xref: /freebsd/sys/vm/vm_object.c (revision aa64588d28258aef88cc33b8043112e8856948d0)
1 /*-
2  * Copyright (c) 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * The Mach Operating System project at Carnegie-Mellon University.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 4. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	from: @(#)vm_object.c	8.5 (Berkeley) 3/22/94
33  *
34  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39  *
40  * Permission to use, copy, modify and distribute this software and
41  * its documentation is hereby granted, provided that both the copyright
42  * notice and this permission notice appear in all copies of the
43  * software, derivative works or modified versions, and any portions
44  * thereof, and that both notices appear in supporting documentation.
45  *
46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49  *
50  * Carnegie Mellon requests users of this software to return to
51  *
52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
53  *  School of Computer Science
54  *  Carnegie Mellon University
55  *  Pittsburgh PA 15213-3890
56  *
57  * any improvements or extensions that they make and grant Carnegie the
58  * rights to redistribute these changes.
59  */
60 
61 /*
62  *	Virtual memory object module.
63  */
64 
65 #include <sys/cdefs.h>
66 __FBSDID("$FreeBSD$");
67 
68 #include "opt_vm.h"
69 
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/lock.h>
73 #include <sys/mman.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/mutex.h>
78 #include <sys/proc.h>		/* for curproc, pageproc */
79 #include <sys/socket.h>
80 #include <sys/resourcevar.h>
81 #include <sys/vnode.h>
82 #include <sys/vmmeter.h>
83 #include <sys/sx.h>
84 
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_pager.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_reserv.h>
97 #include <vm/uma.h>
98 
99 #define EASY_SCAN_FACTOR       8
100 
101 #define MSYNC_FLUSH_HARDSEQ	0x01
102 #define MSYNC_FLUSH_SOFTSEQ	0x02
103 
104 /*
105  * msync / VM object flushing optimizations
106  */
107 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ;
108 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags, CTLFLAG_RW, &msync_flush_flags, 0,
109     "Enable sequential iteration optimization");
110 
111 static int old_msync;
112 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
113     "Use old (insecure) msync behavior");
114 
115 static void	vm_object_qcollapse(vm_object_t object);
116 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
117 static void	vm_object_vndeallocate(vm_object_t object);
118 
119 /*
120  *	Virtual memory objects maintain the actual data
121  *	associated with allocated virtual memory.  A given
122  *	page of memory exists within exactly one object.
123  *
124  *	An object is only deallocated when all "references"
125  *	are given up.  Only one "reference" to a given
126  *	region of an object should be writeable.
127  *
128  *	Associated with each object is a list of all resident
129  *	memory pages belonging to that object; this list is
130  *	maintained by the "vm_page" module, and locked by the object's
131  *	lock.
132  *
133  *	Each object also records a "pager" routine which is
134  *	used to retrieve (and store) pages to the proper backing
135  *	storage.  In addition, objects may be backed by other
136  *	objects from which they were virtual-copied.
137  *
138  *	The only items within the object structure which are
139  *	modified after time of creation are:
140  *		reference count		locked by object's lock
141  *		pager routine		locked by object's lock
142  *
143  */
144 
145 struct object_q vm_object_list;
146 struct mtx vm_object_list_mtx;	/* lock for object list and count */
147 
148 struct vm_object kernel_object_store;
149 struct vm_object kmem_object_store;
150 
151 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats");
152 
153 static long object_collapses;
154 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
155     &object_collapses, 0, "VM object collapses");
156 
157 static long object_bypasses;
158 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
159     &object_bypasses, 0, "VM object bypasses");
160 
161 static uma_zone_t obj_zone;
162 
163 static int vm_object_zinit(void *mem, int size, int flags);
164 
165 #ifdef INVARIANTS
166 static void vm_object_zdtor(void *mem, int size, void *arg);
167 
168 static void
169 vm_object_zdtor(void *mem, int size, void *arg)
170 {
171 	vm_object_t object;
172 
173 	object = (vm_object_t)mem;
174 	KASSERT(TAILQ_EMPTY(&object->memq),
175 	    ("object %p has resident pages",
176 	    object));
177 #if VM_NRESERVLEVEL > 0
178 	KASSERT(LIST_EMPTY(&object->rvq),
179 	    ("object %p has reservations",
180 	    object));
181 #endif
182 	KASSERT(object->cache == NULL,
183 	    ("object %p has cached pages",
184 	    object));
185 	KASSERT(object->paging_in_progress == 0,
186 	    ("object %p paging_in_progress = %d",
187 	    object, object->paging_in_progress));
188 	KASSERT(object->resident_page_count == 0,
189 	    ("object %p resident_page_count = %d",
190 	    object, object->resident_page_count));
191 	KASSERT(object->shadow_count == 0,
192 	    ("object %p shadow_count = %d",
193 	    object, object->shadow_count));
194 }
195 #endif
196 
197 static int
198 vm_object_zinit(void *mem, int size, int flags)
199 {
200 	vm_object_t object;
201 
202 	object = (vm_object_t)mem;
203 	bzero(&object->mtx, sizeof(object->mtx));
204 	VM_OBJECT_LOCK_INIT(object, "standard object");
205 
206 	/* These are true for any object that has been freed */
207 	object->paging_in_progress = 0;
208 	object->resident_page_count = 0;
209 	object->shadow_count = 0;
210 	return (0);
211 }
212 
213 void
214 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
215 {
216 
217 	TAILQ_INIT(&object->memq);
218 	LIST_INIT(&object->shadow_head);
219 
220 	object->root = NULL;
221 	object->type = type;
222 	object->size = size;
223 	object->generation = 1;
224 	object->ref_count = 1;
225 	object->memattr = VM_MEMATTR_DEFAULT;
226 	object->flags = 0;
227 	object->uip = NULL;
228 	object->charge = 0;
229 	if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
230 		object->flags = OBJ_ONEMAPPING;
231 	object->pg_color = 0;
232 	object->handle = NULL;
233 	object->backing_object = NULL;
234 	object->backing_object_offset = (vm_ooffset_t) 0;
235 #if VM_NRESERVLEVEL > 0
236 	LIST_INIT(&object->rvq);
237 #endif
238 	object->cache = NULL;
239 
240 	mtx_lock(&vm_object_list_mtx);
241 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
242 	mtx_unlock(&vm_object_list_mtx);
243 }
244 
245 /*
246  *	vm_object_init:
247  *
248  *	Initialize the VM objects module.
249  */
250 void
251 vm_object_init(void)
252 {
253 	TAILQ_INIT(&vm_object_list);
254 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
255 
256 	VM_OBJECT_LOCK_INIT(&kernel_object_store, "kernel object");
257 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
258 	    kernel_object);
259 #if VM_NRESERVLEVEL > 0
260 	kernel_object->flags |= OBJ_COLORED;
261 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
262 #endif
263 
264 	VM_OBJECT_LOCK_INIT(&kmem_object_store, "kmem object");
265 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
266 	    kmem_object);
267 #if VM_NRESERVLEVEL > 0
268 	kmem_object->flags |= OBJ_COLORED;
269 	kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
270 #endif
271 
272 	/*
273 	 * The lock portion of struct vm_object must be type stable due
274 	 * to vm_pageout_fallback_object_lock locking a vm object
275 	 * without holding any references to it.
276 	 */
277 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
278 #ifdef INVARIANTS
279 	    vm_object_zdtor,
280 #else
281 	    NULL,
282 #endif
283 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
284 }
285 
286 void
287 vm_object_clear_flag(vm_object_t object, u_short bits)
288 {
289 
290 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
291 	object->flags &= ~bits;
292 }
293 
294 /*
295  *	Sets the default memory attribute for the specified object.  Pages
296  *	that are allocated to this object are by default assigned this memory
297  *	attribute.
298  *
299  *	Presently, this function must be called before any pages are allocated
300  *	to the object.  In the future, this requirement may be relaxed for
301  *	"default" and "swap" objects.
302  */
303 int
304 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
305 {
306 
307 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
308 	switch (object->type) {
309 	case OBJT_DEFAULT:
310 	case OBJT_DEVICE:
311 	case OBJT_PHYS:
312 	case OBJT_SG:
313 	case OBJT_SWAP:
314 	case OBJT_VNODE:
315 		if (!TAILQ_EMPTY(&object->memq))
316 			return (KERN_FAILURE);
317 		break;
318 	case OBJT_DEAD:
319 		return (KERN_INVALID_ARGUMENT);
320 	}
321 	object->memattr = memattr;
322 	return (KERN_SUCCESS);
323 }
324 
325 void
326 vm_object_pip_add(vm_object_t object, short i)
327 {
328 
329 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
330 	object->paging_in_progress += i;
331 }
332 
333 void
334 vm_object_pip_subtract(vm_object_t object, short i)
335 {
336 
337 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
338 	object->paging_in_progress -= i;
339 }
340 
341 void
342 vm_object_pip_wakeup(vm_object_t object)
343 {
344 
345 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
346 	object->paging_in_progress--;
347 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
348 		vm_object_clear_flag(object, OBJ_PIPWNT);
349 		wakeup(object);
350 	}
351 }
352 
353 void
354 vm_object_pip_wakeupn(vm_object_t object, short i)
355 {
356 
357 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
358 	if (i)
359 		object->paging_in_progress -= i;
360 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
361 		vm_object_clear_flag(object, OBJ_PIPWNT);
362 		wakeup(object);
363 	}
364 }
365 
366 void
367 vm_object_pip_wait(vm_object_t object, char *waitid)
368 {
369 
370 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
371 	while (object->paging_in_progress) {
372 		object->flags |= OBJ_PIPWNT;
373 		msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
374 	}
375 }
376 
377 /*
378  *	vm_object_allocate:
379  *
380  *	Returns a new object with the given size.
