xref: /freebsd/sys/vm/vm_object.c (revision 4ed925457ab06e83238a5db33e89ccc94b99a713)
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 	vm_page_lock_queues();
721 	while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
722 		KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
723 			("vm_object_terminate: freeing busy page %p "
724 			"p->busy = %d, p->oflags %x\n", p, p->busy, p->oflags));
725 		if (p->wire_count == 0) {
726 			vm_page_free(p);
727 			cnt.v_pfree++;
728 		} else {
729 			vm_page_remove(p);
730 		}
731 	}
732 	vm_page_unlock_queues();
733 
734 #if VM_NRESERVLEVEL > 0
735 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
736 		vm_reserv_break_all(object);
737 #endif
738 	if (__predict_false(object->cache != NULL))
739 		vm_page_cache_free(object, 0, 0);
740 
741 	/*
742 	 * Let the pager know object is dead.
743 	 */
744 	vm_pager_deallocate(object);
745 	VM_OBJECT_UNLOCK(object);
746 
747 	vm_object_destroy(object);
748 }
749 
750 /*
751  *	vm_object_page_clean
752  *
753  *	Clean all dirty pages in the specified range of object.  Leaves page
754  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
755  *	write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
756  *	leaving the object dirty.
757  *
758  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
759  *	synchronous clustering mode implementation.
760  *
761  *	Odd semantics: if start == end, we clean everything.
762  *
763  *	The object must be locked.
764  */
765 void
766 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags)
767 {
768 	vm_page_t p, np;
769 	vm_pindex_t tstart, tend;
770 	vm_pindex_t pi;
771 	int clearobjflags;
772 	int pagerflags;
773 	int curgeneration;
774 
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 	vm_page_lock_queues();
793 	/*
794 	 * If the caller is smart and only msync()s a range he knows is
795 	 * dirty, we may be able to avoid an object scan.  This results in
796 	 * a phenominal improvement in performance.  We cannot do this
797 	 * as a matter of course because the object may be huge - e.g.
798 	 * the size might be in the gigabytes or terrabytes.
799 	 */
800 	if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
801 		vm_pindex_t tscan;
802 		int scanlimit;
803 		int scanreset;
804 
805 		scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
806 		if (scanreset < 16)
807 			scanreset = 16;
808 		pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
809 
810 		scanlimit = scanreset;
811 		tscan = tstart;
812 		while (tscan < tend) {
813 			curgeneration = object->generation;
814 			p = vm_page_lookup(object, tscan);
815 			if (p == NULL || p->valid == 0) {
816 				if (--scanlimit == 0)
817 					break;
818 				++tscan;
819 				continue;
820 			}
821 			vm_page_test_dirty(p);
822 			if (p->dirty == 0) {
823 				if (--scanlimit == 0)
824 					break;
825 				++tscan;
826 				continue;
827 			}
828 			/*
829 			 * If we have been asked to skip nosync pages and
830 			 * this is a nosync page, we can't continue.
831 			 */
832 			if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
833 				if (--scanlimit == 0)
834 					break;
835 				++tscan;
836 				continue;
837 			}
838 			scanlimit = scanreset;
839 
840 			/*
841 			 * This returns 0 if it was unable to busy the first
842 			 * page (i.e. had to sleep).
843 			 */
844 			tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
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_page_unlock_queues();
855 			vm_object_clear_flag(object, OBJ_CLEANING);
856 			return;
857 		}
858 		pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
859 	}
860 
861 	/*
862 	 * Generally set CLEANCHK interlock and make the page read-only so
863 	 * we can then clear the object flags.
864 	 *
865 	 * However, if this is a nosync mmap then the object is likely to
866 	 * stay dirty so do not mess with the page and do not clear the
867 	 * object flags.
868 	 */
869 	clearobjflags = 1;
870 	TAILQ_FOREACH(p, &object->memq, listq) {
871 		p->oflags |= VPO_CLEANCHK;
872 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC))
873 			clearobjflags = 0;
874 		else
875 			pmap_remove_write(p);
876 	}
877 
878 	if (clearobjflags && (tstart == 0) && (tend == object->size))
879 		vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
880 
881 rescan:
882 	curgeneration = object->generation;
883 
884 	for (p = TAILQ_FIRST(&object->memq); p; p = np) {
885 		int n;
886 
887 		np = TAILQ_NEXT(p, listq);
888 
889 again:
890 		pi = p->pindex;
891 		if ((p->oflags & VPO_CLEANCHK) == 0 ||
892 			(pi < tstart) || (pi >= tend) ||
893 		    p->valid == 0) {
894 			p->oflags &= ~VPO_CLEANCHK;
895 			continue;
896 		}
897 
898 		vm_page_test_dirty(p);
899 		if (p->dirty == 0) {
900 			p->oflags &= ~VPO_CLEANCHK;
901 			continue;
902 		}
903 
904 		/*
905 		 * If we have been asked to skip nosync pages and this is a
906 		 * nosync page, skip it.  Note that the object flags were
907 		 * not cleared in this case so we do not have to set them.
908 		 */
909 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
910 			p->oflags &= ~VPO_CLEANCHK;
911 			continue;
912 		}
913 
914 		n = vm_object_page_collect_flush(object, p,
915 			curgeneration, pagerflags);
916 		if (n == 0)
917 			goto rescan;
918 
919 		if (object->generation != curgeneration)
920 			goto rescan;
921 
922 		/*
923 		 * Try to optimize the next page.  If we can't we pick up
924 		 * our (random) scan where we left off.
