xref: /freebsd/sys/vm/vm_object.c (revision aa77200569e397d6ff1fdb4d255d0fa254d0a128)
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 static int old_msync;
100 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
101     "Use old (insecure) msync behavior");
102 
103 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
104 		    int pagerflags, int flags, boolean_t *clearobjflags,
105 		    boolean_t *eio);
106 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
107 		    boolean_t *clearobjflags);
108 static void	vm_object_qcollapse(vm_object_t object);
109 static void	vm_object_vndeallocate(vm_object_t object);
110 
111 /*
112  *	Virtual memory objects maintain the actual data
113  *	associated with allocated virtual memory.  A given
114  *	page of memory exists within exactly one object.
115  *
116  *	An object is only deallocated when all "references"
117  *	are given up.  Only one "reference" to a given
118  *	region of an object should be writeable.
119  *
120  *	Associated with each object is a list of all resident
121  *	memory pages belonging to that object; this list is
122  *	maintained by the "vm_page" module, and locked by the object's
123  *	lock.
124  *
125  *	Each object also records a "pager" routine which is
126  *	used to retrieve (and store) pages to the proper backing
127  *	storage.  In addition, objects may be backed by other
128  *	objects from which they were virtual-copied.
129  *
130  *	The only items within the object structure which are
131  *	modified after time of creation are:
132  *		reference count		locked by object's lock
133  *		pager routine		locked by object's lock
134  *
135  */
136 
137 struct object_q vm_object_list;
138 struct mtx vm_object_list_mtx;	/* lock for object list and count */
139 
140 struct vm_object kernel_object_store;
141 struct vm_object kmem_object_store;
142 
143 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
144     "VM object stats");
145 
146 static long object_collapses;
147 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
148     &object_collapses, 0, "VM object collapses");
149 
150 static long object_bypasses;
151 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
152     &object_bypasses, 0, "VM object bypasses");
153 
154 static uma_zone_t obj_zone;
155 
156 static int vm_object_zinit(void *mem, int size, int flags);
157 
158 #ifdef INVARIANTS
159 static void vm_object_zdtor(void *mem, int size, void *arg);
160 
161 static void
162 vm_object_zdtor(void *mem, int size, void *arg)
163 {
164 	vm_object_t object;
165 
166 	object = (vm_object_t)mem;
167 	KASSERT(TAILQ_EMPTY(&object->memq),
168 	    ("object %p has resident pages",
169 	    object));
170 #if VM_NRESERVLEVEL > 0
171 	KASSERT(LIST_EMPTY(&object->rvq),
172 	    ("object %p has reservations",
173 	    object));
174 #endif
175 	KASSERT(object->cache == NULL,
176 	    ("object %p has cached pages",
177 	    object));
178 	KASSERT(object->paging_in_progress == 0,
179 	    ("object %p paging_in_progress = %d",
180 	    object, object->paging_in_progress));
181 	KASSERT(object->resident_page_count == 0,
182 	    ("object %p resident_page_count = %d",
183 	    object, object->resident_page_count));
184 	KASSERT(object->shadow_count == 0,
185 	    ("object %p shadow_count = %d",
186 	    object, object->shadow_count));
187 }
188 #endif
189 
190 static int
191 vm_object_zinit(void *mem, int size, int flags)
192 {
193 	vm_object_t object;
194 
195 	object = (vm_object_t)mem;
196 	bzero(&object->mtx, sizeof(object->mtx));
197 	VM_OBJECT_LOCK_INIT(object, "standard object");
198 
199 	/* These are true for any object that has been freed */
200 	object->paging_in_progress = 0;
201 	object->resident_page_count = 0;
202 	object->shadow_count = 0;
203 	return (0);
204 }
205 
206 void
207 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
208 {
209 
210 	TAILQ_INIT(&object->memq);
211 	LIST_INIT(&object->shadow_head);
212 
213 	object->root = NULL;
214 	object->type = type;
215 	object->size = size;
216 	object->generation = 1;
217 	object->ref_count = 1;
218 	object->memattr = VM_MEMATTR_DEFAULT;
219 	object->flags = 0;
220 	object->cred = NULL;
221 	object->charge = 0;
222 	if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
223 		object->flags = OBJ_ONEMAPPING;
224 	object->pg_color = 0;
225 	object->handle = NULL;
226 	object->backing_object = NULL;
227 	object->backing_object_offset = (vm_ooffset_t) 0;
228 #if VM_NRESERVLEVEL > 0
229 	LIST_INIT(&object->rvq);
230 #endif
231 	object->cache = NULL;
232 
233 	mtx_lock(&vm_object_list_mtx);
234 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
235 	mtx_unlock(&vm_object_list_mtx);
236 }
237 
238 /*
239  *	vm_object_init:
240  *
241  *	Initialize the VM objects module.
242  */
243 void
244 vm_object_init(void)
245 {
246 	TAILQ_INIT(&vm_object_list);
247 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
248 
249 	VM_OBJECT_LOCK_INIT(kernel_object, "kernel object");
250 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
251 	    kernel_object);
252 #if VM_NRESERVLEVEL > 0
253 	kernel_object->flags |= OBJ_COLORED;
254 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
255 #endif
256 
257 	VM_OBJECT_LOCK_INIT(kmem_object, "kmem object");
258 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
259 	    kmem_object);
260 #if VM_NRESERVLEVEL > 0
261 	kmem_object->flags |= OBJ_COLORED;
262 	kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
263 #endif
264 
265 	/*
266 	 * The lock portion of struct vm_object must be type stable due
267 	 * to vm_pageout_fallback_object_lock locking a vm object
268 	 * without holding any references to it.
269 	 */
270 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
271 #ifdef INVARIANTS
272 	    vm_object_zdtor,
273 #else
274 	    NULL,
275 #endif
276 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
277 }
278 
279 void
280 vm_object_clear_flag(vm_object_t object, u_short bits)
281 {
282 
283 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
284 	object->flags &= ~bits;
285 }
286 
287 /*
288  *	Sets the default memory attribute for the specified object.  Pages
289  *	that are allocated to this object are by default assigned this memory
290  *	attribute.
291  *
292  *	Presently, this function must be called before any pages are allocated
293  *	to the object.  In the future, this requirement may be relaxed for
294  *	"default" and "swap" objects.
295  */
296 int
297 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
298 {
299 
300 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
301 	switch (object->type) {
302 	case OBJT_DEFAULT:
303 	case OBJT_DEVICE:
304 	case OBJT_MGTDEVICE:
305 	case OBJT_PHYS:
306 	case OBJT_SG:
307 	case OBJT_SWAP:
308 	case OBJT_VNODE:
309 		if (!TAILQ_EMPTY(&object->memq))
310 			return (KERN_FAILURE);
311 		break;
312 	case OBJT_DEAD:
313 		return (KERN_INVALID_ARGUMENT);
314 	default:
315 		panic("vm_object_set_memattr: object %p is of undefined type",
316 		    object);
317 	}
318 	object->memattr = memattr;
319 	return (KERN_SUCCESS);
320 }
321 
322 void
323 vm_object_pip_add(vm_object_t object, short i)
324 {
325 
326 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
327 	object->paging_in_progress += i;
328 }
329 
330 void
331 vm_object_pip_subtract(vm_object_t object, short i)
332 {
333 
334 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
335 	object->paging_in_progress -= i;
336 }
337 
338 void
339 vm_object_pip_wakeup(vm_object_t object)
340 {
341 
342 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
343 	object->paging_in_progress--;
344 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
345 		vm_object_clear_flag(object, OBJ_PIPWNT);
346 		wakeup(object);
347 	}
348 }
349 
350 void
351 vm_object_pip_wakeupn(vm_object_t object, short i)
352 {
353 
354 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
355 	if (i)
356 		object->paging_in_progress -= i;
357 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
358 		vm_object_clear_flag(object, OBJ_PIPWNT);
359 		wakeup(object);
360 	}
361 }
362 
363 void
364 vm_object_pip_wait(vm_object_t object, char *waitid)
365 {
366 
367 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
368 	while (object->paging_in_progress) {
369 		object->flags |= OBJ_PIPWNT;
370 		msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
371 	}
372 }
373 
374 /*
375  *	vm_object_allocate:
376  *
377  *	Returns a new object with the given size.
378  */
379 vm_object_t
380 vm_object_allocate(objtype_t type, vm_pindex_t size)
381 {
382 	vm_object_t object;
383 
384 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
385 	_vm_object_allocate(type, size, object);
386 	return (object);
387 }
388 
389 
390 /*
391  *	vm_object_reference:
392  *
393  *	Gets another reference to the given object.  Note: OBJ_DEAD
394  *	objects can be referenced during final cleaning.
395  */
396 void
397 vm_object_reference(vm_object_t object)
398 {
399 	if (object == NULL)
400 		return;
401 	VM_OBJECT_LOCK(object);
402 	vm_object_reference_locked(object);
403 	VM_OBJECT_UNLOCK(object);
404 }
405 
406 /*
407  *	vm_object_reference_locked:
408  *
409  *	Gets another reference to the given object.
410  *
411  *	The object must be locked.
412  */
413 void
414 vm_object_reference_locked(vm_object_t object)
415 {
416 	struct vnode *vp;
417 
418 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
419 	object->ref_count++;
420 	if (object->type == OBJT_VNODE) {
421 		vp = object->handle;
422 		vref(vp);
423 	}
424 }
425 
426 /*
427  * Handle deallocating an object of type OBJT_VNODE.
