xref: /freebsd/sys/vm/vm_object.c (revision 10f0bcab61ef441cb5af32fb706688d8cbd55dc0)
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/vnode.h>
81 #include <sys/vmmeter.h>
82 #include <sys/sx.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_param.h>
86 #include <vm/pmap.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_pager.h>
92 #include <vm/swap_pager.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
95 #include <vm/vm_reserv.h>
96 #include <vm/uma.h>
97 
98 #define EASY_SCAN_FACTOR       8
99 
100 #define MSYNC_FLUSH_HARDSEQ	0x01
101 #define MSYNC_FLUSH_SOFTSEQ	0x02
102 
103 /*
104  * msync / VM object flushing optimizations
105  */
106 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ;
107 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags,
108         CTLFLAG_RW, &msync_flush_flags, 0, "");
109 
110 static int old_msync;
111 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
112     "Use old (insecure) msync behavior");
113 
114 static void	vm_object_qcollapse(vm_object_t object);
115 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
116 static void	vm_object_vndeallocate(vm_object_t object);
117 
118 /*
119  *	Virtual memory objects maintain the actual data
120  *	associated with allocated virtual memory.  A given
121  *	page of memory exists within exactly one object.
122  *
123  *	An object is only deallocated when all "references"
124  *	are given up.  Only one "reference" to a given
125  *	region of an object should be writeable.
126  *
127  *	Associated with each object is a list of all resident
128  *	memory pages belonging to that object; this list is
129  *	maintained by the "vm_page" module, and locked by the object's
130  *	lock.
131  *
132  *	Each object also records a "pager" routine which is
133  *	used to retrieve (and store) pages to the proper backing
134  *	storage.  In addition, objects may be backed by other
135  *	objects from which they were virtual-copied.
136  *
137  *	The only items within the object structure which are
138  *	modified after time of creation are:
139  *		reference count		locked by object's lock
140  *		pager routine		locked by object's lock
141  *
142  */
143 
144 struct object_q vm_object_list;
145 struct mtx vm_object_list_mtx;	/* lock for object list and count */
146 
147 struct vm_object kernel_object_store;
148 struct vm_object kmem_object_store;
149 
150 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats");
151 
152 static long object_collapses;
153 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
154     &object_collapses, 0, "VM object collapses");
155 
156 static long object_bypasses;
157 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
158     &object_bypasses, 0, "VM object bypasses");
159 
160 static uma_zone_t obj_zone;
161 
162 static int vm_object_zinit(void *mem, int size, int flags);
163 
164 #ifdef INVARIANTS
165 static void vm_object_zdtor(void *mem, int size, void *arg);
166 
167 static void
168 vm_object_zdtor(void *mem, int size, void *arg)
169 {
170 	vm_object_t object;
171 
172 	object = (vm_object_t)mem;
173 	KASSERT(TAILQ_EMPTY(&object->memq),
174 	    ("object %p has resident pages",
175 	    object));
176 #if VM_NRESERVLEVEL > 0
177 	KASSERT(LIST_EMPTY(&object->rvq),
178 	    ("object %p has reservations",
179 	    object));
180 #endif
181 	KASSERT(object->cache == NULL,
182 	    ("object %p has cached pages",
183 	    object));
184 	KASSERT(object->paging_in_progress == 0,
185 	    ("object %p paging_in_progress = %d",
186 	    object, object->paging_in_progress));
187 	KASSERT(object->resident_page_count == 0,
188 	    ("object %p resident_page_count = %d",
189 	    object, object->resident_page_count));
190 	KASSERT(object->shadow_count == 0,
191 	    ("object %p shadow_count = %d",
192 	    object, object->shadow_count));
193 }
194 #endif
195 
196 static int
197 vm_object_zinit(void *mem, int size, int flags)
198 {
199 	vm_object_t object;
200 
201 	object = (vm_object_t)mem;
202 	bzero(&object->mtx, sizeof(object->mtx));
203 	VM_OBJECT_LOCK_INIT(object, "standard object");
204 
205 	/* These are true for any object that has been freed */
206 	object->paging_in_progress = 0;
207 	object->resident_page_count = 0;
208 	object->shadow_count = 0;
209 	return (0);
210 }
211 
212 void
213 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
214 {
215 
216 	TAILQ_INIT(&object->memq);
217 	LIST_INIT(&object->shadow_head);
218 
219 	object->root = NULL;
220 	object->type = type;
221 	object->size = size;
222 	object->generation = 1;
223 	object->ref_count = 1;
224 	object->flags = 0;
225 	if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
226 		object->flags = OBJ_ONEMAPPING;
227 	object->pg_color = 0;
228 	object->handle = NULL;
229 	object->backing_object = NULL;
230 	object->backing_object_offset = (vm_ooffset_t) 0;
231 #if VM_NRESERVLEVEL > 0
232 	LIST_INIT(&object->rvq);
233 #endif
234 	object->cache = NULL;
235 
236 	mtx_lock(&vm_object_list_mtx);
237 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
238 	mtx_unlock(&vm_object_list_mtx);
239 }
240 
241 /*
242  *	vm_object_init:
243  *
244  *	Initialize the VM objects module.
245  */
246 void
247 vm_object_init(void)
248 {
249 	TAILQ_INIT(&vm_object_list);
250 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
251 
252 	VM_OBJECT_LOCK_INIT(&kernel_object_store, "kernel object");
253 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
254 	    kernel_object);
255 #if VM_NRESERVLEVEL > 0
256 	kernel_object->flags |= OBJ_COLORED;
257 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
258 #endif
259 
260 	VM_OBJECT_LOCK_INIT(&kmem_object_store, "kmem object");
261 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
262 	    kmem_object);
263 #if VM_NRESERVLEVEL > 0
264 	kmem_object->flags |= OBJ_COLORED;
265 	kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
266 #endif
267 
268 	/*
269 	 * The lock portion of struct vm_object must be type stable due
270 	 * to vm_pageout_fallback_object_lock locking a vm object
271 	 * without holding any references to it.
272 	 */
273 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
274 #ifdef INVARIANTS
275 	    vm_object_zdtor,
276 #else
277 	    NULL,
278 #endif
279 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
280 }
281 
282 void
283 vm_object_clear_flag(vm_object_t object, u_short bits)
284 {
285 
286 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
287 	object->flags &= ~bits;
288 }
289 
290 void
291 vm_object_pip_add(vm_object_t object, short i)
292 {
293 
294 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
295 	object->paging_in_progress += i;
296 }
297 
298 void
299 vm_object_pip_subtract(vm_object_t object, short i)
300 {
301 
302 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
303 	object->paging_in_progress -= i;
304 }
305 
306 void
307 vm_object_pip_wakeup(vm_object_t object)
308 {
309 
310 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
311 	object->paging_in_progress--;
312 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
313 		vm_object_clear_flag(object, OBJ_PIPWNT);
314 		wakeup(object);
315 	}
316 }
317 
318 void
319 vm_object_pip_wakeupn(vm_object_t object, short i)
320 {
321 
322 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
323 	if (i)
324 		object->paging_in_progress -= i;
325 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
326 		vm_object_clear_flag(object, OBJ_PIPWNT);
327 		wakeup(object);
328 	}
329 }
330 
331 void
332 vm_object_pip_wait(vm_object_t object, char *waitid)
333 {
334 
335 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
336 	while (object->paging_in_progress) {
337 		object->flags |= OBJ_PIPWNT;
338 		msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
339 	}
340 }
341 
342 /*
343  *	vm_object_allocate:
344  *
345  *	Returns a new object with the given size.
346  */
347 vm_object_t
348 vm_object_allocate(objtype_t type, vm_pindex_t size)
349 {
350 	vm_object_t object;
351 
352 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
353 	_vm_object_allocate(type, size, object);
354 	return (object);
355 }
356 
357 
358 /*
359  *	vm_object_reference:
360  *
361  *	Gets another reference to the given object.  Note: OBJ_DEAD
362  *	objects can be referenced during final cleaning.
363  */
364 void
365 vm_object_reference(vm_object_t object)
366 {
367 	struct vnode *vp;
368 
369 	if (object == NULL)
370 		return;
371 	VM_OBJECT_LOCK(object);
372 	object->ref_count++;
373 	if (object->type == OBJT_VNODE) {
374 		int vfslocked;
375 
376 		vp = object->handle;
377 		VM_OBJECT_UNLOCK(object);
378 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
379 		vget(vp, LK_RETRY, curthread);
380 		VFS_UNLOCK_GIANT(vfslocked);
381 	} else
382 		VM_OBJECT_UNLOCK(object);
383 }
384 
385 /*
386  *	vm_object_reference_locked:
387  *
388  *	Gets another reference to the given object.
389  *
390  *	The object must be locked.
