1255 	 * Store the offset into the source object, and fix up the offset into
1256 	 * the new object.
1257 	 */
1258 	result->backing_object_offset = *offset;
1259 	if (source != NULL) {
1260 		VM_OBJECT_LOCK(source);
1261 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1262 		source->shadow_count++;
1263 		source->generation++;
1264 #if VM_NRESERVLEVEL > 0
1265 		result->flags |= source->flags & (OBJ_NEEDGIANT | OBJ_COLORED);
1266 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1267 		    ((1 << (VM_NFREEORDER - 1)) - 1);
1268 #else
1269 		result->flags |= source->flags & OBJ_NEEDGIANT;
1270 #endif
1271 		VM_OBJECT_UNLOCK(source);
1272 	}
1273 
1274 
1275 	/*
1276 	 * Return the new things
1277 	 */
1278 	*offset = 0;
1279 	*object = result;
1280 }
1281 
1282 /*
1283  *	vm_object_split:
1284  *
1285  * Split the pages in a map entry into a new object.  This affords
1286  * easier removal of unused pages, and keeps object inheritance from
1287  * being a negative impact on memory usage.
1288  */
1289 void
1290 vm_object_split(vm_map_entry_t entry)
1291 {
1292 	vm_page_t m, m_next;
1293 	vm_object_t orig_object, new_object, source;
1294 	vm_pindex_t idx, offidxstart;
1295 	vm_size_t size;
1296 
1297 	orig_object = entry->object.vm_object;
1298 	if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1299 		return;
1300 	if (orig_object->ref_count <= 1)
1301 		return;
1302 	VM_OBJECT_UNLOCK(orig_object);
1303 
1304 	offidxstart = OFF_TO_IDX(entry->offset);
1305 	size = atop(entry->end - entry->start);
1306 
1307 	/*
1308 	 * If swap_pager_copy() is later called, it will convert new_object
1309 	 * into a swap object.
1310 	 */
1311 	new_object = vm_object_allocate(OBJT_DEFAULT, size);
1312 
1313 	/*
1314 	 * At this point, the new object is still private, so the order in
1315 	 * which the original and new objects are locked does not matter.
1316 	 */
1317 	VM_OBJECT_LOCK(new_object);
1318 	VM_OBJECT_LOCK(orig_object);
1319 	source = orig_object->backing_object;
1320 	if (source != NULL) {
1321 		VM_OBJECT_LOCK(source);
1322 		if ((source->flags & OBJ_DEAD) != 0) {
1323 			VM_OBJECT_UNLOCK(source);
1324 			VM_OBJECT_UNLOCK(orig_object);
1325 			VM_OBJECT_UNLOCK(new_object);
1326 			vm_object_deallocate(new_object);
1327 			VM_OBJECT_LOCK(orig_object);
1328 			return;
1329 		}
1330 		LIST_INSERT_HEAD(&source->shadow_head,
1331 				  new_object, shadow_list);
1332 		source->shadow_count++;
1333 		source->generation++;
1334 		vm_object_reference_locked(source);	/* for new_object */
1335 		vm_object_clear_flag(source, OBJ_ONEMAPPING);
1336 		VM_OBJECT_UNLOCK(source);
1337 		new_object->backing_object_offset =
1338 			orig_object->backing_object_offset + entry->offset;
1339 		new_object->backing_object = source;
1340 	}
1341 	new_object->flags |= orig_object->flags & OBJ_NEEDGIANT;
1342 retry:
1343 	if ((m = TAILQ_FIRST(&orig_object->memq)) != NULL) {
1344 		if (m->pindex < offidxstart) {
1345 			m = vm_page_splay(offidxstart, orig_object->root);
1346 			if ((orig_object->root = m)->pindex < offidxstart)
1347 				m = TAILQ_NEXT(m, listq);
1348 		}
1349 	}
1350 	vm_page_lock_queues();
1351 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1352 	    m = m_next) {
1353 		m_next = TAILQ_NEXT(m, listq);
1354 
1355 		/*
1356 		 * We must wait for pending I/O to complete before we can
1357 		 * rename the page.
1358 		 *
1359 		 * We do not have to VM_PROT_NONE the page as mappings should
1360 		 * not be changed by this operation.