428  */
429 static void
430 vm_object_vndeallocate(vm_object_t object)
431 {
432 	struct vnode *vp = (struct vnode *) object->handle;
433 
434 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
435 	KASSERT(object->type == OBJT_VNODE,
436 	    ("vm_object_vndeallocate: not a vnode object"));
437 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
438 #ifdef INVARIANTS
439 	if (object->ref_count == 0) {
440 		vprint("vm_object_vndeallocate", vp);
441 		panic("vm_object_vndeallocate: bad object reference count");
442 	}
443 #endif
444 
445 	if (object->ref_count > 1) {
446 		object->ref_count--;
447 		VM_OBJECT_UNLOCK(object);
448 		/* vrele may need the vnode lock. */
449 		vrele(vp);
450 	} else {
451 		vhold(vp);
452 		VM_OBJECT_UNLOCK(object);
453 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
454 		vdrop(vp);
455 		VM_OBJECT_LOCK(object);
456 		object->ref_count--;
457 		if (object->type == OBJT_DEAD) {
458 			VM_OBJECT_UNLOCK(object);
459 			VOP_UNLOCK(vp, 0);
460 		} else {
461 			if (object->ref_count == 0)
462 				VOP_UNSET_TEXT(vp);
463 			VM_OBJECT_UNLOCK(object);
464 			vput(vp);
465 		}
466 	}
467 }
468 
469 /*
470  *	vm_object_deallocate:
471  *
472  *	Release a reference to the specified object,
473  *	gained either through a vm_object_allocate
474  *	or a vm_object_reference call.  When all references
475  *	are gone, storage associated with this object
476  *	may be relinquished.
477  *
478  *	No object may be locked.
479  */
480 void
481 vm_object_deallocate(vm_object_t object)
482 {
483 	vm_object_t temp;
484 
485 	while (object != NULL) {
486 		VM_OBJECT_LOCK(object);
487 		if (object->type == OBJT_VNODE) {
488 			vm_object_vndeallocate(object);
489 			return;
490 		}
491 
492 		KASSERT(object->ref_count != 0,
493 			("vm_object_deallocate: object deallocated too many times: %d", object->type));
494 
495 		/*
496 		 * If the reference count goes to 0 we start calling
497 		 * vm_object_terminate() on the object chain.
498 		 * A ref count of 1 may be a special case depending on the
499 		 * shadow count being 0 or 1.
500 		 */
501 		object->ref_count--;
502 		if (object->ref_count > 1) {
503 			VM_OBJECT_UNLOCK(object);
504 			return;
505 		} else if (object->ref_count == 1) {
506 			if (object->shadow_count == 0 &&
507 			    object->handle == NULL &&
508 			    (object->type == OBJT_DEFAULT ||
509 			     object->type == OBJT_SWAP)) {
510 				vm_object_set_flag(object, OBJ_ONEMAPPING);
511 			} else if ((object->shadow_count == 1) &&
512 			    (object->handle == NULL) &&
513 			    (object->type == OBJT_DEFAULT ||
514 			     object->type == OBJT_SWAP)) {
515 				vm_object_t robject;
516 
517 				robject = LIST_FIRST(&object->shadow_head);
518 				KASSERT(robject != NULL,
519 				    ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
520 					 object->ref_count,
521 					 object->shadow_count));
522 				if (!VM_OBJECT_TRYLOCK(robject)) {
523 					/*
524 					 * Avoid a potential deadlock.
525 					 */
526 					object->ref_count++;
527 					VM_OBJECT_UNLOCK(object);
528 					/*
529 					 * More likely than not the thread
530 					 * holding robject's lock has lower
531 					 * priority than the current thread.
532 					 * Let the lower priority thread run.
533 					 */
534 					pause("vmo_de", 1);
535 					continue;
536 				}
537 				/*
538 				 * Collapse object into its shadow unless its
539 				 * shadow is dead.  In that case, object will
540 				 * be deallocated by the thread that is
541 				 * deallocating its shadow.
542 				 */
543 				if ((robject->flags & OBJ_DEAD) == 0 &&
544 				    (robject->handle == NULL) &&
545 				    (robject->type == OBJT_DEFAULT ||
546 				     robject->type == OBJT_SWAP)) {
547 
548 					robject->ref_count++;
549 retry:
550 					if (robject->paging_in_progress) {
551 						VM_OBJECT_UNLOCK(object);
552 						vm_object_pip_wait(robject,
553 						    "objde1");
554 						temp = robject->backing_object;
555 						if (object == temp) {
556 							VM_OBJECT_LOCK(object);
557 							goto retry;
558 						}
559 					} else if (object->paging_in_progress) {
560 						VM_OBJECT_UNLOCK(robject);
561 						object->flags |= OBJ_PIPWNT;
562 						msleep(object,
563 						    VM_OBJECT_MTX(object),
564 						    PDROP | PVM, "objde2", 0);
565 						VM_OBJECT_LOCK(robject);
566 						temp = robject->backing_object;
567 						if (object == temp) {
568 							VM_OBJECT_LOCK(object);
569 							goto retry;
570 						}
571 					} else
572 						VM_OBJECT_UNLOCK(object);
573 
574 					if (robject->ref_count == 1) {
575 						robject->ref_count--;
576 						object = robject;
577 						goto doterm;
578 					}
579 					object = robject;
580 					vm_object_collapse(object);
581 					VM_OBJECT_UNLOCK(object);
582 					continue;
583 				}
584 				VM_OBJECT_UNLOCK(robject);
585 			}
586 			VM_OBJECT_UNLOCK(object);
587 			return;
588 		}
589 doterm:
590 		temp = object->backing_object;
591 		if (temp != NULL) {
592 			VM_OBJECT_LOCK(temp);
593 			LIST_REMOVE(object, shadow_list);
594 			temp->shadow_count--;
595 			VM_OBJECT_UNLOCK(temp);
596 			object->backing_object = NULL;
597 		}
598 		/*
599 		 * Don't double-terminate, we could be in a termination
600 		 * recursion due to the terminate having to sync data
601 		 * to disk.
602 		 */
603 		if ((object->flags & OBJ_DEAD) == 0)
604 			vm_object_terminate(object);
605 		else
606 			VM_OBJECT_UNLOCK(object);
607 		object = temp;
608 	}
609 }
610 
611 /*
612  *	vm_object_destroy removes the object from the global object list
613  *      and frees the space for the object.
614  */
615 void
616 vm_object_destroy(vm_object_t object)
617 {
618 
619 	/*
620 	 * Remove the object from the global object list.
621 	 */
622 	mtx_lock(&vm_object_list_mtx);
623 	TAILQ_REMOVE(&vm_object_list, object, object_list);
624 	mtx_unlock(&vm_object_list_mtx);
625 
626 	/*
627 	 * Release the allocation charge.
628 	 */
629 	if (object->cred != NULL) {
630 		KASSERT(object->type == OBJT_DEFAULT ||
631 		    object->type == OBJT_SWAP,
632 		    ("vm_object_terminate: non-swap obj %p has cred",
633 		     object));
634 		swap_release_by_cred(object->charge, object->cred);
635 		object->charge = 0;
636 		crfree(object->cred);
637 		object->cred = NULL;
638 	}
639 
640 	/*
641 	 * Free the space for the object.
642 	 */
643 	uma_zfree(obj_zone, object);
644 }
645 
646 /*
647  *	vm_object_terminate actually destroys the specified object, freeing
648  *	up all previously used resources.
649  *
650  *	The object must be locked.
651  *	This routine may block.
652  */
653 void
654 vm_object_terminate(vm_object_t object)
655 {
656 	vm_page_t p, p_next;
657 
658 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
659 
660 	/*
661 	 * Make sure no one uses us.
662 	 */
663 	vm_object_set_flag(object, OBJ_DEAD);
664 
665 	/*
666 	 * wait for the pageout daemon to be done with the object
667 	 */
668 	vm_object_pip_wait(object, "objtrm");
669 
670 	KASSERT(!object->paging_in_progress,
671 		("vm_object_terminate: pageout in progress"));
672 
673 	/*
674 	 * Clean and free the pages, as appropriate. All references to the
675 	 * object are gone, so we don't need to lock it.
676 	 */
677 	if (object->type == OBJT_VNODE) {
678 		struct vnode *vp = (struct vnode *)object->handle;
679 
680 		/*
681 		 * Clean pages and flush buffers.
682 		 */
683 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
684 		VM_OBJECT_UNLOCK(object);
685 
686 		vinvalbuf(vp, V_SAVE, 0, 0);
687 
688 		VM_OBJECT_LOCK(object);
689 	}
690 
691 	KASSERT(object->ref_count == 0,
692 		("vm_object_terminate: object with references, ref_count=%d",
693 		object->ref_count));
694 
695 	/*
696 	 * Free any remaining pageable pages.  This also removes them from the
697 	 * paging queues.  However, don't free wired pages, just remove them
698 	 * from the object.  Rather than incrementally removing each page from
699 	 * the object, the page and object are reset to any empty state.
