xref: /freebsd/sys/vm/vm_object.c (revision dd48af360fdbbb9552f9fc6de7abe50d68ad5331)
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 		vm_page_lock_queues();
726 		if (p->wire_count == 0) {
727 			vm_page_free(p);
728 			cnt.v_pfree++;
729 		} else {
730 			vm_page_remove(p);
731 		}
732 		vm_page_unlock_queues();
733 		vm_page_unlock(p);
734 	}
735 
736 #if VM_NRESERVLEVEL > 0
737 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
738 		vm_reserv_break_all(object);
739 #endif
740 	if (__predict_false(object->cache != NULL))
741 		vm_page_cache_free(object, 0, 0);
742 
743 	/*
744 	 * Let the pager know object is dead.
745 	 */
746 	vm_pager_deallocate(object);
747 	VM_OBJECT_UNLOCK(object);
748 
749 	vm_object_destroy(object);
750 }
751 
752 /*
753  *	vm_object_page_clean
754  *
755  *	Clean all dirty pages in the specified range of object.  Leaves page
756  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
757  *	write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
758  *	leaving the object dirty.
759  *
760  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
761  *	synchronous clustering mode implementation.
762  *
763  *	Odd semantics: if start == end, we clean everything.
764  *
765  *	The object must be locked.
766  */
767 void
768 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags)
769 {
770 	vm_page_t p, np;
771 	vm_pindex_t tstart, tend;
772 	vm_pindex_t pi;
773 	int clearobjflags;
774 	int pagerflags;
775 	int curgeneration;
776 
777 	mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
778 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
779 	if ((object->flags & OBJ_MIGHTBEDIRTY) == 0)
780 		return;
781 	KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
782 
783 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
784 	pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
785 
786 	vm_object_set_flag(object, OBJ_CLEANING);
787 
788 	tstart = start;
789 	if (end == 0) {
790 		tend = object->size;
791 	} else {
792 		tend = end;
793 	}
794 
795 	/*
796 	 * If the caller is smart and only msync()s a range he knows is
797 	 * dirty, we may be able to avoid an object scan.  This results in
798 	 * a phenominal improvement in performance.  We cannot do this
799 	 * as a matter of course because the object may be huge - e.g.
800 	 * the size might be in the gigabytes or terrabytes.
801 	 */
802 	if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
803 		vm_pindex_t tscan;
804 		int scanlimit;
805 		int scanreset;
806 
807 		scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
808 		if (scanreset < 16)
809 			scanreset = 16;
810 		pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
811 
812 		scanlimit = scanreset;
813 		tscan = tstart;
814 		while (tscan < tend) {
815 			curgeneration = object->generation;
816 			p = vm_page_lookup(object, tscan);
817 			if (p == NULL || p->valid == 0) {
818 				if (--scanlimit == 0)
819 					break;
820 				++tscan;
821 				continue;
822 			}
823 			vm_page_lock(p);
824 			vm_page_lock_queues();
825 			vm_page_test_dirty(p);
826 			if (p->dirty == 0) {
827 				vm_page_unlock_queues();
828 				vm_page_unlock(p);
829 				if (--scanlimit == 0)
830 					break;
831 				++tscan;
832 				continue;
833 			}
834 			vm_page_unlock_queues();
835 			vm_page_unlock(p);
836 			/*
837 			 * If we have been asked to skip nosync pages and
838 			 * this is a nosync page, we can't continue.
839 			 */
840 			if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
841 				if (--scanlimit == 0)
842 					break;
843 				++tscan;
844 				continue;
845 			}
846 			scanlimit = scanreset;
847 
848 			/*
849 			 * This returns 0 if it was unable to busy the first
850 			 * page (i.e. had to sleep).
851 			 */
852 			tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
853 
854 		}
855 
856 		/*
857 		 * If everything was dirty and we flushed it successfully,
858 		 * and the requested range is not the entire object, we
859 		 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
860 		 * return immediately.
861 		 */
862 		if (tscan >= tend && (tstart || tend < object->size)) {
863 			vm_object_clear_flag(object, OBJ_CLEANING);
864 			return;
865 		}
866 		pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
867 	}
868 
869 	/*
870 	 * Generally set CLEANCHK interlock and make the page read-only so
871 	 * we can then clear the object flags.
872 	 *
873 	 * However, if this is a nosync mmap then the object is likely to
874 	 * stay dirty so do not mess with the page and do not clear the
875 	 * object flags.
876 	 */
877 	clearobjflags = 1;
878 	TAILQ_FOREACH(p, &object->memq, listq) {
879 		p->oflags |= VPO_CLEANCHK;
880 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC))
881 			clearobjflags = 0;
882 		else {
883 			vm_page_lock(p);
884 			vm_page_lock_queues();
885 			pmap_remove_write(p);
886 			vm_page_unlock_queues();
887 			vm_page_unlock(p);
888 		}
889 	}
890 
891 	if (clearobjflags && (tstart == 0) && (tend == object->size))
892 		vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
893 
894 rescan:
895 	curgeneration = object->generation;
896 
897 	for (p = TAILQ_FIRST(&object->memq); p; p = np) {
898 		int n;
899 
900 		np = TAILQ_NEXT(p, listq);
901 
902 again:
903 		pi = p->pindex;
904 		if ((p->oflags & VPO_CLEANCHK) == 0 ||
905 			(pi < tstart) || (pi >= tend) ||
906 		    p->valid == 0) {
907 			p->oflags &= ~VPO_CLEANCHK;
908 			continue;
909 		}
910 
911 		vm_page_lock(p);
912 		vm_page_lock_queues();
913 		vm_page_test_dirty(p);
914 		if (p->dirty == 0) {
915 			vm_page_unlock_queues();
916 			vm_page_unlock(p);
917 			p->oflags &= ~VPO_CLEANCHK;
918 			continue;
919 		}
920 		vm_page_unlock_queues();
921 		vm_page_unlock(p);
922 		/*
923 		 * If we have been asked to skip nosync pages and this is a
924 		 * nosync page, skip it.  Note that the object flags were
925 		 * not cleared in this case so we do not have to set them.