391  */
392 void
393 vm_object_reference_locked(vm_object_t object)
394 {
395 	struct vnode *vp;
396 
397 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
398 	KASSERT((object->flags & OBJ_DEAD) == 0,
399 	    ("vm_object_reference_locked: dead object referenced"));
400 	object->ref_count++;
401 	if (object->type == OBJT_VNODE) {
402 		vp = object->handle;
403 		vref(vp);
404 	}
405 }
406 
407 /*
408  * Handle deallocating an object of type OBJT_VNODE.
409  */
410 static void
411 vm_object_vndeallocate(vm_object_t object)
412 {
413 	struct vnode *vp = (struct vnode *) object->handle;
414 
415 	VFS_ASSERT_GIANT(vp->v_mount);
416 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
417 	KASSERT(object->type == OBJT_VNODE,
418 	    ("vm_object_vndeallocate: not a vnode object"));
419 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
420 #ifdef INVARIANTS
421 	if (object->ref_count == 0) {
422 		vprint("vm_object_vndeallocate", vp);
423 		panic("vm_object_vndeallocate: bad object reference count");
424 	}
425 #endif
426 
427 	object->ref_count--;
428 	if (object->ref_count == 0) {
429 		mp_fixme("Unlocked vflag access.");
430 		vp->v_vflag &= ~VV_TEXT;
431 	}
432 	VM_OBJECT_UNLOCK(object);
433 	/*
434 	 * vrele may need a vop lock
435 	 */
436 	vrele(vp);
437 }
438 
439 /*
440  *	vm_object_deallocate:
441  *
442  *	Release a reference to the specified object,
443  *	gained either through a vm_object_allocate
444  *	or a vm_object_reference call.  When all references
445  *	are gone, storage associated with this object
446  *	may be relinquished.
447  *
448  *	No object may be locked.
449  */
450 void
451 vm_object_deallocate(vm_object_t object)
452 {
453 	vm_object_t temp;
454 
455 	while (object != NULL) {
456 		int vfslocked;
457 
458 		vfslocked = 0;
459 	restart:
460 		VM_OBJECT_LOCK(object);
461 		if (object->type == OBJT_VNODE) {
462 			struct vnode *vp = (struct vnode *) object->handle;
463 
464 			/*
465 			 * Conditionally acquire Giant for a vnode-backed
466 			 * object.  We have to be careful since the type of
467 			 * a vnode object can change while the object is
468 			 * unlocked.
469 			 */
470 			if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
471 				vfslocked = 1;
472 				if (!mtx_trylock(&Giant)) {
473 					VM_OBJECT_UNLOCK(object);
474 					mtx_lock(&Giant);
475 					goto restart;
476 				}
477 			}
478 			vm_object_vndeallocate(object);
479 			VFS_UNLOCK_GIANT(vfslocked);
480 			return;
481 		} else
482 			/*
483 			 * This is to handle the case that the object
484 			 * changed type while we dropped its lock to
485 			 * obtain Giant.
486 			 */
487 			VFS_UNLOCK_GIANT(vfslocked);
488 
489 		KASSERT(object->ref_count != 0,
490 			("vm_object_deallocate: object deallocated too many times: %d", object->type));
491 
492 		/*
493 		 * If the reference count goes to 0 we start calling
494 		 * vm_object_terminate() on the object chain.
495 		 * A ref count of 1 may be a special case depending on the
496 		 * shadow count being 0 or 1.
497 		 */
498 		object->ref_count--;
499 		if (object->ref_count > 1) {
500 			VM_OBJECT_UNLOCK(object);
501 			return;
502 		} else if (object->ref_count == 1) {
503 			if (object->shadow_count == 0 &&
504 			    object->handle == NULL &&
505 			    (object->type == OBJT_DEFAULT ||
506 			     object->type == OBJT_SWAP)) {
507 				vm_object_set_flag(object, OBJ_ONEMAPPING);
508 			} else if ((object->shadow_count == 1) &&
509 			    (object->handle == NULL) &&
510 			    (object->type == OBJT_DEFAULT ||
511 			     object->type == OBJT_SWAP)) {
512 				vm_object_t robject;
513 
514 				robject = LIST_FIRST(&object->shadow_head);
515 				KASSERT(robject != NULL,
516 				    ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
517 					 object->ref_count,
518 					 object->shadow_count));
519 				if (!VM_OBJECT_TRYLOCK(robject)) {
520 					/*
521 					 * Avoid a potential deadlock.
522 					 */
523 					object->ref_count++;
524 					VM_OBJECT_UNLOCK(object);
525 					/*
526 					 * More likely than not the thread
527 					 * holding robject's lock has lower
528 					 * priority than the current thread.
529 					 * Let the lower priority thread run.
530 					 */
531 					pause("vmo_de", 1);
532 					continue;
533 				}
534 				/*
535 				 * Collapse object into its shadow unless its
536 				 * shadow is dead.  In that case, object will
537 				 * be deallocated by the thread that is
538 				 * deallocating its shadow.
539 				 */
540 				if ((robject->flags & OBJ_DEAD) == 0 &&
541 				    (robject->handle == NULL) &&
542 				    (robject->type == OBJT_DEFAULT ||
543 				     robject->type == OBJT_SWAP)) {
544 
545 					robject->ref_count++;
546 retry:
547 					if (robject->paging_in_progress) {
548 						VM_OBJECT_UNLOCK(object);
549 						vm_object_pip_wait(robject,
550 						    "objde1");
551 						temp = robject->backing_object;
552 						if (object == temp) {
553 							VM_OBJECT_LOCK(object);
554 							goto retry;
555 						}
556 					} else if (object->paging_in_progress) {
557 						VM_OBJECT_UNLOCK(robject);
558 						object->flags |= OBJ_PIPWNT;
559 						msleep(object,
560 						    VM_OBJECT_MTX(object),
561 						    PDROP | PVM, "objde2", 0);
562 						VM_OBJECT_LOCK(robject);
563 						temp = robject->backing_object;
564 						if (object == temp) {
565 							VM_OBJECT_LOCK(object);
566 							goto retry;
567 						}
568 					} else
569 						VM_OBJECT_UNLOCK(object);
570 
571 					if (robject->ref_count == 1) {
572 						robject->ref_count--;
573 						object = robject;
574 						goto doterm;
575 					}
576 					object = robject;
577 					vm_object_collapse(object);
578 					VM_OBJECT_UNLOCK(object);
579 					continue;
580 				}
581 				VM_OBJECT_UNLOCK(robject);
582 			}
583 			VM_OBJECT_UNLOCK(object);
584 			return;
585 		}
586 doterm:
587 		temp = object->backing_object;
588 		if (temp != NULL) {
589 			VM_OBJECT_LOCK(temp);
590 			LIST_REMOVE(object, shadow_list);
591 			temp->shadow_count--;
592 			temp->generation++;
593 			VM_OBJECT_UNLOCK(temp);
594 			object->backing_object = NULL;
595 		}
596 		/*
597 		 * Don't double-terminate, we could be in a termination
598 		 * recursion due to the terminate having to sync data
599 		 * to disk.
600 		 */
601 		if ((object->flags & OBJ_DEAD) == 0)
602 			vm_object_terminate(object);
603 		else
604 			VM_OBJECT_UNLOCK(object);
605 		object = temp;
606 	}
607 }
608 
609 /*
610  *	vm_object_terminate actually destroys the specified object, freeing
611  *	up all previously used resources.
612  *
613  *	The object must be locked.
614  *	This routine may block.
615  */
616 void
617 vm_object_terminate(vm_object_t object)
618 {
619 	vm_page_t p;
620 
621 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
622 
623 	/*
624 	 * Make sure no one uses us.
625 	 */
626 	vm_object_set_flag(object, OBJ_DEAD);
627 
628 	/*
629 	 * wait for the pageout daemon to be done with the object
630 	 */
631 	vm_object_pip_wait(object, "objtrm");
632 
633 	KASSERT(!object->paging_in_progress,
634 		("vm_object_terminate: pageout in progress"));
635 
636 	/*
637 	 * Clean and free the pages, as appropriate. All references to the
638 	 * object are gone, so we don't need to lock it.
639 	 */
640 	if (object->type == OBJT_VNODE) {
641 		struct vnode *vp = (struct vnode *)object->handle;
642 
643 		/*
644 		 * Clean pages and flush buffers.
645 		 */
646 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
647 		VM_OBJECT_UNLOCK(object);
648 
649 		vinvalbuf(vp, V_SAVE, NULL, 0, 0);
650 
651 		VM_OBJECT_LOCK(object);
652 	}
653 
654 	KASSERT(object->ref_count == 0,
655 		("vm_object_terminate: object with references, ref_count=%d",
656 		object->ref_count));
657 
658 	/*
659 	 * Now free any remaining pages. For internal objects, this also
660 	 * removes them from paging queues. Don't free wired pages, just
661 	 * remove them from the object.
