xref: /freebsd/sys/vm/vm_object.c (revision 8e6b01171e30297084bb0b4457c4183c2746aacc)
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 1.53 1995/08/26 23:19:48 bde 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/vm_kern.h>
85 
86 static void _vm_object_allocate(objtype_t, vm_size_t, vm_object_t);
87 
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 int vm_object_cache_max;
116 struct object_q vm_object_cached_list;
117 int vm_object_cached;
118 struct object_q vm_object_list;
119 long vm_object_count;
120 vm_object_t kernel_object;
121 vm_object_t kmem_object;
122 struct vm_object kernel_object_store;
123 struct vm_object kmem_object_store;
124 
125 long object_collapses;
126 long object_bypasses;
127 
128 static void
129 _vm_object_allocate(type, size, object)
130 	objtype_t type;
131 	vm_size_t size;
132 	register vm_object_t object;
133 {
134 	TAILQ_INIT(&object->memq);
135 	TAILQ_INIT(&object->shadow_head);
136 
137 	object->type = type;
138 	object->size = size;
139 	object->ref_count = 1;
140 	object->flags = 0;
141 	object->paging_in_progress = 0;
142 	object->resident_page_count = 0;
143 	object->handle = NULL;
144 	object->paging_offset = 0;
145 	object->backing_object = NULL;
146 	object->backing_object_offset = (vm_offset_t) 0;
147 
148 	object->last_read = 0;
149 
150 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
151 	vm_object_count++;
152 }
153 
154 /*
155  *	vm_object_init:
156  *
157  *	Initialize the VM objects module.
158  */
159 void
160 vm_object_init(vm_offset_t nothing)
161 {
162 	register int i;
163 
164 	TAILQ_INIT(&vm_object_cached_list);
165 	TAILQ_INIT(&vm_object_list);
166 	vm_object_count = 0;
167 
168 	vm_object_cache_max = 84;
169 	if (cnt.v_page_count > 1000)
170 		vm_object_cache_max += (cnt.v_page_count - 1000) / 4;
171 
172 	kernel_object = &kernel_object_store;
173 	_vm_object_allocate(OBJT_DEFAULT, VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS,
174 	    kernel_object);
175 
176 	kmem_object = &kmem_object_store;
177 	_vm_object_allocate(OBJT_DEFAULT, VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS,
178 	    kmem_object);
179 }
180 
181 /*
182  *	vm_object_allocate:
183  *
184  *	Returns a new object with the given size.
185  */
186 
187 vm_object_t
188 vm_object_allocate(type, size)
189 	objtype_t type;
190 	vm_size_t size;
191 {
192 	register vm_object_t result;
193 
194 	result = (vm_object_t)
195 	    malloc((u_long) sizeof *result, M_VMOBJ, M_WAITOK);
196 
197 
198 	_vm_object_allocate(type, size, result);
199 
200 	return (result);
201 }
202 
203 
204 /*
205  *	vm_object_reference:
206  *
207  *	Gets another reference to the given object.
208  */
209 inline void
210 vm_object_reference(object)
211 	register vm_object_t object;
212 {
213 	if (object == NULL)
214 		return;
215 
216 	if (object->ref_count == 0) {
217 		if ((object->flags & OBJ_CANPERSIST) == 0)
218 			panic("vm_object_reference: non-persistent object with 0 ref_count");
219 		TAILQ_REMOVE(&vm_object_cached_list, object, cached_list);
220 		vm_object_cached--;
221 	}
222 	object->ref_count++;
223 }
224 
225 /*
226  *	vm_object_deallocate:
227  *
228  *	Release a reference to the specified object,
229  *	gained either through a vm_object_allocate
230  *	or a vm_object_reference call.  When all references
231  *	are gone, storage associated with this object
232  *	may be relinquished.
233  *
234  *	No object may be locked.
235  */
236 void
237 vm_object_deallocate(object)
238 	vm_object_t object;
239 {
240 	vm_object_t temp;
241 
242 	while (object != NULL) {
243 
244 		if (object->ref_count == 0)
245 			panic("vm_object_deallocate: object deallocated too many times");
246 
247 		/*
248 		 * Lose the reference
249 		 */
250 		object->ref_count--;
251 
252 		if (object->ref_count != 0) {
253 			if ((object->ref_count == 1) &&
254 			    (object->handle == NULL) &&
255 			    (object->type == OBJT_DEFAULT ||
256 			     object->type == OBJT_SWAP)) {
257 				vm_object_t robject;
258 				robject = object->shadow_head.tqh_first;
259 				if ((robject != NULL) &&
260 				    (robject->handle == NULL) &&
261 				    (robject->type == OBJT_DEFAULT ||
262 				     robject->type == OBJT_SWAP)) {
263 					int s;
264 					robject->ref_count += 2;
265 					object->ref_count += 2;
266 
267 					do {
268 						s = splhigh();
269 						while (robject->paging_in_progress) {
270 							robject->flags |= OBJ_PIPWNT;
271 							tsleep(robject, PVM, "objde1", 0);
272 						}
273 
274 						while (object->paging_in_progress) {
275 							object->flags |= OBJ_PIPWNT;
276 							tsleep(object, PVM, "objde2", 0);
277 						}
278 						splx(s);
279 
280 					} while( object->paging_in_progress || robject->paging_in_progress);
281 
282 					object->ref_count -= 2;
283 					robject->ref_count -= 2;
284 					if( robject->ref_count == 0) {
285 						robject->ref_count += 1;
286 						object = robject;
287 						continue;
288 					}
289 					vm_object_collapse(robject);
290 					return;
291 				}
292 			}
293 			/*
294 			 * If there are still references, then we are done.
