xref: /freebsd/sys/vm/vm_object.c (revision 5861f9665471e98e544f6fa3ce73c4912229ff82)
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  * 4. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	from: @(#)vm_object.c	8.5 (Berkeley) 3/22/94
33  *
34  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39  *
40  * Permission to use, copy, modify and distribute this software and
41  * its documentation is hereby granted, provided that both the copyright
42  * notice and this permission notice appear in all copies of the
43  * software, derivative works or modified versions, and any portions
44  * thereof, and that both notices appear in supporting documentation.
45  *
46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49  *
50  * Carnegie Mellon requests users of this software to return to
51  *
52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
53  *  School of Computer Science
54  *  Carnegie Mellon University
55  *  Pittsburgh PA 15213-3890
56  *
57  * any improvements or extensions that they make and grant Carnegie the
58  * rights to redistribute these changes.
59  */
60 
61 /*
62  *	Virtual memory object module.
63  */
64 
65 #include <sys/cdefs.h>
66 __FBSDID("$FreeBSD$");
67 
68 #include "opt_vm.h"
69 
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/lock.h>
73 #include <sys/mman.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/mutex.h>
78 #include <sys/proc.h>		/* for curproc, pageproc */
79 #include <sys/socket.h>
80 #include <sys/resourcevar.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/swap_pager.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_reserv.h>
97 #include <vm/uma.h>
98 
99 #define EASY_SCAN_FACTOR       8
100 
101 #define MSYNC_FLUSH_HARDSEQ	0x01
102 #define MSYNC_FLUSH_SOFTSEQ	0x02
103 
104 /*
105  * msync / VM object flushing optimizations
106  */
107 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ;
108 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags, CTLFLAG_RW, &msync_flush_flags, 0,
109     "Enable sequential iteration optimization");
110 
111 static int old_msync;
112 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
113     "Use old (insecure) msync behavior");
114 
115 static void	vm_object_qcollapse(vm_object_t object);
116 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
117 static void	vm_object_vndeallocate(vm_object_t object);
118 
119 /*
120  *	Virtual memory objects maintain the actual data
121  *	associated with allocated virtual memory.  A given
122  *	page of memory exists within exactly one object.
123  *
124  *	An object is only deallocated when all "references"
125  *	are given up.  Only one "reference" to a given
126  *	region of an object should be writeable.
127  *
128  *	Associated with each object is a list of all resident
129  *	memory pages belonging to that object; this list is
130  *	maintained by the "vm_page" module, and locked by the object's
131  *	lock.
132  *
133  *	Each object also records a "pager" routine which is
134  *	used to retrieve (and store) pages to the proper backing
135  *	storage.  In addition, objects may be backed by other
136  *	objects from which they were virtual-copied.
137  *
138  *	The only items within the object structure which are
139  *	modified after time of creation are:
140  *		reference count		locked by object's lock
141  *		pager routine		locked by object's lock
142  *
143  */
144 
145 struct object_q vm_object_list;
146 struct mtx vm_object_list_mtx;	/* lock for object list and count */
147 
148 struct vm_object kernel_object_store;
149 struct vm_object kmem_object_store;
150 
151 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats");
152 
153 static long object_collapses;
154 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
155     &object_collapses, 0, "VM object collapses");
156 
157 static long object_bypasses;
158 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
159     &object_bypasses, 0, "VM object bypasses");
160 
161 static uma_zone_t obj_zone;
162 
163 static int vm_object_zinit(void *mem, int size, int flags);
164 
165 #ifdef INVARIANTS
166 static void vm_object_zdtor(void *mem, int size, void *arg);
167 
168 static void
169 vm_object_zdtor(void *mem, int size, void *arg)
170 {
171 	vm_object_t object;
172 
173 	object = (vm_object_t)mem;
174 	KASSERT(TAILQ_EMPTY(&object->memq),
175 	    ("object %p has resident pages",
176 	    object));
177 #if VM_NRESERVLEVEL > 0
178 	KASSERT(LIST_EMPTY(&object->rvq),
179 	    ("object %p has reservations",
180 	    object));
181 #endif
182 	KASSERT(object->cache == NULL,
183 	    ("object %p has cached pages",
184 	    object));
185 	KASSERT(object->paging_in_progress == 0,
186 	    ("object %p paging_in_progress = %d",
187 	    object, object->paging_in_progress));
188 	KASSERT(object->resident_page_count == 0,
189 	    ("object %p resident_page_count = %d",
190 	    object, object->resident_page_count));
191 	KASSERT(object->shadow_count == 0,
192 	    ("object %p shadow_count = %d",
193 	    object, object->shadow_count));
194 }
195 #endif
196 
197 static int
198 vm_object_zinit(void *mem, int size, int flags)
199 {
200 	vm_object_t object;
201 
202 	object = (vm_object_t)mem;
203 	bzero(&object->mtx, sizeof(object->mtx));
204 	VM_OBJECT_LOCK_INIT(object, "standard object");
205 
206 	/* These are true for any object that has been freed */
207 	object->paging_in_progress = 0;
208 	object->resident_page_count = 0;
209 	object->shadow_count = 0;
210 	return (0);
211 }
212 
213 void
214 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
215 {
216 
217 	TAILQ_INIT(&object->memq);
218 	LIST_INIT(&object->shadow_head);
219 
220 	object->root = NULL;
221 	object->type = type;
222 	object->size = size;
223 	object->generation = 1;
224 	object->ref_count = 1;
225 	object->flags = 0;
226 	object->uip = NULL;
227 	object->charge = 0;
228 	if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
229 		object->flags = OBJ_ONEMAPPING;
230 	object->pg_color = 0;
231 	object->handle = NULL;
232 	object->backing_object = NULL;
233 	object->backing_object_offset = (vm_ooffset_t) 0;
234 #if VM_NRESERVLEVEL > 0
235 	LIST_INIT(&object->rvq);
236 #endif
237 	object->cache = NULL;
238 
239 	mtx_lock(&vm_object_list_mtx);
240 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
241 	mtx_unlock(&vm_object_list_mtx);
242 }
243 
244 /*
245  *	vm_object_init:
246  *
247  *	Initialize the VM objects module.
248  */
249 void
250 vm_object_init(void)
251 {
252 	TAILQ_INIT(&vm_object_list);
253 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
254 
255 	VM_OBJECT_LOCK_INIT(&kernel_object_store, "kernel object");
256 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
257 	    kernel_object);
258 #if VM_NRESERVLEVEL > 0
259 	kernel_object->flags |= OBJ_COLORED;
260 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
261 #endif
262 
263 	VM_OBJECT_LOCK_INIT(&kmem_object_store, "kmem object");
264 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
265 	    kmem_object);
266 #if VM_NRESERVLEVEL > 0
267 	kmem_object->flags |= OBJ_COLORED;
268 	kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
269 #endif
270 
271 	/*
272 	 * The lock portion of struct vm_object must be type stable due
273 	 * to vm_pageout_fallback_object_lock locking a vm object
274 	 * without holding any references to it.
275 	 */
276 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
277 #ifdef INVARIANTS
278 	    vm_object_zdtor,
279 #else
280 	    NULL,
281 #endif
282 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
283 }
284 
285 void
286 vm_object_clear_flag(vm_object_t object, u_short bits)
287 {
288 
289 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
290 	object->flags &= ~bits;
291 }
292 
293 void
294 vm_object_pip_add(vm_object_t object, short i)
295 {
296 
297 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
298 	object->paging_in_progress += i;
299 }
300 
301 void
302 vm_object_pip_subtract(vm_object_t object, short i)
303 {
304 
305 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
306 	object->paging_in_progress -= i;
307 }
308 
309 void
310 vm_object_pip_wakeup(vm_object_t object)
311 {
312 
313 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
314 	object->paging_in_progress--;
315 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
316 		vm_object_clear_flag(object, OBJ_PIPWNT);
317 		wakeup(object);
318 	}
319 }
320 
321 void
322 vm_object_pip_wakeupn(vm_object_t object, short i)
323 {
324 
325 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
326 	if (i)
327 		object->paging_in_progress -= i;
328 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
329 		vm_object_clear_flag(object, OBJ_PIPWNT);
330 		wakeup(object);
331 	}
332 }
333 
334 void
335 vm_object_pip_wait(vm_object_t object, char *waitid)
336 {
337 
338 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
339 	while (object->paging_in_progress) {
340 		object->flags |= OBJ_PIPWNT;
341 		msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
342 	}
343 }
344 
345 /*
346  *	vm_object_allocate:
347  *
348  *	Returns a new object with the given size.
349  */
350 vm_object_t
351 vm_object_allocate(objtype_t type, vm_pindex_t size)
352 {
353 	vm_object_t object;
354 
355 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
356 	_vm_object_allocate(type, size, object);
357 	return (object);
358 }
359 
360 
361 /*
362  *	vm_object_reference:
363  *
364  *	Gets another reference to the given object.  Note: OBJ_DEAD
365  *	objects can be referenced during final cleaning.
366  */
367 void
368 vm_object_reference(vm_object_t object)
369 {
370 	if (object == NULL)
371 		return;
372 	VM_OBJECT_LOCK(object);
373 	vm_object_reference_locked(object);
374 	VM_OBJECT_UNLOCK(object);
375 }
376 
377 /*
378  *	vm_object_reference_locked:
379  *
380  *	Gets another reference to the given object.
381  *
382  *	The object must be locked.
