xref: /freebsd/sys/vm/vm_object.c (revision 721351876cd4d3a8a700f62d2061331fa951a488)
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_destroy removes the object from the global object list
598  *      and frees the space for the object.
599  */
600 void
601 vm_object_destroy(vm_object_t object)
602 {
603 
604 	/*
605 	 * Remove the object from the global object list.
606 	 */
607 	mtx_lock(&vm_object_list_mtx);
608 	TAILQ_REMOVE(&vm_object_list, object, object_list);
609 	mtx_unlock(&vm_object_list_mtx);
610 
611 	/*
612 	 * Free the space for the object.
613 	 */
614 	uma_zfree(obj_zone, object);
615 
616 }
617 
618 
619 
620 /*
621  *	vm_object_terminate actually destroys the specified object, freeing
622  *	up all previously used resources.
623  *
624  *	The object must be locked.
625  *	This routine may block.
626  */
627 void
628 vm_object_terminate(vm_object_t object)
629 {
630 	vm_page_t p;
631 
632 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
633 
634 	/*
635 	 * Make sure no one uses us.
636 	 */
637 	vm_object_set_flag(object, OBJ_DEAD);
638 
639 	/*
640 	 * wait for the pageout daemon to be done with the object
641 	 */
642 	vm_object_pip_wait(object, "objtrm");
643 
644 	KASSERT(!object->paging_in_progress,
645 		("vm_object_terminate: pageout in progress"));
646 
647 	/*
648 	 * Clean and free the pages, as appropriate. All references to the
649 	 * object are gone, so we don't need to lock it.
650 	 */
651 	if (object->type == OBJT_VNODE) {
652 		struct vnode *vp = (struct vnode *)object->handle;
653 
654 		/*
655 		 * Clean pages and flush buffers.
656 		 */
657 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
658 		VM_OBJECT_UNLOCK(object);
659 
660 		vinvalbuf(vp, V_SAVE, NULL, 0, 0);
661 
662 		VM_OBJECT_LOCK(object);
663 	}
664 
665 	KASSERT(object->ref_count == 0,
666 		("vm_object_terminate: object with references, ref_count=%d",
667 		object->ref_count));
668 
669 	/*
670 	 * Now free any remaining pages. For internal objects, this also
671 	 * removes them from paging queues. Don't free wired pages, just
672 	 * remove them from the object.
673 	 */
674 	vm_page_lock_queues();
675 	while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
676 		KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
677 			("vm_object_terminate: freeing busy page %p "
678 			"p->busy = %d, p->flags %x\n", p, p->busy, p->flags));
679 		if (p->wire_count == 0) {
680 			vm_page_free(p);
681 			cnt.v_pfree++;
682 		} else {
683 			vm_page_remove(p);
684 		}
685 	}
686 	vm_page_unlock_queues();
687 
688 #if VM_NRESERVLEVEL > 0
689 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
690 		vm_reserv_break_all(object);
691 #endif
692 	if (__predict_false(object->cache != NULL))
693 		vm_page_cache_free(object, 0, 0);
694 
695 	/*
696 	 * Let the pager know object is dead.
697 	 */
698 	vm_pager_deallocate(object);
699 	VM_OBJECT_UNLOCK(object);
700 
701 	vm_object_destroy(object);
702 }
703 
704 /*
705  *	vm_object_page_clean
706  *
707  *	Clean all dirty pages in the specified range of object.  Leaves page
708  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
709  *	write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
710  *	leaving the object dirty.
711  *
712  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
713  *	synchronous clustering mode implementation.
714  *
715  *	Odd semantics: if start == end, we clean everything.
716  *
717  *	The object must be locked.
718  */
719 void
720 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags)
721 {
722 	vm_page_t p, np;
723 	vm_pindex_t tstart, tend;
724 	vm_pindex_t pi;
725 	int clearobjflags;
726 	int pagerflags;
727 	int curgeneration;
728 
729 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
730 	if (object->type != OBJT_VNODE ||
731 		(object->flags & OBJ_MIGHTBEDIRTY) == 0)
732 		return;
733 
734 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
735 	pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
736 
737 	vm_object_set_flag(object, OBJ_CLEANING);
738 
739 	tstart = start;
740 	if (end == 0) {
741 		tend = object->size;
742 	} else {
743 		tend = end;
744 	}
745 
746 	vm_page_lock_queues();
747 	/*
748 	 * If the caller is smart and only msync()s a range he knows is
749 	 * dirty, we may be able to avoid an object scan.  This results in
750 	 * a phenominal improvement in performance.  We cannot do this
751 	 * as a matter of course because the object may be huge - e.g.
752 	 * the size might be in the gigabytes or terrabytes.
753 	 */
754 	if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
755 		vm_pindex_t tscan;
756 		int scanlimit;
757 		int scanreset;
758 
759 		scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
760 		if (scanreset < 16)
761 			scanreset = 16;
762 		pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
763 
764 		scanlimit = scanreset;
765 		tscan = tstart;
766 		while (tscan < tend) {
767 			curgeneration = object->generation;
768 			p = vm_page_lookup(object, tscan);
769 			if (p == NULL || p->valid == 0) {
770 				if (--scanlimit == 0)
771 					break;
772 				++tscan;
773 				continue;
774 			}
775 			vm_page_test_dirty(p);
776 			if ((p->dirty & p->valid) == 0) {
777 				if (--scanlimit == 0)
778 					break;
779 				++tscan;
780 				continue;
781 			}
782 			/*
783 			 * If we have been asked to skip nosync pages and
784 			 * this is a nosync page, we can't continue.
785 			 */
786 			if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
787 				if (--scanlimit == 0)
788 					break;
789 				++tscan;
790 				continue;
791 			}
792 			scanlimit = scanreset;
793 
794 			/*
795 			 * This returns 0 if it was unable to busy the first
796 			 * page (i.e. had to sleep).
797 			 */
798 			tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
799 		}
800 
801 		/*
802 		 * If everything was dirty and we flushed it successfully,
803 		 * and the requested range is not the entire object, we
804 		 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
805 		 * return immediately.
806 		 */
807 		if (tscan >= tend && (tstart || tend < object->size)) {
808 			vm_page_unlock_queues();
809 			vm_object_clear_flag(object, OBJ_CLEANING);
810 			return;
811 		}
812 		pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
813 	}
814 
815 	/*
816 	 * Generally set CLEANCHK interlock and make the page read-only so
817 	 * we can then clear the object flags.
818 	 *
819 	 * However, if this is a nosync mmap then the object is likely to
820 	 * stay dirty so do not mess with the page and do not clear the
821 	 * object flags.
