xref: /freebsd/sys/vm/vm_object.c (revision a316b26e50bbed7cf655fbba726ab87d8ab7599d)
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  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *	This product includes software developed by the University of
19  *	California, Berkeley and its contributors.
20  * 4. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	from: @(#)vm_object.c	8.5 (Berkeley) 3/22/94
37  *
38  *
39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40  * All rights reserved.
41  *
42  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  *
64  * $Id: vm_object.c,v 1.14 1995/01/05 04:30:40 davidg Exp $
65  */
66 
67 /*
68  *	Virtual memory object module.
69  */
70 
71 #include <sys/param.h>
72 #include <sys/systm.h>
73 #include <sys/kernel.h>
74 #include <sys/proc.h>		/* for curproc, pageproc */
75 #include <sys/malloc.h>
76 #include <sys/vnode.h>
77 #include <sys/mount.h>
78 
79 #include <vm/vm.h>
80 #include <vm/vm_page.h>
81 #include <vm/vm_pageout.h>
82 #include <vm/vm_pager.h>
83 #include <vm/swap_pager.h>
84 #include <vm/vnode_pager.h>
85 
86 static void _vm_object_allocate(vm_size_t, vm_object_t);
87 static void vm_object_rcollapse(vm_object_t, vm_object_t);
88 
89 /*
90  *	Virtual memory objects maintain the actual data
91  *	associated with allocated virtual memory.  A given
92  *	page of memory exists within exactly one object.
93  *
94  *	An object is only deallocated when all "references"
95  *	are given up.  Only one "reference" to a given
96  *	region of an object should be writeable.
97  *
98  *	Associated with each object is a list of all resident
99  *	memory pages belonging to that object; this list is
100  *	maintained by the "vm_page" module, and locked by the object's
101  *	lock.
102  *
103  *	Each object also records a "pager" routine which is
104  *	used to retrieve (and store) pages to the proper backing
105  *	storage.  In addition, objects may be backed by other
106  *	objects from which they were virtual-copied.
107  *
108  *	The only items within the object structure which are
109  *	modified after time of creation are:
110  *		reference count		locked by object's lock
111  *		pager routine		locked by object's lock
112  *
113  */
114 
115 
116 struct vm_object kernel_object_store;
117 struct vm_object kmem_object_store;
118 
119 int vm_object_cache_max;
120 
121 #define	VM_OBJECT_HASH_COUNT	509
122 
123 struct vm_object_hash_head vm_object_hashtable[VM_OBJECT_HASH_COUNT];
124 
125 long object_collapses = 0;
126 long object_bypasses = 0;
127 
128 static void
129 _vm_object_allocate(size, object)
130 	vm_size_t size;
131 	register vm_object_t object;
132 {
133 	bzero(object, sizeof *object);
134 	TAILQ_INIT(&object->memq);
135 	TAILQ_INIT(&object->reverse_shadow_head);
136 	vm_object_lock_init(object);
137 	object->ref_count = 1;
138 	object->resident_page_count = 0;
139 	object->size = size;
140 	object->flags = OBJ_INTERNAL;	/* vm_allocate_with_pager will reset */
141 	object->paging_in_progress = 0;
142 	object->copy = NULL;
143 
144 	/*
145 	 * Object starts out read-write, with no pager.
146 	 */
147 
148 	object->pager = NULL;
149 	object->paging_offset = 0;
150 	object->shadow = NULL;
151 	object->shadow_offset = (vm_offset_t) 0;
152 
153 	simple_lock(&vm_object_list_lock);
154 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
155 	vm_object_count++;
156 	cnt.v_nzfod += atop(size);
157 	simple_unlock(&vm_object_list_lock);
158 }
159 
160 /*
161  *	vm_object_init:
162  *
163  *	Initialize the VM objects module.
164  */
165 void
166 vm_object_init(vm_offset_t nothing)
167 {
168 	register int i;
169 
170 	TAILQ_INIT(&vm_object_cached_list);
171 	TAILQ_INIT(&vm_object_list);
172 	vm_object_count = 0;
173 	simple_lock_init(&vm_cache_lock);
174 	simple_lock_init(&vm_object_list_lock);
175 	vm_object_cache_max = (cnt.v_page_count - 500) / 8;
176 
177 	for (i = 0; i < VM_OBJECT_HASH_COUNT; i++)
178 		TAILQ_INIT(&vm_object_hashtable[i]);
179 
180 	kernel_object = &kernel_object_store;
181 	_vm_object_allocate(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS,
182 	    kernel_object);
183 
184 	kmem_object = &kmem_object_store;
185 	_vm_object_allocate(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS,
186 	    kmem_object);
187 }
188 
189 /*
190  *	vm_object_allocate:
191  *
192  *	Returns a new object with the given size.
193  */
194 
195 vm_object_t
196 vm_object_allocate(size)
197 	vm_size_t size;
198 {
199 	register vm_object_t result;
200 
201 	result = (vm_object_t)
202 	    malloc((u_long) sizeof *result, M_VMOBJ, M_WAITOK);
203 
204 
205 	_vm_object_allocate(size, result);
206 
207 	return (result);
208 }
209 
210 
211 /*
212  *	vm_object_reference:
213  *
214  *	Gets another reference to the given object.
215  */
216 inline void
217 vm_object_reference(object)
218 	register vm_object_t object;
219 {
220 	if (object == NULL)
221 		return;
222 
223 	vm_object_lock(object);
224 	object->ref_count++;
225 	vm_object_unlock(object);
226 }
227 
228 /*
229  *	vm_object_deallocate:
230  *
231  *	Release a reference to the specified object,
232  *	gained either through a vm_object_allocate
233  *	or a vm_object_reference call.  When all references
234  *	are gone, storage associated with this object
235  *	may be relinquished.
236  *
237  *	No object may be locked.
238  */
239 void
240 vm_object_deallocate(object)
241 	vm_object_t object;
242 {
243 	vm_object_t temp;
244 
245 	while (object != NULL) {
246 
247 		/*
248 		 * The cache holds a reference (uncounted) to the object; we
249 		 * must lock it before removing the object.
250 		 */
251 
252 		vm_object_cache_lock();
253 
254 		/*
255 		 * Lose the reference
256 		 */
257 		vm_object_lock(object);
258 		if (--(object->ref_count) != 0) {
259 			if (object->ref_count == 1) {
260 				if (object->reverse_shadow_head.tqh_first) {
261 					++object->reverse_shadow_head.tqh_first->ref_count;
262 					if (vm_object_lock_try(object->reverse_shadow_head.tqh_first)) {
263 						vm_object_rcollapse(object->reverse_shadow_head.tqh_first, object);
264 						vm_object_unlock(object->reverse_shadow_head.tqh_first);
265 					}
266 					vm_object_deallocate(object->reverse_shadow_head.tqh_first);
267 				}
268 			}
269 			vm_object_unlock(object);
270 			/*
271 			 * If there are still references, then we are done.
272 			 */
273 			vm_object_cache_unlock();
274 			return;
275 		}
276 		/*
277 		 * See if this object can persist.  If so, enter it in the
278 		 * cache, then deactivate all of its pages.
279 		 */
280 
281 		if (object->flags & OBJ_CANPERSIST) {
282 
283 			TAILQ_INSERT_TAIL(&vm_object_cached_list, object,
284 			    cached_list);
285 			vm_object_cached++;
286 			vm_object_cache_unlock();
287 
288 			vm_object_unlock(object);
289 
290 			vm_object_cache_trim();
291 			return;
292 		}
293 		/*
294 		 * Make sure no one can look us up now.