383  */
384 void
385 vm_object_reference_locked(vm_object_t object)
386 {
387 	struct vnode *vp;
388 
389 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
390 	object->ref_count++;
391 	if (object->type == OBJT_VNODE) {
392 		vp = object->handle;
393 		vref(vp);
394 	}
395 }
396 
397 /*
398  * Handle deallocating an object of type OBJT_VNODE.
399  */
400 static void
401 vm_object_vndeallocate(vm_object_t object)
402 {
403 	struct vnode *vp = (struct vnode *) object->handle;
404 
405 	VFS_ASSERT_GIANT(vp->v_mount);
406 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
407 	KASSERT(object->type == OBJT_VNODE,
408 	    ("vm_object_vndeallocate: not a vnode object"));
409 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
410 #ifdef INVARIANTS
411 	if (object->ref_count == 0) {
412 		vprint("vm_object_vndeallocate", vp);
413 		panic("vm_object_vndeallocate: bad object reference count");
414 	}
415 #endif
416 
417 	object->ref_count--;
418 	if (object->ref_count == 0) {
419 		mp_fixme("Unlocked vflag access.");
420 		vp->v_vflag &= ~VV_TEXT;
421 	}
422 	VM_OBJECT_UNLOCK(object);
423 	/*
424 	 * vrele may need a vop lock
425 	 */
426 	vrele(vp);
427 }
428 
429 /*
430  *	vm_object_deallocate:
431  *
432  *	Release a reference to the specified object,
433  *	gained either through a vm_object_allocate
434  *	or a vm_object_reference call.  When all references
435  *	are gone, storage associated with this object
436  *	may be relinquished.
437  *
438  *	No object may be locked.
439  */
440 void
441 vm_object_deallocate(vm_object_t object)
442 {
443 	vm_object_t temp;
444 
445 	while (object != NULL) {
446 		int vfslocked;
447 
448 		vfslocked = 0;
449 	restart:
450 		VM_OBJECT_LOCK(object);
451 		if (object->type == OBJT_VNODE) {
452 			struct vnode *vp = (struct vnode *) object->handle;
453 
454 			/*
455 			 * Conditionally acquire Giant for a vnode-backed
456 			 * object.  We have to be careful since the type of
457 			 * a vnode object can change while the object is
458 			 * unlocked.
459 			 */
460 			if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
461 				vfslocked = 1;
462 				if (!mtx_trylock(&Giant)) {
463 					VM_OBJECT_UNLOCK(object);
464 					mtx_lock(&Giant);
465 					goto restart;
466 				}
467 			}
468 			vm_object_vndeallocate(object);
469 			VFS_UNLOCK_GIANT(vfslocked);
470 			return;
471 		} else
472 			/*
473 			 * This is to handle the case that the object
474 			 * changed type while we dropped its lock to
475 			 * obtain Giant.
476 			 */
477 			VFS_UNLOCK_GIANT(vfslocked);
478 
479 		KASSERT(object->ref_count != 0,
480 			("vm_object_deallocate: object deallocated too many times: %d", object->type));
481 
482 		/*
483 		 * If the reference count goes to 0 we start calling
484 		 * vm_object_terminate() on the object chain.
485 		 * A ref count of 1 may be a special case depending on the
486 		 * shadow count being 0 or 1.
487 		 */
488 		object->ref_count--;
489 		if (object->ref_count > 1) {
490 			VM_OBJECT_UNLOCK(object);
491 			return;
492 		} else if (object->ref_count == 1) {
493 			if (object->shadow_count == 0 &&
494 			    object->handle == NULL &&
495 			    (object->type == OBJT_DEFAULT ||
496 			     object->type == OBJT_SWAP)) {
497 				vm_object_set_flag(object, OBJ_ONEMAPPING);
498 			} else if ((object->shadow_count == 1) &&
499 			    (object->handle == NULL) &&
500 			    (object->type == OBJT_DEFAULT ||
501 			     object->type == OBJT_SWAP)) {
502 				vm_object_t robject;
503 
504 				robject = LIST_FIRST(&object->shadow_head);
505 				KASSERT(robject != NULL,
506 				    ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
507 					 object->ref_count,
508 					 object->shadow_count));
509 				if (!VM_OBJECT_TRYLOCK(robject)) {
510 					/*
511 					 * Avoid a potential deadlock.
512 					 */
513 					object->ref_count++;
514 					VM_OBJECT_UNLOCK(object);
515 					/*
516 					 * More likely than not the thread
517 					 * holding robject's lock has lower
518 					 * priority than the current thread.
519 					 * Let the lower priority thread run.
520 					 */
521 					pause("vmo_de", 1);
522 					continue;
523 				}
524 				/*
525 				 * Collapse object into its shadow unless its
526 				 * shadow is dead.  In that case, object will
527 				 * be deallocated by the thread that is
528 				 * deallocating its shadow.
529 				 */
530 				if ((robject->flags & OBJ_DEAD) == 0 &&
531 				    (robject->handle == NULL) &&
532 				    (robject->type == OBJT_DEFAULT ||
533 				     robject->type == OBJT_SWAP)) {
534 
535 					robject->ref_count++;
536 retry:
537 					if (robject->paging_in_progress) {
538 						VM_OBJECT_UNLOCK(object);
539 						vm_object_pip_wait(robject,
540 						    "objde1");
541 						temp = robject->backing_object;
542 						if (object == temp) {
543 							VM_OBJECT_LOCK(object);
544 							goto retry;
545 						}
546 					} else if (object->paging_in_progress) {
547 						VM_OBJECT_UNLOCK(robject);
548 						object->flags |= OBJ_PIPWNT;
549 						msleep(object,
550 						    VM_OBJECT_MTX(object),
551 						    PDROP | PVM, "objde2", 0);
552 						VM_OBJECT_LOCK(robject);
553 						temp = robject->backing_object;
554 						if (object == temp) {
555 							VM_OBJECT_LOCK(object);
556 							goto retry;
557 						}
558 					} else
559 						VM_OBJECT_UNLOCK(object);
560 
561 					if (robject->ref_count == 1) {
562 						robject->ref_count--;
563 						object = robject;
564 						goto doterm;
565 					}
566 					object = robject;
567 					vm_object_collapse(object);
568 					VM_OBJECT_UNLOCK(object);
569 					continue;
570 				}
571 				VM_OBJECT_UNLOCK(robject);
572 			}
573 			VM_OBJECT_UNLOCK(object);
574 			return;
575 		}
576 doterm:
577 		temp = object->backing_object;
578 		if (temp != NULL) {
579 			VM_OBJECT_LOCK(temp);
580 			LIST_REMOVE(object, shadow_list);
581 			temp->shadow_count--;
582 			temp->generation++;
583 			VM_OBJECT_UNLOCK(temp);
584 			object->backing_object = NULL;
585 		}
586 		/*
587 		 * Don't double-terminate, we could be in a termination
588 		 * recursion due to the terminate having to sync data
589 		 * to disk.
590 		 */
591 		if ((object->flags & OBJ_DEAD) == 0)
592 			vm_object_terminate(object);
593 		else
594 			VM_OBJECT_UNLOCK(object);
595 		object = temp;
596 	}
597 }
598 
599 /*
600  *	vm_object_destroy removes the object from the global object list
601  *      and frees the space for the object.
602  */
603 void
604 vm_object_destroy(vm_object_t object)
605 {
606 
607 	/*
608 	 * Remove the object from the global object list.
609 	 */
610 	mtx_lock(&vm_object_list_mtx);
611 	TAILQ_REMOVE(&vm_object_list, object, object_list);
612 	mtx_unlock(&vm_object_list_mtx);
613 
614 	/*
615 	 * Release the allocation charge.
616 	 */
617 	if (object->uip != NULL) {
618 		KASSERT(object->type == OBJT_DEFAULT ||
619 		    object->type == OBJT_SWAP,
620 		    ("vm_object_terminate: non-swap obj %p has uip",
621 		     object));
622 		swap_release_by_uid(object->charge, object->uip);
623 		object->charge = 0;
624 		uifree(object->uip);
625 		object->uip = NULL;
626 	}
627 
628 	/*
629 	 * Free the space for the object.
630 	 */
631 	uma_zfree(obj_zone, object);
632 }
633 
634 /*
635  *	vm_object_terminate actually destroys the specified object, freeing
636  *	up all previously used resources.
637  *
638  *	The object must be locked.
639  *	This routine may block.
640  */
641 void
642 vm_object_terminate(vm_object_t object)
643 {
644 	vm_page_t p;
645 
646 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
647 
648 	/*
649 	 * Make sure no one uses us.
650 	 */
651 	vm_object_set_flag(object, OBJ_DEAD);
652 
653 	/*
654 	 * wait for the pageout daemon to be done with the object
655 	 */
656 	vm_object_pip_wait(object, "objtrm");
657 
658 	KASSERT(!object->paging_in_progress,
659 		("vm_object_terminate: pageout in progress"));
660 
661 	/*
662 	 * Clean and free the pages, as appropriate. All references to the
663 	 * object are gone, so we don't need to lock it.
664 	 */
665 	if (object->type == OBJT_VNODE) {
666 		struct vnode *vp = (struct vnode *)object->handle;
667 
668 		/*
669 		 * Clean pages and flush buffers.
670 		 */
671 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
672 		VM_OBJECT_UNLOCK(object);
673 
674 		vinvalbuf(vp, V_SAVE, 0, 0);
675 
676 		VM_OBJECT_LOCK(object);
677 	}
678 
679 	KASSERT(object->ref_count == 0,
680 		("vm_object_terminate: object with references, ref_count=%d",
681 		object->ref_count));
682 
683 	/*
684 	 * Now free any remaining pages. For internal objects, this also
685 	 * removes them from paging queues. Don't free wired pages, just
686 	 * remove them from the object.
