xref: /freebsd/sys/vm/vm_object.c (revision f2ac424af7b980ba4d858ecfd1644ce197d6869d)
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  * $FreeBSD$
65  */
66 
67 /*
68  *	Virtual memory object module.
69  */
70 
71 #include <sys/param.h>
72 #include <sys/systm.h>
73 #include <sys/lock.h>
74 #include <sys/mman.h>
75 #include <sys/mount.h>
76 #include <sys/kernel.h>
77 #include <sys/sysctl.h>
78 #include <sys/mutex.h>
79 #include <sys/proc.h>		/* for curproc, pageproc */
80 #include <sys/socket.h>
81 #include <sys/vnode.h>
82 #include <sys/vmmeter.h>
83 #include <sys/sx.h>
84 
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_pager.h>
93 #include <vm/vm_zone.h>
94 #include <vm/swap_pager.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_extern.h>
97 
98 #define EASY_SCAN_FACTOR       8
99 
100 #define MSYNC_FLUSH_HARDSEQ	0x01
101 #define MSYNC_FLUSH_SOFTSEQ	0x02
102 
103 /*
104  * msync / VM object flushing optimizations
105  */
106 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ;
107 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags,
108         CTLFLAG_RW, &msync_flush_flags, 0, "");
109 
110 static void	vm_object_qcollapse(vm_object_t object);
111 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
112 
113 /*
114  *	Virtual memory objects maintain the actual data
115  *	associated with allocated virtual memory.  A given
116  *	page of memory exists within exactly one object.
117  *
118  *	An object is only deallocated when all "references"
119  *	are given up.  Only one "reference" to a given
120  *	region of an object should be writeable.
121  *
122  *	Associated with each object is a list of all resident
123  *	memory pages belonging to that object; this list is
124  *	maintained by the "vm_page" module, and locked by the object's
125  *	lock.
126  *
127  *	Each object also records a "pager" routine which is
128  *	used to retrieve (and store) pages to the proper backing
129  *	storage.  In addition, objects may be backed by other
130  *	objects from which they were virtual-copied.
131  *
132  *	The only items within the object structure which are
133  *	modified after time of creation are:
134  *		reference count		locked by object's lock
135  *		pager routine		locked by object's lock
136  *
137  */
138 
139 struct object_q vm_object_list;
140 static struct mtx vm_object_list_mtx;	/* lock for object list and count */
141 static long vm_object_count;		/* count of all objects */
142 vm_object_t kernel_object;
143 vm_object_t kmem_object;
144 static struct vm_object kernel_object_store;
145 static struct vm_object kmem_object_store;
146 extern int vm_pageout_page_count;
147 
148 static long object_collapses;
149 static long object_bypasses;
150 static int next_index;
151 static vm_zone_t obj_zone;
152 static struct vm_zone obj_zone_store;
153 static int object_hash_rand;
154 #define VM_OBJECTS_INIT 256
155 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
156 
157 void
158 _vm_object_allocate(objtype_t type, vm_size_t size, vm_object_t object)
159 {
160 	int incr;
161 
162 	GIANT_REQUIRED;
163 
164 	TAILQ_INIT(&object->memq);
165 	TAILQ_INIT(&object->shadow_head);
166 
167 	object->type = type;
168 	object->size = size;
169 	object->ref_count = 1;
170 	object->flags = 0;
171 	if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
172 		vm_object_set_flag(object, OBJ_ONEMAPPING);
173 	object->paging_in_progress = 0;
174 	object->resident_page_count = 0;
175 	object->shadow_count = 0;
176 	object->pg_color = next_index;
177 	if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
178 		incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
179 	else
180 		incr = size;
181 	next_index = (next_index + incr) & PQ_L2_MASK;
182 	object->handle = NULL;
183 	object->backing_object = NULL;
184 	object->backing_object_offset = (vm_ooffset_t) 0;
185 	/*
186 	 * Try to generate a number that will spread objects out in the
187 	 * hash table.  We 'wipe' new objects across the hash in 128 page
188 	 * increments plus 1 more to offset it a little more by the time
189 	 * it wraps around.
190 	 */
191 	object->hash_rand = object_hash_rand - 129;
192 
193 	object->generation++;
194 
195 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
196 	vm_object_count++;
197 	object_hash_rand = object->hash_rand;
198 }
199 
200 /*
201  *	vm_object_init:
202  *
203  *	Initialize the VM objects module.
204  */
205 void
206 vm_object_init(void)
207 {
208 	GIANT_REQUIRED;
209 
210 	TAILQ_INIT(&vm_object_list);
211 	mtx_init(&vm_object_list_mtx, "vm object_list", MTX_DEF);
212 	vm_object_count = 0;
213 
214 	kernel_object = &kernel_object_store;
215 	_vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
216 	    kernel_object);
217 
218 	kmem_object = &kmem_object_store;
219 	_vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
220 	    kmem_object);
221 
222 	obj_zone = &obj_zone_store;
223 	zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
224 		vm_objects_init, VM_OBJECTS_INIT);
225 }
226 
227 void
228 vm_object_init2(void)
229 {
230 	zinitna(obj_zone, NULL, NULL, 0, 0, 0, 1);
231 }
232 
233 void
234 vm_object_set_flag(vm_object_t object, u_short bits)
235 {
236 	GIANT_REQUIRED;
237 	object->flags |= bits;
238 }
239 
240 void
241 vm_object_clear_flag(vm_object_t object, u_short bits)
242 {
243 	GIANT_REQUIRED;
244 	object->flags &= ~bits;
245 }
246 
247 void
248 vm_object_pip_add(vm_object_t object, short i)
249 {
250 	GIANT_REQUIRED;
251 	object->paging_in_progress += i;
252 }
253 
254 void
255 vm_object_pip_subtract(vm_object_t object, short i)
256 {
257 	GIANT_REQUIRED;
258 	object->paging_in_progress -= i;
259 }
260 
261 void
262 vm_object_pip_wakeup(vm_object_t object)
263 {
264 	GIANT_REQUIRED;
265 	object->paging_in_progress--;
266 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
267 		vm_object_clear_flag(object, OBJ_PIPWNT);
268 		wakeup(object);
269 	}
270 }
271 
272 void
273 vm_object_pip_wakeupn(vm_object_t object, short i)
274 {
275 	GIANT_REQUIRED;
276 	if (i)
277 		object->paging_in_progress -= i;
278 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
279 		vm_object_clear_flag(object, OBJ_PIPWNT);
280 		wakeup(object);
281 	}
282 }
283 
284 void
285 vm_object_pip_sleep(vm_object_t object, char *waitid)
286 {
287 	GIANT_REQUIRED;
288 	if (object->paging_in_progress) {
289 		int s = splvm();
290 		if (object->paging_in_progress) {
291 			vm_object_set_flag(object, OBJ_PIPWNT);
292 			tsleep(object, PVM, waitid, 0);
293 		}
294 		splx(s);
295 	}
296 }
297 
298 void
299 vm_object_pip_wait(vm_object_t object, char *waitid)
300 {
301 	GIANT_REQUIRED;
302 	while (object->paging_in_progress)
303 		vm_object_pip_sleep(object, waitid);
304 }
305 
306 /*
307  *	vm_object_allocate:
308  *
309  *	Returns a new object with the given size.
