xref: /dflybsd-src/sys/vm/vm_object.c (revision f8c7a42d831dbc6ff46a5448591fd28de34b649b)
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  * $FreeBSD: src/sys/vm/vm_object.c,v 1.171.2.8 2003/05/26 19:17:56 alc Exp $
65  * $DragonFly: src/sys/vm/vm_object.c,v 1.27 2006/12/02 23:13:46 dillon Exp $
66  */
67 
68 /*
69  *	Virtual memory object module.
70  */
71 
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/proc.h>		/* for curproc, pageproc */
75 #include <sys/vnode.h>
76 #include <sys/vmmeter.h>
77 #include <sys/mman.h>
78 #include <sys/mount.h>
79 #include <sys/kernel.h>
80 #include <sys/sysctl.h>
81 
82 #include <vm/vm.h>
83 #include <vm/vm_param.h>
84 #include <vm/pmap.h>
85 #include <vm/vm_map.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_pager.h>
90 #include <vm/swap_pager.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_extern.h>
93 #include <vm/vm_zone.h>
94 
95 #define EASY_SCAN_FACTOR	8
96 
97 static void	vm_object_qcollapse(vm_object_t object);
98 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
99 					     int pagerflags);
100 
101 /*
102  *	Virtual memory objects maintain the actual data
103  *	associated with allocated virtual memory.  A given
104  *	page of memory exists within exactly one object.
105  *
106  *	An object is only deallocated when all "references"
107  *	are given up.  Only one "reference" to a given
108  *	region of an object should be writeable.
109  *
110  *	Associated with each object is a list of all resident
111  *	memory pages belonging to that object; this list is
112  *	maintained by the "vm_page" module, and locked by the object's
113  *	lock.
114  *
115  *	Each object also records a "pager" routine which is
116  *	used to retrieve (and store) pages to the proper backing
117  *	storage.  In addition, objects may be backed by other
118  *	objects from which they were virtual-copied.
119  *
120  *	The only items within the object structure which are
121  *	modified after time of creation are:
122  *		reference count		locked by object's lock
123  *		pager routine		locked by object's lock
124  *
125  */
126 
127 struct object_q vm_object_list;
128 static long vm_object_count;		/* count of all objects */
129 vm_object_t kernel_object;
130 vm_object_t kmem_object;
131 static struct vm_object kernel_object_store;
132 static struct vm_object kmem_object_store;
133 extern int vm_pageout_page_count;
134 
135 static long object_collapses;
136 static long object_bypasses;
137 static int next_index;
138 static vm_zone_t obj_zone;
139 static struct vm_zone obj_zone_store;
140 static int object_hash_rand;
141 #define VM_OBJECTS_INIT 256
142 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
143 
144 void
145 _vm_object_allocate(objtype_t type, vm_size_t size, vm_object_t object)
146 {
147 	int incr;
148 	RB_INIT(&object->rb_memq);
149 	LIST_INIT(&object->shadow_head);
150 
151 	object->type = type;
152 	object->size = size;
153 	object->ref_count = 1;
154 	object->flags = 0;
155 	if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
156 		vm_object_set_flag(object, OBJ_ONEMAPPING);
157 	object->paging_in_progress = 0;
158 	object->resident_page_count = 0;
159 	object->shadow_count = 0;
160 	object->pg_color = next_index;
161 	if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
162 		incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
163 	else
164 		incr = size;
165 	next_index = (next_index + incr) & PQ_L2_MASK;
166 	object->handle = NULL;
167 	object->backing_object = NULL;
168 	object->backing_object_offset = (vm_ooffset_t) 0;
169 	/*
170 	 * Try to generate a number that will spread objects out in the
171 	 * hash table.  We 'wipe' new objects across the hash in 128 page
172 	 * increments plus 1 more to offset it a little more by the time
173 	 * it wraps around.
174 	 */
175 	object->hash_rand = object_hash_rand - 129;
176 
177 	object->generation++;
178 
179 	crit_enter();
180 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
181 	vm_object_count++;
182 	object_hash_rand = object->hash_rand;
183 	crit_exit();
184 }
185 
186 /*
187  *	vm_object_init:
188  *
189  *	Initialize the VM objects module.
190  */
191 void
192 vm_object_init(void)
193 {
194 	TAILQ_INIT(&vm_object_list);
195 
196 	kernel_object = &kernel_object_store;
197 	_vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
198 	    kernel_object);
199 
200 	kmem_object = &kmem_object_store;
201 	_vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
202 	    kmem_object);
203 
204 	obj_zone = &obj_zone_store;
205 	zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
206 		vm_objects_init, VM_OBJECTS_INIT);
207 }
208 
209 void
210 vm_object_init2(void)
211 {
212 	zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1);
213 }
214 
215 /*
216  *	vm_object_allocate:
217  *
218  *	Returns a new object with the given size.
219  */
220 
221 vm_object_t
222 vm_object_allocate(objtype_t type, vm_size_t size)
223 {
224 	vm_object_t result;
225 
226 	result = (vm_object_t) zalloc(obj_zone);
227 
228 	_vm_object_allocate(type, size, result);
229 
230 	return (result);
231 }
232 
233 
234 /*
235  *	vm_object_reference:
236  *
237  *	Gets another reference to the given object.
238  */
239 void
240 vm_object_reference(vm_object_t object)
241 {
242 	if (object == NULL)
243 		return;
244 
245 	object->ref_count++;
246 	if (object->type == OBJT_VNODE) {
247 		vref(object->handle);
248 		/* XXX what if the vnode is being destroyed? */
249 	}
250 }
251 
252 static void
253 vm_object_vndeallocate(vm_object_t object)
254 {
255 	struct vnode *vp = (struct vnode *) object->handle;
256 
257 	KASSERT(object->type == OBJT_VNODE,
258 	    ("vm_object_vndeallocate: not a vnode object"));
259 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
260 #ifdef INVARIANTS
261 	if (object->ref_count == 0) {
262 		vprint("vm_object_vndeallocate", vp);
263 		panic("vm_object_vndeallocate: bad object reference count");
264 	}
265 #endif
266 
267 	object->ref_count--;
268 	if (object->ref_count == 0)
269 		vp->v_flag &= ~VTEXT;
270 	vrele(vp);
271 }
272 
273 /*
274  *	vm_object_deallocate:
275  *
276  *	Release a reference to the specified object,
277  *	gained either through a vm_object_allocate
278  *	or a vm_object_reference call.  When all references
279  *	are gone, storage associated with this object
280  *	may be relinquished.
281  *
282  *	No object may be locked.
283  */
284 void
285 vm_object_deallocate(vm_object_t object)
286 {
287 	vm_object_t temp;
288 
289 	while (object != NULL) {
290 		if (object->type == OBJT_VNODE) {
291 			vm_object_vndeallocate(object);
292 			return;
293 		}
294 
295 		if (object->ref_count == 0) {
296 			panic("vm_object_deallocate: object deallocated too many times: %d", object->type);
297 		} else if (object->ref_count > 2) {
298 			object->ref_count--;
299 			return;
300 		}
301 
302 		/*
303 		 * Here on ref_count of one or two, which are special cases for
304 		 * objects.