687 	 */
688 	vm_page_lock_queues();
689 	while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
690 		KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
691 			("vm_object_terminate: freeing busy page %p "
692 			"p->busy = %d, p->oflags %x\n", p, p->busy, p->oflags));
693 		if (p->wire_count == 0) {
694 			vm_page_free(p);
695 			cnt.v_pfree++;
696 		} else {
697 			vm_page_remove(p);
698 		}
699 	}
700 	vm_page_unlock_queues();
701 
702 #if VM_NRESERVLEVEL > 0
703 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
704 		vm_reserv_break_all(object);
705 #endif
706 	if (__predict_false(object->cache != NULL))
707 		vm_page_cache_free(object, 0, 0);
708 
709 	/*
710 	 * Let the pager know object is dead.
711 	 */
712 	vm_pager_deallocate(object);
713 	VM_OBJECT_UNLOCK(object);
714 
715 	vm_object_destroy(object);
716 }
717 
718 /*
719  *	vm_object_page_clean
720  *
721  *	Clean all dirty pages in the specified range of object.  Leaves page
722  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
723  *	write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
724  *	leaving the object dirty.
725  *
726  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
727  *	synchronous clustering mode implementation.
728  *
729  *	Odd semantics: if start == end, we clean everything.
730  *
731  *	The object must be locked.
732  */
733 void
734 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags)
735 {
736 	vm_page_t p, np;
737 	vm_pindex_t tstart, tend;
738 	vm_pindex_t pi;
739 	int clearobjflags;
740 	int pagerflags;
741 	int curgeneration;
742 
743 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
744 	if (object->type != OBJT_VNODE ||
745 		(object->flags & OBJ_MIGHTBEDIRTY) == 0)
746 		return;
747 
748 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
749 	pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
750 
751 	vm_object_set_flag(object, OBJ_CLEANING);
752 
753 	tstart = start;
754 	if (end == 0) {
755 		tend = object->size;
756 	} else {
757 		tend = end;
758 	}
759 
760 	vm_page_lock_queues();
761 	/*
762 	 * If the caller is smart and only msync()s a range he knows is
763 	 * dirty, we may be able to avoid an object scan.  This results in
764 	 * a phenominal improvement in performance.  We cannot do this
765 	 * as a matter of course because the object may be huge - e.g.
766 	 * the size might be in the gigabytes or terrabytes.
767 	 */
768 	if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
769 		vm_pindex_t tscan;
770 		int scanlimit;
771 		int scanreset;
772 
773 		scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
774 		if (scanreset < 16)
775 			scanreset = 16;
776 		pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
777 
778 		scanlimit = scanreset;
779 		tscan = tstart;
780 		while (tscan < tend) {
781 			curgeneration = object->generation;
782 			p = vm_page_lookup(object, tscan);
783 			if (p == NULL || p->valid == 0) {
784 				if (--scanlimit == 0)
785 					break;
786 				++tscan;
787 				continue;
788 			}
789 			vm_page_test_dirty(p);
790 			if (p->dirty == 0) {
791 				if (--scanlimit == 0)
792 					break;
793 				++tscan;
794 				continue;
795 			}
796 			/*
797 			 * If we have been asked to skip nosync pages and
798 			 * this is a nosync page, we can't continue.
799 			 */
800 			if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
801 				if (--scanlimit == 0)
802 					break;
803 				++tscan;
804 				continue;
805 			}
806 			scanlimit = scanreset;
807 
808 			/*
809 			 * This returns 0 if it was unable to busy the first
810 			 * page (i.e. had to sleep).
811 			 */
812 			tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
813 		}
814 
815 		/*
816 		 * If everything was dirty and we flushed it successfully,
817 		 * and the requested range is not the entire object, we
818 		 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
819 		 * return immediately.
820 		 */
821 		if (tscan >= tend && (tstart || tend < object->size)) {
822 			vm_page_unlock_queues();
823 			vm_object_clear_flag(object, OBJ_CLEANING);
824 			return;
825 		}
826 		pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
827 	}
828 
829 	/*
830 	 * Generally set CLEANCHK interlock and make the page read-only so
831 	 * we can then clear the object flags.
832 	 *
833 	 * However, if this is a nosync mmap then the object is likely to
834 	 * stay dirty so do not mess with the page and do not clear the
835 	 * object flags.
836 	 */
837 	clearobjflags = 1;
838 	TAILQ_FOREACH(p, &object->memq, listq) {
839 		p->oflags |= VPO_CLEANCHK;
840 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC))
841 			clearobjflags = 0;
842 		else
843 			pmap_remove_write(p);
844 	}
845 
846 	if (clearobjflags && (tstart == 0) && (tend == object->size)) {
847 		struct vnode *vp;
848 
849 		vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
850 		if (object->type == OBJT_VNODE &&
851 		    (vp = (struct vnode *)object->handle) != NULL) {
852 			VI_LOCK(vp);
853 			if (vp->v_iflag & VI_OBJDIRTY)
854 				vp->v_iflag &= ~VI_OBJDIRTY;
855 			VI_UNLOCK(vp);
856 		}
857 	}
858 
859 rescan:
860 	curgeneration = object->generation;
861 
862 	for (p = TAILQ_FIRST(&object->memq); p; p = np) {
863 		int n;
864 
865 		np = TAILQ_NEXT(p, listq);
866 
867 again:
868 		pi = p->pindex;
869 		if ((p->oflags & VPO_CLEANCHK) == 0 ||
870 			(pi < tstart) || (pi >= tend) ||
871 		    p->valid == 0) {
872 			p->oflags &= ~VPO_CLEANCHK;
873 			continue;
874 		}
875 
876 		vm_page_test_dirty(p);
877 		if (p->dirty == 0) {
878 			p->oflags &= ~VPO_CLEANCHK;
879 			continue;
880 		}
881 
882 		/*
883 		 * If we have been asked to skip nosync pages and this is a
884 		 * nosync page, skip it.  Note that the object flags were
885 		 * not cleared in this case so we do not have to set them.
886 		 */
887 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
888 			p->oflags &= ~VPO_CLEANCHK;
889 			continue;
890 		}
891 
892 		n = vm_object_page_collect_flush(object, p,
893 			curgeneration, pagerflags);
894 		if (n == 0)
895 			goto rescan;
896 
897 		if (object->generation != curgeneration)
898 			goto rescan;
899 
900 		/*
901 		 * Try to optimize the next page.  If we can't we pick up
902 		 * our (random) scan where we left off.
903 		 */
904 		if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
905 			if ((p = vm_page_lookup(object, pi + n)) != NULL)
906 				goto again;
907 		}
908 	}
909 	vm_page_unlock_queues();
910 #if 0
911 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
912 #endif
913 
914 	vm_object_clear_flag(object, OBJ_CLEANING);
915 	return;
916 }
917 
918 static int
919 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
920 {
921 	int runlen;
922 	int maxf;
923 	int chkb;
924 	int maxb;
925 	int i;
926 	vm_pindex_t pi;
927 	vm_page_t maf[vm_pageout_page_count];
928 	vm_page_t mab[vm_pageout_page_count];
929 	vm_page_t ma[vm_pageout_page_count];
930 
931 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
932 	pi = p->pindex;
933 	while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
934 		vm_page_lock_queues();
935 		if (object->generation != curgeneration) {
936 			return(0);
937 		}
938 	}
939 	maxf = 0;
940 	for(i = 1; i < vm_pageout_page_count; i++) {
941 		vm_page_t tp;
942 
943 		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
944 			if ((tp->oflags & VPO_BUSY) ||
945 				((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
946 				 (tp->oflags & VPO_CLEANCHK) == 0) ||
947 				(tp->busy != 0))
948 				break;
949 			vm_page_test_dirty(tp);
950 			if (tp->dirty == 0) {
951 				tp->oflags &= ~VPO_CLEANCHK;
952 				break;
953 			}
954 			maf[ i - 1 ] = tp;
955 			maxf++;
956 			continue;
957 		}
958 		break;
959 	}
960 
961 	maxb = 0;
962 	chkb = vm_pageout_page_count -  maxf;
963 	if (chkb) {
964 		for(i = 1; i < chkb;i++) {
965 			vm_page_t tp;
966 
967 			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
968 				if ((tp->oflags & VPO_BUSY) ||
969 					((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
970 					 (tp->oflags & VPO_CLEANCHK) == 0) ||
971 					(tp->busy != 0))
972 					break;
973 				vm_page_test_dirty(tp);
974 				if (tp->dirty == 0) {
975 					tp->oflags &= ~VPO_CLEANCHK;
976 					break;
977 				}
978 				mab[ i - 1 ] = tp;
979 				maxb++;
980 				continue;
981 			}
982 			break;
983 		}
984 	}
985 
986 	for(i = 0; i < maxb; i++) {
987 		int index = (maxb - i) - 1;
988 		ma[index] = mab[i];
989 		ma[index]->oflags &= ~VPO_CLEANCHK;
990 	}
991 	p->oflags &= ~VPO_CLEANCHK;
992 	ma[maxb] = p;
993 	for(i = 0; i < maxf; i++) {
994 		int index = (maxb + i) + 1;
995 		ma[index] = maf[i];
996 		ma[index]->oflags &= ~VPO_CLEANCHK;
997 	}
998 	runlen = maxb + maxf + 1;
999 
1000 	vm_pageout_flush(ma, runlen, pagerflags);
1001 	for (i = 0; i < runlen; i++) {
1002 		if (ma[i]->dirty) {
1003 			pmap_remove_write(ma[i]);
1004 			ma[i]->oflags |= VPO_CLEANCHK;
1005 
1006 			/*
1007 			 * maxf will end up being the actual number of pages
1008 			 * we wrote out contiguously, non-inclusive of the
1009 			 * first page.  We do not count look-behind pages.