310  */
311 
312 vm_object_t
313 vm_object_allocate(objtype_t type, vm_size_t size)
314 {
315 	vm_object_t result;
316 
317 	GIANT_REQUIRED;
318 
319 	result = (vm_object_t) zalloc(obj_zone);
320 	_vm_object_allocate(type, size, result);
321 
322 	return (result);
323 }
324 
325 
326 /*
327  *	vm_object_reference:
328  *
329  *	Gets another reference to the given object.
330  */
331 void
332 vm_object_reference(vm_object_t object)
333 {
334 	GIANT_REQUIRED;
335 
336 	if (object == NULL)
337 		return;
338 
339 #if 0
340 	/* object can be re-referenced during final cleaning */
341 	KASSERT(!(object->flags & OBJ_DEAD),
342 	    ("vm_object_reference: attempting to reference dead obj"));
343 #endif
344 
345 	object->ref_count++;
346 	if (object->type == OBJT_VNODE) {
347 		while (vget((struct vnode *) object->handle, LK_RETRY|LK_NOOBJ, curthread)) {
348 			printf("vm_object_reference: delay in getting object\n");
349 		}
350 	}
351 }
352 
353 /*
354  * handle deallocating a object of type OBJT_VNODE
355  */
356 void
357 vm_object_vndeallocate(vm_object_t object)
358 {
359 	struct vnode *vp = (struct vnode *) object->handle;
360 
361 	GIANT_REQUIRED;
362 	KASSERT(object->type == OBJT_VNODE,
363 	    ("vm_object_vndeallocate: not a vnode object"));
364 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
365 #ifdef INVARIANTS
366 	if (object->ref_count == 0) {
367 		vprint("vm_object_vndeallocate", vp);
368 		panic("vm_object_vndeallocate: bad object reference count");
369 	}
370 #endif
371 
372 	object->ref_count--;
373 	if (object->ref_count == 0) {
374 		vp->v_flag &= ~VTEXT;
375 		vm_object_clear_flag(object, OBJ_OPT);
376 	}
377 	/*
378 	 * vrele may need a vop lock
379 	 */
380 	vrele(vp);
381 }
382 
383 /*
384  *	vm_object_deallocate:
385  *
386  *	Release a reference to the specified object,
387  *	gained either through a vm_object_allocate
388  *	or a vm_object_reference call.  When all references
389  *	are gone, storage associated with this object
390  *	may be relinquished.
391  *
392  *	No object may be locked.
393  */
394 void
395 vm_object_deallocate(vm_object_t object)
396 {
397 	vm_object_t temp;
398 
399 	GIANT_REQUIRED;
400 
401 	while (object != NULL) {
402 
403 		if (object->type == OBJT_VNODE) {
404 			vm_object_vndeallocate(object);
405 			return;
406 		}
407 
408 		KASSERT(object->ref_count != 0,
409 			("vm_object_deallocate: object deallocated too many times: %d", object->type));
410 
411 		/*
412 		 * If the reference count goes to 0 we start calling
413 		 * vm_object_terminate() on the object chain.
414 		 * A ref count of 1 may be a special case depending on the
415 		 * shadow count being 0 or 1.
416 		 */
417 		object->ref_count--;
418 		if (object->ref_count > 1) {
419 			return;
420 		} else if (object->ref_count == 1) {
421 			if (object->shadow_count == 0) {
422 				vm_object_set_flag(object, OBJ_ONEMAPPING);
423 			} else if ((object->shadow_count == 1) &&
424 			    (object->handle == NULL) &&
425 			    (object->type == OBJT_DEFAULT ||
426 			     object->type == OBJT_SWAP)) {
427 				vm_object_t robject;
428 
429 				robject = TAILQ_FIRST(&object->shadow_head);
430 				KASSERT(robject != NULL,
431 				    ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
432 					 object->ref_count,
433 					 object->shadow_count));
434 				if ((robject->handle == NULL) &&
435 				    (robject->type == OBJT_DEFAULT ||
436 				     robject->type == OBJT_SWAP)) {
437 
438 					robject->ref_count++;
439 
440 					while (
441 						robject->paging_in_progress ||
442 						object->paging_in_progress
443 					) {
444 						vm_object_pip_sleep(robject, "objde1");
445 						vm_object_pip_sleep(object, "objde2");
446 					}
447 
448 					if (robject->ref_count == 1) {
449 						robject->ref_count--;
450 						object = robject;
451 						goto doterm;
452 					}
453 
454 					object = robject;
455 					vm_object_collapse(object);
456 					continue;
457 				}
458 			}
459 
460 			return;
461 
462 		}
463 
464 doterm:
465 
466 		temp = object->backing_object;
467 		if (temp) {
468 			TAILQ_REMOVE(&temp->shadow_head, object, shadow_list);
469 			temp->shadow_count--;
470 			if (temp->ref_count == 0)
471 				vm_object_clear_flag(temp, OBJ_OPT);
472 			temp->generation++;
473 			object->backing_object = NULL;
474 		}
475 		/*
476 		 * Don't double-terminate, we could be in a termination
477 		 * recursion due to the terminate having to sync data
478 		 * to disk.
479 		 */
480 		if ((object->flags & OBJ_DEAD) == 0)
481 			vm_object_terminate(object);
482 		object = temp;
483 	}
484 }
485 
486 /*
487  *	vm_object_terminate actually destroys the specified object, freeing
488  *	up all previously used resources.
489  *
490  *	The object must be locked.
491  *	This routine may block.
492  */
493 void
494 vm_object_terminate(vm_object_t object)
495 {
496 	vm_page_t p;
497 	int s;
498 
499 	GIANT_REQUIRED;
500 
501 	/*
502 	 * Make sure no one uses us.
503 	 */
504 	vm_object_set_flag(object, OBJ_DEAD);
505 
506 	/*
507 	 * wait for the pageout daemon to be done with the object
508 	 */
509 	vm_object_pip_wait(object, "objtrm");
510 
511 	KASSERT(!object->paging_in_progress,
512 		("vm_object_terminate: pageout in progress"));
513 
514 	/*
515 	 * Clean and free the pages, as appropriate. All references to the
516 	 * object are gone, so we don't need to lock it.
517 	 */
518 	if (object->type == OBJT_VNODE) {
519 		struct vnode *vp;
520 
521 		/*
522 		 * Freeze optimized copies.
523 		 */
524 		vm_freeze_copyopts(object, 0, object->size);
525 
526 		/*
527 		 * Clean pages and flush buffers.
528 		 */
529 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
530 
531 		vp = (struct vnode *) object->handle;
532 		vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0);
533 	}
534 
535 	KASSERT(object->ref_count == 0,
536 		("vm_object_terminate: object with references, ref_count=%d",
537 		object->ref_count));
538 
539 	/*
540 	 * Now free any remaining pages. For internal objects, this also
541 	 * removes them from paging queues. Don't free wired pages, just
542 	 * remove them from the object.
543 	 */
544 	s = splvm();
545 	while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
546 		KASSERT(!p->busy && (p->flags & PG_BUSY) == 0,
547 			("vm_object_terminate: freeing busy page %p "
548 			"p->busy = %d, p->flags %x\n", p, p->busy, p->flags));
549 		if (p->wire_count == 0) {
550 			vm_page_busy(p);
551 			vm_page_free(p);
552 			cnt.v_pfree++;
553 		} else {
554 			vm_page_busy(p);
555 			vm_page_remove(p);
556 		}
557 	}
558 	splx(s);
559 
560 	/*
561 	 * Let the pager know object is dead.
562 	 */
563 	vm_pager_deallocate(object);
564 
565 	/*
566 	 * Remove the object from the global object list.