925 		 */
926 		if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
927 			if ((p = vm_page_lookup(object, pi + n)) != NULL)
928 				goto again;
929 		}
930 	}
931 	vm_page_unlock_queues();
932 #if 0
933 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
934 #endif
935 
936 	vm_object_clear_flag(object, OBJ_CLEANING);
937 	return;
938 }
939 
940 static int
941 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
942 {
943 	int runlen;
944 	int maxf;
945 	int chkb;
946 	int maxb;
947 	int i;
948 	vm_pindex_t pi;
949 	vm_page_t maf[vm_pageout_page_count];
950 	vm_page_t mab[vm_pageout_page_count];
951 	vm_page_t ma[vm_pageout_page_count];
952 
953 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
954 	pi = p->pindex;
955 	while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
956 		vm_page_lock_queues();
957 		if (object->generation != curgeneration) {
958 			return(0);
959 		}
960 	}
961 	maxf = 0;
962 	for(i = 1; i < vm_pageout_page_count; i++) {
963 		vm_page_t tp;
964 
965 		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
966 			if ((tp->oflags & VPO_BUSY) ||
967 				((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
968 				 (tp->oflags & VPO_CLEANCHK) == 0) ||
969 				(tp->busy != 0))
970 				break;
971 			vm_page_test_dirty(tp);
972 			if (tp->dirty == 0) {
973 				tp->oflags &= ~VPO_CLEANCHK;
974 				break;
975 			}
976 			maf[ i - 1 ] = tp;
977 			maxf++;
978 			continue;
979 		}
980 		break;
981 	}
982 
983 	maxb = 0;
984 	chkb = vm_pageout_page_count -  maxf;
985 	if (chkb) {
986 		for(i = 1; i < chkb;i++) {
987 			vm_page_t tp;
988 
989 			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
990 				if ((tp->oflags & VPO_BUSY) ||
991 					((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
992 					 (tp->oflags & VPO_CLEANCHK) == 0) ||
993 					(tp->busy != 0))
994 					break;
995 				vm_page_test_dirty(tp);
996 				if (tp->dirty == 0) {
997 					tp->oflags &= ~VPO_CLEANCHK;
998 					break;
999 				}
1000 				mab[ i - 1 ] = tp;
1001 				maxb++;
1002 				continue;
1003 			}
1004 			break;
1005 		}
1006 	}
1007 
1008 	for(i = 0; i < maxb; i++) {
1009 		int index = (maxb - i) - 1;
1010 		ma[index] = mab[i];
1011 		ma[index]->oflags &= ~VPO_CLEANCHK;
1012 	}
1013 	p->oflags &= ~VPO_CLEANCHK;
1014 	ma[maxb] = p;
1015 	for(i = 0; i < maxf; i++) {
1016 		int index = (maxb + i) + 1;
1017 		ma[index] = maf[i];
1018 		ma[index]->oflags &= ~VPO_CLEANCHK;
1019 	}
1020 	runlen = maxb + maxf + 1;
1021 
1022 	vm_pageout_flush(ma, runlen, pagerflags);
1023 	for (i = 0; i < runlen; i++) {
1024 		if (ma[i]->dirty) {
1025 			pmap_remove_write(ma[i]);
1026 			ma[i]->oflags |= VPO_CLEANCHK;
1027 
1028 			/*
1029 			 * maxf will end up being the actual number of pages
1030 			 * we wrote out contiguously, non-inclusive of the
1031 			 * first page.  We do not count look-behind pages.
1032 			 */
1033 			if (i >= maxb + 1 && (maxf > i - maxb - 1))
1034 				maxf = i - maxb - 1;
1035 		}
1036 	}
1037 	return(maxf + 1);
1038 }
1039 
1040 /*
1041  * Note that there is absolutely no sense in writing out
1042  * anonymous objects, so we track down the vnode object
1043  * to write out.
1044  * We invalidate (remove) all pages from the address space
1045  * for semantic correctness.
1046  *
1047  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1048  * may start out with a NULL object.
1049  */
1050 void
1051 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1052     boolean_t syncio, boolean_t invalidate)
1053 {
1054 	vm_object_t backing_object;
1055 	struct vnode *vp;
1056 	struct mount *mp;
1057 	int flags;
1058 
1059 	if (object == NULL)
1060 		return;
1061 	VM_OBJECT_LOCK(object);
1062 	while ((backing_object = object->backing_object) != NULL) {
1063 		VM_OBJECT_LOCK(backing_object);
1064 		offset += object->backing_object_offset;
1065 		VM_OBJECT_UNLOCK(object);
1066 		object = backing_object;
1067 		if (object->size < OFF_TO_IDX(offset + size))
1068 			size = IDX_TO_OFF(object->size) - offset;
1069 	}
1070 	/*
1071 	 * Flush pages if writing is allowed, invalidate them
1072 	 * if invalidation requested.  Pages undergoing I/O
1073 	 * will be ignored by vm_object_page_remove().
1074 	 *
1075 	 * We cannot lock the vnode and then wait for paging
1076 	 * to complete without deadlocking against vm_fault.
1077 	 * Instead we simply call vm_object_page_remove() and
1078 	 * allow it to block internally on a page-by-page
1079 	 * basis when it encounters pages undergoing async
1080 	 * I/O.