1010 			 */
1011 			if (i >= maxb + 1 && (maxf > i - maxb - 1))
1012 				maxf = i - maxb - 1;
1013 		}
1014 	}
1015 	return(maxf + 1);
1016 }
1017 
1018 /*
1019  * Note that there is absolutely no sense in writing out
1020  * anonymous objects, so we track down the vnode object
1021  * to write out.
1022  * We invalidate (remove) all pages from the address space
1023  * for semantic correctness.
1024  *
1025  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1026  * may start out with a NULL object.
1027  */
1028 void
1029 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1030     boolean_t syncio, boolean_t invalidate)
1031 {
1032 	vm_object_t backing_object;
1033 	struct vnode *vp;
1034 	struct mount *mp;
1035 	int flags;
1036 
1037 	if (object == NULL)
1038 		return;
1039 	VM_OBJECT_LOCK(object);
1040 	while ((backing_object = object->backing_object) != NULL) {
1041 		VM_OBJECT_LOCK(backing_object);
1042 		offset += object->backing_object_offset;
1043 		VM_OBJECT_UNLOCK(object);
1044 		object = backing_object;
1045 		if (object->size < OFF_TO_IDX(offset + size))
1046 			size = IDX_TO_OFF(object->size) - offset;
1047 	}
1048 	/*
1049 	 * Flush pages if writing is allowed, invalidate them
1050 	 * if invalidation requested.  Pages undergoing I/O
1051 	 * will be ignored by vm_object_page_remove().
1052 	 *
1053 	 * We cannot lock the vnode and then wait for paging
1054 	 * to complete without deadlocking against vm_fault.
1055 	 * Instead we simply call vm_object_page_remove() and
1056 	 * allow it to block internally on a page-by-page
1057 	 * basis when it encounters pages undergoing async
1058 	 * I/O.
1059 	 */
1060 	if (object->type == OBJT_VNODE &&
1061 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1062 		int vfslocked;
1063 		vp = object->handle;
1064 		VM_OBJECT_UNLOCK(object);
1065 		(void) vn_start_write(vp, &mp, V_WAIT);
1066 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1067 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1068 		flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1069 		flags |= invalidate ? OBJPC_INVAL : 0;
1070 		VM_OBJECT_LOCK(object);
1071 		vm_object_page_clean(object,
1072 		    OFF_TO_IDX(offset),
1073 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1074 		    flags);
1075 		VM_OBJECT_UNLOCK(object);
1076 		VOP_UNLOCK(vp, 0);
1077 		VFS_UNLOCK_GIANT(vfslocked);
1078 		vn_finished_write(mp);
1079 		VM_OBJECT_LOCK(object);
1080 	}
1081 	if ((object->type == OBJT_VNODE ||
1082 	     object->type == OBJT_DEVICE) && invalidate) {
1083 		boolean_t purge;
1084 		purge = old_msync || (object->type == OBJT_DEVICE);
1085 		vm_object_page_remove(object,
1086 		    OFF_TO_IDX(offset),
1087 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1088 		    purge ? FALSE : TRUE);
1089 	}
1090 	VM_OBJECT_UNLOCK(object);
1091 }
1092 
1093 /*
1094  *	vm_object_madvise:
1095  *
1096  *	Implements the madvise function at the object/page level.
1097  *
1098  *	MADV_WILLNEED	(any object)
1099  *
1100  *	    Activate the specified pages if they are resident.
1101  *
1102  *	MADV_DONTNEED	(any object)
1103  *
1104  *	    Deactivate the specified pages if they are resident.
1105  *
1106  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1107  *			 OBJ_ONEMAPPING only)
1108  *
1109  *	    Deactivate and clean the specified pages if they are
1110  *	    resident.  This permits the process to reuse the pages
1111  *	    without faulting or the kernel to reclaim the pages
1112  *	    without I/O.
1113  */
1114 void
1115 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1116 {
1117 	vm_pindex_t end, tpindex;
1118 	vm_object_t backing_object, tobject;
1119 	vm_page_t m;
1120 
1121 	if (object == NULL)
1122 		return;
1123 	VM_OBJECT_LOCK(object);
1124 	end = pindex + count;
1125 	/*
1126 	 * Locate and adjust resident pages
1127 	 */
1128 	for (; pindex < end; pindex += 1) {
1129 relookup:
1130 		tobject = object;
1131 		tpindex = pindex;
1132 shadowlookup:
1133 		/*
1134 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1135 		 * and those pages must be OBJ_ONEMAPPING.
1136 		 */
1137 		if (advise == MADV_FREE) {
1138 			if ((tobject->type != OBJT_DEFAULT &&
1139 			     tobject->type != OBJT_SWAP) ||
1140 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
1141 				goto unlock_tobject;
1142 			}
1143 		}
1144 		m = vm_page_lookup(tobject, tpindex);
1145 		if (m == NULL && advise == MADV_WILLNEED) {
1146 			/*
1147 			 * If the page is cached, reactivate it.
1148 			 */
1149 			m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1150 			    VM_ALLOC_NOBUSY);
1151 		}
1152 		if (m == NULL) {
1153 			/*
1154 			 * There may be swap even if there is no backing page
1155 			 */
1156 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1157 				swap_pager_freespace(tobject, tpindex, 1);
1158 			/*
1159 			 * next object
1160 			 */
1161 			backing_object = tobject->backing_object;
1162 			if (backing_object == NULL)
1163 				goto unlock_tobject;
1164 			VM_OBJECT_LOCK(backing_object);
1165 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1166 			if (tobject != object)
1167 				VM_OBJECT_UNLOCK(tobject);
1168 			tobject = backing_object;
1169 			goto shadowlookup;
1170 		}
1171 		/*
1172 		 * If the page is busy or not in a normal active state,
1173 		 * we skip it.  If the page is not managed there are no
1174 		 * page queues to mess with.  Things can break if we mess
1175 		 * with pages in any of the below states.
1176 		 */
1177 		vm_page_lock_queues();
1178 		if (m->hold_count ||
1179 		    m->wire_count ||
1180 		    (m->flags & PG_UNMANAGED) ||
1181 		    m->valid != VM_PAGE_BITS_ALL) {
1182 			vm_page_unlock_queues();
1183 			goto unlock_tobject;
1184 		}
1185 		if ((m->oflags & VPO_BUSY) || m->busy) {
1186 			vm_page_flag_set(m, PG_REFERENCED);
1187 			vm_page_unlock_queues();
1188 			if (object != tobject)
1189 				VM_OBJECT_UNLOCK(object);
1190 			m->oflags |= VPO_WANTED;
1191 			msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo", 0);
1192 			VM_OBJECT_LOCK(object);
1193   			goto relookup;
1194 		}
1195 		if (advise == MADV_WILLNEED) {
1196 			vm_page_activate(m);
1197 		} else if (advise == MADV_DONTNEED) {
1198 			vm_page_dontneed(m);
1199 		} else if (advise == MADV_FREE) {
1200 			/*
1201 			 * Mark the page clean.  This will allow the page
1202 			 * to be freed up by the system.  However, such pages
1203 			 * are often reused quickly by malloc()/free()
1204 			 * so we do not do anything that would cause
1205 			 * a page fault if we can help it.
1206 			 *
1207 			 * Specifically, we do not try to actually free
1208 			 * the page now nor do we try to put it in the
1209 			 * cache (which would cause a page fault on reuse).
1210 			 *
1211 			 * But we do make the page is freeable as we
1212 			 * can without actually taking the step of unmapping
1213 			 * it.
1214 			 */
1215 			pmap_clear_modify(m);
1216 			m->dirty = 0;
1217 			m->act_count = 0;
1218 			vm_page_dontneed(m);
1219 		}
1220 		vm_page_unlock_queues();
1221 		if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1222 			swap_pager_freespace(tobject, tpindex, 1);
1223 unlock_tobject:
1224 		if (tobject != object)
1225 			VM_OBJECT_UNLOCK(tobject);
1226 	}
1227 	VM_OBJECT_UNLOCK(object);
1228 }
1229 
1230 /*
1231  *	vm_object_shadow:
1232  *
1233  *	Create a new object which is backed by the
1234  *	specified existing object range.  The source
1235  *	object reference is deallocated.
1236  *
1237  *	The new object and offset into that object
1238  *	are returned in the source parameters.
1239  */
1240 void
1241 vm_object_shadow(
1242 	vm_object_t *object,	/* IN/OUT */
1243 	vm_ooffset_t *offset,	/* IN/OUT */
1244 	vm_size_t length)
1245 {
1246 	vm_object_t source;
1247 	vm_object_t result;
1248 
1249 	source = *object;
1250 
1251 	/*
1252 	 * Don't create the new object if the old object isn't shared.