381  */
382 vm_object_t
383 vm_object_allocate(objtype_t type, vm_pindex_t size)
384 {
385 	vm_object_t object;
386 
387 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
388 	_vm_object_allocate(type, size, object);
389 	return (object);
390 }
391 
392 
393 /*
394  *	vm_object_reference:
395  *
396  *	Gets another reference to the given object.  Note: OBJ_DEAD
397  *	objects can be referenced during final cleaning.
398  */
399 void
400 vm_object_reference(vm_object_t object)
401 {
402 	if (object == NULL)
403 		return;
404 	VM_OBJECT_LOCK(object);
405 	vm_object_reference_locked(object);
406 	VM_OBJECT_UNLOCK(object);
407 }
408 
409 /*
410  *	vm_object_reference_locked:
411  *
412  *	Gets another reference to the given object.
413  *
414  *	The object must be locked.
415  */
416 void
417 vm_object_reference_locked(vm_object_t object)
418 {
419 	struct vnode *vp;
420 
421 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
422 	object->ref_count++;
423 	if (object->type == OBJT_VNODE) {
424 		vp = object->handle;
425 		vref(vp);
426 	}
427 }
428 
429 /*
430  * Handle deallocating an object of type OBJT_VNODE.
431  */
432 static void
433 vm_object_vndeallocate(vm_object_t object)
434 {
435 	struct vnode *vp = (struct vnode *) object->handle;
436 
437 	VFS_ASSERT_GIANT(vp->v_mount);
438 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
439 	KASSERT(object->type == OBJT_VNODE,
440 	    ("vm_object_vndeallocate: not a vnode object"));
441 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
442 #ifdef INVARIANTS
443 	if (object->ref_count == 0) {
444 		vprint("vm_object_vndeallocate", vp);
445 		panic("vm_object_vndeallocate: bad object reference count");
446 	}
447 #endif
448 
449 	object->ref_count--;
450 	if (object->ref_count == 0) {
451 		mp_fixme("Unlocked vflag access.");
452 		vp->v_vflag &= ~VV_TEXT;
453 	}
454 	VM_OBJECT_UNLOCK(object);
455 	/*
456 	 * vrele may need a vop lock
457 	 */
458 	vrele(vp);
459 }
460 
461 /*
462  *	vm_object_deallocate:
463  *
464  *	Release a reference to the specified object,
465  *	gained either through a vm_object_allocate
466  *	or a vm_object_reference call.  When all references
467  *	are gone, storage associated with this object
468  *	may be relinquished.
469  *
470  *	No object may be locked.
471  */
472 void
473 vm_object_deallocate(vm_object_t object)
474 {
475 	vm_object_t temp;
476 
477 	while (object != NULL) {
478 		int vfslocked;
479 
480 		vfslocked = 0;
481 	restart:
482 		VM_OBJECT_LOCK(object);
483 		if (object->type == OBJT_VNODE) {
484 			struct vnode *vp = (struct vnode *) object->handle;
485 
486 			/*
487 			 * Conditionally acquire Giant for a vnode-backed
488 			 * object.  We have to be careful since the type of
489 			 * a vnode object can change while the object is
490 			 * unlocked.
491 			 */
492 			if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
493 				vfslocked = 1;
494 				if (!mtx_trylock(&Giant)) {
495 					VM_OBJECT_UNLOCK(object);
496 					mtx_lock(&Giant);
497 					goto restart;
498 				}
499 			}
500 			vm_object_vndeallocate(object);
501 			VFS_UNLOCK_GIANT(vfslocked);
502 			return;
503 		} else
504 			/*
505 			 * This is to handle the case that the object
506 			 * changed type while we dropped its lock to
507 			 * obtain Giant.
508 			 */
509 			VFS_UNLOCK_GIANT(vfslocked);
510 
511 		KASSERT(object->ref_count != 0,
512 			("vm_object_deallocate: object deallocated too many times: %d", object->type));
513 
514 		/*
515 		 * If the reference count goes to 0 we start calling
516 		 * vm_object_terminate() on the object chain.
517 		 * A ref count of 1 may be a special case depending on the
518 		 * shadow count being 0 or 1.
519 		 */
520 		object->ref_count--;
521 		if (object->ref_count > 1) {
522 			VM_OBJECT_UNLOCK(object);
523 			return;
524 		} else if (object->ref_count == 1) {
525 			if (object->shadow_count == 0 &&
526 			    object->handle == NULL &&
527 			    (object->type == OBJT_DEFAULT ||
528 			     object->type == OBJT_SWAP)) {
529 				vm_object_set_flag(object, OBJ_ONEMAPPING);
530 			} else if ((object->shadow_count == 1) &&
531 			    (object->handle == NULL) &&
532 			    (object->type == OBJT_DEFAULT ||
533 			     object->type == OBJT_SWAP)) {
534 				vm_object_t robject;
535 
536 				robject = LIST_FIRST(&object->shadow_head);
537 				KASSERT(robject != NULL,
538 				    ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
539 					 object->ref_count,
540 					 object->shadow_count));
541 				if (!VM_OBJECT_TRYLOCK(robject)) {
542 					/*
543 					 * Avoid a potential deadlock.
544 					 */
545 					object->ref_count++;
546 					VM_OBJECT_UNLOCK(object);
547 					/*
548 					 * More likely than not the thread
549 					 * holding robject's lock has lower
550 					 * priority than the current thread.
551 					 * Let the lower priority thread run.
552 					 */
553 					pause("vmo_de", 1);
554 					continue;
555 				}
556 				/*
557 				 * Collapse object into its shadow unless its
558 				 * shadow is dead.  In that case, object will
559 				 * be deallocated by the thread that is
560 				 * deallocating its shadow.
561 				 */
562 				if ((robject->flags & OBJ_DEAD) == 0 &&
563 				    (robject->handle == NULL) &&
564 				    (robject->type == OBJT_DEFAULT ||
565 				     robject->type == OBJT_SWAP)) {
566 
567 					robject->ref_count++;
568 retry:
569 					if (robject->paging_in_progress) {
570 						VM_OBJECT_UNLOCK(object);
571 						vm_object_pip_wait(robject,
572 						    "objde1");
573 						temp = robject->backing_object;
574 						if (object == temp) {
575 							VM_OBJECT_LOCK(object);
576 							goto retry;
577 						}
578 					} else if (object->paging_in_progress) {
579 						VM_OBJECT_UNLOCK(robject);
580 						object->flags |= OBJ_PIPWNT;
581 						msleep(object,
582 						    VM_OBJECT_MTX(object),
583 						    PDROP | PVM, "objde2", 0);
584 						VM_OBJECT_LOCK(robject);
585 						temp = robject->backing_object;
586 						if (object == temp) {
587 							VM_OBJECT_LOCK(object);
588 							goto retry;
589 						}
590 					} else
591 						VM_OBJECT_UNLOCK(object);
592 
593 					if (robject->ref_count == 1) {
594 						robject->ref_count--;
595 						object = robject;
596 						goto doterm;
597 					}
598 					object = robject;
599 					vm_object_collapse(object);
600 					VM_OBJECT_UNLOCK(object);
601 					continue;
602 				}
603 				VM_OBJECT_UNLOCK(robject);
604 			}
605 			VM_OBJECT_UNLOCK(object);
606 			return;
607 		}
608 doterm:
609 		temp = object->backing_object;
610 		if (temp != NULL) {
611 			VM_OBJECT_LOCK(temp);
612 			LIST_REMOVE(object, shadow_list);
613 			temp->shadow_count--;
614 			temp->generation++;
615 			VM_OBJECT_UNLOCK(temp);
616 			object->backing_object = NULL;
617 		}
618 		/*
619 		 * Don't double-terminate, we could be in a termination
620 		 * recursion due to the terminate having to sync data
621 		 * to disk.
622 		 */
623 		if ((object->flags & OBJ_DEAD) == 0)
624 			vm_object_terminate(object);
625 		else
626 			VM_OBJECT_UNLOCK(object);
627 		object = temp;
628 	}
629 }
630 
631 /*
632  *	vm_object_destroy removes the object from the global object list
633  *      and frees the space for the object.
634  */
635 void
636 vm_object_destroy(vm_object_t object)
637 {
638 
639 	/*
640 	 * Remove the object from the global object list.
641 	 */
642 	mtx_lock(&vm_object_list_mtx);
643 	TAILQ_REMOVE(&vm_object_list, object, object_list);
644 	mtx_unlock(&vm_object_list_mtx);
645 
646 	/*
647 	 * Release the allocation charge.
648 	 */
649 	if (object->uip != NULL) {
650 		KASSERT(object->type == OBJT_DEFAULT ||
651 		    object->type == OBJT_SWAP,
652 		    ("vm_object_terminate: non-swap obj %p has uip",
653 		     object));
654 		swap_release_by_uid(object->charge, object->uip);
655 		object->charge = 0;
656 		uifree(object->uip);
657 		object->uip = NULL;
658 	}
659 
660 	/*
661 	 * Free the space for the object.
662 	 */
663 	uma_zfree(obj_zone, object);
664 }
665 
666 /*
667  *	vm_object_terminate actually destroys the specified object, freeing
668  *	up all previously used resources.
669  *
670  *	The object must be locked.
671  *	This routine may block.
672  */
673 void
674 vm_object_terminate(vm_object_t object)
675 {
676 	vm_page_t p;
677 
678 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
679 
680 	/*
681 	 * Make sure no one uses us.