1081 	 */
1082 	if (object->type == OBJT_VNODE &&
1083 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1084 		int vfslocked;
1085 		vp = object->handle;
1086 		VM_OBJECT_UNLOCK(object);
1087 		(void) vn_start_write(vp, &mp, V_WAIT);
1088 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1089 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1090 		flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1091 		flags |= invalidate ? OBJPC_INVAL : 0;
1092 		VM_OBJECT_LOCK(object);
1093 		vm_object_page_clean(object,
1094 		    OFF_TO_IDX(offset),
1095 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1096 		    flags);
1097 		VM_OBJECT_UNLOCK(object);
1098 		VOP_UNLOCK(vp, 0);
1099 		VFS_UNLOCK_GIANT(vfslocked);
1100 		vn_finished_write(mp);
1101 		VM_OBJECT_LOCK(object);
1102 	}
1103 	if ((object->type == OBJT_VNODE ||
1104 	     object->type == OBJT_DEVICE) && invalidate) {
1105 		boolean_t purge;
1106 		purge = old_msync || (object->type == OBJT_DEVICE);
1107 		vm_object_page_remove(object,
1108 		    OFF_TO_IDX(offset),
1109 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1110 		    purge ? FALSE : TRUE);
1111 	}
1112 	VM_OBJECT_UNLOCK(object);
1113 }
1114 
1115 /*
1116  *	vm_object_madvise:
1117  *
1118  *	Implements the madvise function at the object/page level.
1119  *
1120  *	MADV_WILLNEED	(any object)
1121  *
1122  *	    Activate the specified pages if they are resident.
1123  *
1124  *	MADV_DONTNEED	(any object)
1125  *
1126  *	    Deactivate the specified pages if they are resident.
1127  *
1128  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1129  *			 OBJ_ONEMAPPING only)
1130  *
1131  *	    Deactivate and clean the specified pages if they are
1132  *	    resident.  This permits the process to reuse the pages
1133  *	    without faulting or the kernel to reclaim the pages
1134  *	    without I/O.
1135  */
1136 void
1137 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1138 {
1139 	vm_pindex_t end, tpindex;
1140 	vm_object_t backing_object, tobject;
1141 	vm_page_t m;
1142 
1143 	if (object == NULL)
1144 		return;
1145 	VM_OBJECT_LOCK(object);
1146 	end = pindex + count;
1147 	/*
1148 	 * Locate and adjust resident pages
1149 	 */
1150 	for (; pindex < end; pindex += 1) {
1151 relookup:
1152 		tobject = object;
1153 		tpindex = pindex;
1154 shadowlookup:
1155 		/*
1156 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1157 		 * and those pages must be OBJ_ONEMAPPING.
1158 		 */
1159 		if (advise == MADV_FREE) {
1160 			if ((tobject->type != OBJT_DEFAULT &&
1161 			     tobject->type != OBJT_SWAP) ||
1162 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
1163 				goto unlock_tobject;
1164 			}
1165 		}
1166 		m = vm_page_lookup(tobject, tpindex);
1167 		if (m == NULL && advise == MADV_WILLNEED) {
1168 			/*
1169 			 * If the page is cached, reactivate it.
1170 			 */
1171 			m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1172 			    VM_ALLOC_NOBUSY);
1173 		}
1174 		if (m == NULL) {
1175 			/*
1176 			 * There may be swap even if there is no backing page
1177 			 */
1178 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1179 				swap_pager_freespace(tobject, tpindex, 1);
1180 			/*
1181 			 * next object
1182 			 */
1183 			backing_object = tobject->backing_object;
1184 			if (backing_object == NULL)
1185 				goto unlock_tobject;
1186 			VM_OBJECT_LOCK(backing_object);
1187 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1188 			if (tobject != object)
1189 				VM_OBJECT_UNLOCK(tobject);
1190 			tobject = backing_object;
1191 			goto shadowlookup;
1192 		}
1193 		/*
1194 		 * If the page is busy or not in a normal active state,
1195 		 * we skip it.  If the page is not managed there are no
1196 		 * page queues to mess with.  Things can break if we mess
1197 		 * with pages in any of the below states.
1198 		 */
1199 		vm_page_lock_queues();
1200 		if (m->hold_count ||
1201 		    m->wire_count ||
1202 		    (m->flags & PG_UNMANAGED) ||
1203 		    m->valid != VM_PAGE_BITS_ALL) {
1204 			vm_page_unlock_queues();
1205 			goto unlock_tobject;
1206 		}
1207 		if ((m->oflags & VPO_BUSY) || m->busy) {
1208 			vm_page_flag_set(m, PG_REFERENCED);
1209 			vm_page_unlock_queues();
1210 			if (object != tobject)
1211 				VM_OBJECT_UNLOCK(object);
1212 			m->oflags |= VPO_WANTED;
1213 			msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo", 0);
1214 			VM_OBJECT_LOCK(object);
1215   			goto relookup;
1216 		}
1217 		if (advise == MADV_WILLNEED) {
1218 			vm_page_activate(m);
1219 		} else if (advise == MADV_DONTNEED) {
1220 			vm_page_dontneed(m);
1221 		} else if (advise == MADV_FREE) {
1222 			/*
1223 			 * Mark the page clean.  This will allow the page
1224 			 * to be freed up by the system.  However, such pages
1225 			 * are often reused quickly by malloc()/free()
1226 			 * so we do not do anything that would cause
1227 			 * a page fault if we can help it.
1228 			 *
1229 			 * Specifically, we do not try to actually free
1230 			 * the page now nor do we try to put it in the
1231 			 * cache (which would cause a page fault on reuse).
1232 			 *
1233 			 * But we do make the page is freeable as we
1234 			 * can without actually taking the step of unmapping
1235 			 * it.