1361 		 */
1362 		if ((m->oflags & VPO_BUSY) || m->busy) {
1363 			vm_page_flag_set(m, PG_REFERENCED);
1364 			vm_page_unlock_queues();
1365 			VM_OBJECT_UNLOCK(new_object);
1366 			m->oflags |= VPO_WANTED;
1367 			msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1368 			VM_OBJECT_LOCK(new_object);
1369 			goto retry;
1370 		}
1371 		vm_page_rename(m, new_object, idx);
1372 		/* page automatically made dirty by rename and cache handled */
1373 		vm_page_busy(m);
1374 	}
1375 	vm_page_unlock_queues();
1376 	if (orig_object->type == OBJT_SWAP) {
1377 		/*
1378 		 * swap_pager_copy() can sleep, in which case the orig_object's
1379 		 * and new_object's locks are released and reacquired.
1380 		 */
1381 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1382 
1383 		/*
1384 		 * Transfer any cached pages from orig_object to new_object.
1385 		 */
1386 		if (__predict_false(orig_object->cache != NULL))
1387 			vm_page_cache_transfer(orig_object, offidxstart,
1388 			    new_object);
1389 	}
1390 	VM_OBJECT_UNLOCK(orig_object);
1391 	TAILQ_FOREACH(m, &new_object->memq, listq)
1392 		vm_page_wakeup(m);
1393 	VM_OBJECT_UNLOCK(new_object);
1394 	entry->object.vm_object = new_object;
1395 	entry->offset = 0LL;
1396 	vm_object_deallocate(orig_object);
1397 	VM_OBJECT_LOCK(new_object);
1398 }
1399 
1400 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1401 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1402 #define	OBSC_COLLAPSE_WAIT	0x0004
1403 
1404 static int
1405 vm_object_backing_scan(vm_object_t object, int op)
1406 {
1407 	int r = 1;
1408 	vm_page_t p;
1409 	vm_object_t backing_object;
1410 	vm_pindex_t backing_offset_index;
1411 
1412 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1413 	VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1414 
1415 	backing_object = object->backing_object;
1416 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1417 
1418 	/*
1419 	 * Initial conditions
1420 	 */
1421 	if (op & OBSC_TEST_ALL_SHADOWED) {
1422 		/*
1423 		 * We do not want to have to test for the existence of cache
1424 		 * or swap pages in the backing object.  XXX but with the
1425 		 * new swapper this would be pretty easy to do.
1426 		 *
1427 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1428 		 * been ZFOD faulted yet?  If we do not test for this, the
1429 		 * shadow test may succeed! XXX
1430 		 */
1431 		if (backing_object->type != OBJT_DEFAULT) {
1432 			return (0);
1433 		}
1434 	}
1435 	if (op & OBSC_COLLAPSE_WAIT) {
1436 		vm_object_set_flag(backing_object, OBJ_DEAD);
1437 	}
1438 
1439 	/*
1440 	 * Our scan
1441 	 */
1442 	p = TAILQ_FIRST(&backing_object->memq);
1443 	while (p) {
1444 		vm_page_t next = TAILQ_NEXT(p, listq);
1445 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1446 
1447 		if (op & OBSC_TEST_ALL_SHADOWED) {
1448 			vm_page_t pp;
1449 
1450 			/*
1451 			 * Ignore pages outside the parent object's range
1452 			 * and outside the parent object's mapping of the
1453 			 * backing object.
1454 			 *
1455 			 * note that we do not busy the backing object's
1456 			 * page.
1457 			 */
1458 			if (
1459 			    p->pindex < backing_offset_index ||
1460 			    new_pindex >= object->size
1461 			) {
1462 				p = next;
1463 				continue;
1464 			}
1465 
1466 			/*
1467 			 * See if the parent has the page or if the parent's
1468 			 * object pager has the page.  If the parent has the
1469 			 * page but the page is not valid, the parent's
1470 			 * object pager must have the page.
1471 			 *
1472 			 * If this fails, the parent does not completely shadow
1473 			 * the object and we might as well give up now.