567 	 */
568 	mtx_lock(&vm_object_list_mtx);
569 	TAILQ_REMOVE(&vm_object_list, object, object_list);
570 	mtx_unlock(&vm_object_list_mtx);
571 
572 	wakeup(object);
573 
574 	/*
575 	 * Free the space for the object.
576 	 */
577 	zfree(obj_zone, object);
578 }
579 
580 /*
581  *	vm_object_page_clean
582  *
583  *	Clean all dirty pages in the specified range of object.  Leaves page
584  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
585  *	write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
586  *	leaving the object dirty.
587  *
588  *	Odd semantics: if start == end, we clean everything.
589  *
590  *	The object must be locked.
591  */
592 
593 void
594 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags)
595 {
596 	vm_page_t p, np;
597 	vm_offset_t tstart, tend;
598 	vm_pindex_t pi;
599 	struct vnode *vp;
600 	int clearobjflags;
601 	int pagerflags;
602 	int curgeneration;
603 
604 	GIANT_REQUIRED;
605 
606 	if (object->type != OBJT_VNODE ||
607 		(object->flags & OBJ_MIGHTBEDIRTY) == 0)
608 		return;
609 
610 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : 0;
611 	pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
612 
613 	vp = object->handle;
614 
615 	vm_object_set_flag(object, OBJ_CLEANING);
616 
617 	tstart = start;
618 	if (end == 0) {
619 		tend = object->size;
620 	} else {
621 		tend = end;
622 	}
623 
624 	/*
625 	 * If the caller is smart and only msync()s a range he knows is
626 	 * dirty, we may be able to avoid an object scan.  This results in
627 	 * a phenominal improvement in performance.  We cannot do this
628 	 * as a matter of course because the object may be huge - e.g.
629 	 * the size might be in the gigabytes or terrabytes.
630 	 */
631 	if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
632 		vm_offset_t tscan;
633 		int scanlimit;
634 		int scanreset;
635 
636 		scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
637 		if (scanreset < 16)
638 			scanreset = 16;
639 
640 		scanlimit = scanreset;
641 		tscan = tstart;
642 		while (tscan < tend) {
643 			curgeneration = object->generation;
644 			p = vm_page_lookup(object, tscan);
645 			if (p == NULL || p->valid == 0 ||
646 			    (p->queue - p->pc) == PQ_CACHE) {
647 				if (--scanlimit == 0)
648 					break;
649 				++tscan;
650 				continue;
651 			}
652 			vm_page_test_dirty(p);
653 			if ((p->dirty & p->valid) == 0) {
654 				if (--scanlimit == 0)
655 					break;
656 				++tscan;
657 				continue;
658 			}
659 			/*
660 			 * If we have been asked to skip nosync pages and
661 			 * this is a nosync page, we can't continue.
662 			 */
663 			if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
664 				if (--scanlimit == 0)
665 					break;
666 				++tscan;
667 				continue;
668 			}
669 			scanlimit = scanreset;
670 
671 			/*
672 			 * This returns 0 if it was unable to busy the first
673 			 * page (i.e. had to sleep).
674 			 */
675 			tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
676 		}
677 
678 		/*
679 		 * If everything was dirty and we flushed it successfully,
680 		 * and the requested range is not the entire object, we
681 		 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
682 		 * return immediately.
683 		 */
684 		if (tscan >= tend && (tstart || tend < object->size)) {
685 			vm_object_clear_flag(object, OBJ_CLEANING);
686 			return;
687 		}
688 	}
689 
690 	/*
691 	 * Generally set CLEANCHK interlock and make the page read-only so
692 	 * we can then clear the object flags.
693 	 *
694 	 * However, if this is a nosync mmap then the object is likely to
695 	 * stay dirty so do not mess with the page and do not clear the
696 	 * object flags.
697 	 */
698 
699 	clearobjflags = 1;
700 
701 	TAILQ_FOREACH(p, &object->memq, listq) {
702 		vm_page_flag_set(p, PG_CLEANCHK);
703 		if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
704 			clearobjflags = 0;
705 		else
706 			vm_page_protect(p, VM_PROT_READ);
707 	}
708 
709 	if (clearobjflags && (tstart == 0) && (tend == object->size)) {
710 		struct vnode *vp;
711 
712 		vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
713 		if (object->type == OBJT_VNODE &&
714 		    (vp = (struct vnode *)object->handle) != NULL) {
715 			if (vp->v_flag & VOBJDIRTY) {
716 				mtx_lock(&vp->v_interlock);
717 				vp->v_flag &= ~VOBJDIRTY;
718 				mtx_unlock(&vp->v_interlock);
719 			}
720 		}
721 	}
722 
723 rescan:
724 	curgeneration = object->generation;
725 
726 	for (p = TAILQ_FIRST(&object->memq); p; p = np) {
727 		int n;
728 
729 		np = TAILQ_NEXT(p, listq);
730 
731 again:
732 		pi = p->pindex;
733 		if (((p->flags & PG_CLEANCHK) == 0) ||
734 			(pi < tstart) || (pi >= tend) ||
735 			(p->valid == 0) ||
736 			((p->queue - p->pc) == PQ_CACHE)) {
737 			vm_page_flag_clear(p, PG_CLEANCHK);
738 			continue;
739 		}
740 
741 		vm_page_test_dirty(p);
742 		if ((p->dirty & p->valid) == 0) {
743 			vm_page_flag_clear(p, PG_CLEANCHK);
744 			continue;
745 		}
746 
747 		/*
748 		 * If we have been asked to skip nosync pages and this is a
749 		 * nosync page, skip it.  Note that the object flags were
750 		 * not cleared in this case so we do not have to set them.
751 		 */
752 		if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
753 			vm_page_flag_clear(p, PG_CLEANCHK);
754 			continue;
755 		}
756 
757 		n = vm_object_page_collect_flush(object, p,
758 			curgeneration, pagerflags);
759 		if (n == 0)
760 			goto rescan;
761 
762 		if (object->generation != curgeneration)
763 			goto rescan;
764 
765 		/*
766 		 * Try to optimize the next page.  If we can't we pick up
767 		 * our (random) scan where we left off.
