xref: /freebsd/sys/vm/vm_object.c (revision 48991a368427cadb9cdac39581d1676c29619c52)
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 1.55 1995/11/05 20:46:00 dyson 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 extern int vm_pageout_page_count;
125 
126 long object_collapses;
127 long object_bypasses;
128 
129 static void
130 _vm_object_allocate(type, size, object)
131 	objtype_t type;
132 	vm_size_t size;
133 	register vm_object_t object;
134 {
135 	TAILQ_INIT(&object->memq);
136 	TAILQ_INIT(&object->shadow_head);
137 
138 	object->type = type;
139 	object->size = size;
140 	object->ref_count = 1;
141 	object->flags = 0;
142 	object->paging_in_progress = 0;
143 	object->resident_page_count = 0;
144 	object->handle = NULL;
145 	object->paging_offset = 0;
146 	object->backing_object = NULL;
147 	object->backing_object_offset = (vm_offset_t) 0;
148 
149 	object->last_read = 0;
150 
151 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
152 	vm_object_count++;
153 }
154 
155 /*
156  *	vm_object_init:
157  *
158  *	Initialize the VM objects module.
159  */
160 void
161 vm_object_init(vm_offset_t nothing)
162 {
163 	TAILQ_INIT(&vm_object_cached_list);
164 	TAILQ_INIT(&vm_object_list);
165 	vm_object_count = 0;
166 
167 	vm_object_cache_max = 84;
168 	if (cnt.v_page_count > 1000)
169 		vm_object_cache_max += (cnt.v_page_count - 1000) / 4;
170 
171 	kernel_object = &kernel_object_store;
172 	_vm_object_allocate(OBJT_DEFAULT, VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS,
173 	    kernel_object);
174 
175 	kmem_object = &kmem_object_store;
176 	_vm_object_allocate(OBJT_DEFAULT, VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS,
177 	    kmem_object);
178 }
179 
180 /*
181  *	vm_object_allocate:
182  *
183  *	Returns a new object with the given size.
184  */
185 
186 vm_object_t
187 vm_object_allocate(type, size)
188 	objtype_t type;
189 	vm_size_t size;
190 {
191 	register vm_object_t result;
192 
193 	result = (vm_object_t)
194 	    malloc((u_long) sizeof *result, M_VMOBJ, M_WAITOK);
195 
196 
197 	_vm_object_allocate(type, size, result);
198 
199 	return (result);
200 }
201 
202 
203 /*
204  *	vm_object_reference:
205  *
206  *	Gets another reference to the given object.
207  */
208 inline void
209 vm_object_reference(object)
210 	register vm_object_t object;
211 {
212 	if (object == NULL)
213 		return;
214 
215 	if (object->ref_count == 0) {
216 		if ((object->flags & OBJ_CANPERSIST) == 0)
217 			panic("vm_object_reference: non-persistent object with 0 ref_count");
218 		TAILQ_REMOVE(&vm_object_cached_list, object, cached_list);
219 		vm_object_cached--;
220 	}
221 	object->ref_count++;
222 }
223 
224 /*
225  *	vm_object_deallocate:
226  *
227  *	Release a reference to the specified object,
228  *	gained either through a vm_object_allocate
229  *	or a vm_object_reference call.  When all references
230  *	are gone, storage associated with this object
231  *	may be relinquished.
232  *
233  *	No object may be locked.
234  */
235 void
236 vm_object_deallocate(object)
237 	vm_object_t object;
238 {
239 	vm_object_t temp;
240 
241 	while (object != NULL) {
242 
243 		if (object->ref_count == 0)
244 			panic("vm_object_deallocate: object deallocated too many times");
245 
246 		/*
247 		 * Lose the reference
248 		 */
249 		object->ref_count--;
250 
251 		if (object->ref_count != 0) {
252 			if ((object->ref_count == 1) &&
253 			    (object->handle == NULL) &&
254 			    (object->type == OBJT_DEFAULT ||
255 			     object->type == OBJT_SWAP)) {
256 				vm_object_t robject;
257 				robject = object->shadow_head.tqh_first;
258 				if ((robject != NULL) &&
259 				    (robject->handle == NULL) &&
260 				    (robject->type == OBJT_DEFAULT ||
261 				     robject->type == OBJT_SWAP)) {
262 					int s;
263 					robject->ref_count += 2;
264 					object->ref_count += 2;
265 
266 					do {
267 						s = splhigh();
268 						while (robject->paging_in_progress) {
269 							robject->flags |= OBJ_PIPWNT;
270 							tsleep(robject, PVM, "objde1", 0);
271 						}
272 
273 						while (object->paging_in_progress) {
274 							object->flags |= OBJ_PIPWNT;
275 							tsleep(object, PVM, "objde2", 0);
276 						}
277 						splx(s);
278 
279 					} while( object->paging_in_progress || robject->paging_in_progress);
280 
281 					object->ref_count -= 2;
282 					robject->ref_count -= 2;
283 					if( robject->ref_count == 0) {
284 						robject->ref_count += 1;
285 						object = robject;
286 						continue;
287 					}
288 					vm_object_collapse(robject);
289 					return;
290 				}
291 			}
292 			/*
293 			 * If there are still references, then we are done.