662 	 */
663 	vm_page_lock_queues();
664 	while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
665 		KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
666 			("vm_object_terminate: freeing busy page %p "
667 			"p->busy = %d, p->flags %x\n", p, p->busy, p->flags));
668 		if (p->wire_count == 0) {
669 			vm_page_free(p);
670 			cnt.v_pfree++;
671 		} else {
672 			vm_page_remove(p);
673 		}
674 	}
675 	vm_page_unlock_queues();
676 
677 #if VM_NRESERVLEVEL > 0
678 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
679 		vm_reserv_break_all(object);
680 #endif
681 	if (__predict_false(object->cache != NULL))
682 		vm_page_cache_free(object, 0, 0);
683 
684 	/*
685 	 * Let the pager know object is dead.
686 	 */
687 	vm_pager_deallocate(object);
688 	VM_OBJECT_UNLOCK(object);
689 
690 	/*
691 	 * Remove the object from the global object list.
692 	 */
693 	mtx_lock(&vm_object_list_mtx);
694 	TAILQ_REMOVE(&vm_object_list, object, object_list);
695 	mtx_unlock(&vm_object_list_mtx);
696 
697 	/*
698 	 * Free the space for the object.
699 	 */
700 	uma_zfree(obj_zone, object);
701 }
702 
703 /*
704  *	vm_object_page_clean
705  *
706  *	Clean all dirty pages in the specified range of object.  Leaves page
707  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
708  *	write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
709  *	leaving the object dirty.
710  *
711  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
712  *	synchronous clustering mode implementation.
713  *
714  *	Odd semantics: if start == end, we clean everything.
715  *
716  *	The object must be locked.
717  */
718 void
719 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags)
720 {
721 	vm_page_t p, np;
722 	vm_pindex_t tstart, tend;
723 	vm_pindex_t pi;
724 	int clearobjflags;
725 	int pagerflags;
726 	int curgeneration;
727 
728 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
729 	if (object->type != OBJT_VNODE ||
730 		(object->flags & OBJ_MIGHTBEDIRTY) == 0)
731 		return;
732 
733 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
734 	pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
735 
736 	vm_object_set_flag(object, OBJ_CLEANING);
737 
738 	tstart = start;
739 	if (end == 0) {
740 		tend = object->size;
741 	} else {
742 		tend = end;
743 	}
744 
745 	vm_page_lock_queues();
746 	/*
747 	 * If the caller is smart and only msync()s a range he knows is
748 	 * dirty, we may be able to avoid an object scan.  This results in
749 	 * a phenominal improvement in performance.  We cannot do this
750 	 * as a matter of course because the object may be huge - e.g.
751 	 * the size might be in the gigabytes or terrabytes.
752 	 */
753 	if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
754 		vm_pindex_t tscan;
755 		int scanlimit;
756 		int scanreset;
757 
758 		scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
759 		if (scanreset < 16)
760 			scanreset = 16;
761 		pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
762 
763 		scanlimit = scanreset;
764 		tscan = tstart;
765 		while (tscan < tend) {
766 			curgeneration = object->generation;
767 			p = vm_page_lookup(object, tscan);
768 			if (p == NULL || p->valid == 0) {
769 				if (--scanlimit == 0)
770 					break;
771 				++tscan;
772 				continue;
773 			}
774 			vm_page_test_dirty(p);
775 			if ((p->dirty & p->valid) == 0) {
776 				if (--scanlimit == 0)
777 					break;
778 				++tscan;
779 				continue;
780 			}
781 			/*
782 			 * If we have been asked to skip nosync pages and
783 			 * this is a nosync page, we can't continue.
784 			 */
785 			if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
786 				if (--scanlimit == 0)
787 					break;
788 				++tscan;
789 				continue;
790 			}
791 			scanlimit = scanreset;
792 
793 			/*
794 			 * This returns 0 if it was unable to busy the first
795 			 * page (i.e. had to sleep).
796 			 */
797 			tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
798 		}
799 
800 		/*
801 		 * If everything was dirty and we flushed it successfully,
802 		 * and the requested range is not the entire object, we
803 		 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
804 		 * return immediately.
805 		 */
806 		if (tscan >= tend && (tstart || tend < object->size)) {
807 			vm_page_unlock_queues();
808 			vm_object_clear_flag(object, OBJ_CLEANING);
809 			return;
810 		}
811 		pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
812 	}
813 
814 	/*
815 	 * Generally set CLEANCHK interlock and make the page read-only so
816 	 * we can then clear the object flags.
817 	 *
818 	 * However, if this is a nosync mmap then the object is likely to
819 	 * stay dirty so do not mess with the page and do not clear the
820 	 * object flags.
821 	 */
822 	clearobjflags = 1;
823 	TAILQ_FOREACH(p, &object->memq, listq) {
824 		p->oflags |= VPO_CLEANCHK;
825 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC))
826 			clearobjflags = 0;
827 		else
828 			pmap_remove_write(p);
829 	}
830 
831 	if (clearobjflags && (tstart == 0) && (tend == object->size)) {
832 		struct vnode *vp;
833 
834 		vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
835 		if (object->type == OBJT_VNODE &&
836 		    (vp = (struct vnode *)object->handle) != NULL) {
837 			VI_LOCK(vp);
838 			if (vp->v_iflag & VI_OBJDIRTY)
839 				vp->v_iflag &= ~VI_OBJDIRTY;
840 			VI_UNLOCK(vp);
841 		}
842 	}
843 
844 rescan:
845 	curgeneration = object->generation;
846 
847 	for (p = TAILQ_FIRST(&object->memq); p; p = np) {
848 		int n;
849 
850 		np = TAILQ_NEXT(p, listq);
851 
852 again:
853 		pi = p->pindex;
854 		if ((p->oflags & VPO_CLEANCHK) == 0 ||
855 			(pi < tstart) || (pi >= tend) ||
856 		    p->valid == 0) {
857 			p->oflags &= ~VPO_CLEANCHK;
858 			continue;
859 		}
860 
861 		vm_page_test_dirty(p);
862 		if ((p->dirty & p->valid) == 0) {
863 			p->oflags &= ~VPO_CLEANCHK;
864 			continue;
865 		}
866 
867 		/*
868 		 * If we have been asked to skip nosync pages and this is a
869 		 * nosync page, skip it.  Note that the object flags were
870 		 * not cleared in this case so we do not have to set them.
871 		 */
872 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
873 			p->oflags &= ~VPO_CLEANCHK;
874 			continue;
875 		}
876 
877 		n = vm_object_page_collect_flush(object, p,
878 			curgeneration, pagerflags);
879 		if (n == 0)
880 			goto rescan;
881 
882 		if (object->generation != curgeneration)
883 			goto rescan;
884 
885 		/*
886 		 * Try to optimize the next page.  If we can't we pick up
887 		 * our (random) scan where we left off.
888 		 */
889 		if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
890 			if ((p = vm_page_lookup(object, pi + n)) != NULL)
891 				goto again;
892 		}
893 	}
894 	vm_page_unlock_queues();
895 #if 0
896 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
897 #endif
898 
899 	vm_object_clear_flag(object, OBJ_CLEANING);
900 	return;
901 }
902 
903 static int
904 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
905 {
906 	int runlen;
907 	int maxf;
908 	int chkb;
909 	int maxb;
910 	int i;
911 	vm_pindex_t pi;
912 	vm_page_t maf[vm_pageout_page_count];
913 	vm_page_t mab[vm_pageout_page_count];
914 	vm_page_t ma[vm_pageout_page_count];
915 
916 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
917 	pi = p->pindex;
918 	while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
919 		vm_page_lock_queues();
920 		if (object->generation != curgeneration) {
921 			return(0);
922 		}
923 	}
924 	maxf = 0;
925 	for(i = 1; i < vm_pageout_page_count; i++) {
926 		vm_page_t tp;
927 
928 		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
929 			if ((tp->oflags & VPO_BUSY) ||
930 				((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
931 				 (tp->oflags & VPO_CLEANCHK) == 0) ||
932 				(tp->busy != 0))
933 				break;
934 			vm_page_test_dirty(tp);
935 			if ((tp->dirty & tp->valid) == 0) {
936 				tp->oflags &= ~VPO_CLEANCHK;
937 				break;
938 			}
939 			maf[ i - 1 ] = tp;
940 			maxf++;
941 			continue;
942 		}
943 		break;
944 	}
945 
946 	maxb = 0;
947 	chkb = vm_pageout_page_count -  maxf;
948 	if (chkb) {
949 		for(i = 1; i < chkb;i++) {
950 			vm_page_t tp;
951 
952 			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
953 				if ((tp->oflags & VPO_BUSY) ||
954 					((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
955 					 (tp->oflags & VPO_CLEANCHK) == 0) ||
956 					(tp->busy != 0))
957 					break;
958 				vm_page_test_dirty(tp);
959 				if ((tp->dirty & tp->valid) == 0) {
960 					tp->oflags &= ~VPO_CLEANCHK;
961 					break;
962 				}
963 				mab[ i - 1 ] = tp;
964 				maxb++;
965 				continue;
966 			}
967 			break;
968 		}
969 	}
970 
971 	for(i = 0; i < maxb; i++) {
972 		int index = (maxb - i) - 1;
973 		ma[index] = mab[i];
974 		ma[index]->oflags &= ~VPO_CLEANCHK;
975 	}
976 	p->oflags &= ~VPO_CLEANCHK;
977 	ma[maxb] = p;
978 	for(i = 0; i < maxf; i++) {
979 		int index = (maxb + i) + 1;
980 		ma[index] = maf[i];
981 		ma[index]->oflags &= ~VPO_CLEANCHK;
982 	}
983 	runlen = maxb + maxf + 1;
984 
985 	vm_pageout_flush(ma, runlen, pagerflags);
986 	for (i = 0; i < runlen; i++) {
987 		if (ma[i]->valid & ma[i]->dirty) {
988 			pmap_remove_write(ma[i]);
989 			ma[i]->oflags |= VPO_CLEANCHK;
990 
991 			/*
992 			 * maxf will end up being the actual number of pages
993 			 * we wrote out contiguously, non-inclusive of the
994 			 * first page.  We do not count look-behind pages.