295 			 */
296 			return;
297 		}
298 
299 		if (object->type == OBJT_VNODE) {
300 			struct vnode *vp = object->handle;
301 
302 			vp->v_flag &= ~VTEXT;
303 		}
304 
305 		/*
306 		 * See if this object can persist and has some resident
307 		 * pages.  If so, enter it in the cache.
308 		 */
309 		if (object->flags & OBJ_CANPERSIST) {
310 			if (object->resident_page_count != 0) {
311 				vm_object_page_clean(object, 0, 0 ,TRUE, TRUE);
312 				TAILQ_INSERT_TAIL(&vm_object_cached_list, object,
313 				    cached_list);
314 				vm_object_cached++;
315 
316 				vm_object_cache_trim();
317 				return;
318 			} else {
319 				object->flags &= ~OBJ_CANPERSIST;
320 			}
321 		}
322 
323 		/*
324 		 * Make sure no one uses us.
325 		 */
326 		object->flags |= OBJ_DEAD;
327 
328 		temp = object->backing_object;
329 		if (temp)
330 			TAILQ_REMOVE(&temp->shadow_head, object, shadow_list);
331 		vm_object_terminate(object);
332 		/* unlocks and deallocates object */
333 		object = temp;
334 	}
335 }
336 
337 /*
338  *	vm_object_terminate actually destroys the specified object, freeing
339  *	up all previously used resources.
340  *
341  *	The object must be locked.
342  */
343 void
344 vm_object_terminate(object)
345 	register vm_object_t object;
346 {
347 	register vm_page_t p, next;
348 	vm_object_t backing_object;
349 	int s;
350 
351 	/*
352 	 * wait for the pageout daemon to be done with the object
353 	 */
354 	s = splhigh();
355 	while (object->paging_in_progress) {
356 		object->flags |= OBJ_PIPWNT;
357 		tsleep(object, PVM, "objtrm", 0);
358 	}
359 	splx(s);
360 
361 	if (object->paging_in_progress != 0)
362 		panic("vm_object_deallocate: pageout in progress");
363 
364 	/*
365 	 * Clean and free the pages, as appropriate. All references to the
366 	 * object are gone, so we don't need to lock it.
367 	 */
368 	if (object->type == OBJT_VNODE) {
369 		struct vnode *vp = object->handle;
370 
371 		VOP_LOCK(vp);
372 		vm_object_page_clean(object, 0, 0, TRUE, FALSE);
373 		vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0);
374 		VOP_UNLOCK(vp);
375 	}
376 
377 	/*
378 	 * Now free the pages. For internal objects, this also removes them
379 	 * from paging queues.
380 	 */
381 	while ((p = object->memq.tqh_first) != NULL) {
382 		if (p->flags & PG_BUSY)
383 			printf("vm_object_terminate: freeing busy page\n");
384 		PAGE_WAKEUP(p);
385 		vm_page_free(p);
386 		cnt.v_pfree++;
387 	}
388 
389 	/*
390 	 * Let the pager know object is dead.
391 	 */
392 	vm_pager_deallocate(object);
393 
394 	TAILQ_REMOVE(&vm_object_list, object, object_list);
395 	vm_object_count--;
396 
397 	wakeup(object);
398 
399 	/*
400 	 * Free the space for the object.
401 	 */
402 	free((caddr_t) object, M_VMOBJ);
403 }
404 
405 /*
406  *	vm_object_page_clean
407  *
408  *	Clean all dirty pages in the specified range of object.
409  *	Leaves page on whatever queue it is currently on.
410  *
411  *	Odd semantics: if start == end, we clean everything.
412  *
413  *	The object must be locked.
414  */
415 
416 void
417 vm_object_page_clean(object, start, end, syncio, lockflag)
418 	vm_object_t object;
419 	vm_offset_t start;
420 	vm_offset_t end;
421 	boolean_t syncio;
422 	boolean_t lockflag;
423 {
424 	register vm_page_t p;
425 	register vm_offset_t tstart, tend;
426 	int pass;
427 	int pgcount, s;
428 	int allclean;
429 	int entireobj;
430 	struct vnode *vp;
431 
432 	if (object->type != OBJT_VNODE || (object->flags & OBJ_WRITEABLE) == 0)
433 		return;
434 
435 	vp = object->handle;
436 
437 	if (lockflag)
438 		VOP_LOCK(vp);
439 
440 	if (start != end) {
441 		start = trunc_page(start);
442 		end = round_page(end);
443 	}
444 
445 	pass = 0;
446 startover:
447 	tstart = start;
448 	if (end == 0) {
449 		tend = object->size;
450 	} else {
451 		tend = end;
452 	}
453 	entireobj = 0;
454 	if (tstart == 0 && tend == object->size) {
455 		object->flags &= ~OBJ_WRITEABLE;
456 		entireobj = 1;
457 	}
458 
459 	pgcount = object->resident_page_count;
460 
461 	if (pass == 0 &&