822 	 */
823 	clearobjflags = 1;
824 	TAILQ_FOREACH(p, &object->memq, listq) {
825 		p->oflags |= VPO_CLEANCHK;
826 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC))
827 			clearobjflags = 0;
828 		else
829 			pmap_remove_write(p);
830 	}
831 
832 	if (clearobjflags && (tstart == 0) && (tend == object->size)) {
833 		struct vnode *vp;
834 
835 		vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
836 		if (object->type == OBJT_VNODE &&
837 		    (vp = (struct vnode *)object->handle) != NULL) {
838 			VI_LOCK(vp);
839 			if (vp->v_iflag & VI_OBJDIRTY)
840 				vp->v_iflag &= ~VI_OBJDIRTY;
841 			VI_UNLOCK(vp);
842 		}
843 	}
844 
845 rescan:
846 	curgeneration = object->generation;
847 
848 	for (p = TAILQ_FIRST(&object->memq); p; p = np) {
849 		int n;
850 
851 		np = TAILQ_NEXT(p, listq);
852 
853 again:
854 		pi = p->pindex;
855 		if ((p->oflags & VPO_CLEANCHK) == 0 ||
856 			(pi < tstart) || (pi >= tend) ||
857 		    p->valid == 0) {
858 			p->oflags &= ~VPO_CLEANCHK;
859 			continue;
860 		}
861 
862 		vm_page_test_dirty(p);
863 		if ((p->dirty & p->valid) == 0) {
864 			p->oflags &= ~VPO_CLEANCHK;
865 			continue;
866 		}
867 
868 		/*
869 		 * If we have been asked to skip nosync pages and this is a
870 		 * nosync page, skip it.  Note that the object flags were
871 		 * not cleared in this case so we do not have to set them.
872 		 */
873 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
874 			p->oflags &= ~VPO_CLEANCHK;
875 			continue;
876 		}
877 
878 		n = vm_object_page_collect_flush(object, p,
879 			curgeneration, pagerflags);
880 		if (n == 0)
881 			goto rescan;
882 
883 		if (object->generation != curgeneration)
884 			goto rescan;
885 
886 		/*
887 		 * Try to optimize the next page.  If we can't we pick up
888 		 * our (random) scan where we left off.
889 		 */
890 		if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
891 			if ((p = vm_page_lookup(object, pi + n)) != NULL)
892 				goto again;
893 		}
894 	}
895 	vm_page_unlock_queues();
896 #if 0
897 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
898 #endif
899 
900 	vm_object_clear_flag(object, OBJ_CLEANING);
901 	return;
902 }
903 
904 static int
905 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
906 {
907 	int runlen;
908 	int maxf;
909 	int chkb;
910 	int maxb;
911 	int i;
912 	vm_pindex_t pi;
913 	vm_page_t maf[vm_pageout_page_count];
914 	vm_page_t mab[vm_pageout_page_count];
915 	vm_page_t ma[vm_pageout_page_count];
916 
917 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
918 	pi = p->pindex;
919 	while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
920 		vm_page_lock_queues();
921 		if (object->generation != curgeneration) {
922 			return(0);
923 		}
924 	}
925 	maxf = 0;
926 	for(i = 1; i < vm_pageout_page_count; i++) {
927 		vm_page_t tp;
928 
929 		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
930 			if ((tp->oflags & VPO_BUSY) ||
931 				((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
932 				 (tp->oflags & VPO_CLEANCHK) == 0) ||
933 				(tp->busy != 0))
934 				break;
935 			vm_page_test_dirty(tp);
936 			if ((tp->dirty & tp->valid) == 0) {
937 				tp->oflags &= ~VPO_CLEANCHK;
938 				break;
939 			}
940 			maf[ i - 1 ] = tp;
941 			maxf++;
942 			continue;
943 		}
944 		break;
945 	}
946 
947 	maxb = 0;
948 	chkb = vm_pageout_page_count -  maxf;
949 	if (chkb) {
950 		for(i = 1; i < chkb;i++) {
951 			vm_page_t tp;
952 
953 			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
954 				if ((tp->oflags & VPO_BUSY) ||
955 					((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
956 					 (tp->oflags & VPO_CLEANCHK) == 0) ||
957 					(tp->busy != 0))
958 					break;
959 				vm_page_test_dirty(tp);
960 				if ((tp->dirty & tp->valid) == 0) {
961 					tp->oflags &= ~VPO_CLEANCHK;
962 					break;
963 				}
964 				mab[ i - 1 ] = tp;
965 				maxb++;
966 				continue;
967 			}
968 			break;
969 		}
970 	}
971 
972 	for(i = 0; i < maxb; i++) {
973 		int index = (maxb - i) - 1;
974 		ma[index] = mab[i];
975 		ma[index]->oflags &= ~VPO_CLEANCHK;
976 	}
977 	p->oflags &= ~VPO_CLEANCHK;
978 	ma[maxb] = p;
979 	for(i = 0; i < maxf; i++) {
980 		int index = (maxb + i) + 1;
981 		ma[index] = maf[i];
982 		ma[index]->oflags &= ~VPO_CLEANCHK;
983 	}
984 	runlen = maxb + maxf + 1;
985 
986 	vm_pageout_flush(ma, runlen, pagerflags);
987 	for (i = 0; i < runlen; i++) {
988 		if (ma[i]->valid & ma[i]->dirty) {
989 			pmap_remove_write(ma[i]);
990 			ma[i]->oflags |= VPO_CLEANCHK;
991 
992 			/*
993 			 * maxf will end up being the actual number of pages
994 			 * we wrote out contiguously, non-inclusive of the
995 			 * first page.  We do not count look-behind pages.
996 			 */
997 			if (i >= maxb + 1 && (maxf > i - maxb - 1))
998 				maxf = i - maxb - 1;
999 		}
1000 	}
1001 	return(maxf + 1);
1002 }
1003 
1004 /*
1005  * Note that there is absolutely no sense in writing out
1006  * anonymous objects, so we track down the vnode object
1007  * to write out.
1008  * We invalidate (remove) all pages from the address space
1009  * for semantic correctness.
1010  *
1011  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1012  * may start out with a NULL object.
1013  */
1014 void
1015 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1016     boolean_t syncio, boolean_t invalidate)
1017 {
1018 	vm_object_t backing_object;
1019 	struct vnode *vp;
1020 	struct mount *mp;
1021 	int flags;
1022 
1023 	if (object == NULL)
1024 		return;
1025 	VM_OBJECT_LOCK(object);
1026 	while ((backing_object = object->backing_object) != NULL) {
1027 		VM_OBJECT_LOCK(backing_object);
1028 		offset += object->backing_object_offset;
1029 		VM_OBJECT_UNLOCK(object);
1030 		object = backing_object;
1031 		if (object->size < OFF_TO_IDX(offset + size))
1032 			size = IDX_TO_OFF(object->size) - offset;
1033 	}
1034 	/*
1035 	 * Flush pages if writing is allowed, invalidate them
1036 	 * if invalidation requested.  Pages undergoing I/O
1037 	 * will be ignored by vm_object_page_remove().
1038 	 *
1039 	 * We cannot lock the vnode and then wait for paging
1040 	 * to complete without deadlocking against vm_fault.
1041 	 * Instead we simply call vm_object_page_remove() and
1042 	 * allow it to block internally on a page-by-page
1043 	 * basis when it encounters pages undergoing async
1044 	 * I/O.