926 		 */
927 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
928 			p->oflags &= ~VPO_CLEANCHK;
929 			continue;
930 		}
931 
932 		n = vm_object_page_collect_flush(object, p,
933 			curgeneration, pagerflags);
934 		if (n == 0)
935 			goto rescan;
936 
937 		if (object->generation != curgeneration)
938 			goto rescan;
939 
940 		/*
941 		 * Try to optimize the next page.  If we can't we pick up
942 		 * our (random) scan where we left off.
943 		 */
944 		if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ)
945 			if ((p = vm_page_lookup(object, pi + n)) != NULL)
946 				goto again;
947 	}
948 #if 0
949 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
950 #endif
951 
952 	vm_object_clear_flag(object, OBJ_CLEANING);
953 	return;
954 }
955 
956 static int
957 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
958 {
959 	int runlen;
960 	int maxf;
961 	int chkb;
962 	int maxb;
963 	int i;
964 	vm_pindex_t pi;
965 	vm_page_t maf[vm_pageout_page_count];
966 	vm_page_t mab[vm_pageout_page_count];
967 	vm_page_t ma[vm_pageout_page_count];
968 
969 	mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
970 	vm_page_lock_assert(p, MA_NOTOWNED);
971 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
972 	pi = p->pindex;
973 	while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
974 		if (object->generation != curgeneration) {
975 			return(0);
976 		}
977 	}
978 	maxf = 0;
979 	for(i = 1; i < vm_pageout_page_count; i++) {
980 		vm_page_t tp;
981 
982 		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
983 			if ((tp->oflags & VPO_BUSY) ||
984 				((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
985 				 (tp->oflags & VPO_CLEANCHK) == 0) ||
986 				(tp->busy != 0))
987 				break;
988 			vm_page_lock(tp);
989 			vm_page_lock_queues();
990 			vm_page_test_dirty(tp);
991 			if (tp->dirty == 0) {
992 				vm_page_unlock(tp);
993 				vm_page_unlock_queues();
994 				tp->oflags &= ~VPO_CLEANCHK;
995 				break;
996 			}
997 			vm_page_unlock(tp);
998 			vm_page_unlock_queues();
999 			maf[ i - 1 ] = tp;
1000 			maxf++;
1001 			continue;
1002 		}
1003 		break;
1004 	}
1005 
1006 	maxb = 0;
1007 	chkb = vm_pageout_page_count -  maxf;
1008 	if (chkb) {
1009 		for(i = 1; i < chkb;i++) {
1010 			vm_page_t tp;
1011 
1012 			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
1013 				if ((tp->oflags & VPO_BUSY) ||
1014 					((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1015 					 (tp->oflags & VPO_CLEANCHK) == 0) ||
1016 					(tp->busy != 0))
1017 					break;
1018 				vm_page_lock(tp);
1019 				vm_page_lock_queues();
1020 				vm_page_test_dirty(tp);
1021 				if (tp->dirty == 0) {
1022 					vm_page_unlock_queues();
1023 					vm_page_unlock(tp);
1024 					tp->oflags &= ~VPO_CLEANCHK;
1025 					break;
1026 				}
1027 				vm_page_unlock_queues();
1028 				vm_page_unlock(tp);
1029 				mab[ i - 1 ] = tp;
1030 				maxb++;
1031 				continue;
1032 			}
1033 			break;
1034 		}
1035 	}
1036 
1037 	for(i = 0; i < maxb; i++) {
1038 		int index = (maxb - i) - 1;
1039 		ma[index] = mab[i];
1040 		ma[index]->oflags &= ~VPO_CLEANCHK;
1041 	}
1042 	p->oflags &= ~VPO_CLEANCHK;
1043 	ma[maxb] = p;
1044 	for(i = 0; i < maxf; i++) {
1045 		int index = (maxb + i) + 1;
1046 		ma[index] = maf[i];
1047 		ma[index]->oflags &= ~VPO_CLEANCHK;
1048 	}
1049 	runlen = maxb + maxf + 1;
1050 
1051 	vm_pageout_flush(ma, runlen, pagerflags);
1052 	for (i = 0; i < runlen; i++) {
1053 		if (ma[i]->dirty) {
1054 			vm_page_lock(ma[i]);
1055 			vm_page_lock_queues();
1056 			pmap_remove_write(ma[i]);
1057 			vm_page_unlock_queues();
1058 			vm_page_unlock(ma[i]);
1059 			ma[i]->oflags |= VPO_CLEANCHK;
1060 
1061 			/*
1062 			 * maxf will end up being the actual number of pages
1063 			 * we wrote out contiguously, non-inclusive of the
1064 			 * first page.  We do not count look-behind pages.
1065 			 */
1066 			if (i >= maxb + 1 && (maxf > i - maxb - 1))
1067 				maxf = i - maxb - 1;
1068 		}
1069 	}
1070 	return(maxf + 1);
1071 }
1072 
1073 /*
1074  * Note that there is absolutely no sense in writing out
1075  * anonymous objects, so we track down the vnode object
1076  * to write out.
1077  * We invalidate (remove) all pages from the address space
1078  * for semantic correctness.
1079  *
1080  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1081  * may start out with a NULL object.
1082  */
1083 void
1084 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1085     boolean_t syncio, boolean_t invalidate)
1086 {
1087 	vm_object_t backing_object;
1088 	struct vnode *vp;
1089 	struct mount *mp;
1090 	int flags;
1091 
1092 	if (object == NULL)
1093 		return;
1094 	VM_OBJECT_LOCK(object);
1095 	while ((backing_object = object->backing_object) != NULL) {
1096 		VM_OBJECT_LOCK(backing_object);
1097 		offset += object->backing_object_offset;
1098 		VM_OBJECT_UNLOCK(object);
1099 		object = backing_object;
1100 		if (object->size < OFF_TO_IDX(offset + size))
1101 			size = IDX_TO_OFF(object->size) - offset;
1102 	}
1103 	/*
1104 	 * Flush pages if writing is allowed, invalidate them
1105 	 * if invalidation requested.  Pages undergoing I/O
1106 	 * will be ignored by vm_object_page_remove().
1107 	 *
1108 	 * We cannot lock the vnode and then wait for paging
1109 	 * to complete without deadlocking against vm_fault.
1110 	 * Instead we simply call vm_object_page_remove() and
1111 	 * allow it to block internally on a page-by-page
1112 	 * basis when it encounters pages undergoing async
1113 	 * I/O.