305 		 */
306 		if ((object->ref_count == 2) && (object->shadow_count == 0)) {
307 			vm_object_set_flag(object, OBJ_ONEMAPPING);
308 			object->ref_count--;
309 			return;
310 		} else if ((object->ref_count == 2) && (object->shadow_count == 1)) {
311 			object->ref_count--;
312 			if ((object->handle == NULL) &&
313 			    (object->type == OBJT_DEFAULT ||
314 			     object->type == OBJT_SWAP)) {
315 				vm_object_t robject;
316 
317 				robject = LIST_FIRST(&object->shadow_head);
318 				KASSERT(robject != NULL,
319 				    ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
320 					 object->ref_count,
321 					 object->shadow_count));
322 				if ((robject->handle == NULL) &&
323 				    (robject->type == OBJT_DEFAULT ||
324 				     robject->type == OBJT_SWAP)) {
325 
326 					robject->ref_count++;
327 
328 					while (
329 						robject->paging_in_progress ||
330 						object->paging_in_progress
331 					) {
332 						vm_object_pip_sleep(robject, "objde1");
333 						vm_object_pip_sleep(object, "objde2");
334 					}
335 
336 					if (robject->ref_count == 1) {
337 						robject->ref_count--;
338 						object = robject;
339 						goto doterm;
340 					}
341 
342 					object = robject;
343 					vm_object_collapse(object);
344 					continue;
345 				}
346 			}
347 
348 			return;
349 
350 		} else {
351 			object->ref_count--;
352 			if (object->ref_count != 0)
353 				return;
354 		}
355 
356 doterm:
357 
358 		temp = object->backing_object;
359 		if (temp) {
360 			LIST_REMOVE(object, shadow_list);
361 			temp->shadow_count--;
362 			temp->generation++;
363 			object->backing_object = NULL;
364 		}
365 
366 		/*
367 		 * Don't double-terminate, we could be in a termination
368 		 * recursion due to the terminate having to sync data
369 		 * to disk.
370 		 */
371 		if ((object->flags & OBJ_DEAD) == 0)
372 			vm_object_terminate(object);
373 		object = temp;
374 	}
375 }
376 
377 /*
378  *	vm_object_terminate actually destroys the specified object, freeing
379  *	up all previously used resources.
380  *
381  *	The object must be locked.
382  *	This routine may block.
383  */
384 static int vm_object_terminate_callback(vm_page_t p, void *data);
385 
386 void
387 vm_object_terminate(vm_object_t object)
388 {
389 	/*
390 	 * Make sure no one uses us.
391 	 */
392 	vm_object_set_flag(object, OBJ_DEAD);
393 
394 	/*
395 	 * wait for the pageout daemon to be done with the object
396 	 */
397 	vm_object_pip_wait(object, "objtrm");
398 
399 	KASSERT(!object->paging_in_progress,
400 		("vm_object_terminate: pageout in progress"));
401 
402 	/*
403 	 * Clean and free the pages, as appropriate. All references to the
404 	 * object are gone, so we don't need to lock it.
405 	 */
406 	if (object->type == OBJT_VNODE) {
407 		struct vnode *vp;
408 
409 		/*
410 		 * Clean pages and flush buffers.
411 		 */
412 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
413 
414 		vp = (struct vnode *) object->handle;
415 		vinvalbuf(vp, V_SAVE, 0, 0);
416 	}
417 
418 	/*
419 	 * Wait for any I/O to complete, after which there had better not
420 	 * be any references left on the object.
421 	 */
422 	vm_object_pip_wait(object, "objtrm");
423 
424 	if (object->ref_count != 0)
425 		panic("vm_object_terminate: object with references, ref_count=%d", object->ref_count);
426 
427 	/*
428 	 * Now free any remaining pages. For internal objects, this also
429 	 * removes them from paging queues. Don't free wired pages, just
430 	 * remove them from the object.
431 	 */
432 	crit_enter();
433 	vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
434 				vm_object_terminate_callback, NULL);
435 	crit_exit();
436 
437 	/*
438 	 * Let the pager know object is dead.
439 	 */
440 	vm_pager_deallocate(object);
441 
442 	/*
443 	 * Remove the object from the global object list.
444 	 */
445 	crit_enter();
446 	TAILQ_REMOVE(&vm_object_list, object, object_list);
447 	vm_object_count--;
448 	crit_exit();
449 
450 	wakeup(object);
451 	if (object->ref_count != 0)
452 		panic("vm_object_terminate2: object with references, ref_count=%d", object->ref_count);
453 
454 	/*
455 	 * Free the space for the object.
456 	 */
457 	zfree(obj_zone, object);
458 }
459 
460 static int
461 vm_object_terminate_callback(vm_page_t p, void *data __unused)
462 {
463 	if (p->busy || (p->flags & PG_BUSY))
464 		panic("vm_object_terminate: freeing busy page %p", p);
465 	if (p->wire_count == 0) {
466 		vm_page_busy(p);
467 		vm_page_free(p);
468 		mycpu->gd_cnt.v_pfree++;
469 	} else {
470 		vm_page_busy(p);
471 		vm_page_remove(p);
472 		vm_page_wakeup(p);
473 	}
474 	return(0);
475 }
476 
477 /*
478  *	vm_object_page_clean
479  *
480  *	Clean all dirty pages in the specified range of object.  Leaves page
481  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
482  *	write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
483  *	leaving the object dirty.
484  *
485  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
486  *	synchronous clustering mode implementation.
487  *
488  *	Odd semantics: if start == end, we clean everything.
489  */
490 static int vm_object_page_clean_pass1(struct vm_page *p, void *data);
491 static int vm_object_page_clean_pass2(struct vm_page *p, void *data);
492 
493 void
494 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
495 		     int flags)
496 {
497 	struct rb_vm_page_scan_info info;
498 	struct vnode *vp;
499 	int wholescan;
500 	int pagerflags;
501 	int curgeneration;
502 
503 	if (object->type != OBJT_VNODE ||
504 		(object->flags & OBJ_MIGHTBEDIRTY) == 0)
505 		return;
506 
507 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ?
508 			VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
509 	pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
510 
511 	vp = object->handle;
512 
513 	/*
514 	 * Interlock other major object operations.  This allows us to
515 	 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY.
516 	 */
517 	crit_enter();
518 	vm_object_set_flag(object, OBJ_CLEANING);
519 
520 	/*
521 	 * Handle 'entire object' case
522 	 */
523 	info.start_pindex = start;
524 	if (end == 0) {
525 		info.end_pindex = object->size - 1;
526 	} else {
527 		info.end_pindex = end - 1;
528 	}
529 	wholescan = (start == 0 && info.end_pindex == object->size - 1);
530 	info.limit = flags;
531 	info.pagerflags = pagerflags;
532 	info.object = object;
533 
534 	/*
535 	 * If cleaning the entire object do a pass to mark the pages read-only.
536 	 * If everything worked out ok, clear OBJ_WRITEABLE and
537 	 * OBJ_MIGHTBEDIRTY.
538 	 */
539 	if (wholescan) {
540 		info.error = 0;
541 		vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
542 					vm_object_page_clean_pass1, &info);
543 		if (info.error == 0) {
544 			vm_object_clear_flag(object,
545 					     OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
546 			if (object->type == OBJT_VNODE &&
547 			    (vp = (struct vnode *)object->handle) != NULL) {
548 				if (vp->v_flag & VOBJDIRTY)
549 					vclrflags(vp, VOBJDIRTY);
550 			}
551 		}
552 	}
553 
554 	/*
555 	 * Do a pass to clean all the dirty pages we find.