1045 	 */
1046 	if (object->type == OBJT_VNODE &&
1047 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1048 		int vfslocked;
1049 		vp = object->handle;
1050 		VM_OBJECT_UNLOCK(object);
1051 		(void) vn_start_write(vp, &mp, V_WAIT);
1052 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1053 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1054 		flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1055 		flags |= invalidate ? OBJPC_INVAL : 0;
1056 		VM_OBJECT_LOCK(object);
1057 		vm_object_page_clean(object,
1058 		    OFF_TO_IDX(offset),
1059 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1060 		    flags);
1061 		VM_OBJECT_UNLOCK(object);
1062 		VOP_UNLOCK(vp, 0);
1063 		VFS_UNLOCK_GIANT(vfslocked);
1064 		vn_finished_write(mp);
1065 		VM_OBJECT_LOCK(object);
1066 	}
1067 	if ((object->type == OBJT_VNODE ||
1068 	     object->type == OBJT_DEVICE) && invalidate) {
1069 		boolean_t purge;
1070 		purge = old_msync || (object->type == OBJT_DEVICE);
1071 		vm_object_page_remove(object,
1072 		    OFF_TO_IDX(offset),
1073 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1074 		    purge ? FALSE : TRUE);
1075 	}
1076 	VM_OBJECT_UNLOCK(object);
1077 }
1078 
1079 /*
1080  *	vm_object_madvise:
1081  *
1082  *	Implements the madvise function at the object/page level.
1083  *
1084  *	MADV_WILLNEED	(any object)
1085  *
1086  *	    Activate the specified pages if they are resident.
1087  *
1088  *	MADV_DONTNEED	(any object)
1089  *
1090  *	    Deactivate the specified pages if they are resident.
1091  *
1092  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1093  *			 OBJ_ONEMAPPING only)
1094  *
1095  *	    Deactivate and clean the specified pages if they are
1096  *	    resident.  This permits the process to reuse the pages
1097  *	    without faulting or the kernel to reclaim the pages
1098  *	    without I/O.
1099  */
1100 void
1101 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1102 {
1103 	vm_pindex_t end, tpindex;
1104 	vm_object_t backing_object, tobject;
1105 	vm_page_t m;
1106 
1107 	if (object == NULL)
1108 		return;
1109 	VM_OBJECT_LOCK(object);
1110 	end = pindex + count;
1111 	/*
1112 	 * Locate and adjust resident pages
1113 	 */
1114 	for (; pindex < end; pindex += 1) {
1115 relookup:
1116 		tobject = object;
1117 		tpindex = pindex;
1118 shadowlookup:
1119 		/*
1120 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1121 		 * and those pages must be OBJ_ONEMAPPING.
1122 		 */
1123 		if (advise == MADV_FREE) {
1124 			if ((tobject->type != OBJT_DEFAULT &&
1125 			     tobject->type != OBJT_SWAP) ||
1126 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
1127 				goto unlock_tobject;
1128 			}
1129 		}
1130 		m = vm_page_lookup(tobject, tpindex);
1131 		if (m == NULL && advise == MADV_WILLNEED) {
1132 			/*
1133 			 * If the page is cached, reactivate it.
1134 			 */
1135 			m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1136 			    VM_ALLOC_NOBUSY);
1137 		}
1138 		if (m == NULL) {
1139 			/*
1140 			 * There may be swap even if there is no backing page
1141 			 */
1142 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1143 				swap_pager_freespace(tobject, tpindex, 1);
1144 			/*
1145 			 * next object
1146 			 */
1147 			backing_object = tobject->backing_object;
1148 			if (backing_object == NULL)
1149 				goto unlock_tobject;
1150 			VM_OBJECT_LOCK(backing_object);
1151 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1152 			if (tobject != object)
1153 				VM_OBJECT_UNLOCK(tobject);
1154 			tobject = backing_object;
1155 			goto shadowlookup;
1156 		}
1157 		/*
1158 		 * If the page is busy or not in a normal active state,
1159 		 * we skip it.  If the page is not managed there are no
1160 		 * page queues to mess with.  Things can break if we mess
1161 		 * with pages in any of the below states.
1162 		 */
1163 		vm_page_lock_queues();
1164 		if (m->hold_count ||
1165 		    m->wire_count ||
1166 		    (m->flags & PG_UNMANAGED) ||
1167 		    m->valid != VM_PAGE_BITS_ALL) {
1168 			vm_page_unlock_queues();
1169 			goto unlock_tobject;
1170 		}
1171 		if ((m->oflags & VPO_BUSY) || m->busy) {
1172 			vm_page_flag_set(m, PG_REFERENCED);
1173 			vm_page_unlock_queues();
1174 			if (object != tobject)
1175 				VM_OBJECT_UNLOCK(object);
1176 			m->oflags |= VPO_WANTED;
1177 			msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo", 0);
1178 			VM_OBJECT_LOCK(object);
1179   			goto relookup;
1180 		}
1181 		if (advise == MADV_WILLNEED) {
1182 			vm_page_activate(m);
1183 		} else if (advise == MADV_DONTNEED) {
1184 			vm_page_dontneed(m);
1185 		} else if (advise == MADV_FREE) {
1186 			/*
1187 			 * Mark the page clean.  This will allow the page
1188 			 * to be freed up by the system.  However, such pages
1189 			 * are often reused quickly by malloc()/free()
1190 			 * so we do not do anything that would cause
1191 			 * a page fault if we can help it.
1192 			 *
1193 			 * Specifically, we do not try to actually free
1194 			 * the page now nor do we try to put it in the
1195 			 * cache (which would cause a page fault on reuse).
1196 			 *
1197 			 * But we do make the page is freeable as we
1198 			 * can without actually taking the step of unmapping
1199 			 * it.
1200 			 */
1201 			pmap_clear_modify(m);
1202 			m->dirty = 0;
1203 			m->act_count = 0;
1204 			vm_page_dontneed(m);
1205 		}
1206 		vm_page_unlock_queues();
1207 		if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1208 			swap_pager_freespace(tobject, tpindex, 1);
1209 unlock_tobject:
1210 		if (tobject != object)
1211 			VM_OBJECT_UNLOCK(tobject);
1212 	}
1213 	VM_OBJECT_UNLOCK(object);
1214 }
1215 
1216 /*
1217  *	vm_object_shadow:
1218  *
1219  *	Create a new object which is backed by the
1220  *	specified existing object range.  The source
1221  *	object reference is deallocated.
1222  *
1223  *	The new object and offset into that object
1224  *	are returned in the source parameters.
1225  */
1226 void
1227 vm_object_shadow(
1228 	vm_object_t *object,	/* IN/OUT */
1229 	vm_ooffset_t *offset,	/* IN/OUT */
1230 	vm_size_t length)
1231 {
1232 	vm_object_t source;
1233 	vm_object_t result;
1234 
1235 	source = *object;
1236 
1237 	/*
1238 	 * Don't create the new object if the old object isn't shared.