462 	    (pgcount < 128 || pgcount > (object->size / (8 * PAGE_SIZE)))) {
463 		allclean = 1;
464 		for(; pgcount && (tstart < tend); tstart += PAGE_SIZE) {
465 			p = vm_page_lookup(object, tstart);
466 			if (!p)
467 				continue;
468 			--pgcount;
469 			s = splhigh();
470 			TAILQ_REMOVE(&object->memq, p, listq);
471 			TAILQ_INSERT_TAIL(&object->memq, p, listq);
472 			splx(s);
473 			if (entireobj)
474 				vm_page_protect(p, VM_PROT_READ);
475 			if ((p->flags & (PG_BUSY|PG_CACHE)) || p->busy ||
476 				p->valid == 0) {
477 				continue;
478 			}
479 			vm_page_test_dirty(p);
480 			if ((p->valid & p->dirty) != 0) {
481 				vm_offset_t tincr;
482 				tincr = vm_pageout_clean(p, VM_PAGEOUT_FORCE);
483 				if( tincr) {
484 					pgcount -= (tincr - 1);
485 					tincr *= PAGE_SIZE;
486 					tstart += tincr - PAGE_SIZE;
487 				}
488 				allclean = 0;
489 			}
490 		}
491 		if (!allclean) {
492 			pass = 1;
493 			goto startover;
494 		}
495 		if (lockflag)
496 			VOP_UNLOCK(vp);
497 		return;
498 	}
499 
500 	allclean = 1;
501 	while ((p = object->memq.tqh_first) != NULL && pgcount > 0) {
502 
503 		if (p->flags & PG_CACHE) {
504 			goto donext;
505 		}
506 
507 		if (entireobj || (p->offset >= tstart && p->offset < tend)) {
508 			if (entireobj)
509 				vm_page_protect(p, VM_PROT_READ);
510 
511 			if (p->valid == 0) {
512 				goto donext;
513 			}
514 
515 			s = splhigh();
516 			if ((p->flags & PG_BUSY) || p->busy) {
517 				allclean = 0;
518 				if (pass > 0) {
519 					p->flags |= PG_WANTED;
520 					tsleep(p, PVM, "objpcn", 0);
521 					splx(s);
522 					continue;
523 				} else {
524 					splx(s);
525 					goto donext;
526 				}
527 			}
528 
529 			TAILQ_REMOVE(&object->memq, p, listq);
530 			TAILQ_INSERT_TAIL(&object->memq, p, listq);
531 			splx(s);
532 
533 			pgcount--;
534 			vm_page_test_dirty(p);
535 			if ((p->valid & p->dirty) != 0) {
536 				vm_pageout_clean(p, VM_PAGEOUT_FORCE);
537 				allclean = 0;
538 			}
539 			continue;
540 		}
541 	donext:
542 		TAILQ_REMOVE(&object->memq, p, listq);
543 		TAILQ_INSERT_TAIL(&object->memq, p, listq);
544 		pgcount--;
545 	}
546 	if ((!allclean && (pass == 0)) ||
547 	    (entireobj && (object->flags & OBJ_WRITEABLE))) {
548 		pass = 1;
549 		if (entireobj)
550 			object->flags &= ~OBJ_WRITEABLE;
551 		goto startover;
552 	}
553 	if (lockflag)
554 		VOP_UNLOCK(vp);
555 	return;
556 }
557 
558 /*
559  *	vm_object_deactivate_pages
560  *
561  *	Deactivate all pages in the specified object.  (Keep its pages
562  *	in memory even though it is no longer referenced.)
563  *
564  *	The object must be locked.
565  */
566 void
567 vm_object_deactivate_pages(object)
568 	register vm_object_t object;
569 {
570 	register vm_page_t p, next;
571 
572 	for (p = object->memq.tqh_first; p != NULL; p = next) {
573 		next = p->listq.tqe_next;
574 		vm_page_deactivate(p);
575 	}
576 }
577 
578 /*
579  *	Trim the object cache to size.
580  */
581 void
582 vm_object_cache_trim()
583 {
584 	register vm_object_t object;
585 
586 	while (vm_object_cached > vm_object_cache_max) {
587 		object = vm_object_cached_list.tqh_first;
588 
589 		vm_object_reference(object);
590 		pager_cache(object, FALSE);
591 	}
592 }
593 
594 
595 /*
596  *	vm_object_pmap_copy:
597  *
598  *	Makes all physical pages in the specified
599  *	object range copy-on-write.  No writeable
600  *	references to these pages should remain.
601  *
602  *	The object must *not* be locked.
603  */
604 void
605 vm_object_pmap_copy(object, start, end)
606 	register vm_object_t object;
607 	register vm_offset_t start;
608 	register vm_offset_t end;
609 {
610 	register vm_page_t p;
611 
612 	if (object == NULL)
613 		return;
614 
615 	for (p = object->memq.tqh_first; p != NULL; p = p->listq.tqe_next) {
616 		if ((start <= p->offset) && (p->offset < end)) {
617 			vm_page_protect(p, VM_PROT_READ);
618 		}
619 	}
620 }
621 
622 /*
623  *	vm_object_pmap_remove:
624  *
625  *	Removes all physical pages in the specified
626  *	object range from all physical maps.
627  *
628  *	The object must *not* be locked.