1474 			 */
1475 
1476 			pp = vm_page_lookup(object, new_pindex);
1477 			if (
1478 			    (pp == NULL || pp->valid == 0) &&
1479 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1480 			) {
1481 				r = 0;
1482 				break;
1483 			}
1484 		}
1485 
1486 		/*
1487 		 * Check for busy page
1488 		 */
1489 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1490 			vm_page_t pp;
1491 
1492 			if (op & OBSC_COLLAPSE_NOWAIT) {
1493 				if ((p->oflags & VPO_BUSY) ||
1494 				    !p->valid ||
1495 				    p->busy) {
1496 					p = next;
1497 					continue;
1498 				}
1499 			} else if (op & OBSC_COLLAPSE_WAIT) {
1500 				if ((p->oflags & VPO_BUSY) || p->busy) {
1501 					vm_page_lock_queues();
1502 					vm_page_flag_set(p, PG_REFERENCED);
1503 					vm_page_unlock_queues();
1504 					VM_OBJECT_UNLOCK(object);
1505 					p->oflags |= VPO_WANTED;
1506 					msleep(p, VM_OBJECT_MTX(backing_object),
1507 					    PDROP | PVM, "vmocol", 0);
1508 					VM_OBJECT_LOCK(object);
1509 					VM_OBJECT_LOCK(backing_object);
1510 					/*
1511 					 * If we slept, anything could have
1512 					 * happened.  Since the object is
1513 					 * marked dead, the backing offset
1514 					 * should not have changed so we
1515 					 * just restart our scan.
1516 					 */
1517 					p = TAILQ_FIRST(&backing_object->memq);
1518 					continue;
1519 				}
1520 			}
1521 
1522 			KASSERT(
1523 			    p->object == backing_object,
1524 			    ("vm_object_backing_scan: object mismatch")
1525 			);
1526 
1527 			/*
1528 			 * Destroy any associated swap
1529 			 */
1530 			if (backing_object->type == OBJT_SWAP) {
1531 				swap_pager_freespace(
1532 				    backing_object,
1533 				    p->pindex,
1534 				    1
1535 				);
1536 			}
1537 
1538 			if (
1539 			    p->pindex < backing_offset_index ||
1540 			    new_pindex >= object->size
1541 			) {
1542 				/*
1543 				 * Page is out of the parent object's range, we
1544 				 * can simply destroy it.
1545 				 */
1546 				vm_page_lock_queues();
1547 				KASSERT(!pmap_page_is_mapped(p),
1548 				    ("freeing mapped page %p", p));
1549 				if (p->wire_count == 0)
1550 					vm_page_free(p);
1551 				else
1552 					vm_page_remove(p);
1553 				vm_page_unlock_queues();
1554 				p = next;
1555 				continue;
1556 			}
1557 
1558 			pp = vm_page_lookup(object, new_pindex);
1559 			if (
1560 			    pp != NULL ||
1561 			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1562 			) {
1563 				/*
1564 				 * page already exists in parent OR swap exists
1565 				 * for this location in the parent.  Destroy
1566 				 * the original page from the backing object.
1567 				 *
1568 				 * Leave the parent's page alone
1569 				 */
1570 				vm_page_lock_queues();
1571 				KASSERT(!pmap_page_is_mapped(p),
1572 				    ("freeing mapped page %p", p));
1573 				if (p->wire_count == 0)
1574 					vm_page_free(p);
1575 				else
1576 					vm_page_remove(p);
1577 				vm_page_unlock_queues();
1578 				p = next;
1579 				continue;
1580 			}
1581 
1582 #if VM_NRESERVLEVEL > 0
1583 			/*
1584 			 * Rename the reservation.
1585 			 */
1586 			vm_reserv_rename(p, object, backing_object,
1587 			    backing_offset_index);
1588 #endif
1589 
1590 			/*
1591 			 * Page does not exist in parent, rename the
1592 			 * page from the backing object to the main object.
1593 			 *
1594 			 * If the page was mapped to a process, it can remain
1595 			 * mapped through the rename.
1596 			 */
1597 			vm_page_lock_queues();
1598 			vm_page_rename(p, object, new_pindex);
1599 			vm_page_unlock_queues();
1600 			/* page automatically made dirty by rename */
1601 		}
1602 		p = next;
1603 	}
1604 	return (r);
1605 }
1606 
1607 
1608 /*
1609  * this version of collapse allows the operation to occur earlier and
1610  * when paging_in_progress is true for an object...  This is not a complete
1611  * operation, but should plug 99.9% of the rest of the leaks.