1239 	 */
1240 	if (source != NULL) {
1241 		VM_OBJECT_LOCK(source);
1242 		if (source->ref_count == 1 &&
1243 		    source->handle == NULL &&
1244 		    (source->type == OBJT_DEFAULT ||
1245 		     source->type == OBJT_SWAP)) {
1246 			VM_OBJECT_UNLOCK(source);
1247 			return;
1248 		}
1249 		VM_OBJECT_UNLOCK(source);
1250 	}
1251 
1252 	/*
1253 	 * Allocate a new object with the given length.
1254 	 */
1255 	result = vm_object_allocate(OBJT_DEFAULT, length);
1256 
1257 	/*
1258 	 * The new object shadows the source object, adding a reference to it.
1259 	 * Our caller changes his reference to point to the new object,
1260 	 * removing a reference to the source object.  Net result: no change
1261 	 * of reference count.
1262 	 *
1263 	 * Try to optimize the result object's page color when shadowing
1264 	 * in order to maintain page coloring consistency in the combined
1265 	 * shadowed object.
1266 	 */
1267 	result->backing_object = source;
1268 	/*
1269 	 * Store the offset into the source object, and fix up the offset into
1270 	 * the new object.
1271 	 */
1272 	result->backing_object_offset = *offset;
1273 	if (source != NULL) {
1274 		VM_OBJECT_LOCK(source);
1275 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1276 		source->shadow_count++;
1277 		source->generation++;
1278 #if VM_NRESERVLEVEL > 0
1279 		result->flags |= source->flags & (OBJ_NEEDGIANT | OBJ_COLORED);
1280 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1281 		    ((1 << (VM_NFREEORDER - 1)) - 1);
1282 #else
1283 		result->flags |= source->flags & OBJ_NEEDGIANT;
1284 #endif
1285 		VM_OBJECT_UNLOCK(source);
1286 	}
1287 
1288 
1289 	/*
1290 	 * Return the new things
1291 	 */
1292 	*offset = 0;
1293 	*object = result;
1294 }
1295 
1296 /*
1297  *	vm_object_split:
1298  *
1299  * Split the pages in a map entry into a new object.  This affords
1300  * easier removal of unused pages, and keeps object inheritance from
1301  * being a negative impact on memory usage.
1302  */
1303 void
1304 vm_object_split(vm_map_entry_t entry)
1305 {
1306 	vm_page_t m, m_next;
1307 	vm_object_t orig_object, new_object, source;
1308 	vm_pindex_t idx, offidxstart;
1309 	vm_size_t size;
1310 
1311 	orig_object = entry->object.vm_object;
1312 	if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1313 		return;
1314 	if (orig_object->ref_count <= 1)
1315 		return;
1316 	VM_OBJECT_UNLOCK(orig_object);
1317 
1318 	offidxstart = OFF_TO_IDX(entry->offset);
1319 	size = atop(entry->end - entry->start);
1320 
1321 	/*
1322 	 * If swap_pager_copy() is later called, it will convert new_object
1323 	 * into a swap object.
1324 	 */
1325 	new_object = vm_object_allocate(OBJT_DEFAULT, size);
1326 
1327 	/*
1328 	 * At this point, the new object is still private, so the order in
1329 	 * which the original and new objects are locked does not matter.
1330 	 */
1331 	VM_OBJECT_LOCK(new_object);
1332 	VM_OBJECT_LOCK(orig_object);
1333 	source = orig_object->backing_object;
1334 	if (source != NULL) {
1335 		VM_OBJECT_LOCK(source);
1336 		if ((source->flags & OBJ_DEAD) != 0) {
1337 			VM_OBJECT_UNLOCK(source);
1338 			VM_OBJECT_UNLOCK(orig_object);
1339 			VM_OBJECT_UNLOCK(new_object);
1340 			vm_object_deallocate(new_object);
1341 			VM_OBJECT_LOCK(orig_object);
1342 			return;
1343 		}
1344 		LIST_INSERT_HEAD(&source->shadow_head,
1345 				  new_object, shadow_list);
1346 		source->shadow_count++;
1347 		source->generation++;
1348 		vm_object_reference_locked(source);	/* for new_object */
1349 		vm_object_clear_flag(source, OBJ_ONEMAPPING);
1350 		VM_OBJECT_UNLOCK(source);
1351 		new_object->backing_object_offset =
1352 			orig_object->backing_object_offset + entry->offset;
1353 		new_object->backing_object = source;
1354 	}
1355 	new_object->flags |= orig_object->flags & OBJ_NEEDGIANT;
1356 retry:
1357 	if ((m = TAILQ_FIRST(&orig_object->memq)) != NULL) {
1358 		if (m->pindex < offidxstart) {
1359 			m = vm_page_splay(offidxstart, orig_object->root);
1360 			if ((orig_object->root = m)->pindex < offidxstart)
1361 				m = TAILQ_NEXT(m, listq);
1362 		}
1363 	}
1364 	vm_page_lock_queues();
1365 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1366 	    m = m_next) {
1367 		m_next = TAILQ_NEXT(m, listq);
1368 
1369 		/*
1370 		 * We must wait for pending I/O to complete before we can
1371 		 * rename the page.
1372 		 *
1373 		 * We do not have to VM_PROT_NONE the page as mappings should
1374 		 * not be changed by this operation.
1375 		 */
1376 		if ((m->oflags & VPO_BUSY) || m->busy) {
1377 			vm_page_flag_set(m, PG_REFERENCED);
1378 			vm_page_unlock_queues();
1379 			VM_OBJECT_UNLOCK(new_object);
1380 			m->oflags |= VPO_WANTED;
1381 			msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1382 			VM_OBJECT_LOCK(new_object);
1383 			goto retry;
1384 		}
1385 		vm_page_rename(m, new_object, idx);
1386 		/* page automatically made dirty by rename and cache handled */
1387 		vm_page_busy(m);
1388 	}
1389 	vm_page_unlock_queues();
1390 	if (orig_object->type == OBJT_SWAP) {
1391 		/*
1392 		 * swap_pager_copy() can sleep, in which case the orig_object's
1393 		 * and new_object's locks are released and reacquired.
1394 		 */
1395 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1396 
1397 		/*
1398 		 * Transfer any cached pages from orig_object to new_object.
1399 		 */
1400 		if (__predict_false(orig_object->cache != NULL))
1401 			vm_page_cache_transfer(orig_object, offidxstart,
1402 			    new_object);
1403 	}
1404 	VM_OBJECT_UNLOCK(orig_object);
1405 	TAILQ_FOREACH(m, &new_object->memq, listq)
1406 		vm_page_wakeup(m);
1407 	VM_OBJECT_UNLOCK(new_object);
1408 	entry->object.vm_object = new_object;
1409 	entry->offset = 0LL;
1410 	vm_object_deallocate(orig_object);
1411 	VM_OBJECT_LOCK(new_object);
1412 }
1413 
1414 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1415 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1416 #define	OBSC_COLLAPSE_WAIT	0x0004
1417 
1418 static int
1419 vm_object_backing_scan(vm_object_t object, int op)
1420 {
1421 	int r = 1;
1422 	vm_page_t p;
1423 	vm_object_t backing_object;
1424 	vm_pindex_t backing_offset_index;
1425 
1426 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1427 	VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1428 
1429 	backing_object = object->backing_object;
1430 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1431 
1432 	/*
1433 	 * Initial conditions
1434 	 */
1435 	if (op & OBSC_TEST_ALL_SHADOWED) {
1436 		/*
1437 		 * We do not want to have to test for the existence of cache
1438 		 * or swap pages in the backing object.  XXX but with the
1439 		 * new swapper this would be pretty easy to do.