995 			 */
996 			if (i >= maxb + 1 && (maxf > i - maxb - 1))
997 				maxf = i - maxb - 1;
998 		}
999 	}
1000 	return(maxf + 1);
1001 }
1002 
1003 /*
1004  * Note that there is absolutely no sense in writing out
1005  * anonymous objects, so we track down the vnode object
1006  * to write out.
1007  * We invalidate (remove) all pages from the address space
1008  * for semantic correctness.
1009  *
1010  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1011  * may start out with a NULL object.
1012  */
1013 void
1014 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1015     boolean_t syncio, boolean_t invalidate)
1016 {
1017 	vm_object_t backing_object;
1018 	struct vnode *vp;
1019 	struct mount *mp;
1020 	int flags;
1021 
1022 	if (object == NULL)
1023 		return;
1024 	VM_OBJECT_LOCK(object);
1025 	while ((backing_object = object->backing_object) != NULL) {
1026 		VM_OBJECT_LOCK(backing_object);
1027 		offset += object->backing_object_offset;
1028 		VM_OBJECT_UNLOCK(object);
1029 		object = backing_object;
1030 		if (object->size < OFF_TO_IDX(offset + size))
1031 			size = IDX_TO_OFF(object->size) - offset;
1032 	}
1033 	/*
1034 	 * Flush pages if writing is allowed, invalidate them
1035 	 * if invalidation requested.  Pages undergoing I/O
1036 	 * will be ignored by vm_object_page_remove().
1037 	 *
1038 	 * We cannot lock the vnode and then wait for paging
1039 	 * to complete without deadlocking against vm_fault.
1040 	 * Instead we simply call vm_object_page_remove() and
1041 	 * allow it to block internally on a page-by-page
1042 	 * basis when it encounters pages undergoing async
1043 	 * I/O.
1044 	 */
1045 	if (object->type == OBJT_VNODE &&
1046 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1047 		int vfslocked;
1048 		vp = object->handle;
1049 		VM_OBJECT_UNLOCK(object);
1050 		(void) vn_start_write(vp, &mp, V_WAIT);
1051 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1052 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1053 		flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1054 		flags |= invalidate ? OBJPC_INVAL : 0;
1055 		VM_OBJECT_LOCK(object);
1056 		vm_object_page_clean(object,
1057 		    OFF_TO_IDX(offset),
1058 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1059 		    flags);
1060 		VM_OBJECT_UNLOCK(object);
1061 		VOP_UNLOCK(vp, 0);
1062 		VFS_UNLOCK_GIANT(vfslocked);
1063 		vn_finished_write(mp);
1064 		VM_OBJECT_LOCK(object);
1065 	}
1066 	if ((object->type == OBJT_VNODE ||
1067 	     object->type == OBJT_DEVICE) && invalidate) {
1068 		boolean_t purge;
1069 		purge = old_msync || (object->type == OBJT_DEVICE);
1070 		vm_object_page_remove(object,
1071 		    OFF_TO_IDX(offset),
1072 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1073 		    purge ? FALSE : TRUE);
1074 	}
1075 	VM_OBJECT_UNLOCK(object);
1076 }
1077 
1078 /*
1079  *	vm_object_madvise:
1080  *
1081  *	Implements the madvise function at the object/page level.
1082  *
1083  *	MADV_WILLNEED	(any object)
1084  *
1085  *	    Activate the specified pages if they are resident.
1086  *
1087  *	MADV_DONTNEED	(any object)
1088  *
1089  *	    Deactivate the specified pages if they are resident.
1090  *
1091  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1092  *			 OBJ_ONEMAPPING only)
1093  *
1094  *	    Deactivate and clean the specified pages if they are
1095  *	    resident.  This permits the process to reuse the pages
1096  *	    without faulting or the kernel to reclaim the pages
1097  *	    without I/O.
1098  */
1099 void
1100 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1101 {
1102 	vm_pindex_t end, tpindex;
1103 	vm_object_t backing_object, tobject;
1104 	vm_page_t m;
1105 
1106 	if (object == NULL)
1107 		return;
1108 	VM_OBJECT_LOCK(object);
1109 	end = pindex + count;
1110 	/*
1111 	 * Locate and adjust resident pages
1112 	 */
1113 	for (; pindex < end; pindex += 1) {
1114 relookup:
1115 		tobject = object;
1116 		tpindex = pindex;
1117 shadowlookup:
1118 		/*
1119 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1120 		 * and those pages must be OBJ_ONEMAPPING.
1121 		 */
1122 		if (advise == MADV_FREE) {
1123 			if ((tobject->type != OBJT_DEFAULT &&
1124 			     tobject->type != OBJT_SWAP) ||
1125 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
1126 				goto unlock_tobject;
1127 			}
1128 		}
1129 		m = vm_page_lookup(tobject, tpindex);
1130 		if (m == NULL && advise == MADV_WILLNEED) {
1131 			/*
1132 			 * If the page is cached, reactivate it.
1133 			 */
1134 			m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1135 			    VM_ALLOC_NOBUSY);
1136 		}
1137 		if (m == NULL) {
1138 			/*
1139 			 * There may be swap even if there is no backing page
1140 			 */
1141 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1142 				swap_pager_freespace(tobject, tpindex, 1);
1143 			/*
1144 			 * next object
1145 			 */
1146 			backing_object = tobject->backing_object;
1147 			if (backing_object == NULL)
1148 				goto unlock_tobject;
1149 			VM_OBJECT_LOCK(backing_object);
1150 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1151 			if (tobject != object)
1152 				VM_OBJECT_UNLOCK(tobject);
1153 			tobject = backing_object;
1154 			goto shadowlookup;
1155 		}
1156 		/*
1157 		 * If the page is busy or not in a normal active state,
1158 		 * we skip it.  If the page is not managed there are no
1159 		 * page queues to mess with.  Things can break if we mess
1160 		 * with pages in any of the below states.
1161 		 */
1162 		vm_page_lock_queues();
1163 		if (m->hold_count ||
1164 		    m->wire_count ||
1165 		    (m->flags & PG_UNMANAGED) ||
1166 		    m->valid != VM_PAGE_BITS_ALL) {
1167 			vm_page_unlock_queues();
1168 			goto unlock_tobject;
1169 		}
1170 		if ((m->oflags & VPO_BUSY) || m->busy) {
1171 			vm_page_flag_set(m, PG_REFERENCED);
1172 			vm_page_unlock_queues();
1173 			if (object != tobject)
1174 				VM_OBJECT_UNLOCK(object);
1175 			m->oflags |= VPO_WANTED;
1176 			msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo", 0);
1177 			VM_OBJECT_LOCK(object);
1178   			goto relookup;
1179 		}
1180 		if (advise == MADV_WILLNEED) {
1181 			vm_page_activate(m);
1182 		} else if (advise == MADV_DONTNEED) {
1183 			vm_page_dontneed(m);
1184 		} else if (advise == MADV_FREE) {
1185 			/*
1186 			 * Mark the page clean.  This will allow the page
1187 			 * to be freed up by the system.  However, such pages
1188 			 * are often reused quickly by malloc()/free()
1189 			 * so we do not do anything that would cause
1190 			 * a page fault if we can help it.
1191 			 *
1192 			 * Specifically, we do not try to actually free
1193 			 * the page now nor do we try to put it in the
1194 			 * cache (which would cause a page fault on reuse).
1195 			 *
1196 			 * But we do make the page is freeable as we
1197 			 * can without actually taking the step of unmapping
1198 			 * it.
1199 			 */
1200 			pmap_clear_modify(m);
1201 			m->dirty = 0;
1202 			m->act_count = 0;
1203 			vm_page_dontneed(m);
1204 		}
1205 		vm_page_unlock_queues();
1206 		if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1207 			swap_pager_freespace(tobject, tpindex, 1);
1208 unlock_tobject:
1209 		if (tobject != object)
1210 			VM_OBJECT_UNLOCK(tobject);
1211 	}
1212 	VM_OBJECT_UNLOCK(object);
1213 }
1214 
1215 /*
1216  *	vm_object_shadow:
1217  *
1218  *	Create a new object which is backed by the
1219  *	specified existing object range.  The source
1220  *	object reference is deallocated.