700 	 */
701 	TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
702 		KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
703 		    ("vm_object_terminate: freeing busy page %p", p));
704 		vm_page_lock(p);
705 		/*
706 		 * Optimize the page's removal from the object by resetting
707 		 * its "object" field.  Specifically, if the page is not
708 		 * wired, then the effect of this assignment is that
709 		 * vm_page_free()'s call to vm_page_remove() will return
710 		 * immediately without modifying the page or the object.
711 		 */
712 		p->object = NULL;
713 		if (p->wire_count == 0) {
714 			vm_page_free(p);
715 			PCPU_INC(cnt.v_pfree);
716 		}
717 		vm_page_unlock(p);
718 	}
719 	/*
720 	 * If the object contained any pages, then reset it to an empty state.
721 	 * None of the object's fields, including "resident_page_count", were
722 	 * modified by the preceding loop.
723 	 */
724 	if (object->resident_page_count != 0) {
725 		object->root = NULL;
726 		TAILQ_INIT(&object->memq);
727 		object->resident_page_count = 0;
728 		if (object->type == OBJT_VNODE)
729 			vdrop(object->handle);
730 	}
731 
732 #if VM_NRESERVLEVEL > 0
733 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
734 		vm_reserv_break_all(object);
735 #endif
736 	if (__predict_false(object->cache != NULL))
737 		vm_page_cache_free(object, 0, 0);
738 
739 	/*
740 	 * Let the pager know object is dead.
741 	 */
742 	vm_pager_deallocate(object);
743 	VM_OBJECT_UNLOCK(object);
744 
745 	vm_object_destroy(object);
746 }
747 
748 /*
749  * Make the page read-only so that we can clear the object flags.  However, if
750  * this is a nosync mmap then the object is likely to stay dirty so do not
751  * mess with the page and do not clear the object flags.  Returns TRUE if the
752  * page should be flushed, and FALSE otherwise.
753  */
754 static boolean_t
755 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
756 {
757 
758 	/*
759 	 * If we have been asked to skip nosync pages and this is a
760 	 * nosync page, skip it.  Note that the object flags were not
761 	 * cleared in this case so we do not have to set them.
762 	 */
763 	if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
764 		*clearobjflags = FALSE;
765 		return (FALSE);
766 	} else {
767 		pmap_remove_write(p);
768 		return (p->dirty != 0);
769 	}
770 }
771 
772 /*
773  *	vm_object_page_clean
774  *
775  *	Clean all dirty pages in the specified range of object.  Leaves page
776  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
777  *	write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
778  *	leaving the object dirty.
779  *
780  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
781  *	synchronous clustering mode implementation.
782  *
783  *	Odd semantics: if start == end, we clean everything.
784  *
785  *	The object must be locked.
786  *
787  *	Returns FALSE if some page from the range was not written, as
788  *	reported by the pager, and TRUE otherwise.
789  */
790 boolean_t
791 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
792     int flags)
793 {
794 	vm_page_t np, p;
795 	vm_pindex_t pi, tend, tstart;
796 	int curgeneration, n, pagerflags;
797 	boolean_t clearobjflags, eio, res;
798 
799 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
800 	KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
801 	if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
802 	    object->resident_page_count == 0)
803 		return (TRUE);
804 
805 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
806 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
807 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
808 
809 	tstart = OFF_TO_IDX(start);
810 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
811 	clearobjflags = tstart == 0 && tend >= object->size;
812 	res = TRUE;
813 
814 rescan:
815 	curgeneration = object->generation;
816 
817 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
818 		pi = p->pindex;
819 		if (pi >= tend)
820 			break;
821 		np = TAILQ_NEXT(p, listq);
822 		if (p->valid == 0)
823 			continue;
824 		if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
825 			if (object->generation != curgeneration) {
826 				if ((flags & OBJPC_SYNC) != 0)
827 					goto rescan;
828 				else
829 					clearobjflags = FALSE;
830 			}
831 			np = vm_page_find_least(object, pi);
832 			continue;
833 		}
834 		if (!vm_object_page_remove_write(p, flags, &clearobjflags))
835 			continue;
836 
837 		n = vm_object_page_collect_flush(object, p, pagerflags,
838 		    flags, &clearobjflags, &eio);
839 		if (eio) {
840 			res = FALSE;
841 			clearobjflags = FALSE;
842 		}
843 		if (object->generation != curgeneration) {
844 			if ((flags & OBJPC_SYNC) != 0)
845 				goto rescan;
846 			else
847 				clearobjflags = FALSE;
848 		}
849 
850 		/*
851 		 * If the VOP_PUTPAGES() did a truncated write, so
852 		 * that even the first page of the run is not fully
853 		 * written, vm_pageout_flush() returns 0 as the run
854 		 * length.  Since the condition that caused truncated
855 		 * write may be permanent, e.g. exhausted free space,
856 		 * accepting n == 0 would cause an infinite loop.
857 		 *
858 		 * Forwarding the iterator leaves the unwritten page
859 		 * behind, but there is not much we can do there if
860 		 * filesystem refuses to write it.
861 		 */
862 		if (n == 0) {
863 			n = 1;
864 			clearobjflags = FALSE;
865 		}
866 		np = vm_page_find_least(object, pi + n);
867 	}
868 #if 0
869 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
870 #endif
871 
872 	if (clearobjflags)
873 		vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
874 	return (res);
875 }
876 
877 static int
878 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
879     int flags, boolean_t *clearobjflags, boolean_t *eio)
880 {
881 	vm_page_t ma[vm_pageout_page_count], p_first, tp;
882 	int count, i, mreq, runlen;
883 
884 	vm_page_lock_assert(p, MA_NOTOWNED);
885 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
886 
887 	count = 1;
888 	mreq = 0;
889 
890 	for (tp = p; count < vm_pageout_page_count; count++) {
891 		tp = vm_page_next(tp);
892 		if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
893 			break;
894 		if (!vm_object_page_remove_write(tp, flags, clearobjflags))
895 			break;
896 	}
897 
898 	for (p_first = p; count < vm_pageout_page_count; count++) {
899 		tp = vm_page_prev(p_first);
900 		if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
901 			break;
902 		if (!vm_object_page_remove_write(tp, flags, clearobjflags))
903 			break;
904 		p_first = tp;
905 		mreq++;
906 	}
907 
908 	for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
909 		ma[i] = tp;
910 
911 	vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
912 	return (runlen);
913 }
914 
915 /*
916  * Note that there is absolutely no sense in writing out
917  * anonymous objects, so we track down the vnode object
918  * to write out.
919  * We invalidate (remove) all pages from the address space
920  * for semantic correctness.
921  *
922  * If the backing object is a device object with unmanaged pages, then any
923  * mappings to the specified range of pages must be removed before this
924  * function is called.
925  *
926  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
927  * may start out with a NULL object.
928  */
929 boolean_t
930 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
931     boolean_t syncio, boolean_t invalidate)
932 {
933 	vm_object_t backing_object;
934 	struct vnode *vp;
935 	struct mount *mp;
936 	int error, flags, fsync_after;
937 	boolean_t res;
938 
939 	if (object == NULL)
940 		return (TRUE);
941 	res = TRUE;
942 	error = 0;
943 	VM_OBJECT_LOCK(object);
944 	while ((backing_object = object->backing_object) != NULL) {
945 		VM_OBJECT_LOCK(backing_object);
946 		offset += object->backing_object_offset;
947 		VM_OBJECT_UNLOCK(object);
948 		object = backing_object;
949 		if (object->size < OFF_TO_IDX(offset + size))
950 			size = IDX_TO_OFF(object->size) - offset;
951 	}
952 	/*
953 	 * Flush pages if writing is allowed, invalidate them
954 	 * if invalidation requested.  Pages undergoing I/O
955 	 * will be ignored by vm_object_page_remove().
956 	 *
957 	 * We cannot lock the vnode and then wait for paging
958 	 * to complete without deadlocking against vm_fault.
959 	 * Instead we simply call vm_object_page_remove() and
960 	 * allow it to block internally on a page-by-page
961 	 * basis when it encounters pages undergoing async
962 	 * I/O.
963 	 */
964 	if (object->type == OBJT_VNODE &&
965 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
966 		vp = object->handle;
967 		VM_OBJECT_UNLOCK(object);
968 		(void) vn_start_write(vp, &mp, V_WAIT);
969 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
970 		if (syncio && !invalidate && offset == 0 &&
971 		    OFF_TO_IDX(size) == object->size) {
972 			/*
973 			 * If syncing the whole mapping of the file,
974 			 * it is faster to schedule all the writes in
975 			 * async mode, also allowing the clustering,
976 			 * and then wait for i/o to complete.
977 			 */
978 			flags = 0;
979 			fsync_after = TRUE;
980 		} else {
981 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
982 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
983 			fsync_after = FALSE;
984 		}
985 		VM_OBJECT_LOCK(object);
986 		res = vm_object_page_clean(object, offset, offset + size,
987 		    flags);
988 		VM_OBJECT_UNLOCK(object);
989 		if (fsync_after)
990 			error = VOP_FSYNC(vp, MNT_WAIT, curthread);
991 		VOP_UNLOCK(vp, 0);
992 		vn_finished_write(mp);
993 		if (error != 0)
994 			res = FALSE;
995 		VM_OBJECT_LOCK(object);
996 	}
997 	if ((object->type == OBJT_VNODE ||
998 	     object->type == OBJT_DEVICE) && invalidate) {
999 		if (object->type == OBJT_DEVICE)
1000 			/*
1001 			 * The option OBJPR_NOTMAPPED must be passed here
1002 			 * because vm_object_page_remove() cannot remove
1003 			 * unmanaged mappings.
