xref: /dragonfly/sys/vm/vm_object.c (revision 984263bc)
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  */
66 
67 /*
68  *	Virtual memory object module.
69  */
70 
71 #include <sys/param.h>
72 #include <sys/systm.h>
73 #include <sys/proc.h>		/* for curproc, pageproc */
74 #include <sys/vnode.h>
75 #include <sys/vmmeter.h>
76 #include <sys/mman.h>
77 #include <sys/mount.h>
78 #include <sys/kernel.h>
79 #include <sys/sysctl.h>
80 
81 #include <vm/vm.h>
82 #include <vm/vm_param.h>
83 #include <vm/pmap.h>
84 #include <vm/vm_map.h>
85 #include <vm/vm_object.h>
86 #include <vm/vm_page.h>
87 #include <vm/vm_pageout.h>
88 #include <vm/vm_pager.h>
89 #include <vm/swap_pager.h>
90 #include <vm/vm_kern.h>
91 #include <vm/vm_extern.h>
92 #include <vm/vm_zone.h>
93 
94 #define EASY_SCAN_FACTOR	8
95 
96 #define MSYNC_FLUSH_HARDSEQ	0x01
97 #define MSYNC_FLUSH_SOFTSEQ	0x02
98 
99 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ;
100 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags,
101         CTLFLAG_RW, &msync_flush_flags, 0, "");
102 
103 static void	vm_object_qcollapse (vm_object_t object);
104 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
105 
106 /*
107  *	Virtual memory objects maintain the actual data
108  *	associated with allocated virtual memory.  A given
109  *	page of memory exists within exactly one object.
110  *
111  *	An object is only deallocated when all "references"
112  *	are given up.  Only one "reference" to a given
113  *	region of an object should be writeable.
114  *
115  *	Associated with each object is a list of all resident
116  *	memory pages belonging to that object; this list is
117  *	maintained by the "vm_page" module, and locked by the object's
118  *	lock.
119  *
120  *	Each object also records a "pager" routine which is
121  *	used to retrieve (and store) pages to the proper backing
122  *	storage.  In addition, objects may be backed by other
123  *	objects from which they were virtual-copied.
124  *
125  *	The only items within the object structure which are
126  *	modified after time of creation are:
127  *		reference count		locked by object's lock
128  *		pager routine		locked by object's lock
129  *
130  */
131 
132 struct object_q vm_object_list;
133 #ifndef NULL_SIMPLELOCKS
134 static struct simplelock vm_object_list_lock;
135 #endif
136 static long vm_object_count;		/* count of all objects */
137 vm_object_t kernel_object;
138 vm_object_t kmem_object;
139 static struct vm_object kernel_object_store;
140 static struct vm_object kmem_object_store;
141 extern int vm_pageout_page_count;
142 
143 static long object_collapses;
144 static long object_bypasses;
145 static int next_index;
146 static vm_zone_t obj_zone;
147 static struct vm_zone obj_zone_store;
148 static int object_hash_rand;
149 #define VM_OBJECTS_INIT 256
150 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
151 
152 void
153 _vm_object_allocate(type, size, object)
154 	objtype_t type;
155 	vm_size_t size;
156 	vm_object_t object;
157 {
158 	int incr;
159 	TAILQ_INIT(&object->memq);
160 	LIST_INIT(&object->shadow_head);
161 
162 	object->type = type;
163 	object->size = size;
164 	object->ref_count = 1;
165 	object->flags = 0;
166 	if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
167 		vm_object_set_flag(object, OBJ_ONEMAPPING);
168 	object->paging_in_progress = 0;
169 	object->resident_page_count = 0;
170 	object->shadow_count = 0;
171 	object->pg_color = next_index;
172 	if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
173 		incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
174 	else
175 		incr = size;
176 	next_index = (next_index + incr) & PQ_L2_MASK;
177 	object->handle = NULL;
178 	object->backing_object = NULL;
179 	object->backing_object_offset = (vm_ooffset_t) 0;
180 	/*
181 	 * Try to generate a number that will spread objects out in the
182 	 * hash table.  We 'wipe' new objects across the hash in 128 page
183 	 * increments plus 1 more to offset it a little more by the time
184 	 * it wraps around.
185 	 */
186 	object->hash_rand = object_hash_rand - 129;
187 
188 	object->generation++;
189 
190 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
191 	vm_object_count++;
192 	object_hash_rand = object->hash_rand;
193 }
194 
195 /*
196  *	vm_object_init:
197  *
198  *	Initialize the VM objects module.
199  */
200 void
201 vm_object_init()
202 {
203 	TAILQ_INIT(&vm_object_list);
204 	simple_lock_init(&vm_object_list_lock);
205 	vm_object_count = 0;
206 
207 	kernel_object = &kernel_object_store;
208 	_vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
209 	    kernel_object);
210 
211 	kmem_object = &kmem_object_store;
212 	_vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
213 	    kmem_object);
214 
215 	obj_zone = &obj_zone_store;
216 	zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
217 		vm_objects_init, VM_OBJECTS_INIT);
218 }
219 
220 void
221 vm_object_init2() {
222 	zinitna(obj_zone, NULL, NULL, 0, 0, 0, 1);
223 }
224 
225 /*
226  *	vm_object_allocate:
227  *
228  *	Returns a new object with the given size.
229  */
230 
231 vm_object_t
232 vm_object_allocate(type, size)
233 	objtype_t type;
234 	vm_size_t size;
235 {
236 	vm_object_t result;
237 
238 	result = (vm_object_t) zalloc(obj_zone);
239 
240 	_vm_object_allocate(type, size, result);
241 
242 	return (result);
243 }
244 
245 
246 /*
247  *	vm_object_reference:
248  *
249  *	Gets another reference to the given object.
250  */
251 void
252 vm_object_reference(object)
253 	vm_object_t object;
254 {
255 	if (object == NULL)
256 		return;
257 
258 #if 0
259 	/* object can be re-referenced during final cleaning */
260 	KASSERT(!(object->flags & OBJ_DEAD),
261 	    ("vm_object_reference: attempting to reference dead obj"));
262 #endif
263 
264 	object->ref_count++;
265 	if (object->type == OBJT_VNODE) {
266 		while (vget((struct vnode *) object->handle, LK_RETRY|LK_NOOBJ, curproc)) {
267 			printf("vm_object_reference: delay in getting object\n");
268 		}
269 	}
270 }
271 
272 void
273 vm_object_vndeallocate(object)
274 	vm_object_t object;
275 {
276 	struct vnode *vp = (struct vnode *) object->handle;
277 
278 	KASSERT(object->type == OBJT_VNODE,
279 	    ("vm_object_vndeallocate: not a vnode object"));
280 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
281 #ifdef INVARIANTS
282 	if (object->ref_count == 0) {
283 		vprint("vm_object_vndeallocate", vp);
284 		panic("vm_object_vndeallocate: bad object reference count");
285 	}
286 #endif
287 
288 	object->ref_count--;
289 	if (object->ref_count == 0) {
290 		vp->v_flag &= ~VTEXT;
291 		vm_object_clear_flag(object, OBJ_OPT);
292 	}
293 	vrele(vp);
294 }
295 
296 /*
297  *	vm_object_deallocate:
298  *
299  *	Release a reference to the specified object,
300  *	gained either through a vm_object_allocate
301  *	or a vm_object_reference call.  When all references
302  *	are gone, storage associated with this object
303  *	may be relinquished.