1612  */
1613 static void
1614 vm_object_qcollapse(vm_object_t object)
1615 {
1616 	vm_object_t backing_object = object->backing_object;
1617 
1618 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1619 	VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1620 
1621 	if (backing_object->ref_count != 1)
1622 		return;
1623 
1624 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1625 }
1626 
1627 /*
1628  *	vm_object_collapse:
1629  *
1630  *	Collapse an object with the object backing it.
1631  *	Pages in the backing object are moved into the
1632  *	parent, and the backing object is deallocated.
1633  */
1634 void
1635 vm_object_collapse(vm_object_t object)
1636 {
1637 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1638 
1639 	while (TRUE) {
1640 		vm_object_t backing_object;
1641 
1642 		/*
1643 		 * Verify that the conditions are right for collapse:
1644 		 *
1645 		 * The object exists and the backing object exists.
1646 		 */
1647 		if ((backing_object = object->backing_object) == NULL)
1648 			break;
1649 
1650 		/*
1651 		 * we check the backing object first, because it is most likely
1652 		 * not collapsable.
1653 		 */
1654 		VM_OBJECT_LOCK(backing_object);
1655 		if (backing_object->handle != NULL ||
1656 		    (backing_object->type != OBJT_DEFAULT &&
1657 		     backing_object->type != OBJT_SWAP) ||
1658 		    (backing_object->flags & OBJ_DEAD) ||
1659 		    object->handle != NULL ||
1660 		    (object->type != OBJT_DEFAULT &&
1661 		     object->type != OBJT_SWAP) ||
1662 		    (object->flags & OBJ_DEAD)) {
1663 			VM_OBJECT_UNLOCK(backing_object);
1664 			break;
1665 		}
1666 
1667 		if (
1668 		    object->paging_in_progress != 0 ||
1669 		    backing_object->paging_in_progress != 0
1670 		) {
1671 			vm_object_qcollapse(object);
1672 			VM_OBJECT_UNLOCK(backing_object);
1673 			break;
1674 		}
1675 		/*
1676 		 * We know that we can either collapse the backing object (if
1677 		 * the parent is the only reference to it) or (perhaps) have
1678 		 * the parent bypass the object if the parent happens to shadow
1679 		 * all the resident pages in the entire backing object.
1680 		 *
1681 		 * This is ignoring pager-backed pages such as swap pages.
1682 		 * vm_object_backing_scan fails the shadowing test in this
1683 		 * case.
1684 		 */
1685 		if (backing_object->ref_count == 1) {
1686 			/*
1687 			 * If there is exactly one reference to the backing
1688 			 * object, we can collapse it into the parent.
1689 			 */
1690 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1691 
1692 #if VM_NRESERVLEVEL > 0
1693 			/*
1694 			 * Break any reservations from backing_object.
1695 			 */
1696 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1697 				vm_reserv_break_all(backing_object);
1698 #endif
1699 
1700 			/*
1701 			 * Move the pager from backing_object to object.
1702 			 */
1703 			if (backing_object->type == OBJT_SWAP) {
1704 				/*
1705 				 * swap_pager_copy() can sleep, in which case
1706 				 * the backing_object's and object's locks are
1707 				 * released and reacquired.
1708 				 */
1709 				swap_pager_copy(
1710 				    backing_object,
1711 				    object,
1712 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1713 
1714 				/*
1715 				 * Free any cached pages from backing_object.
1716 				 */
1717 				if (__predict_false(backing_object->cache != NULL))
1718 					vm_page_cache_free(backing_object, 0, 0);
1719 			}
1720 			/*
1721 			 * Object now shadows whatever backing_object did.
1722 			 * Note that the reference to
1723 			 * backing_object->backing_object moves from within
1724 			 * backing_object to within object.
1725 			 */
1726 			LIST_REMOVE(object, shadow_list);
1727 			backing_object->shadow_count--;
1728 			backing_object->generation++;
1729 			if (backing_object->backing_object) {
1730 				VM_OBJECT_LOCK(backing_object->backing_object);
1731 				LIST_REMOVE(backing_object, shadow_list);
1732 				LIST_INSERT_HEAD(
1733 				    &backing_object->backing_object->shadow_head,
1734 				    object, shadow_list);
1735 				/*
1736 				 * The shadow_count has not changed.
1737 				 */
1738 				backing_object->backing_object->generation++;
1739 				VM_OBJECT_UNLOCK(backing_object->backing_object);
1740 			}
1741 			object->backing_object = backing_object->backing_object;
1742 			object->backing_object_offset +=
1743 			    backing_object->backing_object_offset;
1744 
1745 			/*
1746 			 * Discard backing_object.