768 		 */
769 		if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
770 			if ((p = vm_page_lookup(object, pi + n)) != NULL)
771 				goto again;
772 		}
773 	}
774 
775 #if 0
776 	VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
777 #endif
778 
779 	vm_object_clear_flag(object, OBJ_CLEANING);
780 	return;
781 }
782 
783 static int
784 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
785 {
786 	int runlen;
787 	int s;
788 	int maxf;
789 	int chkb;
790 	int maxb;
791 	int i;
792 	vm_pindex_t pi;
793 	vm_page_t maf[vm_pageout_page_count];
794 	vm_page_t mab[vm_pageout_page_count];
795 	vm_page_t ma[vm_pageout_page_count];
796 
797 	s = splvm();
798 	pi = p->pindex;
799 	while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
800 		if (object->generation != curgeneration) {
801 			splx(s);
802 			return(0);
803 		}
804 	}
805 
806 	maxf = 0;
807 	for(i = 1; i < vm_pageout_page_count; i++) {
808 		vm_page_t tp;
809 
810 		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
811 			if ((tp->flags & PG_BUSY) ||
812 				(tp->flags & PG_CLEANCHK) == 0 ||
813 				(tp->busy != 0))
814 				break;
815 			if((tp->queue - tp->pc) == PQ_CACHE) {
816 				vm_page_flag_clear(tp, PG_CLEANCHK);
817 				break;
818 			}
819 			vm_page_test_dirty(tp);
820 			if ((tp->dirty & tp->valid) == 0) {
821 				vm_page_flag_clear(tp, PG_CLEANCHK);
822 				break;
823 			}
824 			maf[ i - 1 ] = tp;
825 			maxf++;
826 			continue;
827 		}
828 		break;
829 	}
830 
831 	maxb = 0;
832 	chkb = vm_pageout_page_count -  maxf;
833 	if (chkb) {
834 		for(i = 1; i < chkb;i++) {
835 			vm_page_t tp;
836 
837 			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
838 				if ((tp->flags & PG_BUSY) ||
839 					(tp->flags & PG_CLEANCHK) == 0 ||
840 					(tp->busy != 0))
841 					break;
842 				if((tp->queue - tp->pc) == PQ_CACHE) {
843 					vm_page_flag_clear(tp, PG_CLEANCHK);
844 					break;
845 				}
846 				vm_page_test_dirty(tp);
847 				if ((tp->dirty & tp->valid) == 0) {
848 					vm_page_flag_clear(tp, PG_CLEANCHK);
849 					break;
850 				}
851 				mab[ i - 1 ] = tp;
852 				maxb++;
853 				continue;
854 			}
855 			break;
856 		}
857 	}
858 
859 	for(i = 0; i < maxb; i++) {
860 		int index = (maxb - i) - 1;
861 		ma[index] = mab[i];
862 		vm_page_flag_clear(ma[index], PG_CLEANCHK);
863 	}
864 	vm_page_flag_clear(p, PG_CLEANCHK);
865 	ma[maxb] = p;
866 	for(i = 0; i < maxf; i++) {
867 		int index = (maxb + i) + 1;
868 		ma[index] = maf[i];
869 		vm_page_flag_clear(ma[index], PG_CLEANCHK);
870 	}
871 	runlen = maxb + maxf + 1;
872 
873 	splx(s);
874 	vm_pageout_flush(ma, runlen, pagerflags);
875 	for (i = 0; i < runlen; i++) {
876 		if (ma[i]->valid & ma[i]->dirty) {
877 			vm_page_protect(ma[i], VM_PROT_READ);
878 			vm_page_flag_set(ma[i], PG_CLEANCHK);
879 
880 			/*
881 			 * maxf will end up being the actual number of pages
882 			 * we wrote out contiguously, non-inclusive of the
883 			 * first page.  We do not count look-behind pages.
884 			 */
885 			if (i >= maxb + 1 && (maxf > i - maxb - 1))
886 				maxf = i - maxb - 1;
887 		}
888 	}
889 	return(maxf + 1);
890 }
891 
892 /*
893  * Same as vm_object_pmap_copy, except range checking really
894  * works, and is meant for small sections of an object.
895  *
896  * This code protects resident pages by making them read-only
897  * and is typically called on a fork or split when a page
898  * is converted to copy-on-write.
899  *
900  * NOTE: If the page is already at VM_PROT_NONE, calling
901  * vm_page_protect will have no effect.
902  */
903 
904 void
905 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
906 {
907 	vm_pindex_t idx;
908 	vm_page_t p;
909 
910 	GIANT_REQUIRED;
911 
912 	if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
913 		return;
914 
915 	for (idx = start; idx < end; idx++) {
916 		p = vm_page_lookup(object, idx);
917 		if (p == NULL)
918 			continue;
919 		vm_page_protect(p, VM_PROT_READ);
920 	}
921 }
922 
923 /*
924  *	vm_object_pmap_remove:
925  *
926  *	Removes all physical pages in the specified
927  *	object range from all physical maps.
928  *
929  *	The object must *not* be locked.
930  */
931 void
932 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
933 {
934 	vm_page_t p;
935 
936 	GIANT_REQUIRED;
937 	if (object == NULL)
938 		return;
939 	TAILQ_FOREACH(p, &object->memq, listq) {
940 		if (p->pindex >= start && p->pindex < end)
941 			vm_page_protect(p, VM_PROT_NONE);
942 	}
943 	if ((start == 0) && (object->size == end))
944 		vm_object_clear_flag(object, OBJ_WRITEABLE);
945 }
946 
947 /*
948  *	vm_object_madvise:
949  *
950  *	Implements the madvise function at the object/page level.
951  *
952  *	MADV_WILLNEED	(any object)
953  *
954  *	    Activate the specified pages if they are resident.
955  *
956  *	MADV_DONTNEED	(any object)
957  *
958  *	    Deactivate the specified pages if they are resident.
959  *
960  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
961  *			 OBJ_ONEMAPPING only)
962  *
963  *	    Deactivate and clean the specified pages if they are
964  *	    resident.  This permits the process to reuse the pages
965  *	    without faulting or the kernel to reclaim the pages
966  *	    without I/O.
967  */
968 void
969 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
970 {
971 	vm_pindex_t end, tpindex;
972 	vm_object_t tobject;
973 	vm_page_t m;
974 
975 	GIANT_REQUIRED;
976 	if (object == NULL)
977 		return;
978 
979 	end = pindex + count;
980 
981 	/*
982 	 * Locate and adjust resident pages
983 	 */
984 
985 	for (; pindex < end; pindex += 1) {
986 relookup:
987 		tobject = object;
988 		tpindex = pindex;
989 shadowlookup:
990 		/*
991 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
992 		 * and those pages must be OBJ_ONEMAPPING.
993 		 */
994 		if (advise == MADV_FREE) {
995 			if ((tobject->type != OBJT_DEFAULT &&
996 			     tobject->type != OBJT_SWAP) ||
997 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
998 				continue;
999 			}
1000 		}
1001 
1002 		m = vm_page_lookup(tobject, tpindex);
1003 
1004 		if (m == NULL) {
1005 			/*
1006 			 * There may be swap even if there is no backing page
1007 			 */
1008 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1009 				swap_pager_freespace(tobject, tpindex, 1);
1010 
1011 			/*
1012 			 * next object
1013 			 */
1014 			tobject = tobject->backing_object;
1015 			if (tobject == NULL)
1016 				continue;
1017 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1018 			goto shadowlookup;
1019 		}
1020 
1021 		/*
1022 		 * If the page is busy or not in a normal active state,
1023 		 * we skip it.  If the page is not managed there are no
1024 		 * page queues to mess with.  Things can break if we mess
1025 		 * with pages in any of the below states.
1026 		 */
1027 		if (
1028 		    m->hold_count ||
1029 		    m->wire_count ||
1030 		    (m->flags & PG_UNMANAGED) ||
1031 		    m->valid != VM_PAGE_BITS_ALL
1032 		) {
1033 			continue;
1034 		}
1035 
1036  		if (vm_page_sleep_busy(m, TRUE, "madvpo"))
1037   			goto relookup;
1038 
1039 		if (advise == MADV_WILLNEED) {
1040 			vm_page_activate(m);
1041 		} else if (advise == MADV_DONTNEED) {
1042 			vm_page_dontneed(m);
1043 		} else if (advise == MADV_FREE) {
1044 			/*
1045 			 * Mark the page clean.  This will allow the page
1046 			 * to be freed up by the system.  However, such pages
1047 			 * are often reused quickly by malloc()/free()
1048 			 * so we do not do anything that would cause
1049 			 * a page fault if we can help it.
1050 			 *
1051 			 * Specifically, we do not try to actually free
1052 			 * the page now nor do we try to put it in the
1053 			 * cache (which would cause a page fault on reuse).
1054 			 *
1055 			 * But we do make the page is freeable as we
1056 			 * can without actually taking the step of unmapping
1057 			 * it.