294 			 */
295 			return;
296 		}
297 
298 		if (object->type == OBJT_VNODE) {
299 			struct vnode *vp = object->handle;
300 
301 			vp->v_flag &= ~VTEXT;
302 		}
303 
304 		/*
305 		 * See if this object can persist and has some resident
306 		 * pages.  If so, enter it in the cache.
307 		 */
308 		if (object->flags & OBJ_CANPERSIST) {
309 			if (object->resident_page_count != 0) {
310 				vm_object_page_clean(object, 0, 0 ,TRUE, TRUE);
311 				TAILQ_INSERT_TAIL(&vm_object_cached_list, object,
312 				    cached_list);
313 				vm_object_cached++;
314 
315 				vm_object_cache_trim();
316 				return;
317 			} else {
318 				object->flags &= ~OBJ_CANPERSIST;
319 			}
320 		}
321 
322 		/*
323 		 * Make sure no one uses us.
324 		 */
325 		object->flags |= OBJ_DEAD;
326 
327 		temp = object->backing_object;
328 		if (temp)
329 			TAILQ_REMOVE(&temp->shadow_head, object, shadow_list);
330 		vm_object_terminate(object);
331 		/* unlocks and deallocates object */
332 		object = temp;
333 	}
334 }
335 
336 /*
337  *	vm_object_terminate actually destroys the specified object, freeing
338  *	up all previously used resources.
339  *
340  *	The object must be locked.
341  */
342 void
343 vm_object_terminate(object)
344 	register vm_object_t object;
345 {
346 	register vm_page_t p;
347 	int s;
348 
349 	/*
350 	 * wait for the pageout daemon to be done with the object
351 	 */
352 	s = splhigh();
353 	while (object->paging_in_progress) {
354 		object->flags |= OBJ_PIPWNT;
355 		tsleep(object, PVM, "objtrm", 0);
356 	}
357 	splx(s);
358 
359 	if (object->paging_in_progress != 0)
360 		panic("vm_object_deallocate: pageout in progress");
361 
362 	/*
363 	 * Clean and free the pages, as appropriate. All references to the
364 	 * object are gone, so we don't need to lock it.
365 	 */
366 	if (object->type == OBJT_VNODE) {
367 		struct vnode *vp = object->handle;
368 
369 		VOP_LOCK(vp);
370 		vm_object_page_clean(object, 0, 0, TRUE, FALSE);
371 		vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0);
372 		VOP_UNLOCK(vp);
373 	}
374 
375 	/*
376 	 * Now free the pages. For internal objects, this also removes them
377 	 * from paging queues.
378 	 */
379 	while ((p = object->memq.tqh_first) != NULL) {
380 		if (p->flags & PG_BUSY)
381 			printf("vm_object_terminate: freeing busy page\n");
382 		PAGE_WAKEUP(p);
383 		vm_page_free(p);
384 		cnt.v_pfree++;
385 	}
386 
387 	/*
388 	 * Let the pager know object is dead.
389 	 */
390 	vm_pager_deallocate(object);
391 
392 	TAILQ_REMOVE(&vm_object_list, object, object_list);
393 	vm_object_count--;
394 
395 	wakeup(object);
396 
397 	/*
398 	 * Free the space for the object.
399 	 */
400 	free((caddr_t) object, M_VMOBJ);
401 }
402 
403 /*
404  *	vm_object_page_clean
405  *
406  *	Clean all dirty pages in the specified range of object.
407  *	Leaves page on whatever queue it is currently on.
408  *
409  *	Odd semantics: if start == end, we clean everything.
410  *
411  *	The object must be locked.