1747 			 *
1748 			 * Since the backing object has no pages, no pager left,
1749 			 * and no object references within it, all that is
1750 			 * necessary is to dispose of it.
1751 			 */
1752 			KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1753 			VM_OBJECT_UNLOCK(backing_object);
1754 
1755 			mtx_lock(&vm_object_list_mtx);
1756 			TAILQ_REMOVE(
1757 			    &vm_object_list,
1758 			    backing_object,
1759 			    object_list
1760 			);
1761 			mtx_unlock(&vm_object_list_mtx);
1762 
1763 			uma_zfree(obj_zone, backing_object);
1764 
1765 			object_collapses++;
1766 		} else {
1767 			vm_object_t new_backing_object;
1768 
1769 			/*
1770 			 * If we do not entirely shadow the backing object,
1771 			 * there is nothing we can do so we give up.
1772 			 */
1773 			if (object->resident_page_count != object->size &&
1774 			    vm_object_backing_scan(object,
1775 			    OBSC_TEST_ALL_SHADOWED) == 0) {
1776 				VM_OBJECT_UNLOCK(backing_object);
1777 				break;
1778 			}
1779 
1780 			/*
1781 			 * Make the parent shadow the next object in the
1782 			 * chain.  Deallocating backing_object will not remove
1783 			 * it, since its reference count is at least 2.
1784 			 */
1785 			LIST_REMOVE(object, shadow_list);
1786 			backing_object->shadow_count--;
1787 			backing_object->generation++;
1788 
1789 			new_backing_object = backing_object->backing_object;
1790 			if ((object->backing_object = new_backing_object) != NULL) {
1791 				VM_OBJECT_LOCK(new_backing_object);
1792 				LIST_INSERT_HEAD(
1793 				    &new_backing_object->shadow_head,
1794 				    object,
1795 				    shadow_list
1796 				);
1797 				new_backing_object->shadow_count++;
1798 				new_backing_object->generation++;
1799 				vm_object_reference_locked(new_backing_object);
1800 				VM_OBJECT_UNLOCK(new_backing_object);
1801 				object->backing_object_offset +=
1802 					backing_object->backing_object_offset;
1803 			}
1804 
1805 			/*
1806 			 * Drop the reference count on backing_object. Since
1807 			 * its ref_count was at least 2, it will not vanish.
1808 			 */
1809 			backing_object->ref_count--;
1810 			VM_OBJECT_UNLOCK(backing_object);
1811 			object_bypasses++;
1812 		}
1813 
1814 		/*
1815 		 * Try again with this object's new backing object.
1816 		 */
1817 	}
1818 }
1819 
1820 /*
1821  *	vm_object_page_remove:
1822  *
1823  *	For the given object, either frees or invalidates each of the
1824  *	specified pages.  In general, a page is freed.  However, if a
1825  *	page is wired for any reason other than the existence of a
1826  *	managed, wired mapping, then it may be invalidated but not
1827  *	removed from the object.  Pages are specified by the given
1828  *	range ["start", "end") and Boolean "clean_only".  As a
1829  *	special case, if "end" is zero, then the range extends from
1830  *	"start" to the end of the object.  If "clean_only" is TRUE,
1831  *	then only the non-dirty pages within the specified range are
1832  *	affected.
1833  *
1834  *	In general, this operation should only be performed on objects
1835  *	that contain managed pages.  There are two exceptions.  First,
1836  *	it may be performed on the kernel and kmem objects.  Second,
1837  *	it may be used by msync(..., MS_INVALIDATE) to invalidate
1838  *	device-backed pages.
1839  *
1840  *	The object must be locked.
1841  */
1842 void
1843 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1844     boolean_t clean_only)
1845 {
1846 	vm_page_t p, next;
1847 	int wirings;
1848 
1849 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1850 	if (object->resident_page_count == 0)
1851 		goto skipmemq;
1852 
1853 	/*
1854 	 * Since physically-backed objects do not use managed pages, we can't
1855 	 * remove pages from the object (we must instead remove the page
1856 	 * references, and then destroy the object).