682 	 */
683 	vm_object_set_flag(object, OBJ_DEAD);
684 
685 	/*
686 	 * wait for the pageout daemon to be done with the object
687 	 */
688 	vm_object_pip_wait(object, "objtrm");
689 
690 	KASSERT(!object->paging_in_progress,
691 		("vm_object_terminate: pageout in progress"));
692 
693 	/*
694 	 * Clean and free the pages, as appropriate. All references to the
695 	 * object are gone, so we don't need to lock it.
696 	 */
697 	if (object->type == OBJT_VNODE) {
698 		struct vnode *vp = (struct vnode *)object->handle;
699 
700 		/*
701 		 * Clean pages and flush buffers.
702 		 */
703 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
704 		VM_OBJECT_UNLOCK(object);
705 
706 		vinvalbuf(vp, V_SAVE, 0, 0);
707 
708 		VM_OBJECT_LOCK(object);
709 	}
710 
711 	KASSERT(object->ref_count == 0,
712 		("vm_object_terminate: object with references, ref_count=%d",
713 		object->ref_count));
714 
715 	/*
716 	 * Now free any remaining pages. For internal objects, this also
717 	 * removes them from paging queues. Don't free wired pages, just
718 	 * remove them from the object.
719 	 */
720 	while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
721 		KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
722 			("vm_object_terminate: freeing busy page %p "
723 			"p->busy = %d, p->oflags %x\n", p, p->busy, p->oflags));
724 		vm_page_lock(p);
725 		if (p->wire_count == 0) {
726 			vm_page_free(p);
727 			PCPU_INC(cnt.v_pfree);
728 		} else
729 			vm_page_remove(p);
730 		vm_page_unlock(p);
731 	}
732 
733 #if VM_NRESERVLEVEL > 0
734 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
735 		vm_reserv_break_all(object);
736 #endif
737 	if (__predict_false(object->cache != NULL))
738 		vm_page_cache_free(object, 0, 0);
739 
740 	/*
741 	 * Let the pager know object is dead.
742 	 */
743 	vm_pager_deallocate(object);
744 	VM_OBJECT_UNLOCK(object);
745 
746 	vm_object_destroy(object);
747 }
748 
749 /*
750  *	vm_object_page_clean
751  *
752  *	Clean all dirty pages in the specified range of object.  Leaves page
753  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
754  *	write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
755  *	leaving the object dirty.
756  *
757  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
758  *	synchronous clustering mode implementation.
759  *
760  *	Odd semantics: if start == end, we clean everything.
761  *
762  *	The object must be locked.
763  */
764 void
765 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags)
766 {
767 	vm_page_t p, np;
768 	vm_pindex_t tstart, tend;
769 	vm_pindex_t pi;
770 	int clearobjflags;
771 	int pagerflags;
772 	int curgeneration;
773 
774 	mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
775 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
776 	if ((object->flags & OBJ_MIGHTBEDIRTY) == 0)
777 		return;
778 	KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
779 
780 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
781 	pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
782 
783 	vm_object_set_flag(object, OBJ_CLEANING);
784 
785 	tstart = start;
786 	if (end == 0) {
787 		tend = object->size;
788 	} else {
789 		tend = end;
790 	}
791 
792 	/*
793 	 * If the caller is smart and only msync()s a range he knows is
794 	 * dirty, we may be able to avoid an object scan.  This results in
795 	 * a phenominal improvement in performance.  We cannot do this
796 	 * as a matter of course because the object may be huge - e.g.
797 	 * the size might be in the gigabytes or terrabytes.
798 	 */
799 	if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
800 		vm_pindex_t tscan;
801 		int scanlimit;
802 		int scanreset;
803 
804 		scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
805 		if (scanreset < 16)
806 			scanreset = 16;
807 		pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
808 
809 		scanlimit = scanreset;
810 		tscan = tstart;
811 		while (tscan < tend) {
812 			curgeneration = object->generation;
813 			p = vm_page_lookup(object, tscan);
814 			if (p == NULL || p->valid == 0) {
815 				if (--scanlimit == 0)
816 					break;
817 				++tscan;
818 				continue;
819 			}
820 			vm_page_test_dirty(p);
821 			if (p->dirty == 0) {
822 				if (--scanlimit == 0)
823 					break;
824 				++tscan;
825 				continue;
826 			}
827 			/*
828 			 * If we have been asked to skip nosync pages and
829 			 * this is a nosync page, we can't continue.
830 			 */
831 			if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
832 				if (--scanlimit == 0)
833 					break;
834 				++tscan;
835 				continue;
836 			}
837 			scanlimit = scanreset;
838 
839 			/*
840 			 * This returns 0 if it was unable to busy the first
841 			 * page (i.e. had to sleep).
842 			 */
843 			tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
844 
845 		}
846 
847 		/*
848 		 * If everything was dirty and we flushed it successfully,
849 		 * and the requested range is not the entire object, we
850 		 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
851 		 * return immediately.
852 		 */
853 		if (tscan >= tend && (tstart || tend < object->size)) {
854 			vm_object_clear_flag(object, OBJ_CLEANING);
855 			return;
856 		}
857 		pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
858 	}
859 
860 	/*
861 	 * Generally set CLEANCHK interlock and make the page read-only so
862 	 * we can then clear the object flags.
863 	 *
864 	 * However, if this is a nosync mmap then the object is likely to
865 	 * stay dirty so do not mess with the page and do not clear the
866 	 * object flags.
867 	 */
868 	clearobjflags = 1;
869 	TAILQ_FOREACH(p, &object->memq, listq) {
870 		p->oflags |= VPO_CLEANCHK;
871 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC))
872 			clearobjflags = 0;
873 		else
874 			pmap_remove_write(p);
875 	}
876 
877 	if (clearobjflags && (tstart == 0) && (tend == object->size))
878 		vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
879 
880 rescan:
881 	curgeneration = object->generation;
882 
883 	for (p = TAILQ_FIRST(&object->memq); p; p = np) {
884 		int n;
885 
886 		np = TAILQ_NEXT(p, listq);
887 
888 again:
889 		pi = p->pindex;
890 		if ((p->oflags & VPO_CLEANCHK) == 0 ||
891 			(pi < tstart) || (pi >= tend) ||
892 		    p->valid == 0) {
893 			p->oflags &= ~VPO_CLEANCHK;
894 			continue;
895 		}
896 
897 		vm_page_test_dirty(p);
898 		if (p->dirty == 0) {
899 			p->oflags &= ~VPO_CLEANCHK;
900 			continue;
901 		}
902 		/*
903 		 * If we have been asked to skip nosync pages and this is a
904 		 * nosync page, skip it.  Note that the object flags were
905 		 * not cleared in this case so we do not have to set them.
906 		 */
907 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
908 			p->oflags &= ~VPO_CLEANCHK;
909 			continue;
910 		}
911 
912 		n = vm_object_page_collect_flush(object, p,
913 			curgeneration, pagerflags);
914 		if (n == 0)
915 			goto rescan;
916 
917 		if (object->generation != curgeneration)
918 			goto rescan;
919 
920 		/*
921 		 * Try to optimize the next page.  If we can't we pick up
922 		 * our (random) scan where we left off.
923 		 */
924 		if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ)
925 			if ((p = vm_page_lookup(object, pi + n)) != NULL)
926 				goto again;
927 	}
928 #if 0
929 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
930 #endif
931 
932 	vm_object_clear_flag(object, OBJ_CLEANING);
933 	return;
934 }
935 
936 static int
937 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
938 {
939 	int runlen;
940 	int maxf;
941 	int chkb;
942 	int maxb;
943 	int i;
944 	vm_pindex_t pi;
945 	vm_page_t maf[vm_pageout_page_count];
946 	vm_page_t mab[vm_pageout_page_count];
947 	vm_page_t ma[vm_pageout_page_count];
948 
949 	mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
950 	vm_page_lock_assert(p, MA_NOTOWNED);
951 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
952 	pi = p->pindex;
953 	while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
954 		if (object->generation != curgeneration) {
955 			return(0);
956 		}
957 	}
958 	maxf = 0;
959 	for(i = 1; i < vm_pageout_page_count; i++) {
960 		vm_page_t tp;
961 
962 		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
963 			if ((tp->oflags & VPO_BUSY) ||
964 				((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
965 				 (tp->oflags & VPO_CLEANCHK) == 0) ||
966 				(tp->busy != 0))
967 				break;
968 			vm_page_test_dirty(tp);
969 			if (tp->dirty == 0) {
970 				tp->oflags &= ~VPO_CLEANCHK;
971 				break;
972 			}
973 			maf[ i - 1 ] = tp;
974 			maxf++;
975 			continue;
976 		}
977 		break;
978 	}
979 
980 	maxb = 0;
981 	chkb = vm_pageout_page_count -  maxf;
982 	if (chkb) {
983 		for(i = 1; i < chkb;i++) {
984 			vm_page_t tp;
985 
986 			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
987 				if ((tp->oflags & VPO_BUSY) ||
988 					((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
989 					 (tp->oflags & VPO_CLEANCHK) == 0) ||
990 					(tp->busy != 0))
991 					break;
992 				vm_page_test_dirty(tp);
993 				if (tp->dirty == 0) {
994 					tp->oflags &= ~VPO_CLEANCHK;
995 					break;
996 				}
997 				mab[ i - 1 ] = tp;
998 				maxb++;
999 				continue;
1000 			}
1001 			break;
1002 		}
1003 	}
1004 
1005 	for(i = 0; i < maxb; i++) {
1006 		int index = (maxb - i) - 1;
1007 		ma[index] = mab[i];
1008 		ma[index]->oflags &= ~VPO_CLEANCHK;
1009 	}
1010 	p->oflags &= ~VPO_CLEANCHK;
1011 	ma[maxb] = p;
1012 	for(i = 0; i < maxf; i++) {
1013 		int index = (maxb + i) + 1;
1014 		ma[index] = maf[i];
1015 		ma[index]->oflags &= ~VPO_CLEANCHK;
1016 	}
1017 	runlen = maxb + maxf + 1;
1018 
1019 	vm_pageout_flush(ma, runlen, pagerflags);
1020 	for (i = 0; i < runlen; i++) {
1021 		if (ma[i]->dirty) {
1022 			pmap_remove_write(ma[i]);
1023 			ma[i]->oflags |= VPO_CLEANCHK;
1024 
1025 			/*
1026 			 * maxf will end up being the actual number of pages
1027 			 * we wrote out contiguously, non-inclusive of the
1028 			 * first page.  We do not count look-behind pages.