629  */
630 void
631 vm_object_pmap_remove(object, start, end)
632 	register vm_object_t object;
633 	register vm_offset_t start;
634 	register vm_offset_t end;
635 {
636 	register vm_page_t p;
637 	int s;
638 
639 	if (object == NULL)
640 		return;
641 	++object->paging_in_progress;
642 
643 again:
644 	for (p = object->memq.tqh_first; p != NULL; p = p->listq.tqe_next) {
645 		if ((start <= p->offset) && (p->offset < end)) {
646 			s = splhigh();
647 			if ((p->flags & PG_BUSY) || p->busy) {
648 				p->flags |= PG_WANTED;
649 				tsleep(p, PVM, "vmopmr", 0);
650 				splx(s);
651 				goto again;
652 			}
653 			splx(s);
654 			vm_page_protect(p, VM_PROT_NONE);
655 		}
656 	}
657 	vm_object_pip_wakeup(object);
658 }
659 
660 /*
661  *	vm_object_copy:
662  *
663  *	Create a new object which is a copy of an existing
664  *	object, and mark all of the pages in the existing
665  *	object 'copy-on-write'.  The new object has one reference.
666  *	Returns the new object.
667  *
668  *	May defer the copy until later if the object is not backed
669  *	up by a non-default pager.
670  */
671 void
672 vm_object_copy(src_object, src_offset, size,
673     dst_object, dst_offset, src_needs_copy)
674 	register vm_object_t src_object;
675 	vm_offset_t src_offset;
676 	vm_size_t size;
677 	vm_object_t *dst_object;/* OUT */
678 	vm_offset_t *dst_offset;/* OUT */
679 	boolean_t *src_needs_copy;	/* OUT */
680 {
681 	register vm_object_t new_copy;
682 	register vm_object_t old_copy;
683 	vm_offset_t new_start, new_end;
684 
685 	register vm_page_t p;
686 
687 	if (src_object == NULL) {
688 		/*
689 		 * Nothing to copy
690 		 */
691 		*dst_object = NULL;
692 		*dst_offset = 0;
693 		*src_needs_copy = FALSE;
694 		return;
695 	}
696 
697 	/*
698 	 * Try to collapse the object before copying it.
699 	 */
700 	if (src_object->handle == NULL &&
701 	    (src_object->type == OBJT_DEFAULT ||
702 	     src_object->type == OBJT_SWAP))
703 		vm_object_collapse(src_object);
704 
705 
706 	/*
707 	 * Make another reference to the object
708 	 */
709 	src_object->ref_count++;
710 
711 	*dst_object = src_object;
712 	*dst_offset = src_offset;
713 
714 	/*
715 	 * Must make a shadow when write is desired
716 	 */
717 	*src_needs_copy = TRUE;
718 	return;
719 }
720 
721 /*
722  *	vm_object_shadow:
723  *
724  *	Create a new object which is backed by the
725  *	specified existing object range.  The source
726  *	object reference is deallocated.
727  *
728  *	The new object and offset into that object
729  *	are returned in the source parameters.
730  */
731 
732 void
733 vm_object_shadow(object, offset, length)
734 	vm_object_t *object;	/* IN/OUT */
735 	vm_offset_t *offset;	/* IN/OUT */
736 	vm_size_t length;
737 {
738 	register vm_object_t source;
739 	register vm_object_t result;
740 
741 	source = *object;
742 
743 	/*
744 	 * Allocate a new object with the given length
745 	 */
746 
747 	if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
748 		panic("vm_object_shadow: no object for shadowing");
749 
750 	/*
751 	 * The new object shadows the source object, adding a reference to it.
752 	 * Our caller changes his reference to point to the new object,
753 	 * removing a reference to the source object.  Net result: no change
754 	 * of reference count.
755 	 */
756 	result->backing_object = source;
757 	if (source)
758 		TAILQ_INSERT_TAIL(&result->backing_object->shadow_head, result, shadow_list);
759 
760 	/*
761 	 * Store the offset into the source object, and fix up the offset into
762 	 * the new object.
763 	 */
764 
765 	result->backing_object_offset = *offset;
766 
767 	/*
768 	 * Return the new things
769 	 */
770 
771 	*offset = 0;
772 	*object = result;
773 }
774 
775 
776 /*
777  * this version of collapse allows the operation to occur earlier and
778  * when paging_in_progress is true for an object...  This is not a complete
779  * operation, but should plug 99.9% of the rest of the leaks.
780  */
781 static void
782 vm_object_qcollapse(object)
783 	register vm_object_t object;
784 {
785 	register vm_object_t backing_object;
786 	register vm_offset_t backing_offset, new_offset;
787 	register vm_page_t p, pp;
788 	register vm_size_t size;
789 
790 	backing_object = object->backing_object;
791 	if (backing_object->ref_count != 1)
792 		return;
793 
794 	backing_object->ref_count += 2;
795 
796 	backing_offset = object->backing_object_offset;
797 	size = object->size;
798 	p = backing_object->memq.tqh_first;
799 	while (p) {
800 		vm_page_t next;
801 
802 		next = p->listq.tqe_next;
803 		if ((p->flags & (PG_BUSY | PG_FICTITIOUS | PG_CACHE)) ||
804 		    !p->valid || p->hold_count || p->wire_count || p->busy) {
805 			p = next;
806 			continue;
807 		}
808 		vm_page_protect(p, VM_PROT_NONE);
809 		new_offset = (p->offset - backing_offset);
810 		if (p->offset < backing_offset ||
811 		    new_offset >= size) {
812 			if (backing_object->type == OBJT_SWAP)
813 				swap_pager_freespace(backing_object,
814 				    backing_object->paging_offset + p->offset, PAGE_SIZE);
815 			vm_page_free(p);
816 		} else {
817 			pp = vm_page_lookup(object, new_offset);
818 			if (pp != NULL || (object->type == OBJT_SWAP && vm_pager_has_page(object,
819 				    object->paging_offset + new_offset, NULL, NULL))) {
820 				if (backing_object->type == OBJT_SWAP)
821 					swap_pager_freespace(backing_object,
822 					    backing_object->paging_offset + p->offset, PAGE_SIZE);
823 				vm_page_free(p);
824 			} else {
825 				if (backing_object->type == OBJT_SWAP)
826 					swap_pager_freespace(backing_object,
827 					    backing_object->paging_offset + p->offset, PAGE_SIZE);
828 				vm_page_rename(p, object, new_offset);
829 				p->dirty = VM_PAGE_BITS_ALL;
830 			}
831 		}
832 		p = next;
833 	}
834 	backing_object->ref_count -= 2;
835 }
836 
837 /*
838  *	vm_object_collapse:
839  *
840  *	Collapse an object with the object backing it.