1114 	 */
1115 	if (object->type == OBJT_VNODE &&
1116 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1117 		int vfslocked;
1118 		vp = object->handle;
1119 		VM_OBJECT_UNLOCK(object);
1120 		(void) vn_start_write(vp, &mp, V_WAIT);
1121 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1122 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1123 		flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1124 		flags |= invalidate ? OBJPC_INVAL : 0;
1125 		VM_OBJECT_LOCK(object);
1126 		vm_object_page_clean(object,
1127 		    OFF_TO_IDX(offset),
1128 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1129 		    flags);
1130 		VM_OBJECT_UNLOCK(object);
1131 		VOP_UNLOCK(vp, 0);
1132 		VFS_UNLOCK_GIANT(vfslocked);
1133 		vn_finished_write(mp);
1134 		VM_OBJECT_LOCK(object);
1135 	}
1136 	if ((object->type == OBJT_VNODE ||
1137 	     object->type == OBJT_DEVICE) && invalidate) {
1138 		boolean_t purge;
1139 		purge = old_msync || (object->type == OBJT_DEVICE);
1140 		vm_object_page_remove(object,
1141 		    OFF_TO_IDX(offset),
1142 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1143 		    purge ? FALSE : TRUE);
1144 	}
1145 	VM_OBJECT_UNLOCK(object);
1146 }
1147 
1148 /*
1149  *	vm_object_madvise:
1150  *
1151  *	Implements the madvise function at the object/page level.
1152  *
1153  *	MADV_WILLNEED	(any object)
1154  *
1155  *	    Activate the specified pages if they are resident.
1156  *
1157  *	MADV_DONTNEED	(any object)
1158  *
1159  *	    Deactivate the specified pages if they are resident.
1160  *
1161  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1162  *			 OBJ_ONEMAPPING only)
1163  *
1164  *	    Deactivate and clean the specified pages if they are
1165  *	    resident.  This permits the process to reuse the pages
1166  *	    without faulting or the kernel to reclaim the pages
1167  *	    without I/O.
1168  */
1169 void
1170 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1171 {
1172 	vm_pindex_t end, tpindex;
1173 	vm_object_t backing_object, tobject;
1174 	vm_page_t m;
1175 
1176 	if (object == NULL)
1177 		return;
1178 	VM_OBJECT_LOCK(object);
1179 	end = pindex + count;
1180 	/*
1181 	 * Locate and adjust resident pages
1182 	 */
1183 	for (; pindex < end; pindex += 1) {
1184 relookup:
1185 		tobject = object;
1186 		tpindex = pindex;
1187 shadowlookup:
1188 		/*
1189 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1190 		 * and those pages must be OBJ_ONEMAPPING.
1191 		 */
1192 		if (advise == MADV_FREE) {
1193 			if ((tobject->type != OBJT_DEFAULT &&
1194 			     tobject->type != OBJT_SWAP) ||
1195 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
1196 				goto unlock_tobject;
1197 			}
1198 		} else if (tobject->type == OBJT_PHYS)
1199 			goto unlock_tobject;
1200 		m = vm_page_lookup(tobject, tpindex);
1201 		if (m == NULL && advise == MADV_WILLNEED) {
1202 			/*
1203 			 * If the page is cached, reactivate it.
1204 			 */
1205 			m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1206 			    VM_ALLOC_NOBUSY);
1207 		}
1208 		if (m == NULL) {
1209 			/*
1210 			 * There may be swap even if there is no backing page
1211 			 */
1212 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1213 				swap_pager_freespace(tobject, tpindex, 1);
1214 			/*
1215 			 * next object
1216 			 */
1217 			backing_object = tobject->backing_object;
1218 			if (backing_object == NULL)
1219 				goto unlock_tobject;
1220 			VM_OBJECT_LOCK(backing_object);
1221 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1222 			if (tobject != object)
1223 				VM_OBJECT_UNLOCK(tobject);
1224 			tobject = backing_object;
1225 			goto shadowlookup;
1226 		} else if (m->valid != VM_PAGE_BITS_ALL)
1227 			goto unlock_tobject;
1228 		/*
1229 		 * If the page is not in a normal state, skip it.
1230 		 */
1231 		vm_page_lock(m);
1232 		vm_page_lock_queues();
1233 		if (m->hold_count != 0 || m->wire_count != 0) {
1234 			vm_page_unlock_queues();
1235 			vm_page_unlock(m);
1236 			goto unlock_tobject;
1237 		}
1238 		if ((m->oflags & VPO_BUSY) || m->busy) {
1239 			if (advise == MADV_WILLNEED)
1240 				/*
1241 				 * Reference the page before unlocking and
1242 				 * sleeping so that the page daemon is less
1243 				 * likely to reclaim it.
1244 				 */
1245 				vm_page_flag_set(m, PG_REFERENCED);
1246 			vm_page_unlock_queues();
1247 			vm_page_unlock(m);
1248 			if (object != tobject)
1249 				VM_OBJECT_UNLOCK(object);
1250 			m->oflags |= VPO_WANTED;
1251 			msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1252 			    0);
1253 			VM_OBJECT_LOCK(object);
1254   			goto relookup;
1255 		}
1256 		if (advise == MADV_WILLNEED) {
1257 			vm_page_activate(m);
1258 		} else if (advise == MADV_DONTNEED) {
1259 			vm_page_dontneed(m);
1260 		} else if (advise == MADV_FREE) {
1261 			/*
1262 			 * Mark the page clean.  This will allow the page
1263 			 * to be freed up by the system.  However, such pages
1264 			 * are often reused quickly by malloc()/free()
1265 			 * so we do not do anything that would cause
1266 			 * a page fault if we can help it.
1267 			 *
1268 			 * Specifically, we do not try to actually free
1269 			 * the page now nor do we try to put it in the
1270 			 * cache (which would cause a page fault on reuse).
1271 			 *
1272 			 * But we do make the page is freeable as we
1273 			 * can without actually taking the step of unmapping
1274 			 * it.