412  */
413 
414 void
415 vm_object_page_clean(object, start, end, syncio, lockflag)
416 	vm_object_t object;
417 	vm_offset_t start;
418 	vm_offset_t end;
419 	boolean_t syncio;
420 	boolean_t lockflag;
421 {
422 	register vm_page_t p;
423 	register vm_offset_t tstart, tend;
424 	int s;
425 	struct vnode *vp;
426 	int runlen;
427 	vm_page_t ma[vm_pageout_page_count];
428 
429 	if (object->type != OBJT_VNODE ||
430 		(object->flags & OBJ_MIGHTBEDIRTY) == 0)
431 		return;
432 
433 	vp = object->handle;
434 
435 	if (lockflag)
436 		VOP_LOCK(vp);
437 	object->flags |= OBJ_CLEANING;
438 
439 	if (start != end) {
440 		start = trunc_page(start);
441 		end = round_page(end);
442 	}
443 
444 	tstart = start;
445 	if (end == 0) {
446 		tend = object->size;
447 	} else {
448 		tend = end;
449 	}
450 	if (tstart == 0 && tend == object->size) {
451 		object->flags &= ~(OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
452 	}
453 
454 	runlen = 0;
455 	for(;tstart < tend; tstart += PAGE_SIZE) {
456 relookup:
457 		p = vm_page_lookup(object, tstart);
458 		if (!p) {
459 			if (runlen > 0) {
460 				vm_pageout_flush(ma, runlen, syncio);
461 				runlen = 0;
462 			}
463 			continue;
464 		}
465 		if (p->valid == 0 || (p->flags & PG_CACHE)) {
466 			if (runlen > 0) {
467 				vm_pageout_flush(ma, runlen, syncio);
468 				runlen = 0;
469 			}
470 			continue;
471 		}
472 
473 		vm_page_protect(p, VM_PROT_READ);
474 
475 		s = splhigh();
476 		while ((p->flags & PG_BUSY) || p->busy) {
477 			if (runlen > 0) {
478 				splx(s);
479 				vm_pageout_flush(ma, runlen, syncio);
480 				runlen = 0;
481 				goto relookup;
482 			}
483 			p->flags |= PG_WANTED|PG_REFERENCED;
484 			tsleep(p, PVM, "vpcwai", 0);
485 			splx(s);
486 			goto relookup;
487 		}
488 		splx(s);
489 
490 		if (p->dirty == 0)
491 			vm_page_test_dirty(p);
492 
493 		if ((p->valid & p->dirty) != 0) {
494 			ma[runlen] = p;
495 			p->flags |= PG_BUSY;
496 			runlen++;
497 			if (runlen >= vm_pageout_page_count) {
498 				vm_pageout_flush(ma, runlen, syncio);
499 				runlen = 0;
500 			}
501 		} else if (runlen > 0) {
502 			vm_pageout_flush(ma, runlen, syncio);
503 			runlen = 0;
504 		}
505 
506 	}
507 	if (runlen > 0) {
508 		vm_pageout_flush(ma, runlen, syncio);
509 	}
510 
511 	VOP_FSYNC(vp, NULL, syncio, curproc);
512 
513 	if (lockflag)
514 		VOP_UNLOCK(vp);
515 	object->flags &= ~OBJ_CLEANING;
516 	return;
517 }
518 
519 /*
520  *	vm_object_deactivate_pages
521  *
522  *	Deactivate all pages in the specified object.  (Keep its pages
523  *	in memory even though it is no longer referenced.)
524  *
525  *	The object must be locked.
526  */
527 void
528 vm_object_deactivate_pages(object)
529 	register vm_object_t object;
530 {
531 	register vm_page_t p, next;
532 
533 	for (p = object->memq.tqh_first; p != NULL; p = next) {
534 		next = p->listq.tqe_next;
535 		vm_page_deactivate(p);
536 	}
537 }
538 
539 /*
540  *	Trim the object cache to size.
541  */
542 void
543 vm_object_cache_trim()
544 {
545 	register vm_object_t object;
546 
547 	while (vm_object_cached > vm_object_cache_max) {
548 		object = vm_object_cached_list.tqh_first;
549 
550 		vm_object_reference(object);
551 		pager_cache(object, FALSE);
552 	}
553 }
554 
555 
556 /*
557  *	vm_object_pmap_copy:
558  *
559  *	Makes all physical pages in the specified
560  *	object range copy-on-write.  No writeable
561  *	references to these pages should remain.
562  *
563  *	The object must *not* be locked.
564  */
565 void
566 vm_object_pmap_copy(object, start, end)
567 	register vm_object_t object;
568 	register vm_offset_t start;
569 	register vm_offset_t end;
570 {
571 	register vm_page_t p;
572 
573 	if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
574 		return;
575 
576 	for (p = object->memq.tqh_first; p != NULL; p = p->listq.tqe_next) {
577 		vm_page_protect(p, VM_PROT_READ);
578 	}
579 
580 	object->flags &= ~OBJ_WRITEABLE;
581 }
582 
583 /*
584  *	vm_object_pmap_remove:
585  *
586  *	Removes all physical pages in the specified
587  *	object range from all physical maps.
588  *
589  *	The object must *not* be locked.
590  */
591 void
592 vm_object_pmap_remove(object, start, end)
593 	register vm_object_t object;
594 	register vm_offset_t start;
595 	register vm_offset_t end;
596 {
597 	register vm_page_t p;
598 
599 	if (object == NULL)
600 		return;
601 	for (p = object->memq.tqh_first; p != NULL; p = p->listq.tqe_next) {
602 		vm_page_protect(p, VM_PROT_NONE);
603 	}
604 }
605 
606 /*
607  *	vm_object_copy:
608  *
609  *	Create a new object which is a copy of an existing
610  *	object, and mark all of the pages in the existing
611  *	object 'copy-on-write'.  The new object has one reference.
612  *	Returns the new object.
613  *
614  *	May defer the copy until later if the object is not backed
615  *	up by a non-default pager.
616  */
617 void
618 vm_object_copy(src_object, src_offset, size,
619     dst_object, dst_offset, src_needs_copy)
620 	register vm_object_t src_object;
621 	vm_offset_t src_offset;
622 	vm_size_t size;
623 	vm_object_t *dst_object;/* OUT */
624 	vm_offset_t *dst_offset;/* OUT */
625 	boolean_t *src_needs_copy;	/* OUT */
626 {
627 	if (src_object == NULL) {
628 		/*
629 		 * Nothing to copy
630 		 */
631 		*dst_object = NULL;
632 		*dst_offset = 0;
633 		*src_needs_copy = FALSE;
634 		return;
635 	}
636 
637 	/*
638 	 * Try to collapse the object before copying it.