1058 			 */
1059 			pmap_clear_modify(m);
1060 			m->dirty = 0;
1061 			m->act_count = 0;
1062 			vm_page_dontneed(m);
1063 			if (tobject->type == OBJT_SWAP)
1064 				swap_pager_freespace(tobject, tpindex, 1);
1065 		}
1066 	}
1067 }
1068 
1069 /*
1070  *	vm_object_shadow:
1071  *
1072  *	Create a new object which is backed by the
1073  *	specified existing object range.  The source
1074  *	object reference is deallocated.
1075  *
1076  *	The new object and offset into that object
1077  *	are returned in the source parameters.
1078  */
1079 
1080 void
1081 vm_object_shadow(
1082 	vm_object_t *object,	/* IN/OUT */
1083 	vm_ooffset_t *offset,	/* IN/OUT */
1084 	vm_size_t length)
1085 {
1086 	vm_object_t source;
1087 	vm_object_t result;
1088 
1089 	GIANT_REQUIRED;
1090 	source = *object;
1091 
1092 	/*
1093 	 * Don't create the new object if the old object isn't shared.
1094 	 */
1095 
1096 	if (source != NULL &&
1097 	    source->ref_count == 1 &&
1098 	    source->handle == NULL &&
1099 	    (source->type == OBJT_DEFAULT ||
1100 	     source->type == OBJT_SWAP))
1101 		return;
1102 
1103 	/*
1104 	 * Allocate a new object with the given length
1105 	 */
1106 	result = vm_object_allocate(OBJT_DEFAULT, length);
1107 	KASSERT(result != NULL, ("vm_object_shadow: no object for shadowing"));
1108 
1109 	/*
1110 	 * The new object shadows the source object, adding a reference to it.
1111 	 * Our caller changes his reference to point to the new object,
1112 	 * removing a reference to the source object.  Net result: no change
1113 	 * of reference count.
1114 	 *
1115 	 * Try to optimize the result object's page color when shadowing
1116 	 * in order to maintain page coloring consistency in the combined
1117 	 * shadowed object.
1118 	 */
1119 	result->backing_object = source;
1120 	if (source) {
1121 		TAILQ_INSERT_TAIL(&source->shadow_head, result, shadow_list);
1122 		source->shadow_count++;
1123 		source->generation++;
1124 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1125 	}
1126 
1127 	/*
1128 	 * Store the offset into the source object, and fix up the offset into
1129 	 * the new object.
1130 	 */
1131 
1132 	result->backing_object_offset = *offset;
1133 
1134 	/*
1135 	 * Return the new things
1136 	 */
1137 
1138 	*offset = 0;
1139 	*object = result;
1140 }
1141 
1142 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1143 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1144 #define	OBSC_COLLAPSE_WAIT	0x0004
1145 
1146 static __inline int
1147 vm_object_backing_scan(vm_object_t object, int op)
1148 {
1149 	int s;
1150 	int r = 1;
1151 	vm_page_t p;
1152 	vm_object_t backing_object;
1153 	vm_pindex_t backing_offset_index;
1154 
1155 	s = splvm();
1156 	GIANT_REQUIRED;
1157 
1158 	backing_object = object->backing_object;
1159 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1160 
1161 	/*
1162 	 * Initial conditions
1163 	 */
1164 
1165 	if (op & OBSC_TEST_ALL_SHADOWED) {
1166 		/*
1167 		 * We do not want to have to test for the existence of
1168 		 * swap pages in the backing object.  XXX but with the
1169 		 * new swapper this would be pretty easy to do.
1170 		 *
1171 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1172 		 * been ZFOD faulted yet?  If we do not test for this, the
1173 		 * shadow test may succeed! XXX
1174 		 */
1175 		if (backing_object->type != OBJT_DEFAULT) {
1176 			splx(s);
1177 			return(0);
1178 		}
1179 	}
1180 	if (op & OBSC_COLLAPSE_WAIT) {
1181 		vm_object_set_flag(backing_object, OBJ_DEAD);
1182 	}
1183 
1184 	/*
1185 	 * Our scan
1186 	 */
1187 
1188 	p = TAILQ_FIRST(&backing_object->memq);
1189 	while (p) {
1190 		vm_page_t next = TAILQ_NEXT(p, listq);
1191 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1192 
1193 		if (op & OBSC_TEST_ALL_SHADOWED) {
1194 			vm_page_t pp;
1195 
1196 			/*
1197 			 * Ignore pages outside the parent object's range
1198 			 * and outside the parent object's mapping of the
1199 			 * backing object.
1200 			 *
1201 			 * note that we do not busy the backing object's
1202 			 * page.
1203 			 */
1204 
1205 			if (
1206 			    p->pindex < backing_offset_index ||
1207 			    new_pindex >= object->size
1208 			) {
1209 				p = next;
1210 				continue;
1211 			}
1212 
1213 			/*
1214 			 * See if the parent has the page or if the parent's
1215 			 * object pager has the page.  If the parent has the
1216 			 * page but the page is not valid, the parent's
1217 			 * object pager must have the page.
1218 			 *
1219 			 * If this fails, the parent does not completely shadow
1220 			 * the object and we might as well give up now.
1221 			 */
1222 
1223 			pp = vm_page_lookup(object, new_pindex);
1224 			if (
1225 			    (pp == NULL || pp->valid == 0) &&
1226 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1227 			) {
1228 				r = 0;
1229 				break;
1230 			}
1231 		}
1232 
1233 		/*
1234 		 * Check for busy page
1235 		 */
1236 
1237 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1238 			vm_page_t pp;
1239 
1240 			if (op & OBSC_COLLAPSE_NOWAIT) {
1241 				if (
1242 				    (p->flags & PG_BUSY) ||
1243 				    !p->valid ||
1244 				    p->hold_count ||
1245 				    p->wire_count ||
1246 				    p->busy
1247 				) {
1248 					p = next;
1249 					continue;
1250 				}
1251 			} else if (op & OBSC_COLLAPSE_WAIT) {
1252 				if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1253 					/*
1254 					 * If we slept, anything could have
1255 					 * happened.  Since the object is
1256 					 * marked dead, the backing offset
1257 					 * should not have changed so we
1258 					 * just restart our scan.
1259 					 */
1260 					p = TAILQ_FIRST(&backing_object->memq);
1261 					continue;
1262 				}
1263 			}
1264 
1265 			/*
1266 			 * Busy the page
1267 			 */
1268 			vm_page_busy(p);
1269 
1270 			KASSERT(
1271 			    p->object == backing_object,
1272 			    ("vm_object_qcollapse(): object mismatch")
1273 			);
1274 
1275 			/*
1276 			 * Destroy any associated swap
1277 			 */
1278 			if (backing_object->type == OBJT_SWAP) {
1279 				swap_pager_freespace(
1280 				    backing_object,
1281 				    p->pindex,
1282 				    1
1283 				);
1284 			}
1285 
1286 			if (
1287 			    p->pindex < backing_offset_index ||
1288 			    new_pindex >= object->size
1289 			) {
1290 				/*
1291 				 * Page is out of the parent object's range, we
1292 				 * can simply destroy it.
1293 				 */
1294 				vm_page_protect(p, VM_PROT_NONE);
1295 				vm_page_free(p);
1296 				p = next;
1297 				continue;
1298 			}
1299 
1300 			pp = vm_page_lookup(object, new_pindex);
1301 			if (
1302 			    pp != NULL ||
1303 			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1304 			) {
1305 				/*
1306 				 * page already exists in parent OR swap exists
1307 				 * for this location in the parent.  Destroy
1308 				 * the original page from the backing object.