1275 			 */
1276 			pmap_clear_modify(m);
1277 			m->dirty = 0;
1278 			m->act_count = 0;
1279 			vm_page_dontneed(m);
1280 		}
1281 		vm_page_unlock_queues();
1282 		vm_page_unlock(m);
1283 		if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1284 			swap_pager_freespace(tobject, tpindex, 1);
1285 unlock_tobject:
1286 		if (tobject != object)
1287 			VM_OBJECT_UNLOCK(tobject);
1288 	}
1289 	VM_OBJECT_UNLOCK(object);
1290 }
1291 
1292 /*
1293  *	vm_object_shadow:
1294  *
1295  *	Create a new object which is backed by the
1296  *	specified existing object range.  The source
1297  *	object reference is deallocated.
1298  *
1299  *	The new object and offset into that object
1300  *	are returned in the source parameters.
1301  */
1302 void
1303 vm_object_shadow(
1304 	vm_object_t *object,	/* IN/OUT */
1305 	vm_ooffset_t *offset,	/* IN/OUT */
1306 	vm_size_t length)
1307 {
1308 	vm_object_t source;
1309 	vm_object_t result;
1310 
1311 	source = *object;
1312 
1313 	/*
1314 	 * Don't create the new object if the old object isn't shared.
1315 	 */
1316 	if (source != NULL) {
1317 		VM_OBJECT_LOCK(source);
1318 		if (source->ref_count == 1 &&
1319 		    source->handle == NULL &&
1320 		    (source->type == OBJT_DEFAULT ||
1321 		     source->type == OBJT_SWAP)) {
1322 			VM_OBJECT_UNLOCK(source);
1323 			return;
1324 		}
1325 		VM_OBJECT_UNLOCK(source);
1326 	}
1327 
1328 	/*
1329 	 * Allocate a new object with the given length.
1330 	 */
1331 	result = vm_object_allocate(OBJT_DEFAULT, length);
1332 
1333 	/*
1334 	 * The new object shadows the source object, adding a reference to it.
1335 	 * Our caller changes his reference to point to the new object,
1336 	 * removing a reference to the source object.  Net result: no change
1337 	 * of reference count.
1338 	 *
1339 	 * Try to optimize the result object's page color when shadowing
1340 	 * in order to maintain page coloring consistency in the combined
1341 	 * shadowed object.
1342 	 */
1343 	result->backing_object = source;
1344 	/*
1345 	 * Store the offset into the source object, and fix up the offset into
1346 	 * the new object.
1347 	 */
1348 	result->backing_object_offset = *offset;
1349 	if (source != NULL) {
1350 		VM_OBJECT_LOCK(source);
1351 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1352 		source->shadow_count++;
1353 		source->generation++;
1354 #if VM_NRESERVLEVEL > 0
1355 		result->flags |= source->flags & OBJ_COLORED;
1356 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1357 		    ((1 << (VM_NFREEORDER - 1)) - 1);
1358 #endif
1359 		VM_OBJECT_UNLOCK(source);
1360 	}
1361 
1362 
1363 	/*
1364 	 * Return the new things
1365 	 */
1366 	*offset = 0;
1367 	*object = result;
1368 }
1369 
1370 /*
1371  *	vm_object_split:
1372  *
1373  * Split the pages in a map entry into a new object.  This affords
1374  * easier removal of unused pages, and keeps object inheritance from
1375  * being a negative impact on memory usage.
1376  */
1377 void
1378 vm_object_split(vm_map_entry_t entry)
1379 {
1380 	vm_page_t m, m_next;
1381 	vm_object_t orig_object, new_object, source;
1382 	vm_pindex_t idx, offidxstart;
1383 	vm_size_t size;
1384 
1385 	orig_object = entry->object.vm_object;
1386 	if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1387 		return;
1388 	if (orig_object->ref_count <= 1)
1389 		return;
1390 	VM_OBJECT_UNLOCK(orig_object);
1391 
1392 	offidxstart = OFF_TO_IDX(entry->offset);
1393 	size = atop(entry->end - entry->start);
1394 
1395 	/*
1396 	 * If swap_pager_copy() is later called, it will convert new_object
1397 	 * into a swap object.
1398 	 */
1399 	new_object = vm_object_allocate(OBJT_DEFAULT, size);
1400 
1401 	/*
1402 	 * At this point, the new object is still private, so the order in
1403 	 * which the original and new objects are locked does not matter.
1404 	 */
1405 	VM_OBJECT_LOCK(new_object);
1406 	VM_OBJECT_LOCK(orig_object);
1407 	source = orig_object->backing_object;
1408 	if (source != NULL) {
1409 		VM_OBJECT_LOCK(source);
1410 		if ((source->flags & OBJ_DEAD) != 0) {
1411 			VM_OBJECT_UNLOCK(source);
1412 			VM_OBJECT_UNLOCK(orig_object);
1413 			VM_OBJECT_UNLOCK(new_object);
1414 			vm_object_deallocate(new_object);
1415 			VM_OBJECT_LOCK(orig_object);
1416 			return;
1417 		}
1418 		LIST_INSERT_HEAD(&source->shadow_head,
1419 				  new_object, shadow_list);
1420 		source->shadow_count++;
1421 		source->generation++;
1422 		vm_object_reference_locked(source);	/* for new_object */
1423 		vm_object_clear_flag(source, OBJ_ONEMAPPING);
1424 		VM_OBJECT_UNLOCK(source);
1425 		new_object->backing_object_offset =
1426 			orig_object->backing_object_offset + entry->offset;
1427 		new_object->backing_object = source;
1428 	}
1429 	if (orig_object->uip != NULL) {
1430 		new_object->uip = orig_object->uip;
1431 		uihold(orig_object->uip);
1432 		new_object->charge = ptoa(size);
1433 		KASSERT(orig_object->charge >= ptoa(size),
1434 		    ("orig_object->charge < 0"));
1435 		orig_object->charge -= ptoa(size);
1436 	}
1437 retry:
1438 	if ((m = TAILQ_FIRST(&orig_object->memq)) != NULL) {
1439 		if (m->pindex < offidxstart) {
1440 			m = vm_page_splay(offidxstart, orig_object->root);
1441 			if ((orig_object->root = m)->pindex < offidxstart)
1442 				m = TAILQ_NEXT(m, listq);
1443 		}
1444 	}
1445 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1446 	    m = m_next) {
1447 		m_next = TAILQ_NEXT(m, listq);
1448 
1449 		/*
1450 		 * We must wait for pending I/O to complete before we can
1451 		 * rename the page.
1452 		 *
1453 		 * We do not have to VM_PROT_NONE the page as mappings should
1454 		 * not be changed by this operation.