556 	 */
557 	do {
558 		info.error = 0;
559 		curgeneration = object->generation;
560 		vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
561 					vm_object_page_clean_pass2, &info);
562 	} while (info.error || curgeneration != object->generation);
563 
564 	vm_object_clear_flag(object, OBJ_CLEANING);
565 	crit_exit();
566 }
567 
568 static
569 int
570 vm_object_page_clean_pass1(struct vm_page *p, void *data)
571 {
572 	struct rb_vm_page_scan_info *info = data;
573 
574 	vm_page_flag_set(p, PG_CLEANCHK);
575 	if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
576 		info->error = 1;
577 	else
578 		vm_page_protect(p, VM_PROT_READ);
579 	return(0);
580 }
581 
582 static
583 int
584 vm_object_page_clean_pass2(struct vm_page *p, void *data)
585 {
586 	struct rb_vm_page_scan_info *info = data;
587 	int n;
588 
589 	/*
590 	 * Do not mess with pages that were inserted after we started
591 	 * the cleaning pass.
592 	 */
593 	if ((p->flags & PG_CLEANCHK) == 0)
594 		return(0);
595 
596 	/*
597 	 * Before wasting time traversing the pmaps, check for trivial
598 	 * cases where the page cannot be dirty.
599 	 */
600 	if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) {
601 		KKASSERT((p->dirty & p->valid) == 0);
602 		return(0);
603 	}
604 
605 	/*
606 	 * Check whether the page is dirty or not.  The page has been set
607 	 * to be read-only so the check will not race a user dirtying the
608 	 * page.
609 	 */
610 	vm_page_test_dirty(p);
611 	if ((p->dirty & p->valid) == 0) {
612 		vm_page_flag_clear(p, PG_CLEANCHK);
613 		return(0);
614 	}
615 
616 	/*
617 	 * If we have been asked to skip nosync pages and this is a
618 	 * nosync page, skip it.  Note that the object flags were
619 	 * not cleared in this case (because pass1 will have returned an
620 	 * error), so we do not have to set them.
621 	 */
622 	if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
623 		vm_page_flag_clear(p, PG_CLEANCHK);
624 		return(0);
625 	}
626 
627 	/*
628 	 * Flush as many pages as we can.  PG_CLEANCHK will be cleared on
629 	 * the pages that get successfully flushed.  Set info->error if
630 	 * we raced an object modification.
631 	 */
632 	n = vm_object_page_collect_flush(info->object, p, info->pagerflags);
633 	if (n == 0)
634 		info->error = 1;
635 	return(0);
636 }
637 
638 /*
639  * This routine must be called within a critical section to properly avoid
640  * an interrupt unbusy/free race that can occur prior to the busy check.
641  *
642  * Using the object generation number here to detect page ripout is not
643  * the best idea in the world. XXX
644  *
645  * NOTE: we operate under the assumption that a page found to not be busy
646  * will not be ripped out from under us by an interrupt.  XXX we should
647  * recode this to explicitly busy the pages.
648  */
649 static int
650 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags)
651 {
652 	int runlen;
653 	int maxf;
654 	int chkb;
655 	int maxb;
656 	int i;
657 	int curgeneration;
658 	vm_pindex_t pi;
659 	vm_page_t maf[vm_pageout_page_count];
660 	vm_page_t mab[vm_pageout_page_count];
661 	vm_page_t ma[vm_pageout_page_count];
662 
663 	curgeneration = object->generation;
664 
665 	pi = p->pindex;
666 	while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
667 		if (object->generation != curgeneration) {
668 			return(0);
669 		}
670 	}
671 	KKASSERT(p->object == object && p->pindex == pi);
672 
673 	maxf = 0;
674 	for(i = 1; i < vm_pageout_page_count; i++) {
675 		vm_page_t tp;
676 
677 		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
678 			if ((tp->flags & PG_BUSY) ||
679 				((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
680 				 (tp->flags & PG_CLEANCHK) == 0) ||
681 				(tp->busy != 0))
682 				break;
683 			if((tp->queue - tp->pc) == PQ_CACHE) {
684 				vm_page_flag_clear(tp, PG_CLEANCHK);
685 				break;
686 			}
687 			vm_page_test_dirty(tp);
688 			if ((tp->dirty & tp->valid) == 0) {
689 				vm_page_flag_clear(tp, PG_CLEANCHK);
690 				break;
691 			}
692 			maf[ i - 1 ] = tp;
693 			maxf++;
694 			continue;
695 		}
696 		break;
697 	}
698 
699 	maxb = 0;
700 	chkb = vm_pageout_page_count -  maxf;
701 	if (chkb) {
702 		for(i = 1; i < chkb;i++) {
703 			vm_page_t tp;
704 
705 			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
706 				if ((tp->flags & PG_BUSY) ||
707 					((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
708 					 (tp->flags & PG_CLEANCHK) == 0) ||
709 					(tp->busy != 0))
710 					break;
711 				if((tp->queue - tp->pc) == PQ_CACHE) {
712 					vm_page_flag_clear(tp, PG_CLEANCHK);
713 					break;
714 				}
715 				vm_page_test_dirty(tp);
716 				if ((tp->dirty & tp->valid) == 0) {
717 					vm_page_flag_clear(tp, PG_CLEANCHK);
718 					break;
719 				}
720 				mab[ i - 1 ] = tp;
721 				maxb++;
722 				continue;
723 			}
724 			break;
725 		}
726 	}
727 
728 	for(i = 0; i < maxb; i++) {
729 		int index = (maxb - i) - 1;
730 		ma[index] = mab[i];
731 		vm_page_flag_clear(ma[index], PG_CLEANCHK);
732 	}
733 	vm_page_flag_clear(p, PG_CLEANCHK);
734 	ma[maxb] = p;
735 	for(i = 0; i < maxf; i++) {
736 		int index = (maxb + i) + 1;
737 		ma[index] = maf[i];
738 		vm_page_flag_clear(ma[index], PG_CLEANCHK);
739 	}
740 	runlen = maxb + maxf + 1;
741 
742 	vm_pageout_flush(ma, runlen, pagerflags);
743 	for (i = 0; i < runlen; i++) {
744 		if (ma[i]->valid & ma[i]->dirty) {
745 			vm_page_protect(ma[i], VM_PROT_READ);
746 			vm_page_flag_set(ma[i], PG_CLEANCHK);
747 
748 			/*
749 			 * maxf will end up being the actual number of pages
750 			 * we wrote out contiguously, non-inclusive of the
751 			 * first page.  We do not count look-behind pages.
752 			 */
753 			if (i >= maxb + 1 && (maxf > i - maxb - 1))
754 				maxf = i - maxb - 1;
755 		}
756 	}
757 	return(maxf + 1);
758 }
759 
760 #ifdef not_used
761 /* XXX I cannot tell if this should be an exported symbol */
762 /*
763  *	vm_object_deactivate_pages
764  *
765  *	Deactivate all pages in the specified object.  (Keep its pages
766  *	in memory even though it is no longer referenced.)
767  *
768  *	The object must be locked.
769  */
770 static int vm_object_deactivate_pages_callback(vm_page_t p, void *data);
771 
772 static void
773 vm_object_deactivate_pages(vm_object_t object)
774 {
775 	crit_enter();
776 	vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
777 				vm_object_deactivate_pages_callback, NULL);
778 	crit_exit();
779 }
780 
781 static int
782 vm_object_deactivate_pages_callback(vm_page_t p, void *data __unused)
783 {
784 	vm_page_deactivate(p);
785 	return(0);
786 }
787 
788 #endif
789 
790 /*
791  * Same as vm_object_pmap_copy, except range checking really
792  * works, and is meant for small sections of an object.