1004 			 */
1005 			flags = OBJPR_NOTMAPPED;
1006 		else if (old_msync)
1007 			flags = 0;
1008 		else
1009 			flags = OBJPR_CLEANONLY;
1010 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1011 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1012 	}
1013 	VM_OBJECT_UNLOCK(object);
1014 	return (res);
1015 }
1016 
1017 /*
1018  *	vm_object_madvise:
1019  *
1020  *	Implements the madvise function at the object/page level.
1021  *
1022  *	MADV_WILLNEED	(any object)
1023  *
1024  *	    Activate the specified pages if they are resident.
1025  *
1026  *	MADV_DONTNEED	(any object)
1027  *
1028  *	    Deactivate the specified pages if they are resident.
1029  *
1030  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1031  *			 OBJ_ONEMAPPING only)
1032  *
1033  *	    Deactivate and clean the specified pages if they are
1034  *	    resident.  This permits the process to reuse the pages
1035  *	    without faulting or the kernel to reclaim the pages
1036  *	    without I/O.
1037  */
1038 void
1039 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1040     int advise)
1041 {
1042 	vm_pindex_t tpindex;
1043 	vm_object_t backing_object, tobject;
1044 	vm_page_t m;
1045 
1046 	if (object == NULL)
1047 		return;
1048 	VM_OBJECT_LOCK(object);
1049 	/*
1050 	 * Locate and adjust resident pages
1051 	 */
1052 	for (; pindex < end; pindex += 1) {
1053 relookup:
1054 		tobject = object;
1055 		tpindex = pindex;
1056 shadowlookup:
1057 		/*
1058 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1059 		 * and those pages must be OBJ_ONEMAPPING.
1060 		 */
1061 		if (advise == MADV_FREE) {
1062 			if ((tobject->type != OBJT_DEFAULT &&
1063 			     tobject->type != OBJT_SWAP) ||
1064 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
1065 				goto unlock_tobject;
1066 			}
1067 		} else if (tobject->type == OBJT_PHYS)
1068 			goto unlock_tobject;
1069 		m = vm_page_lookup(tobject, tpindex);
1070 		if (m == NULL && advise == MADV_WILLNEED) {
1071 			/*
1072 			 * If the page is cached, reactivate it.
1073 			 */
1074 			m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1075 			    VM_ALLOC_NOBUSY);
1076 		}
1077 		if (m == NULL) {
1078 			/*
1079 			 * There may be swap even if there is no backing page
1080 			 */
1081 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1082 				swap_pager_freespace(tobject, tpindex, 1);
1083 			/*
1084 			 * next object
1085 			 */
1086 			backing_object = tobject->backing_object;
1087 			if (backing_object == NULL)
1088 				goto unlock_tobject;
1089 			VM_OBJECT_LOCK(backing_object);
1090 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1091 			if (tobject != object)
1092 				VM_OBJECT_UNLOCK(tobject);
1093 			tobject = backing_object;
1094 			goto shadowlookup;
1095 		} else if (m->valid != VM_PAGE_BITS_ALL)
1096 			goto unlock_tobject;
1097 		/*
1098 		 * If the page is not in a normal state, skip it.
1099 		 */
1100 		vm_page_lock(m);
1101 		if (m->hold_count != 0 || m->wire_count != 0) {
1102 			vm_page_unlock(m);
1103 			goto unlock_tobject;
1104 		}
1105 		KASSERT((m->flags & PG_FICTITIOUS) == 0,
1106 		    ("vm_object_madvise: page %p is fictitious", m));
1107 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1108 		    ("vm_object_madvise: page %p is not managed", m));
1109 		if ((m->oflags & VPO_BUSY) || m->busy) {
1110 			if (advise == MADV_WILLNEED) {
1111 				/*
1112 				 * Reference the page before unlocking and
1113 				 * sleeping so that the page daemon is less
1114 				 * likely to reclaim it.
1115 				 */
1116 				vm_page_aflag_set(m, PGA_REFERENCED);
1117 			}
1118 			vm_page_unlock(m);
1119 			if (object != tobject)
1120 				VM_OBJECT_UNLOCK(object);
1121 			m->oflags |= VPO_WANTED;
1122 			msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1123 			    0);
1124 			VM_OBJECT_LOCK(object);
1125   			goto relookup;
1126 		}
1127 		if (advise == MADV_WILLNEED) {
1128 			vm_page_activate(m);
1129 		} else if (advise == MADV_DONTNEED) {
1130 			vm_page_dontneed(m);
1131 		} else if (advise == MADV_FREE) {
1132 			/*
1133 			 * Mark the page clean.  This will allow the page
1134 			 * to be freed up by the system.  However, such pages
1135 			 * are often reused quickly by malloc()/free()
1136 			 * so we do not do anything that would cause
1137 			 * a page fault if we can help it.
1138 			 *
1139 			 * Specifically, we do not try to actually free
1140 			 * the page now nor do we try to put it in the
1141 			 * cache (which would cause a page fault on reuse).
1142 			 *
1143 			 * But we do make the page is freeable as we
1144 			 * can without actually taking the step of unmapping
1145 			 * it.
1146 			 */
1147 			pmap_clear_modify(m);
1148 			m->dirty = 0;
1149 			m->act_count = 0;
1150 			vm_page_dontneed(m);
1151 		}
1152 		vm_page_unlock(m);
1153 		if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1154 			swap_pager_freespace(tobject, tpindex, 1);
1155 unlock_tobject:
1156 		if (tobject != object)
1157 			VM_OBJECT_UNLOCK(tobject);
1158 	}
1159 	VM_OBJECT_UNLOCK(object);
1160 }
1161 
1162 /*
1163  *	vm_object_shadow:
1164  *
1165  *	Create a new object which is backed by the
1166  *	specified existing object range.  The source
1167  *	object reference is deallocated.
1168  *
1169  *	The new object and offset into that object
1170  *	are returned in the source parameters.
1171  */
1172 void
1173 vm_object_shadow(
1174 	vm_object_t *object,	/* IN/OUT */
1175 	vm_ooffset_t *offset,	/* IN/OUT */
1176 	vm_size_t length)
1177 {
1178 	vm_object_t source;
1179 	vm_object_t result;
1180 
1181 	source = *object;
1182 
1183 	/*
1184 	 * Don't create the new object if the old object isn't shared.
1185 	 */
1186 	if (source != NULL) {
1187 		VM_OBJECT_LOCK(source);
1188 		if (source->ref_count == 1 &&
1189 		    source->handle == NULL &&
1190 		    (source->type == OBJT_DEFAULT ||
1191 		     source->type == OBJT_SWAP)) {
1192 			VM_OBJECT_UNLOCK(source);
1193 			return;
1194 		}
1195 		VM_OBJECT_UNLOCK(source);
1196 	}
1197 
1198 	/*
1199 	 * Allocate a new object with the given length.
1200 	 */
1201 	result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1202 
1203 	/*
1204 	 * The new object shadows the source object, adding a reference to it.
1205 	 * Our caller changes his reference to point to the new object,
1206 	 * removing a reference to the source object.  Net result: no change
1207 	 * of reference count.
1208 	 *
1209 	 * Try to optimize the result object's page color when shadowing
1210 	 * in order to maintain page coloring consistency in the combined
1211 	 * shadowed object.
1212 	 */
1213 	result->backing_object = source;
1214 	/*
1215 	 * Store the offset into the source object, and fix up the offset into
1216 	 * the new object.
1217 	 */
1218 	result->backing_object_offset = *offset;
1219 	if (source != NULL) {
1220 		VM_OBJECT_LOCK(source);
1221 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1222 		source->shadow_count++;
1223 #if VM_NRESERVLEVEL > 0
1224 		result->flags |= source->flags & OBJ_COLORED;
1225 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1226 		    ((1 << (VM_NFREEORDER - 1)) - 1);
1227 #endif
1228 		VM_OBJECT_UNLOCK(source);
1229 	}
1230 
1231 
1232 	/*
1233 	 * Return the new things
1234 	 */
1235 	*offset = 0;
1236 	*object = result;
1237 }
1238 
1239 /*
1240  *	vm_object_split:
1241  *
1242  * Split the pages in a map entry into a new object.  This affords
1243  * easier removal of unused pages, and keeps object inheritance from
1244  * being a negative impact on memory usage.
1245  */
1246 void
1247 vm_object_split(vm_map_entry_t entry)
1248 {
1249 	vm_page_t m, m_next;
1250 	vm_object_t orig_object, new_object, source;
1251 	vm_pindex_t idx, offidxstart;
1252 	vm_size_t size;
1253 
1254 	orig_object = entry->object.vm_object;
1255 	if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1256 		return;
1257 	if (orig_object->ref_count <= 1)
1258 		return;
1259 	VM_OBJECT_UNLOCK(orig_object);
1260 
1261 	offidxstart = OFF_TO_IDX(entry->offset);
1262 	size = atop(entry->end - entry->start);
1263 
1264 	/*
1265 	 * If swap_pager_copy() is later called, it will convert new_object
1266 	 * into a swap object.
1267 	 */
1268 	new_object = vm_object_allocate(OBJT_DEFAULT, size);
1269 
1270 	/*
1271 	 * At this point, the new object is still private, so the order in
1272 	 * which the original and new objects are locked does not matter.