295 		 */
296 		object->flags |= OBJ_DEAD;
297 		vm_object_remove(object->pager);
298 		vm_object_cache_unlock();
299 
300 		temp = object->shadow;
301 		if (temp)
302 			TAILQ_REMOVE(&temp->reverse_shadow_head, object, reverse_shadow_list);
303 		vm_object_terminate(object);
304 		/* unlocks and deallocates object */
305 		object = temp;
306 	}
307 }
308 
309 /*
310  *	vm_object_terminate actually destroys the specified object, freeing
311  *	up all previously used resources.
312  *
313  *	The object must be locked.
314  */
315 void
316 vm_object_terminate(object)
317 	register vm_object_t object;
318 {
319 	register vm_page_t p, next;
320 	vm_object_t shadow_object;
321 	int s;
322 	struct vnode *vp = NULL;
323 
324 	/*
325 	 * Detach the object from its shadow if we are the shadow's copy.
326 	 */
327 	if ((shadow_object = object->shadow) != NULL) {
328 		vm_object_lock(shadow_object);
329 		if (shadow_object->copy == object)
330 			shadow_object->copy = NULL;
331 /*
332 		else if (shadow_object->copy != NULL)
333 			panic("vm_object_terminate: copy/shadow inconsistency");
334 */
335 		vm_object_unlock(shadow_object);
336 	}
337 	if (object->pager && (object->pager->pg_type == PG_VNODE)) {
338 		vn_pager_t vnp = object->pager->pg_data;
339 
340 		vp = vnp->vnp_vp;
341 		VOP_FSYNC(vp, NOCRED, MNT_WAIT, NULL);
342 		vinvalbuf(vp, 0, NOCRED, NULL, 0, 0);
343 	}
344 	/*
345 	 * Wait until the pageout daemon is through with the object.
346 	 */
347 
348 	s = splhigh();
349 	while (object->paging_in_progress) {
350 		vm_object_unlock(object);
351 		tsleep((caddr_t) object, PVM, "objtrm", 0);
352 		vm_object_lock(object);
353 	}
354 	splx(s);
355 
356 	/*
357 	 * While the paging system is locked, pull the object's pages off the
358 	 * active and inactive queues.  This keeps the pageout daemon from
359 	 * playing with them during vm_pager_deallocate.
360 	 *
361 	 * We can't free the pages yet, because the object's pager may have to
362 	 * write them out before deallocating the paging space.
363 	 */
364 
365 	for (p = object->memq.tqh_first; p; p = next) {
366 		VM_PAGE_CHECK(p);
367 		next = p->listq.tqe_next;
368 
369 		vm_page_lock_queues();
370 		if (p->flags & PG_CACHE)
371 			vm_page_free(p);
372 		else
373 			vm_page_unqueue(p);
374 		vm_page_unlock_queues();
375 		p = next;
376 	}
377 
378 	if (object->paging_in_progress != 0)
379 		panic("vm_object_deallocate: pageout in progress");
380 
381 	/*
382 	 * Clean and free the pages, as appropriate. All references to the
383 	 * object are gone, so we don't need to lock it.
384 	 */
385 
386 	if ((object->flags & OBJ_INTERNAL) == 0) {
387 		(void) vm_object_page_clean(object, 0, 0, TRUE, TRUE);
388 	}
389 	/*
390 	 * one last time -- get rid of buffers that might have been created
391 	 * for the vm_object_page_clean
392 	 */
393 	if (vp != NULL) {
394 		vm_object_unlock(object);
395 		vinvalbuf(vp, 0, NOCRED, NULL, 0, 0);
396 		vm_object_lock(object);
397 	}
398 	/*
399 	 * Now free the pages. For internal objects, this also removes them
400 	 * from paging queues.
401 	 */
402 	while ((p = object->memq.tqh_first) != NULL) {
403 		VM_PAGE_CHECK(p);
404 		vm_page_lock_queues();
405 		PAGE_WAKEUP(p);
406 		vm_page_free(p);
407 		cnt.v_pfree++;
408 		vm_page_unlock_queues();
409 	}
410 	vm_object_unlock(object);
411 
412 	/*
413 	 * Let the pager know object is dead.
414 	 */
415 	if (object->pager != NULL)
416 		vm_pager_deallocate(object->pager);
417 
418 	simple_lock(&vm_object_list_lock);
419 	TAILQ_REMOVE(&vm_object_list, object, object_list);
420 	vm_object_count--;
421 	simple_unlock(&vm_object_list_lock);
422 
423 	/*
424 	 * Free the space for the object.
425 	 */
426 	free((caddr_t) object, M_VMOBJ);
427 }
428 
429 /*
430  *	vm_object_page_clean
431  *
432  *	Clean all dirty pages in the specified range of object.
433  *	Leaves page on whatever queue it is currently on.
434  *
435  *	Odd semantics: if start == end, we clean everything.
436  *
437  *	The object must be locked.
438  */
439 #if 1
440 boolean_t
441 vm_object_page_clean(object, start, end, syncio, de_queue)
442 	register vm_object_t object;
443 	register vm_offset_t start;
444 	register vm_offset_t end;
445 	boolean_t syncio;
446 	boolean_t de_queue;
447 {
448 	register vm_page_t p, nextp;
449 	int size;
450 
451 	if (object->pager == NULL)
452 		return 1;
453 
454 	if (start != end) {
455 		start = trunc_page(start);
456 		end = round_page(end);
457 	}
458 	size = end - start;
459 
460 again:
461 	/*
462 	 * Wait until the pageout daemon is through with the object.
463 	 */
464 	while (object->paging_in_progress) {
465 		tsleep(object, PVM, "objpcw", 0);
466 	}
467 
468 	nextp = object->memq.tqh_first;
469 	while ((p = nextp) && ((start == end) || (size != 0))) {
470 		nextp = p->listq.tqe_next;
471 		if (start == end || (p->offset >= start && p->offset < end)) {
472 			if ((p->flags & PG_BUSY) || p->busy) {
473 				int s = splhigh();
474 
475 				p->flags |= PG_WANTED;
476 				tsleep(p, PVM, "objpcn", 0);
477 				splx(s);
478 				goto again;
479 			}
480 			size -= PAGE_SIZE;
481 
482 			vm_page_test_dirty(p);
483 
484 			if ((p->dirty & p->valid) != 0) {
485 				vm_pageout_clean(p, VM_PAGEOUT_FORCE);
486 				goto again;
487 			}
488 		}
489 	}
490 	wakeup((caddr_t) object);
491 	return 1;
492 }
493 #endif
494 /*
495  *	vm_object_page_clean
496  *
497  *	Clean all dirty pages in the specified range of object.
498  *	If syncio is TRUE, page cleaning is done synchronously.
499  *	If de_queue is TRUE, pages are removed from any paging queue
500  *	they were on, otherwise they are left on whatever queue they
501  *	were on before the cleaning operation began.
502  *
503  *	Odd semantics: if start == end, we clean everything.
504  *
505  *	The object must be locked.
506  *
507  *	Returns TRUE if all was well, FALSE if there was a pager error
508  *	somewhere.  We attempt to clean (and dequeue) all pages regardless
509  *	of where an error occurs.
510  */
511 #if 0
512 boolean_t
513 vm_object_page_clean(object, start, end, syncio, de_queue)
514 	register vm_object_t object;
515 	register vm_offset_t start;
516 	register vm_offset_t end;
517 	boolean_t syncio;
518 	boolean_t de_queue;
519 {
520 	register vm_page_t p;
521 	int onqueue;
522 	boolean_t noerror = TRUE;
523 
524 	if (object == NULL)
525 		return (TRUE);
526 
527 	/*
528 	 * If it is an internal object and there is no pager, attempt to
529 	 * allocate one.  Note that vm_object_collapse may relocate one from a
530 	 * collapsed object so we must recheck afterward.