841  *	Pages in the backing object are moved into the
842  *	parent, and the backing object is deallocated.
843  */
844 void
845 vm_object_collapse(object)
846 	vm_object_t object;
847 
848 {
849 	vm_object_t backing_object;
850 	vm_offset_t backing_offset;
851 	vm_size_t size;
852 	vm_offset_t new_offset;
853 	vm_page_t p, pp;
854 
855 	while (TRUE) {
856 		/*
857 		 * Verify that the conditions are right for collapse:
858 		 *
859 		 * The object exists and no pages in it are currently being paged
860 		 * out.
861 		 */
862 		if (object == NULL)
863 			return;
864 
865 		/*
866 		 * Make sure there is a backing object.
867 		 */
868 		if ((backing_object = object->backing_object) == NULL)
869 			return;
870 
871 		/*
872 		 * we check the backing object first, because it is most likely
873 		 * not collapsable.
874 		 */
875 		if (backing_object->handle != NULL ||
876 		    (backing_object->type != OBJT_DEFAULT &&
877 		     backing_object->type != OBJT_SWAP) ||
878 		    (backing_object->flags & OBJ_DEAD) ||
879 		    object->handle != NULL ||
880 		    (object->type != OBJT_DEFAULT &&
881 		     object->type != OBJT_SWAP) ||
882 		    (object->flags & OBJ_DEAD)) {
883 			return;
884 		}
885 
886 		if (object->paging_in_progress != 0 ||
887 		    backing_object->paging_in_progress != 0) {
888 			vm_object_qcollapse(object);
889 			return;
890 		}
891 
892 		/*
893 		 * We know that we can either collapse the backing object (if
894 		 * the parent is the only reference to it) or (perhaps) remove
895 		 * the parent's reference to it.
896 		 */
897 
898 		backing_offset = object->backing_object_offset;
899 		size = object->size;
900 
901 		/*
902 		 * If there is exactly one reference to the backing object, we
903 		 * can collapse it into the parent.
904 		 */
905 
906 		if (backing_object->ref_count == 1) {
907 
908 			backing_object->flags |= OBJ_DEAD;
909 			/*
910 			 * We can collapse the backing object.
911 			 *
912 			 * Move all in-memory pages from backing_object to the
913 			 * parent.  Pages that have been paged out will be
914 			 * overwritten by any of the parent's pages that
915 			 * shadow them.
916 			 */
917 
918 			while ((p = backing_object->memq.tqh_first) != 0) {
919 
920 				new_offset = (p->offset - backing_offset);
921 
922 				/*
923 				 * If the parent has a page here, or if this
924 				 * page falls outside the parent, dispose of
925 				 * it.
926 				 *
927 				 * Otherwise, move it as planned.
928 				 */
929 
930 				if (p->offset < backing_offset ||
931 				    new_offset >= size) {
932 					vm_page_protect(p, VM_PROT_NONE);
933 					PAGE_WAKEUP(p);
934 					vm_page_free(p);
935 				} else {
936 					pp = vm_page_lookup(object, new_offset);
937 					if (pp != NULL || (object->type == OBJT_SWAP && vm_pager_has_page(object,
938 					    object->paging_offset + new_offset, NULL, NULL))) {
939 						vm_page_protect(p, VM_PROT_NONE);
940 						PAGE_WAKEUP(p);
941 						vm_page_free(p);
942 					} else {
943 						vm_page_rename(p, object, new_offset);
944 					}
945 				}
946 			}
947 
948 			/*
949 			 * Move the pager from backing_object to object.
950 			 */
951 
952 			if (backing_object->type == OBJT_SWAP) {
953 				backing_object->paging_in_progress++;
954 				if (object->type == OBJT_SWAP) {
955 					object->paging_in_progress++;
956 					/*
957 					 * copy shadow object pages into ours
958 					 * and destroy unneeded pages in
959 					 * shadow object.
960 					 */
961 					swap_pager_copy(
962 					    backing_object, backing_object->paging_offset,
963 					    object, object->paging_offset,
964 					    object->backing_object_offset);
965 					vm_object_pip_wakeup(object);
966 				} else {
967 					object->paging_in_progress++;
968 					/*
969 					 * move the shadow backing_object's pager data to
970 					 * "object" and convert "object" type to OBJT_SWAP.