1273 	 */
1274 	VM_OBJECT_LOCK(new_object);
1275 	VM_OBJECT_LOCK(orig_object);
1276 	source = orig_object->backing_object;
1277 	if (source != NULL) {
1278 		VM_OBJECT_LOCK(source);
1279 		if ((source->flags & OBJ_DEAD) != 0) {
1280 			VM_OBJECT_UNLOCK(source);
1281 			VM_OBJECT_UNLOCK(orig_object);
1282 			VM_OBJECT_UNLOCK(new_object);
1283 			vm_object_deallocate(new_object);
1284 			VM_OBJECT_LOCK(orig_object);
1285 			return;
1286 		}
1287 		LIST_INSERT_HEAD(&source->shadow_head,
1288 				  new_object, shadow_list);
1289 		source->shadow_count++;
1290 		vm_object_reference_locked(source);	/* for new_object */
1291 		vm_object_clear_flag(source, OBJ_ONEMAPPING);
1292 		VM_OBJECT_UNLOCK(source);
1293 		new_object->backing_object_offset =
1294 			orig_object->backing_object_offset + entry->offset;
1295 		new_object->backing_object = source;
1296 	}
1297 	if (orig_object->cred != NULL) {
1298 		new_object->cred = orig_object->cred;
1299 		crhold(orig_object->cred);
1300 		new_object->charge = ptoa(size);
1301 		KASSERT(orig_object->charge >= ptoa(size),
1302 		    ("orig_object->charge < 0"));
1303 		orig_object->charge -= ptoa(size);
1304 	}
1305 retry:
1306 	m = vm_page_find_least(orig_object, offidxstart);
1307 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1308 	    m = m_next) {
1309 		m_next = TAILQ_NEXT(m, listq);
1310 
1311 		/*
1312 		 * We must wait for pending I/O to complete before we can
1313 		 * rename the page.
1314 		 *
1315 		 * We do not have to VM_PROT_NONE the page as mappings should
1316 		 * not be changed by this operation.
1317 		 */
1318 		if ((m->oflags & VPO_BUSY) || m->busy) {
1319 			VM_OBJECT_UNLOCK(new_object);
1320 			m->oflags |= VPO_WANTED;
1321 			msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1322 			VM_OBJECT_LOCK(new_object);
1323 			goto retry;
1324 		}
1325 #if VM_NRESERVLEVEL > 0
1326 		/*
1327 		 * If some of the reservation's allocated pages remain with
1328 		 * the original object, then transferring the reservation to
1329 		 * the new object is neither particularly beneficial nor
1330 		 * particularly harmful as compared to leaving the reservation
1331 		 * with the original object.  If, however, all of the
1332 		 * reservation's allocated pages are transferred to the new
1333 		 * object, then transferring the reservation is typically
1334 		 * beneficial.  Determining which of these two cases applies
1335 		 * would be more costly than unconditionally renaming the
1336 		 * reservation.
1337 		 */
1338 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1339 #endif
1340 		vm_page_lock(m);
1341 		vm_page_rename(m, new_object, idx);
1342 		vm_page_unlock(m);
1343 		/* page automatically made dirty by rename and cache handled */
1344 		vm_page_busy(m);
1345 	}
1346 	if (orig_object->type == OBJT_SWAP) {
1347 		/*
1348 		 * swap_pager_copy() can sleep, in which case the orig_object's
1349 		 * and new_object's locks are released and reacquired.
1350 		 */
1351 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1352 
1353 		/*
1354 		 * Transfer any cached pages from orig_object to new_object.
1355 		 * If swap_pager_copy() found swapped out pages within the
1356 		 * specified range of orig_object, then it changed
1357 		 * new_object's type to OBJT_SWAP when it transferred those
1358 		 * pages to new_object.  Otherwise, new_object's type
1359 		 * should still be OBJT_DEFAULT and orig_object should not
1360 		 * contain any cached pages within the specified range.
1361 		 */
1362 		if (__predict_false(orig_object->cache != NULL))
1363 			vm_page_cache_transfer(orig_object, offidxstart,
1364 			    new_object);
1365 	}
1366 	VM_OBJECT_UNLOCK(orig_object);
1367 	TAILQ_FOREACH(m, &new_object->memq, listq)
1368 		vm_page_wakeup(m);
1369 	VM_OBJECT_UNLOCK(new_object);
1370 	entry->object.vm_object = new_object;
1371 	entry->offset = 0LL;
1372 	vm_object_deallocate(orig_object);
1373 	VM_OBJECT_LOCK(new_object);
1374 }
1375 
1376 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1377 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1378 #define	OBSC_COLLAPSE_WAIT	0x0004
1379 
1380 static int
1381 vm_object_backing_scan(vm_object_t object, int op)
1382 {
1383 	int r = 1;
1384 	vm_page_t p;
1385 	vm_object_t backing_object;
1386 	vm_pindex_t backing_offset_index;
1387 
1388 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1389 	VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1390 
1391 	backing_object = object->backing_object;
1392 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1393 
1394 	/*
1395 	 * Initial conditions
1396 	 */
1397 	if (op & OBSC_TEST_ALL_SHADOWED) {
1398 		/*
1399 		 * We do not want to have to test for the existence of cache
1400 		 * or swap pages in the backing object.  XXX but with the
1401 		 * new swapper this would be pretty easy to do.
1402 		 *
1403 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1404 		 * been ZFOD faulted yet?  If we do not test for this, the
1405 		 * shadow test may succeed! XXX
1406 		 */
1407 		if (backing_object->type != OBJT_DEFAULT) {
1408 			return (0);
1409 		}
1410 	}
1411 	if (op & OBSC_COLLAPSE_WAIT) {
1412 		vm_object_set_flag(backing_object, OBJ_DEAD);
1413 	}
1414 
1415 	/*
1416 	 * Our scan
1417 	 */
1418 	p = TAILQ_FIRST(&backing_object->memq);
1419 	while (p) {
1420 		vm_page_t next = TAILQ_NEXT(p, listq);
1421 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1422 
1423 		if (op & OBSC_TEST_ALL_SHADOWED) {
1424 			vm_page_t pp;
1425 
1426 			/*
1427 			 * Ignore pages outside the parent object's range
1428 			 * and outside the parent object's mapping of the
1429 			 * backing object.
1430 			 *
1431 			 * note that we do not busy the backing object's
1432 			 * page.
1433 			 */
1434 			if (
1435 			    p->pindex < backing_offset_index ||
1436 			    new_pindex >= object->size
1437 			) {
1438 				p = next;
1439 				continue;
1440 			}
1441 
1442 			/*
1443 			 * See if the parent has the page or if the parent's
1444 			 * object pager has the page.  If the parent has the
1445 			 * page but the page is not valid, the parent's
1446 			 * object pager must have the page.
1447 			 *
1448 			 * If this fails, the parent does not completely shadow
1449 			 * the object and we might as well give up now.
1450 			 */
1451 
1452 			pp = vm_page_lookup(object, new_pindex);
1453 			if (
1454 			    (pp == NULL || pp->valid == 0) &&
1455 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1456 			) {
1457 				r = 0;
1458 				break;
1459 			}
1460 		}
1461 
1462 		/*
1463 		 * Check for busy page
1464 		 */
1465 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1466 			vm_page_t pp;
1467 
1468 			if (op & OBSC_COLLAPSE_NOWAIT) {
1469 				if ((p->oflags & VPO_BUSY) ||
1470 				    !p->valid ||
1471 				    p->busy) {
1472 					p = next;
1473 					continue;
1474 				}
1475 			} else if (op & OBSC_COLLAPSE_WAIT) {
1476 				if ((p->oflags & VPO_BUSY) || p->busy) {
1477 					VM_OBJECT_UNLOCK(object);
1478 					p->oflags |= VPO_WANTED;
1479 					msleep(p, VM_OBJECT_MTX(backing_object),
1480 					    PDROP | PVM, "vmocol", 0);
1481 					VM_OBJECT_LOCK(object);
1482 					VM_OBJECT_LOCK(backing_object);
1483 					/*
1484 					 * If we slept, anything could have
1485 					 * happened.  Since the object is
1486 					 * marked dead, the backing offset
1487 					 * should not have changed so we
1488 					 * just restart our scan.
1489 					 */
1490 					p = TAILQ_FIRST(&backing_object->memq);
1491 					continue;
1492 				}
1493 			}
1494 
1495 			KASSERT(
1496 			    p->object == backing_object,
1497 			    ("vm_object_backing_scan: object mismatch")
1498 			);
1499 
1500 			/*
1501 			 * Destroy any associated swap
1502 			 */
1503 			if (backing_object->type == OBJT_SWAP) {
1504 				swap_pager_freespace(
1505 				    backing_object,
1506 				    p->pindex,
1507 				    1
1508 				);
1509 			}
1510 
1511 			if (
1512 			    p->pindex < backing_offset_index ||
1513 			    new_pindex >= object->size
1514 			) {
1515 				/*
1516 				 * Page is out of the parent object's range, we
1517 				 * can simply destroy it.