1440 		 *
1441 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1442 		 * been ZFOD faulted yet?  If we do not test for this, the
1443 		 * shadow test may succeed! XXX
1444 		 */
1445 		if (backing_object->type != OBJT_DEFAULT) {
1446 			return (0);
1447 		}
1448 	}
1449 	if (op & OBSC_COLLAPSE_WAIT) {
1450 		vm_object_set_flag(backing_object, OBJ_DEAD);
1451 	}
1452 
1453 	/*
1454 	 * Our scan
1455 	 */
1456 	p = TAILQ_FIRST(&backing_object->memq);
1457 	while (p) {
1458 		vm_page_t next = TAILQ_NEXT(p, listq);
1459 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1460 
1461 		if (op & OBSC_TEST_ALL_SHADOWED) {
1462 			vm_page_t pp;
1463 
1464 			/*
1465 			 * Ignore pages outside the parent object's range
1466 			 * and outside the parent object's mapping of the
1467 			 * backing object.
1468 			 *
1469 			 * note that we do not busy the backing object's
1470 			 * page.
1471 			 */
1472 			if (
1473 			    p->pindex < backing_offset_index ||
1474 			    new_pindex >= object->size
1475 			) {
1476 				p = next;
1477 				continue;
1478 			}
1479 
1480 			/*
1481 			 * See if the parent has the page or if the parent's
1482 			 * object pager has the page.  If the parent has the
1483 			 * page but the page is not valid, the parent's
1484 			 * object pager must have the page.
1485 			 *
1486 			 * If this fails, the parent does not completely shadow
1487 			 * the object and we might as well give up now.
1488 			 */
1489 
1490 			pp = vm_page_lookup(object, new_pindex);
1491 			if (
1492 			    (pp == NULL || pp->valid == 0) &&
1493 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1494 			) {
1495 				r = 0;
1496 				break;
1497 			}
1498 		}
1499 
1500 		/*
1501 		 * Check for busy page
1502 		 */
1503 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1504 			vm_page_t pp;
1505 
1506 			if (op & OBSC_COLLAPSE_NOWAIT) {
1507 				if ((p->oflags & VPO_BUSY) ||
1508 				    !p->valid ||
1509 				    p->busy) {
1510 					p = next;
1511 					continue;
1512 				}
1513 			} else if (op & OBSC_COLLAPSE_WAIT) {
1514 				if ((p->oflags & VPO_BUSY) || p->busy) {
1515 					vm_page_lock_queues();
1516 					vm_page_flag_set(p, PG_REFERENCED);
1517 					vm_page_unlock_queues();
1518 					VM_OBJECT_UNLOCK(object);
1519 					p->oflags |= VPO_WANTED;
1520 					msleep(p, VM_OBJECT_MTX(backing_object),
1521 					    PDROP | PVM, "vmocol", 0);
1522 					VM_OBJECT_LOCK(object);
1523 					VM_OBJECT_LOCK(backing_object);
1524 					/*
1525 					 * If we slept, anything could have
1526 					 * happened.  Since the object is
1527 					 * marked dead, the backing offset
1528 					 * should not have changed so we
1529 					 * just restart our scan.
1530 					 */
1531 					p = TAILQ_FIRST(&backing_object->memq);
1532 					continue;
1533 				}
1534 			}
1535 
1536 			KASSERT(
1537 			    p->object == backing_object,
1538 			    ("vm_object_backing_scan: object mismatch")
1539 			);
1540 
1541 			/*
1542 			 * Destroy any associated swap
1543 			 */
1544 			if (backing_object->type == OBJT_SWAP) {
1545 				swap_pager_freespace(
1546 				    backing_object,
1547 				    p->pindex,
1548 				    1
1549 				);
1550 			}
1551 
1552 			if (
1553 			    p->pindex < backing_offset_index ||
1554 			    new_pindex >= object->size
1555 			) {
1556 				/*
1557 				 * Page is out of the parent object's range, we
1558 				 * can simply destroy it.
1559 				 */
1560 				vm_page_lock_queues();
1561 				KASSERT(!pmap_page_is_mapped(p),
1562 				    ("freeing mapped page %p", p));
1563 				if (p->wire_count == 0)
1564 					vm_page_free(p);
1565 				else
1566 					vm_page_remove(p);
1567 				vm_page_unlock_queues();
1568 				p = next;
1569 				continue;
1570 			}
1571 
1572 			pp = vm_page_lookup(object, new_pindex);
1573 			if (
1574 			    pp != NULL ||
1575 			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1576 			) {
1577 				/*
1578 				 * page already exists in parent OR swap exists
1579 				 * for this location in the parent.  Destroy
1580 				 * the original page from the backing object.
1581 				 *
1582 				 * Leave the parent's page alone
1583 				 */
1584 				vm_page_lock_queues();
1585 				KASSERT(!pmap_page_is_mapped(p),
1586 				    ("freeing mapped page %p", p));
1587 				if (p->wire_count == 0)
1588 					vm_page_free(p);
1589 				else
1590 					vm_page_remove(p);
1591 				vm_page_unlock_queues();
1592 				p = next;
1593 				continue;
1594 			}
1595 
1596 #if VM_NRESERVLEVEL > 0
1597 			/*
1598 			 * Rename the reservation.
1599 			 */
1600 			vm_reserv_rename(p, object, backing_object,
1601 			    backing_offset_index);
1602 #endif
1603 
1604 			/*
1605 			 * Page does not exist in parent, rename the
1606 			 * page from the backing object to the main object.
1607 			 *
1608 			 * If the page was mapped to a process, it can remain
1609 			 * mapped through the rename.
1610 			 */
1611 			vm_page_lock_queues();
1612 			vm_page_rename(p, object, new_pindex);
1613 			vm_page_unlock_queues();
1614 			/* page automatically made dirty by rename */
1615 		}
1616 		p = next;
1617 	}
1618 	return (r);
1619 }
1620 
1621 
1622 /*
1623  * this version of collapse allows the operation to occur earlier and
1624  * when paging_in_progress is true for an object...  This is not a complete
1625  * operation, but should plug 99.9% of the rest of the leaks.
1626  */
1627 static void
1628 vm_object_qcollapse(vm_object_t object)
1629 {
1630 	vm_object_t backing_object = object->backing_object;
1631 
1632 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1633 	VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1634 
1635 	if (backing_object->ref_count != 1)
1636 		return;
1637 
1638 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1639 }
1640 
1641 /*
1642  *	vm_object_collapse:
1643  *
1644  *	Collapse an object with the object backing it.
1645  *	Pages in the backing object are moved into the
1646  *	parent, and the backing object is deallocated.