1309 				 *
1310 				 * Leave the parent's page alone
1311 				 */
1312 				vm_page_protect(p, VM_PROT_NONE);
1313 				vm_page_free(p);
1314 				p = next;
1315 				continue;
1316 			}
1317 
1318 			/*
1319 			 * Page does not exist in parent, rename the
1320 			 * page from the backing object to the main object.
1321 			 *
1322 			 * If the page was mapped to a process, it can remain
1323 			 * mapped through the rename.
1324 			 */
1325 			if ((p->queue - p->pc) == PQ_CACHE)
1326 				vm_page_deactivate(p);
1327 
1328 			vm_page_rename(p, object, new_pindex);
1329 			/* page automatically made dirty by rename */
1330 		}
1331 		p = next;
1332 	}
1333 	splx(s);
1334 	return(r);
1335 }
1336 
1337 
1338 /*
1339  * this version of collapse allows the operation to occur earlier and
1340  * when paging_in_progress is true for an object...  This is not a complete
1341  * operation, but should plug 99.9% of the rest of the leaks.
1342  */
1343 static void
1344 vm_object_qcollapse(vm_object_t object)
1345 {
1346 	vm_object_t backing_object = object->backing_object;
1347 
1348 	GIANT_REQUIRED;
1349 
1350 	if (backing_object->ref_count != 1)
1351 		return;
1352 
1353 	backing_object->ref_count += 2;
1354 
1355 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1356 
1357 	backing_object->ref_count -= 2;
1358 }
1359 
1360 /*
1361  *	vm_object_collapse:
1362  *
1363  *	Collapse an object with the object backing it.
1364  *	Pages in the backing object are moved into the
1365  *	parent, and the backing object is deallocated.
1366  */
1367 void
1368 vm_object_collapse(vm_object_t object)
1369 {
1370 	GIANT_REQUIRED;
1371 
1372 	while (TRUE) {
1373 		vm_object_t backing_object;
1374 
1375 		/*
1376 		 * Verify that the conditions are right for collapse:
1377 		 *
1378 		 * The object exists and the backing object exists.
1379 		 */
1380 		if (object == NULL)
1381 			break;
1382 
1383 		if ((backing_object = object->backing_object) == NULL)
1384 			break;
1385 
1386 		/*
1387 		 * we check the backing object first, because it is most likely
1388 		 * not collapsable.
1389 		 */
1390 		if (backing_object->handle != NULL ||
1391 		    (backing_object->type != OBJT_DEFAULT &&
1392 		     backing_object->type != OBJT_SWAP) ||
1393 		    (backing_object->flags & OBJ_DEAD) ||
1394 		    object->handle != NULL ||
1395 		    (object->type != OBJT_DEFAULT &&
1396 		     object->type != OBJT_SWAP) ||
1397 		    (object->flags & OBJ_DEAD)) {
1398 			break;
1399 		}
1400 
1401 		if (
1402 		    object->paging_in_progress != 0 ||
1403 		    backing_object->paging_in_progress != 0
1404 		) {
1405 			vm_object_qcollapse(object);
1406 			break;
1407 		}
1408 
1409 		/*
1410 		 * We know that we can either collapse the backing object (if
1411 		 * the parent is the only reference to it) or (perhaps) have
1412 		 * the parent bypass the object if the parent happens to shadow
1413 		 * all the resident pages in the entire backing object.
1414 		 *
1415 		 * This is ignoring pager-backed pages such as swap pages.
1416 		 * vm_object_backing_scan fails the shadowing test in this
1417 		 * case.
1418 		 */
1419 
1420 		if (backing_object->ref_count == 1) {
1421 			/*
1422 			 * If there is exactly one reference to the backing
1423 			 * object, we can collapse it into the parent.
1424 			 */
1425 
1426 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1427 
1428 			/*
1429 			 * Move the pager from backing_object to object.
1430 			 */
1431 
1432 			if (backing_object->type == OBJT_SWAP) {
1433 				vm_object_pip_add(backing_object, 1);
1434 
1435 				/*
1436 				 * scrap the paging_offset junk and do a
1437 				 * discrete copy.  This also removes major
1438 				 * assumptions about how the swap-pager
1439 				 * works from where it doesn't belong.  The
1440 				 * new swapper is able to optimize the
1441 				 * destroy-source case.
1442 				 */
1443 
1444 				vm_object_pip_add(object, 1);
1445 				swap_pager_copy(
1446 				    backing_object,
1447 				    object,
1448 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1449 				vm_object_pip_wakeup(object);
1450 
1451 				vm_object_pip_wakeup(backing_object);
1452 			}
1453 			/*
1454 			 * Object now shadows whatever backing_object did.
1455 			 * Note that the reference to
1456 			 * backing_object->backing_object moves from within
1457 			 * backing_object to within object.
1458 			 */
1459 
1460 			TAILQ_REMOVE(
1461 			    &object->backing_object->shadow_head,
1462 			    object,
1463 			    shadow_list
1464 			);
1465 			object->backing_object->shadow_count--;
1466 			object->backing_object->generation++;
1467 			if (backing_object->backing_object) {
1468 				TAILQ_REMOVE(
1469 				    &backing_object->backing_object->shadow_head,
1470 				    backing_object,
1471 				    shadow_list
1472 				);
1473 				backing_object->backing_object->shadow_count--;
1474 				backing_object->backing_object->generation++;
1475 			}
1476 			object->backing_object = backing_object->backing_object;
1477 			if (object->backing_object) {
1478 				TAILQ_INSERT_TAIL(
1479 				    &object->backing_object->shadow_head,
1480 				    object,
1481 				    shadow_list
1482 				);
1483 				object->backing_object->shadow_count++;
1484 				object->backing_object->generation++;
1485 			}
1486 
1487 			object->backing_object_offset +=
1488 			    backing_object->backing_object_offset;
1489 
1490 			/*
1491 			 * Discard backing_object.
1492 			 *
1493 			 * Since the backing object has no pages, no pager left,
1494 			 * and no object references within it, all that is
1495 			 * necessary is to dispose of it.
1496 			 */
1497 			KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1498 			KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object));
1499 
1500 			TAILQ_REMOVE(
1501 			    &vm_object_list,
1502 			    backing_object,
1503 			    object_list
1504 			);
1505 			vm_object_count--;
1506 
1507 			zfree(obj_zone, backing_object);
1508 
1509 			object_collapses++;
1510 		} else {
1511 			vm_object_t new_backing_object;
1512 
1513 			/*
1514 			 * If we do not entirely shadow the backing object,
1515 			 * there is nothing we can do so we give up.
1516 			 */
1517 
1518 			if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1519 				break;
1520 			}
1521 
1522 			/*
1523 			 * Make the parent shadow the next object in the
1524 			 * chain.  Deallocating backing_object will not remove
1525 			 * it, since its reference count is at least 2.