1518 				 */
1519 				vm_page_lock(p);
1520 				KASSERT(!pmap_page_is_mapped(p),
1521 				    ("freeing mapped page %p", p));
1522 				if (p->wire_count == 0)
1523 					vm_page_free(p);
1524 				else
1525 					vm_page_remove(p);
1526 				vm_page_unlock(p);
1527 				p = next;
1528 				continue;
1529 			}
1530 
1531 			pp = vm_page_lookup(object, new_pindex);
1532 			if (
1533 			    (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1534 			    (pp != NULL && pp->valid == 0)
1535 			) {
1536 				/*
1537 				 * The page in the parent is not (yet) valid.
1538 				 * We don't know anything about the state of
1539 				 * the original page.  It might be mapped,
1540 				 * so we must avoid the next if here.
1541 				 *
1542 				 * This is due to a race in vm_fault() where
1543 				 * we must unbusy the original (backing_obj)
1544 				 * page before we can (re)lock the parent.
1545 				 * Hence we can get here.
1546 				 */
1547 				p = next;
1548 				continue;
1549 			}
1550 			if (
1551 			    pp != NULL ||
1552 			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1553 			) {
1554 				/*
1555 				 * page already exists in parent OR swap exists
1556 				 * for this location in the parent.  Destroy
1557 				 * the original page from the backing object.
1558 				 *
1559 				 * Leave the parent's page alone
1560 				 */
1561 				vm_page_lock(p);
1562 				KASSERT(!pmap_page_is_mapped(p),
1563 				    ("freeing mapped page %p", p));
1564 				if (p->wire_count == 0)
1565 					vm_page_free(p);
1566 				else
1567 					vm_page_remove(p);
1568 				vm_page_unlock(p);
1569 				p = next;
1570 				continue;
1571 			}
1572 
1573 #if VM_NRESERVLEVEL > 0
1574 			/*
1575 			 * Rename the reservation.
1576 			 */
1577 			vm_reserv_rename(p, object, backing_object,
1578 			    backing_offset_index);
1579 #endif
1580 
1581 			/*
1582 			 * Page does not exist in parent, rename the
1583 			 * page from the backing object to the main object.
1584 			 *
1585 			 * If the page was mapped to a process, it can remain
1586 			 * mapped through the rename.
1587 			 */
1588 			vm_page_lock(p);
1589 			vm_page_rename(p, object, new_pindex);
1590 			vm_page_unlock(p);
1591 			/* page automatically made dirty by rename */
1592 		}
1593 		p = next;
1594 	}
1595 	return (r);
1596 }
1597 
1598 
1599 /*
1600  * this version of collapse allows the operation to occur earlier and
1601  * when paging_in_progress is true for an object...  This is not a complete
1602  * operation, but should plug 99.9% of the rest of the leaks.
1603  */
1604 static void
1605 vm_object_qcollapse(vm_object_t object)
1606 {
1607 	vm_object_t backing_object = object->backing_object;
1608 
1609 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1610 	VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1611 
1612 	if (backing_object->ref_count != 1)
1613 		return;
1614 
1615 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1616 }
1617 
1618 /*
1619  *	vm_object_collapse:
1620  *
1621  *	Collapse an object with the object backing it.
1622  *	Pages in the backing object are moved into the
1623  *	parent, and the backing object is deallocated.
1624  */
1625 void
1626 vm_object_collapse(vm_object_t object)
1627 {
1628 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1629 
1630 	while (TRUE) {
1631 		vm_object_t backing_object;
1632 
1633 		/*
1634 		 * Verify that the conditions are right for collapse:
1635 		 *
1636 		 * The object exists and the backing object exists.
1637 		 */
1638 		if ((backing_object = object->backing_object) == NULL)
1639 			break;
1640 
1641 		/*
1642 		 * we check the backing object first, because it is most likely
1643 		 * not collapsable.
1644 		 */
1645 		VM_OBJECT_LOCK(backing_object);
1646 		if (backing_object->handle != NULL ||
1647 		    (backing_object->type != OBJT_DEFAULT &&
1648 		     backing_object->type != OBJT_SWAP) ||
1649 		    (backing_object->flags & OBJ_DEAD) ||
1650 		    object->handle != NULL ||
1651 		    (object->type != OBJT_DEFAULT &&
1652 		     object->type != OBJT_SWAP) ||
1653 		    (object->flags & OBJ_DEAD)) {
1654 			VM_OBJECT_UNLOCK(backing_object);
1655 			break;
1656 		}
1657 
1658 		if (
1659 		    object->paging_in_progress != 0 ||
1660 		    backing_object->paging_in_progress != 0
1661 		) {
1662 			vm_object_qcollapse(object);
1663 			VM_OBJECT_UNLOCK(backing_object);
1664 			break;
1665 		}
1666 		/*
1667 		 * We know that we can either collapse the backing object (if
1668 		 * the parent is the only reference to it) or (perhaps) have
1669 		 * the parent bypass the object if the parent happens to shadow
1670 		 * all the resident pages in the entire backing object.
1671 		 *
1672 		 * This is ignoring pager-backed pages such as swap pages.
1673 		 * vm_object_backing_scan fails the shadowing test in this
1674 		 * case.
1675 		 */
1676 		if (backing_object->ref_count == 1) {
1677 			/*
1678 			 * If there is exactly one reference to the backing
1679 			 * object, we can collapse it into the parent.
1680 			 */
1681 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1682 
1683 #if VM_NRESERVLEVEL > 0
1684 			/*
1685 			 * Break any reservations from backing_object.
1686 			 */
1687 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1688 				vm_reserv_break_all(backing_object);
1689 #endif
1690 
1691 			/*
1692 			 * Move the pager from backing_object to object.
1693 			 */
1694 			if (backing_object->type == OBJT_SWAP) {
1695 				/*
1696 				 * swap_pager_copy() can sleep, in which case
1697 				 * the backing_object's and object's locks are
1698 				 * released and reacquired.
1699 				 * Since swap_pager_copy() is being asked to
1700 				 * destroy the source, it will change the
1701 				 * backing_object's type to OBJT_DEFAULT.
1702 				 */
1703 				swap_pager_copy(
1704 				    backing_object,
1705 				    object,
1706 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1707 
1708 				/*
1709 				 * Free any cached pages from backing_object.
1710 				 */
1711 				if (__predict_false(backing_object->cache != NULL))
1712 					vm_page_cache_free(backing_object, 0, 0);
1713 			}
1714 			/*
1715 			 * Object now shadows whatever backing_object did.
1716 			 * Note that the reference to
1717 			 * backing_object->backing_object moves from within
1718 			 * backing_object to within object.
1719 			 */
1720 			LIST_REMOVE(object, shadow_list);
1721 			backing_object->shadow_count--;
1722 			if (backing_object->backing_object) {
1723 				VM_OBJECT_LOCK(backing_object->backing_object);
1724 				LIST_REMOVE(backing_object, shadow_list);
1725 				LIST_INSERT_HEAD(
1726 				    &backing_object->backing_object->shadow_head,
1727 				    object, shadow_list);
1728 				/*
1729 				 * The shadow_count has not changed.
1730 				 */
1731 				VM_OBJECT_UNLOCK(backing_object->backing_object);
1732 			}
1733 			object->backing_object = backing_object->backing_object;
1734 			object->backing_object_offset +=
1735 			    backing_object->backing_object_offset;
1736 
1737 			/*
1738 			 * Discard backing_object.
1739 			 *
1740 			 * Since the backing object has no pages, no pager left,
1741 			 * and no object references within it, all that is
1742 			 * necessary is to dispose of it.
1743 			 */
1744 			KASSERT(backing_object->ref_count == 1, (
1745 "backing_object %p was somehow re-referenced during collapse!",
1746 			    backing_object));
1747 			VM_OBJECT_UNLOCK(backing_object);
1748 			vm_object_destroy(backing_object);
1749 
1750 			object_collapses++;
1751 		} else {
1752 			vm_object_t new_backing_object;
1753 
1754 			/*
1755 			 * If we do not entirely shadow the backing object,
1756 			 * there is nothing we can do so we give up.
1757 			 */
1758 			if (object->resident_page_count != object->size &&
1759 			    vm_object_backing_scan(object,
1760 			    OBSC_TEST_ALL_SHADOWED) == 0) {
1761 				VM_OBJECT_UNLOCK(backing_object);
1762 				break;
1763 			}
1764 
1765 			/*
1766 			 * Make the parent shadow the next object in the
1767 			 * chain.  Deallocating backing_object will not remove
1768 			 * it, since its reference count is at least 2.
1769 			 */
1770 			LIST_REMOVE(object, shadow_list);
1771 			backing_object->shadow_count--;
1772 
1773 			new_backing_object = backing_object->backing_object;
1774 			if ((object->backing_object = new_backing_object) != NULL) {
1775 				VM_OBJECT_LOCK(new_backing_object);
1776 				LIST_INSERT_HEAD(
1777 				    &new_backing_object->shadow_head,
1778 				    object,
1779 				    shadow_list
1780 				);
1781 				new_backing_object->shadow_count++;
1782 				vm_object_reference_locked(new_backing_object);
1783 				VM_OBJECT_UNLOCK(new_backing_object);
1784 				object->backing_object_offset +=
1785 					backing_object->backing_object_offset;
1786 			}
1787 
1788 			/*
1789 			 * Drop the reference count on backing_object. Since
1790 			 * its ref_count was at least 2, it will not vanish.