639 	 */
640 	if (src_object->handle == NULL &&
641 	    (src_object->type == OBJT_DEFAULT ||
642 	     src_object->type == OBJT_SWAP))
643 		vm_object_collapse(src_object);
644 
645 
646 	/*
647 	 * Make another reference to the object
648 	 */
649 	src_object->ref_count++;
650 
651 	*dst_object = src_object;
652 	*dst_offset = src_offset;
653 
654 	/*
655 	 * Must make a shadow when write is desired
656 	 */
657 	*src_needs_copy = TRUE;
658 	return;
659 }
660 
661 /*
662  *	vm_object_shadow:
663  *
664  *	Create a new object which is backed by the
665  *	specified existing object range.  The source
666  *	object reference is deallocated.
667  *
668  *	The new object and offset into that object
669  *	are returned in the source parameters.
670  */
671 
672 void
673 vm_object_shadow(object, offset, length)
674 	vm_object_t *object;	/* IN/OUT */
675 	vm_offset_t *offset;	/* IN/OUT */
676 	vm_size_t length;
677 {
678 	register vm_object_t source;
679 	register vm_object_t result;
680 
681 	source = *object;
682 
683 	/*
684 	 * Allocate a new object with the given length
685 	 */
686 
687 	if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
688 		panic("vm_object_shadow: no object for shadowing");
689 
690 	/*
691 	 * The new object shadows the source object, adding a reference to it.
692 	 * Our caller changes his reference to point to the new object,
693 	 * removing a reference to the source object.  Net result: no change
694 	 * of reference count.
695 	 */
696 	result->backing_object = source;
697 	if (source)
698 		TAILQ_INSERT_TAIL(&result->backing_object->shadow_head, result, shadow_list);
699 
700 	/*
701 	 * Store the offset into the source object, and fix up the offset into
702 	 * the new object.
703 	 */
704 
705 	result->backing_object_offset = *offset;
706 
707 	/*
708 	 * Return the new things
709 	 */
710 
711 	*offset = 0;
712 	*object = result;
713 }
714 
715 
716 /*
717  * this version of collapse allows the operation to occur earlier and
718  * when paging_in_progress is true for an object...  This is not a complete
719  * operation, but should plug 99.9% of the rest of the leaks.
720  */
721 static void
722 vm_object_qcollapse(object)
723 	register vm_object_t object;
724 {
725 	register vm_object_t backing_object;
726 	register vm_offset_t backing_offset, new_offset;
727 	register vm_page_t p, pp;
728 	register vm_size_t size;
729 
730 	backing_object = object->backing_object;
731 	if (backing_object->ref_count != 1)
732 		return;
733 
734 	backing_object->ref_count += 2;
735 
736 	backing_offset = object->backing_object_offset;
737 	size = object->size;
738 	p = backing_object->memq.tqh_first;
739 	while (p) {
740 		vm_page_t next;
741 
742 		next = p->listq.tqe_next;
743 		if ((p->flags & (PG_BUSY | PG_FICTITIOUS | PG_CACHE)) ||
744 		    !p->valid || p->hold_count || p->wire_count || p->busy) {
745 			p = next;
746 			continue;
747 		}
748 		vm_page_protect(p, VM_PROT_NONE);
749 		new_offset = (p->offset - backing_offset);
750 		if (p->offset < backing_offset ||
751 		    new_offset >= size) {
752 			if (backing_object->type == OBJT_SWAP)
753 				swap_pager_freespace(backing_object,
754 				    backing_object->paging_offset + p->offset, PAGE_SIZE);
755 			vm_page_free(p);
756 		} else {
757 			pp = vm_page_lookup(object, new_offset);
758 			if (pp != NULL || (object->type == OBJT_SWAP && vm_pager_has_page(object,
759 				    object->paging_offset + new_offset, NULL, NULL))) {
760 				if (backing_object->type == OBJT_SWAP)
761 					swap_pager_freespace(backing_object,
762 					    backing_object->paging_offset + p->offset, PAGE_SIZE);
763 				vm_page_free(p);
764 			} else {
765 				if (backing_object->type == OBJT_SWAP)
766 					swap_pager_freespace(backing_object,
767 					    backing_object->paging_offset + p->offset, PAGE_SIZE);
768 				vm_page_rename(p, object, new_offset);
769 				p->dirty = VM_PAGE_BITS_ALL;
770 			}
771 		}
772 		p = next;
773 	}
774 	backing_object->ref_count -= 2;
775 }
776 
777 /*
778  *	vm_object_collapse:
779  *
780  *	Collapse an object with the object backing it.
781  *	Pages in the backing object are moved into the
782  *	parent, and the backing object is deallocated.
783  */
784 void
785 vm_object_collapse(object)
786 	vm_object_t object;
787 
788 {
789 	vm_object_t backing_object;
790 	vm_offset_t backing_offset;
791 	vm_size_t size;
792 	vm_offset_t new_offset;
793 	vm_page_t p, pp;
794 
795 	while (TRUE) {
796 		/*
797 		 * Verify that the conditions are right for collapse:
798 		 *
799 		 * The object exists and no pages in it are currently being paged
800 		 * out.