1029 			 */
1030 			if (i >= maxb + 1 && (maxf > i - maxb - 1))
1031 				maxf = i - maxb - 1;
1032 		}
1033 	}
1034 	return(maxf + 1);
1035 }
1036 
1037 /*
1038  * Note that there is absolutely no sense in writing out
1039  * anonymous objects, so we track down the vnode object
1040  * to write out.
1041  * We invalidate (remove) all pages from the address space
1042  * for semantic correctness.
1043  *
1044  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1045  * may start out with a NULL object.
1046  */
1047 void
1048 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1049     boolean_t syncio, boolean_t invalidate)
1050 {
1051 	vm_object_t backing_object;
1052 	struct vnode *vp;
1053 	struct mount *mp;
1054 	int flags;
1055 
1056 	if (object == NULL)
1057 		return;
1058 	VM_OBJECT_LOCK(object);
1059 	while ((backing_object = object->backing_object) != NULL) {
1060 		VM_OBJECT_LOCK(backing_object);
1061 		offset += object->backing_object_offset;
1062 		VM_OBJECT_UNLOCK(object);
1063 		object = backing_object;
1064 		if (object->size < OFF_TO_IDX(offset + size))
1065 			size = IDX_TO_OFF(object->size) - offset;
1066 	}
1067 	/*
1068 	 * Flush pages if writing is allowed, invalidate them
1069 	 * if invalidation requested.  Pages undergoing I/O
1070 	 * will be ignored by vm_object_page_remove().
1071 	 *
1072 	 * We cannot lock the vnode and then wait for paging
1073 	 * to complete without deadlocking against vm_fault.
1074 	 * Instead we simply call vm_object_page_remove() and
1075 	 * allow it to block internally on a page-by-page
1076 	 * basis when it encounters pages undergoing async
1077 	 * I/O.
1078 	 */
1079 	if (object->type == OBJT_VNODE &&
1080 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1081 		int vfslocked;
1082 		vp = object->handle;
1083 		VM_OBJECT_UNLOCK(object);
1084 		(void) vn_start_write(vp, &mp, V_WAIT);
1085 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1086 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1087 		flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1088 		flags |= invalidate ? OBJPC_INVAL : 0;
1089 		VM_OBJECT_LOCK(object);
1090 		vm_object_page_clean(object,
1091 		    OFF_TO_IDX(offset),
1092 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1093 		    flags);
1094 		VM_OBJECT_UNLOCK(object);
1095 		VOP_UNLOCK(vp, 0);
1096 		VFS_UNLOCK_GIANT(vfslocked);
1097 		vn_finished_write(mp);
1098 		VM_OBJECT_LOCK(object);
1099 	}
1100 	if ((object->type == OBJT_VNODE ||
1101 	     object->type == OBJT_DEVICE) && invalidate) {
1102 		boolean_t purge;
1103 		purge = old_msync || (object->type == OBJT_DEVICE);
1104 		vm_object_page_remove(object,
1105 		    OFF_TO_IDX(offset),
1106 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1107 		    purge ? FALSE : TRUE);
1108 	}
1109 	VM_OBJECT_UNLOCK(object);
1110 }
1111 
1112 /*
1113  *	vm_object_madvise:
1114  *
1115  *	Implements the madvise function at the object/page level.
1116  *
1117  *	MADV_WILLNEED	(any object)
1118  *
1119  *	    Activate the specified pages if they are resident.
1120  *
1121  *	MADV_DONTNEED	(any object)
1122  *
1123  *	    Deactivate the specified pages if they are resident.
1124  *
1125  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1126  *			 OBJ_ONEMAPPING only)
1127  *
1128  *	    Deactivate and clean the specified pages if they are
1129  *	    resident.  This permits the process to reuse the pages
1130  *	    without faulting or the kernel to reclaim the pages
1131  *	    without I/O.
1132  */
1133 void
1134 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1135 {
1136 	vm_pindex_t end, tpindex;
1137 	vm_object_t backing_object, tobject;
1138 	vm_page_t m;
1139 
1140 	if (object == NULL)
1141 		return;
1142 	VM_OBJECT_LOCK(object);
1143 	end = pindex + count;
1144 	/*
1145 	 * Locate and adjust resident pages
1146 	 */
1147 	for (; pindex < end; pindex += 1) {
1148 relookup:
1149 		tobject = object;
1150 		tpindex = pindex;
1151 shadowlookup:
1152 		/*
1153 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1154 		 * and those pages must be OBJ_ONEMAPPING.
1155 		 */
1156 		if (advise == MADV_FREE) {
1157 			if ((tobject->type != OBJT_DEFAULT &&
1158 			     tobject->type != OBJT_SWAP) ||
1159 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
1160 				goto unlock_tobject;
1161 			}
1162 		} else if (tobject->type == OBJT_PHYS)
1163 			goto unlock_tobject;
1164 		m = vm_page_lookup(tobject, tpindex);
1165 		if (m == NULL && advise == MADV_WILLNEED) {
1166 			/*
1167 			 * If the page is cached, reactivate it.
1168 			 */
1169 			m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1170 			    VM_ALLOC_NOBUSY);
1171 		}
1172 		if (m == NULL) {
1173 			/*
1174 			 * There may be swap even if there is no backing page
1175 			 */
1176 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1177 				swap_pager_freespace(tobject, tpindex, 1);
1178 			/*
1179 			 * next object
1180 			 */
1181 			backing_object = tobject->backing_object;
1182 			if (backing_object == NULL)
1183 				goto unlock_tobject;
1184 			VM_OBJECT_LOCK(backing_object);
1185 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1186 			if (tobject != object)
1187 				VM_OBJECT_UNLOCK(tobject);
1188 			tobject = backing_object;
1189 			goto shadowlookup;
1190 		} else if (m->valid != VM_PAGE_BITS_ALL)
1191 			goto unlock_tobject;
1192 		/*
1193 		 * If the page is not in a normal state, skip it.
1194 		 */
1195 		vm_page_lock(m);
1196 		if (m->hold_count != 0 || m->wire_count != 0) {
1197 			vm_page_unlock(m);
1198 			goto unlock_tobject;
1199 		}
1200 		KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1201 		    ("vm_object_madvise: page %p is not managed", m));
1202 		if ((m->oflags & VPO_BUSY) || m->busy) {
1203 			if (advise == MADV_WILLNEED) {
1204 				/*
1205 				 * Reference the page before unlocking and
1206 				 * sleeping so that the page daemon is less
1207 				 * likely to reclaim it.
1208 				 */
1209 				vm_page_lock_queues();
1210 				vm_page_flag_set(m, PG_REFERENCED);
1211 				vm_page_unlock_queues();
1212 			}
1213 			vm_page_unlock(m);
1214 			if (object != tobject)
1215 				VM_OBJECT_UNLOCK(object);
1216 			m->oflags |= VPO_WANTED;
1217 			msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1218 			    0);
1219 			VM_OBJECT_LOCK(object);
1220   			goto relookup;
1221 		}
1222 		if (advise == MADV_WILLNEED) {
1223 			vm_page_activate(m);
1224 		} else if (advise == MADV_DONTNEED) {
1225 			vm_page_dontneed(m);
1226 		} else if (advise == MADV_FREE) {
1227 			/*
1228 			 * Mark the page clean.  This will allow the page
1229 			 * to be freed up by the system.  However, such pages
1230 			 * are often reused quickly by malloc()/free()
1231 			 * so we do not do anything that would cause
1232 			 * a page fault if we can help it.
1233 			 *
1234 			 * Specifically, we do not try to actually free
1235 			 * the page now nor do we try to put it in the
1236 			 * cache (which would cause a page fault on reuse).
1237 			 *
1238 			 * But we do make the page is freeable as we
1239 			 * can without actually taking the step of unmapping
1240 			 * it.
1241 			 */
1242 			pmap_clear_modify(m);
1243 			m->dirty = 0;
1244 			m->act_count = 0;
1245 			vm_page_dontneed(m);
1246 		}
1247 		vm_page_unlock(m);
1248 		if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1249 			swap_pager_freespace(tobject, tpindex, 1);
1250 unlock_tobject:
1251 		if (tobject != object)
1252 			VM_OBJECT_UNLOCK(tobject);
1253 	}
1254 	VM_OBJECT_UNLOCK(object);
1255 }
1256 
1257 /*
1258  *	vm_object_shadow:
1259  *
1260  *	Create a new object which is backed by the
1261  *	specified existing object range.  The source
1262  *	object reference is deallocated.