1236 			 */
1237 			pmap_clear_modify(m);
1238 			m->dirty = 0;
1239 			m->act_count = 0;
1240 			vm_page_dontneed(m);
1241 		}
1242 		vm_page_unlock_queues();
1243 		if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1244 			swap_pager_freespace(tobject, tpindex, 1);
1245 unlock_tobject:
1246 		if (tobject != object)
1247 			VM_OBJECT_UNLOCK(tobject);
1248 	}
1249 	VM_OBJECT_UNLOCK(object);
1250 }
1251 
1252 /*
1253  *	vm_object_shadow:
1254  *
1255  *	Create a new object which is backed by the
1256  *	specified existing object range.  The source
1257  *	object reference is deallocated.
1258  *
1259  *	The new object and offset into that object
1260  *	are returned in the source parameters.
1261  */
1262 void
1263 vm_object_shadow(
1264 	vm_object_t *object,	/* IN/OUT */
1265 	vm_ooffset_t *offset,	/* IN/OUT */
1266 	vm_size_t length)
1267 {
1268 	vm_object_t source;
1269 	vm_object_t result;
1270 
1271 	source = *object;
1272 
1273 	/*
1274 	 * Don't create the new object if the old object isn't shared.
1275 	 */
1276 	if (source != NULL) {
1277 		VM_OBJECT_LOCK(source);
1278 		if (source->ref_count == 1 &&
1279 		    source->handle == NULL &&
1280 		    (source->type == OBJT_DEFAULT ||
1281 		     source->type == OBJT_SWAP)) {
1282 			VM_OBJECT_UNLOCK(source);
1283 			return;
1284 		}
1285 		VM_OBJECT_UNLOCK(source);
1286 	}
1287 
1288 	/*
1289 	 * Allocate a new object with the given length.
1290 	 */
1291 	result = vm_object_allocate(OBJT_DEFAULT, length);
1292 
1293 	/*
1294 	 * The new object shadows the source object, adding a reference to it.
1295 	 * Our caller changes his reference to point to the new object,
1296 	 * removing a reference to the source object.  Net result: no change
1297 	 * of reference count.
1298 	 *
1299 	 * Try to optimize the result object's page color when shadowing
1300 	 * in order to maintain page coloring consistency in the combined
1301 	 * shadowed object.
1302 	 */
1303 	result->backing_object = source;
1304 	/*
1305 	 * Store the offset into the source object, and fix up the offset into
1306 	 * the new object.
1307 	 */
1308 	result->backing_object_offset = *offset;
1309 	if (source != NULL) {
1310 		VM_OBJECT_LOCK(source);
1311 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1312 		source->shadow_count++;
1313 		source->generation++;
1314 #if VM_NRESERVLEVEL > 0
1315 		result->flags |= source->flags & OBJ_COLORED;
1316 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1317 		    ((1 << (VM_NFREEORDER - 1)) - 1);
1318 #endif
1319 		VM_OBJECT_UNLOCK(source);
1320 	}
1321 
1322 
1323 	/*
1324 	 * Return the new things
1325 	 */
1326 	*offset = 0;
1327 	*object = result;
1328 }
1329 
1330 /*
1331  *	vm_object_split:
1332  *
1333  * Split the pages in a map entry into a new object.  This affords
1334  * easier removal of unused pages, and keeps object inheritance from
1335  * being a negative impact on memory usage.
1336  */
1337 void
1338 vm_object_split(vm_map_entry_t entry)
1339 {
1340 	vm_page_t m, m_next;
1341 	vm_object_t orig_object, new_object, source;
1342 	vm_pindex_t idx, offidxstart;
1343 	vm_size_t size;
1344 
1345 	orig_object = entry->object.vm_object;
1346 	if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1347 		return;
1348 	if (orig_object->ref_count <= 1)
1349 		return;
1350 	VM_OBJECT_UNLOCK(orig_object);
1351 
1352 	offidxstart = OFF_TO_IDX(entry->offset);
1353 	size = atop(entry->end - entry->start);
1354 
1355 	/*
1356 	 * If swap_pager_copy() is later called, it will convert new_object
1357 	 * into a swap object.
1358 	 */
1359 	new_object = vm_object_allocate(OBJT_DEFAULT, size);
1360 
1361 	/*
1362 	 * At this point, the new object is still private, so the order in
1363 	 * which the original and new objects are locked does not matter.
1364 	 */
1365 	VM_OBJECT_LOCK(new_object);
1366 	VM_OBJECT_LOCK(orig_object);
1367 	source = orig_object->backing_object;
1368 	if (source != NULL) {
1369 		VM_OBJECT_LOCK(source);
1370 		if ((source->flags & OBJ_DEAD) != 0) {
1371 			VM_OBJECT_UNLOCK(source);
1372 			VM_OBJECT_UNLOCK(orig_object);
1373 			VM_OBJECT_UNLOCK(new_object);
1374 			vm_object_deallocate(new_object);
1375 			VM_OBJECT_LOCK(orig_object);
1376 			return;
1377 		}
1378 		LIST_INSERT_HEAD(&source->shadow_head,
1379 				  new_object, shadow_list);
1380 		source->shadow_count++;
1381 		source->generation++;
1382 		vm_object_reference_locked(source);	/* for new_object */
1383 		vm_object_clear_flag(source, OBJ_ONEMAPPING);
1384 		VM_OBJECT_UNLOCK(source);
1385 		new_object->backing_object_offset =
1386 			orig_object->backing_object_offset + entry->offset;
1387 		new_object->backing_object = source;
1388 	}
1389 	if (orig_object->uip != NULL) {
1390 		new_object->uip = orig_object->uip;
1391 		uihold(orig_object->uip);
1392 		new_object->charge = ptoa(size);
1393 		KASSERT(orig_object->charge >= ptoa(size),
1394 		    ("orig_object->charge < 0"));
1395 		orig_object->charge -= ptoa(size);
1396 	}
1397 retry:
1398 	if ((m = TAILQ_FIRST(&orig_object->memq)) != NULL) {
1399 		if (m->pindex < offidxstart) {
1400 			m = vm_page_splay(offidxstart, orig_object->root);
1401 			if ((orig_object->root = m)->pindex < offidxstart)
1402 				m = TAILQ_NEXT(m, listq);
1403 		}
1404 	}
1405 	vm_page_lock_queues();
1406 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1407 	    m = m_next) {
1408 		m_next = TAILQ_NEXT(m, listq);
1409 
1410 		/*
1411 		 * We must wait for pending I/O to complete before we can
1412 		 * rename the page.