1221  *
1222  *	The new object and offset into that object
1223  *	are returned in the source parameters.
1224  */
1225 void
1226 vm_object_shadow(
1227 	vm_object_t *object,	/* IN/OUT */
1228 	vm_ooffset_t *offset,	/* IN/OUT */
1229 	vm_size_t length)
1230 {
1231 	vm_object_t source;
1232 	vm_object_t result;
1233 
1234 	source = *object;
1235 
1236 	/*
1237 	 * Don't create the new object if the old object isn't shared.
1238 	 */
1239 	if (source != NULL) {
1240 		VM_OBJECT_LOCK(source);
1241 		if (source->ref_count == 1 &&
1242 		    source->handle == NULL &&
1243 		    (source->type == OBJT_DEFAULT ||
1244 		     source->type == OBJT_SWAP)) {
1245 			VM_OBJECT_UNLOCK(source);
1246 			return;
1247 		}
1248 		VM_OBJECT_UNLOCK(source);
1249 	}
1250 
1251 	/*
1252 	 * Allocate a new object with the given length.
1253 	 */
1254 	result = vm_object_allocate(OBJT_DEFAULT, length);
1255 
1256 	/*
1257 	 * The new object shadows the source object, adding a reference to it.
1258 	 * Our caller changes his reference to point to the new object,
1259 	 * removing a reference to the source object.  Net result: no change
1260 	 * of reference count.
1261 	 *
1262 	 * Try to optimize the result object's page color when shadowing
1263 	 * in order to maintain page coloring consistency in the combined
1264 	 * shadowed object.
1265 	 */
1266 	result->backing_object = source;
1267 	/*
1268 	 * Store the offset into the source object, and fix up the offset into
1269 	 * the new object.
1270 	 */
1271 	result->backing_object_offset = *offset;
1272 	if (source != NULL) {
1273 		VM_OBJECT_LOCK(source);
1274 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1275 		source->shadow_count++;
1276 		source->generation++;
1277 #if VM_NRESERVLEVEL > 0
1278 		result->flags |= source->flags & (OBJ_NEEDGIANT | OBJ_COLORED);
1279 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1280 		    ((1 << (VM_NFREEORDER - 1)) - 1);
1281 #else
1282 		result->flags |= source->flags & OBJ_NEEDGIANT;
1283 #endif
1284 		VM_OBJECT_UNLOCK(source);
1285 	}
1286 
1287 
1288 	/*
1289 	 * Return the new things
1290 	 */
1291 	*offset = 0;
1292 	*object = result;
1293 }
1294 
1295 /*
1296  *	vm_object_split:
1297  *
1298  * Split the pages in a map entry into a new object.  This affords
1299  * easier removal of unused pages, and keeps object inheritance from
1300  * being a negative impact on memory usage.
1301  */
1302 void
1303 vm_object_split(vm_map_entry_t entry)
1304 {
1305 	vm_page_t m, m_next;
1306 	vm_object_t orig_object, new_object, source;
1307 	vm_pindex_t idx, offidxstart;
1308 	vm_size_t size;
1309 
1310 	orig_object = entry->object.vm_object;
1311 	if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1312 		return;
1313 	if (orig_object->ref_count <= 1)
1314 		return;
1315 	VM_OBJECT_UNLOCK(orig_object);
1316 
1317 	offidxstart = OFF_TO_IDX(entry->offset);
1318 	size = atop(entry->end - entry->start);
1319 
1320 	/*
1321 	 * If swap_pager_copy() is later called, it will convert new_object
1322 	 * into a swap object.
1323 	 */
1324 	new_object = vm_object_allocate(OBJT_DEFAULT, size);
1325 
1326 	/*
1327 	 * At this point, the new object is still private, so the order in
1328 	 * which the original and new objects are locked does not matter.
1329 	 */
1330 	VM_OBJECT_LOCK(new_object);
1331 	VM_OBJECT_LOCK(orig_object);
1332 	source = orig_object->backing_object;
1333 	if (source != NULL) {
1334 		VM_OBJECT_LOCK(source);
1335 		if ((source->flags & OBJ_DEAD) != 0) {
1336 			VM_OBJECT_UNLOCK(source);
1337 			VM_OBJECT_UNLOCK(orig_object);
1338 			VM_OBJECT_UNLOCK(new_object);
1339 			vm_object_deallocate(new_object);
1340 			VM_OBJECT_LOCK(orig_object);
1341 			return;
1342 		}
1343 		LIST_INSERT_HEAD(&source->shadow_head,
1344 				  new_object, shadow_list);
1345 		source->shadow_count++;
1346 		source->generation++;
1347 		vm_object_reference_locked(source);	/* for new_object */
1348 		vm_object_clear_flag(source, OBJ_ONEMAPPING);
1349 		VM_OBJECT_UNLOCK(source);
1350 		new_object->backing_object_offset =
1351 			orig_object->backing_object_offset + entry->offset;
1352 		new_object->backing_object = source;
1353 	}
1354 	new_object->flags |= orig_object->flags & OBJ_NEEDGIANT;
1355 retry:
1356 	if ((m = TAILQ_FIRST(&orig_object->memq)) != NULL) {
1357 		if (m->pindex < offidxstart) {
1358 			m = vm_page_splay(offidxstart, orig_object->root);
1359 			if ((orig_object->root = m)->pindex < offidxstart)
1360 				m = TAILQ_NEXT(m, listq);
1361 		}
1362 	}
1363 	vm_page_lock_queues();
1364 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1365 	    m = m_next) {
1366 		m_next = TAILQ_NEXT(m, listq);
1367 
1368 		/*
1369 		 * We must wait for pending I/O to complete before we can
1370 		 * rename the page.
1371 		 *
1372 		 * We do not have to VM_PROT_NONE the page as mappings should
1373 		 * not be changed by this operation.
1374 		 */
1375 		if ((m->oflags & VPO_BUSY) || m->busy) {
1376 			vm_page_flag_set(m, PG_REFERENCED);
1377 			vm_page_unlock_queues();
1378 			VM_OBJECT_UNLOCK(new_object);
1379 			m->oflags |= VPO_WANTED;
1380 			msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1381 			VM_OBJECT_LOCK(new_object);
1382 			goto retry;
1383 		}
1384 		vm_page_rename(m, new_object, idx);
1385 		/* page automatically made dirty by rename and cache handled */
1386 		vm_page_busy(m);
1387 	}
1388 	vm_page_unlock_queues();
1389 	if (orig_object->type == OBJT_SWAP) {
1390 		/*
1391 		 * swap_pager_copy() can sleep, in which case the orig_object's
1392 		 * and new_object's locks are released and reacquired.
1393 		 */
1394 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1395 
1396 		/*
1397 		 * Transfer any cached pages from orig_object to new_object.
1398 		 */
1399 		if (__predict_false(orig_object->cache != NULL))
1400 			vm_page_cache_transfer(orig_object, offidxstart,
1401 			    new_object);
1402 	}
1403 	VM_OBJECT_UNLOCK(orig_object);
1404 	TAILQ_FOREACH(m, &new_object->memq, listq)
1405 		vm_page_wakeup(m);
1406 	VM_OBJECT_UNLOCK(new_object);
1407 	entry->object.vm_object = new_object;
1408 	entry->offset = 0LL;
1409 	vm_object_deallocate(orig_object);
1410 	VM_OBJECT_LOCK(new_object);
1411 }
1412 
1413 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1414 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1415 #define	OBSC_COLLAPSE_WAIT	0x0004
1416 
1417 static int
1418 vm_object_backing_scan(vm_object_t object, int op)
1419 {
1420 	int r = 1;
1421 	vm_page_t p;
1422 	vm_object_t backing_object;
1423 	vm_pindex_t backing_offset_index;
1424 
1425 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1426 	VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1427 
1428 	backing_object = object->backing_object;
1429 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1430 
1431 	/*
1432 	 * Initial conditions
1433 	 */
1434 	if (op & OBSC_TEST_ALL_SHADOWED) {
1435 		/*
1436 		 * We do not want to have to test for the existence of cache
1437 		 * or swap pages in the backing object.  XXX but with the
1438 		 * new swapper this would be pretty easy to do.
1439 		 *
1440 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1441 		 * been ZFOD faulted yet?  If we do not test for this, the
1442 		 * shadow test may succeed! XXX
1443 		 */
1444 		if (backing_object->type != OBJT_DEFAULT) {
1445 			return (0);
1446 		}
1447 	}
1448 	if (op & OBSC_COLLAPSE_WAIT) {
1449 		vm_object_set_flag(backing_object, OBJ_DEAD);
1450 	}
1451 
1452 	/*
1453 	 * Our scan
1454 	 */
1455 	p = TAILQ_FIRST(&backing_object->memq);
1456 	while (p) {
1457 		vm_page_t next = TAILQ_NEXT(p, listq);
1458 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1459 
1460 		if (op & OBSC_TEST_ALL_SHADOWED) {
1461 			vm_page_t pp;
1462 
1463 			/*
1464 			 * Ignore pages outside the parent object's range
1465 			 * and outside the parent object's mapping of the
1466 			 * backing object.