1263  *
1264  *	The new object and offset into that object
1265  *	are returned in the source parameters.
1266  */
1267 void
1268 vm_object_shadow(
1269 	vm_object_t *object,	/* IN/OUT */
1270 	vm_ooffset_t *offset,	/* IN/OUT */
1271 	vm_size_t length)
1272 {
1273 	vm_object_t source;
1274 	vm_object_t result;
1275 
1276 	source = *object;
1277 
1278 	/*
1279 	 * Don't create the new object if the old object isn't shared.
1280 	 */
1281 	if (source != NULL) {
1282 		VM_OBJECT_LOCK(source);
1283 		if (source->ref_count == 1 &&
1284 		    source->handle == NULL &&
1285 		    (source->type == OBJT_DEFAULT ||
1286 		     source->type == OBJT_SWAP)) {
1287 			VM_OBJECT_UNLOCK(source);
1288 			return;
1289 		}
1290 		VM_OBJECT_UNLOCK(source);
1291 	}
1292 
1293 	/*
1294 	 * Allocate a new object with the given length.
1295 	 */
1296 	result = vm_object_allocate(OBJT_DEFAULT, length);
1297 
1298 	/*
1299 	 * The new object shadows the source object, adding a reference to it.
1300 	 * Our caller changes his reference to point to the new object,
1301 	 * removing a reference to the source object.  Net result: no change
1302 	 * of reference count.
1303 	 *
1304 	 * Try to optimize the result object's page color when shadowing
1305 	 * in order to maintain page coloring consistency in the combined
1306 	 * shadowed object.
1307 	 */
1308 	result->backing_object = source;
1309 	/*
1310 	 * Store the offset into the source object, and fix up the offset into
1311 	 * the new object.
1312 	 */
1313 	result->backing_object_offset = *offset;
1314 	if (source != NULL) {
1315 		VM_OBJECT_LOCK(source);
1316 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1317 		source->shadow_count++;
1318 		source->generation++;
1319 #if VM_NRESERVLEVEL > 0
1320 		result->flags |= source->flags & OBJ_COLORED;
1321 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1322 		    ((1 << (VM_NFREEORDER - 1)) - 1);
1323 #endif
1324 		VM_OBJECT_UNLOCK(source);
1325 	}
1326 
1327 
1328 	/*
1329 	 * Return the new things
1330 	 */
1331 	*offset = 0;
1332 	*object = result;
1333 }
1334 
1335 /*
1336  *	vm_object_split:
1337  *
1338  * Split the pages in a map entry into a new object.  This affords
1339  * easier removal of unused pages, and keeps object inheritance from
1340  * being a negative impact on memory usage.
1341  */
1342 void
1343 vm_object_split(vm_map_entry_t entry)
1344 {
1345 	vm_page_t m, m_next;
1346 	vm_object_t orig_object, new_object, source;
1347 	vm_pindex_t idx, offidxstart;
1348 	vm_size_t size;
1349 
1350 	orig_object = entry->object.vm_object;
1351 	if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1352 		return;
1353 	if (orig_object->ref_count <= 1)
1354 		return;
1355 	VM_OBJECT_UNLOCK(orig_object);
1356 
1357 	offidxstart = OFF_TO_IDX(entry->offset);
1358 	size = atop(entry->end - entry->start);
1359 
1360 	/*
1361 	 * If swap_pager_copy() is later called, it will convert new_object
1362 	 * into a swap object.
1363 	 */
1364 	new_object = vm_object_allocate(OBJT_DEFAULT, size);
1365 
1366 	/*
1367 	 * At this point, the new object is still private, so the order in
1368 	 * which the original and new objects are locked does not matter.
1369 	 */
1370 	VM_OBJECT_LOCK(new_object);
1371 	VM_OBJECT_LOCK(orig_object);
1372 	source = orig_object->backing_object;
1373 	if (source != NULL) {
1374 		VM_OBJECT_LOCK(source);
1375 		if ((source->flags & OBJ_DEAD) != 0) {
1376 			VM_OBJECT_UNLOCK(source);
1377 			VM_OBJECT_UNLOCK(orig_object);
1378 			VM_OBJECT_UNLOCK(new_object);
1379 			vm_object_deallocate(new_object);
1380 			VM_OBJECT_LOCK(orig_object);
1381 			return;
1382 		}
1383 		LIST_INSERT_HEAD(&source->shadow_head,
1384 				  new_object, shadow_list);
1385 		source->shadow_count++;
1386 		source->generation++;
1387 		vm_object_reference_locked(source);	/* for new_object */
1388 		vm_object_clear_flag(source, OBJ_ONEMAPPING);
1389 		VM_OBJECT_UNLOCK(source);
1390 		new_object->backing_object_offset =
1391 			orig_object->backing_object_offset + entry->offset;
1392 		new_object->backing_object = source;
1393 	}
1394 	if (orig_object->uip != NULL) {
1395 		new_object->uip = orig_object->uip;
1396 		uihold(orig_object->uip);
1397 		new_object->charge = ptoa(size);
1398 		KASSERT(orig_object->charge >= ptoa(size),
1399 		    ("orig_object->charge < 0"));
1400 		orig_object->charge -= ptoa(size);
1401 	}
1402 retry:
1403 	if ((m = TAILQ_FIRST(&orig_object->memq)) != NULL) {
1404 		if (m->pindex < offidxstart) {
1405 			m = vm_page_splay(offidxstart, orig_object->root);
1406 			if ((orig_object->root = m)->pindex < offidxstart)
1407 				m = TAILQ_NEXT(m, listq);
1408 		}
1409 	}
1410 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1411 	    m = m_next) {
1412 		m_next = TAILQ_NEXT(m, listq);
1413 
1414 		/*
1415 		 * We must wait for pending I/O to complete before we can
1416 		 * rename the page.
1417 		 *
1418 		 * We do not have to VM_PROT_NONE the page as mappings should
1419 		 * not be changed by this operation.
1420 		 */
1421 		if ((m->oflags & VPO_BUSY) || m->busy) {
1422 			VM_OBJECT_UNLOCK(new_object);
1423 			m->oflags |= VPO_WANTED;
1424 			msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1425 			VM_OBJECT_LOCK(new_object);
1426 			goto retry;
1427 		}
1428 		vm_page_lock(m);
1429 		vm_page_rename(m, new_object, idx);
1430 		vm_page_unlock(m);
1431 		/* page automatically made dirty by rename and cache handled */
1432 		vm_page_busy(m);
1433 	}
1434 	if (orig_object->type == OBJT_SWAP) {
1435 		/*
1436 		 * swap_pager_copy() can sleep, in which case the orig_object's
1437 		 * and new_object's locks are released and reacquired.
1438 		 */
1439 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1440 
1441 		/*
1442 		 * Transfer any cached pages from orig_object to new_object.
1443 		 */
1444 		if (__predict_false(orig_object->cache != NULL))
1445 			vm_page_cache_transfer(orig_object, offidxstart,
1446 			    new_object);
1447 	}
1448 	VM_OBJECT_UNLOCK(orig_object);
1449 	TAILQ_FOREACH(m, &new_object->memq, listq)
1450 		vm_page_wakeup(m);
1451 	VM_OBJECT_UNLOCK(new_object);
1452 	entry->object.vm_object = new_object;
1453 	entry->offset = 0LL;
1454 	vm_object_deallocate(orig_object);
1455 	VM_OBJECT_LOCK(new_object);
1456 }
1457 
1458 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1459 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1460 #define	OBSC_COLLAPSE_WAIT	0x0004
1461 
1462 static int
1463 vm_object_backing_scan(vm_object_t object, int op)
1464 {
1465 	int r = 1;
1466 	vm_page_t p;
1467 	vm_object_t backing_object;
1468 	vm_pindex_t backing_offset_index;
1469 
1470 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1471 	VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1472 
1473 	backing_object = object->backing_object;
1474 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1475 
1476 	/*
1477 	 * Initial conditions
1478 	 */
1479 	if (op & OBSC_TEST_ALL_SHADOWED) {
1480 		/*
1481 		 * We do not want to have to test for the existence of cache
1482 		 * or swap pages in the backing object.  XXX but with the
1483 		 * new swapper this would be pretty easy to do.
1484 		 *
1485 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1486 		 * been ZFOD faulted yet?  If we do not test for this, the
1487 		 * shadow test may succeed! XXX
1488 		 */
1489 		if (backing_object->type != OBJT_DEFAULT) {
1490 			return (0);
1491 		}
1492 	}
1493 	if (op & OBSC_COLLAPSE_WAIT) {
1494 		vm_object_set_flag(backing_object, OBJ_DEAD);
1495 	}
1496 
1497 	/*
1498 	 * Our scan
1499 	 */
1500 	p = TAILQ_FIRST(&backing_object->memq);
1501 	while (p) {
1502 		vm_page_t next = TAILQ_NEXT(p, listq);
1503 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1504 
1505 		if (op & OBSC_TEST_ALL_SHADOWED) {
1506 			vm_page_t pp;
1507 
1508 			/*
1509 			 * Ignore pages outside the parent object's range
1510 			 * and outside the parent object's mapping of the
1511 			 * backing object.
1512 			 *
1513 			 * note that we do not busy the backing object's
1514 			 * page.
1515 			 */
1516 			if (
1517 			    p->pindex < backing_offset_index ||
1518 			    new_pindex >= object->size
1519 			) {
1520 				p = next;
1521 				continue;
1522 			}
1523 
1524 			/*
1525 			 * See if the parent has the page or if the parent's
1526 			 * object pager has the page.  If the parent has the
1527 			 * page but the page is not valid, the parent's
1528 			 * object pager must have the page.