1526 			 */
1527 
1528 			TAILQ_REMOVE(
1529 			    &backing_object->shadow_head,
1530 			    object,
1531 			    shadow_list
1532 			);
1533 			backing_object->shadow_count--;
1534 			backing_object->generation++;
1535 
1536 			new_backing_object = backing_object->backing_object;
1537 			if ((object->backing_object = new_backing_object) != NULL) {
1538 				vm_object_reference(new_backing_object);
1539 				TAILQ_INSERT_TAIL(
1540 				    &new_backing_object->shadow_head,
1541 				    object,
1542 				    shadow_list
1543 				);
1544 				new_backing_object->shadow_count++;
1545 				new_backing_object->generation++;
1546 				object->backing_object_offset +=
1547 					backing_object->backing_object_offset;
1548 			}
1549 
1550 			/*
1551 			 * Drop the reference count on backing_object. Since
1552 			 * its ref_count was at least 2, it will not vanish;
1553 			 * so we don't need to call vm_object_deallocate, but
1554 			 * we do anyway.
1555 			 */
1556 			vm_object_deallocate(backing_object);
1557 			object_bypasses++;
1558 		}
1559 
1560 		/*
1561 		 * Try again with this object's new backing object.
1562 		 */
1563 	}
1564 }
1565 
1566 /*
1567  *	vm_object_page_remove: [internal]
1568  *
1569  *	Removes all physical pages in the specified
1570  *	object range from the object's list of pages.
1571  *
1572  *	The object must be locked.
1573  */
1574 void
1575 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, boolean_t clean_only)
1576 {
1577 	vm_page_t p, next;
1578 	unsigned int size;
1579 	int all;
1580 
1581 	GIANT_REQUIRED;
1582 
1583 	if (object == NULL ||
1584 	    object->resident_page_count == 0)
1585 		return;
1586 
1587 	all = ((end == 0) && (start == 0));
1588 
1589 	/*
1590 	 * Since physically-backed objects do not use managed pages, we can't
1591 	 * remove pages from the object (we must instead remove the page
1592 	 * references, and then destroy the object).
1593 	 */
1594 	KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object"));
1595 
1596 	vm_object_pip_add(object, 1);
1597 again:
1598 	size = end - start;
1599 	if (all || size > object->resident_page_count / 4) {
1600 		for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
1601 			next = TAILQ_NEXT(p, listq);
1602 			if (all || ((start <= p->pindex) && (p->pindex < end))) {
1603 				if (p->wire_count != 0) {
1604 					vm_page_protect(p, VM_PROT_NONE);
1605 					if (!clean_only)
1606 						p->valid = 0;
1607 					continue;
1608 				}
1609 
1610 				/*
1611 				 * The busy flags are only cleared at
1612 				 * interrupt -- minimize the spl transitions
1613 				 */
1614 
1615  				if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1616  					goto again;
1617 
1618 				if (clean_only && p->valid) {
1619 					vm_page_test_dirty(p);
1620 					if (p->valid & p->dirty)
1621 						continue;
1622 				}
1623 
1624 				vm_page_busy(p);
1625 				vm_page_protect(p, VM_PROT_NONE);
1626 				vm_page_free(p);
1627 			}
1628 		}
1629 	} else {
1630 		while (size > 0) {
1631 			if ((p = vm_page_lookup(object, start)) != 0) {
1632 
1633 				if (p->wire_count != 0) {
1634 					vm_page_protect(p, VM_PROT_NONE);
1635 					if (!clean_only)
1636 						p->valid = 0;
1637 					start += 1;
1638 					size -= 1;
1639 					continue;
1640 				}
1641 
1642 				/*
1643 				 * The busy flags are only cleared at
1644 				 * interrupt -- minimize the spl transitions
1645 				 */
1646  				if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1647 					goto again;
1648 
1649 				if (clean_only && p->valid) {
1650 					vm_page_test_dirty(p);
1651 					if (p->valid & p->dirty) {
1652 						start += 1;
1653 						size -= 1;
1654 						continue;
1655 					}
1656 				}
1657 
1658 				vm_page_busy(p);
1659 				vm_page_protect(p, VM_PROT_NONE);
1660 				vm_page_free(p);
1661 			}
1662 			start += 1;
1663 			size -= 1;
1664 		}
1665 	}
1666 	vm_object_pip_wakeup(object);
1667 }
1668 
1669 /*
1670  *	Routine:	vm_object_coalesce
1671  *	Function:	Coalesces two objects backing up adjoining
1672  *			regions of memory into a single object.
1673  *
1674  *	returns TRUE if objects were combined.
1675  *
1676  *	NOTE:	Only works at the moment if the second object is NULL -
1677  *		if it's not, which object do we lock first?
1678  *
1679  *	Parameters:
1680  *		prev_object	First object to coalesce
1681  *		prev_offset	Offset into prev_object
1682  *		next_object	Second object into coalesce
1683  *		next_offset	Offset into next_object
1684  *
1685  *		prev_size	Size of reference to prev_object
1686  *		next_size	Size of reference to next_object
1687  *
1688  *	Conditions:
1689  *	The object must *not* be locked.
1690  */
1691 boolean_t
1692 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, vm_size_t prev_size, vm_size_t next_size)
1693 {
1694 	vm_pindex_t next_pindex;
1695 
1696 	GIANT_REQUIRED;
1697 
1698 	if (prev_object == NULL) {
1699 		return (TRUE);
1700 	}
1701 
1702 	if (prev_object->type != OBJT_DEFAULT &&
1703 	    prev_object->type != OBJT_SWAP) {
1704 		return (FALSE);
1705 	}
1706 
1707 	/*
1708 	 * Try to collapse the object first
1709 	 */
1710 	vm_object_collapse(prev_object);
1711 
1712 	/*
1713 	 * Can't coalesce if: . more than one reference . paged out . shadows
1714 	 * another object . has a copy elsewhere (any of which mean that the
1715 	 * pages not mapped to prev_entry may be in use anyway)
1716 	 */
1717 
1718 	if (prev_object->backing_object != NULL) {
1719 		return (FALSE);
1720 	}
1721 
1722 	prev_size >>= PAGE_SHIFT;
1723 	next_size >>= PAGE_SHIFT;
1724 	next_pindex = prev_pindex + prev_size;
1725 
1726 	if ((prev_object->ref_count > 1) &&
1727 	    (prev_object->size != next_pindex)) {
1728 		return (FALSE);
1729 	}
1730 
1731 	/*
1732 	 * Remove any pages that may still be in the object from a previous
1733 	 * deallocation.
1734 	 */
1735 	if (next_pindex < prev_object->size) {
1736 		vm_object_page_remove(prev_object,
1737 				      next_pindex,
1738 				      next_pindex + next_size, FALSE);
1739 		if (prev_object->type == OBJT_SWAP)
1740 			swap_pager_freespace(prev_object,
1741 					     next_pindex, next_size);
1742 	}
1743 
1744 	/*
1745 	 * Extend the object if necessary.