531 	 */
532 	if ((object->flags & OBJ_INTERNAL) && object->pager == NULL) {
533 		vm_object_collapse(object);
534 		if (object->pager == NULL) {
535 			vm_pager_t pager;
536 
537 			vm_object_unlock(object);
538 			pager = vm_pager_allocate(PG_DFLT, (caddr_t) 0,
539 			    object->size, VM_PROT_ALL,
540 			    (vm_offset_t) 0);
541 			if (pager)
542 				vm_object_setpager(object, pager, 0, FALSE);
543 			vm_object_lock(object);
544 		}
545 	}
546 	if (object->pager == NULL)
547 		return (FALSE);
548 
549 again:
550 	/*
551 	 * Wait until the pageout daemon is through with the object.
552 	 */
553 	while (object->paging_in_progress) {
554 		vm_object_sleep((int) object, object, FALSE);
555 		vm_object_lock(object);
556 	}
557 	/*
558 	 * Loop through the object page list cleaning as necessary.
559 	 */
560 	for (p = object->memq.tqh_first; p != NULL; p = p->listq.tqe_next) {
561 		onqueue = 0;
562 		if ((start == end || p->offset >= start && p->offset < end) &&
563 		    !(p->flags & PG_FICTITIOUS)) {
564 			vm_page_test_dirty(p);
565 			/*
566 			 * Remove the page from any paging queue. This needs
567 			 * to be done if either we have been explicitly asked
568 			 * to do so or it is about to be cleaned (see comment
569 			 * below).
570 			 */
571 			if (de_queue || (p->dirty & p->valid)) {
572 				vm_page_lock_queues();
573 				if (p->flags & PG_ACTIVE) {
574 					TAILQ_REMOVE(&vm_page_queue_active,
575 					    p, pageq);
576 					p->flags &= ~PG_ACTIVE;
577 					cnt.v_active_count--;
578 					onqueue = 1;
579 				} else if (p->flags & PG_INACTIVE) {
580 					TAILQ_REMOVE(&vm_page_queue_inactive,
581 					    p, pageq);
582 					p->flags &= ~PG_INACTIVE;
583 					cnt.v_inactive_count--;
584 					onqueue = -1;
585 				} else
586 					onqueue = 0;
587 				vm_page_unlock_queues();
588 			}
589 			/*
590 			 * To ensure the state of the page doesn't change
591 			 * during the clean operation we do two things. First
592 			 * we set the busy bit and write-protect all mappings
593 			 * to ensure that write accesses to the page block (in
594 			 * vm_fault).  Second, we remove the page from any
595 			 * paging queue to foil the pageout daemon
596 			 * (vm_pageout_scan).
597 			 */
598 			pmap_page_protect(VM_PAGE_TO_PHYS(p), VM_PROT_READ);
599 			if (p->dirty & p->valid) {
600 				p->flags |= PG_BUSY;
601 				object->paging_in_progress++;
602 				vm_object_unlock(object);
603 				/*
604 				 * XXX if put fails we mark the page as clean
605 				 * to avoid an infinite loop. Will loose
606 				 * changes to the page.
607 				 */
608 				if (vm_pager_put(object->pager, p, syncio)) {
609 					printf("%s: pager_put error\n",
610 					    "vm_object_page_clean");
611 					p->dirty = 0;
612 					noerror = FALSE;
613 				}
614 				vm_object_lock(object);
615 				object->paging_in_progress--;
616 				if (!de_queue && onqueue) {
617 					vm_page_lock_queues();
618 					if (onqueue > 0)
619 						vm_page_activate(p);
620 					else
621 						vm_page_deactivate(p);
622 					vm_page_unlock_queues();
623 				}
624 				PAGE_WAKEUP(p);
625 				goto again;
626 			}
627 		}
628 	}
629 	return (noerror);
630 }
631 #endif
632 
633 /*
634  *	vm_object_deactivate_pages
635  *
636  *	Deactivate all pages in the specified object.  (Keep its pages
637  *	in memory even though it is no longer referenced.)
638  *
639  *	The object must be locked.
640  */
641 void
642 vm_object_deactivate_pages(object)
643 	register vm_object_t object;
644 {
645 	register vm_page_t p, next;
646 
647 	for (p = object->memq.tqh_first; p != NULL; p = next) {
648 		next = p->listq.tqe_next;
649 		vm_page_lock_queues();
650 		vm_page_deactivate(p);
651 		vm_page_unlock_queues();
652 	}
653 }
654 
655 /*
656  *	Trim the object cache to size.
657  */
658 void
659 vm_object_cache_trim()
660 {
661 	register vm_object_t object;
662 
663 	vm_object_cache_lock();
664 	while (vm_object_cached > vm_object_cache_max) {
665 		object = vm_object_cached_list.tqh_first;
666 		vm_object_cache_unlock();
667 
668 		if (object != vm_object_lookup(object->pager))
669 			panic("vm_object_cache_trim: I'm sooo confused.");
670 
671 		pager_cache(object, FALSE);
672 
673 		vm_object_cache_lock();
674 	}
675 	vm_object_cache_unlock();
676 }
677 
678 
679 /*
680  *	vm_object_pmap_copy:
681  *
682  *	Makes all physical pages in the specified
683  *	object range copy-on-write.  No writeable
684  *	references to these pages should remain.
685  *
686  *	The object must *not* be locked.
687  */
688 void
689 vm_object_pmap_copy(object, start, end)
690 	register vm_object_t object;
691 	register vm_offset_t start;
692 	register vm_offset_t end;
693 {
694 	register vm_page_t p;
695 
696 	if (object == NULL)
697 		return;
698 
699 	vm_object_lock(object);
700 	for (p = object->memq.tqh_first; p != NULL; p = p->listq.tqe_next) {
701 		if ((start <= p->offset) && (p->offset < end)) {
702 			pmap_page_protect(VM_PAGE_TO_PHYS(p), VM_PROT_READ);
703 			p->flags |= PG_COPYONWRITE;
704 		}
705 	}
706 	vm_object_unlock(object);
707 }
708 
709 /*
710  *	vm_object_pmap_remove:
711  *
712  *	Removes all physical pages in the specified
713  *	object range from all physical maps.
714  *
715  *	The object must *not* be locked.
716  */
717 void
718 vm_object_pmap_remove(object, start, end)
719 	register vm_object_t object;
720 	register vm_offset_t start;
721 	register vm_offset_t end;
722 {
723 	register vm_page_t p;
724 	int s;
725 
726 	if (object == NULL)
727 		return;
728 	++object->paging_in_progress;
729 
730 	vm_object_lock(object);
731 again:
732 	for (p = object->memq.tqh_first; p != NULL; p = p->listq.tqe_next) {
733 		if ((start <= p->offset) && (p->offset < end)) {
734 			s = splhigh();
735 			if ((p->flags & PG_BUSY) || p->busy) {
736 				p->flags |= PG_WANTED;
737 				tsleep((caddr_t) p, PVM, "vmopmr", 0);
738 				splx(s);
739 				goto again;
740 			}
741 			splx(s);
742 			pmap_page_protect(VM_PAGE_TO_PHYS(p), VM_PROT_NONE);
743 		}
744 	}
745 	vm_object_unlock(object);
746 	--object->paging_in_progress;
747 	if (object->paging_in_progress == 0)
748 		wakeup((caddr_t) object);
749 }
750 
751 /*
752  *	vm_object_copy:
753  *
754  *	Create a new object which is a copy of an existing
755  *	object, and mark all of the pages in the existing
756  *	object 'copy-on-write'.  The new object has one reference.