1529 			 *
1530 			 * If this fails, the parent does not completely shadow
1531 			 * the object and we might as well give up now.
1532 			 */
1533 
1534 			pp = vm_page_lookup(object, new_pindex);
1535 			if (
1536 			    (pp == NULL || pp->valid == 0) &&
1537 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1538 			) {
1539 				r = 0;
1540 				break;
1541 			}
1542 		}
1543 
1544 		/*
1545 		 * Check for busy page
1546 		 */
1547 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1548 			vm_page_t pp;
1549 
1550 			if (op & OBSC_COLLAPSE_NOWAIT) {
1551 				if ((p->oflags & VPO_BUSY) ||
1552 				    !p->valid ||
1553 				    p->busy) {
1554 					p = next;
1555 					continue;
1556 				}
1557 			} else if (op & OBSC_COLLAPSE_WAIT) {
1558 				if ((p->oflags & VPO_BUSY) || p->busy) {
1559 					VM_OBJECT_UNLOCK(object);
1560 					p->oflags |= VPO_WANTED;
1561 					msleep(p, VM_OBJECT_MTX(backing_object),
1562 					    PDROP | PVM, "vmocol", 0);
1563 					VM_OBJECT_LOCK(object);
1564 					VM_OBJECT_LOCK(backing_object);
1565 					/*
1566 					 * If we slept, anything could have
1567 					 * happened.  Since the object is
1568 					 * marked dead, the backing offset
1569 					 * should not have changed so we
1570 					 * just restart our scan.
1571 					 */
1572 					p = TAILQ_FIRST(&backing_object->memq);
1573 					continue;
1574 				}
1575 			}
1576 
1577 			KASSERT(
1578 			    p->object == backing_object,
1579 			    ("vm_object_backing_scan: object mismatch")
1580 			);
1581 
1582 			/*
1583 			 * Destroy any associated swap
1584 			 */
1585 			if (backing_object->type == OBJT_SWAP) {
1586 				swap_pager_freespace(
1587 				    backing_object,
1588 				    p->pindex,
1589 				    1
1590 				);
1591 			}
1592 
1593 			if (
1594 			    p->pindex < backing_offset_index ||
1595 			    new_pindex >= object->size
1596 			) {
1597 				/*
1598 				 * Page is out of the parent object's range, we
1599 				 * can simply destroy it.
1600 				 */
1601 				vm_page_lock(p);
1602 				KASSERT(!pmap_page_is_mapped(p),
1603 				    ("freeing mapped page %p", p));
1604 				if (p->wire_count == 0)
1605 					vm_page_free(p);
1606 				else
1607 					vm_page_remove(p);
1608 				vm_page_unlock(p);
1609 				p = next;
1610 				continue;
1611 			}
1612 
1613 			pp = vm_page_lookup(object, new_pindex);
1614 			if (
1615 			    pp != NULL ||
1616 			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1617 			) {
1618 				/*
1619 				 * page already exists in parent OR swap exists
1620 				 * for this location in the parent.  Destroy
1621 				 * the original page from the backing object.
1622 				 *
1623 				 * Leave the parent's page alone
1624 				 */
1625 				vm_page_lock(p);
1626 				KASSERT(!pmap_page_is_mapped(p),
1627 				    ("freeing mapped page %p", p));
1628 				if (p->wire_count == 0)
1629 					vm_page_free(p);
1630 				else
1631 					vm_page_remove(p);
1632 				vm_page_unlock(p);
1633 				p = next;
1634 				continue;
1635 			}
1636 
1637 #if VM_NRESERVLEVEL > 0
1638 			/*
1639 			 * Rename the reservation.
1640 			 */
1641 			vm_reserv_rename(p, object, backing_object,
1642 			    backing_offset_index);
1643 #endif
1644 
1645 			/*
1646 			 * Page does not exist in parent, rename the
1647 			 * page from the backing object to the main object.
1648 			 *
1649 			 * If the page was mapped to a process, it can remain
1650 			 * mapped through the rename.
1651 			 */
1652 			vm_page_lock(p);
1653 			vm_page_rename(p, object, new_pindex);
1654 			vm_page_unlock(p);
1655 			/* page automatically made dirty by rename */
1656 		}
1657 		p = next;
1658 	}
1659 	return (r);
1660 }
1661 
1662 
1663 /*
1664  * this version of collapse allows the operation to occur earlier and
1665  * when paging_in_progress is true for an object...  This is not a complete
1666  * operation, but should plug 99.9% of the rest of the leaks.
1667  */
1668 static void
1669 vm_object_qcollapse(vm_object_t object)
1670 {
1671 	vm_object_t backing_object = object->backing_object;
1672 
1673 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1674 	VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1675 
1676 	if (backing_object->ref_count != 1)
1677 		return;
1678 
1679 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1680 }
1681 
1682 /*
1683  *	vm_object_collapse:
1684  *
1685  *	Collapse an object with the object backing it.
1686  *	Pages in the backing object are moved into the
1687  *	parent, and the backing object is deallocated.
1688  */
1689 void
1690 vm_object_collapse(vm_object_t object)
1691 {
1692 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1693 
1694 	while (TRUE) {
1695 		vm_object_t backing_object;
1696 
1697 		/*
1698 		 * Verify that the conditions are right for collapse:
1699 		 *
1700 		 * The object exists and the backing object exists.
1701 		 */
1702 		if ((backing_object = object->backing_object) == NULL)
1703 			break;
1704 
1705 		/*
1706 		 * we check the backing object first, because it is most likely
1707 		 * not collapsable.
1708 		 */
1709 		VM_OBJECT_LOCK(backing_object);
1710 		if (backing_object->handle != NULL ||
1711 		    (backing_object->type != OBJT_DEFAULT &&
1712 		     backing_object->type != OBJT_SWAP) ||
1713 		    (backing_object->flags & OBJ_DEAD) ||
1714 		    object->handle != NULL ||
1715 		    (object->type != OBJT_DEFAULT &&
1716 		     object->type != OBJT_SWAP) ||
1717 		    (object->flags & OBJ_DEAD)) {
1718 			VM_OBJECT_UNLOCK(backing_object);
1719 			break;
1720 		}
1721 
1722 		if (
1723 		    object->paging_in_progress != 0 ||
1724 		    backing_object->paging_in_progress != 0
1725 		) {
1726 			vm_object_qcollapse(object);
1727 			VM_OBJECT_UNLOCK(backing_object);
1728 			break;
1729 		}
1730 		/*
1731 		 * We know that we can either collapse the backing object (if
1732 		 * the parent is the only reference to it) or (perhaps) have
1733 		 * the parent bypass the object if the parent happens to shadow
1734 		 * all the resident pages in the entire backing object.
1735 		 *
1736 		 * This is ignoring pager-backed pages such as swap pages.
1737 		 * vm_object_backing_scan fails the shadowing test in this
1738 		 * case.
1739 		 */
1740 		if (backing_object->ref_count == 1) {
1741 			/*
1742 			 * If there is exactly one reference to the backing
1743 			 * object, we can collapse it into the parent.
1744 			 */
1745 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1746 
1747 #if VM_NRESERVLEVEL > 0
1748 			/*
1749 			 * Break any reservations from backing_object.
1750 			 */
1751 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1752 				vm_reserv_break_all(backing_object);
1753 #endif
1754 
1755 			/*
1756 			 * Move the pager from backing_object to object.
1757 			 */
1758 			if (backing_object->type == OBJT_SWAP) {
1759 				/*
1760 				 * swap_pager_copy() can sleep, in which case
1761 				 * the backing_object's and object's locks are
1762 				 * released and reacquired.
1763 				 */
1764 				swap_pager_copy(
1765 				    backing_object,
1766 				    object,
1767 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1768 
1769 				/*
1770 				 * Free any cached pages from backing_object.
1771 				 */
1772 				if (__predict_false(backing_object->cache != NULL))
1773 					vm_page_cache_free(backing_object, 0, 0);
1774 			}
1775 			/*
1776 			 * Object now shadows whatever backing_object did.
1777 			 * Note that the reference to
1778 			 * backing_object->backing_object moves from within
1779 			 * backing_object to within object.
1780 			 */
1781 			LIST_REMOVE(object, shadow_list);
1782 			backing_object->shadow_count--;
1783 			backing_object->generation++;
1784 			if (backing_object->backing_object) {
1785 				VM_OBJECT_LOCK(backing_object->backing_object);
1786 				LIST_REMOVE(backing_object, shadow_list);
1787 				LIST_INSERT_HEAD(
1788 				    &backing_object->backing_object->shadow_head,
1789 				    object, shadow_list);
1790 				/*
1791 				 * The shadow_count has not changed.
1792 				 */
1793 				backing_object->backing_object->generation++;
1794 				VM_OBJECT_UNLOCK(backing_object->backing_object);
1795 			}
1796 			object->backing_object = backing_object->backing_object;
1797 			object->backing_object_offset +=
1798 			    backing_object->backing_object_offset;
1799 
1800 			/*
1801 			 * Discard backing_object.
1802 			 *
1803 			 * Since the backing object has no pages, no pager left,
1804 			 * and no object references within it, all that is
1805 			 * necessary is to dispose of it.
1806 			 */
1807 			KASSERT(backing_object->ref_count == 1, (
1808 "backing_object %p was somehow re-referenced during collapse!",
1809 			    backing_object));
1810 			VM_OBJECT_UNLOCK(backing_object);
1811 			vm_object_destroy(backing_object);
1812 
1813 			object_collapses++;
1814 		} else {
1815 			vm_object_t new_backing_object;
1816 
1817 			/*
1818 			 * If we do not entirely shadow the backing object,
1819 			 * there is nothing we can do so we give up.