304  *
305  *	No object may be locked.
306  */
307 void
308 vm_object_deallocate(object)
309 	vm_object_t object;
310 {
311 	vm_object_t temp;
312 
313 	while (object != NULL) {
314 
315 		if (object->type == OBJT_VNODE) {
316 			vm_object_vndeallocate(object);
317 			return;
318 		}
319 
320 		if (object->ref_count == 0) {
321 			panic("vm_object_deallocate: object deallocated too many times: %d", object->type);
322 		} else if (object->ref_count > 2) {
323 			object->ref_count--;
324 			return;
325 		}
326 
327 		/*
328 		 * Here on ref_count of one or two, which are special cases for
329 		 * objects.
330 		 */
331 		if ((object->ref_count == 2) && (object->shadow_count == 0)) {
332 			vm_object_set_flag(object, OBJ_ONEMAPPING);
333 			object->ref_count--;
334 			return;
335 		} else if ((object->ref_count == 2) && (object->shadow_count == 1)) {
336 			object->ref_count--;
337 			if ((object->handle == NULL) &&
338 			    (object->type == OBJT_DEFAULT ||
339 			     object->type == OBJT_SWAP)) {
340 				vm_object_t robject;
341 
342 				robject = LIST_FIRST(&object->shadow_head);
343 				KASSERT(robject != NULL,
344 				    ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
345 					 object->ref_count,
346 					 object->shadow_count));
347 				if ((robject->handle == NULL) &&
348 				    (robject->type == OBJT_DEFAULT ||
349 				     robject->type == OBJT_SWAP)) {
350 
351 					robject->ref_count++;
352 
353 					while (
354 						robject->paging_in_progress ||
355 						object->paging_in_progress
356 					) {
357 						vm_object_pip_sleep(robject, "objde1");
358 						vm_object_pip_sleep(object, "objde2");
359 					}
360 
361 					if (robject->ref_count == 1) {
362 						robject->ref_count--;
363 						object = robject;
364 						goto doterm;
365 					}
366 
367 					object = robject;
368 					vm_object_collapse(object);
369 					continue;
370 				}
371 			}
372 
373 			return;
374 
375 		} else {
376 			object->ref_count--;
377 			if (object->ref_count != 0)
378 				return;
379 		}
380 
381 doterm:
382 
383 		temp = object->backing_object;
384 		if (temp) {
385 			LIST_REMOVE(object, shadow_list);
386 			temp->shadow_count--;
387 			if (temp->ref_count == 0)
388 				vm_object_clear_flag(temp, OBJ_OPT);
389 			temp->generation++;
390 			object->backing_object = NULL;
391 		}
392 
393 		/*
394 		 * Don't double-terminate, we could be in a termination
395 		 * recursion due to the terminate having to sync data
396 		 * to disk.
397 		 */
398 		if ((object->flags & OBJ_DEAD) == 0)
399 			vm_object_terminate(object);
400 		object = temp;
401 	}
402 }
403 
404 /*
405  *	vm_object_terminate actually destroys the specified object, freeing
406  *	up all previously used resources.
407  *
408  *	The object must be locked.
409  *	This routine may block.
410  */
411 void
412 vm_object_terminate(object)
413 	vm_object_t object;
414 {
415 	vm_page_t p;
416 	int s;
417 
418 	/*
419 	 * Make sure no one uses us.
420 	 */
421 	vm_object_set_flag(object, OBJ_DEAD);
422 
423 	/*
424 	 * wait for the pageout daemon to be done with the object
425 	 */
426 	vm_object_pip_wait(object, "objtrm");
427 
428 	KASSERT(!object->paging_in_progress,
429 		("vm_object_terminate: pageout in progress"));
430 
431 	/*
432 	 * Clean and free the pages, as appropriate. All references to the
433 	 * object are gone, so we don't need to lock it.
434 	 */
435 	if (object->type == OBJT_VNODE) {
436 		struct vnode *vp;
437 
438 		/*
439 		 * Freeze optimized copies.
440 		 */
441 		vm_freeze_copyopts(object, 0, object->size);
442 
443 		/*
444 		 * Clean pages and flush buffers.
445 		 */
446 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
447 
448 		vp = (struct vnode *) object->handle;
449 		vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0);
450 	}
451 
452 	/*
453 	 * Wait for any I/O to complete, after which there had better not
454 	 * be any references left on the object.
455 	 */
456 	vm_object_pip_wait(object, "objtrm");
457 
458 	if (object->ref_count != 0)
459 		panic("vm_object_terminate: object with references, ref_count=%d", object->ref_count);
460 
461 	/*
462 	 * Now free any remaining pages. For internal objects, this also
463 	 * removes them from paging queues. Don't free wired pages, just
464 	 * remove them from the object.
465 	 */
466 	s = splvm();
467 	while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
468 		if (p->busy || (p->flags & PG_BUSY))
469 			panic("vm_object_terminate: freeing busy page %p\n", p);
470 		if (p->wire_count == 0) {
471 			vm_page_busy(p);
472 			vm_page_free(p);
473 			cnt.v_pfree++;
474 		} else {
475 			vm_page_busy(p);
476 			vm_page_remove(p);
477 		}
478 	}
479 	splx(s);
480 
481 	/*
482 	 * Let the pager know object is dead.
483 	 */
484 	vm_pager_deallocate(object);
485 
486 	/*
487 	 * Remove the object from the global object list.
488 	 */
489 	simple_lock(&vm_object_list_lock);
490 	TAILQ_REMOVE(&vm_object_list, object, object_list);
491 	simple_unlock(&vm_object_list_lock);
492 
493 	wakeup(object);
494 
495 	/*
496 	 * Free the space for the object.
497 	 */
498 	zfree(obj_zone, object);
499 }
500 
501 /*
502  *	vm_object_page_clean
503  *
504  *	Clean all dirty pages in the specified range of object.  Leaves page
505  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
506  *	write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
507  *	leaving the object dirty.
508  *
509  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
510  *	synchronous clustering mode implementation.
511  *
512  *	Odd semantics: if start == end, we clean everything.
513  *
514  *	The object must be locked.
515  */
516 
517 void
518 vm_object_page_clean(object, start, end, flags)
519 	vm_object_t object;
520 	vm_pindex_t start;
521 	vm_pindex_t end;
522 	int flags;
523 {
524 	vm_page_t p, np;
525 	vm_offset_t tstart, tend;
526 	vm_pindex_t pi;
527 	struct vnode *vp;
528 	int clearobjflags;
529 	int pagerflags;
530 	int curgeneration;
531 
532 	if (object->type != OBJT_VNODE ||
533 		(object->flags & OBJ_MIGHTBEDIRTY) == 0)
534 		return;
535 
536 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
537 	pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
538 
539 	vp = object->handle;
540 
541 	vm_object_set_flag(object, OBJ_CLEANING);
542 
543 	/*
544 	 * Handle 'entire object' case
545 	 */
546 	tstart = start;
547 	if (end == 0) {
548 		tend = object->size;
549 	} else {
550 		tend = end;
551 	}
552 
553 	/*
554 	 * If the caller is smart and only msync()s a range he knows is
555 	 * dirty, we may be able to avoid an object scan.  This results in
556 	 * a phenominal improvement in performance.  We cannot do this
557 	 * as a matter of course because the object may be huge - e.g.
558 	 * the size might be in the gigabytes or terrabytes.