801 		 */
802 		if (object == NULL)
803 			return;
804 
805 		/*
806 		 * Make sure there is a backing object.
807 		 */
808 		if ((backing_object = object->backing_object) == NULL)
809 			return;
810 
811 		/*
812 		 * we check the backing object first, because it is most likely
813 		 * not collapsable.
814 		 */
815 		if (backing_object->handle != NULL ||
816 		    (backing_object->type != OBJT_DEFAULT &&
817 		     backing_object->type != OBJT_SWAP) ||
818 		    (backing_object->flags & OBJ_DEAD) ||
819 		    object->handle != NULL ||
820 		    (object->type != OBJT_DEFAULT &&
821 		     object->type != OBJT_SWAP) ||
822 		    (object->flags & OBJ_DEAD)) {
823 			return;
824 		}
825 
826 		if (object->paging_in_progress != 0 ||
827 		    backing_object->paging_in_progress != 0) {
828 			vm_object_qcollapse(object);
829 			return;
830 		}
831 
832 		/*
833 		 * We know that we can either collapse the backing object (if
834 		 * the parent is the only reference to it) or (perhaps) remove
835 		 * the parent's reference to it.
836 		 */
837 
838 		backing_offset = object->backing_object_offset;
839 		size = object->size;
840 
841 		/*
842 		 * If there is exactly one reference to the backing object, we
843 		 * can collapse it into the parent.
844 		 */
845 
846 		if (backing_object->ref_count == 1) {
847 
848 			backing_object->flags |= OBJ_DEAD;
849 			/*
850 			 * We can collapse the backing object.
851 			 *
852 			 * Move all in-memory pages from backing_object to the
853 			 * parent.  Pages that have been paged out will be
854 			 * overwritten by any of the parent's pages that
855 			 * shadow them.
856 			 */
857 
858 			while ((p = backing_object->memq.tqh_first) != 0) {
859 
860 				new_offset = (p->offset - backing_offset);
861 
862 				/*
863 				 * If the parent has a page here, or if this
864 				 * page falls outside the parent, dispose of
865 				 * it.
866 				 *
867 				 * Otherwise, move it as planned.
868 				 */
869 
870 				if (p->offset < backing_offset ||
871 				    new_offset >= size) {
872 					vm_page_protect(p, VM_PROT_NONE);
873 					PAGE_WAKEUP(p);
874 					vm_page_free(p);
875 				} else {
876 					pp = vm_page_lookup(object, new_offset);
877 					if (pp != NULL || (object->type == OBJT_SWAP && vm_pager_has_page(object,
878 					    object->paging_offset + new_offset, NULL, NULL))) {
879 						vm_page_protect(p, VM_PROT_NONE);
880 						PAGE_WAKEUP(p);
881 						vm_page_free(p);
882 					} else {
883 						vm_page_rename(p, object, new_offset);
884 					}
885 				}
886 			}
887 
888 			/*
889 			 * Move the pager from backing_object to object.
890 			 */
891 
892 			if (backing_object->type == OBJT_SWAP) {
893 				backing_object->paging_in_progress++;
894 				if (object->type == OBJT_SWAP) {
895 					object->paging_in_progress++;
896 					/*
897 					 * copy shadow object pages into ours
898 					 * and destroy unneeded pages in
899 					 * shadow object.
900 					 */
901 					swap_pager_copy(
902 					    backing_object, backing_object->paging_offset,
903 					    object, object->paging_offset,
904 					    object->backing_object_offset);
905 					vm_object_pip_wakeup(object);
906 				} else {
907 					object->paging_in_progress++;
908 					/*
909 					 * move the shadow backing_object's pager data to
910 					 * "object" and convert "object" type to OBJT_SWAP.
911 					 */
912 					object->type = OBJT_SWAP;
913 					object->un_pager.swp.swp_nblocks =
914 					    backing_object->un_pager.swp.swp_nblocks;
915 					object->un_pager.swp.swp_allocsize =
916 					    backing_object->un_pager.swp.swp_allocsize;
917 					object->un_pager.swp.swp_blocks =
918 					    backing_object->un_pager.swp.swp_blocks;
919 					object->un_pager.swp.swp_poip =		/* XXX */
920 					    backing_object->un_pager.swp.swp_poip;
921 					object->paging_offset = backing_object->paging_offset + backing_offset;
922 					TAILQ_INSERT_TAIL(&swap_pager_un_object_list, object, pager_object_list);
923 
924 					/*
925 					 * Convert backing object from OBJT_SWAP to
926 					 * OBJT_DEFAULT. XXX - only the TAILQ_REMOVE is
927 					 * actually necessary.
928 					 */
929 					backing_object->type = OBJT_DEFAULT;
930 					TAILQ_REMOVE(&swap_pager_un_object_list, backing_object, pager_object_list);
931 					/*
932 					 * free unnecessary blocks
933 					 */
934 					swap_pager_freespace(object, 0, object->paging_offset);
935 					vm_object_pip_wakeup(object);
936 				}
937 
938 				vm_object_pip_wakeup(backing_object);
939 			}
940 			/*
941 			 * Object now shadows whatever backing_object did.