1746 	 */
1747 	if (next_pindex + next_size > prev_object->size)
1748 		prev_object->size = next_pindex + next_size;
1749 
1750 	return (TRUE);
1751 }
1752 
1753 void
1754 vm_object_set_writeable_dirty(vm_object_t object)
1755 {
1756 	struct vnode *vp;
1757 
1758 	vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1759 	if (object->type == OBJT_VNODE &&
1760 	    (vp = (struct vnode *)object->handle) != NULL) {
1761 		if ((vp->v_flag & VOBJDIRTY) == 0) {
1762 			mtx_lock(&vp->v_interlock);
1763 			vp->v_flag |= VOBJDIRTY;
1764 			mtx_unlock(&vp->v_interlock);
1765 		}
1766 	}
1767 }
1768 
1769 #include "opt_ddb.h"
1770 #ifdef DDB
1771 #include <sys/kernel.h>
1772 
1773 #include <sys/cons.h>
1774 
1775 #include <ddb/ddb.h>
1776 
1777 static int
1778 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
1779 {
1780 	vm_map_t tmpm;
1781 	vm_map_entry_t tmpe;
1782 	vm_object_t obj;
1783 	int entcount;
1784 
1785 	if (map == 0)
1786 		return 0;
1787 
1788 	if (entry == 0) {
1789 		tmpe = map->header.next;
1790 		entcount = map->nentries;
1791 		while (entcount-- && (tmpe != &map->header)) {
1792 			if( _vm_object_in_map(map, object, tmpe)) {
1793 				return 1;
1794 			}
1795 			tmpe = tmpe->next;
1796 		}
1797 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
1798 		tmpm = entry->object.sub_map;
1799 		tmpe = tmpm->header.next;
1800 		entcount = tmpm->nentries;
1801 		while (entcount-- && tmpe != &tmpm->header) {
1802 			if( _vm_object_in_map(tmpm, object, tmpe)) {
1803 				return 1;
1804 			}
1805 			tmpe = tmpe->next;
1806 		}
1807 	} else if ((obj = entry->object.vm_object) != NULL) {
1808 		for (; obj; obj = obj->backing_object)
1809 			if( obj == object) {
1810 				return 1;
1811 			}
1812 	}
1813 	return 0;
1814 }
1815 
1816 static int
1817 vm_object_in_map(vm_object_t object)
1818 {
1819 	struct proc *p;
1820 
1821 	/* sx_slock(&allproc_lock); */
1822 	LIST_FOREACH(p, &allproc, p_list) {
1823 		if( !p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1824 			continue;
1825 		if( _vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
1826 			/* sx_sunlock(&allproc_lock); */
1827 			return 1;
1828 		}
1829 	}
1830 	/* sx_sunlock(&allproc_lock); */
1831 	if( _vm_object_in_map( kernel_map, object, 0))
1832 		return 1;
1833 	if( _vm_object_in_map( kmem_map, object, 0))
1834 		return 1;
1835 	if( _vm_object_in_map( pager_map, object, 0))
1836 		return 1;
1837 	if( _vm_object_in_map( buffer_map, object, 0))
1838 		return 1;
1839 	return 0;
1840 }
1841 
1842 DB_SHOW_COMMAND(vmochk, vm_object_check)
1843 {
1844 	vm_object_t object;
1845 
1846 	/*
1847 	 * make sure that internal objs are in a map somewhere
1848 	 * and none have zero ref counts.
1849 	 */
1850 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
1851 		if (object->handle == NULL &&
1852 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1853 			if (object->ref_count == 0) {
1854 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
1855 					(long)object->size);
1856 			}
1857 			if (!vm_object_in_map(object)) {
1858 				db_printf(
1859 			"vmochk: internal obj is not in a map: "
1860 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1861 				    object->ref_count, (u_long)object->size,
1862 				    (u_long)object->size,
1863 				    (void *)object->backing_object);
1864 			}
1865 		}
1866 	}
1867 }
1868 
1869 /*
1870  *	vm_object_print:	[ debug ]
1871  */
1872 DB_SHOW_COMMAND(object, vm_object_print_static)
1873 {
1874 	/* XXX convert args. */
1875 	vm_object_t object = (vm_object_t)addr;
1876 	boolean_t full = have_addr;
1877 
1878 	vm_page_t p;
1879 
1880 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
1881 #define	count	was_count
1882 
1883 	int count;
1884 
1885 	if (object == NULL)
1886 		return;
1887 
1888 	db_iprintf(
1889 	    "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1890 	    object, (int)object->type, (u_long)object->size,
1891 	    object->resident_page_count, object->ref_count, object->flags);
1892 	/*
1893 	 * XXX no %qd in kernel.  Truncate object->backing_object_offset.
1894 	 */
1895 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1896 	    object->shadow_count,
1897 	    object->backing_object ? object->backing_object->ref_count : 0,
1898 	    object->backing_object, (long)object->backing_object_offset);
1899 
1900 	if (!full)
1901 		return;
1902 
1903 	db_indent += 2;
1904 	count = 0;
1905 	TAILQ_FOREACH(p, &object->memq, listq) {
1906 		if (count == 0)
1907 			db_iprintf("memory:=");
1908 		else if (count == 6) {
1909 			db_printf("\n");
1910 			db_iprintf(" ...");
1911 			count = 0;
1912 		} else
1913 			db_printf(",");
1914 		count++;
1915 
1916 		db_printf("(off=0x%lx,page=0x%lx)",
1917 		    (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1918 	}
1919 	if (count != 0)
1920 		db_printf("\n");
1921 	db_indent -= 2;
1922 }
1923 
1924 /* XXX. */
1925 #undef count
1926 
1927 /* XXX need this non-static entry for calling from vm_map_print. */
1928 void
1929 vm_object_print(
1930         /* db_expr_t */ long addr,
1931 	boolean_t have_addr,
1932 	/* db_expr_t */ long count,
1933 	char *modif)
1934 {
1935 	vm_object_print_static(addr, have_addr, count, modif);
1936 }
1937 
1938 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1939 {
1940 	vm_object_t object;
1941 	int nl = 0;
1942 	int c;
1943 
1944 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
1945 		vm_pindex_t idx, fidx;
1946 		vm_pindex_t osize;
1947 		vm_offset_t pa = -1, padiff;
1948 		int rcount;
1949 		vm_page_t m;
1950 
1951 		db_printf("new object: %p\n", (void *)object);
1952 		if ( nl > 18) {
1953 			c = cngetc();
1954 			if (c != ' ')
1955 				return;
1956 			nl = 0;
1957 		}
1958 		nl++;
1959 		rcount = 0;
1960 		fidx = 0;
1961 		osize = object->size;
1962 		if (osize > 128)
1963 			osize = 128;
1964 		for (idx = 0; idx < osize; idx++) {
1965 			m = vm_page_lookup(object, idx);
1966 			if (m == NULL) {
1967 				if (rcount) {
1968 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1969 						(long)fidx, rcount, (long)pa);
1970 					if ( nl > 18) {
1971 						c = cngetc();
1972 						if (c != ' ')
1973 							return;
1974 						nl = 0;
1975 					}
1976 					nl++;
1977 					rcount = 0;
1978 				}
1979 				continue;
1980 			}
1981 
1982 
1983 			if (rcount &&
1984 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1985 				++rcount;
1986 				continue;
1987 			}
1988 			if (rcount) {
1989 				padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1990 				padiff >>= PAGE_SHIFT;
1991 				padiff &= PQ_L2_MASK;
1992 				if (padiff == 0) {
1993 					pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1994 					++rcount;
1995 					continue;
1996 				}
1997 				db_printf(" index(%ld)run(%d)pa(0x%lx)",
1998 					(long)fidx, rcount, (long)pa);
1999 				db_printf("pd(%ld)\n", (long)padiff);
2000 				if ( nl > 18) {
2001 					c = cngetc();
2002 					if (c != ' ')
2003 						return;
2004 					nl = 0;
2005 				}
2006 				nl++;
2007 			}
2008 			fidx = idx;
2009 			pa = VM_PAGE_TO_PHYS(m);
2010 			rcount = 1;
2011 		}
2012 		if (rcount) {
2013 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2014 				(long)fidx, rcount, (long)pa);
2015 			if ( nl > 18) {
2016 				c = cngetc();
2017 				if (c != ' ')
2018 					return;
2019 				nl = 0;
2020 			}
2021 			nl++;
2022 		}
2023 	}
2024 }
2025 #endif /* DDB */
2026