1413 		 *
1414 		 * We do not have to VM_PROT_NONE the page as mappings should
1415 		 * not be changed by this operation.
1416 		 */
1417 		if ((m->oflags & VPO_BUSY) || m->busy) {
1418 			vm_page_flag_set(m, PG_REFERENCED);
1419 			vm_page_unlock_queues();
1420 			VM_OBJECT_UNLOCK(new_object);
1421 			m->oflags |= VPO_WANTED;
1422 			msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1423 			VM_OBJECT_LOCK(new_object);
1424 			goto retry;
1425 		}
1426 		vm_page_rename(m, new_object, idx);
1427 		/* page automatically made dirty by rename and cache handled */
1428 		vm_page_busy(m);
1429 	}
1430 	vm_page_unlock_queues();
1431 	if (orig_object->type == OBJT_SWAP) {
1432 		/*
1433 		 * swap_pager_copy() can sleep, in which case the orig_object's
1434 		 * and new_object's locks are released and reacquired.
1435 		 */
1436 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1437 
1438 		/*
1439 		 * Transfer any cached pages from orig_object to new_object.
1440 		 */
1441 		if (__predict_false(orig_object->cache != NULL))
1442 			vm_page_cache_transfer(orig_object, offidxstart,
1443 			    new_object);
1444 	}
1445 	VM_OBJECT_UNLOCK(orig_object);
1446 	TAILQ_FOREACH(m, &new_object->memq, listq)
1447 		vm_page_wakeup(m);
1448 	VM_OBJECT_UNLOCK(new_object);
1449 	entry->object.vm_object = new_object;
1450 	entry->offset = 0LL;
1451 	vm_object_deallocate(orig_object);
1452 	VM_OBJECT_LOCK(new_object);
1453 }
1454 
1455 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1456 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1457 #define	OBSC_COLLAPSE_WAIT	0x0004
1458 
1459 static int
1460 vm_object_backing_scan(vm_object_t object, int op)
1461 {
1462 	int r = 1;
1463 	vm_page_t p;
1464 	vm_object_t backing_object;
1465 	vm_pindex_t backing_offset_index;
1466 
1467 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1468 	VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1469 
1470 	backing_object = object->backing_object;
1471 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1472 
1473 	/*
1474 	 * Initial conditions
1475 	 */
1476 	if (op & OBSC_TEST_ALL_SHADOWED) {
1477 		/*
1478 		 * We do not want to have to test for the existence of cache
1479 		 * or swap pages in the backing object.  XXX but with the
1480 		 * new swapper this would be pretty easy to do.
1481 		 *
1482 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1483 		 * been ZFOD faulted yet?  If we do not test for this, the
1484 		 * shadow test may succeed! XXX
1485 		 */
1486 		if (backing_object->type != OBJT_DEFAULT) {
1487 			return (0);
1488 		}
1489 	}
1490 	if (op & OBSC_COLLAPSE_WAIT) {
1491 		vm_object_set_flag(backing_object, OBJ_DEAD);
1492 	}
1493 
1494 	/*
1495 	 * Our scan
1496 	 */
1497 	p = TAILQ_FIRST(&backing_object->memq);
1498 	while (p) {
1499 		vm_page_t next = TAILQ_NEXT(p, listq);
1500 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1501 
1502 		if (op & OBSC_TEST_ALL_SHADOWED) {
1503 			vm_page_t pp;
1504 
1505 			/*
1506 			 * Ignore pages outside the parent object's range
1507 			 * and outside the parent object's mapping of the
1508 			 * backing object.
1509 			 *
1510 			 * note that we do not busy the backing object's
1511 			 * page.
1512 			 */
1513 			if (
1514 			    p->pindex < backing_offset_index ||
1515 			    new_pindex >= object->size
1516 			) {
1517 				p = next;
1518 				continue;
1519 			}
1520 
1521 			/*
1522 			 * See if the parent has the page or if the parent's
1523 			 * object pager has the page.  If the parent has the
1524 			 * page but the page is not valid, the parent's
1525 			 * object pager must have the page.
1526 			 *
1527 			 * If this fails, the parent does not completely shadow
1528 			 * the object and we might as well give up now.
1529 			 */
1530 
1531 			pp = vm_page_lookup(object, new_pindex);
1532 			if (
1533 			    (pp == NULL || pp->valid == 0) &&
1534 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1535 			) {
1536 				r = 0;
1537 				break;
1538 			}
1539 		}
1540 
1541 		/*
1542 		 * Check for busy page
1543 		 */
1544 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1545 			vm_page_t pp;
1546 
1547 			if (op & OBSC_COLLAPSE_NOWAIT) {
1548 				if ((p->oflags & VPO_BUSY) ||
1549 				    !p->valid ||
1550 				    p->busy) {
1551 					p = next;
1552 					continue;
1553 				}
1554 			} else if (op & OBSC_COLLAPSE_WAIT) {
1555 				if ((p->oflags & VPO_BUSY) || p->busy) {
1556 					vm_page_lock_queues();
1557 					vm_page_flag_set(p, PG_REFERENCED);
1558 					vm_page_unlock_queues();
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_queues();
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_queues();
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_queues();
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_queues();
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_queues();
1653 			vm_page_rename(p, object, new_pindex);
1654 			vm_page_unlock_queues();
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.