559 	 */
560 	if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
561 		vm_offset_t tscan;
562 		int scanlimit;
563 		int scanreset;
564 
565 		scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
566 		if (scanreset < 16)
567 			scanreset = 16;
568 		pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
569 
570 		scanlimit = scanreset;
571 		tscan = tstart;
572 		while (tscan < tend) {
573 			curgeneration = object->generation;
574 			p = vm_page_lookup(object, tscan);
575 			if (p == NULL || p->valid == 0 ||
576 			    (p->queue - p->pc) == PQ_CACHE) {
577 				if (--scanlimit == 0)
578 					break;
579 				++tscan;
580 				continue;
581 			}
582 			vm_page_test_dirty(p);
583 			if ((p->dirty & p->valid) == 0) {
584 				if (--scanlimit == 0)
585 					break;
586 				++tscan;
587 				continue;
588 			}
589 			/*
590 			 * If we have been asked to skip nosync pages and
591 			 * this is a nosync page, we can't continue.
592 			 */
593 			if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
594 				if (--scanlimit == 0)
595 					break;
596 				++tscan;
597 				continue;
598 			}
599 			scanlimit = scanreset;
600 
601 			/*
602 			 * This returns 0 if it was unable to busy the first
603 			 * page (i.e. had to sleep).
604 			 */
605 			tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
606 		}
607 
608 		/*
609 		 * If everything was dirty and we flushed it successfully,
610 		 * and the requested range is not the entire object, we
611 		 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
612 		 * return immediately.
613 		 */
614 		if (tscan >= tend && (tstart || tend < object->size)) {
615 			vm_object_clear_flag(object, OBJ_CLEANING);
616 			return;
617 		}
618 		pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
619 	}
620 
621 	/*
622 	 * Generally set CLEANCHK interlock and make the page read-only so
623 	 * we can then clear the object flags.
624 	 *
625 	 * However, if this is a nosync mmap then the object is likely to
626 	 * stay dirty so do not mess with the page and do not clear the
627 	 * object flags.
628 	 */
629 
630 	clearobjflags = 1;
631 
632 	for(p = TAILQ_FIRST(&object->memq); p; p = TAILQ_NEXT(p, listq)) {
633 		vm_page_flag_set(p, PG_CLEANCHK);
634 		if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
635 			clearobjflags = 0;
636 		else
637 			vm_page_protect(p, VM_PROT_READ);
638 	}
639 
640 	if (clearobjflags && (tstart == 0) && (tend == object->size)) {
641 		struct vnode *vp;
642 
643 		vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
644                 if (object->type == OBJT_VNODE &&
645                     (vp = (struct vnode *)object->handle) != NULL) {
646                         if (vp->v_flag & VOBJDIRTY) {
647                                 simple_lock(&vp->v_interlock);
648                                 vp->v_flag &= ~VOBJDIRTY;
649                                 simple_unlock(&vp->v_interlock);
650                         }
651                 }
652 	}
653 
654 rescan:
655 	curgeneration = object->generation;
656 
657 	for(p = TAILQ_FIRST(&object->memq); p; p = np) {
658 		int n;
659 
660 		np = TAILQ_NEXT(p, listq);
661 
662 again:
663 		pi = p->pindex;
664 		if (((p->flags & PG_CLEANCHK) == 0) ||
665 			(pi < tstart) || (pi >= tend) ||
666 			(p->valid == 0) ||
667 			((p->queue - p->pc) == PQ_CACHE)) {
668 			vm_page_flag_clear(p, PG_CLEANCHK);
669 			continue;
670 		}
671 
672 		vm_page_test_dirty(p);
673 		if ((p->dirty & p->valid) == 0) {
674 			vm_page_flag_clear(p, PG_CLEANCHK);
675 			continue;
676 		}
677 
678 		/*
679 		 * If we have been asked to skip nosync pages and this is a
680 		 * nosync page, skip it.  Note that the object flags were
681 		 * not cleared in this case so we do not have to set them.
682 		 */
683 		if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
684 			vm_page_flag_clear(p, PG_CLEANCHK);
685 			continue;
686 		}
687 
688 		n = vm_object_page_collect_flush(object, p,
689 			curgeneration, pagerflags);
690 		if (n == 0)
691 			goto rescan;
692 		if (object->generation != curgeneration)
693 			goto rescan;
694 
695 		/*
696 		 * Try to optimize the next page.  If we can't we pick up
697 		 * our (random) scan where we left off.
698 		 */
699 		if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
700 			if ((p = vm_page_lookup(object, pi + n)) != NULL)
701 				goto again;
702 		}
703 	}
704 
705 #if 0
706 	VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
707 #endif
708 
709 	vm_object_clear_flag(object, OBJ_CLEANING);
710 	return;
711 }
712 
713 static int
714 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
715 {
716 	int runlen;
717 	int s;
718 	int maxf;
719 	int chkb;
720 	int maxb;
721 	int i;
722 	vm_pindex_t pi;
723 	vm_page_t maf[vm_pageout_page_count];
724 	vm_page_t mab[vm_pageout_page_count];
725 	vm_page_t ma[vm_pageout_page_count];
726 
727 	s = splvm();
728 	pi = p->pindex;
729 	while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
730 		if (object->generation != curgeneration) {
731 			splx(s);
732 			return(0);
733 		}
734 	}
735 
736 	maxf = 0;
737 	for(i = 1; i < vm_pageout_page_count; i++) {
738 		vm_page_t tp;
739 
740 		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
741 			if ((tp->flags & PG_BUSY) ||
742 				((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
743 				 (tp->flags & PG_CLEANCHK) == 0) ||
744 				(tp->busy != 0))
745 				break;
746 			if((tp->queue - tp->pc) == PQ_CACHE) {
747 				vm_page_flag_clear(tp, PG_CLEANCHK);
748 				break;
749 			}
750 			vm_page_test_dirty(tp);
751 			if ((tp->dirty & tp->valid) == 0) {
752 				vm_page_flag_clear(tp, PG_CLEANCHK);
753 				break;
754 			}
755 			maf[ i - 1 ] = tp;
756 			maxf++;
757 			continue;
758 		}
759 		break;
760 	}
761 
762 	maxb = 0;
763 	chkb = vm_pageout_page_count -  maxf;
764 	if (chkb) {
765 		for(i = 1; i < chkb;i++) {
766 			vm_page_t tp;
767 
768 			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
769 				if ((tp->flags & PG_BUSY) ||
770 					((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
771 					 (tp->flags & PG_CLEANCHK) == 0) ||
772 					(tp->busy != 0))
773 					break;
774 				if((tp->queue - tp->pc) == PQ_CACHE) {
775 					vm_page_flag_clear(tp, PG_CLEANCHK);
776 					break;
777 				}
778 				vm_page_test_dirty(tp);
779 				if ((tp->dirty & tp->valid) == 0) {
780 					vm_page_flag_clear(tp, PG_CLEANCHK);
781 					break;
782 				}
783 				mab[ i - 1 ] = tp;
784 				maxb++;
785 				continue;
786 			}
787 			break;
788 		}
789 	}
790 
791 	for(i = 0; i < maxb; i++) {
792 		int index = (maxb - i) - 1;
793 		ma[index] = mab[i];
794 		vm_page_flag_clear(ma[index], PG_CLEANCHK);
795 	}
796 	vm_page_flag_clear(p, PG_CLEANCHK);
797 	ma[maxb] = p;
798 	for(i = 0; i < maxf; i++) {
799 		int index = (maxb + i) + 1;
800 		ma[index] = maf[i];
801 		vm_page_flag_clear(ma[index], PG_CLEANCHK);
802 	}
803 	runlen = maxb + maxf + 1;
804 
805 	splx(s);
806 	vm_pageout_flush(ma, runlen, pagerflags);
807 	for (i = 0; i < runlen; i++) {
808 		if (ma[i]->valid & ma[i]->dirty) {
809 			vm_page_protect(ma[i], VM_PROT_READ);
810 			vm_page_flag_set(ma[i], PG_CLEANCHK);
811 
812 			/*
813 			 * maxf will end up being the actual number of pages
814 			 * we wrote out contiguously, non-inclusive of the
815 			 * first page.  We do not count look-behind pages.