1253 	 */
1254 	if (source != NULL) {
1255 		VM_OBJECT_LOCK(source);
1256 		if (source->ref_count == 1 &&
1257 		    source->handle == NULL &&
1258 		    (source->type == OBJT_DEFAULT ||
1259 		     source->type == OBJT_SWAP)) {
1260 			VM_OBJECT_UNLOCK(source);
1261 			return;
1262 		}
1263 		VM_OBJECT_UNLOCK(source);
1264 	}
1265 
1266 	/*
1267 	 * Allocate a new object with the given length.
1268 	 */
1269 	result = vm_object_allocate(OBJT_DEFAULT, length);
1270 
1271 	/*
1272 	 * The new object shadows the source object, adding a reference to it.
1273 	 * Our caller changes his reference to point to the new object,
1274 	 * removing a reference to the source object.  Net result: no change
1275 	 * of reference count.
1276 	 *
1277 	 * Try to optimize the result object's page color when shadowing
1278 	 * in order to maintain page coloring consistency in the combined
1279 	 * shadowed object.
1280 	 */
1281 	result->backing_object = source;
1282 	/*
1283 	 * Store the offset into the source object, and fix up the offset into
1284 	 * the new object.
1285 	 */
1286 	result->backing_object_offset = *offset;
1287 	if (source != NULL) {
1288 		VM_OBJECT_LOCK(source);
1289 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1290 		source->shadow_count++;
1291 		source->generation++;
1292 #if VM_NRESERVLEVEL > 0
1293 		result->flags |= source->flags & OBJ_COLORED;
1294 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1295 		    ((1 << (VM_NFREEORDER - 1)) - 1);
1296 #endif
1297 		VM_OBJECT_UNLOCK(source);
1298 	}
1299 
1300 
1301 	/*
1302 	 * Return the new things
1303 	 */
1304 	*offset = 0;
1305 	*object = result;
1306 }
1307 
1308 /*
1309  *	vm_object_split:
1310  *
1311  * Split the pages in a map entry into a new object.  This affords
1312  * easier removal of unused pages, and keeps object inheritance from
1313  * being a negative impact on memory usage.
1314  */
1315 void
1316 vm_object_split(vm_map_entry_t entry)
1317 {
1318 	vm_page_t m, m_next;
1319 	vm_object_t orig_object, new_object, source;
1320 	vm_pindex_t idx, offidxstart;
1321 	vm_size_t size;
1322 
1323 	orig_object = entry->object.vm_object;
1324 	if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1325 		return;
1326 	if (orig_object->ref_count <= 1)
1327 		return;
1328 	VM_OBJECT_UNLOCK(orig_object);
1329 
1330 	offidxstart = OFF_TO_IDX(entry->offset);
1331 	size = atop(entry->end - entry->start);
1332 
1333 	/*
1334 	 * If swap_pager_copy() is later called, it will convert new_object
1335 	 * into a swap object.
1336 	 */
1337 	new_object = vm_object_allocate(OBJT_DEFAULT, size);
1338 
1339 	/*
1340 	 * At this point, the new object is still private, so the order in
1341 	 * which the original and new objects are locked does not matter.
1342 	 */
1343 	VM_OBJECT_LOCK(new_object);
1344 	VM_OBJECT_LOCK(orig_object);
1345 	source = orig_object->backing_object;
1346 	if (source != NULL) {
1347 		VM_OBJECT_LOCK(source);
1348 		if ((source->flags & OBJ_DEAD) != 0) {
1349 			VM_OBJECT_UNLOCK(source);
1350 			VM_OBJECT_UNLOCK(orig_object);
1351 			VM_OBJECT_UNLOCK(new_object);
1352 			vm_object_deallocate(new_object);
1353 			VM_OBJECT_LOCK(orig_object);
1354 			return;
1355 		}
1356 		LIST_INSERT_HEAD(&source->shadow_head,
1357 				  new_object, shadow_list);
1358 		source->shadow_count++;
1359 		source->generation++;
1360 		vm_object_reference_locked(source);	/* for new_object */
1361 		vm_object_clear_flag(source, OBJ_ONEMAPPING);
1362 		VM_OBJECT_UNLOCK(source);
1363 		new_object->backing_object_offset =
1364 			orig_object->backing_object_offset + entry->offset;
1365 		new_object->backing_object = source;
1366 	}
1367 	if (orig_object->uip != NULL) {
1368 		new_object->uip = orig_object->uip;
1369 		uihold(orig_object->uip);
1370 		new_object->charge = ptoa(size);
1371 		KASSERT(orig_object->charge >= ptoa(size),
1372 		    ("orig_object->charge < 0"));
1373 		orig_object->charge -= ptoa(size);
1374 	}
1375 retry:
1376 	if ((m = TAILQ_FIRST(&orig_object->memq)) != NULL) {
1377 		if (m->pindex < offidxstart) {
1378 			m = vm_page_splay(offidxstart, orig_object->root);
1379 			if ((orig_object->root = m)->pindex < offidxstart)
1380 				m = TAILQ_NEXT(m, listq);
1381 		}
1382 	}
1383 	vm_page_lock_queues();
1384 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1385 	    m = m_next) {
1386 		m_next = TAILQ_NEXT(m, listq);
1387 
1388 		/*
1389 		 * We must wait for pending I/O to complete before we can
1390 		 * rename the page.
1391 		 *
1392 		 * We do not have to VM_PROT_NONE the page as mappings should
1393 		 * not be changed by this operation.
1394 		 */
1395 		if ((m->oflags & VPO_BUSY) || m->busy) {
1396 			vm_page_flag_set(m, PG_REFERENCED);
1397 			vm_page_unlock_queues();
1398 			VM_OBJECT_UNLOCK(new_object);
1399 			m->oflags |= VPO_WANTED;
1400 			msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1401 			VM_OBJECT_LOCK(new_object);
1402 			goto retry;
1403 		}
1404 		vm_page_rename(m, new_object, idx);
1405 		/* page automatically made dirty by rename and cache handled */
1406 		vm_page_busy(m);
1407 	}
1408 	vm_page_unlock_queues();
1409 	if (orig_object->type == OBJT_SWAP) {
1410 		/*
1411 		 * swap_pager_copy() can sleep, in which case the orig_object's
1412 		 * and new_object's locks are released and reacquired.
1413 		 */
1414 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1415 
1416 		/*
1417 		 * Transfer any cached pages from orig_object to new_object.
1418 		 */
1419 		if (__predict_false(orig_object->cache != NULL))
1420 			vm_page_cache_transfer(orig_object, offidxstart,
1421 			    new_object);
1422 	}
1423 	VM_OBJECT_UNLOCK(orig_object);
1424 	TAILQ_FOREACH(m, &new_object->memq, listq)
1425 		vm_page_wakeup(m);
1426 	VM_OBJECT_UNLOCK(new_object);
1427 	entry->object.vm_object = new_object;
1428 	entry->offset = 0LL;
1429 	vm_object_deallocate(orig_object);
1430 	VM_OBJECT_LOCK(new_object);
1431 }
1432 
1433 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1434 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1435 #define	OBSC_COLLAPSE_WAIT	0x0004
1436 
1437 static int
1438 vm_object_backing_scan(vm_object_t object, int op)
1439 {
1440 	int r = 1;
1441 	vm_page_t p;
1442 	vm_object_t backing_object;
1443 	vm_pindex_t backing_offset_index;
1444 
1445 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1446 	VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1447 
1448 	backing_object = object->backing_object;
1449 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1450 
1451 	/*
1452 	 * Initial conditions
1453 	 */
1454 	if (op & OBSC_TEST_ALL_SHADOWED) {
1455 		/*
1456 		 * We do not want to have to test for the existence of cache
1457 		 * or swap pages in the backing object.  XXX but with the
1458 		 * new swapper this would be pretty easy to do.
1459 		 *
1460 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1461 		 * been ZFOD faulted yet?  If we do not test for this, the
1462 		 * shadow test may succeed! XXX
1463 		 */
1464 		if (backing_object->type != OBJT_DEFAULT) {
1465 			return (0);
1466 		}
1467 	}
1468 	if (op & OBSC_COLLAPSE_WAIT) {
1469 		vm_object_set_flag(backing_object, OBJ_DEAD);
1470 	}
1471 
1472 	/*
1473 	 * Our scan
1474 	 */
1475 	p = TAILQ_FIRST(&backing_object->memq);
1476 	while (p) {
1477 		vm_page_t next = TAILQ_NEXT(p, listq);
1478 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1479 
1480 		if (op & OBSC_TEST_ALL_SHADOWED) {
1481 			vm_page_t pp;
1482 
1483 			/*
1484 			 * Ignore pages outside the parent object's range
1485 			 * and outside the parent object's mapping of the
1486 			 * backing object.
1487 			 *
1488 			 * note that we do not busy the backing object's
1489 			 * page.
1490 			 */
1491 			if (
1492 			    p->pindex < backing_offset_index ||
1493 			    new_pindex >= object->size
1494 			) {
1495 				p = next;
1496 				continue;
1497 			}
1498 
1499 			/*
1500 			 * See if the parent has the page or if the parent's
1501 			 * object pager has the page.  If the parent has the
1502 			 * page but the page is not valid, the parent's
1503 			 * object pager must have the page.
1504 			 *
1505 			 * If this fails, the parent does not completely shadow
1506 			 * the object and we might as well give up now.