942 			 * Note that the reference to backing_object->backing_object
943 			 * moves from within backing_object to within object.
944 			 */
945 
946 			TAILQ_REMOVE(&object->backing_object->shadow_head, object,
947 			    shadow_list);
948 			if (backing_object->backing_object)
949 				TAILQ_REMOVE(&backing_object->backing_object->shadow_head,
950 				    backing_object, shadow_list);
951 			object->backing_object = backing_object->backing_object;
952 			if (object->backing_object)
953 				TAILQ_INSERT_TAIL(&object->backing_object->shadow_head,
954 				    object, shadow_list);
955 
956 			object->backing_object_offset += backing_object->backing_object_offset;
957 			/*
958 			 * Discard backing_object.
959 			 *
960 			 * Since the backing object has no pages, no pager left,
961 			 * and no object references within it, all that is
962 			 * necessary is to dispose of it.
963 			 */
964 
965 			TAILQ_REMOVE(&vm_object_list, backing_object,
966 			    object_list);
967 			vm_object_count--;
968 
969 			free((caddr_t) backing_object, M_VMOBJ);
970 
971 			object_collapses++;
972 		} else {
973 			/*
974 			 * If all of the pages in the backing object are
975 			 * shadowed by the parent object, the parent object no
976 			 * longer has to shadow the backing object; it can
977 			 * shadow the next one in the chain.
978 			 *
979 			 * The backing object must not be paged out - we'd have
980 			 * to check all of the paged-out pages, as well.
981 			 */
982 
983 			if (backing_object->type != OBJT_DEFAULT) {
984 				return;
985 			}
986 			/*
987 			 * Should have a check for a 'small' number of pages
988 			 * here.
989 			 */
990 
991 			for (p = backing_object->memq.tqh_first; p; p = p->listq.tqe_next) {
992 				new_offset = (p->offset - backing_offset);
993 
994 				/*
995 				 * If the parent has a page here, or if this
996 				 * page falls outside the parent, keep going.
997 				 *
998 				 * Otherwise, the backing_object must be left in
999 				 * the chain.
1000 				 */
1001 
1002 				if (p->offset >= backing_offset && new_offset <= size) {
1003 
1004 					pp = vm_page_lookup(object, new_offset);
1005 
1006 					if ((pp == NULL || pp->valid == 0) &&
1007 				   	    !vm_pager_has_page(object, object->paging_offset + new_offset, NULL, NULL)) {
1008 
1009 						/*
1010 						 * Page still needed. Can't go any
1011 						 * further.
1012 						 */
1013 						return;
1014 					}
1015 				}
1016 			}
1017 
1018 			/*
1019 			 * Make the parent shadow the next object in the
1020 			 * chain.  Deallocating backing_object will not remove
1021 			 * it, since its reference count is at least 2.
1022 			 */
1023 
1024 			TAILQ_REMOVE(&object->backing_object->shadow_head,
1025 			    object, shadow_list);
1026 			vm_object_reference(object->backing_object = backing_object->backing_object);
1027 			if (object->backing_object)
1028 				TAILQ_INSERT_TAIL(&object->backing_object->shadow_head,
1029 				    object, shadow_list);
1030 			object->backing_object_offset += backing_object->backing_object_offset;
1031 
1032 			/*
1033 			 * Drop the reference count on backing_object. Since
1034 			 * its ref_count was at least 2, it will not vanish;
1035 			 * so we don't need to call vm_object_deallocate.
1036 			 */
1037 			if (backing_object->ref_count == 1)
1038 				printf("should have called obj deallocate\n");
1039 			backing_object->ref_count--;
1040 
1041 			object_bypasses++;
1042 
1043 		}
1044 
1045 		/*
1046 		 * Try again with this object's new backing object.
1047 		 */
1048 	}
1049 }
1050 
1051 /*
1052  *	vm_object_page_remove: [internal]
1053  *
1054  *	Removes all physical pages in the specified
1055  *	object range from the object's list of pages.
1056  *
1057  *	The object must be locked.