1820 			 */
1821 			if (object->resident_page_count != object->size &&
1822 			    vm_object_backing_scan(object,
1823 			    OBSC_TEST_ALL_SHADOWED) == 0) {
1824 				VM_OBJECT_UNLOCK(backing_object);
1825 				break;
1826 			}
1827 
1828 			/*
1829 			 * Make the parent shadow the next object in the
1830 			 * chain.  Deallocating backing_object will not remove
1831 			 * it, since its reference count is at least 2.
1832 			 */
1833 			LIST_REMOVE(object, shadow_list);
1834 			backing_object->shadow_count--;
1835 			backing_object->generation++;
1836 
1837 			new_backing_object = backing_object->backing_object;
1838 			if ((object->backing_object = new_backing_object) != NULL) {
1839 				VM_OBJECT_LOCK(new_backing_object);
1840 				LIST_INSERT_HEAD(
1841 				    &new_backing_object->shadow_head,
1842 				    object,
1843 				    shadow_list
1844 				);
1845 				new_backing_object->shadow_count++;
1846 				new_backing_object->generation++;
1847 				vm_object_reference_locked(new_backing_object);
1848 				VM_OBJECT_UNLOCK(new_backing_object);
1849 				object->backing_object_offset +=
1850 					backing_object->backing_object_offset;
1851 			}
1852 
1853 			/*
1854 			 * Drop the reference count on backing_object. Since
1855 			 * its ref_count was at least 2, it will not vanish.
1856 			 */
1857 			backing_object->ref_count--;
1858 			VM_OBJECT_UNLOCK(backing_object);
1859 			object_bypasses++;
1860 		}
1861 
1862 		/*
1863 		 * Try again with this object's new backing object.
1864 		 */
1865 	}
1866 }
1867 
1868 /*
1869  *	vm_object_page_remove:
1870  *
1871  *	For the given object, either frees or invalidates each of the
1872  *	specified pages.  In general, a page is freed.  However, if a
1873  *	page is wired for any reason other than the existence of a
1874  *	managed, wired mapping, then it may be invalidated but not
1875  *	removed from the object.  Pages are specified by the given
1876  *	range ["start", "end") and Boolean "clean_only".  As a
1877  *	special case, if "end" is zero, then the range extends from
1878  *	"start" to the end of the object.  If "clean_only" is TRUE,
1879  *	then only the non-dirty pages within the specified range are
1880  *	affected.
1881  *
1882  *	In general, this operation should only be performed on objects
1883  *	that contain managed pages.  There are two exceptions.  First,
1884  *	it may be performed on the kernel and kmem objects.  Second,
1885  *	it may be used by msync(..., MS_INVALIDATE) to invalidate
1886  *	device-backed pages.  In both of these cases, "clean_only"
1887  *	must be FALSE.
1888  *
1889  *	The object must be locked.
1890  */
1891 void
1892 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1893     boolean_t clean_only)
1894 {
1895 	vm_page_t p, next;
1896 	int wirings;
1897 
1898 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1899 	if (object->resident_page_count == 0)
1900 		goto skipmemq;
1901 
1902 	/*
1903 	 * Since physically-backed objects do not use managed pages, we can't
1904 	 * remove pages from the object (we must instead remove the page
1905 	 * references, and then destroy the object).
1906 	 */
1907 	KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1908 	    object == kmem_object,
1909 	    ("attempt to remove pages from a physical object"));
1910 
1911 	vm_object_pip_add(object, 1);
1912 again:
1913 	if ((p = TAILQ_FIRST(&object->memq)) != NULL) {
1914 		if (p->pindex < start) {
1915 			p = vm_page_splay(start, object->root);
1916 			if ((object->root = p)->pindex < start)
1917 				p = TAILQ_NEXT(p, listq);
1918 		}
1919 	}
1920 
1921 	/*
1922 	 * Assert: the variable p is either (1) the page with the
1923 	 * least pindex greater than or equal to the parameter pindex
1924 	 * or (2) NULL.
1925 	 */
1926 	for (;
1927 	     p != NULL && (p->pindex < end || end == 0);
1928 	     p = next) {
1929 		next = TAILQ_NEXT(p, listq);
1930 
1931 		/*
1932 		 * If the page is wired for any reason besides the
1933 		 * existence of managed, wired mappings, then it cannot
1934 		 * be freed.  For example, fictitious pages, which
1935 		 * represent device memory, are inherently wired and
1936 		 * cannot be freed.  They can, however, be invalidated
1937 		 * if "clean_only" is FALSE.
1938 		 */
1939 		vm_page_lock(p);
1940 		if ((wirings = p->wire_count) != 0 &&
1941 		    (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1942 			/* Fictitious pages do not have managed mappings. */
1943 			if ((p->flags & PG_FICTITIOUS) == 0)
1944 				pmap_remove_all(p);
1945 			/* Account for removal of managed, wired mappings. */
1946 			p->wire_count -= wirings;
1947 			if (!clean_only) {
1948 				p->valid = 0;
1949 				vm_page_undirty(p);
1950 			}
1951 			vm_page_unlock(p);
1952 			continue;
1953 		}
1954 		if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1955 			goto again;
1956 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
1957 		    ("vm_object_page_remove: page %p is fictitious", p));
1958 		if (clean_only && p->valid) {
1959 			pmap_remove_write(p);
1960 			if (p->dirty) {
1961 				vm_page_unlock(p);
1962 				continue;
1963 			}
1964 		}
1965 		pmap_remove_all(p);
1966 		/* Account for removal of managed, wired mappings. */
1967 		if (wirings != 0)
1968 			p->wire_count -= wirings;
1969 		vm_page_free(p);
1970 		vm_page_unlock(p);
1971 	}
1972 	vm_object_pip_wakeup(object);
1973 skipmemq:
1974 	if (__predict_false(object->cache != NULL))
1975 		vm_page_cache_free(object, start, end);
1976 }
1977 
1978 /*
1979  *	Populate the specified range of the object with valid pages.  Returns
1980  *	TRUE if the range is successfully populated and FALSE otherwise.
1981  *
1982  *	Note: This function should be optimized to pass a larger array of
1983  *	pages to vm_pager_get_pages() before it is applied to a non-
1984  *	OBJT_DEVICE object.
1985  *
1986  *	The object must be locked.
1987  */
1988 boolean_t
1989 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1990 {
1991 	vm_page_t m, ma[1];
1992 	vm_pindex_t pindex;
1993 	int rv;
1994 
1995 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1996 	for (pindex = start; pindex < end; pindex++) {
1997 		m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1998 		    VM_ALLOC_RETRY);
1999 		if (m->valid != VM_PAGE_BITS_ALL) {
2000 			ma[0] = m;
2001 			rv = vm_pager_get_pages(object, ma, 1, 0);
2002 			m = vm_page_lookup(object, pindex);
2003 			if (m == NULL)
2004 				break;
2005 			if (rv != VM_PAGER_OK) {
2006 				vm_page_lock(m);
2007 				vm_page_free(m);
2008 				vm_page_unlock(m);
2009 				break;
2010 			}
2011 		}
2012 		/*
2013 		 * Keep "m" busy because a subsequent iteration may unlock
2014 		 * the object.
2015 		 */
2016 	}
2017 	if (pindex > start) {
2018 		m = vm_page_lookup(object, start);
2019 		while (m != NULL && m->pindex < pindex) {
2020 			vm_page_wakeup(m);
2021 			m = TAILQ_NEXT(m, listq);
2022 		}
2023 	}
2024 	return (pindex == end);
2025 }
2026 
2027 /*
2028  *	Routine:	vm_object_coalesce
2029  *	Function:	Coalesces two objects backing up adjoining
2030  *			regions of memory into a single object.
2031  *
2032  *	returns TRUE if objects were combined.
2033  *
2034  *	NOTE:	Only works at the moment if the second object is NULL -
2035  *		if it's not, which object do we lock first?
2036  *
2037  *	Parameters:
2038  *		prev_object	First object to coalesce
2039  *		prev_offset	Offset into prev_object
2040  *		prev_size	Size of reference to prev_object
2041  *		next_size	Size of reference to the second object
2042  *		reserved	Indicator that extension region has
2043  *				swap accounted for
2044  *
2045  *	Conditions:
2046  *	The object must *not* be locked.
2047  */
2048 boolean_t
2049 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2050     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2051 {
2052 	vm_pindex_t next_pindex;
2053 
2054 	if (prev_object == NULL)
2055 		return (TRUE);
2056 	VM_OBJECT_LOCK(prev_object);
2057 	if (prev_object->type != OBJT_DEFAULT &&
2058 	    prev_object->type != OBJT_SWAP) {
2059 		VM_OBJECT_UNLOCK(prev_object);
2060 		return (FALSE);
2061 	}
2062 
2063 	/*
2064 	 * Try to collapse the object first
2065 	 */
2066 	vm_object_collapse(prev_object);
2067 
2068 	/*
2069 	 * Can't coalesce if: . more than one reference . paged out . shadows
2070 	 * another object . has a copy elsewhere (any of which mean that the
2071 	 * pages not mapped to prev_entry may be in use anyway)
2072 	 */
2073 	if (prev_object->backing_object != NULL) {
2074 		VM_OBJECT_UNLOCK(prev_object);
2075 		return (FALSE);
2076 	}
2077 
2078 	prev_size >>= PAGE_SHIFT;
2079 	next_size >>= PAGE_SHIFT;
2080 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2081 
2082 	if ((prev_object->ref_count > 1) &&
2083 	    (prev_object->size != next_pindex)) {
2084 		VM_OBJECT_UNLOCK(prev_object);
2085 		return (FALSE);
2086 	}
2087 
2088 	/*
2089 	 * Account for the charge.