1467 			 *
1468 			 * note that we do not busy the backing object's
1469 			 * page.
1470 			 */
1471 			if (
1472 			    p->pindex < backing_offset_index ||
1473 			    new_pindex >= object->size
1474 			) {
1475 				p = next;
1476 				continue;
1477 			}
1478 
1479 			/*
1480 			 * See if the parent has the page or if the parent's
1481 			 * object pager has the page.  If the parent has the
1482 			 * page but the page is not valid, the parent's
1483 			 * object pager must have the page.
1484 			 *
1485 			 * If this fails, the parent does not completely shadow
1486 			 * the object and we might as well give up now.
1487 			 */
1488 
1489 			pp = vm_page_lookup(object, new_pindex);
1490 			if (
1491 			    (pp == NULL || pp->valid == 0) &&
1492 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1493 			) {
1494 				r = 0;
1495 				break;
1496 			}
1497 		}
1498 
1499 		/*
1500 		 * Check for busy page
1501 		 */
1502 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1503 			vm_page_t pp;
1504 
1505 			if (op & OBSC_COLLAPSE_NOWAIT) {
1506 				if ((p->oflags & VPO_BUSY) ||
1507 				    !p->valid ||
1508 				    p->busy) {
1509 					p = next;
1510 					continue;
1511 				}
1512 			} else if (op & OBSC_COLLAPSE_WAIT) {
1513 				if ((p->oflags & VPO_BUSY) || p->busy) {
1514 					vm_page_lock_queues();
1515 					vm_page_flag_set(p, PG_REFERENCED);
1516 					vm_page_unlock_queues();
1517 					VM_OBJECT_UNLOCK(object);
1518 					p->oflags |= VPO_WANTED;
1519 					msleep(p, VM_OBJECT_MTX(backing_object),
1520 					    PDROP | PVM, "vmocol", 0);
1521 					VM_OBJECT_LOCK(object);
1522 					VM_OBJECT_LOCK(backing_object);
1523 					/*
1524 					 * If we slept, anything could have
1525 					 * happened.  Since the object is
1526 					 * marked dead, the backing offset
1527 					 * should not have changed so we
1528 					 * just restart our scan.
1529 					 */
1530 					p = TAILQ_FIRST(&backing_object->memq);
1531 					continue;
1532 				}
1533 			}
1534 
1535 			KASSERT(
1536 			    p->object == backing_object,
1537 			    ("vm_object_backing_scan: object mismatch")
1538 			);
1539 
1540 			/*
1541 			 * Destroy any associated swap
1542 			 */
1543 			if (backing_object->type == OBJT_SWAP) {
1544 				swap_pager_freespace(
1545 				    backing_object,
1546 				    p->pindex,
1547 				    1
1548 				);
1549 			}
1550 
1551 			if (
1552 			    p->pindex < backing_offset_index ||
1553 			    new_pindex >= object->size
1554 			) {
1555 				/*
1556 				 * Page is out of the parent object's range, we
1557 				 * can simply destroy it.
1558 				 */
1559 				vm_page_lock_queues();
1560 				KASSERT(!pmap_page_is_mapped(p),
1561 				    ("freeing mapped page %p", p));
1562 				if (p->wire_count == 0)
1563 					vm_page_free(p);
1564 				else
1565 					vm_page_remove(p);
1566 				vm_page_unlock_queues();
1567 				p = next;
1568 				continue;
1569 			}
1570 
1571 			pp = vm_page_lookup(object, new_pindex);
1572 			if (
1573 			    pp != NULL ||
1574 			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1575 			) {
1576 				/*
1577 				 * page already exists in parent OR swap exists
1578 				 * for this location in the parent.  Destroy
1579 				 * the original page from the backing object.
1580 				 *
1581 				 * Leave the parent's page alone
1582 				 */
1583 				vm_page_lock_queues();
1584 				KASSERT(!pmap_page_is_mapped(p),
1585 				    ("freeing mapped page %p", p));
1586 				if (p->wire_count == 0)
1587 					vm_page_free(p);
1588 				else
1589 					vm_page_remove(p);
1590 				vm_page_unlock_queues();
1591 				p = next;
1592 				continue;
1593 			}
1594 
1595 #if VM_NRESERVLEVEL > 0
1596 			/*
1597 			 * Rename the reservation.
1598 			 */
1599 			vm_reserv_rename(p, object, backing_object,
1600 			    backing_offset_index);
1601 #endif
1602 
1603 			/*
1604 			 * Page does not exist in parent, rename the
1605 			 * page from the backing object to the main object.
1606 			 *
1607 			 * If the page was mapped to a process, it can remain
1608 			 * mapped through the rename.
1609 			 */
1610 			vm_page_lock_queues();
1611 			vm_page_rename(p, object, new_pindex);
1612 			vm_page_unlock_queues();
1613 			/* page automatically made dirty by rename */
1614 		}
1615 		p = next;
1616 	}
1617 	return (r);
1618 }
1619 
1620 
1621 /*
1622  * this version of collapse allows the operation to occur earlier and
1623  * when paging_in_progress is true for an object...  This is not a complete
1624  * operation, but should plug 99.9% of the rest of the leaks.
1625  */
1626 static void
1627 vm_object_qcollapse(vm_object_t object)
1628 {
1629 	vm_object_t backing_object = object->backing_object;
1630 
1631 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1632 	VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1633 
1634 	if (backing_object->ref_count != 1)
1635 		return;
1636 
1637 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1638 }
1639 
1640 /*
1641  *	vm_object_collapse:
1642  *
1643  *	Collapse an object with the object backing it.
1644  *	Pages in the backing object are moved into the
1645  *	parent, and the backing object is deallocated.
1646  */
1647 void
1648 vm_object_collapse(vm_object_t object)
1649 {
1650 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1651 
1652 	while (TRUE) {
1653 		vm_object_t backing_object;
1654 
1655 		/*
1656 		 * Verify that the conditions are right for collapse:
1657 		 *
1658 		 * The object exists and the backing object exists.
1659 		 */
1660 		if ((backing_object = object->backing_object) == NULL)
1661 			break;
1662 
1663 		/*
1664 		 * we check the backing object first, because it is most likely
1665 		 * not collapsable.
1666 		 */
1667 		VM_OBJECT_LOCK(backing_object);
1668 		if (backing_object->handle != NULL ||
1669 		    (backing_object->type != OBJT_DEFAULT &&
1670 		     backing_object->type != OBJT_SWAP) ||
1671 		    (backing_object->flags & OBJ_DEAD) ||
1672 		    object->handle != NULL ||
1673 		    (object->type != OBJT_DEFAULT &&
1674 		     object->type != OBJT_SWAP) ||
1675 		    (object->flags & OBJ_DEAD)) {
1676 			VM_OBJECT_UNLOCK(backing_object);
1677 			break;
1678 		}
1679 
1680 		if (
1681 		    object->paging_in_progress != 0 ||
1682 		    backing_object->paging_in_progress != 0
1683 		) {
1684 			vm_object_qcollapse(object);
1685 			VM_OBJECT_UNLOCK(backing_object);
1686 			break;
1687 		}
1688 		/*
1689 		 * We know that we can either collapse the backing object (if
1690 		 * the parent is the only reference to it) or (perhaps) have
1691 		 * the parent bypass the object if the parent happens to shadow
1692 		 * all the resident pages in the entire backing object.
1693 		 *
1694 		 * This is ignoring pager-backed pages such as swap pages.
1695 		 * vm_object_backing_scan fails the shadowing test in this
1696 		 * case.
1697 		 */
1698 		if (backing_object->ref_count == 1) {
1699 			/*
1700 			 * If there is exactly one reference to the backing
1701 			 * object, we can collapse it into the parent.
1702 			 */
1703 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1704 
1705 #if VM_NRESERVLEVEL > 0
1706 			/*
1707 			 * Break any reservations from backing_object.
1708 			 */
1709 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1710 				vm_reserv_break_all(backing_object);
1711 #endif
1712 
1713 			/*
1714 			 * Move the pager from backing_object to object.
1715 			 */
1716 			if (backing_object->type == OBJT_SWAP) {
1717 				/*
1718 				 * swap_pager_copy() can sleep, in which case
1719 				 * the backing_object's and object's locks are
1720 				 * released and reacquired.
1721 				 */
1722 				swap_pager_copy(
1723 				    backing_object,
1724 				    object,
1725 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1726 
1727 				/*
1728 				 * Free any cached pages from backing_object.
1729 				 */
1730 				if (__predict_false(backing_object->cache != NULL))
1731 					vm_page_cache_free(backing_object, 0, 0);
1732 			}
1733 			/*
1734 			 * Object now shadows whatever backing_object did.