1455 		 */
1456 		if ((m->oflags & VPO_BUSY) || m->busy) {
1457 			VM_OBJECT_UNLOCK(new_object);
1458 			m->oflags |= VPO_WANTED;
1459 			msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1460 			VM_OBJECT_LOCK(new_object);
1461 			goto retry;
1462 		}
1463 		vm_page_lock(m);
1464 		vm_page_lock_queues();
1465 		vm_page_rename(m, new_object, idx);
1466 		vm_page_unlock_queues();
1467 		vm_page_unlock(m);
1468 		/* page automatically made dirty by rename and cache handled */
1469 		vm_page_busy(m);
1470 	}
1471 	if (orig_object->type == OBJT_SWAP) {
1472 		/*
1473 		 * swap_pager_copy() can sleep, in which case the orig_object's
1474 		 * and new_object's locks are released and reacquired.
1475 		 */
1476 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1477 
1478 		/*
1479 		 * Transfer any cached pages from orig_object to new_object.
1480 		 */
1481 		if (__predict_false(orig_object->cache != NULL))
1482 			vm_page_cache_transfer(orig_object, offidxstart,
1483 			    new_object);
1484 	}
1485 	VM_OBJECT_UNLOCK(orig_object);
1486 	TAILQ_FOREACH(m, &new_object->memq, listq)
1487 		vm_page_wakeup(m);
1488 	VM_OBJECT_UNLOCK(new_object);
1489 	entry->object.vm_object = new_object;
1490 	entry->offset = 0LL;
1491 	vm_object_deallocate(orig_object);
1492 	VM_OBJECT_LOCK(new_object);
1493 }
1494 
1495 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1496 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1497 #define	OBSC_COLLAPSE_WAIT	0x0004
1498 
1499 static int
1500 vm_object_backing_scan(vm_object_t object, int op)
1501 {
1502 	int r = 1;
1503 	vm_page_t p;
1504 	vm_object_t backing_object;
1505 	vm_pindex_t backing_offset_index;
1506 
1507 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1508 	VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1509 
1510 	backing_object = object->backing_object;
1511 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1512 
1513 	/*
1514 	 * Initial conditions
1515 	 */
1516 	if (op & OBSC_TEST_ALL_SHADOWED) {
1517 		/*
1518 		 * We do not want to have to test for the existence of cache
1519 		 * or swap pages in the backing object.  XXX but with the
1520 		 * new swapper this would be pretty easy to do.
1521 		 *
1522 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1523 		 * been ZFOD faulted yet?  If we do not test for this, the
1524 		 * shadow test may succeed! XXX
1525 		 */
1526 		if (backing_object->type != OBJT_DEFAULT) {
1527 			return (0);
1528 		}
1529 	}
1530 	if (op & OBSC_COLLAPSE_WAIT) {
1531 		vm_object_set_flag(backing_object, OBJ_DEAD);
1532 	}
1533 
1534 	/*
1535 	 * Our scan
1536 	 */
1537 	p = TAILQ_FIRST(&backing_object->memq);
1538 	while (p) {
1539 		vm_page_t next = TAILQ_NEXT(p, listq);
1540 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1541 
1542 		if (op & OBSC_TEST_ALL_SHADOWED) {
1543 			vm_page_t pp;
1544 
1545 			/*
1546 			 * Ignore pages outside the parent object's range
1547 			 * and outside the parent object's mapping of the
1548 			 * backing object.
1549 			 *
1550 			 * note that we do not busy the backing object's
1551 			 * page.
1552 			 */
1553 			if (
1554 			    p->pindex < backing_offset_index ||
1555 			    new_pindex >= object->size
1556 			) {
1557 				p = next;
1558 				continue;
1559 			}
1560 
1561 			/*
1562 			 * See if the parent has the page or if the parent's
1563 			 * object pager has the page.  If the parent has the
1564 			 * page but the page is not valid, the parent's
1565 			 * object pager must have the page.
1566 			 *
1567 			 * If this fails, the parent does not completely shadow
1568 			 * the object and we might as well give up now.
1569 			 */
1570 
1571 			pp = vm_page_lookup(object, new_pindex);
1572 			if (
1573 			    (pp == NULL || pp->valid == 0) &&
1574 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1575 			) {
1576 				r = 0;
1577 				break;
1578 			}
1579 		}
1580 
1581 		/*
1582 		 * Check for busy page
1583 		 */
1584 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1585 			vm_page_t pp;
1586 
1587 			if (op & OBSC_COLLAPSE_NOWAIT) {
1588 				if ((p->oflags & VPO_BUSY) ||
1589 				    !p->valid ||
1590 				    p->busy) {
1591 					p = next;
1592 					continue;
1593 				}
1594 			} else if (op & OBSC_COLLAPSE_WAIT) {
1595 				if ((p->oflags & VPO_BUSY) || p->busy) {
1596 					VM_OBJECT_UNLOCK(object);
1597 					p->oflags |= VPO_WANTED;
1598 					msleep(p, VM_OBJECT_MTX(backing_object),
1599 					    PDROP | PVM, "vmocol", 0);
1600 					VM_OBJECT_LOCK(object);
1601 					VM_OBJECT_LOCK(backing_object);
1602 					/*
1603 					 * If we slept, anything could have
1604 					 * happened.  Since the object is
1605 					 * marked dead, the backing offset
1606 					 * should not have changed so we
1607 					 * just restart our scan.
1608 					 */
1609 					p = TAILQ_FIRST(&backing_object->memq);
1610 					continue;
1611 				}
1612 			}
1613 
1614 			KASSERT(
1615 			    p->object == backing_object,
1616 			    ("vm_object_backing_scan: object mismatch")
1617 			);
1618 
1619 			/*
1620 			 * Destroy any associated swap
1621 			 */
1622 			if (backing_object->type == OBJT_SWAP) {
1623 				swap_pager_freespace(
1624 				    backing_object,
1625 				    p->pindex,
1626 				    1
1627 				);
1628 			}
1629 
1630 			if (
1631 			    p->pindex < backing_offset_index ||
1632 			    new_pindex >= object->size
1633 			) {
1634 				/*
1635 				 * Page is out of the parent object's range, we
1636 				 * can simply destroy it.