1647  */
1648 void
1649 vm_object_collapse(vm_object_t object)
1650 {
1651 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1652 
1653 	while (TRUE) {
1654 		vm_object_t backing_object;
1655 
1656 		/*
1657 		 * Verify that the conditions are right for collapse:
1658 		 *
1659 		 * The object exists and the backing object exists.
1660 		 */
1661 		if ((backing_object = object->backing_object) == NULL)
1662 			break;
1663 
1664 		/*
1665 		 * we check the backing object first, because it is most likely
1666 		 * not collapsable.
1667 		 */
1668 		VM_OBJECT_LOCK(backing_object);
1669 		if (backing_object->handle != NULL ||
1670 		    (backing_object->type != OBJT_DEFAULT &&
1671 		     backing_object->type != OBJT_SWAP) ||
1672 		    (backing_object->flags & OBJ_DEAD) ||
1673 		    object->handle != NULL ||
1674 		    (object->type != OBJT_DEFAULT &&
1675 		     object->type != OBJT_SWAP) ||
1676 		    (object->flags & OBJ_DEAD)) {
1677 			VM_OBJECT_UNLOCK(backing_object);
1678 			break;
1679 		}
1680 
1681 		if (
1682 		    object->paging_in_progress != 0 ||
1683 		    backing_object->paging_in_progress != 0
1684 		) {
1685 			vm_object_qcollapse(object);
1686 			VM_OBJECT_UNLOCK(backing_object);
1687 			break;
1688 		}
1689 		/*
1690 		 * We know that we can either collapse the backing object (if
1691 		 * the parent is the only reference to it) or (perhaps) have
1692 		 * the parent bypass the object if the parent happens to shadow
1693 		 * all the resident pages in the entire backing object.
1694 		 *
1695 		 * This is ignoring pager-backed pages such as swap pages.
1696 		 * vm_object_backing_scan fails the shadowing test in this
1697 		 * case.
1698 		 */
1699 		if (backing_object->ref_count == 1) {
1700 			/*
1701 			 * If there is exactly one reference to the backing
1702 			 * object, we can collapse it into the parent.
1703 			 */
1704 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1705 
1706 #if VM_NRESERVLEVEL > 0
1707 			/*
1708 			 * Break any reservations from backing_object.
1709 			 */
1710 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1711 				vm_reserv_break_all(backing_object);
1712 #endif
1713 
1714 			/*
1715 			 * Move the pager from backing_object to object.
1716 			 */
1717 			if (backing_object->type == OBJT_SWAP) {
1718 				/*
1719 				 * swap_pager_copy() can sleep, in which case
1720 				 * the backing_object's and object's locks are
1721 				 * released and reacquired.
1722 				 */
1723 				swap_pager_copy(
1724 				    backing_object,
1725 				    object,
1726 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1727 
1728 				/*
1729 				 * Free any cached pages from backing_object.
1730 				 */
1731 				if (__predict_false(backing_object->cache != NULL))
1732 					vm_page_cache_free(backing_object, 0, 0);
1733 			}
1734 			/*
1735 			 * Object now shadows whatever backing_object did.
1736 			 * Note that the reference to
1737 			 * backing_object->backing_object moves from within
1738 			 * backing_object to within object.
1739 			 */
1740 			LIST_REMOVE(object, shadow_list);
1741 			backing_object->shadow_count--;
1742 			backing_object->generation++;
1743 			if (backing_object->backing_object) {
1744 				VM_OBJECT_LOCK(backing_object->backing_object);
1745 				LIST_REMOVE(backing_object, shadow_list);
1746 				LIST_INSERT_HEAD(
1747 				    &backing_object->backing_object->shadow_head,
1748 				    object, shadow_list);
1749 				/*
1750 				 * The shadow_count has not changed.
1751 				 */
1752 				backing_object->backing_object->generation++;
1753 				VM_OBJECT_UNLOCK(backing_object->backing_object);
1754 			}
1755 			object->backing_object = backing_object->backing_object;
1756 			object->backing_object_offset +=
1757 			    backing_object->backing_object_offset;
1758 
1759 			/*
1760 			 * Discard backing_object.
1761 			 *
1762 			 * Since the backing object has no pages, no pager left,
1763 			 * and no object references within it, all that is
1764 			 * necessary is to dispose of it.
1765 			 */
1766 			KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1767 			VM_OBJECT_UNLOCK(backing_object);
1768 
1769 			mtx_lock(&vm_object_list_mtx);
1770 			TAILQ_REMOVE(
1771 			    &vm_object_list,
1772 			    backing_object,
1773 			    object_list
1774 			);
1775 			mtx_unlock(&vm_object_list_mtx);
1776 
1777 			uma_zfree(obj_zone, backing_object);
1778 
1779 			object_collapses++;
1780 		} else {
1781 			vm_object_t new_backing_object;
1782 
1783 			/*
1784 			 * If we do not entirely shadow the backing object,
1785 			 * there is nothing we can do so we give up.
1786 			 */
1787 			if (object->resident_page_count != object->size &&
1788 			    vm_object_backing_scan(object,
1789 			    OBSC_TEST_ALL_SHADOWED) == 0) {
1790 				VM_OBJECT_UNLOCK(backing_object);
1791 				break;
1792 			}
1793 
1794 			/*
1795 			 * Make the parent shadow the next object in the
1796 			 * chain.  Deallocating backing_object will not remove
1797 			 * it, since its reference count is at least 2.
1798 			 */
1799 			LIST_REMOVE(object, shadow_list);
1800 			backing_object->shadow_count--;
1801 			backing_object->generation++;
1802 
1803 			new_backing_object = backing_object->backing_object;
1804 			if ((object->backing_object = new_backing_object) != NULL) {
1805 				VM_OBJECT_LOCK(new_backing_object);
1806 				LIST_INSERT_HEAD(
1807 				    &new_backing_object->shadow_head,
1808 				    object,
1809 				    shadow_list
1810 				);
1811 				new_backing_object->shadow_count++;
1812 				new_backing_object->generation++;
1813 				vm_object_reference_locked(new_backing_object);
1814 				VM_OBJECT_UNLOCK(new_backing_object);
1815 				object->backing_object_offset +=
1816 					backing_object->backing_object_offset;
1817 			}
1818 
1819 			/*
1820 			 * Drop the reference count on backing_object. Since
1821 			 * its ref_count was at least 2, it will not vanish.
1822 			 */
1823 			backing_object->ref_count--;
1824 			VM_OBJECT_UNLOCK(backing_object);
1825 			object_bypasses++;
1826 		}
1827 
1828 		/*
1829 		 * Try again with this object's new backing object.
1830 		 */
1831 	}
1832 }
1833 
1834 /*
1835  *	vm_object_page_remove:
1836  *
1837  *	For the given object, either frees or invalidates each of the
1838  *	specified pages.  In general, a page is freed.  However, if a
1839  *	page is wired for any reason other than the existence of a
1840  *	managed, wired mapping, then it may be invalidated but not
1841  *	removed from the object.  Pages are specified by the given
1842  *	range ["start", "end") and Boolean "clean_only".  As a
1843  *	special case, if "end" is zero, then the range extends from
1844  *	"start" to the end of the object.  If "clean_only" is TRUE,
1845  *	then only the non-dirty pages within the specified range are
1846  *	affected.