1507 			 */
1508 
1509 			pp = vm_page_lookup(object, new_pindex);
1510 			if (
1511 			    (pp == NULL || pp->valid == 0) &&
1512 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1513 			) {
1514 				r = 0;
1515 				break;
1516 			}
1517 		}
1518 
1519 		/*
1520 		 * Check for busy page
1521 		 */
1522 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1523 			vm_page_t pp;
1524 
1525 			if (op & OBSC_COLLAPSE_NOWAIT) {
1526 				if ((p->oflags & VPO_BUSY) ||
1527 				    !p->valid ||
1528 				    p->busy) {
1529 					p = next;
1530 					continue;
1531 				}
1532 			} else if (op & OBSC_COLLAPSE_WAIT) {
1533 				if ((p->oflags & VPO_BUSY) || p->busy) {
1534 					vm_page_lock_queues();
1535 					vm_page_flag_set(p, PG_REFERENCED);
1536 					vm_page_unlock_queues();
1537 					VM_OBJECT_UNLOCK(object);
1538 					p->oflags |= VPO_WANTED;
1539 					msleep(p, VM_OBJECT_MTX(backing_object),
1540 					    PDROP | PVM, "vmocol", 0);
1541 					VM_OBJECT_LOCK(object);
1542 					VM_OBJECT_LOCK(backing_object);
1543 					/*
1544 					 * If we slept, anything could have
1545 					 * happened.  Since the object is
1546 					 * marked dead, the backing offset
1547 					 * should not have changed so we
1548 					 * just restart our scan.
1549 					 */
1550 					p = TAILQ_FIRST(&backing_object->memq);
1551 					continue;
1552 				}
1553 			}
1554 
1555 			KASSERT(
1556 			    p->object == backing_object,
1557 			    ("vm_object_backing_scan: object mismatch")
1558 			);
1559 
1560 			/*
1561 			 * Destroy any associated swap
1562 			 */
1563 			if (backing_object->type == OBJT_SWAP) {
1564 				swap_pager_freespace(
1565 				    backing_object,
1566 				    p->pindex,
1567 				    1
1568 				);
1569 			}
1570 
1571 			if (
1572 			    p->pindex < backing_offset_index ||
1573 			    new_pindex >= object->size
1574 			) {
1575 				/*
1576 				 * Page is out of the parent object's range, we
1577 				 * can simply destroy it.
1578 				 */
1579 				vm_page_lock_queues();
1580 				KASSERT(!pmap_page_is_mapped(p),
1581 				    ("freeing mapped page %p", p));
1582 				if (p->wire_count == 0)
1583 					vm_page_free(p);
1584 				else
1585 					vm_page_remove(p);
1586 				vm_page_unlock_queues();
1587 				p = next;
1588 				continue;
1589 			}
1590 
1591 			pp = vm_page_lookup(object, new_pindex);
1592 			if (
1593 			    pp != NULL ||
1594 			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1595 			) {
1596 				/*
1597 				 * page already exists in parent OR swap exists
1598 				 * for this location in the parent.  Destroy
1599 				 * the original page from the backing object.
1600 				 *
1601 				 * Leave the parent's page alone
1602 				 */
1603 				vm_page_lock_queues();
1604 				KASSERT(!pmap_page_is_mapped(p),
1605 				    ("freeing mapped page %p", p));
1606 				if (p->wire_count == 0)
1607 					vm_page_free(p);
1608 				else
1609 					vm_page_remove(p);
1610 				vm_page_unlock_queues();
1611 				p = next;
1612 				continue;
1613 			}
1614 
1615 #if VM_NRESERVLEVEL > 0
1616 			/*
1617 			 * Rename the reservation.
1618 			 */
1619 			vm_reserv_rename(p, object, backing_object,
1620 			    backing_offset_index);
1621 #endif
1622 
1623 			/*
1624 			 * Page does not exist in parent, rename the
1625 			 * page from the backing object to the main object.
1626 			 *
1627 			 * If the page was mapped to a process, it can remain
1628 			 * mapped through the rename.
1629 			 */
1630 			vm_page_lock_queues();
1631 			vm_page_rename(p, object, new_pindex);
1632 			vm_page_unlock_queues();
1633 			/* page automatically made dirty by rename */
1634 		}
1635 		p = next;
1636 	}
1637 	return (r);
1638 }
1639 
1640 
1641 /*
1642  * this version of collapse allows the operation to occur earlier and
1643  * when paging_in_progress is true for an object...  This is not a complete
1644  * operation, but should plug 99.9% of the rest of the leaks.
1645  */
1646 static void
1647 vm_object_qcollapse(vm_object_t object)
1648 {
1649 	vm_object_t backing_object = object->backing_object;
1650 
1651 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1652 	VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1653 
1654 	if (backing_object->ref_count != 1)
1655 		return;
1656 
1657 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1658 }
1659 
1660 /*
1661  *	vm_object_collapse:
1662  *
1663  *	Collapse an object with the object backing it.
1664  *	Pages in the backing object are moved into the
1665  *	parent, and the backing object is deallocated.
1666  */
1667 void
1668 vm_object_collapse(vm_object_t object)
1669 {
1670 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1671 
1672 	while (TRUE) {
1673 		vm_object_t backing_object;
1674 
1675 		/*
1676 		 * Verify that the conditions are right for collapse:
1677 		 *
1678 		 * The object exists and the backing object exists.
1679 		 */
1680 		if ((backing_object = object->backing_object) == NULL)
1681 			break;
1682 
1683 		/*
1684 		 * we check the backing object first, because it is most likely
1685 		 * not collapsable.
1686 		 */
1687 		VM_OBJECT_LOCK(backing_object);
1688 		if (backing_object->handle != NULL ||
1689 		    (backing_object->type != OBJT_DEFAULT &&
1690 		     backing_object->type != OBJT_SWAP) ||
1691 		    (backing_object->flags & OBJ_DEAD) ||
1692 		    object->handle != NULL ||
1693 		    (object->type != OBJT_DEFAULT &&
1694 		     object->type != OBJT_SWAP) ||
1695 		    (object->flags & OBJ_DEAD)) {
1696 			VM_OBJECT_UNLOCK(backing_object);
1697 			break;
1698 		}
1699 
1700 		if (
1701 		    object->paging_in_progress != 0 ||
1702 		    backing_object->paging_in_progress != 0
1703 		) {
1704 			vm_object_qcollapse(object);
1705 			VM_OBJECT_UNLOCK(backing_object);
1706 			break;
1707 		}
1708 		/*
1709 		 * We know that we can either collapse the backing object (if
1710 		 * the parent is the only reference to it) or (perhaps) have
1711 		 * the parent bypass the object if the parent happens to shadow
1712 		 * all the resident pages in the entire backing object.
1713 		 *
1714 		 * This is ignoring pager-backed pages such as swap pages.
1715 		 * vm_object_backing_scan fails the shadowing test in this
1716 		 * case.
1717 		 */
1718 		if (backing_object->ref_count == 1) {
1719 			/*
1720 			 * If there is exactly one reference to the backing
1721 			 * object, we can collapse it into the parent.
1722 			 */
1723 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1724 
1725 #if VM_NRESERVLEVEL > 0
1726 			/*
1727 			 * Break any reservations from backing_object.
1728 			 */
1729 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1730 				vm_reserv_break_all(backing_object);
1731 #endif
1732 
1733 			/*
1734 			 * Move the pager from backing_object to object.
1735 			 */
1736 			if (backing_object->type == OBJT_SWAP) {
1737 				/*
1738 				 * swap_pager_copy() can sleep, in which case
1739 				 * the backing_object's and object's locks are
1740 				 * released and reacquired.
1741 				 */
1742 				swap_pager_copy(
1743 				    backing_object,
1744 				    object,
1745 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1746 
1747 				/*
1748 				 * Free any cached pages from backing_object.
1749 				 */
1750 				if (__predict_false(backing_object->cache != NULL))
1751 					vm_page_cache_free(backing_object, 0, 0);
1752 			}
1753 			/*
1754 			 * Object now shadows whatever backing_object did.
1755 			 * Note that the reference to
1756 			 * backing_object->backing_object moves from within
1757 			 * backing_object to within object.
1758 			 */
1759 			LIST_REMOVE(object, shadow_list);
1760 			backing_object->shadow_count--;
1761 			backing_object->generation++;
1762 			if (backing_object->backing_object) {
1763 				VM_OBJECT_LOCK(backing_object->backing_object);
1764 				LIST_REMOVE(backing_object, shadow_list);
1765 				LIST_INSERT_HEAD(
1766 				    &backing_object->backing_object->shadow_head,
1767 				    object, shadow_list);
1768 				/*
1769 				 * The shadow_count has not changed.
1770 				 */
1771 				backing_object->backing_object->generation++;
1772 				VM_OBJECT_UNLOCK(backing_object->backing_object);
1773 			}
1774 			object->backing_object = backing_object->backing_object;
1775 			object->backing_object_offset +=
1776 			    backing_object->backing_object_offset;
1777 
1778 			/*
1779 			 * Discard backing_object.
1780 			 *
1781 			 * Since the backing object has no pages, no pager left,
1782 			 * and no object references within it, all that is
1783 			 * necessary is to dispose of it.
1784 			 */
1785 			KASSERT(backing_object->ref_count == 1, (
1786 "backing_object %p was somehow re-referenced during collapse!",
1787 			    backing_object));
1788 			VM_OBJECT_UNLOCK(backing_object);
1789 			vm_object_destroy(backing_object);
1790 
1791 			object_collapses++;
1792 		} else {
1793 			vm_object_t new_backing_object;
1794 
1795 			/*
1796 			 * If we do not entirely shadow the backing object,
1797 			 * there is nothing we can do so we give up.