1637 				 */
1638 				vm_page_lock(p);
1639 				vm_page_lock_queues();
1640 				KASSERT(!pmap_page_is_mapped(p),
1641 				    ("freeing mapped page %p", p));
1642 				if (p->wire_count == 0)
1643 					vm_page_free(p);
1644 				else
1645 					vm_page_remove(p);
1646 				vm_page_unlock_queues();
1647 				vm_page_unlock(p);
1648 				p = next;
1649 				continue;
1650 			}
1651 
1652 			pp = vm_page_lookup(object, new_pindex);
1653 			if (
1654 			    pp != NULL ||
1655 			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1656 			) {
1657 				/*
1658 				 * page already exists in parent OR swap exists
1659 				 * for this location in the parent.  Destroy
1660 				 * the original page from the backing object.
1661 				 *
1662 				 * Leave the parent's page alone
1663 				 */
1664 				vm_page_lock(p);
1665 				vm_page_lock_queues();
1666 				KASSERT(!pmap_page_is_mapped(p),
1667 				    ("freeing mapped page %p", p));
1668 				if (p->wire_count == 0)
1669 					vm_page_free(p);
1670 				else
1671 					vm_page_remove(p);
1672 				vm_page_unlock_queues();
1673 				vm_page_unlock(p);
1674 				p = next;
1675 				continue;
1676 			}
1677 
1678 #if VM_NRESERVLEVEL > 0
1679 			/*
1680 			 * Rename the reservation.
1681 			 */
1682 			vm_reserv_rename(p, object, backing_object,
1683 			    backing_offset_index);
1684 #endif
1685 
1686 			/*
1687 			 * Page does not exist in parent, rename the
1688 			 * page from the backing object to the main object.
1689 			 *
1690 			 * If the page was mapped to a process, it can remain
1691 			 * mapped through the rename.
1692 			 */
1693 			vm_page_lock(p);
1694 			vm_page_lock_queues();
1695 			vm_page_rename(p, object, new_pindex);
1696 			vm_page_unlock_queues();
1697 			vm_page_unlock(p);
1698 			/* page automatically made dirty by rename */
1699 		}
1700 		p = next;
1701 	}
1702 	return (r);
1703 }
1704 
1705 
1706 /*
1707  * this version of collapse allows the operation to occur earlier and
1708  * when paging_in_progress is true for an object...  This is not a complete
1709  * operation, but should plug 99.9% of the rest of the leaks.
1710  */
1711 static void
1712 vm_object_qcollapse(vm_object_t object)
1713 {
1714 	vm_object_t backing_object = object->backing_object;
1715 
1716 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1717 	VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1718 
1719 	if (backing_object->ref_count != 1)
1720 		return;
1721 
1722 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1723 }
1724 
1725 /*
1726  *	vm_object_collapse:
1727  *
1728  *	Collapse an object with the object backing it.
1729  *	Pages in the backing object are moved into the
1730  *	parent, and the backing object is deallocated.
1731  */
1732 void
1733 vm_object_collapse(vm_object_t object)
1734 {
1735 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1736 
1737 	while (TRUE) {
1738 		vm_object_t backing_object;
1739 
1740 		/*
1741 		 * Verify that the conditions are right for collapse:
1742 		 *
1743 		 * The object exists and the backing object exists.
1744 		 */
1745 		if ((backing_object = object->backing_object) == NULL)
1746 			break;
1747 
1748 		/*
1749 		 * we check the backing object first, because it is most likely
1750 		 * not collapsable.
1751 		 */
1752 		VM_OBJECT_LOCK(backing_object);
1753 		if (backing_object->handle != NULL ||
1754 		    (backing_object->type != OBJT_DEFAULT &&
1755 		     backing_object->type != OBJT_SWAP) ||
1756 		    (backing_object->flags & OBJ_DEAD) ||
1757 		    object->handle != NULL ||
1758 		    (object->type != OBJT_DEFAULT &&
1759 		     object->type != OBJT_SWAP) ||
1760 		    (object->flags & OBJ_DEAD)) {
1761 			VM_OBJECT_UNLOCK(backing_object);
1762 			break;
1763 		}
1764 
1765 		if (
1766 		    object->paging_in_progress != 0 ||
1767 		    backing_object->paging_in_progress != 0
1768 		) {
1769 			vm_object_qcollapse(object);
1770 			VM_OBJECT_UNLOCK(backing_object);
1771 			break;
1772 		}
1773 		/*
1774 		 * We know that we can either collapse the backing object (if
1775 		 * the parent is the only reference to it) or (perhaps) have
1776 		 * the parent bypass the object if the parent happens to shadow
1777 		 * all the resident pages in the entire backing object.
1778 		 *
1779 		 * This is ignoring pager-backed pages such as swap pages.
1780 		 * vm_object_backing_scan fails the shadowing test in this
1781 		 * case.
1782 		 */
1783 		if (backing_object->ref_count == 1) {
1784 			/*
1785 			 * If there is exactly one reference to the backing
1786 			 * object, we can collapse it into the parent.
1787 			 */
1788 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1789 
1790 #if VM_NRESERVLEVEL > 0
1791 			/*
1792 			 * Break any reservations from backing_object.
1793 			 */
1794 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1795 				vm_reserv_break_all(backing_object);
1796 #endif
1797 
1798 			/*
1799 			 * Move the pager from backing_object to object.
1800 			 */
1801 			if (backing_object->type == OBJT_SWAP) {
1802 				/*
1803 				 * swap_pager_copy() can sleep, in which case
1804 				 * the backing_object's and object's locks are
1805 				 * released and reacquired.
1806 				 */
1807 				swap_pager_copy(
1808 				    backing_object,
1809 				    object,
1810 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1811 
1812 				/*
1813 				 * Free any cached pages from backing_object.
1814 				 */
1815 				if (__predict_false(backing_object->cache != NULL))
1816 					vm_page_cache_free(backing_object, 0, 0);
1817 			}
1818 			/*
1819 			 * Object now shadows whatever backing_object did.
1820 			 * Note that the reference to
1821 			 * backing_object->backing_object moves from within
1822 			 * backing_object to within object.