1857 	 */
1858 	KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1859 	    object == kmem_object,
1860 	    ("attempt to remove pages from a physical object"));
1861 
1862 	vm_object_pip_add(object, 1);
1863 again:
1864 	vm_page_lock_queues();
1865 	if ((p = TAILQ_FIRST(&object->memq)) != NULL) {
1866 		if (p->pindex < start) {
1867 			p = vm_page_splay(start, object->root);
1868 			if ((object->root = p)->pindex < start)
1869 				p = TAILQ_NEXT(p, listq);
1870 		}
1871 	}
1872 	/*
1873 	 * Assert: the variable p is either (1) the page with the
1874 	 * least pindex greater than or equal to the parameter pindex
1875 	 * or (2) NULL.
1876 	 */
1877 	for (;
1878 	     p != NULL && (p->pindex < end || end == 0);
1879 	     p = next) {
1880 		next = TAILQ_NEXT(p, listq);
1881 
1882 		/*
1883 		 * If the page is wired for any reason besides the
1884 		 * existence of managed, wired mappings, then it cannot
1885 		 * be freed.  For example, fictitious pages, which
1886 		 * represent device memory, are inherently wired and
1887 		 * cannot be freed.  They can, however, be invalidated
1888 		 * if "clean_only" is FALSE.
1889 		 */
1890 		if ((wirings = p->wire_count) != 0 &&
1891 		    (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1892 			/* Fictitious pages do not have managed mappings. */
1893 			if ((p->flags & PG_FICTITIOUS) == 0)
1894 				pmap_remove_all(p);
1895 			/* Account for removal of managed, wired mappings. */
1896 			p->wire_count -= wirings;
1897 			if (!clean_only)
1898 				p->valid = 0;
1899 			continue;
1900 		}
1901 		if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1902 			goto again;
1903 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
1904 		    ("vm_object_page_remove: page %p is fictitious", p));
1905 		if (clean_only && p->valid) {
1906 			pmap_remove_write(p);
1907 			if (p->valid & p->dirty)
1908 				continue;
1909 		}
1910 		pmap_remove_all(p);
1911 		/* Account for removal of managed, wired mappings. */
1912 		if (wirings != 0)
1913 			p->wire_count -= wirings;
1914 		vm_page_free(p);
1915 	}
1916 	vm_page_unlock_queues();
1917 	vm_object_pip_wakeup(object);
1918 skipmemq:
1919 	if (__predict_false(object->cache != NULL))
1920 		vm_page_cache_free(object, start, end);
1921 }
1922 
1923 /*
1924  *	Routine:	vm_object_coalesce
1925  *	Function:	Coalesces two objects backing up adjoining
1926  *			regions of memory into a single object.
1927  *
1928  *	returns TRUE if objects were combined.
1929  *
1930  *	NOTE:	Only works at the moment if the second object is NULL -
1931  *		if it's not, which object do we lock first?
1932  *
1933  *	Parameters:
1934  *		prev_object	First object to coalesce
1935  *		prev_offset	Offset into prev_object
1936  *		prev_size	Size of reference to prev_object
1937  *		next_size	Size of reference to the second object
1938  *
1939  *	Conditions:
1940  *	The object must *not* be locked.
1941  */
1942 boolean_t
1943 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
1944 	vm_size_t prev_size, vm_size_t next_size)
1945 {
1946 	vm_pindex_t next_pindex;
1947 
1948 	if (prev_object == NULL)
1949 		return (TRUE);
1950 	VM_OBJECT_LOCK(prev_object);
1951 	if (prev_object->type != OBJT_DEFAULT &&
1952 	    prev_object->type != OBJT_SWAP) {
1953 		VM_OBJECT_UNLOCK(prev_object);
1954 		return (FALSE);
1955 	}
1956 
1957 	/*
1958 	 * Try to collapse the object first
1959 	 */
1960 	vm_object_collapse(prev_object);
1961 
1962 	/*
1963 	 * Can't coalesce if: . more than one reference . paged out . shadows
1964 	 * another object . has a copy elsewhere (any of which mean that the
1965 	 * pages not mapped to prev_entry may be in use anyway)
1966 	 */
1967 	if (prev_object->backing_object != NULL) {
1968 		VM_OBJECT_UNLOCK(prev_object);
1969 		return (FALSE);
1970 	}
1971 
1972 	prev_size >>= PAGE_SHIFT;
1973 	next_size >>= PAGE_SHIFT;
1974 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
1975 
1976 	if ((prev_object->ref_count > 1) &&
1977 	    (prev_object->size != next_pindex)) {
1978 		VM_OBJECT_UNLOCK(prev_object);
1979 		return (FALSE);
1980 	}
1981 
1982 	/*
1983 	 * Remove any pages that may still be in the object from a previous
1984 	 * deallocation.