1887  *
1888  *	The object must be locked.
1889  */
1890 void
1891 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1892     boolean_t clean_only)
1893 {
1894 	vm_page_t p, next;
1895 	int wirings;
1896 
1897 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1898 	if (object->resident_page_count == 0)
1899 		goto skipmemq;
1900 
1901 	/*
1902 	 * Since physically-backed objects do not use managed pages, we can't
1903 	 * remove pages from the object (we must instead remove the page
1904 	 * references, and then destroy the object).
1905 	 */
1906 	KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1907 	    object == kmem_object,
1908 	    ("attempt to remove pages from a physical object"));
1909 
1910 	vm_object_pip_add(object, 1);
1911 again:
1912 	if ((p = TAILQ_FIRST(&object->memq)) != NULL) {
1913 		if (p->pindex < start) {
1914 			p = vm_page_splay(start, object->root);
1915 			if ((object->root = p)->pindex < start)
1916 				p = TAILQ_NEXT(p, listq);
1917 		}
1918 	}
1919 	vm_page_lock_queues();
1920 	/*
1921 	 * Assert: the variable p is either (1) the page with the
1922 	 * least pindex greater than or equal to the parameter pindex
1923 	 * or (2) NULL.
1924 	 */
1925 	for (;
1926 	     p != NULL && (p->pindex < end || end == 0);
1927 	     p = next) {
1928 		next = TAILQ_NEXT(p, listq);
1929 
1930 		/*
1931 		 * If the page is wired for any reason besides the
1932 		 * existence of managed, wired mappings, then it cannot
1933 		 * be freed.  For example, fictitious pages, which
1934 		 * represent device memory, are inherently wired and
1935 		 * cannot be freed.  They can, however, be invalidated
1936 		 * if "clean_only" is FALSE.
1937 		 */
1938 		if ((wirings = p->wire_count) != 0 &&
1939 		    (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1940 			/* Fictitious pages do not have managed mappings. */
1941 			if ((p->flags & PG_FICTITIOUS) == 0)
1942 				pmap_remove_all(p);
1943 			/* Account for removal of managed, wired mappings. */
1944 			p->wire_count -= wirings;
1945 			if (!clean_only) {
1946 				p->valid = 0;
1947 				vm_page_undirty(p);
1948 			}
1949 			continue;
1950 		}
1951 		if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1952 			goto again;
1953 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
1954 		    ("vm_object_page_remove: page %p is fictitious", p));
1955 		if (clean_only && p->valid) {
1956 			pmap_remove_write(p);
1957 			if (p->dirty)
1958 				continue;
1959 		}
1960 		pmap_remove_all(p);
1961 		/* Account for removal of managed, wired mappings. */
1962 		if (wirings != 0)
1963 			p->wire_count -= wirings;
1964 		vm_page_free(p);
1965 	}
1966 	vm_page_unlock_queues();
1967 	vm_object_pip_wakeup(object);
1968 skipmemq:
1969 	if (__predict_false(object->cache != NULL))
1970 		vm_page_cache_free(object, start, end);
1971 }
1972 
1973 /*
1974  *	Populate the specified range of the object with valid pages.  Returns
1975  *	TRUE if the range is successfully populated and FALSE otherwise.
1976  *
1977  *	Note: This function should be optimized to pass a larger array of
1978  *	pages to vm_pager_get_pages() before it is applied to a non-
1979  *	OBJT_DEVICE object.
1980  *
1981  *	The object must be locked.
1982  */
1983 boolean_t
1984 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1985 {
1986 	vm_page_t m, ma[1];
1987 	vm_pindex_t pindex;
1988 	int rv;
1989 
1990 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1991 	for (pindex = start; pindex < end; pindex++) {
1992 		m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1993 		    VM_ALLOC_RETRY);
1994 		if (m->valid != VM_PAGE_BITS_ALL) {
1995 			ma[0] = m;
1996 			rv = vm_pager_get_pages(object, ma, 1, 0);
1997 			m = vm_page_lookup(object, pindex);
1998 			if (m == NULL)
1999 				break;
2000 			if (rv != VM_PAGER_OK) {
2001 				vm_page_lock_queues();
2002 				vm_page_free(m);
2003 				vm_page_unlock_queues();
2004 				break;
2005 			}
2006 		}
2007 		/*
2008 		 * Keep "m" busy because a subsequent iteration may unlock
2009 		 * the object.
2010 		 */
2011 	}
2012 	if (pindex > start) {
2013 		m = vm_page_lookup(object, start);
2014 		while (m != NULL && m->pindex < pindex) {
2015 			vm_page_wakeup(m);
2016 			m = TAILQ_NEXT(m, listq);
2017 		}
2018 	}
2019 	return (pindex == end);
2020 }
2021 
2022 /*
2023  *	Routine:	vm_object_coalesce
2024  *	Function:	Coalesces two objects backing up adjoining
2025  *			regions of memory into a single object.
2026  *
2027  *	returns TRUE if objects were combined.
2028  *
2029  *	NOTE:	Only works at the moment if the second object is NULL -
2030  *		if it's not, which object do we lock first?