1735 			 * Note that the reference to
1736 			 * backing_object->backing_object moves from within
1737 			 * backing_object to within object.
1738 			 */
1739 			LIST_REMOVE(object, shadow_list);
1740 			backing_object->shadow_count--;
1741 			backing_object->generation++;
1742 			if (backing_object->backing_object) {
1743 				VM_OBJECT_LOCK(backing_object->backing_object);
1744 				LIST_REMOVE(backing_object, shadow_list);
1745 				LIST_INSERT_HEAD(
1746 				    &backing_object->backing_object->shadow_head,
1747 				    object, shadow_list);
1748 				/*
1749 				 * The shadow_count has not changed.
1750 				 */
1751 				backing_object->backing_object->generation++;
1752 				VM_OBJECT_UNLOCK(backing_object->backing_object);
1753 			}
1754 			object->backing_object = backing_object->backing_object;
1755 			object->backing_object_offset +=
1756 			    backing_object->backing_object_offset;
1757 
1758 			/*
1759 			 * Discard backing_object.
1760 			 *
1761 			 * Since the backing object has no pages, no pager left,
1762 			 * and no object references within it, all that is
1763 			 * necessary is to dispose of it.
1764 			 */
1765 			KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1766 			VM_OBJECT_UNLOCK(backing_object);
1767 
1768 			mtx_lock(&vm_object_list_mtx);
1769 			TAILQ_REMOVE(
1770 			    &vm_object_list,
1771 			    backing_object,
1772 			    object_list
1773 			);
1774 			mtx_unlock(&vm_object_list_mtx);
1775 
1776 			uma_zfree(obj_zone, backing_object);
1777 
1778 			object_collapses++;
1779 		} else {
1780 			vm_object_t new_backing_object;
1781 
1782 			/*
1783 			 * If we do not entirely shadow the backing object,
1784 			 * there is nothing we can do so we give up.
1785 			 */
1786 			if (object->resident_page_count != object->size &&
1787 			    vm_object_backing_scan(object,
1788 			    OBSC_TEST_ALL_SHADOWED) == 0) {
1789 				VM_OBJECT_UNLOCK(backing_object);
1790 				break;
1791 			}
1792 
1793 			/*
1794 			 * Make the parent shadow the next object in the
1795 			 * chain.  Deallocating backing_object will not remove
1796 			 * it, since its reference count is at least 2.
1797 			 */
1798 			LIST_REMOVE(object, shadow_list);
1799 			backing_object->shadow_count--;
1800 			backing_object->generation++;
1801 
1802 			new_backing_object = backing_object->backing_object;
1803 			if ((object->backing_object = new_backing_object) != NULL) {
1804 				VM_OBJECT_LOCK(new_backing_object);
1805 				LIST_INSERT_HEAD(
1806 				    &new_backing_object->shadow_head,
1807 				    object,
1808 				    shadow_list
1809 				);
1810 				new_backing_object->shadow_count++;
1811 				new_backing_object->generation++;
1812 				vm_object_reference_locked(new_backing_object);
1813 				VM_OBJECT_UNLOCK(new_backing_object);
1814 				object->backing_object_offset +=
1815 					backing_object->backing_object_offset;
1816 			}
1817 
1818 			/*
1819 			 * Drop the reference count on backing_object. Since
1820 			 * its ref_count was at least 2, it will not vanish.
1821 			 */
1822 			backing_object->ref_count--;
1823 			VM_OBJECT_UNLOCK(backing_object);
1824 			object_bypasses++;
1825 		}
1826 
1827 		/*
1828 		 * Try again with this object's new backing object.
1829 		 */
1830 	}
1831 }
1832 
1833 /*
1834  *	vm_object_page_remove:
1835  *
1836  *	For the given object, either frees or invalidates each of the
1837  *	specified pages.  In general, a page is freed.  However, if a
1838  *	page is wired for any reason other than the existence of a
1839  *	managed, wired mapping, then it may be invalidated but not
1840  *	removed from the object.  Pages are specified by the given
1841  *	range ["start", "end") and Boolean "clean_only".  As a
1842  *	special case, if "end" is zero, then the range extends from
1843  *	"start" to the end of the object.  If "clean_only" is TRUE,
1844  *	then only the non-dirty pages within the specified range are
1845  *	affected.
1846  *
1847  *	In general, this operation should only be performed on objects
1848  *	that contain managed pages.  There are two exceptions.  First,
1849  *	it may be performed on the kernel and kmem objects.  Second,
1850  *	it may be used by msync(..., MS_INVALIDATE) to invalidate
1851  *	device-backed pages.
1852  *
1853  *	The object must be locked.
1854  */
1855 void
1856 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1857     boolean_t clean_only)
1858 {
1859 	vm_page_t p, next;
1860 	int wirings;
1861 
1862 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1863 	if (object->resident_page_count == 0)
1864 		goto skipmemq;
1865 
1866 	/*
1867 	 * Since physically-backed objects do not use managed pages, we can't
1868 	 * remove pages from the object (we must instead remove the page
1869 	 * references, and then destroy the object).
1870 	 */
1871 	KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1872 	    object == kmem_object,
1873 	    ("attempt to remove pages from a physical object"));
1874 
1875 	vm_object_pip_add(object, 1);
1876 again:
1877 	vm_page_lock_queues();
1878 	if ((p = TAILQ_FIRST(&object->memq)) != NULL) {
1879 		if (p->pindex < start) {
1880 			p = vm_page_splay(start, object->root);
1881 			if ((object->root = p)->pindex < start)
1882 				p = TAILQ_NEXT(p, listq);
1883 		}
1884 	}
1885 	/*
1886 	 * Assert: the variable p is either (1) the page with the
1887 	 * least pindex greater than or equal to the parameter pindex
1888 	 * or (2) NULL.
1889 	 */
1890 	for (;
1891 	     p != NULL && (p->pindex < end || end == 0);
1892 	     p = next) {
1893 		next = TAILQ_NEXT(p, listq);
1894 
1895 		/*
1896 		 * If the page is wired for any reason besides the
1897 		 * existence of managed, wired mappings, then it cannot
1898 		 * be freed.  For example, fictitious pages, which
1899 		 * represent device memory, are inherently wired and
1900 		 * cannot be freed.  They can, however, be invalidated
1901 		 * if "clean_only" is FALSE.
1902 		 */
1903 		if ((wirings = p->wire_count) != 0 &&
1904 		    (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1905 			/* Fictitious pages do not have managed mappings. */
1906 			if ((p->flags & PG_FICTITIOUS) == 0)
1907 				pmap_remove_all(p);
1908 			/* Account for removal of managed, wired mappings. */
1909 			p->wire_count -= wirings;
1910 			if (!clean_only)
1911 				p->valid = 0;
1912 			continue;
1913 		}
1914 		if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1915 			goto again;
1916 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
1917 		    ("vm_object_page_remove: page %p is fictitious", p));
1918 		if (clean_only && p->valid) {
1919 			pmap_remove_write(p);
1920 			if (p->valid & p->dirty)
1921 				continue;
1922 		}
1923 		pmap_remove_all(p);
1924 		/* Account for removal of managed, wired mappings. */
1925 		if (wirings != 0)
1926 			p->wire_count -= wirings;
1927 		vm_page_free(p);
1928 	}
1929 	vm_page_unlock_queues();
1930 	vm_object_pip_wakeup(object);
1931 skipmemq:
1932 	if (__predict_false(object->cache != NULL))
1933 		vm_page_cache_free(object, start, end);
1934 }
1935 
1936 /*
1937  *	Routine:	vm_object_coalesce
1938  *	Function:	Coalesces two objects backing up adjoining
1939  *			regions of memory into a single object.
1940  *
1941  *	returns TRUE if objects were combined.
1942  *
1943  *	NOTE:	Only works at the moment if the second object is NULL -
1944  *		if it's not, which object do we lock first?
1945  *
1946  *	Parameters:
1947  *		prev_object	First object to coalesce
1948  *		prev_offset	Offset into prev_object
1949  *		prev_size	Size of reference to prev_object
1950  *		next_size	Size of reference to the second object
1951  *
1952  *	Conditions:
1953  *	The object must *not* be locked.
1954  */
1955 boolean_t
1956 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
1957 	vm_size_t prev_size, vm_size_t next_size)
1958 {
1959 	vm_pindex_t next_pindex;
1960 
1961 	if (prev_object == NULL)
1962 		return (TRUE);
1963 	VM_OBJECT_LOCK(prev_object);
1964 	if (prev_object->type != OBJT_DEFAULT &&
1965 	    prev_object->type != OBJT_SWAP) {
1966 		VM_OBJECT_UNLOCK(prev_object);
1967 		return (FALSE);
1968 	}
1969 
1970 	/*
1971 	 * Try to collapse the object first
1972 	 */
1973 	vm_object_collapse(prev_object);
1974 
1975 	/*
1976 	 * Can't coalesce if: . more than one reference . paged out . shadows
1977 	 * another object . has a copy elsewhere (any of which mean that the
1978 	 * pages not mapped to prev_entry may be in use anyway)
1979 	 */
1980 	if (prev_object->backing_object != NULL) {
1981 		VM_OBJECT_UNLOCK(prev_object);
1982 		return (FALSE);
1983 	}
1984 
1985 	prev_size >>= PAGE_SHIFT;
1986 	next_size >>= PAGE_SHIFT;
1987 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
1988 
1989 	if ((prev_object->ref_count > 1) &&
1990 	    (prev_object->size != next_pindex)) {
1991 		VM_OBJECT_UNLOCK(prev_object);
1992 		return (FALSE);
1993 	}
1994 
1995 	/*
1996 	 * Remove any pages that may still be in the object from a previous
1997 	 * deallocation.