793  *
794  * This code protects resident pages by making them read-only
795  * and is typically called on a fork or split when a page
796  * is converted to copy-on-write.
797  *
798  * NOTE: If the page is already at VM_PROT_NONE, calling
799  * vm_page_protect will have no effect.
800  */
801 void
802 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
803 {
804 	vm_pindex_t idx;
805 	vm_page_t p;
806 
807 	if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
808 		return;
809 
810 	/*
811 	 * spl protection needed to prevent races between the lookup,
812 	 * an interrupt unbusy/free, and our protect call.
813 	 */
814 	crit_enter();
815 	for (idx = start; idx < end; idx++) {
816 		p = vm_page_lookup(object, idx);
817 		if (p == NULL)
818 			continue;
819 		vm_page_protect(p, VM_PROT_READ);
820 	}
821 	crit_exit();
822 }
823 
824 /*
825  *	vm_object_pmap_remove:
826  *
827  *	Removes all physical pages in the specified
828  *	object range from all physical maps.
829  *
830  *	The object must *not* be locked.
831  */
832 
833 static int vm_object_pmap_remove_callback(vm_page_t p, void *data);
834 
835 void
836 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
837 {
838 	struct rb_vm_page_scan_info info;
839 
840 	if (object == NULL)
841 		return;
842 	info.start_pindex = start;
843 	info.end_pindex = end - 1;
844 	crit_enter();
845 	vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
846 				vm_object_pmap_remove_callback, &info);
847 	if (start == 0 && end == object->size)
848 		vm_object_clear_flag(object, OBJ_WRITEABLE);
849 	crit_exit();
850 }
851 
852 static int
853 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused)
854 {
855 	vm_page_protect(p, VM_PROT_NONE);
856 	return(0);
857 }
858 
859 /*
860  *	vm_object_madvise:
861  *
862  *	Implements the madvise function at the object/page level.
863  *
864  *	MADV_WILLNEED	(any object)
865  *
866  *	    Activate the specified pages if they are resident.
867  *
868  *	MADV_DONTNEED	(any object)
869  *
870  *	    Deactivate the specified pages if they are resident.
871  *
872  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
873  *			 OBJ_ONEMAPPING only)
874  *
875  *	    Deactivate and clean the specified pages if they are
876  *	    resident.  This permits the process to reuse the pages
877  *	    without faulting or the kernel to reclaim the pages
878  *	    without I/O.
879  */
880 void
881 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
882 {
883 	vm_pindex_t end, tpindex;
884 	vm_object_t tobject;
885 	vm_page_t m;
886 
887 	if (object == NULL)
888 		return;
889 
890 	end = pindex + count;
891 
892 	/*
893 	 * Locate and adjust resident pages
894 	 */
895 
896 	for (; pindex < end; pindex += 1) {
897 relookup:
898 		tobject = object;
899 		tpindex = pindex;
900 shadowlookup:
901 		/*
902 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
903 		 * and those pages must be OBJ_ONEMAPPING.
904 		 */
905 		if (advise == MADV_FREE) {
906 			if ((tobject->type != OBJT_DEFAULT &&
907 			     tobject->type != OBJT_SWAP) ||
908 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
909 				continue;
910 			}
911 		}
912 
913 		/*
914 		 * spl protection is required to avoid a race between the
915 		 * lookup, an interrupt unbusy/free, and our busy check.
916 		 */
917 
918 		crit_enter();
919 		m = vm_page_lookup(tobject, tpindex);
920 
921 		if (m == NULL) {
922 			/*
923 			 * There may be swap even if there is no backing page
924 			 */
925 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
926 				swap_pager_freespace(tobject, tpindex, 1);
927 
928 			/*
929 			 * next object
930 			 */
931 			crit_exit();
932 			if (tobject->backing_object == NULL)
933 				continue;
934 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
935 			tobject = tobject->backing_object;
936 			goto shadowlookup;
937 		}
938 
939 		/*
940 		 * If the page is busy or not in a normal active state,
941 		 * we skip it.  If the page is not managed there are no
942 		 * page queues to mess with.  Things can break if we mess
943 		 * with pages in any of the below states.
944 		 */
945 		if (
946 		    m->hold_count ||
947 		    m->wire_count ||
948 		    (m->flags & PG_UNMANAGED) ||
949 		    m->valid != VM_PAGE_BITS_ALL
950 		) {
951 			crit_exit();
952 			continue;
953 		}
954 
955  		if (vm_page_sleep_busy(m, TRUE, "madvpo")) {
956 			crit_exit();
957   			goto relookup;
958 		}
959 		crit_exit();
960 
961 		/*
962 		 * Theoretically once a page is known not to be busy, an
963 		 * interrupt cannot come along and rip it out from under us.
964 		 */
965 
966 		if (advise == MADV_WILLNEED) {
967 			vm_page_activate(m);
968 		} else if (advise == MADV_DONTNEED) {
969 			vm_page_dontneed(m);
970 		} else if (advise == MADV_FREE) {
971 			/*
972 			 * Mark the page clean.  This will allow the page
973 			 * to be freed up by the system.  However, such pages
974 			 * are often reused quickly by malloc()/free()
975 			 * so we do not do anything that would cause
976 			 * a page fault if we can help it.
977 			 *
978 			 * Specifically, we do not try to actually free
979 			 * the page now nor do we try to put it in the
980 			 * cache (which would cause a page fault on reuse).
981 			 *
982 			 * But we do make the page is freeable as we
983 			 * can without actually taking the step of unmapping
984 			 * it.
985 			 */
986 			pmap_clear_modify(m);
987 			m->dirty = 0;
988 			m->act_count = 0;
989 			vm_page_dontneed(m);
990 			if (tobject->type == OBJT_SWAP)
991 				swap_pager_freespace(tobject, tpindex, 1);
992 		}
993 	}
994 }
995 
996 /*
997  *	vm_object_shadow:
998  *
999  *	Create a new object which is backed by the
1000  *	specified existing object range.  The source
1001  *	object reference is deallocated.
1002  *
1003  *	The new object and offset into that object
1004  *	are returned in the source parameters.
1005  */
1006 
1007 void
1008 vm_object_shadow(vm_object_t *object,	/* IN/OUT */
1009 		 vm_ooffset_t *offset,	/* IN/OUT */
1010 		 vm_size_t length)
1011 {
1012 	vm_object_t source;
1013 	vm_object_t result;
1014 
1015 	source = *object;
1016 
1017 	/*
1018 	 * Don't create the new object if the old object isn't shared.
1019 	 */
1020 
1021 	if (source != NULL &&
1022 	    source->ref_count == 1 &&
1023 	    source->handle == NULL &&
1024 	    (source->type == OBJT_DEFAULT ||
1025 	     source->type == OBJT_SWAP))
1026 		return;
1027 
1028 	/*
1029 	 * Allocate a new object with the given length
1030 	 */
1031 
1032 	if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1033 		panic("vm_object_shadow: no object for shadowing");
1034 
1035 	/*
1036 	 * The new object shadows the source object, adding a reference to it.
1037 	 * Our caller changes his reference to point to the new object,
1038 	 * removing a reference to the source object.  Net result: no change
1039 	 * of reference count.
1040 	 *
1041 	 * Try to optimize the result object's page color when shadowing
1042 	 * in order to maintain page coloring consistency in the combined
1043 	 * shadowed object.