2031  *
2032  *	Parameters:
2033  *		prev_object	First object to coalesce
2034  *		prev_offset	Offset into prev_object
2035  *		prev_size	Size of reference to prev_object
2036  *		next_size	Size of reference to the second object
2037  *		reserved	Indicator that extension region has
2038  *				swap accounted for
2039  *
2040  *	Conditions:
2041  *	The object must *not* be locked.
2042  */
2043 boolean_t
2044 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2045     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2046 {
2047 	vm_pindex_t next_pindex;
2048 
2049 	if (prev_object == NULL)
2050 		return (TRUE);
2051 	VM_OBJECT_LOCK(prev_object);
2052 	if (prev_object->type != OBJT_DEFAULT &&
2053 	    prev_object->type != OBJT_SWAP) {
2054 		VM_OBJECT_UNLOCK(prev_object);
2055 		return (FALSE);
2056 	}
2057 
2058 	/*
2059 	 * Try to collapse the object first
2060 	 */
2061 	vm_object_collapse(prev_object);
2062 
2063 	/*
2064 	 * Can't coalesce if: . more than one reference . paged out . shadows
2065 	 * another object . has a copy elsewhere (any of which mean that the
2066 	 * pages not mapped to prev_entry may be in use anyway)
2067 	 */
2068 	if (prev_object->backing_object != NULL) {
2069 		VM_OBJECT_UNLOCK(prev_object);
2070 		return (FALSE);
2071 	}
2072 
2073 	prev_size >>= PAGE_SHIFT;
2074 	next_size >>= PAGE_SHIFT;
2075 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2076 
2077 	if ((prev_object->ref_count > 1) &&
2078 	    (prev_object->size != next_pindex)) {
2079 		VM_OBJECT_UNLOCK(prev_object);
2080 		return (FALSE);
2081 	}
2082 
2083 	/*
2084 	 * Account for the charge.
2085 	 */
2086 	if (prev_object->uip != NULL) {
2087 
2088 		/*
2089 		 * If prev_object was charged, then this mapping,
2090 		 * althought not charged now, may become writable
2091 		 * later. Non-NULL uip in the object would prevent
2092 		 * swap reservation during enabling of the write
2093 		 * access, so reserve swap now. Failed reservation
2094 		 * cause allocation of the separate object for the map
2095 		 * entry, and swap reservation for this entry is
2096 		 * managed in appropriate time.
2097 		 */
2098 		if (!reserved && !swap_reserve_by_uid(ptoa(next_size),
2099 		    prev_object->uip)) {
2100 			return (FALSE);
2101 		}
2102 		prev_object->charge += ptoa(next_size);
2103 	}
2104 
2105 	/*
2106 	 * Remove any pages that may still be in the object from a previous
2107 	 * deallocation.
2108 	 */
2109 	if (next_pindex < prev_object->size) {
2110 		vm_object_page_remove(prev_object,
2111 				      next_pindex,
2112 				      next_pindex + next_size, FALSE);
2113 		if (prev_object->type == OBJT_SWAP)
2114 			swap_pager_freespace(prev_object,
2115 					     next_pindex, next_size);
2116 #if 0
2117 		if (prev_object->uip != NULL) {
2118 			KASSERT(prev_object->charge >=
2119 			    ptoa(prev_object->size - next_pindex),
2120 			    ("object %p overcharged 1 %jx %jx", prev_object,
2121 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2122 			prev_object->charge -= ptoa(prev_object->size -
2123 			    next_pindex);
2124 		}
2125 #endif
2126 	}
2127 
2128 	/*
2129 	 * Extend the object if necessary.
2130 	 */
2131 	if (next_pindex + next_size > prev_object->size)
2132 		prev_object->size = next_pindex + next_size;
2133 
2134 	VM_OBJECT_UNLOCK(prev_object);
2135 	return (TRUE);
2136 }
2137 
2138 void
2139 vm_object_set_writeable_dirty(vm_object_t object)
2140 {
2141 
2142 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2143 	if (object->type != OBJT_VNODE ||
2144 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0)
2145 		return;
2146 	vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2147 }
2148 
2149 #include "opt_ddb.h"
2150 #ifdef DDB
2151 #include <sys/kernel.h>
2152 
2153 #include <sys/cons.h>
2154 
2155 #include <ddb/ddb.h>
2156 
2157 static int
2158 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2159 {
2160 	vm_map_t tmpm;
2161 	vm_map_entry_t tmpe;
2162 	vm_object_t obj;
2163 	int entcount;
2164 
2165 	if (map == 0)
2166 		return 0;
2167 
2168 	if (entry == 0) {
2169 		tmpe = map->header.next;
2170 		entcount = map->nentries;
2171 		while (entcount-- && (tmpe != &map->header)) {
2172 			if (_vm_object_in_map(map, object, tmpe)) {
2173 				return 1;
2174 			}
2175 			tmpe = tmpe->next;
2176 		}
2177 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2178 		tmpm = entry->object.sub_map;
2179 		tmpe = tmpm->header.next;
2180 		entcount = tmpm->nentries;
2181 		while (entcount-- && tmpe != &tmpm->header) {
2182 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2183 				return 1;
2184 			}
2185 			tmpe = tmpe->next;
2186 		}
2187 	} else if ((obj = entry->object.vm_object) != NULL) {
2188 		for (; obj; obj = obj->backing_object)
2189 			if (obj == object) {
2190 				return 1;
2191 			}
2192 	}
2193 	return 0;
2194 }
2195 
2196 static int
2197 vm_object_in_map(vm_object_t object)
2198 {
2199 	struct proc *p;
2200 
2201 	/* sx_slock(&allproc_lock); */
2202 	FOREACH_PROC_IN_SYSTEM(p) {
2203 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2204 			continue;
2205 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2206 			/* sx_sunlock(&allproc_lock); */
2207 			return 1;
2208 		}
2209 	}
2210 	/* sx_sunlock(&allproc_lock); */
2211 	if (_vm_object_in_map(kernel_map, object, 0))
2212 		return 1;
2213 	if (_vm_object_in_map(kmem_map, object, 0))
2214 		return 1;
2215 	if (_vm_object_in_map(pager_map, object, 0))
2216 		return 1;
2217 	if (_vm_object_in_map(buffer_map, object, 0))
2218 		return 1;
2219 	return 0;
2220 }
2221 
2222 DB_SHOW_COMMAND(vmochk, vm_object_check)
2223 {
2224 	vm_object_t object;
2225 
2226 	/*
2227 	 * make sure that internal objs are in a map somewhere
2228 	 * and none have zero ref counts.