1058  */
1059 void
1060 vm_object_page_remove(object, start, end, clean_only)
1061 	register vm_object_t object;
1062 	register vm_offset_t start;
1063 	register vm_offset_t end;
1064 	boolean_t clean_only;
1065 {
1066 	register vm_page_t p, next;
1067 	vm_offset_t size;
1068 	int s;
1069 
1070 	if (object == NULL)
1071 		return;
1072 
1073 	object->paging_in_progress++;
1074 	start = trunc_page(start);
1075 	end = round_page(end);
1076 again:
1077 	size = end - start;
1078 	if (size > 4 * PAGE_SIZE || size >= object->size / 4) {
1079 		for (p = object->memq.tqh_first; p != NULL; p = next) {
1080 			next = p->listq.tqe_next;
1081 			if ((start <= p->offset) && (p->offset < end)) {
1082 				s = splhigh();
1083 				if (p->bmapped) {
1084 					splx(s);
1085 					continue;
1086 				}
1087 				if ((p->flags & PG_BUSY) || p->busy) {
1088 					p->flags |= PG_WANTED;
1089 					tsleep(p, PVM, "vmopar", 0);
1090 					splx(s);
1091 					goto again;
1092 				}
1093 				splx(s);
1094 				if (clean_only) {
1095 					vm_page_test_dirty(p);
1096 					if (p->valid & p->dirty)
1097 						continue;
1098 				}
1099 				vm_page_protect(p, VM_PROT_NONE);
1100 				PAGE_WAKEUP(p);
1101 				vm_page_free(p);
1102 			}
1103 		}
1104 	} else {
1105 		while (size > 0) {
1106 			while ((p = vm_page_lookup(object, start)) != 0) {
1107 				s = splhigh();
1108 				if (p->bmapped) {
1109 					splx(s);
1110 					break;
1111 				}
1112 				if ((p->flags & PG_BUSY) || p->busy) {
1113 					p->flags |= PG_WANTED;
1114 					tsleep(p, PVM, "vmopar", 0);
1115 					splx(s);
1116 					goto again;
1117 				}
1118 				splx(s);
1119 				if (clean_only) {
1120 					vm_page_test_dirty(p);
1121 					if (p->valid & p->dirty)
1122 						continue;
1123 				}
1124 				vm_page_protect(p, VM_PROT_NONE);
1125 				PAGE_WAKEUP(p);
1126 				vm_page_free(p);
1127 			}
1128 			start += PAGE_SIZE;
1129 			size -= PAGE_SIZE;
1130 		}
1131 	}
1132 	vm_object_pip_wakeup(object);
1133 }
1134 
1135 /*
1136  *	Routine:	vm_object_coalesce
1137  *	Function:	Coalesces two objects backing up adjoining
1138  *			regions of memory into a single object.
1139  *
1140  *	returns TRUE if objects were combined.
1141  *
1142  *	NOTE:	Only works at the moment if the second object is NULL -
1143  *		if it's not, which object do we lock first?
1144  *
1145  *	Parameters:
1146  *		prev_object	First object to coalesce
1147  *		prev_offset	Offset into prev_object
1148  *		next_object	Second object into coalesce
1149  *		next_offset	Offset into next_object
1150  *
1151  *		prev_size	Size of reference to prev_object
1152  *		next_size	Size of reference to next_object
1153  *
1154  *	Conditions:
1155  *	The object must *not* be locked.
1156  */
1157 boolean_t
1158 vm_object_coalesce(prev_object, next_object,
1159     prev_offset, next_offset,
1160     prev_size, next_size)
1161 	register vm_object_t prev_object;
1162 	vm_object_t next_object;
1163 	vm_offset_t prev_offset, next_offset;
1164 	vm_size_t prev_size, next_size;
1165 {
1166 	vm_size_t newsize;
1167 
1168 	if (next_object != NULL) {
1169 		return (FALSE);
1170 	}
1171 	if (prev_object == NULL) {
1172 		return (TRUE);
1173 	}
1174 
1175 	/*
1176 	 * Try to collapse the object first
1177 	 */
1178 	vm_object_collapse(prev_object);
1179 
1180 	/*
1181 	 * Can't coalesce if: . more than one reference . paged out . shadows
1182 	 * another object . has a copy elsewhere (any of which mean that the
1183 	 * pages not mapped to prev_entry may be in use anyway)
1184 	 */
1185 
1186 	if (prev_object->ref_count > 1 ||
1187 	    prev_object->type != OBJT_DEFAULT ||
1188 	    prev_object->backing_object != NULL) {
1189 		return (FALSE);
1190 	}
1191 	/*
1192 	 * Remove any pages that may still be in the object from a previous
1193 	 * deallocation.
1194 	 */
1195 
1196 	vm_object_page_remove(prev_object,
1197 	    prev_offset + prev_size,
1198 	    prev_offset + prev_size + next_size, FALSE);
1199 
1200 	/*
1201 	 * Extend the object if necessary.