1798 			 */
1799 			if (object->resident_page_count != object->size &&
1800 			    vm_object_backing_scan(object,
1801 			    OBSC_TEST_ALL_SHADOWED) == 0) {
1802 				VM_OBJECT_UNLOCK(backing_object);
1803 				break;
1804 			}
1805 
1806 			/*
1807 			 * Make the parent shadow the next object in the
1808 			 * chain.  Deallocating backing_object will not remove
1809 			 * it, since its reference count is at least 2.
1810 			 */
1811 			LIST_REMOVE(object, shadow_list);
1812 			backing_object->shadow_count--;
1813 			backing_object->generation++;
1814 
1815 			new_backing_object = backing_object->backing_object;
1816 			if ((object->backing_object = new_backing_object) != NULL) {
1817 				VM_OBJECT_LOCK(new_backing_object);
1818 				LIST_INSERT_HEAD(
1819 				    &new_backing_object->shadow_head,
1820 				    object,
1821 				    shadow_list
1822 				);
1823 				new_backing_object->shadow_count++;
1824 				new_backing_object->generation++;
1825 				vm_object_reference_locked(new_backing_object);
1826 				VM_OBJECT_UNLOCK(new_backing_object);
1827 				object->backing_object_offset +=
1828 					backing_object->backing_object_offset;
1829 			}
1830 
1831 			/*
1832 			 * Drop the reference count on backing_object. Since
1833 			 * its ref_count was at least 2, it will not vanish.
1834 			 */
1835 			backing_object->ref_count--;
1836 			VM_OBJECT_UNLOCK(backing_object);
1837 			object_bypasses++;
1838 		}
1839 
1840 		/*
1841 		 * Try again with this object's new backing object.
1842 		 */
1843 	}
1844 }
1845 
1846 /*
1847  *	vm_object_page_remove:
1848  *
1849  *	For the given object, either frees or invalidates each of the
1850  *	specified pages.  In general, a page is freed.  However, if a
1851  *	page is wired for any reason other than the existence of a
1852  *	managed, wired mapping, then it may be invalidated but not
1853  *	removed from the object.  Pages are specified by the given
1854  *	range ["start", "end") and Boolean "clean_only".  As a
1855  *	special case, if "end" is zero, then the range extends from
1856  *	"start" to the end of the object.  If "clean_only" is TRUE,
1857  *	then only the non-dirty pages within the specified range are
1858  *	affected.
1859  *
1860  *	In general, this operation should only be performed on objects
1861  *	that contain managed pages.  There are two exceptions.  First,
1862  *	it may be performed on the kernel and kmem objects.  Second,
1863  *	it may be used by msync(..., MS_INVALIDATE) to invalidate
1864  *	device-backed pages.
1865  *
1866  *	The object must be locked.
1867  */
1868 void
1869 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1870     boolean_t clean_only)
1871 {
1872 	vm_page_t p, next;
1873 	int wirings;
1874 
1875 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1876 	if (object->resident_page_count == 0)
1877 		goto skipmemq;
1878 
1879 	/*
1880 	 * Since physically-backed objects do not use managed pages, we can't
1881 	 * remove pages from the object (we must instead remove the page
1882 	 * references, and then destroy the object).
1883 	 */
1884 	KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1885 	    object == kmem_object,
1886 	    ("attempt to remove pages from a physical object"));
1887 
1888 	vm_object_pip_add(object, 1);
1889 again:
1890 	if ((p = TAILQ_FIRST(&object->memq)) != NULL) {
1891 		if (p->pindex < start) {
1892 			p = vm_page_splay(start, object->root);
1893 			if ((object->root = p)->pindex < start)
1894 				p = TAILQ_NEXT(p, listq);
1895 		}
1896 	}
1897 	vm_page_lock_queues();
1898 	/*
1899 	 * Assert: the variable p is either (1) the page with the
1900 	 * least pindex greater than or equal to the parameter pindex
1901 	 * or (2) NULL.
1902 	 */
1903 	for (;
1904 	     p != NULL && (p->pindex < end || end == 0);
1905 	     p = next) {
1906 		next = TAILQ_NEXT(p, listq);
1907 
1908 		/*
1909 		 * If the page is wired for any reason besides the
1910 		 * existence of managed, wired mappings, then it cannot
1911 		 * be freed.  For example, fictitious pages, which
1912 		 * represent device memory, are inherently wired and
1913 		 * cannot be freed.  They can, however, be invalidated
1914 		 * if "clean_only" is FALSE.
1915 		 */
1916 		if ((wirings = p->wire_count) != 0 &&
1917 		    (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1918 			/* Fictitious pages do not have managed mappings. */
1919 			if ((p->flags & PG_FICTITIOUS) == 0)
1920 				pmap_remove_all(p);
1921 			/* Account for removal of managed, wired mappings. */
1922 			p->wire_count -= wirings;
1923 			if (!clean_only) {
1924 				p->valid = 0;
1925 				vm_page_undirty(p);
1926 			}
1927 			continue;
1928 		}
1929 		if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1930 			goto again;
1931 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
1932 		    ("vm_object_page_remove: page %p is fictitious", p));
1933 		if (clean_only && p->valid) {
1934 			pmap_remove_write(p);
1935 			if (p->dirty)
1936 				continue;
1937 		}
1938 		pmap_remove_all(p);
1939 		/* Account for removal of managed, wired mappings. */
1940 		if (wirings != 0)
1941 			p->wire_count -= wirings;
1942 		vm_page_free(p);
1943 	}
1944 	vm_page_unlock_queues();
1945 	vm_object_pip_wakeup(object);
1946 skipmemq:
1947 	if (__predict_false(object->cache != NULL))
1948 		vm_page_cache_free(object, start, end);
1949 }
1950 
1951 /*
1952  *	Populate the specified range of the object with valid pages.  Returns
1953  *	TRUE if the range is successfully populated and FALSE otherwise.
1954  *
1955  *	Note: This function should be optimized to pass a larger array of
1956  *	pages to vm_pager_get_pages() before it is applied to a non-
1957  *	OBJT_DEVICE object.
1958  *
1959  *	The object must be locked.
1960  */
1961 boolean_t
1962 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1963 {
1964 	vm_page_t m, ma[1];
1965 	vm_pindex_t pindex;
1966 	int rv;
1967 
1968 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1969 	for (pindex = start; pindex < end; pindex++) {
1970 		m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1971 		    VM_ALLOC_RETRY);
1972 		if (m->valid != VM_PAGE_BITS_ALL) {
1973 			ma[0] = m;
1974 			rv = vm_pager_get_pages(object, ma, 1, 0);
1975 			m = vm_page_lookup(object, pindex);
1976 			if (m == NULL)
1977 				break;
1978 			if (rv != VM_PAGER_OK) {
1979 				vm_page_lock_queues();
1980 				vm_page_free(m);
1981 				vm_page_unlock_queues();
1982 				break;
1983 			}
1984 		}
1985 		/*
1986 		 * Keep "m" busy because a subsequent iteration may unlock
1987 		 * the object.
1988 		 */
1989 	}
1990 	if (pindex > start) {
1991 		m = vm_page_lookup(object, start);
1992 		while (m != NULL && m->pindex < pindex) {
1993 			vm_page_wakeup(m);
1994 			m = TAILQ_NEXT(m, listq);
1995 		}
1996 	}
1997 	return (pindex == end);
1998 }
1999 
2000 /*
2001  *	Routine:	vm_object_coalesce
2002  *	Function:	Coalesces two objects backing up adjoining
2003  *			regions of memory into a single object.
2004  *
2005  *	returns TRUE if objects were combined.
2006  *
2007  *	NOTE:	Only works at the moment if the second object is NULL -
2008  *		if it's not, which object do we lock first?
2009  *
2010  *	Parameters:
2011  *		prev_object	First object to coalesce
2012  *		prev_offset	Offset into prev_object
2013  *		prev_size	Size of reference to prev_object
2014  *		next_size	Size of reference to the second object
2015  *		reserved	Indicator that extension region has
2016  *				swap accounted for
2017  *
2018  *	Conditions:
2019  *	The object must *not* be locked.
2020  */
2021 boolean_t
2022 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2023     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2024 {
2025 	vm_pindex_t next_pindex;
2026 
2027 	if (prev_object == NULL)
2028 		return (TRUE);
2029 	VM_OBJECT_LOCK(prev_object);
2030 	if (prev_object->type != OBJT_DEFAULT &&
2031 	    prev_object->type != OBJT_SWAP) {
2032 		VM_OBJECT_UNLOCK(prev_object);
2033 		return (FALSE);
2034 	}
2035 
2036 	/*
2037 	 * Try to collapse the object first
2038 	 */
2039 	vm_object_collapse(prev_object);
2040 
2041 	/*
2042 	 * Can't coalesce if: . more than one reference . paged out . shadows
2043 	 * another object . has a copy elsewhere (any of which mean that the
2044 	 * pages not mapped to prev_entry may be in use anyway)
2045 	 */
2046 	if (prev_object->backing_object != NULL) {
2047 		VM_OBJECT_UNLOCK(prev_object);
2048 		return (FALSE);
2049 	}
2050 
2051 	prev_size >>= PAGE_SHIFT;
2052 	next_size >>= PAGE_SHIFT;
2053 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2054 
2055 	if ((prev_object->ref_count > 1) &&
2056 	    (prev_object->size != next_pindex)) {
2057 		VM_OBJECT_UNLOCK(prev_object);
2058 		return (FALSE);
2059 	}
2060 
2061 	/*
2062 	 * Account for the charge.