1985 	 */
1986 	if (next_pindex < prev_object->size) {
1987 		vm_object_page_remove(prev_object,
1988 				      next_pindex,
1989 				      next_pindex + next_size, FALSE);
1990 		if (prev_object->type == OBJT_SWAP)
1991 			swap_pager_freespace(prev_object,
1992 					     next_pindex, next_size);
1993 	}
1994 
1995 	/*
1996 	 * Extend the object if necessary.
1997 	 */
1998 	if (next_pindex + next_size > prev_object->size)
1999 		prev_object->size = next_pindex + next_size;
2000 
2001 	VM_OBJECT_UNLOCK(prev_object);
2002 	return (TRUE);
2003 }
2004 
2005 void
2006 vm_object_set_writeable_dirty(vm_object_t object)
2007 {
2008 	struct vnode *vp;
2009 
2010 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2011 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2012 		return;
2013 	vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2014 	if (object->type == OBJT_VNODE &&
2015 	    (vp = (struct vnode *)object->handle) != NULL) {
2016 		VI_LOCK(vp);
2017 		vp->v_iflag |= VI_OBJDIRTY;
2018 		VI_UNLOCK(vp);
2019 	}
2020 }
2021 
2022 #include "opt_ddb.h"
2023 #ifdef DDB
2024 #include <sys/kernel.h>
2025 
2026 #include <sys/cons.h>
2027 
2028 #include <ddb/ddb.h>
2029 
2030 static int
2031 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2032 {
2033 	vm_map_t tmpm;
2034 	vm_map_entry_t tmpe;
2035 	vm_object_t obj;
2036 	int entcount;
2037 
2038 	if (map == 0)
2039 		return 0;
2040 
2041 	if (entry == 0) {
2042 		tmpe = map->header.next;
2043 		entcount = map->nentries;
2044 		while (entcount-- && (tmpe != &map->header)) {
2045 			if (_vm_object_in_map(map, object, tmpe)) {
2046 				return 1;
2047 			}
2048 			tmpe = tmpe->next;
2049 		}
2050 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2051 		tmpm = entry->object.sub_map;
2052 		tmpe = tmpm->header.next;
2053 		entcount = tmpm->nentries;
2054 		while (entcount-- && tmpe != &tmpm->header) {
2055 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2056 				return 1;
2057 			}
2058 			tmpe = tmpe->next;
2059 		}
2060 	} else if ((obj = entry->object.vm_object) != NULL) {
2061 		for (; obj; obj = obj->backing_object)
2062 			if (obj == object) {
2063 				return 1;
2064 			}
2065 	}
2066 	return 0;
2067 }
2068 
2069 static int
2070 vm_object_in_map(vm_object_t object)
2071 {
2072 	struct proc *p;
2073 
2074 	/* sx_slock(&allproc_lock); */
2075 	FOREACH_PROC_IN_SYSTEM(p) {
2076 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2077 			continue;
2078 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2079 			/* sx_sunlock(&allproc_lock); */
2080 			return 1;
2081 		}
2082 	}
2083 	/* sx_sunlock(&allproc_lock); */
2084 	if (_vm_object_in_map(kernel_map, object, 0))
2085 		return 1;
2086 	if (_vm_object_in_map(kmem_map, object, 0))
2087 		return 1;
2088 	if (_vm_object_in_map(pager_map, object, 0))
2089 		return 1;
2090 	if (_vm_object_in_map(buffer_map, object, 0))
2091 		return 1;
2092 	return 0;
2093 }
2094 
2095 DB_SHOW_COMMAND(vmochk, vm_object_check)
2096 {
2097 	vm_object_t object;
2098 
2099 	/*
2100 	 * make sure that internal objs are in a map somewhere
2101 	 * and none have zero ref counts.