1202 	 */
1203 	newsize = prev_offset + prev_size + next_size;
1204 	if (newsize > prev_object->size)
1205 		prev_object->size = newsize;
1206 
1207 	return (TRUE);
1208 }
1209 
1210 /*
1211  * returns page after looking up in shadow chain
1212  */
1213 
1214 vm_page_t
1215 vm_object_page_lookup(object, offset)
1216 	vm_object_t object;
1217 	vm_offset_t offset;
1218 {
1219 	vm_page_t m;
1220 
1221 	if (!(m = vm_page_lookup(object, offset))) {
1222 		if (!object->backing_object)
1223 			return 0;
1224 		else
1225 			return vm_object_page_lookup(object->backing_object, offset + object->backing_object_offset);
1226 	}
1227 	return m;
1228 }
1229 
1230 #ifdef DDB
1231 
1232 int
1233 _vm_object_in_map(map, object, entry)
1234 	vm_map_t map;
1235 	vm_object_t object;
1236 	vm_map_entry_t entry;
1237 {
1238 	vm_map_t tmpm;
1239 	vm_map_entry_t tmpe;
1240 	vm_object_t obj;
1241 	int entcount;
1242 
1243 	if (map == 0)
1244 		return 0;
1245 
1246 	if (entry == 0) {
1247 		tmpe = map->header.next;
1248 		entcount = map->nentries;
1249 		while (entcount-- && (tmpe != &map->header)) {
1250 			if( _vm_object_in_map(map, object, tmpe)) {
1251 				return 1;
1252 			}
1253 			tmpe = tmpe->next;
1254 		}
1255 	} else if (entry->is_sub_map || entry->is_a_map) {
1256 		tmpm = entry->object.share_map;
1257 		tmpe = tmpm->header.next;
1258 		entcount = tmpm->nentries;
1259 		while (entcount-- && tmpe != &tmpm->header) {
1260 			if( _vm_object_in_map(tmpm, object, tmpe)) {
1261 				return 1;
1262 			}
1263 			tmpe = tmpe->next;
1264 		}
1265 	} else if (obj = entry->object.vm_object) {
1266 		for(; obj; obj=obj->backing_object)
1267 			if( obj == object) {
1268 				return 1;
1269 			}
1270 	}
1271 	return 0;
1272 }
1273 
1274 int
1275 vm_object_in_map( object)
1276 	vm_object_t object;
1277 {
1278 	struct proc *p;
1279 	for (p = (struct proc *) allproc; p != NULL; p = p->p_next) {
1280 		if( !p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1281 			continue;
1282 /*
1283 		if (p->p_stat != SRUN && p->p_stat != SSLEEP) {
1284 			continue;
1285 		}
1286 */
1287 		if( _vm_object_in_map(&p->p_vmspace->vm_map, object, 0))
1288 			return 1;
1289 	}
1290 	if( _vm_object_in_map( kernel_map, object, 0))
1291 		return 1;
1292 	if( _vm_object_in_map( kmem_map, object, 0))
1293 		return 1;
1294 	if( _vm_object_in_map( pager_map, object, 0))
1295 		return 1;
1296 	if( _vm_object_in_map( buffer_map, object, 0))
1297 		return 1;
1298 	if( _vm_object_in_map( io_map, object, 0))
1299 		return 1;
1300 	if( _vm_object_in_map( phys_map, object, 0))
1301 		return 1;
1302 	if( _vm_object_in_map( mb_map, object, 0))
1303 		return 1;
1304 	if( _vm_object_in_map( u_map, object, 0))
1305 		return 1;
1306 	return 0;
1307 }
1308 
1309 
1310 void
1311 vm_object_check() {
1312 	int i;
1313 	int maxhash = 0;
1314 	vm_object_t object;
1315 
1316 	/*
1317 	 * make sure that internal objs are in a map somewhere
1318 	 * and none have zero ref counts.
1319 	 */
1320 	for (object = vm_object_list.tqh_first;
1321 			object != NULL;
1322 			object = object->object_list.tqe_next) {
1323 		if (object->handle == NULL &&
1324 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1325 			if (object->ref_count == 0) {
1326 				printf("vmochk: internal obj has zero ref count: %d\n",
1327 					object->size);
1328 			}
1329 			if (!vm_object_in_map(object)) {
1330 				printf("vmochk: internal obj is not in a map: "
1331 		"ref: %d, size: %d: 0x%x, backing_object: 0x%x\n",
1332 				    object->ref_count, object->size,
1333 				    object->size, object->backing_object);
1334 			}
1335 		}
1336 	}
1337 }
1338 
1339 /*
1340  *	vm_object_print:	[ debug ]
1341  */
1342 void
1343 vm_object_print(iobject, full, dummy3, dummy4)
1344 	/* db_expr_t */ int iobject;
1345 	boolean_t full;
1346 	/* db_expr_t */ int dummy3;
1347 	char *dummy4;
1348 {
1349 	vm_object_t object = (vm_object_t)iobject;	/* XXX */
1350 	register vm_page_t p;
1351 
1352 	register int count;
1353 
1354 	if (object == NULL)
1355 		return;
1356 
1357 	iprintf("Object 0x%x: size=0x%x, res=%d, ref=%d, ",
1358 	    (int) object, (int) object->size,
1359 	    object->resident_page_count, object->ref_count);
1360 	printf("offset=0x%x, backing_object=(0x%x)+0x%x\n",
1361 	    (int) object->paging_offset,
1362 	    (int) object->backing_object, (int) object->backing_object_offset);
1363 	printf("cache: next=%p, prev=%p\n",
1364 	    object->cached_list.tqe_next, object->cached_list.tqe_prev);
1365 
1366 	if (!full)
1367 		return;
1368 
1369 	indent += 2;
1370 	count = 0;
1371 	for (p = object->memq.tqh_first; p != NULL; p = p->listq.tqe_next) {
1372 		if (count == 0)
1373 			iprintf("memory:=");
1374 		else if (count == 6) {
1375 			printf("\n");
1376 			iprintf(" ...");
1377 			count = 0;
1378 		} else
1379 			printf(",");
1380 		count++;
1381 
1382 		printf("(off=0x%lx,page=0x%lx)",
1383 		    (u_long) p->offset, (u_long) VM_PAGE_TO_PHYS(p));
1384 	}
1385 	if (count != 0)
1386 		printf("\n");
1387 	indent -= 2;
1388 }
1389 #endif /* DDB */
1390