1791 			 */
1792 			backing_object->ref_count--;
1793 			VM_OBJECT_UNLOCK(backing_object);
1794 			object_bypasses++;
1795 		}
1796 
1797 		/*
1798 		 * Try again with this object's new backing object.
1799 		 */
1800 	}
1801 }
1802 
1803 /*
1804  *	vm_object_page_remove:
1805  *
1806  *	For the given object, either frees or invalidates each of the
1807  *	specified pages.  In general, a page is freed.  However, if a page is
1808  *	wired for any reason other than the existence of a managed, wired
1809  *	mapping, then it may be invalidated but not removed from the object.
1810  *	Pages are specified by the given range ["start", "end") and the option
1811  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
1812  *	extends from "start" to the end of the object.  If the option
1813  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1814  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
1815  *	specified, then the pages within the specified range must have no
1816  *	mappings.  Otherwise, if this option is not specified, any mappings to
1817  *	the specified pages are removed before the pages are freed or
1818  *	invalidated.
1819  *
1820  *	In general, this operation should only be performed on objects that
1821  *	contain managed pages.  There are, however, two exceptions.  First, it
1822  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
1823  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1824  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
1825  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
1826  *
1827  *	The object must be locked.
1828  */
1829 void
1830 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1831     int options)
1832 {
1833 	vm_page_t p, next;
1834 	int wirings;
1835 
1836 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1837 	KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_PHYS) ||
1838 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1839 	    ("vm_object_page_remove: illegal options for object %p", object));
1840 	if (object->resident_page_count == 0)
1841 		goto skipmemq;
1842 	vm_object_pip_add(object, 1);
1843 again:
1844 	p = vm_page_find_least(object, start);
1845 
1846 	/*
1847 	 * Here, the variable "p" is either (1) the page with the least pindex
1848 	 * greater than or equal to the parameter "start" or (2) NULL.
1849 	 */
1850 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1851 		next = TAILQ_NEXT(p, listq);
1852 
1853 		/*
1854 		 * If the page is wired for any reason besides the existence
1855 		 * of managed, wired mappings, then it cannot be freed.  For
1856 		 * example, fictitious pages, which represent device memory,
1857 		 * are inherently wired and cannot be freed.  They can,
1858 		 * however, be invalidated if the option OBJPR_CLEANONLY is
1859 		 * not specified.
1860 		 */
1861 		vm_page_lock(p);
1862 		if ((wirings = p->wire_count) != 0 &&
1863 		    (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1864 			if ((options & OBJPR_NOTMAPPED) == 0) {
1865 				pmap_remove_all(p);
1866 				/* Account for removal of wired mappings. */
1867 				if (wirings != 0)
1868 					p->wire_count -= wirings;
1869 			}
1870 			if ((options & OBJPR_CLEANONLY) == 0) {
1871 				p->valid = 0;
1872 				vm_page_undirty(p);
1873 			}
1874 			vm_page_unlock(p);
1875 			continue;
1876 		}
1877 		if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1878 			goto again;
1879 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
1880 		    ("vm_object_page_remove: page %p is fictitious", p));
1881 		if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1882 			if ((options & OBJPR_NOTMAPPED) == 0)
1883 				pmap_remove_write(p);
1884 			if (p->dirty) {
1885 				vm_page_unlock(p);
1886 				continue;
1887 			}
1888 		}
1889 		if ((options & OBJPR_NOTMAPPED) == 0) {
1890 			pmap_remove_all(p);
1891 			/* Account for removal of wired mappings. */
1892 			if (wirings != 0) {
1893 				KASSERT(p->wire_count == wirings,
1894 				    ("inconsistent wire count %d %d %p",
1895 				    p->wire_count, wirings, p));
1896 				p->wire_count = 0;
1897 				atomic_subtract_int(&cnt.v_wire_count, 1);
1898 			}
1899 		}
1900 		vm_page_free(p);
1901 		vm_page_unlock(p);
1902 	}
1903 	vm_object_pip_wakeup(object);
1904 skipmemq:
1905 	if (__predict_false(object->cache != NULL))
1906 		vm_page_cache_free(object, start, end);
1907 }
1908 
1909 /*
1910  *	vm_object_page_cache:
1911  *
1912  *	For the given object, attempt to move the specified clean
1913  *	pages to the cache queue.  If a page is wired for any reason,
1914  *	then it will not be changed.  Pages are specified by the given
1915  *	range ["start", "end").  As a special case, if "end" is zero,
1916  *	then the range extends from "start" to the end of the object.
1917  *	Any mappings to the specified pages are removed before the
1918  *	pages are moved to the cache queue.
1919  *
1920  *	This operation should only be performed on objects that
1921  *	contain managed pages.
1922  *
1923  *	The object must be locked.
1924  */
1925 void
1926 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1927 {
1928 	struct mtx *mtx, *new_mtx;
1929 	vm_page_t p, next;
1930 
1931 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1932 	KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_SG &&
1933 	    object->type != OBJT_PHYS),
1934 	    ("vm_object_page_cache: illegal object %p", object));
1935 	if (object->resident_page_count == 0)
1936 		return;
1937 	p = vm_page_find_least(object, start);
1938 
1939 	/*
1940 	 * Here, the variable "p" is either (1) the page with the least pindex
1941 	 * greater than or equal to the parameter "start" or (2) NULL.
1942 	 */
1943 	mtx = NULL;
1944 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1945 		next = TAILQ_NEXT(p, listq);
1946 
1947 		/*
1948 		 * Avoid releasing and reacquiring the same page lock.
1949 		 */
1950 		new_mtx = vm_page_lockptr(p);
1951 		if (mtx != new_mtx) {
1952 			if (mtx != NULL)
1953 				mtx_unlock(mtx);
1954 			mtx = new_mtx;
1955 			mtx_lock(mtx);
1956 		}
1957 		vm_page_try_to_cache(p);
1958 	}
1959 	if (mtx != NULL)
1960 		mtx_unlock(mtx);
1961 }
1962 
1963 /*
1964  *	Populate the specified range of the object with valid pages.  Returns
1965  *	TRUE if the range is successfully populated and FALSE otherwise.
1966  *
1967  *	Note: This function should be optimized to pass a larger array of
1968  *	pages to vm_pager_get_pages() before it is applied to a non-
1969  *	OBJT_DEVICE object.
1970  *
1971  *	The object must be locked.
1972  */
1973 boolean_t
1974 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1975 {
1976 	vm_page_t m, ma[1];
1977 	vm_pindex_t pindex;
1978 	int rv;
1979 
1980 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1981 	for (pindex = start; pindex < end; pindex++) {
1982 		m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1983 		    VM_ALLOC_RETRY);
1984 		if (m->valid != VM_PAGE_BITS_ALL) {
1985 			ma[0] = m;
1986 			rv = vm_pager_get_pages(object, ma, 1, 0);
1987 			m = vm_page_lookup(object, pindex);
1988 			if (m == NULL)
1989 				break;
1990 			if (rv != VM_PAGER_OK) {
1991 				vm_page_lock(m);
1992 				vm_page_free(m);
1993 				vm_page_unlock(m);
1994 				break;
1995 			}
1996 		}
1997 		/*
1998 		 * Keep "m" busy because a subsequent iteration may unlock
1999 		 * the object.
2000 		 */
2001 	}
2002 	if (pindex > start) {
2003 		m = vm_page_lookup(object, start);
2004 		while (m != NULL && m->pindex < pindex) {
2005 			vm_page_wakeup(m);
2006 			m = TAILQ_NEXT(m, listq);
2007 		}
2008 	}
2009 	return (pindex == end);
2010 }
2011 
2012 /*
2013  *	Routine:	vm_object_coalesce
2014  *	Function:	Coalesces two objects backing up adjoining
2015  *			regions of memory into a single object.
2016  *
2017  *	returns TRUE if objects were combined.
2018  *
2019  *	NOTE:	Only works at the moment if the second object is NULL -
2020  *		if it's not, which object do we lock first?
2021  *
2022  *	Parameters:
2023  *		prev_object	First object to coalesce
2024  *		prev_offset	Offset into prev_object
2025  *		prev_size	Size of reference to prev_object
2026  *		next_size	Size of reference to the second object
2027  *		reserved	Indicator that extension region has
2028  *				swap accounted for
2029  *
2030  *	Conditions:
2031  *	The object must *not* be locked.
2032  */
2033 boolean_t
2034 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2035     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2036 {
2037 	vm_pindex_t next_pindex;
2038 
2039 	if (prev_object == NULL)
2040 		return (TRUE);
2041 	VM_OBJECT_LOCK(prev_object);
2042 	if (prev_object->type != OBJT_DEFAULT &&
2043 	    prev_object->type != OBJT_SWAP) {
2044 		VM_OBJECT_UNLOCK(prev_object);
2045 		return (FALSE);
2046 	}
2047 
2048 	/*
2049 	 * Try to collapse the object first
2050 	 */
2051 	vm_object_collapse(prev_object);
2052 
2053 	/*
2054 	 * Can't coalesce if: . more than one reference . paged out . shadows
2055 	 * another object . has a copy elsewhere (any of which mean that the
2056 	 * pages not mapped to prev_entry may be in use anyway)
2057 	 */
2058 	if (prev_object->backing_object != NULL) {
2059 		VM_OBJECT_UNLOCK(prev_object);
2060 		return (FALSE);
2061 	}
2062 
2063 	prev_size >>= PAGE_SHIFT;
2064 	next_size >>= PAGE_SHIFT;
2065 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2066 
2067 	if ((prev_object->ref_count > 1) &&
2068 	    (prev_object->size != next_pindex)) {
2069 		VM_OBJECT_UNLOCK(prev_object);
2070 		return (FALSE);
2071 	}
2072 
2073 	/*
2074 	 * Account for the charge.