2102 	 */
2103 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2104 		if (object->handle == NULL &&
2105 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2106 			if (object->ref_count == 0) {
2107 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2108 					(long)object->size);
2109 			}
2110 			if (!vm_object_in_map(object)) {
2111 				db_printf(
2112 			"vmochk: internal obj is not in a map: "
2113 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2114 				    object->ref_count, (u_long)object->size,
2115 				    (u_long)object->size,
2116 				    (void *)object->backing_object);
2117 			}
2118 		}
2119 	}
2120 }
2121 
2122 /*
2123  *	vm_object_print:	[ debug ]
2124  */
2125 DB_SHOW_COMMAND(object, vm_object_print_static)
2126 {
2127 	/* XXX convert args. */
2128 	vm_object_t object = (vm_object_t)addr;
2129 	boolean_t full = have_addr;
2130 
2131 	vm_page_t p;
2132 
2133 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2134 #define	count	was_count
2135 
2136 	int count;
2137 
2138 	if (object == NULL)
2139 		return;
2140 
2141 	db_iprintf(
2142 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x\n",
2143 	    object, (int)object->type, (uintmax_t)object->size,
2144 	    object->resident_page_count, object->ref_count, object->flags);
2145 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2146 	    object->shadow_count,
2147 	    object->backing_object ? object->backing_object->ref_count : 0,
2148 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2149 
2150 	if (!full)
2151 		return;
2152 
2153 	db_indent += 2;
2154 	count = 0;
2155 	TAILQ_FOREACH(p, &object->memq, listq) {
2156 		if (count == 0)
2157 			db_iprintf("memory:=");
2158 		else if (count == 6) {
2159 			db_printf("\n");
2160 			db_iprintf(" ...");
2161 			count = 0;
2162 		} else
2163 			db_printf(",");
2164 		count++;
2165 
2166 		db_printf("(off=0x%jx,page=0x%jx)",
2167 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2168 	}
2169 	if (count != 0)
2170 		db_printf("\n");
2171 	db_indent -= 2;
2172 }
2173 
2174 /* XXX. */
2175 #undef count
2176 
2177 /* XXX need this non-static entry for calling from vm_map_print. */
2178 void
2179 vm_object_print(
2180         /* db_expr_t */ long addr,
2181 	boolean_t have_addr,
2182 	/* db_expr_t */ long count,
2183 	char *modif)
2184 {
2185 	vm_object_print_static(addr, have_addr, count, modif);
2186 }
2187 
2188 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2189 {
2190 	vm_object_t object;
2191 	int nl = 0;
2192 	int c;
2193 
2194 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2195 		vm_pindex_t idx, fidx;
2196 		vm_pindex_t osize;
2197 		vm_paddr_t pa = -1;
2198 		int rcount;
2199 		vm_page_t m;
2200 
2201 		db_printf("new object: %p\n", (void *)object);
2202 		if (nl > 18) {
2203 			c = cngetc();
2204 			if (c != ' ')
2205 				return;
2206 			nl = 0;
2207 		}
2208 		nl++;
2209 		rcount = 0;
2210 		fidx = 0;
2211 		osize = object->size;
2212 		if (osize > 128)
2213 			osize = 128;
2214 		for (idx = 0; idx < osize; idx++) {
2215 			m = vm_page_lookup(object, idx);
2216 			if (m == NULL) {
2217 				if (rcount) {
2218 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2219 						(long)fidx, rcount, (long)pa);
2220 					if (nl > 18) {
2221 						c = cngetc();
2222 						if (c != ' ')
2223 							return;
2224 						nl = 0;
2225 					}
2226 					nl++;
2227 					rcount = 0;
2228 				}
2229 				continue;
2230 			}
2231 
2232 
2233 			if (rcount &&
2234 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2235 				++rcount;
2236 				continue;
2237 			}
2238 			if (rcount) {
2239 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2240 					(long)fidx, rcount, (long)pa);
2241 				if (nl > 18) {
2242 					c = cngetc();
2243 					if (c != ' ')
2244 						return;
2245 					nl = 0;
2246 				}
2247 				nl++;
2248 			}
2249 			fidx = idx;
2250 			pa = VM_PAGE_TO_PHYS(m);
2251 			rcount = 1;
2252 		}
2253 		if (rcount) {
2254 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2255 				(long)fidx, rcount, (long)pa);
2256 			if (nl > 18) {
2257 				c = cngetc();
2258 				if (c != ' ')
2259 					return;
2260 				nl = 0;
2261 			}
2262 			nl++;
2263 		}
2264 	}
2265 }
2266 #endif /* DDB */
2267