816 			 */
817 			if (i >= maxb + 1 && (maxf > i - maxb - 1))
818 				maxf = i - maxb - 1;
819 		}
820 	}
821 	return(maxf + 1);
822 }
823 
824 #ifdef not_used
825 /* XXX I cannot tell if this should be an exported symbol */
826 /*
827  *	vm_object_deactivate_pages
828  *
829  *	Deactivate all pages in the specified object.  (Keep its pages
830  *	in memory even though it is no longer referenced.)
831  *
832  *	The object must be locked.
833  */
834 static void
835 vm_object_deactivate_pages(object)
836 	vm_object_t object;
837 {
838 	vm_page_t p, next;
839 
840 	for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
841 		next = TAILQ_NEXT(p, listq);
842 		vm_page_deactivate(p);
843 	}
844 }
845 #endif
846 
847 /*
848  * Same as vm_object_pmap_copy, except range checking really
849  * works, and is meant for small sections of an object.
850  *
851  * This code protects resident pages by making them read-only
852  * and is typically called on a fork or split when a page
853  * is converted to copy-on-write.
854  *
855  * NOTE: If the page is already at VM_PROT_NONE, calling
856  * vm_page_protect will have no effect.
857  */
858 
859 void
860 vm_object_pmap_copy_1(object, start, end)
861 	vm_object_t object;
862 	vm_pindex_t start;
863 	vm_pindex_t end;
864 {
865 	vm_pindex_t idx;
866 	vm_page_t p;
867 
868 	if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
869 		return;
870 
871 	for (idx = start; idx < end; idx++) {
872 		p = vm_page_lookup(object, idx);
873 		if (p == NULL)
874 			continue;
875 		vm_page_protect(p, VM_PROT_READ);
876 	}
877 }
878 
879 /*
880  *	vm_object_pmap_remove:
881  *
882  *	Removes all physical pages in the specified
883  *	object range from all physical maps.
884  *
885  *	The object must *not* be locked.
886  */
887 void
888 vm_object_pmap_remove(object, start, end)
889 	vm_object_t object;
890 	vm_pindex_t start;
891 	vm_pindex_t end;
892 {
893 	vm_page_t p;
894 
895 	if (object == NULL)
896 		return;
897 	for (p = TAILQ_FIRST(&object->memq);
898 		p != NULL;
899 		p = TAILQ_NEXT(p, listq)) {
900 		if (p->pindex >= start && p->pindex < end)
901 			vm_page_protect(p, VM_PROT_NONE);
902 	}
903 	if ((start == 0) && (object->size == end))
904 		vm_object_clear_flag(object, OBJ_WRITEABLE);
905 }
906 
907 /*
908  *	vm_object_madvise:
909  *
910  *	Implements the madvise function at the object/page level.
911  *
912  *	MADV_WILLNEED	(any object)
913  *
914  *	    Activate the specified pages if they are resident.
915  *
916  *	MADV_DONTNEED	(any object)
917  *
918  *	    Deactivate the specified pages if they are resident.
919  *
920  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
921  *			 OBJ_ONEMAPPING only)
922  *
923  *	    Deactivate and clean the specified pages if they are
924  *	    resident.  This permits the process to reuse the pages
925  *	    without faulting or the kernel to reclaim the pages
926  *	    without I/O.
927  */
928 void
929 vm_object_madvise(object, pindex, count, advise)
930 	vm_object_t object;
931 	vm_pindex_t pindex;
932 	int count;
933 	int advise;
934 {
935 	vm_pindex_t end, tpindex;
936 	vm_object_t tobject;
937 	vm_page_t m;
938 
939 	if (object == NULL)
940 		return;
941 
942 	end = pindex + count;
943 
944 	/*
945 	 * Locate and adjust resident pages
946 	 */
947 
948 	for (; pindex < end; pindex += 1) {
949 relookup:
950 		tobject = object;
951 		tpindex = pindex;
952 shadowlookup:
953 		/*
954 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
955 		 * and those pages must be OBJ_ONEMAPPING.
956 		 */
957 		if (advise == MADV_FREE) {
958 			if ((tobject->type != OBJT_DEFAULT &&
959 			     tobject->type != OBJT_SWAP) ||
960 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
961 				continue;
962 			}
963 		}
964 
965 		m = vm_page_lookup(tobject, tpindex);
966 
967 		if (m == NULL) {
968 			/*
969 			 * There may be swap even if there is no backing page
970 			 */
971 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
972 				swap_pager_freespace(tobject, tpindex, 1);
973 
974 			/*
975 			 * next object
976 			 */
977 			tobject = tobject->backing_object;
978 			if (tobject == NULL)
979 				continue;
980 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
981 			goto shadowlookup;
982 		}
983 
984 		/*
985 		 * If the page is busy or not in a normal active state,
986 		 * we skip it.  If the page is not managed there are no
987 		 * page queues to mess with.  Things can break if we mess
988 		 * with pages in any of the below states.
989 		 */
990 		if (
991 		    m->hold_count ||
992 		    m->wire_count ||
993 		    (m->flags & PG_UNMANAGED) ||
994 		    m->valid != VM_PAGE_BITS_ALL
995 		) {
996 			continue;
997 		}
998 
999  		if (vm_page_sleep_busy(m, TRUE, "madvpo"))
1000   			goto relookup;
1001 
1002 		if (advise == MADV_WILLNEED) {
1003 			vm_page_activate(m);
1004 		} else if (advise == MADV_DONTNEED) {
1005 			vm_page_dontneed(m);
1006 		} else if (advise == MADV_FREE) {
1007 			/*
1008 			 * Mark the page clean.  This will allow the page
1009 			 * to be freed up by the system.  However, such pages
1010 			 * are often reused quickly by malloc()/free()
1011 			 * so we do not do anything that would cause
1012 			 * a page fault if we can help it.
1013 			 *
1014 			 * Specifically, we do not try to actually free
1015 			 * the page now nor do we try to put it in the
1016 			 * cache (which would cause a page fault on reuse).
1017 			 *
1018 			 * But we do make the page is freeable as we
1019 			 * can without actually taking the step of unmapping
1020 			 * it.
1021 			 */
1022 			pmap_clear_modify(m);
1023 			m->dirty = 0;
1024 			m->act_count = 0;
1025 			vm_page_dontneed(m);
1026 			if (tobject->type == OBJT_SWAP)
1027 				swap_pager_freespace(tobject, tpindex, 1);
1028 		}
1029 	}
1030 }
1031 
1032 /*
1033  *	vm_object_shadow:
1034  *
1035  *	Create a new object which is backed by the
1036  *	specified existing object range.  The source
1037  *	object reference is deallocated.