757  *	Returns the new object.
758  *
759  *	May defer the copy until later if the object is not backed
760  *	up by a non-default pager.
761  */
762 void
763 vm_object_copy(src_object, src_offset, size,
764     dst_object, dst_offset, src_needs_copy)
765 	register vm_object_t src_object;
766 	vm_offset_t src_offset;
767 	vm_size_t size;
768 	vm_object_t *dst_object;/* OUT */
769 	vm_offset_t *dst_offset;/* OUT */
770 	boolean_t *src_needs_copy;	/* OUT */
771 {
772 	register vm_object_t new_copy;
773 	register vm_object_t old_copy;
774 	vm_offset_t new_start, new_end;
775 
776 	register vm_page_t p;
777 
778 	if (src_object == NULL) {
779 		/*
780 		 * Nothing to copy
781 		 */
782 		*dst_object = NULL;
783 		*dst_offset = 0;
784 		*src_needs_copy = FALSE;
785 		return;
786 	}
787 	/*
788 	 * If the object's pager is null_pager or the default pager, we don't
789 	 * have to make a copy of it.  Instead, we set the needs copy flag and
790 	 * make a shadow later.
791 	 */
792 
793 	vm_object_lock(src_object);
794 
795 	/*
796 	 * Try to collapse the object before copying it.
797 	 */
798 
799 	vm_object_collapse(src_object);
800 
801 	if (src_object->pager == NULL ||
802 	    src_object->pager->pg_type == PG_SWAP ||
803 	    (src_object->flags & OBJ_INTERNAL)) {
804 
805 		/*
806 		 * Make another reference to the object
807 		 */
808 		src_object->ref_count++;
809 
810 		/*
811 		 * Mark all of the pages copy-on-write.
812 		 */
813 		for (p = src_object->memq.tqh_first; p; p = p->listq.tqe_next)
814 			if (src_offset <= p->offset &&
815 			    p->offset < src_offset + size)
816 				p->flags |= PG_COPYONWRITE;
817 		vm_object_unlock(src_object);
818 
819 		*dst_object = src_object;
820 		*dst_offset = src_offset;
821 
822 		/*
823 		 * Must make a shadow when write is desired
824 		 */
825 		*src_needs_copy = TRUE;
826 		return;
827 	}
828 	/*
829 	 * If the object has a pager, the pager wants to see all of the
830 	 * changes.  We need a copy-object for the changed pages.
831 	 *
832 	 * If there is a copy-object, and it is empty, no changes have been made
833 	 * to the object since the copy-object was made.  We can use the same
834 	 * copy- object.
835 	 */
836 
837 Retry1:
838 	old_copy = src_object->copy;
839 	if (old_copy != NULL) {
840 		/*
841 		 * Try to get the locks (out of order)
842 		 */
843 		if (!vm_object_lock_try(old_copy)) {
844 			vm_object_unlock(src_object);
845 
846 			/* should spin a bit here... */
847 			tsleep((caddr_t) old_copy, PVM, "cpylck", 1);
848 			vm_object_lock(src_object);
849 			goto Retry1;
850 		}
851 		if (old_copy->resident_page_count == 0 &&
852 		    old_copy->pager == NULL) {
853 			/*
854 			 * Return another reference to the existing
855 			 * copy-object.
856 			 */
857 			old_copy->ref_count++;
858 			vm_object_unlock(old_copy);
859 			vm_object_unlock(src_object);
860 			*dst_object = old_copy;
861 			*dst_offset = src_offset;
862 			*src_needs_copy = FALSE;
863 			return;
864 		}
865 		vm_object_unlock(old_copy);
866 	}
867 	vm_object_unlock(src_object);
868 
869 	/*
870 	 * If the object has a pager, the pager wants to see all of the
871 	 * changes.  We must make a copy-object and put the changed pages
872 	 * there.
873 	 *
874 	 * The copy-object is always made large enough to completely shadow the
875 	 * original object, since it may have several users who want to shadow
876 	 * the original object at different points.
877 	 */
878 
879 	new_copy = vm_object_allocate(src_object->size);
880 
881 Retry2:
882 	vm_object_lock(src_object);
883 	/*
884 	 * Copy object may have changed while we were unlocked
885 	 */
886 	old_copy = src_object->copy;
887 	if (old_copy != NULL) {
888 		/*
889 		 * Try to get the locks (out of order)
890 		 */
891 		if (!vm_object_lock_try(old_copy)) {
892 			vm_object_unlock(src_object);
893 			tsleep((caddr_t) old_copy, PVM, "cpylck", 1);
894 			goto Retry2;
895 		}
896 		/*
897 		 * Consistency check
898 		 */
899 		if (old_copy->shadow != src_object ||
900 		    old_copy->shadow_offset != (vm_offset_t) 0)
901 			panic("vm_object_copy: copy/shadow inconsistency");
902 
903 		/*
904 		 * Make the old copy-object shadow the new one. It will
905 		 * receive no more pages from the original object.
906 		 */
907 
908 		src_object->ref_count--;	/* remove ref. from old_copy */
909 		if (old_copy->shadow)
910 			TAILQ_REMOVE(&old_copy->shadow->reverse_shadow_head, old_copy, reverse_shadow_list);
911 		old_copy->shadow = new_copy;
912 		TAILQ_INSERT_TAIL(&old_copy->shadow->reverse_shadow_head, old_copy, reverse_shadow_list);
913 		new_copy->ref_count++;	/* locking not needed - we have the
914 					 * only pointer */
915 		vm_object_unlock(old_copy);	/* done with old_copy */
916 	}
917 	new_start = (vm_offset_t) 0;	/* always shadow original at 0 */
918 	new_end = (vm_offset_t) new_copy->size;	/* for the whole object */
919 
920 	/*
921 	 * Point the new copy at the existing object.
922 	 */
923 
924 	new_copy->shadow = src_object;
925 	TAILQ_INSERT_TAIL(&new_copy->shadow->reverse_shadow_head, new_copy, reverse_shadow_list);
926 	new_copy->shadow_offset = new_start;
927 	src_object->ref_count++;
928 	src_object->copy = new_copy;
929 
930 	/*
931 	 * Mark all the affected pages of the existing object copy-on-write.
932 	 */
933 	for (p = src_object->memq.tqh_first; p != NULL; p = p->listq.tqe_next)
934 		if ((new_start <= p->offset) && (p->offset < new_end))
935 			p->flags |= PG_COPYONWRITE;
936 
937 	vm_object_unlock(src_object);
938 
939 	*dst_object = new_copy;
940 	*dst_offset = src_offset - new_start;
941 	*src_needs_copy = FALSE;
942 }
943 
944 /*
945  *	vm_object_shadow:
946  *
947  *	Create a new object which is backed by the
948  *	specified existing object range.  The source
949  *	object reference is deallocated.
950  *
951  *	The new object and offset into that object
952  *	are returned in the source parameters.
953  */
954 
955 void
956 vm_object_shadow(object, offset, length)
957 	vm_object_t *object;	/* IN/OUT */
958 	vm_offset_t *offset;	/* IN/OUT */
959 	vm_size_t length;
960 {
961 	register vm_object_t source;
962 	register vm_object_t result;
963 
964 	source = *object;
965 
966 	/*
967 	 * Allocate a new object with the given length
968 	 */
969 
970 	if ((result = vm_object_allocate(length)) == NULL)
971 		panic("vm_object_shadow: no object for shadowing");
972 
973 	/*
974 	 * The new object shadows the source object, adding a reference to it.