2090 	 */
2091 	if (prev_object->uip != NULL) {
2092 
2093 		/*
2094 		 * If prev_object was charged, then this mapping,
2095 		 * althought not charged now, may become writable
2096 		 * later. Non-NULL uip in the object would prevent
2097 		 * swap reservation during enabling of the write
2098 		 * access, so reserve swap now. Failed reservation
2099 		 * cause allocation of the separate object for the map
2100 		 * entry, and swap reservation for this entry is
2101 		 * managed in appropriate time.
2102 		 */
2103 		if (!reserved && !swap_reserve_by_uid(ptoa(next_size),
2104 		    prev_object->uip)) {
2105 			return (FALSE);
2106 		}
2107 		prev_object->charge += ptoa(next_size);
2108 	}
2109 
2110 	/*
2111 	 * Remove any pages that may still be in the object from a previous
2112 	 * deallocation.
2113 	 */
2114 	if (next_pindex < prev_object->size) {
2115 		vm_object_page_remove(prev_object,
2116 				      next_pindex,
2117 				      next_pindex + next_size, FALSE);
2118 		if (prev_object->type == OBJT_SWAP)
2119 			swap_pager_freespace(prev_object,
2120 					     next_pindex, next_size);
2121 #if 0
2122 		if (prev_object->uip != NULL) {
2123 			KASSERT(prev_object->charge >=
2124 			    ptoa(prev_object->size - next_pindex),
2125 			    ("object %p overcharged 1 %jx %jx", prev_object,
2126 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2127 			prev_object->charge -= ptoa(prev_object->size -
2128 			    next_pindex);
2129 		}
2130 #endif
2131 	}
2132 
2133 	/*
2134 	 * Extend the object if necessary.
2135 	 */
2136 	if (next_pindex + next_size > prev_object->size)
2137 		prev_object->size = next_pindex + next_size;
2138 
2139 	VM_OBJECT_UNLOCK(prev_object);
2140 	return (TRUE);
2141 }
2142 
2143 void
2144 vm_object_set_writeable_dirty(vm_object_t object)
2145 {
2146 
2147 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2148 	if (object->type != OBJT_VNODE ||
2149 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0)
2150 		return;
2151 	vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2152 }
2153 
2154 #include "opt_ddb.h"
2155 #ifdef DDB
2156 #include <sys/kernel.h>
2157 
2158 #include <sys/cons.h>
2159 
2160 #include <ddb/ddb.h>
2161 
2162 static int
2163 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2164 {
2165 	vm_map_t tmpm;
2166 	vm_map_entry_t tmpe;
2167 	vm_object_t obj;
2168 	int entcount;
2169 
2170 	if (map == 0)
2171 		return 0;
2172 
2173 	if (entry == 0) {
2174 		tmpe = map->header.next;
2175 		entcount = map->nentries;
2176 		while (entcount-- && (tmpe != &map->header)) {
2177 			if (_vm_object_in_map(map, object, tmpe)) {
2178 				return 1;
2179 			}
2180 			tmpe = tmpe->next;
2181 		}
2182 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2183 		tmpm = entry->object.sub_map;
2184 		tmpe = tmpm->header.next;
2185 		entcount = tmpm->nentries;
2186 		while (entcount-- && tmpe != &tmpm->header) {
2187 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2188 				return 1;
2189 			}
2190 			tmpe = tmpe->next;
2191 		}
2192 	} else if ((obj = entry->object.vm_object) != NULL) {
2193 		for (; obj; obj = obj->backing_object)
2194 			if (obj == object) {
2195 				return 1;
2196 			}
2197 	}
2198 	return 0;
2199 }
2200 
2201 static int
2202 vm_object_in_map(vm_object_t object)
2203 {
2204 	struct proc *p;
2205 
2206 	/* sx_slock(&allproc_lock); */
2207 	FOREACH_PROC_IN_SYSTEM(p) {
2208 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2209 			continue;
2210 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2211 			/* sx_sunlock(&allproc_lock); */
2212 			return 1;
2213 		}
2214 	}
2215 	/* sx_sunlock(&allproc_lock); */
2216 	if (_vm_object_in_map(kernel_map, object, 0))
2217 		return 1;
2218 	if (_vm_object_in_map(kmem_map, object, 0))
2219 		return 1;
2220 	if (_vm_object_in_map(pager_map, object, 0))
2221 		return 1;
2222 	if (_vm_object_in_map(buffer_map, object, 0))
2223 		return 1;
2224 	return 0;
2225 }
2226 
2227 DB_SHOW_COMMAND(vmochk, vm_object_check)
2228 {
2229 	vm_object_t object;
2230 
2231 	/*
2232 	 * make sure that internal objs are in a map somewhere
2233 	 * and none have zero ref counts.
2234 	 */
2235 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2236 		if (object->handle == NULL &&
2237 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2238 			if (object->ref_count == 0) {
2239 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2240 					(long)object->size);
2241 			}
2242 			if (!vm_object_in_map(object)) {
2243 				db_printf(
2244 			"vmochk: internal obj is not in a map: "
2245 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2246 				    object->ref_count, (u_long)object->size,
2247 				    (u_long)object->size,
2248 				    (void *)object->backing_object);
2249 			}
2250 		}
2251 	}
2252 }
2253 
2254 /*
2255  *	vm_object_print:	[ debug ]
2256  */
2257 DB_SHOW_COMMAND(object, vm_object_print_static)
2258 {
2259 	/* XXX convert args. */
2260 	vm_object_t object = (vm_object_t)addr;
2261 	boolean_t full = have_addr;
2262 
2263 	vm_page_t p;
2264 
2265 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2266 #define	count	was_count
2267 
2268 	int count;
2269 
2270 	if (object == NULL)
2271 		return;
2272 
2273 	db_iprintf(
2274 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x uip %d charge %jx\n",
2275 	    object, (int)object->type, (uintmax_t)object->size,
2276 	    object->resident_page_count, object->ref_count, object->flags,
2277 	    object->uip ? object->uip->ui_uid : -1, (uintmax_t)object->charge);
2278 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2279 	    object->shadow_count,
2280 	    object->backing_object ? object->backing_object->ref_count : 0,
2281 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2282 
2283 	if (!full)
2284 		return;
2285 
2286 	db_indent += 2;
2287 	count = 0;
2288 	TAILQ_FOREACH(p, &object->memq, listq) {
2289 		if (count == 0)
2290 			db_iprintf("memory:=");
2291 		else if (count == 6) {
2292 			db_printf("\n");
2293 			db_iprintf(" ...");
2294 			count = 0;
2295 		} else
2296 			db_printf(",");
2297 		count++;
2298 
2299 		db_printf("(off=0x%jx,page=0x%jx)",
2300 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2301 	}
2302 	if (count != 0)
2303 		db_printf("\n");
2304 	db_indent -= 2;
2305 }
2306 
2307 /* XXX. */
2308 #undef count
2309 
2310 /* XXX need this non-static entry for calling from vm_map_print. */
2311 void
2312 vm_object_print(
2313         /* db_expr_t */ long addr,
2314 	boolean_t have_addr,
2315 	/* db_expr_t */ long count,
2316 	char *modif)
2317 {
2318 	vm_object_print_static(addr, have_addr, count, modif);
2319 }
2320 
2321 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2322 {
2323 	vm_object_t object;
2324 	vm_pindex_t fidx;
2325 	vm_paddr_t pa;
2326 	vm_page_t m, prev_m;
2327 	int rcount, nl, c;
2328 
2329 	nl = 0;
2330 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2331 		db_printf("new object: %p\n", (void *)object);
2332 		if (nl > 18) {
2333 			c = cngetc();
2334 			if (c != ' ')
2335 				return;
2336 			nl = 0;
2337 		}
2338 		nl++;
2339 		rcount = 0;
2340 		fidx = 0;
2341 		pa = -1;
2342 		TAILQ_FOREACH(m, &object->memq, listq) {
2343 			if (m->pindex > 128)
2344 				break;
2345 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2346 			    prev_m->pindex + 1 != m->pindex) {
2347 				if (rcount) {
2348 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2349 						(long)fidx, rcount, (long)pa);
2350 					if (nl > 18) {
2351 						c = cngetc();
2352 						if (c != ' ')
2353 							return;
2354 						nl = 0;
2355 					}
2356 					nl++;
2357 					rcount = 0;
2358 				}
2359 			}
2360 			if (rcount &&
2361 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2362 				++rcount;
2363 				continue;
2364 			}
2365 			if (rcount) {
2366 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2367 					(long)fidx, rcount, (long)pa);
2368 				if (nl > 18) {
2369 					c = cngetc();
2370 					if (c != ' ')
2371 						return;
2372 					nl = 0;
2373 				}
2374 				nl++;
2375 			}
2376 			fidx = m->pindex;
2377 			pa = VM_PAGE_TO_PHYS(m);
2378 			rcount = 1;
2379 		}
2380 		if (rcount) {
2381 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2382 				(long)fidx, rcount, (long)pa);
2383 			if (nl > 18) {
2384 				c = cngetc();
2385 				if (c != ' ')
2386 					return;
2387 				nl = 0;
2388 			}
2389 			nl++;
2390 		}
2391 	}
2392 }
2393 #endif /* DDB */
2394