971 					 */
972 					object->type = OBJT_SWAP;
973 					object->un_pager.swp.swp_nblocks =
974 					    backing_object->un_pager.swp.swp_nblocks;
975 					object->un_pager.swp.swp_allocsize =
976 					    backing_object->un_pager.swp.swp_allocsize;
977 					object->un_pager.swp.swp_blocks =
978 					    backing_object->un_pager.swp.swp_blocks;
979 					object->un_pager.swp.swp_poip =		/* XXX */
980 					    backing_object->un_pager.swp.swp_poip;
981 					object->paging_offset = backing_object->paging_offset + backing_offset;
982 					TAILQ_INSERT_TAIL(&swap_pager_un_object_list, object, pager_object_list);
983 
984 					/*
985 					 * Convert backing object from OBJT_SWAP to
986 					 * OBJT_DEFAULT. XXX - only the TAILQ_REMOVE is
987 					 * actually necessary.
988 					 */
989 					backing_object->type = OBJT_DEFAULT;
990 					TAILQ_REMOVE(&swap_pager_un_object_list, backing_object, pager_object_list);
991 					/*
992 					 * free unnecessary blocks
993 					 */
994 					swap_pager_freespace(object, 0, object->paging_offset);
995 					vm_object_pip_wakeup(object);
996 				}
997 
998 				vm_object_pip_wakeup(backing_object);
999 			}
1000 			/*
1001 			 * Object now shadows whatever backing_object did.
1002 			 * Note that the reference to backing_object->backing_object
1003 			 * moves from within backing_object to within object.
1004 			 */
1005 
1006 			TAILQ_REMOVE(&object->backing_object->shadow_head, object,
1007 			    shadow_list);
1008 			if (backing_object->backing_object)
1009 				TAILQ_REMOVE(&backing_object->backing_object->shadow_head,
1010 				    backing_object, shadow_list);
1011 			object->backing_object = backing_object->backing_object;
1012 			if (object->backing_object)
1013 				TAILQ_INSERT_TAIL(&object->backing_object->shadow_head,
1014 				    object, shadow_list);
1015 
1016 			object->backing_object_offset += backing_object->backing_object_offset;
1017 			/*
1018 			 * Discard backing_object.
1019 			 *
1020 			 * Since the backing object has no pages, no pager left,
1021 			 * and no object references within it, all that is
1022 			 * necessary is to dispose of it.
1023 			 */
1024 
1025 			TAILQ_REMOVE(&vm_object_list, backing_object,
1026 			    object_list);
1027 			vm_object_count--;
1028 
1029 			free((caddr_t) backing_object, M_VMOBJ);
1030 
1031 			object_collapses++;
1032 		} else {
1033 			/*
1034 			 * If all of the pages in the backing object are
1035 			 * shadowed by the parent object, the parent object no
1036 			 * longer has to shadow the backing object; it can
1037 			 * shadow the next one in the chain.
1038 			 *
1039 			 * The backing object must not be paged out - we'd have
1040 			 * to check all of the paged-out pages, as well.
1041 			 */
1042 
1043 			if (backing_object->type != OBJT_DEFAULT) {
1044 				return;
1045 			}
1046 			/*
1047 			 * Should have a check for a 'small' number of pages
1048 			 * here.
1049 			 */
1050 
1051 			for (p = backing_object->memq.tqh_first; p; p = p->listq.tqe_next) {
1052 				new_offset = (p->offset - backing_offset);
1053 
1054 				/*
1055 				 * If the parent has a page here, or if this
1056 				 * page falls outside the parent, keep going.
1057 				 *
1058 				 * Otherwise, the backing_object must be left in
1059 				 * the chain.
1060 				 */
1061 
1062 				if (p->offset >= backing_offset && new_offset <= size) {
1063 
1064 					pp = vm_page_lookup(object, new_offset);
1065 
1066 					if ((pp == NULL || pp->valid == 0) &&
1067 				   	    !vm_pager_has_page(object, object->paging_offset + new_offset, NULL, NULL)) {
1068 
1069 						/*
1070 						 * Page still needed. Can't go any
1071 						 * further.
1072 						 */
1073 						return;
1074 					}
1075 				}
1076 			}
1077 
1078 			/*
1079 			 * Make the parent shadow the next object in the
1080 			 * chain.  Deallocating backing_object will not remove
1081 			 * it, since its reference count is at least 2.
1082 			 */
1083 
1084 			TAILQ_REMOVE(&object->backing_object->shadow_head,
1085 			    object, shadow_list);
1086 			vm_object_reference(object->backing_object = backing_object->backing_object);
1087 			if (object->backing_object)
1088 				TAILQ_INSERT_TAIL(&object->backing_object->shadow_head,
1089 				    object, shadow_list);
1090 			object->backing_object_offset += backing_object->backing_object_offset;
1091 
1092 			/*
1093 			 * Drop the reference count on backing_object. Since
1094 			 * its ref_count was at least 2, it will not vanish;
1095 			 * so we don't need to call vm_object_deallocate.
1096 			 */
1097 			if (backing_object->ref_count == 1)
1098 				printf("should have called obj deallocate\n");
1099 			backing_object->ref_count--;
1100 
1101 			object_bypasses++;
1102 
1103 		}
1104 
1105 		/*
1106 		 * Try again with this object's new backing object.
1107 		 */
1108 	}
1109 }
1110 
1111 /*
1112  *	vm_object_page_remove: [internal]
1113  *
1114  *	Removes all physical pages in the specified
1115  *	object range from the object's list of pages.
1116  *
1117  *	The object must be locked.