975 	 * Our caller changes his reference to point to the new object,
976 	 * removing a reference to the source object.  Net result: no change
977 	 * of reference count.
978 	 */
979 	result->shadow = source;
980 	if (source)
981 		TAILQ_INSERT_TAIL(&result->shadow->reverse_shadow_head, result, reverse_shadow_list);
982 
983 	/*
984 	 * Store the offset into the source object, and fix up the offset into
985 	 * the new object.
986 	 */
987 
988 	result->shadow_offset = *offset;
989 
990 	/*
991 	 * Return the new things
992 	 */
993 
994 	*offset = 0;
995 	*object = result;
996 }
997 
998 /*
999  *	Set the specified object's pager to the specified pager.
1000  */
1001 
1002 void
1003 vm_object_setpager(object, pager, paging_offset,
1004     read_only)
1005 	vm_object_t object;
1006 	vm_pager_t pager;
1007 	vm_offset_t paging_offset;
1008 	boolean_t read_only;
1009 {
1010 	vm_object_lock(object);	/* XXX ? */
1011 	if (object->pager && object->pager != pager) {
1012 		panic("!!!pager already allocated!!!\n");
1013 	}
1014 	object->pager = pager;
1015 	object->paging_offset = paging_offset;
1016 	vm_object_unlock(object);	/* XXX ? */
1017 }
1018 
1019 /*
1020  *	vm_object_hash hashes the pager/id pair.
1021  */
1022 
1023 #define vm_object_hash(pager) \
1024 	(((unsigned)pager >> 5)%VM_OBJECT_HASH_COUNT)
1025 
1026 /*
1027  *	vm_object_lookup looks in the object cache for an object with the
1028  *	specified pager and paging id.
1029  */
1030 
1031 vm_object_t
1032 vm_object_lookup(pager)
1033 	vm_pager_t pager;
1034 {
1035 	register vm_object_hash_entry_t entry;
1036 	vm_object_t object;
1037 
1038 	cnt.v_lookups++;
1039 	vm_object_cache_lock();
1040 
1041 	for (entry = vm_object_hashtable[vm_object_hash(pager)].tqh_first;
1042 	    entry != NULL;
1043 	    entry = entry->hash_links.tqe_next) {
1044 		object = entry->object;
1045 		if (object->pager == pager) {
1046 			vm_object_lock(object);
1047 			if (object->ref_count == 0) {
1048 				TAILQ_REMOVE(&vm_object_cached_list, object,
1049 				    cached_list);
1050 				vm_object_cached--;
1051 			}
1052 			object->ref_count++;
1053 			vm_object_unlock(object);
1054 			vm_object_cache_unlock();
1055 			cnt.v_hits++;
1056 			return (object);
1057 		}
1058 	}
1059 
1060 	vm_object_cache_unlock();
1061 	return (NULL);
1062 }
1063 
1064 /*
1065  *	vm_object_enter enters the specified object/pager/id into
1066  *	the hash table.
1067  */
1068 
1069 void
1070 vm_object_enter(object, pager)
1071 	vm_object_t object;
1072 	vm_pager_t pager;
1073 {
1074 	struct vm_object_hash_head *bucket;
1075 	register vm_object_hash_entry_t entry;
1076 
1077 	/*
1078 	 * We don't cache null objects, and we can't cache objects with the
1079 	 * null pager.
1080 	 */
1081 
1082 	if (object == NULL)
1083 		return;
1084 	if (pager == NULL)
1085 		return;
1086 
1087 	bucket = &vm_object_hashtable[vm_object_hash(pager)];
1088 	entry = (vm_object_hash_entry_t)
1089 	    malloc((u_long) sizeof *entry, M_VMOBJHASH, M_WAITOK);
1090 	entry->object = object;
1091 	object->flags |= OBJ_CANPERSIST;
1092 
1093 	vm_object_cache_lock();
1094 	TAILQ_INSERT_TAIL(bucket, entry, hash_links);
1095 	vm_object_cache_unlock();
1096 }
1097 
1098 /*
1099  *	vm_object_remove:
1100  *
1101  *	Remove the pager from the hash table.
1102  *	Note:  This assumes that the object cache
1103  *	is locked.  XXX this should be fixed
1104  *	by reorganizing vm_object_deallocate.
1105  */
1106 void
1107 vm_object_remove(pager)
1108 	register vm_pager_t pager;
1109 {
1110 	struct vm_object_hash_head *bucket;
1111 	register vm_object_hash_entry_t entry;
1112 	register vm_object_t object;
1113 
1114 	bucket = &vm_object_hashtable[vm_object_hash(pager)];
1115 
1116 	for (entry = bucket->tqh_first;
1117 	    entry != NULL;
1118 	    entry = entry->hash_links.tqe_next) {
1119 		object = entry->object;
1120 		if (object->pager == pager) {
1121 			TAILQ_REMOVE(bucket, entry, hash_links);
1122 			free((caddr_t) entry, M_VMOBJHASH);
1123 			break;
1124 		}
1125 	}
1126 }
1127 
1128 static void
1129 vm_object_rcollapse(object, sobject)
1130 	register vm_object_t object, sobject;
1131 {
1132 	register vm_object_t backing_object;
1133 	register vm_offset_t backing_offset, new_offset;
1134 	register vm_page_t p, pp;
1135 	register vm_size_t size;
1136 	int s;
1137 
1138 	if (!object)
1139 		return;
1140 	backing_object = object->shadow;
1141 	if (backing_object != sobject) {
1142 		printf("backing obj != sobject!!!\n");
1143 		return;
1144 	}
1145 	if (!backing_object)
1146 		return;
1147 	if ((backing_object->flags & OBJ_INTERNAL) == 0)
1148 		return;
1149 	if (backing_object->shadow != NULL &&
1150 	    backing_object->shadow->copy == backing_object)
1151 		return;
1152 	if (backing_object->ref_count != 1)
1153 		return;
1154 
1155 	backing_object->ref_count += 2;
1156 	s = splbio();
1157 	while (backing_object->paging_in_progress) {
1158 		tsleep(backing_object, PVM, "rcolow", 0);
1159 	}
1160 	splx(s);
1161 
1162 	backing_offset = object->shadow_offset;
1163 	size = object->size;
1164 	p = backing_object->memq.tqh_first;
1165 	while (p) {
1166 		vm_page_t next;
1167 
1168 		next = p->listq.tqe_next;
1169 		pmap_page_protect(VM_PAGE_TO_PHYS(p), VM_PROT_NONE);
1170 		new_offset = (p->offset - backing_offset);
1171 		if (p->offset < backing_offset ||
1172 		    new_offset >= size) {
1173 			if (backing_object->pager)
1174 				swap_pager_freespace(backing_object->pager,
1175 				    backing_object->paging_offset + p->offset, PAGE_SIZE);
1176 			vm_page_lock_queues();
1177 			vm_page_free(p);
1178 			vm_page_unlock_queues();
1179 		} else {
1180 			pp = vm_page_lookup(object, new_offset);
1181 			if (pp != NULL ||
1182 			    (object->pager &&
1183 			    vm_pager_has_page(object->pager, object->paging_offset + new_offset))) {
1184 				if (backing_object->pager)
1185 					swap_pager_freespace(backing_object->pager,
1186 					    backing_object->paging_offset + p->offset, PAGE_SIZE);
1187 				vm_page_lock_queues();
1188 				vm_page_free(p);
1189 				vm_page_unlock_queues();
1190 			} else {
1191 				if (!backing_object->pager ||
1192 				    !vm_pager_has_page(backing_object->pager, backing_object->paging_offset + p->offset))
1193 					vm_page_rename(p, object, new_offset);
1194 			}
1195 		}
1196 		p = next;
1197 	}
1198 	backing_object->ref_count -= 2;
1199 }
1200 
1201 /*
1202  * this version of collapse allows the operation to occur earlier and
1203  * when paging_in_progress is true for an object...  This is not a complete
1204  * operation, but should plug 99.9% of the rest of the leaks.