2063 	 */
2064 	if (prev_object->uip != NULL) {
2065 
2066 		/*
2067 		 * If prev_object was charged, then this mapping,
2068 		 * althought not charged now, may become writable
2069 		 * later. Non-NULL uip in the object would prevent
2070 		 * swap reservation during enabling of the write
2071 		 * access, so reserve swap now. Failed reservation
2072 		 * cause allocation of the separate object for the map
2073 		 * entry, and swap reservation for this entry is
2074 		 * managed in appropriate time.
2075 		 */
2076 		if (!reserved && !swap_reserve_by_uid(ptoa(next_size),
2077 		    prev_object->uip)) {
2078 			return (FALSE);
2079 		}
2080 		prev_object->charge += ptoa(next_size);
2081 	}
2082 
2083 	/*
2084 	 * Remove any pages that may still be in the object from a previous
2085 	 * deallocation.
2086 	 */
2087 	if (next_pindex < prev_object->size) {
2088 		vm_object_page_remove(prev_object,
2089 				      next_pindex,
2090 				      next_pindex + next_size, FALSE);
2091 		if (prev_object->type == OBJT_SWAP)
2092 			swap_pager_freespace(prev_object,
2093 					     next_pindex, next_size);
2094 #if 0
2095 		if (prev_object->uip != NULL) {
2096 			KASSERT(prev_object->charge >=
2097 			    ptoa(prev_object->size - next_pindex),
2098 			    ("object %p overcharged 1 %jx %jx", prev_object,
2099 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2100 			prev_object->charge -= ptoa(prev_object->size -
2101 			    next_pindex);
2102 		}
2103 #endif
2104 	}
2105 
2106 	/*
2107 	 * Extend the object if necessary.
2108 	 */
2109 	if (next_pindex + next_size > prev_object->size)
2110 		prev_object->size = next_pindex + next_size;
2111 
2112 	VM_OBJECT_UNLOCK(prev_object);
2113 	return (TRUE);
2114 }
2115 
2116 void
2117 vm_object_set_writeable_dirty(vm_object_t object)
2118 {
2119 	struct vnode *vp;
2120 
2121 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2122 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2123 		return;
2124 	vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2125 	if (object->type == OBJT_VNODE &&
2126 	    (vp = (struct vnode *)object->handle) != NULL) {
2127 		VI_LOCK(vp);
2128 		vp->v_iflag |= VI_OBJDIRTY;
2129 		VI_UNLOCK(vp);
2130 	}
2131 }
2132 
2133 #include "opt_ddb.h"
2134 #ifdef DDB
2135 #include <sys/kernel.h>
2136 
2137 #include <sys/cons.h>
2138 
2139 #include <ddb/ddb.h>
2140 
2141 static int
2142 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2143 {
2144 	vm_map_t tmpm;
2145 	vm_map_entry_t tmpe;
2146 	vm_object_t obj;
2147 	int entcount;
2148 
2149 	if (map == 0)
2150 		return 0;
2151 
2152 	if (entry == 0) {
2153 		tmpe = map->header.next;
2154 		entcount = map->nentries;
2155 		while (entcount-- && (tmpe != &map->header)) {
2156 			if (_vm_object_in_map(map, object, tmpe)) {
2157 				return 1;
2158 			}
2159 			tmpe = tmpe->next;
2160 		}
2161 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2162 		tmpm = entry->object.sub_map;
2163 		tmpe = tmpm->header.next;
2164 		entcount = tmpm->nentries;
2165 		while (entcount-- && tmpe != &tmpm->header) {
2166 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2167 				return 1;
2168 			}
2169 			tmpe = tmpe->next;
2170 		}
2171 	} else if ((obj = entry->object.vm_object) != NULL) {
2172 		for (; obj; obj = obj->backing_object)
2173 			if (obj == object) {
2174 				return 1;
2175 			}
2176 	}
2177 	return 0;
2178 }
2179 
2180 static int
2181 vm_object_in_map(vm_object_t object)
2182 {
2183 	struct proc *p;
2184 
2185 	/* sx_slock(&allproc_lock); */
2186 	FOREACH_PROC_IN_SYSTEM(p) {
2187 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2188 			continue;
2189 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2190 			/* sx_sunlock(&allproc_lock); */
2191 			return 1;
2192 		}
2193 	}
2194 	/* sx_sunlock(&allproc_lock); */
2195 	if (_vm_object_in_map(kernel_map, object, 0))
2196 		return 1;
2197 	if (_vm_object_in_map(kmem_map, object, 0))
2198 		return 1;
2199 	if (_vm_object_in_map(pager_map, object, 0))
2200 		return 1;
2201 	if (_vm_object_in_map(buffer_map, object, 0))
2202 		return 1;
2203 	return 0;
2204 }
2205 
2206 DB_SHOW_COMMAND(vmochk, vm_object_check)
2207 {
2208 	vm_object_t object;
2209 
2210 	/*
2211 	 * make sure that internal objs are in a map somewhere
2212 	 * and none have zero ref counts.
2213 	 */
2214 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2215 		if (object->handle == NULL &&
2216 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2217 			if (object->ref_count == 0) {
2218 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2219 					(long)object->size);
2220 			}
2221 			if (!vm_object_in_map(object)) {
2222 				db_printf(
2223 			"vmochk: internal obj is not in a map: "
2224 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2225 				    object->ref_count, (u_long)object->size,
2226 				    (u_long)object->size,
2227 				    (void *)object->backing_object);
2228 			}
2229 		}
2230 	}
2231 }
2232 
2233 /*
2234  *	vm_object_print:	[ debug ]
2235  */
2236 DB_SHOW_COMMAND(object, vm_object_print_static)
2237 {
2238 	/* XXX convert args. */
2239 	vm_object_t object = (vm_object_t)addr;
2240 	boolean_t full = have_addr;
2241 
2242 	vm_page_t p;
2243 
2244 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2245 #define	count	was_count
2246 
2247 	int count;
2248 
2249 	if (object == NULL)
2250 		return;
2251 
2252 	db_iprintf(
2253 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x uip %d charge %jx\n",
2254 	    object, (int)object->type, (uintmax_t)object->size,
2255 	    object->resident_page_count, object->ref_count, object->flags,
2256 	    object->uip ? object->uip->ui_uid : -1, (uintmax_t)object->charge);
2257 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2258 	    object->shadow_count,
2259 	    object->backing_object ? object->backing_object->ref_count : 0,
2260 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2261 
2262 	if (!full)
2263 		return;
2264 
2265 	db_indent += 2;
2266 	count = 0;
2267 	TAILQ_FOREACH(p, &object->memq, listq) {
2268 		if (count == 0)
2269 			db_iprintf("memory:=");
2270 		else if (count == 6) {
2271 			db_printf("\n");
2272 			db_iprintf(" ...");
2273 			count = 0;
2274 		} else
2275 			db_printf(",");
2276 		count++;
2277 
2278 		db_printf("(off=0x%jx,page=0x%jx)",
2279 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2280 	}
2281 	if (count != 0)
2282 		db_printf("\n");
2283 	db_indent -= 2;
2284 }
2285 
2286 /* XXX. */
2287 #undef count
2288 
2289 /* XXX need this non-static entry for calling from vm_map_print. */
2290 void
2291 vm_object_print(
2292         /* db_expr_t */ long addr,
2293 	boolean_t have_addr,
2294 	/* db_expr_t */ long count,
2295 	char *modif)
2296 {
2297 	vm_object_print_static(addr, have_addr, count, modif);
2298 }
2299 
2300 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2301 {
2302 	vm_object_t object;
2303 	vm_pindex_t fidx;
2304 	vm_paddr_t pa;
2305 	vm_page_t m, prev_m;
2306 	int rcount, nl, c;
2307 
2308 	nl = 0;
2309 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2310 		db_printf("new object: %p\n", (void *)object);
2311 		if (nl > 18) {
2312 			c = cngetc();
2313 			if (c != ' ')
2314 				return;
2315 			nl = 0;
2316 		}
2317 		nl++;
2318 		rcount = 0;
2319 		fidx = 0;
2320 		pa = -1;
2321 		TAILQ_FOREACH(m, &object->memq, listq) {
2322 			if (m->pindex > 128)
2323 				break;
2324 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2325 			    prev_m->pindex + 1 != m->pindex) {
2326 				if (rcount) {
2327 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2328 						(long)fidx, rcount, (long)pa);
2329 					if (nl > 18) {
2330 						c = cngetc();
2331 						if (c != ' ')
2332 							return;
2333 						nl = 0;
2334 					}
2335 					nl++;
2336 					rcount = 0;
2337 				}
2338 			}
2339 			if (rcount &&
2340 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2341 				++rcount;
2342 				continue;
2343 			}
2344 			if (rcount) {
2345 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2346 					(long)fidx, rcount, (long)pa);
2347 				if (nl > 18) {
2348 					c = cngetc();
2349 					if (c != ' ')
2350 						return;
2351 					nl = 0;
2352 				}
2353 				nl++;
2354 			}
2355 			fidx = m->pindex;
2356 			pa = VM_PAGE_TO_PHYS(m);
2357 			rcount = 1;
2358 		}
2359 		if (rcount) {
2360 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2361 				(long)fidx, rcount, (long)pa);
2362 			if (nl > 18) {
2363 				c = cngetc();
2364 				if (c != ' ')
2365 					return;
2366 				nl = 0;
2367 			}
2368 			nl++;
2369 		}
2370 	}
2371 }
2372 #endif /* DDB */
2373