1118  */
1119 void
1120 vm_object_page_remove(object, start, end, clean_only)
1121 	register vm_object_t object;
1122 	register vm_offset_t start;
1123 	register vm_offset_t end;
1124 	boolean_t clean_only;
1125 {
1126 	register vm_page_t p, next;
1127 	vm_offset_t size;
1128 	int s;
1129 
1130 	if (object == NULL)
1131 		return;
1132 
1133 	object->paging_in_progress++;
1134 	start = trunc_page(start);
1135 	end = round_page(end);
1136 again:
1137 	size = end - start;
1138 	if (size > 4 * PAGE_SIZE || size >= object->size / 4) {
1139 		for (p = object->memq.tqh_first; p != NULL; p = next) {
1140 			next = p->listq.tqe_next;
1141 			if ((start <= p->offset) && (p->offset < end)) {
1142 				s = splhigh();
1143 				if (p->bmapped) {
1144 					splx(s);
1145 					continue;
1146 				}
1147 				if ((p->flags & PG_BUSY) || p->busy) {
1148 					p->flags |= PG_WANTED;
1149 					tsleep(p, PVM, "vmopar", 0);
1150 					splx(s);
1151 					goto again;
1152 				}
1153 				splx(s);
1154 				if (clean_only) {
1155 					vm_page_test_dirty(p);
1156 					if (p->valid & p->dirty)
1157 						continue;
1158 				}
1159 				vm_page_protect(p, VM_PROT_NONE);
1160 				PAGE_WAKEUP(p);
1161 				vm_page_free(p);
1162 			}
1163 		}
1164 	} else {
1165 		while (size > 0) {
1166 			while ((p = vm_page_lookup(object, start)) != 0) {
1167 				s = splhigh();
1168 				if (p->bmapped) {
1169 					splx(s);
1170 					break;
1171 				}
1172 				if ((p->flags & PG_BUSY) || p->busy) {
1173 					p->flags |= PG_WANTED;
1174 					tsleep(p, PVM, "vmopar", 0);
1175 					splx(s);
1176 					goto again;
1177 				}
1178 				splx(s);
1179 				if (clean_only) {
1180 					vm_page_test_dirty(p);
1181 					if (p->valid & p->dirty)
1182 						continue;
1183 				}
1184 				vm_page_protect(p, VM_PROT_NONE);
1185 				PAGE_WAKEUP(p);
1186 				vm_page_free(p);
1187 			}
1188 			start += PAGE_SIZE;
1189 			size -= PAGE_SIZE;
1190 		}
1191 	}
1192 	vm_object_pip_wakeup(object);
1193 }
1194 
1195 /*
1196  *	Routine:	vm_object_coalesce
1197  *	Function:	Coalesces two objects backing up adjoining
1198  *			regions of memory into a single object.
1199  *
1200  *	returns TRUE if objects were combined.
1201  *
1202  *	NOTE:	Only works at the moment if the second object is NULL -
1203  *		if it's not, which object do we lock first?
1204  *
1205  *	Parameters:
1206  *		prev_object	First object to coalesce
1207  *		prev_offset	Offset into prev_object
1208  *		next_object	Second object into coalesce
1209  *		next_offset	Offset into next_object
1210  *
1211  *		prev_size	Size of reference to prev_object
1212  *		next_size	Size of reference to next_object
1213  *
1214  *	Conditions:
1215  *	The object must *not* be locked.
1216  */
1217 boolean_t
1218 vm_object_coalesce(prev_object, next_object,
1219     prev_offset, next_offset,
1220     prev_size, next_size)
1221 	register vm_object_t prev_object;
1222 	vm_object_t next_object;
1223 	vm_offset_t prev_offset, next_offset;
1224 	vm_size_t prev_size, next_size;
1225 {
1226 	vm_size_t newsize;
1227 
1228 	if (next_object != NULL) {
1229 		return (FALSE);
1230 	}
1231 	if (prev_object == NULL) {
1232 		return (TRUE);
1233 	}
1234 
1235 	/*
1236 	 * Try to collapse the object first
1237 	 */
1238 	vm_object_collapse(prev_object);
1239 
1240 	/*
1241 	 * Can't coalesce if: . more than one reference . paged out . shadows
1242 	 * another object . has a copy elsewhere (any of which mean that the
1243 	 * pages not mapped to prev_entry may be in use anyway)
1244 	 */
1245 
1246 	if (prev_object->ref_count > 1 ||
1247 	    prev_object->type != OBJT_DEFAULT ||
1248 	    prev_object->backing_object != NULL) {
1249 		return (FALSE);
1250 	}
1251 	/*
1252 	 * Remove any pages that may still be in the object from a previous
1253 	 * deallocation.
1254 	 */
1255 
1256 	vm_object_page_remove(prev_object,
1257 	    prev_offset + prev_size,
1258 	    prev_offset + prev_size + next_size, FALSE);
1259 
1260 	/*
1261 	 * Extend the object if necessary.