1205  */
1206 static void
1207 vm_object_qcollapse(object)
1208 	register vm_object_t object;
1209 {
1210 	register vm_object_t backing_object;
1211 	register vm_offset_t backing_offset, new_offset;
1212 	register vm_page_t p, pp;
1213 	register vm_size_t size;
1214 
1215 	backing_object = object->shadow;
1216 	if (!backing_object)
1217 		return;
1218 	if ((backing_object->flags & OBJ_INTERNAL) == 0)
1219 		return;
1220 	if (backing_object->shadow != NULL &&
1221 	    backing_object->shadow->copy == backing_object)
1222 		return;
1223 	if (backing_object->ref_count != 1)
1224 		return;
1225 
1226 	backing_object->ref_count += 2;
1227 
1228 	backing_offset = object->shadow_offset;
1229 	size = object->size;
1230 	p = backing_object->memq.tqh_first;
1231 	while (p) {
1232 		vm_page_t next;
1233 
1234 		next = p->listq.tqe_next;
1235 		if ((p->flags & (PG_BUSY | PG_FICTITIOUS | PG_CACHE)) ||
1236 		    !p->valid || p->hold_count || p->wire_count || p->busy || p->bmapped) {
1237 			p = next;
1238 			continue;
1239 		}
1240 		pmap_page_protect(VM_PAGE_TO_PHYS(p), VM_PROT_NONE);
1241 		new_offset = (p->offset - backing_offset);
1242 		if (p->offset < backing_offset ||
1243 		    new_offset >= size) {
1244 			if (backing_object->pager)
1245 				swap_pager_freespace(backing_object->pager,
1246 				    backing_object->paging_offset + p->offset, PAGE_SIZE);
1247 			vm_page_lock_queues();
1248 			vm_page_free(p);
1249 			vm_page_unlock_queues();
1250 		} else {
1251 			pp = vm_page_lookup(object, new_offset);
1252 			if (pp != NULL || (object->pager && vm_pager_has_page(object->pager,
1253 				    object->paging_offset + new_offset))) {
1254 				if (backing_object->pager)
1255 					swap_pager_freespace(backing_object->pager,
1256 					    backing_object->paging_offset + p->offset, PAGE_SIZE);
1257 				vm_page_lock_queues();
1258 				vm_page_free(p);
1259 				vm_page_unlock_queues();
1260 			} else {
1261 				if (!backing_object->pager ||
1262 				    !vm_pager_has_page(backing_object->pager, backing_object->paging_offset + p->offset))
1263 					vm_page_rename(p, object, new_offset);
1264 			}
1265 		}
1266 		p = next;
1267 	}
1268 	backing_object->ref_count -= 2;
1269 }
1270 
1271 boolean_t vm_object_collapse_allowed = TRUE;
1272 
1273 /*
1274  *	vm_object_collapse:
1275  *
1276  *	Collapse an object with the object backing it.
1277  *	Pages in the backing object are moved into the
1278  *	parent, and the backing object is deallocated.
1279  *
1280  *	Requires that the object be locked and the page
1281  *	queues be unlocked.
1282  *
1283  *	This routine has significant changes by John S. Dyson
1284  *	to fix some swap memory leaks.  18 Dec 93
1285  *
1286  */
1287 void
1288 vm_object_collapse(object)
1289 	register vm_object_t object;
1290 
1291 {
1292 	register vm_object_t backing_object;
1293 	register vm_offset_t backing_offset;
1294 	register vm_size_t size;
1295 	register vm_offset_t new_offset;
1296 	register vm_page_t p, pp;
1297 
1298 	if (!vm_object_collapse_allowed)
1299 		return;
1300 
1301 	while (TRUE) {
1302 		/*
1303 		 * Verify that the conditions are right for collapse:
1304 		 *
1305 		 * The object exists and no pages in it are currently being paged
1306 		 * out.
1307 		 */
1308 		if (object == NULL)
1309 			return;
1310 		if (object->paging_in_progress != 0) {
1311 			if (object->shadow)
1312 				vm_object_qcollapse(object);
1313 			return;
1314 		}
1315 		/*
1316 		 * There is a backing object, and
1317 		 */
1318 
1319 		if ((backing_object = object->shadow) == NULL)
1320 			return;
1321 
1322 		vm_object_lock(backing_object);
1323 		/*
1324 		 * ... The backing object is not read_only, and no pages in
1325 		 * the backing object are currently being paged out. The
1326 		 * backing object is internal.
1327 		 */
1328 
1329 		if ((backing_object->flags & OBJ_INTERNAL) == 0 ||
1330 		    backing_object->paging_in_progress != 0) {
1331 			vm_object_unlock(backing_object);
1332 			vm_object_qcollapse(object);
1333 			return;
1334 		}
1335 		/*
1336 		 * The backing object can't be a copy-object: the
1337 		 * shadow_offset for the copy-object must stay as 0.
1338 		 * Furthermore (for the 'we have all the pages' case), if we
1339 		 * bypass backing_object and just shadow the next object in
1340 		 * the chain, old pages from that object would then have to be
1341 		 * copied BOTH into the (former) backing_object and into the
1342 		 * parent object.
1343 		 */
1344 		if (backing_object->shadow != NULL &&
1345 		    backing_object->shadow->copy == backing_object) {
1346 			vm_object_unlock(backing_object);
1347 			return;
1348 		}
1349 		/*
1350 		 * we can deal only with the swap pager
1351 		 */
1352 		if ((object->pager &&
1353 			object->pager->pg_type != PG_SWAP) ||
1354 		    (backing_object->pager &&
1355 			backing_object->pager->pg_type != PG_SWAP)) {
1356 			vm_object_unlock(backing_object);
1357 			return;
1358 		}
1359 		/*
1360 		 * We know that we can either collapse the backing object (if
1361 		 * the parent is the only reference to it) or (perhaps) remove
1362 		 * the parent's reference to it.
1363 		 */
1364 
1365 		backing_offset = object->shadow_offset;
1366 		size = object->size;
1367 
1368 		/*
1369 		 * If there is exactly one reference to the backing object, we
1370 		 * can collapse it into the parent.
1371 		 */
1372 
1373 		if (backing_object->ref_count == 1) {
1374 
1375 			/*
1376 			 * We can collapse the backing object.
1377 			 *
1378 			 * Move all in-memory pages from backing_object to the
1379 			 * parent.  Pages that have been paged out will be
1380 			 * overwritten by any of the parent's pages that
1381 			 * shadow them.
1382 			 */
1383 
1384 			while ((p = backing_object->memq.tqh_first) != 0) {
1385 
1386 				new_offset = (p->offset - backing_offset);
1387 
1388 				/*
1389 				 * If the parent has a page here, or if this
1390 				 * page falls outside the parent, dispose of
1391 				 * it.
1392 				 *
1393 				 * Otherwise, move it as planned.