2075 	 */
2076 	if (prev_object->cred != NULL) {
2077 
2078 		/*
2079 		 * If prev_object was charged, then this mapping,
2080 		 * althought not charged now, may become writable
2081 		 * later. Non-NULL cred in the object would prevent
2082 		 * swap reservation during enabling of the write
2083 		 * access, so reserve swap now. Failed reservation
2084 		 * cause allocation of the separate object for the map
2085 		 * entry, and swap reservation for this entry is
2086 		 * managed in appropriate time.
2087 		 */
2088 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2089 		    prev_object->cred)) {
2090 			return (FALSE);
2091 		}
2092 		prev_object->charge += ptoa(next_size);
2093 	}
2094 
2095 	/*
2096 	 * Remove any pages that may still be in the object from a previous
2097 	 * deallocation.
2098 	 */
2099 	if (next_pindex < prev_object->size) {
2100 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2101 		    next_size, 0);
2102 		if (prev_object->type == OBJT_SWAP)
2103 			swap_pager_freespace(prev_object,
2104 					     next_pindex, next_size);
2105 #if 0
2106 		if (prev_object->cred != NULL) {
2107 			KASSERT(prev_object->charge >=
2108 			    ptoa(prev_object->size - next_pindex),
2109 			    ("object %p overcharged 1 %jx %jx", prev_object,
2110 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2111 			prev_object->charge -= ptoa(prev_object->size -
2112 			    next_pindex);
2113 		}
2114 #endif
2115 	}
2116 
2117 	/*
2118 	 * Extend the object if necessary.
2119 	 */
2120 	if (next_pindex + next_size > prev_object->size)
2121 		prev_object->size = next_pindex + next_size;
2122 
2123 	VM_OBJECT_UNLOCK(prev_object);
2124 	return (TRUE);
2125 }
2126 
2127 void
2128 vm_object_set_writeable_dirty(vm_object_t object)
2129 {
2130 
2131 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2132 	if (object->type != OBJT_VNODE)
2133 		return;
2134 	object->generation++;
2135 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2136 		return;
2137 	vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2138 }
2139 
2140 #include "opt_ddb.h"
2141 #ifdef DDB
2142 #include <sys/kernel.h>
2143 
2144 #include <sys/cons.h>
2145 
2146 #include <ddb/ddb.h>
2147 
2148 static int
2149 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2150 {
2151 	vm_map_t tmpm;
2152 	vm_map_entry_t tmpe;
2153 	vm_object_t obj;
2154 	int entcount;
2155 
2156 	if (map == 0)
2157 		return 0;
2158 
2159 	if (entry == 0) {
2160 		tmpe = map->header.next;
2161 		entcount = map->nentries;
2162 		while (entcount-- && (tmpe != &map->header)) {
2163 			if (_vm_object_in_map(map, object, tmpe)) {
2164 				return 1;
2165 			}
2166 			tmpe = tmpe->next;
2167 		}
2168 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2169 		tmpm = entry->object.sub_map;
2170 		tmpe = tmpm->header.next;
2171 		entcount = tmpm->nentries;
2172 		while (entcount-- && tmpe != &tmpm->header) {
2173 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2174 				return 1;
2175 			}
2176 			tmpe = tmpe->next;
2177 		}
2178 	} else if ((obj = entry->object.vm_object) != NULL) {
2179 		for (; obj; obj = obj->backing_object)
2180 			if (obj == object) {
2181 				return 1;
2182 			}
2183 	}
2184 	return 0;
2185 }
2186 
2187 static int
2188 vm_object_in_map(vm_object_t object)
2189 {
2190 	struct proc *p;
2191 
2192 	/* sx_slock(&allproc_lock); */
2193 	FOREACH_PROC_IN_SYSTEM(p) {
2194 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2195 			continue;
2196 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2197 			/* sx_sunlock(&allproc_lock); */
2198 			return 1;
2199 		}
2200 	}
2201 	/* sx_sunlock(&allproc_lock); */
2202 	if (_vm_object_in_map(kernel_map, object, 0))
2203 		return 1;
2204 	if (_vm_object_in_map(kmem_map, object, 0))
2205 		return 1;
2206 	if (_vm_object_in_map(pager_map, object, 0))
2207 		return 1;
2208 	if (_vm_object_in_map(buffer_map, object, 0))
2209 		return 1;
2210 	return 0;
2211 }
2212 
2213 DB_SHOW_COMMAND(vmochk, vm_object_check)
2214 {
2215 	vm_object_t object;
2216 
2217 	/*
2218 	 * make sure that internal objs are in a map somewhere
2219 	 * and none have zero ref counts.
2220 	 */
2221 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2222 		if (object->handle == NULL &&
2223 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2224 			if (object->ref_count == 0) {
2225 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2226 					(long)object->size);
2227 			}
2228 			if (!vm_object_in_map(object)) {
2229 				db_printf(
2230 			"vmochk: internal obj is not in a map: "
2231 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2232 				    object->ref_count, (u_long)object->size,
2233 				    (u_long)object->size,
2234 				    (void *)object->backing_object);
2235 			}
2236 		}
2237 	}
2238 }
2239 
2240 /*
2241  *	vm_object_print:	[ debug ]
2242  */
2243 DB_SHOW_COMMAND(object, vm_object_print_static)
2244 {
2245 	/* XXX convert args. */
2246 	vm_object_t object = (vm_object_t)addr;
2247 	boolean_t full = have_addr;
2248 
2249 	vm_page_t p;
2250 
2251 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2252 #define	count	was_count
2253 
2254 	int count;
2255 
2256 	if (object == NULL)
2257 		return;
2258 
2259 	db_iprintf(
2260 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2261 	    object, (int)object->type, (uintmax_t)object->size,
2262 	    object->resident_page_count, object->ref_count, object->flags,
2263 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2264 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2265 	    object->shadow_count,
2266 	    object->backing_object ? object->backing_object->ref_count : 0,
2267 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2268 
2269 	if (!full)
2270 		return;
2271 
2272 	db_indent += 2;
2273 	count = 0;
2274 	TAILQ_FOREACH(p, &object->memq, listq) {
2275 		if (count == 0)
2276 			db_iprintf("memory:=");
2277 		else if (count == 6) {
2278 			db_printf("\n");
2279 			db_iprintf(" ...");
2280 			count = 0;
2281 		} else
2282 			db_printf(",");
2283 		count++;
2284 
2285 		db_printf("(off=0x%jx,page=0x%jx)",
2286 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2287 	}
2288 	if (count != 0)
2289 		db_printf("\n");
2290 	db_indent -= 2;
2291 }
2292 
2293 /* XXX. */
2294 #undef count
2295 
2296 /* XXX need this non-static entry for calling from vm_map_print. */
2297 void
2298 vm_object_print(
2299         /* db_expr_t */ long addr,
2300 	boolean_t have_addr,
2301 	/* db_expr_t */ long count,
2302 	char *modif)
2303 {
2304 	vm_object_print_static(addr, have_addr, count, modif);
2305 }
2306 
2307 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2308 {
2309 	vm_object_t object;
2310 	vm_pindex_t fidx;
2311 	vm_paddr_t pa;
2312 	vm_page_t m, prev_m;
2313 	int rcount, nl, c;
2314 
2315 	nl = 0;
2316 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2317 		db_printf("new object: %p\n", (void *)object);
2318 		if (nl > 18) {
2319 			c = cngetc();
2320 			if (c != ' ')
2321 				return;
2322 			nl = 0;
2323 		}
2324 		nl++;
2325 		rcount = 0;
2326 		fidx = 0;
2327 		pa = -1;
2328 		TAILQ_FOREACH(m, &object->memq, listq) {
2329 			if (m->pindex > 128)
2330 				break;
2331 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2332 			    prev_m->pindex + 1 != m->pindex) {
2333 				if (rcount) {
2334 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2335 						(long)fidx, rcount, (long)pa);
2336 					if (nl > 18) {
2337 						c = cngetc();
2338 						if (c != ' ')
2339 							return;
2340 						nl = 0;
2341 					}
2342 					nl++;
2343 					rcount = 0;
2344 				}
2345 			}
2346 			if (rcount &&
2347 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2348 				++rcount;
2349 				continue;
2350 			}
2351 			if (rcount) {
2352 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2353 					(long)fidx, rcount, (long)pa);
2354 				if (nl > 18) {
2355 					c = cngetc();
2356 					if (c != ' ')
2357 						return;
2358 					nl = 0;
2359 				}
2360 				nl++;
2361 			}
2362 			fidx = m->pindex;
2363 			pa = VM_PAGE_TO_PHYS(m);
2364 			rcount = 1;
2365 		}
2366 		if (rcount) {
2367 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2368 				(long)fidx, rcount, (long)pa);
2369 			if (nl > 18) {
2370 				c = cngetc();
2371 				if (c != ' ')
2372 					return;
2373 				nl = 0;
2374 			}
2375 			nl++;
2376 		}
2377 	}
2378 }
2379 #endif /* DDB */
2380