1038  *
1039  *	The new object and offset into that object
1040  *	are returned in the source parameters.
1041  */
1042 
1043 void
1044 vm_object_shadow(object, offset, length)
1045 	vm_object_t *object;	/* IN/OUT */
1046 	vm_ooffset_t *offset;	/* IN/OUT */
1047 	vm_size_t length;
1048 {
1049 	vm_object_t source;
1050 	vm_object_t result;
1051 
1052 	source = *object;
1053 
1054 	/*
1055 	 * Don't create the new object if the old object isn't shared.
1056 	 */
1057 
1058 	if (source != NULL &&
1059 	    source->ref_count == 1 &&
1060 	    source->handle == NULL &&
1061 	    (source->type == OBJT_DEFAULT ||
1062 	     source->type == OBJT_SWAP))
1063 		return;
1064 
1065 	/*
1066 	 * Allocate a new object with the given length
1067 	 */
1068 
1069 	if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1070 		panic("vm_object_shadow: no object for shadowing");
1071 
1072 	/*
1073 	 * The new object shadows the source object, adding a reference to it.
1074 	 * Our caller changes his reference to point to the new object,
1075 	 * removing a reference to the source object.  Net result: no change
1076 	 * of reference count.
1077 	 *
1078 	 * Try to optimize the result object's page color when shadowing
1079 	 * in order to maintain page coloring consistency in the combined
1080 	 * shadowed object.
1081 	 */
1082 	result->backing_object = source;
1083 	if (source) {
1084 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1085 		source->shadow_count++;
1086 		source->generation++;
1087 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1088 	}
1089 
1090 	/*
1091 	 * Store the offset into the source object, and fix up the offset into
1092 	 * the new object.
1093 	 */
1094 
1095 	result->backing_object_offset = *offset;
1096 
1097 	/*
1098 	 * Return the new things
1099 	 */
1100 
1101 	*offset = 0;
1102 	*object = result;
1103 }
1104 
1105 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1106 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1107 #define	OBSC_COLLAPSE_WAIT	0x0004
1108 
1109 static __inline int
1110 vm_object_backing_scan(vm_object_t object, int op)
1111 {
1112 	int s;
1113 	int r = 1;
1114 	vm_page_t p;
1115 	vm_object_t backing_object;
1116 	vm_pindex_t backing_offset_index;
1117 
1118 	s = splvm();
1119 
1120 	backing_object = object->backing_object;
1121 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1122 
1123 	/*
1124 	 * Initial conditions
1125 	 */
1126 
1127 	if (op & OBSC_TEST_ALL_SHADOWED) {
1128 		/*
1129 		 * We do not want to have to test for the existence of
1130 		 * swap pages in the backing object.  XXX but with the
1131 		 * new swapper this would be pretty easy to do.
1132 		 *
1133 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1134 		 * been ZFOD faulted yet?  If we do not test for this, the
1135 		 * shadow test may succeed! XXX
1136 		 */
1137 		if (backing_object->type != OBJT_DEFAULT) {
1138 			splx(s);
1139 			return(0);
1140 		}
1141 	}
1142 	if (op & OBSC_COLLAPSE_WAIT) {
1143 		vm_object_set_flag(backing_object, OBJ_DEAD);
1144 	}
1145 
1146 	/*
1147 	 * Our scan
1148 	 */
1149 
1150 	p = TAILQ_FIRST(&backing_object->memq);
1151 	while (p) {
1152 		vm_page_t next = TAILQ_NEXT(p, listq);
1153 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1154 
1155 		if (op & OBSC_TEST_ALL_SHADOWED) {
1156 			vm_page_t pp;
1157 
1158 			/*
1159 			 * Ignore pages outside the parent object's range
1160 			 * and outside the parent object's mapping of the
1161 			 * backing object.
1162 			 *
1163 			 * note that we do not busy the backing object's
1164 			 * page.
1165 			 */
1166 
1167 			if (
1168 			    p->pindex < backing_offset_index ||
1169 			    new_pindex >= object->size
1170 			) {
1171 				p = next;
1172 				continue;
1173 			}
1174 
1175 			/*
1176 			 * See if the parent has the page or if the parent's
1177 			 * object pager has the page.  If the parent has the
1178 			 * page but the page is not valid, the parent's
1179 			 * object pager must have the page.
1180 			 *
1181 			 * If this fails, the parent does not completely shadow
1182 			 * the object and we might as well give up now.
1183 			 */
1184 
1185 			pp = vm_page_lookup(object, new_pindex);
1186 			if (
1187 			    (pp == NULL || pp->valid == 0) &&
1188 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1189 			) {
1190 				r = 0;
1191 				break;
1192 			}
1193 		}
1194 
1195 		/*
1196 		 * Check for busy page
1197 		 */
1198 
1199 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1200 			vm_page_t pp;
1201 
1202 			if (op & OBSC_COLLAPSE_NOWAIT) {
1203 				if (
1204 				    (p->flags & PG_BUSY) ||
1205 				    !p->valid ||
1206 				    p->hold_count ||
1207 				    p->wire_count ||
1208 				    p->busy
1209 				) {
1210 					p = next;
1211 					continue;
1212 				}
1213 			} else if (op & OBSC_COLLAPSE_WAIT) {
1214 				if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1215 					/*
1216 					 * If we slept, anything could have
1217 					 * happened.  Since the object is
1218 					 * marked dead, the backing offset
1219 					 * should not have changed so we
1220 					 * just restart our scan.
1221 					 */
1222 					p = TAILQ_FIRST(&backing_object->memq);
1223 					continue;
1224 				}
1225 			}
1226 
1227 			/*
1228 			 * Busy the page
1229 			 */
1230 			vm_page_busy(p);
1231 
1232 			KASSERT(
1233 			    p->object == backing_object,
1234 			    ("vm_object_qcollapse(): object mismatch")
1235 			);
1236 
1237 			/*
1238 			 * Destroy any associated swap
1239 			 */
1240 			if (backing_object->type == OBJT_SWAP) {
1241 				swap_pager_freespace(
1242 				    backing_object,
1243 				    p->pindex,
1244 				    1
1245 				);
1246 			}
1247 
1248 			if (
1249 			    p->pindex < backing_offset_index ||
1250 			    new_pindex >= object->size
1251 			) {
1252 				/*
1253 				 * Page is out of the parent object's range, we
1254 				 * can simply destroy it.
1255 				 */
1256 				vm_page_protect(p, VM_PROT_NONE);
1257 				vm_page_free(p);
1258 				p = next;
1259 				continue;
1260 			}
1261 
1262 			pp = vm_page_lookup(object, new_pindex);
1263 			if (
1264 			    pp != NULL ||
1265 			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1266 			) {
1267 				/*
1268 				 * page already exists in parent OR swap exists
1269 				 * for this location in the parent.  Destroy
1270 				 * the original page from the backing object.
1271 				 *
1272 				 * Leave the parent's page alone
1273 				 */
1274 				vm_page_protect(p, VM_PROT_NONE);
1275 				vm_page_free(p);
1276 				p = next;
1277 				continue;
1278 			}
1279 
1280 			/*
1281 			 * Page does not exist in parent, rename the
1282 			 * page from the backing object to the main object.
1283 			 *
1284 			 * If the page was mapped to a process, it can remain
1285 			 * mapped through the rename.