1044 	 */
1045 	result->backing_object = source;
1046 	if (source) {
1047 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1048 		source->shadow_count++;
1049 		source->generation++;
1050 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1051 	}
1052 
1053 	/*
1054 	 * Store the offset into the source object, and fix up the offset into
1055 	 * the new object.
1056 	 */
1057 
1058 	result->backing_object_offset = *offset;
1059 
1060 	/*
1061 	 * Return the new things
1062 	 */
1063 
1064 	*offset = 0;
1065 	*object = result;
1066 }
1067 
1068 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1069 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1070 #define	OBSC_COLLAPSE_WAIT	0x0004
1071 
1072 static int vm_object_backing_scan_callback(vm_page_t p, void *data);
1073 
1074 static __inline int
1075 vm_object_backing_scan(vm_object_t object, int op)
1076 {
1077 	struct rb_vm_page_scan_info info;
1078 	vm_object_t backing_object;
1079 
1080 	/*
1081 	 * spl protection is required to avoid races between the memq/lookup,
1082 	 * an interrupt doing an unbusy/free, and our busy check.  Amoung
1083 	 * other things.
1084 	 */
1085 	crit_enter();
1086 
1087 	backing_object = object->backing_object;
1088 	info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1089 
1090 	/*
1091 	 * Initial conditions
1092 	 */
1093 
1094 	if (op & OBSC_TEST_ALL_SHADOWED) {
1095 		/*
1096 		 * We do not want to have to test for the existence of
1097 		 * swap pages in the backing object.  XXX but with the
1098 		 * new swapper this would be pretty easy to do.
1099 		 *
1100 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1101 		 * been ZFOD faulted yet?  If we do not test for this, the
1102 		 * shadow test may succeed! XXX
1103 		 */
1104 		if (backing_object->type != OBJT_DEFAULT) {
1105 			crit_exit();
1106 			return(0);
1107 		}
1108 	}
1109 	if (op & OBSC_COLLAPSE_WAIT) {
1110 		KKASSERT((backing_object->flags & OBJ_DEAD) == 0);
1111 		vm_object_set_flag(backing_object, OBJ_DEAD);
1112 	}
1113 
1114 	/*
1115 	 * Our scan.   We have to retry if a negative error code is returned,
1116 	 * otherwise 0 or 1 will be returned in info.error.  0 Indicates that
1117 	 * the scan had to be stopped because the parent does not completely
1118 	 * shadow the child.
1119 	 */
1120 	info.object = object;
1121 	info.backing_object = backing_object;
1122 	info.limit = op;
1123 	do {
1124 		info.error = 1;
1125 		vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL,
1126 					vm_object_backing_scan_callback,
1127 					&info);
1128 	} while (info.error < 0);
1129 	crit_exit();
1130 	return(info.error);
1131 }
1132 
1133 static int
1134 vm_object_backing_scan_callback(vm_page_t p, void *data)
1135 {
1136 	struct rb_vm_page_scan_info *info = data;
1137 	vm_object_t backing_object;
1138 	vm_object_t object;
1139 	vm_pindex_t new_pindex;
1140 	vm_pindex_t backing_offset_index;
1141 	int op;
1142 
1143 	new_pindex = p->pindex - info->backing_offset_index;
1144 	op = info->limit;
1145 	object = info->object;
1146 	backing_object = info->backing_object;
1147 	backing_offset_index = info->backing_offset_index;
1148 
1149 	if (op & OBSC_TEST_ALL_SHADOWED) {
1150 		vm_page_t pp;
1151 
1152 		/*
1153 		 * Ignore pages outside the parent object's range
1154 		 * and outside the parent object's mapping of the
1155 		 * backing object.
1156 		 *
1157 		 * note that we do not busy the backing object's
1158 		 * page.
1159 		 */
1160 		if (
1161 		    p->pindex < backing_offset_index ||
1162 		    new_pindex >= object->size
1163 		) {
1164 			return(0);
1165 		}
1166 
1167 		/*
1168 		 * See if the parent has the page or if the parent's
1169 		 * object pager has the page.  If the parent has the
1170 		 * page but the page is not valid, the parent's
1171 		 * object pager must have the page.
1172 		 *
1173 		 * If this fails, the parent does not completely shadow
1174 		 * the object and we might as well give up now.
1175 		 */
1176 
1177 		pp = vm_page_lookup(object, new_pindex);
1178 		if (
1179 		    (pp == NULL || pp->valid == 0) &&
1180 		    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1181 		) {
1182 			info->error = 0;	/* problemo */
1183 			return(-1);		/* stop the scan */
1184 		}
1185 	}
1186 
1187 	/*
1188 	 * Check for busy page
1189 	 */
1190 
1191 	if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1192 		vm_page_t pp;
1193 
1194 		if (op & OBSC_COLLAPSE_NOWAIT) {
1195 			if (
1196 			    (p->flags & PG_BUSY) ||
1197 			    !p->valid ||
1198 			    p->hold_count ||
1199 			    p->wire_count ||
1200 			    p->busy
1201 			) {
1202 				return(0);
1203 			}
1204 		} else if (op & OBSC_COLLAPSE_WAIT) {
1205 			if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1206 				/*
1207 				 * If we slept, anything could have
1208 				 * happened.   Ask that the scan be restarted.
1209 				 *
1210 				 * Since the object is marked dead, the
1211 				 * backing offset should not have changed.
1212 				 */
1213 				info->error = -1;
1214 				return(-1);
1215 			}
1216 		}
1217 
1218 		/*
1219 		 * Busy the page
1220 		 */
1221 		vm_page_busy(p);
1222 
1223 		KASSERT(
1224 		    p->object == backing_object,
1225 		    ("vm_object_qcollapse(): object mismatch")
1226 		);
1227 
1228 		/*
1229 		 * Destroy any associated swap
1230 		 */
1231 		if (backing_object->type == OBJT_SWAP) {
1232 			swap_pager_freespace(
1233 			    backing_object,
1234 			    p->pindex,
1235 			    1
1236 			);
1237 		}
1238 
1239 		if (
1240 		    p->pindex < backing_offset_index ||
1241 		    new_pindex >= object->size
1242 		) {
1243 			/*
1244 			 * Page is out of the parent object's range, we
1245 			 * can simply destroy it.
1246 			 */
1247 			vm_page_protect(p, VM_PROT_NONE);
1248 			vm_page_free(p);
1249 			return(0);
1250 		}
1251 
1252 		pp = vm_page_lookup(object, new_pindex);
1253 		if (
1254 		    pp != NULL ||
1255 		    vm_pager_has_page(object, new_pindex, NULL, NULL)
1256 		) {
1257 			/*
1258 			 * page already exists in parent OR swap exists
1259 			 * for this location in the parent.  Destroy
1260 			 * the original page from the backing object.
1261 			 *
1262 			 * Leave the parent's page alone
1263 			 */
1264 			vm_page_protect(p, VM_PROT_NONE);
1265 			vm_page_free(p);
1266 			return(0);
1267 		}
1268 
1269 		/*
1270 		 * Page does not exist in parent, rename the
1271 		 * page from the backing object to the main object.
1272 		 *
1273 		 * If the page was mapped to a process, it can remain
1274 		 * mapped through the rename.
1275 		 */
1276 		if ((p->queue - p->pc) == PQ_CACHE)
1277 			vm_page_deactivate(p);
1278 
1279 		vm_page_rename(p, object, new_pindex);
1280 		/* page automatically made dirty by rename */
1281 	}
1282 	return(0);
1283 }
1284 
1285 /*
1286  * this version of collapse allows the operation to occur earlier and
1287  * when paging_in_progress is true for an object...  This is not a complete
1288  * operation, but should plug 99.9% of the rest of the leaks.