1998 	 */
1999 	if (next_pindex < prev_object->size) {
2000 		vm_object_page_remove(prev_object,
2001 				      next_pindex,
2002 				      next_pindex + next_size, FALSE);
2003 		if (prev_object->type == OBJT_SWAP)
2004 			swap_pager_freespace(prev_object,
2005 					     next_pindex, next_size);
2006 	}
2007 
2008 	/*
2009 	 * Extend the object if necessary.
2010 	 */
2011 	if (next_pindex + next_size > prev_object->size)
2012 		prev_object->size = next_pindex + next_size;
2013 
2014 	VM_OBJECT_UNLOCK(prev_object);
2015 	return (TRUE);
2016 }
2017 
2018 void
2019 vm_object_set_writeable_dirty(vm_object_t object)
2020 {
2021 	struct vnode *vp;
2022 
2023 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2024 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2025 		return;
2026 	vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2027 	if (object->type == OBJT_VNODE &&
2028 	    (vp = (struct vnode *)object->handle) != NULL) {
2029 		VI_LOCK(vp);
2030 		vp->v_iflag |= VI_OBJDIRTY;
2031 		VI_UNLOCK(vp);
2032 	}
2033 }
2034 
2035 #include "opt_ddb.h"
2036 #ifdef DDB
2037 #include <sys/kernel.h>
2038 
2039 #include <sys/cons.h>
2040 
2041 #include <ddb/ddb.h>
2042 
2043 static int
2044 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2045 {
2046 	vm_map_t tmpm;
2047 	vm_map_entry_t tmpe;
2048 	vm_object_t obj;
2049 	int entcount;
2050 
2051 	if (map == 0)
2052 		return 0;
2053 
2054 	if (entry == 0) {
2055 		tmpe = map->header.next;
2056 		entcount = map->nentries;
2057 		while (entcount-- && (tmpe != &map->header)) {
2058 			if (_vm_object_in_map(map, object, tmpe)) {
2059 				return 1;
2060 			}
2061 			tmpe = tmpe->next;
2062 		}
2063 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2064 		tmpm = entry->object.sub_map;
2065 		tmpe = tmpm->header.next;
2066 		entcount = tmpm->nentries;
2067 		while (entcount-- && tmpe != &tmpm->header) {
2068 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2069 				return 1;
2070 			}
2071 			tmpe = tmpe->next;
2072 		}
2073 	} else if ((obj = entry->object.vm_object) != NULL) {
2074 		for (; obj; obj = obj->backing_object)
2075 			if (obj == object) {
2076 				return 1;
2077 			}
2078 	}
2079 	return 0;
2080 }
2081 
2082 static int
2083 vm_object_in_map(vm_object_t object)
2084 {
2085 	struct proc *p;
2086 
2087 	/* sx_slock(&allproc_lock); */
2088 	FOREACH_PROC_IN_SYSTEM(p) {
2089 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2090 			continue;
2091 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2092 			/* sx_sunlock(&allproc_lock); */
2093 			return 1;
2094 		}
2095 	}
2096 	/* sx_sunlock(&allproc_lock); */
2097 	if (_vm_object_in_map(kernel_map, object, 0))
2098 		return 1;
2099 	if (_vm_object_in_map(kmem_map, object, 0))
2100 		return 1;
2101 	if (_vm_object_in_map(pager_map, object, 0))
2102 		return 1;
2103 	if (_vm_object_in_map(buffer_map, object, 0))
2104 		return 1;
2105 	return 0;
2106 }
2107 
2108 DB_SHOW_COMMAND(vmochk, vm_object_check)
2109 {
2110 	vm_object_t object;
2111 
2112 	/*
2113 	 * make sure that internal objs are in a map somewhere
2114 	 * and none have zero ref counts.
2115 	 */
2116 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2117 		if (object->handle == NULL &&
2118 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2119 			if (object->ref_count == 0) {
2120 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2121 					(long)object->size);
2122 			}
2123 			if (!vm_object_in_map(object)) {
2124 				db_printf(
2125 			"vmochk: internal obj is not in a map: "
2126 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2127 				    object->ref_count, (u_long)object->size,
2128 				    (u_long)object->size,
2129 				    (void *)object->backing_object);
2130 			}
2131 		}
2132 	}
2133 }
2134 
2135 /*
2136  *	vm_object_print:	[ debug ]
2137  */
2138 DB_SHOW_COMMAND(object, vm_object_print_static)
2139 {
2140 	/* XXX convert args. */
2141 	vm_object_t object = (vm_object_t)addr;
2142 	boolean_t full = have_addr;
2143 
2144 	vm_page_t p;
2145 
2146 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2147 #define	count	was_count
2148 
2149 	int count;
2150 
2151 	if (object == NULL)
2152 		return;
2153 
2154 	db_iprintf(
2155 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x\n",
2156 	    object, (int)object->type, (uintmax_t)object->size,
2157 	    object->resident_page_count, object->ref_count, object->flags);
2158 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2159 	    object->shadow_count,
2160 	    object->backing_object ? object->backing_object->ref_count : 0,
2161 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2162 
2163 	if (!full)
2164 		return;
2165 
2166 	db_indent += 2;
2167 	count = 0;
2168 	TAILQ_FOREACH(p, &object->memq, listq) {
2169 		if (count == 0)
2170 			db_iprintf("memory:=");
2171 		else if (count == 6) {
2172 			db_printf("\n");
2173 			db_iprintf(" ...");
2174 			count = 0;
2175 		} else
2176 			db_printf(",");
2177 		count++;
2178 
2179 		db_printf("(off=0x%jx,page=0x%jx)",
2180 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2181 	}
2182 	if (count != 0)
2183 		db_printf("\n");
2184 	db_indent -= 2;
2185 }
2186 
2187 /* XXX. */
2188 #undef count
2189 
2190 /* XXX need this non-static entry for calling from vm_map_print. */
2191 void
2192 vm_object_print(
2193         /* db_expr_t */ long addr,
2194 	boolean_t have_addr,
2195 	/* db_expr_t */ long count,
2196 	char *modif)
2197 {
2198 	vm_object_print_static(addr, have_addr, count, modif);
2199 }
2200 
2201 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2202 {
2203 	vm_object_t object;
2204 	int nl = 0;
2205 	int c;
2206 
2207 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2208 		vm_pindex_t idx, fidx;
2209 		vm_pindex_t osize;
2210 		vm_paddr_t pa = -1;
2211 		int rcount;
2212 		vm_page_t m;
2213 
2214 		db_printf("new object: %p\n", (void *)object);
2215 		if (nl > 18) {
2216 			c = cngetc();
2217 			if (c != ' ')
2218 				return;
2219 			nl = 0;
2220 		}
2221 		nl++;
2222 		rcount = 0;
2223 		fidx = 0;
2224 		osize = object->size;
2225 		if (osize > 128)
2226 			osize = 128;
2227 		for (idx = 0; idx < osize; idx++) {
2228 			m = vm_page_lookup(object, idx);
2229 			if (m == NULL) {
2230 				if (rcount) {
2231 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2232 						(long)fidx, rcount, (long)pa);
2233 					if (nl > 18) {
2234 						c = cngetc();
2235 						if (c != ' ')
2236 							return;
2237 						nl = 0;
2238 					}
2239 					nl++;
2240 					rcount = 0;
2241 				}
2242 				continue;
2243 			}
2244 
2245 
2246 			if (rcount &&
2247 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2248 				++rcount;
2249 				continue;
2250 			}
2251 			if (rcount) {
2252 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2253 					(long)fidx, rcount, (long)pa);
2254 				if (nl > 18) {
2255 					c = cngetc();
2256 					if (c != ' ')
2257 						return;
2258 					nl = 0;
2259 				}
2260 				nl++;
2261 			}
2262 			fidx = idx;
2263 			pa = VM_PAGE_TO_PHYS(m);
2264 			rcount = 1;
2265 		}
2266 		if (rcount) {
2267 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2268 				(long)fidx, rcount, (long)pa);
2269 			if (nl > 18) {
2270 				c = cngetc();
2271 				if (c != ' ')
2272 					return;
2273 				nl = 0;
2274 			}
2275 			nl++;
2276 		}
2277 	}
2278 }
2279 #endif /* DDB */
2280