1394 				 */
1395 
1396 				if (p->offset < backing_offset ||
1397 				    new_offset >= size) {
1398 					vm_page_lock_queues();
1399 					pmap_page_protect(VM_PAGE_TO_PHYS(p), VM_PROT_NONE);
1400 					PAGE_WAKEUP(p);
1401 					vm_page_free(p);
1402 					vm_page_unlock_queues();
1403 				} else {
1404 					pp = vm_page_lookup(object, new_offset);
1405 					if (pp != NULL || (object->pager && vm_pager_has_page(object->pager,
1406 					    object->paging_offset + new_offset))) {
1407 						vm_page_lock_queues();
1408 						pmap_page_protect(VM_PAGE_TO_PHYS(p), VM_PROT_NONE);
1409 						PAGE_WAKEUP(p);
1410 						vm_page_free(p);
1411 						vm_page_unlock_queues();
1412 					} else {
1413 						vm_page_rename(p, object, new_offset);
1414 					}
1415 				}
1416 			}
1417 
1418 			/*
1419 			 * Move the pager from backing_object to object.
1420 			 */
1421 
1422 			if (backing_object->pager) {
1423 				backing_object->paging_in_progress++;
1424 				if (object->pager) {
1425 					vm_pager_t bopager;
1426 
1427 					object->paging_in_progress++;
1428 					/*
1429 					 * copy shadow object pages into ours
1430 					 * and destroy unneeded pages in
1431 					 * shadow object.
1432 					 */
1433 					bopager = backing_object->pager;
1434 					backing_object->pager = NULL;
1435 					vm_object_remove(backing_object->pager);
1436 					swap_pager_copy(
1437 					    bopager, backing_object->paging_offset,
1438 					    object->pager, object->paging_offset,
1439 					    object->shadow_offset);
1440 					object->paging_in_progress--;
1441 					if (object->paging_in_progress == 0)
1442 						wakeup((caddr_t) object);
1443 				} else {
1444 					object->paging_in_progress++;
1445 					/*
1446 					 * grab the shadow objects pager
1447 					 */
1448 					object->pager = backing_object->pager;
1449 					object->paging_offset = backing_object->paging_offset + backing_offset;
1450 					vm_object_remove(backing_object->pager);
1451 					backing_object->pager = NULL;
1452 					/*
1453 					 * free unnecessary blocks
1454 					 */
1455 					swap_pager_freespace(object->pager, 0, object->paging_offset);
1456 					object->paging_in_progress--;
1457 					if (object->paging_in_progress == 0)
1458 						wakeup((caddr_t) object);
1459 				}
1460 				backing_object->paging_in_progress--;
1461 				if (backing_object->paging_in_progress == 0)
1462 					wakeup((caddr_t) backing_object);
1463 			}
1464 			/*
1465 			 * Object now shadows whatever backing_object did.
1466 			 * Note that the reference to backing_object->shadow
1467 			 * moves from within backing_object to within object.
1468 			 */
1469 
1470 			TAILQ_REMOVE(&object->shadow->reverse_shadow_head, object,
1471 			    reverse_shadow_list);
1472 			if (backing_object->shadow)
1473 				TAILQ_REMOVE(&backing_object->shadow->reverse_shadow_head,
1474 				    backing_object, reverse_shadow_list);
1475 			object->shadow = backing_object->shadow;
1476 			if (object->shadow)
1477 				TAILQ_INSERT_TAIL(&object->shadow->reverse_shadow_head,
1478 				    object, reverse_shadow_list);
1479 
1480 			object->shadow_offset += backing_object->shadow_offset;
1481 			if (object->shadow != NULL &&
1482 			    object->shadow->copy != NULL) {
1483 				panic("vm_object_collapse: we collapsed a copy-object!");
1484 			}
1485 			/*
1486 			 * Discard backing_object.
1487 			 *
1488 			 * Since the backing object has no pages, no pager left,
1489 			 * and no object references within it, all that is
1490 			 * necessary is to dispose of it.
1491 			 */
1492 
1493 			vm_object_unlock(backing_object);
1494 
1495 			simple_lock(&vm_object_list_lock);
1496 			TAILQ_REMOVE(&vm_object_list, backing_object,
1497 			    object_list);
1498 			vm_object_count--;
1499 			simple_unlock(&vm_object_list_lock);
1500 
1501 			free((caddr_t) backing_object, M_VMOBJ);
1502 
1503 			object_collapses++;
1504 		} else {
1505 			/*
1506 			 * If all of the pages in the backing object are
1507 			 * shadowed by the parent object, the parent object no
1508 			 * longer has to shadow the backing object; it can
1509 			 * shadow the next one in the chain.
1510 			 *
1511 			 * The backing object must not be paged out - we'd have
1512 			 * to check all of the paged-out pages, as well.
1513 			 */
1514 
1515 			if (backing_object->pager != NULL) {
1516 				vm_object_unlock(backing_object);
1517 				return;
1518 			}
1519 			/*
1520 			 * Should have a check for a 'small' number of pages
1521 			 * here.
1522 			 */
1523 
1524 			for (p = backing_object->memq.tqh_first; p; p = p->listq.tqe_next) {
1525 				new_offset = (p->offset - backing_offset);
1526 
1527 				/*
1528 				 * If the parent has a page here, or if this
1529 				 * page falls outside the parent, keep going.
1530 				 *
1531 				 * Otherwise, the backing_object must be left in
1532 				 * the chain.
1533 				 */
1534 
1535 				if (p->offset >= backing_offset &&
1536 				    new_offset <= size &&
1537 				    ((pp = vm_page_lookup(object, new_offset)) == NULL ||
1538 					!pp->valid) &&
1539 				    (!object->pager || !vm_pager_has_page(object->pager, object->paging_offset + new_offset))) {
1540 					/*
1541 					 * Page still needed. Can't go any
1542 					 * further.
1543 					 */
1544 					vm_object_unlock(backing_object);
1545 					return;
1546 				}
1547 			}
1548 
1549 			/*
1550 			 * Make the parent shadow the next object in the
1551 			 * chain.  Deallocating backing_object will not remove
1552 			 * it, since its reference count is at least 2.
1553 			 */
1554 
1555 			TAILQ_REMOVE(&object->shadow->reverse_shadow_head,
1556 			    object, reverse_shadow_list);
1557 			vm_object_reference(object->shadow = backing_object->shadow);
1558 			if (object->shadow)
1559 				TAILQ_INSERT_TAIL(&object->shadow->reverse_shadow_head,
1560 				    object, reverse_shadow_list);
1561 			object->shadow_offset += backing_object->shadow_offset;
1562 
1563 			/*
1564 			 * Backing object might have had a copy pointer to us.
1565 			 * If it did, clear it.
1566 			 */
1567 			if (backing_object->copy == object) {
1568 				backing_object->copy = NULL;
1569 			}
1570 			/*
1571 			 * Drop the reference count on backing_object. Since
1572 			 * its ref_count was at least 2, it will not vanish;
1573 			 * so we don't need to call vm_object_deallocate.
1574 			 */
1575 			if (backing_object->ref_count == 1)
1576 				printf("should have called obj deallocate\n");
1577 			backing_object->ref_count--;
1578 			vm_object_unlock(backing_object);
1579 
1580 			object_bypasses++;
1581 
1582 		}
1583 
1584 		/*
1585 		 * Try again with this object's new backing object.
1586 		 */
1587 	}
1588 }
1589 
1590 /*
1591  *	vm_object_page_remove: [internal]
1592  *
1593  *	Removes all physical pages in the specified
1594  *	object range from the object's list of pages.
1595  *
1596  *	The object must be locked.