1289  */
1290 static void
1291 vm_object_qcollapse(vm_object_t object)
1292 {
1293 	vm_object_t backing_object = object->backing_object;
1294 
1295 	if (backing_object->ref_count != 1)
1296 		return;
1297 
1298 	backing_object->ref_count += 2;
1299 
1300 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1301 
1302 	backing_object->ref_count -= 2;
1303 }
1304 
1305 /*
1306  *	vm_object_collapse:
1307  *
1308  *	Collapse an object with the object backing it.
1309  *	Pages in the backing object are moved into the
1310  *	parent, and the backing object is deallocated.
1311  */
1312 void
1313 vm_object_collapse(vm_object_t object)
1314 {
1315 	while (TRUE) {
1316 		vm_object_t backing_object;
1317 
1318 		/*
1319 		 * Verify that the conditions are right for collapse:
1320 		 *
1321 		 * The object exists and the backing object exists.
1322 		 */
1323 		if (object == NULL)
1324 			break;
1325 
1326 		if ((backing_object = object->backing_object) == NULL)
1327 			break;
1328 
1329 		/*
1330 		 * we check the backing object first, because it is most likely
1331 		 * not collapsable.
1332 		 */
1333 		if (backing_object->handle != NULL ||
1334 		    (backing_object->type != OBJT_DEFAULT &&
1335 		     backing_object->type != OBJT_SWAP) ||
1336 		    (backing_object->flags & OBJ_DEAD) ||
1337 		    object->handle != NULL ||
1338 		    (object->type != OBJT_DEFAULT &&
1339 		     object->type != OBJT_SWAP) ||
1340 		    (object->flags & OBJ_DEAD)) {
1341 			break;
1342 		}
1343 
1344 		if (
1345 		    object->paging_in_progress != 0 ||
1346 		    backing_object->paging_in_progress != 0
1347 		) {
1348 			vm_object_qcollapse(object);
1349 			break;
1350 		}
1351 
1352 		/*
1353 		 * We know that we can either collapse the backing object (if
1354 		 * the parent is the only reference to it) or (perhaps) have
1355 		 * the parent bypass the object if the parent happens to shadow
1356 		 * all the resident pages in the entire backing object.
1357 		 *
1358 		 * This is ignoring pager-backed pages such as swap pages.
1359 		 * vm_object_backing_scan fails the shadowing test in this
1360 		 * case.
1361 		 */
1362 
1363 		if (backing_object->ref_count == 1) {
1364 			/*
1365 			 * If there is exactly one reference to the backing
1366 			 * object, we can collapse it into the parent.
1367 			 */
1368 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1369 
1370 			/*
1371 			 * Move the pager from backing_object to object.
1372 			 */
1373 
1374 			if (backing_object->type == OBJT_SWAP) {
1375 				vm_object_pip_add(backing_object, 1);
1376 
1377 				/*
1378 				 * scrap the paging_offset junk and do a
1379 				 * discrete copy.  This also removes major
1380 				 * assumptions about how the swap-pager
1381 				 * works from where it doesn't belong.  The
1382 				 * new swapper is able to optimize the
1383 				 * destroy-source case.
1384 				 */
1385 
1386 				vm_object_pip_add(object, 1);
1387 				swap_pager_copy(
1388 				    backing_object,
1389 				    object,
1390 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1391 				vm_object_pip_wakeup(object);
1392 
1393 				vm_object_pip_wakeup(backing_object);
1394 			}
1395 			/*
1396 			 * Object now shadows whatever backing_object did.
1397 			 * Note that the reference to
1398 			 * backing_object->backing_object moves from within
1399 			 * backing_object to within object.
1400 			 */
1401 
1402 			LIST_REMOVE(object, shadow_list);
1403 			object->backing_object->shadow_count--;
1404 			object->backing_object->generation++;
1405 			if (backing_object->backing_object) {
1406 				LIST_REMOVE(backing_object, shadow_list);
1407 				backing_object->backing_object->shadow_count--;
1408 				backing_object->backing_object->generation++;
1409 			}
1410 			object->backing_object = backing_object->backing_object;
1411 			if (object->backing_object) {
1412 				LIST_INSERT_HEAD(
1413 				    &object->backing_object->shadow_head,
1414 				    object,
1415 				    shadow_list
1416 				);
1417 				object->backing_object->shadow_count++;
1418 				object->backing_object->generation++;
1419 			}
1420 
1421 			object->backing_object_offset +=
1422 			    backing_object->backing_object_offset;
1423 
1424 			/*
1425 			 * Discard backing_object.
1426 			 *
1427 			 * Since the backing object has no pages, no pager left,
1428 			 * and no object references within it, all that is
1429 			 * necessary is to dispose of it.
1430 			 */
1431 
1432 			KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1433 			KASSERT(RB_EMPTY(&backing_object->rb_memq), ("backing_object %p somehow has left over pages during collapse!", backing_object));
1434 			crit_enter();
1435 			TAILQ_REMOVE(
1436 			    &vm_object_list,
1437 			    backing_object,
1438 			    object_list
1439 			);
1440 			vm_object_count--;
1441 			crit_exit();
1442 
1443 			zfree(obj_zone, backing_object);
1444 
1445 			object_collapses++;
1446 		} else {
1447 			vm_object_t new_backing_object;
1448 
1449 			/*
1450 			 * If we do not entirely shadow the backing object,
1451 			 * there is nothing we can do so we give up.
1452 			 */
1453 
1454 			if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1455 				break;
1456 			}
1457 
1458 			/*
1459 			 * Make the parent shadow the next object in the
1460 			 * chain.  Deallocating backing_object will not remove
1461 			 * it, since its reference count is at least 2.
1462 			 */
1463 
1464 			LIST_REMOVE(object, shadow_list);
1465 			backing_object->shadow_count--;
1466 			backing_object->generation++;
1467 
1468 			new_backing_object = backing_object->backing_object;
1469 			if ((object->backing_object = new_backing_object) != NULL) {
1470 				vm_object_reference(new_backing_object);
1471 				LIST_INSERT_HEAD(
1472 				    &new_backing_object->shadow_head,
1473 				    object,
1474 				    shadow_list
1475 				);
1476 				new_backing_object->shadow_count++;
1477 				new_backing_object->generation++;
1478 				object->backing_object_offset +=
1479 					backing_object->backing_object_offset;
1480 			}
1481 
1482 			/*
1483 			 * Drop the reference count on backing_object. Since
1484 			 * its ref_count was at least 2, it will not vanish;
1485 			 * so we don't need to call vm_object_deallocate, but
1486 			 * we do anyway.
1487 			 */
1488 			vm_object_deallocate(backing_object);
1489 			object_bypasses++;
1490 		}
1491 
1492 		/*
1493 		 * Try again with this object's new backing object.
1494 		 */
1495 	}
1496 }
1497 
1498 /*
1499  *	vm_object_page_remove: [internal]
1500  *
1501  *	Removes all physical pages in the specified
1502  *	object range from the object's list of pages.