1847  *
1848  *	In general, this operation should only be performed on objects
1849  *	that contain managed pages.  There are two exceptions.  First,
1850  *	it may be performed on the kernel and kmem objects.  Second,
1851  *	it may be used by msync(..., MS_INVALIDATE) to invalidate
1852  *	device-backed pages.
1853  *
1854  *	The object must be locked.
1855  */
1856 void
1857 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1858     boolean_t clean_only)
1859 {
1860 	vm_page_t p, next;
1861 	int wirings;
1862 
1863 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1864 	if (object->resident_page_count == 0)
1865 		goto skipmemq;
1866 
1867 	/*
1868 	 * Since physically-backed objects do not use managed pages, we can't
1869 	 * remove pages from the object (we must instead remove the page
1870 	 * references, and then destroy the object).
1871 	 */
1872 	KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1873 	    object == kmem_object,
1874 	    ("attempt to remove pages from a physical object"));
1875 
1876 	vm_object_pip_add(object, 1);
1877 again:
1878 	vm_page_lock_queues();
1879 	if ((p = TAILQ_FIRST(&object->memq)) != NULL) {
1880 		if (p->pindex < start) {
1881 			p = vm_page_splay(start, object->root);
1882 			if ((object->root = p)->pindex < start)
1883 				p = TAILQ_NEXT(p, listq);
1884 		}
1885 	}
1886 	/*
1887 	 * Assert: the variable p is either (1) the page with the
1888 	 * least pindex greater than or equal to the parameter pindex
1889 	 * or (2) NULL.
1890 	 */
1891 	for (;
1892 	     p != NULL && (p->pindex < end || end == 0);
1893 	     p = next) {
1894 		next = TAILQ_NEXT(p, listq);
1895 
1896 		/*
1897 		 * If the page is wired for any reason besides the
1898 		 * existence of managed, wired mappings, then it cannot
1899 		 * be freed.  For example, fictitious pages, which
1900 		 * represent device memory, are inherently wired and
1901 		 * cannot be freed.  They can, however, be invalidated
1902 		 * if "clean_only" is FALSE.
1903 		 */
1904 		if ((wirings = p->wire_count) != 0 &&
1905 		    (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1906 			/* Fictitious pages do not have managed mappings. */
1907 			if ((p->flags & PG_FICTITIOUS) == 0)
1908 				pmap_remove_all(p);
1909 			/* Account for removal of managed, wired mappings. */
1910 			p->wire_count -= wirings;
1911 			if (!clean_only)
1912 				p->valid = 0;
1913 			continue;
1914 		}
1915 		if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1916 			goto again;
1917 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
1918 		    ("vm_object_page_remove: page %p is fictitious", p));
1919 		if (clean_only && p->valid) {
1920 			pmap_remove_write(p);
1921 			if (p->valid & p->dirty)
1922 				continue;
1923 		}
1924 		pmap_remove_all(p);
1925 		/* Account for removal of managed, wired mappings. */
1926 		if (wirings != 0)
1927 			p->wire_count -= wirings;
1928 		vm_page_free(p);
1929 	}
1930 	vm_page_unlock_queues();
1931 	vm_object_pip_wakeup(object);
1932 skipmemq:
1933 	if (__predict_false(object->cache != NULL))
1934 		vm_page_cache_free(object, start, end);
1935 }
1936 
1937 /*
1938  *	Routine:	vm_object_coalesce
1939  *	Function:	Coalesces two objects backing up adjoining
1940  *			regions of memory into a single object.
1941  *
1942  *	returns TRUE if objects were combined.
1943  *
1944  *	NOTE:	Only works at the moment if the second object is NULL -
1945  *		if it's not, which object do we lock first?
1946  *
1947  *	Parameters:
1948  *		prev_object	First object to coalesce
1949  *		prev_offset	Offset into prev_object
1950  *		prev_size	Size of reference to prev_object
1951  *		next_size	Size of reference to the second object
1952  *
1953  *	Conditions:
1954  *	The object must *not* be locked.
1955  */
1956 boolean_t
1957 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
1958 	vm_size_t prev_size, vm_size_t next_size)
1959 {
1960 	vm_pindex_t next_pindex;
1961 
1962 	if (prev_object == NULL)
1963 		return (TRUE);
1964 	VM_OBJECT_LOCK(prev_object);
1965 	if (prev_object->type != OBJT_DEFAULT &&
1966 	    prev_object->type != OBJT_SWAP) {
1967 		VM_OBJECT_UNLOCK(prev_object);
1968 		return (FALSE);
1969 	}
1970 
1971 	/*
1972 	 * Try to collapse the object first
1973 	 */
1974 	vm_object_collapse(prev_object);
1975 
1976 	/*
1977 	 * Can't coalesce if: . more than one reference . paged out . shadows
1978 	 * another object . has a copy elsewhere (any of which mean that the
1979 	 * pages not mapped to prev_entry may be in use anyway)
1980 	 */
1981 	if (prev_object->backing_object != NULL) {
1982 		VM_OBJECT_UNLOCK(prev_object);
1983 		return (FALSE);
1984 	}
1985 
1986 	prev_size >>= PAGE_SHIFT;
1987 	next_size >>= PAGE_SHIFT;
1988 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
1989 
1990 	if ((prev_object->ref_count > 1) &&
1991 	    (prev_object->size != next_pindex)) {
1992 		VM_OBJECT_UNLOCK(prev_object);
1993 		return (FALSE);
1994 	}
1995 
1996 	/*
1997 	 * Remove any pages that may still be in the object from a previous
1998 	 * deallocation.
1999 	 */
2000 	if (next_pindex < prev_object->size) {
2001 		vm_object_page_remove(prev_object,
2002 				      next_pindex,
2003 				      next_pindex + next_size, FALSE);
2004 		if (prev_object->type == OBJT_SWAP)
2005 			swap_pager_freespace(prev_object,
2006 					     next_pindex, next_size);
2007 	}
2008 
2009 	/*
2010 	 * Extend the object if necessary.