1597  */
1598 void
1599 vm_object_page_remove(object, start, end)
1600 	register vm_object_t object;
1601 	register vm_offset_t start;
1602 	register vm_offset_t end;
1603 {
1604 	register vm_page_t p, next;
1605 	vm_offset_t size;
1606 	int s;
1607 
1608 	if (object == NULL)
1609 		return;
1610 
1611 	object->paging_in_progress++;
1612 	start = trunc_page(start);
1613 	end = round_page(end);
1614 again:
1615 	size = end - start;
1616 	if (size > 4 * PAGE_SIZE || size >= object->size / 4) {
1617 		for (p = object->memq.tqh_first; p != NULL; p = next) {
1618 			next = p->listq.tqe_next;
1619 			if ((start <= p->offset) && (p->offset < end)) {
1620 				s = splhigh();
1621 				if (p->bmapped) {
1622 					splx(s);
1623 					continue;
1624 				}
1625 				if ((p->flags & PG_BUSY) || p->busy) {
1626 					p->flags |= PG_WANTED;
1627 					tsleep((caddr_t) p, PVM, "vmopar", 0);
1628 					splx(s);
1629 					goto again;
1630 				}
1631 				splx(s);
1632 				pmap_page_protect(VM_PAGE_TO_PHYS(p), VM_PROT_NONE);
1633 				vm_page_lock_queues();
1634 				PAGE_WAKEUP(p);
1635 				vm_page_free(p);
1636 				vm_page_unlock_queues();
1637 			}
1638 		}
1639 	} else {
1640 		while (size > 0) {
1641 			while ((p = vm_page_lookup(object, start)) != 0) {
1642 				s = splhigh();
1643 				if (p->bmapped) {
1644 					splx(s);
1645 					break;
1646 				}
1647 				if ((p->flags & PG_BUSY) || p->busy) {
1648 					p->flags |= PG_WANTED;
1649 					tsleep((caddr_t) p, PVM, "vmopar", 0);
1650 					splx(s);
1651 					goto again;
1652 				}
1653 				splx(s);
1654 				pmap_page_protect(VM_PAGE_TO_PHYS(p), VM_PROT_NONE);
1655 				vm_page_lock_queues();
1656 				PAGE_WAKEUP(p);
1657 				vm_page_free(p);
1658 				vm_page_unlock_queues();
1659 			}
1660 			start += PAGE_SIZE;
1661 			size -= PAGE_SIZE;
1662 		}
1663 	}
1664 	--object->paging_in_progress;
1665 	if (object->paging_in_progress == 0)
1666 		wakeup((caddr_t) object);
1667 }
1668 
1669 /*
1670  *	Routine:	vm_object_coalesce
1671  *	Function:	Coalesces two objects backing up adjoining
1672  *			regions of memory into a single object.
1673  *
1674  *	returns TRUE if objects were combined.
1675  *
1676  *	NOTE:	Only works at the moment if the second object is NULL -
1677  *		if it's not, which object do we lock first?
1678  *
1679  *	Parameters:
1680  *		prev_object	First object to coalesce
1681  *		prev_offset	Offset into prev_object
1682  *		next_object	Second object into coalesce
1683  *		next_offset	Offset into next_object
1684  *
1685  *		prev_size	Size of reference to prev_object
1686  *		next_size	Size of reference to next_object
1687  *
1688  *	Conditions:
1689  *	The object must *not* be locked.
1690  */
1691 boolean_t
1692 vm_object_coalesce(prev_object, next_object,
1693     prev_offset, next_offset,
1694     prev_size, next_size)
1695 	register vm_object_t prev_object;
1696 	vm_object_t next_object;
1697 	vm_offset_t prev_offset, next_offset;
1698 	vm_size_t prev_size, next_size;
1699 {
1700 	vm_size_t newsize;
1701 
1702 	if (next_object != NULL) {
1703 		return (FALSE);
1704 	}
1705 	if (prev_object == NULL) {
1706 		return (TRUE);
1707 	}
1708 	vm_object_lock(prev_object);
1709 
1710 	/*
1711 	 * Try to collapse the object first
1712 	 */
1713 	vm_object_collapse(prev_object);
1714 
1715 	/*
1716 	 * Can't coalesce if: . more than one reference . paged out . shadows
1717 	 * another object . has a copy elsewhere (any of which mean that the
1718 	 * pages not mapped to prev_entry may be in use anyway)
1719 	 */
1720 
1721 	if (prev_object->ref_count > 1 ||
1722 	    prev_object->pager != NULL ||
1723 	    prev_object->shadow != NULL ||
1724 	    prev_object->copy != NULL) {
1725 		vm_object_unlock(prev_object);
1726 		return (FALSE);
1727 	}
1728 	/*
1729 	 * Remove any pages that may still be in the object from a previous
1730 	 * deallocation.
1731 	 */
1732 
1733 	vm_object_page_remove(prev_object,
1734 	    prev_offset + prev_size,
1735 	    prev_offset + prev_size + next_size);
1736 
1737 	/*
1738 	 * Extend the object if necessary.
1739 	 */
1740 	newsize = prev_offset + prev_size + next_size;
1741 	if (newsize > prev_object->size)
1742 		prev_object->size = newsize;
1743 
1744 	vm_object_unlock(prev_object);
1745 	return (TRUE);
1746 }
1747 
1748 /*
1749  * returns page after looking up in shadow chain
1750  */
1751 
1752 vm_page_t
1753 vm_object_page_lookup(object, offset)
1754 	vm_object_t object;
1755 	vm_offset_t offset;
1756 {
1757 	vm_page_t m;
1758 
1759 	if (!(m = vm_page_lookup(object, offset))) {
1760 		if (!object->shadow)
1761 			return 0;
1762 		else
1763 			return vm_object_page_lookup(object->shadow, offset + object->shadow_offset);
1764 	}
1765 	return m;
1766 }
1767 
1768 #define DEBUG
1769 #if defined(DEBUG) || defined(DDB)
1770 /*
1771  *	vm_object_print:	[ debug ]
1772  */
1773 void
1774 vm_object_print(object, full)
1775 	vm_object_t object;
1776 	boolean_t full;
1777 {
1778 	register vm_page_t p;
1779 	extern indent;
1780 
1781 	register int count;
1782 
1783 	if (object == NULL)
1784 		return;
1785 
1786 	iprintf("Object 0x%x: size=0x%x, res=%d, ref=%d, ",
1787 	    (int) object, (int) object->size,
1788 	    object->resident_page_count, object->ref_count);
1789 	printf("pager=0x%x+0x%x, shadow=(0x%x)+0x%x\n",
1790 	    (int) object->pager, (int) object->paging_offset,
1791 	    (int) object->shadow, (int) object->shadow_offset);
1792 	printf("cache: next=%p, prev=%p\n",
1793 	    object->cached_list.tqe_next, object->cached_list.tqe_prev);
1794 
1795 	if (!full)
1796 		return;
1797 
1798 	indent += 2;
1799 	count = 0;
1800 	for (p = object->memq.tqh_first; p != NULL; p = p->listq.tqe_next) {
1801 		if (count == 0)
1802 			iprintf("memory:=");
1803 		else if (count == 6) {
1804 			printf("\n");
1805 			iprintf(" ...");
1806 			count = 0;
1807 		} else
1808 			printf(",");
1809 		count++;
1810 
1811 		printf("(off=0x%lx,page=0x%lx)",
1812 		    (u_long) p->offset, (u_long) VM_PAGE_TO_PHYS(p));
1813 	}
1814 	if (count != 0)
1815 		printf("\n");
1816 	indent -= 2;
1817 }
1818 #endif				/* defined(DEBUG) || defined(DDB) */
1819