1823 			 */
1824 			LIST_REMOVE(object, shadow_list);
1825 			backing_object->shadow_count--;
1826 			backing_object->generation++;
1827 			if (backing_object->backing_object) {
1828 				VM_OBJECT_LOCK(backing_object->backing_object);
1829 				LIST_REMOVE(backing_object, shadow_list);
1830 				LIST_INSERT_HEAD(
1831 				    &backing_object->backing_object->shadow_head,
1832 				    object, shadow_list);
1833 				/*
1834 				 * The shadow_count has not changed.
1835 				 */
1836 				backing_object->backing_object->generation++;
1837 				VM_OBJECT_UNLOCK(backing_object->backing_object);
1838 			}
1839 			object->backing_object = backing_object->backing_object;
1840 			object->backing_object_offset +=
1841 			    backing_object->backing_object_offset;
1842 
1843 			/*
1844 			 * Discard backing_object.
1845 			 *
1846 			 * Since the backing object has no pages, no pager left,
1847 			 * and no object references within it, all that is
1848 			 * necessary is to dispose of it.
1849 			 */
1850 			KASSERT(backing_object->ref_count == 1, (
1851 "backing_object %p was somehow re-referenced during collapse!",
1852 			    backing_object));
1853 			VM_OBJECT_UNLOCK(backing_object);
1854 			vm_object_destroy(backing_object);
1855 
1856 			object_collapses++;
1857 		} else {
1858 			vm_object_t new_backing_object;
1859 
1860 			/*
1861 			 * If we do not entirely shadow the backing object,
1862 			 * there is nothing we can do so we give up.
1863 			 */
1864 			if (object->resident_page_count != object->size &&
1865 			    vm_object_backing_scan(object,
1866 			    OBSC_TEST_ALL_SHADOWED) == 0) {
1867 				VM_OBJECT_UNLOCK(backing_object);
1868 				break;
1869 			}
1870 
1871 			/*
1872 			 * Make the parent shadow the next object in the
1873 			 * chain.  Deallocating backing_object will not remove
1874 			 * it, since its reference count is at least 2.
1875 			 */
1876 			LIST_REMOVE(object, shadow_list);
1877 			backing_object->shadow_count--;
1878 			backing_object->generation++;
1879 
1880 			new_backing_object = backing_object->backing_object;
1881 			if ((object->backing_object = new_backing_object) != NULL) {
1882 				VM_OBJECT_LOCK(new_backing_object);
1883 				LIST_INSERT_HEAD(
1884 				    &new_backing_object->shadow_head,
1885 				    object,
1886 				    shadow_list
1887 				);
1888 				new_backing_object->shadow_count++;
1889 				new_backing_object->generation++;
1890 				vm_object_reference_locked(new_backing_object);
1891 				VM_OBJECT_UNLOCK(new_backing_object);
1892 				object->backing_object_offset +=
1893 					backing_object->backing_object_offset;
1894 			}
1895 
1896 			/*
1897 			 * Drop the reference count on backing_object. Since
1898 			 * its ref_count was at least 2, it will not vanish.
1899 			 */
1900 			backing_object->ref_count--;
1901 			VM_OBJECT_UNLOCK(backing_object);
1902 			object_bypasses++;
1903 		}
1904 
1905 		/*
1906 		 * Try again with this object's new backing object.
1907 		 */
1908 	}
1909 }
1910 
1911 /*
1912  *	vm_object_page_remove:
1913  *
1914  *	For the given object, either frees or invalidates each of the
1915  *	specified pages.  In general, a page is freed.  However, if a
1916  *	page is wired for any reason other than the existence of a
1917  *	managed, wired mapping, then it may be invalidated but not
1918  *	removed from the object.  Pages are specified by the given
1919  *	range ["start", "end") and Boolean "clean_only".  As a
1920  *	special case, if "end" is zero, then the range extends from
1921  *	"start" to the end of the object.  If "clean_only" is TRUE,
1922  *	then only the non-dirty pages within the specified range are
1923  *	affected.
1924  *
1925  *	In general, this operation should only be performed on objects
1926  *	that contain managed pages.  There are two exceptions.  First,
1927  *	it may be performed on the kernel and kmem objects.  Second,
1928  *	it may be used by msync(..., MS_INVALIDATE) to invalidate
1929  *	device-backed pages.
1930  *
1931  *	The object must be locked.
1932  */
1933 void
1934 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1935     boolean_t clean_only)
1936 {
1937 	vm_page_t p, next;
1938 	int wirings;
1939 
1940 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1941 	if (object->resident_page_count == 0)
1942 		goto skipmemq;
1943 
1944 	/*
1945 	 * Since physically-backed objects do not use managed pages, we can't
1946 	 * remove pages from the object (we must instead remove the page
1947 	 * references, and then destroy the object).
1948 	 */
1949 	KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1950 	    object == kmem_object,
1951 	    ("attempt to remove pages from a physical object"));
1952 
1953 	vm_object_pip_add(object, 1);
1954 again:
1955 	if ((p = TAILQ_FIRST(&object->memq)) != NULL) {
1956 		if (p->pindex < start) {
1957 			p = vm_page_splay(start, object->root);
1958 			if ((object->root = p)->pindex < start)
1959 				p = TAILQ_NEXT(p, listq);
1960 		}
1961 	}
1962 
1963 	/*
1964 	 * Assert: the variable p is either (1) the page with the
1965 	 * least pindex greater than or equal to the parameter pindex
1966 	 * or (2) NULL.
1967 	 */
1968 	for (;
1969 	     p != NULL && (p->pindex < end || end == 0);
1970 	     p = next) {
1971 		next = TAILQ_NEXT(p, listq);
1972 
1973 		/*
1974 		 * If the page is wired for any reason besides the
1975 		 * existence of managed, wired mappings, then it cannot
1976 		 * be freed.  For example, fictitious pages, which
1977 		 * represent device memory, are inherently wired and
1978 		 * cannot be freed.  They can, however, be invalidated
1979 		 * if "clean_only" is FALSE.