2229 	 */
2230 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2231 		if (object->handle == NULL &&
2232 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2233 			if (object->ref_count == 0) {
2234 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2235 					(long)object->size);
2236 			}
2237 			if (!vm_object_in_map(object)) {
2238 				db_printf(
2239 			"vmochk: internal obj is not in a map: "
2240 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2241 				    object->ref_count, (u_long)object->size,
2242 				    (u_long)object->size,
2243 				    (void *)object->backing_object);
2244 			}
2245 		}
2246 	}
2247 }
2248 
2249 /*
2250  *	vm_object_print:	[ debug ]
2251  */
2252 DB_SHOW_COMMAND(object, vm_object_print_static)
2253 {
2254 	/* XXX convert args. */
2255 	vm_object_t object = (vm_object_t)addr;
2256 	boolean_t full = have_addr;
2257 
2258 	vm_page_t p;
2259 
2260 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2261 #define	count	was_count
2262 
2263 	int count;
2264 
2265 	if (object == NULL)
2266 		return;
2267 
2268 	db_iprintf(
2269 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x uip %d charge %jx\n",
2270 	    object, (int)object->type, (uintmax_t)object->size,
2271 	    object->resident_page_count, object->ref_count, object->flags,
2272 	    object->uip ? object->uip->ui_uid : -1, (uintmax_t)object->charge);
2273 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2274 	    object->shadow_count,
2275 	    object->backing_object ? object->backing_object->ref_count : 0,
2276 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2277 
2278 	if (!full)
2279 		return;
2280 
2281 	db_indent += 2;
2282 	count = 0;
2283 	TAILQ_FOREACH(p, &object->memq, listq) {
2284 		if (count == 0)
2285 			db_iprintf("memory:=");
2286 		else if (count == 6) {
2287 			db_printf("\n");
2288 			db_iprintf(" ...");
2289 			count = 0;
2290 		} else
2291 			db_printf(",");
2292 		count++;
2293 
2294 		db_printf("(off=0x%jx,page=0x%jx)",
2295 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2296 	}
2297 	if (count != 0)
2298 		db_printf("\n");
2299 	db_indent -= 2;
2300 }
2301 
2302 /* XXX. */
2303 #undef count
2304 
2305 /* XXX need this non-static entry for calling from vm_map_print. */
2306 void
2307 vm_object_print(
2308         /* db_expr_t */ long addr,
2309 	boolean_t have_addr,
2310 	/* db_expr_t */ long count,
2311 	char *modif)
2312 {
2313 	vm_object_print_static(addr, have_addr, count, modif);
2314 }
2315 
2316 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2317 {
2318 	vm_object_t object;
2319 	vm_pindex_t fidx;
2320 	vm_paddr_t pa;
2321 	vm_page_t m, prev_m;
2322 	int rcount, nl, c;
2323 
2324 	nl = 0;
2325 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2326 		db_printf("new object: %p\n", (void *)object);
2327 		if (nl > 18) {
2328 			c = cngetc();
2329 			if (c != ' ')
2330 				return;
2331 			nl = 0;
2332 		}
2333 		nl++;
2334 		rcount = 0;
2335 		fidx = 0;
2336 		pa = -1;
2337 		TAILQ_FOREACH(m, &object->memq, listq) {
2338 			if (m->pindex > 128)
2339 				break;
2340 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2341 			    prev_m->pindex + 1 != m->pindex) {
2342 				if (rcount) {
2343 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2344 						(long)fidx, rcount, (long)pa);
2345 					if (nl > 18) {
2346 						c = cngetc();
2347 						if (c != ' ')
2348 							return;
2349 						nl = 0;
2350 					}
2351 					nl++;
2352 					rcount = 0;
2353 				}
2354 			}
2355 			if (rcount &&
2356 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2357 				++rcount;
2358 				continue;
2359 			}
2360 			if (rcount) {
2361 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2362 					(long)fidx, rcount, (long)pa);
2363 				if (nl > 18) {
2364 					c = cngetc();
2365 					if (c != ' ')
2366 						return;
2367 					nl = 0;
2368 				}
2369 				nl++;
2370 			}
2371 			fidx = m->pindex;
2372 			pa = VM_PAGE_TO_PHYS(m);
2373 			rcount = 1;
2374 		}
2375 		if (rcount) {
2376 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2377 				(long)fidx, rcount, (long)pa);
2378 			if (nl > 18) {
2379 				c = cngetc();
2380 				if (c != ' ')
2381 					return;
2382 				nl = 0;
2383 			}
2384 			nl++;
2385 		}
2386 	}
2387 }
2388 #endif /* DDB */
2389