1286 			 */
1287 			if ((p->queue - p->pc) == PQ_CACHE)
1288 				vm_page_deactivate(p);
1289 
1290 			vm_page_rename(p, object, new_pindex);
1291 			/* page automatically made dirty by rename */
1292 		}
1293 		p = next;
1294 	}
1295 	splx(s);
1296 	return(r);
1297 }
1298 
1299 
1300 /*
1301  * this version of collapse allows the operation to occur earlier and
1302  * when paging_in_progress is true for an object...  This is not a complete
1303  * operation, but should plug 99.9% of the rest of the leaks.
1304  */
1305 static void
1306 vm_object_qcollapse(object)
1307 	vm_object_t object;
1308 {
1309 	vm_object_t backing_object = object->backing_object;
1310 
1311 	if (backing_object->ref_count != 1)
1312 		return;
1313 
1314 	backing_object->ref_count += 2;
1315 
1316 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1317 
1318 	backing_object->ref_count -= 2;
1319 }
1320 
1321 /*
1322  *	vm_object_collapse:
1323  *
1324  *	Collapse an object with the object backing it.
1325  *	Pages in the backing object are moved into the
1326  *	parent, and the backing object is deallocated.
1327  */
1328 void
1329 vm_object_collapse(object)
1330 	vm_object_t object;
1331 {
1332 	while (TRUE) {
1333 		vm_object_t backing_object;
1334 
1335 		/*
1336 		 * Verify that the conditions are right for collapse:
1337 		 *
1338 		 * The object exists and the backing object exists.
1339 		 */
1340 		if (object == NULL)
1341 			break;
1342 
1343 		if ((backing_object = object->backing_object) == NULL)
1344 			break;
1345 
1346 		/*
1347 		 * we check the backing object first, because it is most likely
1348 		 * not collapsable.
1349 		 */
1350 		if (backing_object->handle != NULL ||
1351 		    (backing_object->type != OBJT_DEFAULT &&
1352 		     backing_object->type != OBJT_SWAP) ||
1353 		    (backing_object->flags & OBJ_DEAD) ||
1354 		    object->handle != NULL ||
1355 		    (object->type != OBJT_DEFAULT &&
1356 		     object->type != OBJT_SWAP) ||
1357 		    (object->flags & OBJ_DEAD)) {
1358 			break;
1359 		}
1360 
1361 		if (
1362 		    object->paging_in_progress != 0 ||
1363 		    backing_object->paging_in_progress != 0
1364 		) {
1365 			vm_object_qcollapse(object);
1366 			break;
1367 		}
1368 
1369 		/*
1370 		 * We know that we can either collapse the backing object (if
1371 		 * the parent is the only reference to it) or (perhaps) have
1372 		 * the parent bypass the object if the parent happens to shadow
1373 		 * all the resident pages in the entire backing object.
1374 		 *
1375 		 * This is ignoring pager-backed pages such as swap pages.
1376 		 * vm_object_backing_scan fails the shadowing test in this
1377 		 * case.
1378 		 */
1379 
1380 		if (backing_object->ref_count == 1) {
1381 			/*
1382 			 * If there is exactly one reference to the backing
1383 			 * object, we can collapse it into the parent.
1384 			 */
1385 
1386 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1387 
1388 			/*
1389 			 * Move the pager from backing_object to object.
1390 			 */
1391 
1392 			if (backing_object->type == OBJT_SWAP) {
1393 				vm_object_pip_add(backing_object, 1);
1394 
1395 				/*
1396 				 * scrap the paging_offset junk and do a
1397 				 * discrete copy.  This also removes major
1398 				 * assumptions about how the swap-pager
1399 				 * works from where it doesn't belong.  The
1400 				 * new swapper is able to optimize the
1401 				 * destroy-source case.
1402 				 */
1403 
1404 				vm_object_pip_add(object, 1);
1405 				swap_pager_copy(
1406 				    backing_object,
1407 				    object,
1408 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1409 				vm_object_pip_wakeup(object);
1410 
1411 				vm_object_pip_wakeup(backing_object);
1412 			}
1413 			/*
1414 			 * Object now shadows whatever backing_object did.
1415 			 * Note that the reference to
1416 			 * backing_object->backing_object moves from within
1417 			 * backing_object to within object.
1418 			 */
1419 
1420 			LIST_REMOVE(object, shadow_list);
1421 			object->backing_object->shadow_count--;
1422 			object->backing_object->generation++;
1423 			if (backing_object->backing_object) {
1424 				LIST_REMOVE(backing_object, shadow_list);
1425 				backing_object->backing_object->shadow_count--;
1426 				backing_object->backing_object->generation++;
1427 			}
1428 			object->backing_object = backing_object->backing_object;
1429 			if (object->backing_object) {
1430 				LIST_INSERT_HEAD(
1431 				    &object->backing_object->shadow_head,
1432 				    object,
1433 				    shadow_list
1434 				);
1435 				object->backing_object->shadow_count++;
1436 				object->backing_object->generation++;
1437 			}
1438 
1439 			object->backing_object_offset +=
1440 			    backing_object->backing_object_offset;
1441 
1442 			/*
1443 			 * Discard backing_object.
1444 			 *
1445 			 * Since the backing object has no pages, no pager left,
1446 			 * and no object references within it, all that is
1447 			 * necessary is to dispose of it.
1448 			 */
1449 
1450 			KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1451 			KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object));
1452 			TAILQ_REMOVE(
1453 			    &vm_object_list,
1454 			    backing_object,
1455 			    object_list
1456 			);
1457 			vm_object_count--;
1458 
1459 			zfree(obj_zone, backing_object);
1460 
1461 			object_collapses++;
1462 		} else {
1463 			vm_object_t new_backing_object;
1464 
1465 			/*
1466 			 * If we do not entirely shadow the backing object,
1467 			 * there is nothing we can do so we give up.
1468 			 */
1469 
1470 			if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1471 				break;
1472 			}
1473 
1474 			/*
1475 			 * Make the parent shadow the next object in the
1476 			 * chain.  Deallocating backing_object will not remove
1477 			 * it, since its reference count is at least 2.
1478 			 */
1479 
1480 			LIST_REMOVE(object, shadow_list);
1481 			backing_object->shadow_count--;
1482 			backing_object->generation++;
1483 
1484 			new_backing_object = backing_object->backing_object;
1485 			if ((object->backing_object = new_backing_object) != NULL) {
1486 				vm_object_reference(new_backing_object);
1487 				LIST_INSERT_HEAD(
1488 				    &new_backing_object->shadow_head,
1489 				    object,
1490 				    shadow_list
1491 				);
1492 				new_backing_object->shadow_count++;
1493 				new_backing_object->generation++;
1494 				object->backing_object_offset +=
1495 					backing_object->backing_object_offset;
1496 			}
1497 
1498 			/*
1499 			 * Drop the reference count on backing_object. Since
1500 			 * its ref_count was at least 2, it will not vanish;
1501 			 * so we don't need to call vm_object_deallocate, but
1502 			 * we do anyway.
1503 			 */
1504 			vm_object_deallocate(backing_object);
1505 			object_bypasses++;
1506 		}
1507 
1508 		/*
1509 		 * Try again with this object's new backing object.
1510 		 */
1511 	}
1512 }
1513 
1514 /*
1515  *	vm_object_page_remove: [internal]
1516  *
1517  *	Removes all physical pages in the specified
1518  *	object range from the object's list of pages.
1519  *
1520  *	The object must be locked.