1503  */
1504 static int vm_object_page_remove_callback(vm_page_t p, void *data);
1505 
1506 void
1507 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1508 		      boolean_t clean_only)
1509 {
1510 	struct rb_vm_page_scan_info info;
1511 	int all;
1512 
1513 	/*
1514 	 * Degenerate cases and assertions
1515 	 */
1516 	if (object == NULL || object->resident_page_count == 0)
1517 		return;
1518 	KASSERT(object->type != OBJT_PHYS,
1519 		("attempt to remove pages from a physical object"));
1520 
1521 	/*
1522 	 * Indicate that paging is occuring on the object
1523 	 */
1524 	crit_enter();
1525 	vm_object_pip_add(object, 1);
1526 
1527 	/*
1528 	 * Figure out the actual removal range and whether we are removing
1529 	 * the entire contents of the object or not.  If removing the entire
1530 	 * contents, be sure to get all pages, even those that might be
1531 	 * beyond the end of the object.
1532 	 */
1533 	info.start_pindex = start;
1534 	if (end == 0)
1535 		info.end_pindex = (vm_pindex_t)-1;
1536 	else
1537 		info.end_pindex = end - 1;
1538 	info.limit = clean_only;
1539 	all = (start == 0 && info.end_pindex >= object->size - 1);
1540 
1541 	/*
1542 	 * Loop until we are sure we have gotten them all.
1543 	 */
1544 	do {
1545 		info.error = 0;
1546 		vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1547 					vm_object_page_remove_callback, &info);
1548 	} while (info.error);
1549 
1550 	/*
1551 	 * Cleanup
1552 	 */
1553 	vm_object_pip_wakeup(object);
1554 	crit_exit();
1555 }
1556 
1557 static int
1558 vm_object_page_remove_callback(vm_page_t p, void *data)
1559 {
1560 	struct rb_vm_page_scan_info *info = data;
1561 
1562 	/*
1563 	 * Wired pages cannot be destroyed, but they can be invalidated
1564 	 * and we do so if clean_only (limit) is not set.
1565 	 */
1566 	if (p->wire_count != 0) {
1567 		vm_page_protect(p, VM_PROT_NONE);
1568 		if (info->limit == 0)
1569 			p->valid = 0;
1570 		return(0);
1571 	}
1572 
1573 	/*
1574 	 * The busy flags are only cleared at
1575 	 * interrupt -- minimize the spl transitions
1576 	 */
1577 
1578 	if (vm_page_sleep_busy(p, TRUE, "vmopar")) {
1579 		info->error = 1;
1580 		return(0);
1581 	}
1582 
1583 	/*
1584 	 * limit is our clean_only flag.  If set and the page is dirty, do
1585 	 * not free it.
1586 	 */
1587 	if (info->limit && p->valid) {
1588 		vm_page_test_dirty(p);
1589 		if (p->valid & p->dirty)
1590 			return(0);
1591 	}
1592 
1593 	/*
1594 	 * Destroy the page
1595 	 */
1596 	vm_page_busy(p);
1597 	vm_page_protect(p, VM_PROT_NONE);
1598 	vm_page_free(p);
1599 	return(0);
1600 }
1601 
1602 /*
1603  *	Routine:	vm_object_coalesce
1604  *	Function:	Coalesces two objects backing up adjoining
1605  *			regions of memory into a single object.
1606  *
1607  *	returns TRUE if objects were combined.
1608  *
1609  *	NOTE:	Only works at the moment if the second object is NULL -
1610  *		if it's not, which object do we lock first?
1611  *
1612  *	Parameters:
1613  *		prev_object	First object to coalesce
1614  *		prev_offset	Offset into prev_object
1615  *		next_object	Second object into coalesce
1616  *		next_offset	Offset into next_object
1617  *
1618  *		prev_size	Size of reference to prev_object
1619  *		next_size	Size of reference to next_object
1620  *
1621  *	Conditions:
1622  *	The object must *not* be locked.
1623  */
1624 boolean_t
1625 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
1626     vm_size_t prev_size, vm_size_t next_size)
1627 {
1628 	vm_pindex_t next_pindex;
1629 
1630 	if (prev_object == NULL) {
1631 		return (TRUE);
1632 	}
1633 
1634 	if (prev_object->type != OBJT_DEFAULT &&
1635 	    prev_object->type != OBJT_SWAP) {
1636 		return (FALSE);
1637 	}
1638 
1639 	/*
1640 	 * Try to collapse the object first
1641 	 */
1642 	vm_object_collapse(prev_object);
1643 
1644 	/*
1645 	 * Can't coalesce if: . more than one reference . paged out . shadows
1646 	 * another object . has a copy elsewhere (any of which mean that the
1647 	 * pages not mapped to prev_entry may be in use anyway)
1648 	 */
1649 
1650 	if (prev_object->backing_object != NULL) {
1651 		return (FALSE);
1652 	}
1653 
1654 	prev_size >>= PAGE_SHIFT;
1655 	next_size >>= PAGE_SHIFT;
1656 	next_pindex = prev_pindex + prev_size;
1657 
1658 	if ((prev_object->ref_count > 1) &&
1659 	    (prev_object->size != next_pindex)) {
1660 		return (FALSE);
1661 	}
1662 
1663 	/*
1664 	 * Remove any pages that may still be in the object from a previous
1665 	 * deallocation.
1666 	 */
1667 	if (next_pindex < prev_object->size) {
1668 		vm_object_page_remove(prev_object,
1669 				      next_pindex,
1670 				      next_pindex + next_size, FALSE);
1671 		if (prev_object->type == OBJT_SWAP)
1672 			swap_pager_freespace(prev_object,
1673 					     next_pindex, next_size);
1674 	}
1675 
1676 	/*
1677 	 * Extend the object if necessary.