2011 	 */
2012 	if (next_pindex + next_size > prev_object->size)
2013 		prev_object->size = next_pindex + next_size;
2014 
2015 	VM_OBJECT_UNLOCK(prev_object);
2016 	return (TRUE);
2017 }
2018 
2019 void
2020 vm_object_set_writeable_dirty(vm_object_t object)
2021 {
2022 	struct vnode *vp;
2023 
2024 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2025 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2026 		return;
2027 	vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2028 	if (object->type == OBJT_VNODE &&
2029 	    (vp = (struct vnode *)object->handle) != NULL) {
2030 		VI_LOCK(vp);
2031 		vp->v_iflag |= VI_OBJDIRTY;
2032 		VI_UNLOCK(vp);
2033 	}
2034 }
2035 
2036 #include "opt_ddb.h"
2037 #ifdef DDB
2038 #include <sys/kernel.h>
2039 
2040 #include <sys/cons.h>
2041 
2042 #include <ddb/ddb.h>
2043 
2044 static int
2045 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2046 {
2047 	vm_map_t tmpm;
2048 	vm_map_entry_t tmpe;
2049 	vm_object_t obj;
2050 	int entcount;
2051 
2052 	if (map == 0)
2053 		return 0;
2054 
2055 	if (entry == 0) {
2056 		tmpe = map->header.next;
2057 		entcount = map->nentries;
2058 		while (entcount-- && (tmpe != &map->header)) {
2059 			if (_vm_object_in_map(map, object, tmpe)) {
2060 				return 1;
2061 			}
2062 			tmpe = tmpe->next;
2063 		}
2064 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2065 		tmpm = entry->object.sub_map;
2066 		tmpe = tmpm->header.next;
2067 		entcount = tmpm->nentries;
2068 		while (entcount-- && tmpe != &tmpm->header) {
2069 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2070 				return 1;
2071 			}
2072 			tmpe = tmpe->next;
2073 		}
2074 	} else if ((obj = entry->object.vm_object) != NULL) {
2075 		for (; obj; obj = obj->backing_object)
2076 			if (obj == object) {
2077 				return 1;
2078 			}
2079 	}
2080 	return 0;
2081 }
2082 
2083 static int
2084 vm_object_in_map(vm_object_t object)
2085 {
2086 	struct proc *p;
2087 
2088 	/* sx_slock(&allproc_lock); */
2089 	FOREACH_PROC_IN_SYSTEM(p) {
2090 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2091 			continue;
2092 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2093 			/* sx_sunlock(&allproc_lock); */
2094 			return 1;
2095 		}
2096 	}
2097 	/* sx_sunlock(&allproc_lock); */
2098 	if (_vm_object_in_map(kernel_map, object, 0))
2099 		return 1;
2100 	if (_vm_object_in_map(kmem_map, object, 0))
2101 		return 1;
2102 	if (_vm_object_in_map(pager_map, object, 0))
2103 		return 1;
2104 	if (_vm_object_in_map(buffer_map, object, 0))
2105 		return 1;
2106 	return 0;
2107 }
2108 
2109 DB_SHOW_COMMAND(vmochk, vm_object_check)
2110 {
2111 	vm_object_t object;
2112 
2113 	/*
2114 	 * make sure that internal objs are in a map somewhere
2115 	 * and none have zero ref counts.
2116 	 */
2117 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2118 		if (object->handle == NULL &&
2119 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2120 			if (object->ref_count == 0) {
2121 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2122 					(long)object->size);
2123 			}
2124 			if (!vm_object_in_map(object)) {
2125 				db_printf(
2126 			"vmochk: internal obj is not in a map: "
2127 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2128 				    object->ref_count, (u_long)object->size,
2129 				    (u_long)object->size,
2130 				    (void *)object->backing_object);
2131 			}
2132 		}
2133 	}
2134 }
2135 
2136 /*
2137  *	vm_object_print:	[ debug ]
2138  */
2139 DB_SHOW_COMMAND(object, vm_object_print_static)
2140 {
2141 	/* XXX convert args. */
2142 	vm_object_t object = (vm_object_t)addr;
2143 	boolean_t full = have_addr;
2144 
2145 	vm_page_t p;
2146 
2147 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2148 #define	count	was_count
2149 
2150 	int count;
2151 
2152 	if (object == NULL)
2153 		return;
2154 
2155 	db_iprintf(
2156 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x\n",
2157 	    object, (int)object->type, (uintmax_t)object->size,
2158 	    object->resident_page_count, object->ref_count, object->flags);
2159 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2160 	    object->shadow_count,
2161 	    object->backing_object ? object->backing_object->ref_count : 0,
2162 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2163 
2164 	if (!full)
2165 		return;
2166 
2167 	db_indent += 2;
2168 	count = 0;
2169 	TAILQ_FOREACH(p, &object->memq, listq) {
2170 		if (count == 0)
2171 			db_iprintf("memory:=");
2172 		else if (count == 6) {
2173 			db_printf("\n");
2174 			db_iprintf(" ...");
2175 			count = 0;
2176 		} else
2177 			db_printf(",");
2178 		count++;
2179 
2180 		db_printf("(off=0x%jx,page=0x%jx)",
2181 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2182 	}
2183 	if (count != 0)
2184 		db_printf("\n");
2185 	db_indent -= 2;
2186 }
2187 
2188 /* XXX. */
2189 #undef count
2190 
2191 /* XXX need this non-static entry for calling from vm_map_print. */
2192 void
2193 vm_object_print(
2194         /* db_expr_t */ long addr,
2195 	boolean_t have_addr,
2196 	/* db_expr_t */ long count,
2197 	char *modif)
2198 {
2199 	vm_object_print_static(addr, have_addr, count, modif);
2200 }
2201 
2202 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2203 {
2204 	vm_object_t object;
2205 	int nl = 0;
2206 	int c;
2207 
2208 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2209 		vm_pindex_t idx, fidx;
2210 		vm_pindex_t osize;
2211 		vm_paddr_t pa = -1;
2212 		int rcount;
2213 		vm_page_t m;
2214 
2215 		db_printf("new object: %p\n", (void *)object);
2216 		if (nl > 18) {
2217 			c = cngetc();
2218 			if (c != ' ')
2219 				return;
2220 			nl = 0;
2221 		}
2222 		nl++;
2223 		rcount = 0;
2224 		fidx = 0;
2225 		osize = object->size;
2226 		if (osize > 128)
2227 			osize = 128;
2228 		for (idx = 0; idx < osize; idx++) {
2229 			m = vm_page_lookup(object, idx);
2230 			if (m == NULL) {
2231 				if (rcount) {
2232 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2233 						(long)fidx, rcount, (long)pa);
2234 					if (nl > 18) {
2235 						c = cngetc();
2236 						if (c != ' ')
2237 							return;
2238 						nl = 0;
2239 					}
2240 					nl++;
2241 					rcount = 0;
2242 				}
2243 				continue;
2244 			}
2245 
2246 
2247 			if (rcount &&
2248 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2249 				++rcount;
2250 				continue;
2251 			}
2252 			if (rcount) {
2253 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2254 					(long)fidx, rcount, (long)pa);
2255 				if (nl > 18) {
2256 					c = cngetc();
2257 					if (c != ' ')
2258 						return;
2259 					nl = 0;
2260 				}
2261 				nl++;
2262 			}
2263 			fidx = idx;
2264 			pa = VM_PAGE_TO_PHYS(m);
2265 			rcount = 1;
2266 		}
2267 		if (rcount) {
2268 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2269 				(long)fidx, rcount, (long)pa);
2270 			if (nl > 18) {
2271 				c = cngetc();
2272 				if (c != ' ')
2273 					return;
2274 				nl = 0;
2275 			}
2276 			nl++;
2277 		}
2278 	}
2279 }
2280 #endif /* DDB */
2281