1521  */
1522 void
1523 vm_object_page_remove(object, start, end, clean_only)
1524 	vm_object_t object;
1525 	vm_pindex_t start;
1526 	vm_pindex_t end;
1527 	boolean_t clean_only;
1528 {
1529 	vm_page_t p, next;
1530 	unsigned int size;
1531 	int all;
1532 
1533 	if (object == NULL ||
1534 	    object->resident_page_count == 0)
1535 		return;
1536 
1537 	all = ((end == 0) && (start == 0));
1538 
1539 	/*
1540 	 * Since physically-backed objects do not use managed pages, we can't
1541 	 * remove pages from the object (we must instead remove the page
1542 	 * references, and then destroy the object).
1543 	 */
1544 	KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object"));
1545 
1546 	vm_object_pip_add(object, 1);
1547 again:
1548 	size = end - start;
1549 	if (all || size > object->resident_page_count / 4) {
1550 		for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
1551 			next = TAILQ_NEXT(p, listq);
1552 			if (all || ((start <= p->pindex) && (p->pindex < end))) {
1553 				if (p->wire_count != 0) {
1554 					vm_page_protect(p, VM_PROT_NONE);
1555 					if (!clean_only)
1556 						p->valid = 0;
1557 					continue;
1558 				}
1559 
1560 				/*
1561 				 * The busy flags are only cleared at
1562 				 * interrupt -- minimize the spl transitions
1563 				 */
1564 
1565  				if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1566  					goto again;
1567 
1568 				if (clean_only && p->valid) {
1569 					vm_page_test_dirty(p);
1570 					if (p->valid & p->dirty)
1571 						continue;
1572 				}
1573 
1574 				vm_page_busy(p);
1575 				vm_page_protect(p, VM_PROT_NONE);
1576 				vm_page_free(p);
1577 			}
1578 		}
1579 	} else {
1580 		while (size > 0) {
1581 			if ((p = vm_page_lookup(object, start)) != 0) {
1582 
1583 				if (p->wire_count != 0) {
1584 					vm_page_protect(p, VM_PROT_NONE);
1585 					if (!clean_only)
1586 						p->valid = 0;
1587 					start += 1;
1588 					size -= 1;
1589 					continue;
1590 				}
1591 
1592 				/*
1593 				 * The busy flags are only cleared at
1594 				 * interrupt -- minimize the spl transitions
1595 				 */
1596  				if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1597 					goto again;
1598 
1599 				if (clean_only && p->valid) {
1600 					vm_page_test_dirty(p);
1601 					if (p->valid & p->dirty) {
1602 						start += 1;
1603 						size -= 1;
1604 						continue;
1605 					}
1606 				}
1607 
1608 				vm_page_busy(p);
1609 				vm_page_protect(p, VM_PROT_NONE);
1610 				vm_page_free(p);
1611 			}
1612 			start += 1;
1613 			size -= 1;
1614 		}
1615 	}
1616 	vm_object_pip_wakeup(object);
1617 }
1618 
1619 /*
1620  *	Routine:	vm_object_coalesce
1621  *	Function:	Coalesces two objects backing up adjoining
1622  *			regions of memory into a single object.
1623  *
1624  *	returns TRUE if objects were combined.
1625  *
1626  *	NOTE:	Only works at the moment if the second object is NULL -
1627  *		if it's not, which object do we lock first?
1628  *
1629  *	Parameters:
1630  *		prev_object	First object to coalesce
1631  *		prev_offset	Offset into prev_object
1632  *		next_object	Second object into coalesce
1633  *		next_offset	Offset into next_object
1634  *
1635  *		prev_size	Size of reference to prev_object
1636  *		next_size	Size of reference to next_object
1637  *
1638  *	Conditions:
1639  *	The object must *not* be locked.
1640  */
1641 boolean_t
1642 vm_object_coalesce(prev_object, prev_pindex, prev_size, next_size)
1643 	vm_object_t prev_object;
1644 	vm_pindex_t prev_pindex;
1645 	vm_size_t prev_size, next_size;
1646 {
1647 	vm_pindex_t next_pindex;
1648 
1649 	if (prev_object == NULL) {
1650 		return (TRUE);
1651 	}
1652 
1653 	if (prev_object->type != OBJT_DEFAULT &&
1654 	    prev_object->type != OBJT_SWAP) {
1655 		return (FALSE);
1656 	}
1657 
1658 	/*
1659 	 * Try to collapse the object first
1660 	 */
1661 	vm_object_collapse(prev_object);
1662 
1663 	/*
1664 	 * Can't coalesce if: . more than one reference . paged out . shadows
1665 	 * another object . has a copy elsewhere (any of which mean that the
1666 	 * pages not mapped to prev_entry may be in use anyway)
1667 	 */
1668 
1669 	if (prev_object->backing_object != NULL) {
1670 		return (FALSE);
1671 	}
1672 
1673 	prev_size >>= PAGE_SHIFT;
1674 	next_size >>= PAGE_SHIFT;
1675 	next_pindex = prev_pindex + prev_size;
1676 
1677 	if ((prev_object->ref_count > 1) &&
1678 	    (prev_object->size != next_pindex)) {
1679 		return (FALSE);
1680 	}
1681 
1682 	/*
1683 	 * Remove any pages that may still be in the object from a previous
1684 	 * deallocation.
1685 	 */
1686 	if (next_pindex < prev_object->size) {
1687 		vm_object_page_remove(prev_object,
1688 				      next_pindex,
1689 				      next_pindex + next_size, FALSE);
1690 		if (prev_object->type == OBJT_SWAP)
1691 			swap_pager_freespace(prev_object,
1692 					     next_pindex, next_size);
1693 	}
1694 
1695 	/*
1696 	 * Extend the object if necessary.