1262 	 */
1263 	newsize = prev_offset + prev_size + next_size;
1264 	if (newsize > prev_object->size)
1265 		prev_object->size = newsize;
1266 
1267 	return (TRUE);
1268 }
1269 
1270 /*
1271  * returns page after looking up in shadow chain
1272  */
1273 
1274 vm_page_t
1275 vm_object_page_lookup(object, offset)
1276 	vm_object_t object;
1277 	vm_offset_t offset;
1278 {
1279 	vm_page_t m;
1280 
1281 	if (!(m = vm_page_lookup(object, offset))) {
1282 		if (!object->backing_object)
1283 			return 0;
1284 		else
1285 			return vm_object_page_lookup(object->backing_object, offset + object->backing_object_offset);
1286 	}
1287 	return m;
1288 }
1289 
1290 #ifdef DDB
1291 
1292 int
1293 _vm_object_in_map(map, object, entry)
1294 	vm_map_t map;
1295 	vm_object_t object;
1296 	vm_map_entry_t entry;
1297 {
1298 	vm_map_t tmpm;
1299 	vm_map_entry_t tmpe;
1300 	vm_object_t obj;
1301 	int entcount;
1302 
1303 	if (map == 0)
1304 		return 0;
1305 
1306 	if (entry == 0) {
1307 		tmpe = map->header.next;
1308 		entcount = map->nentries;
1309 		while (entcount-- && (tmpe != &map->header)) {
1310 			if( _vm_object_in_map(map, object, tmpe)) {
1311 				return 1;
1312 			}
1313 			tmpe = tmpe->next;
1314 		}
1315 	} else if (entry->is_sub_map || entry->is_a_map) {
1316 		tmpm = entry->object.share_map;
1317 		tmpe = tmpm->header.next;
1318 		entcount = tmpm->nentries;
1319 		while (entcount-- && tmpe != &tmpm->header) {
1320 			if( _vm_object_in_map(tmpm, object, tmpe)) {
1321 				return 1;
1322 			}
1323 			tmpe = tmpe->next;
1324 		}
1325 	} else if (obj = entry->object.vm_object) {
1326 		for(; obj; obj=obj->backing_object)
1327 			if( obj == object) {
1328 				return 1;
1329 			}
1330 	}
1331 	return 0;
1332 }
1333 
1334 int
1335 vm_object_in_map( object)
1336 	vm_object_t object;
1337 {
1338 	struct proc *p;
1339 	for (p = (struct proc *) allproc; p != NULL; p = p->p_next) {
1340 		if( !p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1341 			continue;
1342 /*
1343 		if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
1344 			continue;
1345 		}
1346 */
1347 		if( _vm_object_in_map(&p->p_vmspace->vm_map, object, 0))
1348 			return 1;
1349 	}
1350 	if( _vm_object_in_map( kernel_map, object, 0))
1351 		return 1;
1352 	if( _vm_object_in_map( kmem_map, object, 0))
1353 		return 1;
1354 	if( _vm_object_in_map( pager_map, object, 0))
1355 		return 1;
1356 	if( _vm_object_in_map( buffer_map, object, 0))
1357 		return 1;
1358 	if( _vm_object_in_map( io_map, object, 0))
1359 		return 1;
1360 	if( _vm_object_in_map( phys_map, object, 0))
1361 		return 1;
1362 	if( _vm_object_in_map( mb_map, object, 0))
1363 		return 1;
1364 	if( _vm_object_in_map( u_map, object, 0))
1365 		return 1;
1366 	return 0;
1367 }
1368 
1369 
1370 void
1371 vm_object_check() {
1372 	int i;
1373 	int maxhash = 0;
1374 	vm_object_t object;
1375 
1376 	/*
1377 	 * make sure that internal objs are in a map somewhere
1378 	 * and none have zero ref counts.
1379 	 */
1380 	for (object = vm_object_list.tqh_first;
1381 			object != NULL;
1382 			object = object->object_list.tqe_next) {
1383 		if (object->handle == NULL &&
1384 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1385 			if (object->ref_count == 0) {
1386 				printf("vmochk: internal obj has zero ref count: %d\n",
1387 					object->size);
1388 			}
1389 			if (!vm_object_in_map(object)) {
1390 				printf("vmochk: internal obj is not in a map: ref: %d, size: %d: 0x%x, backing_object: 0x%x\n",
1391 				    object->ref_count, object->size, object->backing_object);
1392 			}
1393 		}
1394 	}
1395 }
1396 
1397 /*
1398  *	vm_object_print:	[ debug ]
1399  */
1400 void
1401 vm_object_print(iobject, full, dummy3, dummy4)
1402 	/* db_expr_t */ int iobject;
1403 	boolean_t full;
1404 	/* db_expr_t */ int dummy3;
1405 	char *dummy4;
1406 {
1407 	vm_object_t object = (vm_object_t)iobject;	/* XXX */
1408 	register vm_page_t p;
1409 
1410 	register int count;
1411 
1412 	if (object == NULL)
1413 		return;
1414 
1415 	iprintf("Object 0x%x: size=0x%x, res=%d, ref=%d, ",
1416 	    (int) object, (int) object->size,
1417 	    object->resident_page_count, object->ref_count);
1418 	printf("offset=0x%x, backing_object=(0x%x)+0x%x\n",
1419 	    (int) object->paging_offset,
1420 	    (int) object->backing_object, (int) object->backing_object_offset);
1421 	printf("cache: next=%p, prev=%p\n",
1422 	    object->cached_list.tqe_next, object->cached_list.tqe_prev);
1423 
1424 	if (!full)
1425 		return;
1426 
1427 	indent += 2;
1428 	count = 0;
1429 	for (p = object->memq.tqh_first; p != NULL; p = p->listq.tqe_next) {
1430 		if (count == 0)
1431 			iprintf("memory:=");
1432 		else if (count == 6) {
1433 			printf("\n");
1434 			iprintf(" ...");
1435 			count = 0;
1436 		} else
1437 			printf(",");
1438 		count++;
1439 
1440 		printf("(off=0x%lx,page=0x%lx)",
1441 		    (u_long) p->offset, (u_long) VM_PAGE_TO_PHYS(p));
1442 	}
1443 	if (count != 0)
1444 		printf("\n");
1445 	indent -= 2;
1446 }
1447 #endif /* DDB */
1448