1678 	 */
1679 	if (next_pindex + next_size > prev_object->size)
1680 		prev_object->size = next_pindex + next_size;
1681 
1682 	return (TRUE);
1683 }
1684 
1685 void
1686 vm_object_set_writeable_dirty(vm_object_t object)
1687 {
1688 	struct vnode *vp;
1689 
1690 	vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1691 	if (object->type == OBJT_VNODE &&
1692 	    (vp = (struct vnode *)object->handle) != NULL) {
1693 		if ((vp->v_flag & VOBJDIRTY) == 0) {
1694 			vsetflags(vp, VOBJDIRTY);
1695 		}
1696 	}
1697 }
1698 
1699 
1700 
1701 #include "opt_ddb.h"
1702 #ifdef DDB
1703 #include <sys/kernel.h>
1704 
1705 #include <sys/cons.h>
1706 
1707 #include <ddb/ddb.h>
1708 
1709 static int	_vm_object_in_map (vm_map_t map, vm_object_t object,
1710 				       vm_map_entry_t entry);
1711 static int	vm_object_in_map (vm_object_t object);
1712 
1713 static int
1714 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
1715 {
1716 	vm_map_t tmpm;
1717 	vm_map_entry_t tmpe;
1718 	vm_object_t obj;
1719 	int entcount;
1720 
1721 	if (map == 0)
1722 		return 0;
1723 	if (entry == 0) {
1724 		tmpe = map->header.next;
1725 		entcount = map->nentries;
1726 		while (entcount-- && (tmpe != &map->header)) {
1727 			if( _vm_object_in_map(map, object, tmpe)) {
1728 				return 1;
1729 			}
1730 			tmpe = tmpe->next;
1731 		}
1732 		return (0);
1733 	}
1734 	switch(entry->maptype) {
1735 	case VM_MAPTYPE_SUBMAP:
1736 		tmpm = entry->object.sub_map;
1737 		tmpe = tmpm->header.next;
1738 		entcount = tmpm->nentries;
1739 		while (entcount-- && tmpe != &tmpm->header) {
1740 			if( _vm_object_in_map(tmpm, object, tmpe)) {
1741 				return 1;
1742 			}
1743 			tmpe = tmpe->next;
1744 		}
1745 		break;
1746 	case VM_MAPTYPE_NORMAL:
1747 	case VM_MAPTYPE_VPAGETABLE:
1748 		obj = entry->object.vm_object;
1749 		while (obj) {
1750 			if (obj == object)
1751 				return 1;
1752 			obj = obj->backing_object;
1753 		}
1754 		break;
1755 	default:
1756 		break;
1757 	}
1758 	return 0;
1759 }
1760 
1761 static int vm_object_in_map_callback(struct proc *p, void *data);
1762 
1763 struct vm_object_in_map_info {
1764 	vm_object_t object;
1765 	int rv;
1766 };
1767 
1768 static int
1769 vm_object_in_map(vm_object_t object)
1770 {
1771 	struct vm_object_in_map_info info;
1772 
1773 	info.rv = 0;
1774 	info.object = object;
1775 
1776 	allproc_scan(vm_object_in_map_callback, &info);
1777 	if (info.rv)
1778 		return 1;
1779 	if( _vm_object_in_map( kernel_map, object, 0))
1780 		return 1;
1781 	if( _vm_object_in_map( pager_map, object, 0))
1782 		return 1;
1783 	if( _vm_object_in_map( buffer_map, object, 0))
1784 		return 1;
1785 	return 0;
1786 }
1787 
1788 static int
1789 vm_object_in_map_callback(struct proc *p, void *data)
1790 {
1791 	struct vm_object_in_map_info *info = data;
1792 
1793 	if (p->p_vmspace) {
1794 		if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) {
1795 			info->rv = 1;
1796 			return -1;
1797 		}
1798 	}
1799 	return (0);
1800 }
1801 
1802 DB_SHOW_COMMAND(vmochk, vm_object_check)
1803 {
1804 	vm_object_t object;
1805 
1806 	/*
1807 	 * make sure that internal objs are in a map somewhere
1808 	 * and none have zero ref counts.
1809 	 */
1810 	for (object = TAILQ_FIRST(&vm_object_list);
1811 			object != NULL;
1812 			object = TAILQ_NEXT(object, object_list)) {
1813 		if (object->handle == NULL &&
1814 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1815 			if (object->ref_count == 0) {
1816 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
1817 					(long)object->size);
1818 			}
1819 			if (!vm_object_in_map(object)) {
1820 				db_printf(
1821 			"vmochk: internal obj is not in a map: "
1822 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1823 				    object->ref_count, (u_long)object->size,
1824 				    (u_long)object->size,
1825 				    (void *)object->backing_object);
1826 			}
1827 		}
1828 	}
1829 }
1830 
1831 /*
1832  *	vm_object_print:	[ debug ]
1833  */
1834 DB_SHOW_COMMAND(object, vm_object_print_static)
1835 {
1836 	/* XXX convert args. */
1837 	vm_object_t object = (vm_object_t)addr;
1838 	boolean_t full = have_addr;
1839 
1840 	vm_page_t p;
1841 
1842 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
1843 #define	count	was_count
1844 
1845 	int count;
1846 
1847 	if (object == NULL)
1848 		return;
1849 
1850 	db_iprintf(
1851 	    "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1852 	    object, (int)object->type, (u_long)object->size,
1853 	    object->resident_page_count, object->ref_count, object->flags);
1854 	/*
1855 	 * XXX no %qd in kernel.  Truncate object->backing_object_offset.
1856 	 */
1857 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1858 	    object->shadow_count,
1859 	    object->backing_object ? object->backing_object->ref_count : 0,
1860 	    object->backing_object, (long)object->backing_object_offset);
1861 
1862 	if (!full)
1863 		return;
1864 
1865 	db_indent += 2;
1866 	count = 0;
1867 	RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) {
1868 		if (count == 0)
1869 			db_iprintf("memory:=");
1870 		else if (count == 6) {
1871 			db_printf("\n");
1872 			db_iprintf(" ...");
1873 			count = 0;
1874 		} else
1875 			db_printf(",");
1876 		count++;
1877 
1878 		db_printf("(off=0x%lx,page=0x%lx)",
1879 		    (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1880 	}
1881 	if (count != 0)
1882 		db_printf("\n");
1883 	db_indent -= 2;
1884 }
1885 
1886 /* XXX. */
1887 #undef count
1888 
1889 /* XXX need this non-static entry for calling from vm_map_print. */
1890 void
1891 vm_object_print(/* db_expr_t */ long addr,
1892 		boolean_t have_addr,
1893 		/* db_expr_t */ long count,
1894 		char *modif)
1895 {
1896 	vm_object_print_static(addr, have_addr, count, modif);
1897 }
1898 
1899 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1900 {
1901 	vm_object_t object;
1902 	int nl = 0;
1903 	int c;
1904 	for (object = TAILQ_FIRST(&vm_object_list);
1905 			object != NULL;
1906 			object = TAILQ_NEXT(object, object_list)) {
1907 		vm_pindex_t idx, fidx;
1908 		vm_pindex_t osize;
1909 		vm_paddr_t pa = -1, padiff;
1910 		int rcount;
1911 		vm_page_t m;
1912 
1913 		db_printf("new object: %p\n", (void *)object);
1914 		if ( nl > 18) {
1915 			c = cngetc();
1916 			if (c != ' ')
1917 				return;
1918 			nl = 0;
1919 		}
1920 		nl++;
1921 		rcount = 0;
1922 		fidx = 0;
1923 		osize = object->size;
1924 		if (osize > 128)
1925 			osize = 128;
1926 		for (idx = 0; idx < osize; idx++) {
1927 			m = vm_page_lookup(object, idx);
1928 			if (m == NULL) {
1929 				if (rcount) {
1930 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1931 						(long)fidx, rcount, (long)pa);
1932 					if ( nl > 18) {
1933 						c = cngetc();
1934 						if (c != ' ')
1935 							return;
1936 						nl = 0;
1937 					}
1938 					nl++;
1939 					rcount = 0;
1940 				}
1941 				continue;
1942 			}
1943 
1944 
1945 			if (rcount &&
1946 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1947 				++rcount;
1948 				continue;
1949 			}
1950 			if (rcount) {
1951 				padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1952 				padiff >>= PAGE_SHIFT;
1953 				padiff &= PQ_L2_MASK;
1954 				if (padiff == 0) {
1955 					pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1956 					++rcount;
1957 					continue;
1958 				}
1959 				db_printf(" index(%ld)run(%d)pa(0x%lx)",
1960 					(long)fidx, rcount, (long)pa);
1961 				db_printf("pd(%ld)\n", (long)padiff);
1962 				if ( nl > 18) {
1963 					c = cngetc();
1964 					if (c != ' ')
1965 						return;
1966 					nl = 0;
1967 				}
1968 				nl++;
1969 			}
1970 			fidx = idx;
1971 			pa = VM_PAGE_TO_PHYS(m);
1972 			rcount = 1;
1973 		}
1974 		if (rcount) {
1975 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1976 				(long)fidx, rcount, (long)pa);
1977 			if ( nl > 18) {
1978 				c = cngetc();
1979 				if (c != ' ')
1980 					return;
1981 				nl = 0;
1982 			}
1983 			nl++;
1984 		}
1985 	}
1986 }
1987 #endif /* DDB */
1988