1697 	 */
1698 	if (next_pindex + next_size > prev_object->size)
1699 		prev_object->size = next_pindex + next_size;
1700 
1701 	return (TRUE);
1702 }
1703 
1704 void
1705 vm_object_set_writeable_dirty(vm_object_t object)
1706 {
1707 	struct vnode *vp;
1708 
1709 	vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1710 	if (object->type == OBJT_VNODE &&
1711 	    (vp = (struct vnode *)object->handle) != NULL) {
1712 		if ((vp->v_flag & VOBJDIRTY) == 0) {
1713 			simple_lock(&vp->v_interlock);
1714 			vp->v_flag |= VOBJDIRTY;
1715 			simple_unlock(&vp->v_interlock);
1716 		}
1717 	}
1718 }
1719 
1720 
1721 
1722 #include "opt_ddb.h"
1723 #ifdef DDB
1724 #include <sys/kernel.h>
1725 
1726 #include <sys/cons.h>
1727 
1728 #include <ddb/ddb.h>
1729 
1730 static int	_vm_object_in_map __P((vm_map_t map, vm_object_t object,
1731 				       vm_map_entry_t entry));
1732 static int	vm_object_in_map __P((vm_object_t object));
1733 
1734 static int
1735 _vm_object_in_map(map, object, entry)
1736 	vm_map_t map;
1737 	vm_object_t object;
1738 	vm_map_entry_t entry;
1739 {
1740 	vm_map_t tmpm;
1741 	vm_map_entry_t tmpe;
1742 	vm_object_t obj;
1743 	int entcount;
1744 
1745 	if (map == 0)
1746 		return 0;
1747 
1748 	if (entry == 0) {
1749 		tmpe = map->header.next;
1750 		entcount = map->nentries;
1751 		while (entcount-- && (tmpe != &map->header)) {
1752 			if( _vm_object_in_map(map, object, tmpe)) {
1753 				return 1;
1754 			}
1755 			tmpe = tmpe->next;
1756 		}
1757 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
1758 		tmpm = entry->object.sub_map;
1759 		tmpe = tmpm->header.next;
1760 		entcount = tmpm->nentries;
1761 		while (entcount-- && tmpe != &tmpm->header) {
1762 			if( _vm_object_in_map(tmpm, object, tmpe)) {
1763 				return 1;
1764 			}
1765 			tmpe = tmpe->next;
1766 		}
1767 	} else if ((obj = entry->object.vm_object) != NULL) {
1768 		for(; obj; obj=obj->backing_object)
1769 			if( obj == object) {
1770 				return 1;
1771 			}
1772 	}
1773 	return 0;
1774 }
1775 
1776 static int
1777 vm_object_in_map( object)
1778 	vm_object_t object;
1779 {
1780 	struct proc *p;
1781 	for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1782 		if( !p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1783 			continue;
1784 		if( _vm_object_in_map(&p->p_vmspace->vm_map, object, 0))
1785 			return 1;
1786 	}
1787 	if( _vm_object_in_map( kernel_map, object, 0))
1788 		return 1;
1789 	if( _vm_object_in_map( kmem_map, object, 0))
1790 		return 1;
1791 	if( _vm_object_in_map( pager_map, object, 0))
1792 		return 1;
1793 	if( _vm_object_in_map( buffer_map, object, 0))
1794 		return 1;
1795 	if( _vm_object_in_map( mb_map, object, 0))
1796 		return 1;
1797 	return 0;
1798 }
1799 
1800 DB_SHOW_COMMAND(vmochk, vm_object_check)
1801 {
1802 	vm_object_t object;
1803 
1804 	/*
1805 	 * make sure that internal objs are in a map somewhere
1806 	 * and none have zero ref counts.
1807 	 */
1808 	for (object = TAILQ_FIRST(&vm_object_list);
1809 			object != NULL;
1810 			object = TAILQ_NEXT(object, object_list)) {
1811 		if (object->handle == NULL &&
1812 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1813 			if (object->ref_count == 0) {
1814 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
1815 					(long)object->size);
1816 			}
1817 			if (!vm_object_in_map(object)) {
1818 				db_printf(
1819 			"vmochk: internal obj is not in a map: "
1820 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1821 				    object->ref_count, (u_long)object->size,
1822 				    (u_long)object->size,
1823 				    (void *)object->backing_object);
1824 			}
1825 		}
1826 	}
1827 }
1828 
1829 /*
1830  *	vm_object_print:	[ debug ]
1831  */
1832 DB_SHOW_COMMAND(object, vm_object_print_static)
1833 {
1834 	/* XXX convert args. */
1835 	vm_object_t object = (vm_object_t)addr;
1836 	boolean_t full = have_addr;
1837 
1838 	vm_page_t p;
1839 
1840 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
1841 #define	count	was_count
1842 
1843 	int count;
1844 
1845 	if (object == NULL)
1846 		return;
1847 
1848 	db_iprintf(
1849 	    "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1850 	    object, (int)object->type, (u_long)object->size,
1851 	    object->resident_page_count, object->ref_count, object->flags);
1852 	/*
1853 	 * XXX no %qd in kernel.  Truncate object->backing_object_offset.
1854 	 */
1855 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1856 	    object->shadow_count,
1857 	    object->backing_object ? object->backing_object->ref_count : 0,
1858 	    object->backing_object, (long)object->backing_object_offset);
1859 
1860 	if (!full)
1861 		return;
1862 
1863 	db_indent += 2;
1864 	count = 0;
1865 	for (p = TAILQ_FIRST(&object->memq); p != NULL; p = TAILQ_NEXT(p, listq)) {
1866 		if (count == 0)
1867 			db_iprintf("memory:=");
1868 		else if (count == 6) {
1869 			db_printf("\n");
1870 			db_iprintf(" ...");
1871 			count = 0;
1872 		} else
1873 			db_printf(",");
1874 		count++;
1875 
1876 		db_printf("(off=0x%lx,page=0x%lx)",
1877 		    (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1878 	}
1879 	if (count != 0)
1880 		db_printf("\n");
1881 	db_indent -= 2;
1882 }
1883 
1884 /* XXX. */
1885 #undef count
1886 
1887 /* XXX need this non-static entry for calling from vm_map_print. */
1888 void
1889 vm_object_print(addr, have_addr, count, modif)
1890         /* db_expr_t */ long addr;
1891 	boolean_t have_addr;
1892 	/* db_expr_t */ long count;
1893 	char *modif;
1894 {
1895 	vm_object_print_static(addr, have_addr, count, modif);
1896 }
1897 
1898 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1899 {
1900 	vm_object_t object;
1901 	int nl = 0;
1902 	int c;
1903 	for (object = TAILQ_FIRST(&vm_object_list);
1904 			object != NULL;
1905 			object = TAILQ_NEXT(object, object_list)) {
1906 		vm_pindex_t idx, fidx;
1907 		vm_pindex_t osize;
1908 		vm_offset_t pa = -1, padiff;
1909 		int rcount;
1910 		vm_page_t m;
1911 
1912 		db_printf("new object: %p\n", (void *)object);
1913 		if ( nl > 18) {
1914 			c = cngetc();
1915 			if (c != ' ')
1916 				return;
1917 			nl = 0;
1918 		}
1919 		nl++;
1920 		rcount = 0;
1921 		fidx = 0;
1922 		osize = object->size;
1923 		if (osize > 128)
1924 			osize = 128;
1925 		for(idx=0;idx<osize;idx++) {
1926 			m = vm_page_lookup(object, idx);
1927 			if (m == NULL) {
1928 				if (rcount) {
1929 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1930 						(long)fidx, rcount, (long)pa);
1931 					if ( nl > 18) {
1932 						c = cngetc();
1933 						if (c != ' ')
1934 							return;
1935 						nl = 0;
1936 					}
1937 					nl++;
1938 					rcount = 0;
1939 				}
1940 				continue;
1941 			}
1942 
1943 
1944 			if (rcount &&
1945 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1946 				++rcount;
1947 				continue;
1948 			}
1949 			if (rcount) {
1950 				padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1951 				padiff >>= PAGE_SHIFT;
1952 				padiff &= PQ_L2_MASK;
1953 				if (padiff == 0) {
1954 					pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1955 					++rcount;
1956 					continue;
1957 				}
1958 				db_printf(" index(%ld)run(%d)pa(0x%lx)",
1959 					(long)fidx, rcount, (long)pa);
1960 				db_printf("pd(%ld)\n", (long)padiff);
1961 				if ( nl > 18) {
1962 					c = cngetc();
1963 					if (c != ' ')
1964 						return;
1965 					nl = 0;
1966 				}
1967 				nl++;
1968 			}
1969 			fidx = idx;
1970 			pa = VM_PAGE_TO_PHYS(m);
1971 			rcount = 1;
1972 		}
1973 		if (rcount) {
1974 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1975 				(long)fidx, rcount, (long)pa);
1976 			if ( nl > 18) {
1977 				c = cngetc();
1978 				if (c != ' ')
1979 					return;
1980 				nl = 0;
1981 			}
1982 			nl++;
1983 		}
1984 	}
1985 }
1986 #endif /* DDB */
1987