1980 		 */
1981 		vm_page_lock(p);
1982 		vm_page_lock_queues();
1983 		if ((wirings = p->wire_count) != 0 &&
1984 		    (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1985 			/* Fictitious pages do not have managed mappings. */
1986 			if ((p->flags & PG_FICTITIOUS) == 0)
1987 				pmap_remove_all(p);
1988 			/* Account for removal of managed, wired mappings. */
1989 			p->wire_count -= wirings;
1990 			if (!clean_only) {
1991 				p->valid = 0;
1992 				vm_page_undirty(p);
1993 			}
1994 			vm_page_unlock_queues();
1995 			vm_page_unlock(p);
1996 			continue;
1997 		}
1998 		if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1999 			goto again;
2000 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
2001 		    ("vm_object_page_remove: page %p is fictitious", p));
2002 		if (clean_only && p->valid) {
2003 			pmap_remove_write(p);
2004 			if (p->dirty) {
2005 				vm_page_unlock_queues();
2006 				vm_page_unlock(p);
2007 				continue;
2008 			}
2009 		}
2010 		pmap_remove_all(p);
2011 		/* Account for removal of managed, wired mappings. */
2012 		if (wirings != 0)
2013 			p->wire_count -= wirings;
2014 		vm_page_free(p);
2015 		vm_page_unlock_queues();
2016 		vm_page_unlock(p);
2017 	}
2018 	vm_object_pip_wakeup(object);
2019 skipmemq:
2020 	if (__predict_false(object->cache != NULL))
2021 		vm_page_cache_free(object, start, end);
2022 }
2023 
2024 /*
2025  *	Populate the specified range of the object with valid pages.  Returns
2026  *	TRUE if the range is successfully populated and FALSE otherwise.
2027  *
2028  *	Note: This function should be optimized to pass a larger array of
2029  *	pages to vm_pager_get_pages() before it is applied to a non-
2030  *	OBJT_DEVICE object.
2031  *
2032  *	The object must be locked.
2033  */
2034 boolean_t
2035 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2036 {
2037 	vm_page_t m, ma[1];
2038 	vm_pindex_t pindex;
2039 	int rv;
2040 
2041 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2042 	for (pindex = start; pindex < end; pindex++) {
2043 		m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
2044 		    VM_ALLOC_RETRY);
2045 		if (m->valid != VM_PAGE_BITS_ALL) {
2046 			ma[0] = m;
2047 			rv = vm_pager_get_pages(object, ma, 1, 0);
2048 			m = vm_page_lookup(object, pindex);
2049 			if (m == NULL)
2050 				break;
2051 			if (rv != VM_PAGER_OK) {
2052 				vm_page_lock(m);
2053 				vm_page_lock_queues();
2054 				vm_page_free(m);
2055 				vm_page_unlock_queues();
2056 				vm_page_unlock(m);
2057 				break;
2058 			}
2059 		}
2060 		/*
2061 		 * Keep "m" busy because a subsequent iteration may unlock
2062 		 * the object.
2063 		 */
2064 	}
2065 	if (pindex > start) {
2066 		m = vm_page_lookup(object, start);
2067 		while (m != NULL && m->pindex < pindex) {
2068 			vm_page_wakeup(m);
2069 			m = TAILQ_NEXT(m, listq);
2070 		}
2071 	}
2072 	return (pindex == end);
2073 }
2074 
2075 /*
2076  *	Routine:	vm_object_coalesce
2077  *	Function:	Coalesces two objects backing up adjoining
2078  *			regions of memory into a single object.
2079  *
2080  *	returns TRUE if objects were combined.
2081  *
2082  *	NOTE:	Only works at the moment if the second object is NULL -
2083  *		if it's not, which object do we lock first?
2084  *
2085  *	Parameters:
2086  *		prev_object	First object to coalesce
2087  *		prev_offset	Offset into prev_object
2088  *		prev_size	Size of reference to prev_object
2089  *		next_size	Size of reference to the second object
2090  *		reserved	Indicator that extension region has
2091  *				swap accounted for
2092  *
2093  *	Conditions:
2094  *	The object must *not* be locked.
2095  */
2096 boolean_t
2097 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2098     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2099 {
2100 	vm_pindex_t next_pindex;
2101 
2102 	if (prev_object == NULL)
2103 		return (TRUE);
2104 	VM_OBJECT_LOCK(prev_object);
2105 	if (prev_object->type != OBJT_DEFAULT &&
2106 	    prev_object->type != OBJT_SWAP) {
2107 		VM_OBJECT_UNLOCK(prev_object);
2108 		return (FALSE);
2109 	}
2110 
2111 	/*
2112 	 * Try to collapse the object first
2113 	 */
2114 	vm_object_collapse(prev_object);
2115 
2116 	/*
2117 	 * Can't coalesce if: . more than one reference . paged out . shadows
2118 	 * another object . has a copy elsewhere (any of which mean that the
2119 	 * pages not mapped to prev_entry may be in use anyway)
2120 	 */
2121 	if (prev_object->backing_object != NULL) {
2122 		VM_OBJECT_UNLOCK(prev_object);
2123 		return (FALSE);
2124 	}
2125 
2126 	prev_size >>= PAGE_SHIFT;
2127 	next_size >>= PAGE_SHIFT;
2128 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2129 
2130 	if ((prev_object->ref_count > 1) &&
2131 	    (prev_object->size != next_pindex)) {
2132 		VM_OBJECT_UNLOCK(prev_object);
2133 		return (FALSE);
2134 	}
2135 
2136 	/*
2137 	 * Account for the charge.
2138 	 */
2139 	if (prev_object->uip != NULL) {
2140 
2141 		/*
2142 		 * If prev_object was charged, then this mapping,
2143 		 * althought not charged now, may become writable
2144 		 * later. Non-NULL uip in the object would prevent
2145 		 * swap reservation during enabling of the write
2146 		 * access, so reserve swap now. Failed reservation
2147 		 * cause allocation of the separate object for the map
2148 		 * entry, and swap reservation for this entry is
2149 		 * managed in appropriate time.
2150 		 */
2151 		if (!reserved && !swap_reserve_by_uid(ptoa(next_size),
2152 		    prev_object->uip)) {
2153 			return (FALSE);
2154 		}
2155 		prev_object->charge += ptoa(next_size);
2156 	}
2157 
2158 	/*
2159 	 * Remove any pages that may still be in the object from a previous
2160 	 * deallocation.
2161 	 */
2162 	if (next_pindex < prev_object->size) {
2163 		vm_object_page_remove(prev_object,
2164 				      next_pindex,
2165 				      next_pindex + next_size, FALSE);
2166 		if (prev_object->type == OBJT_SWAP)
2167 			swap_pager_freespace(prev_object,
2168 					     next_pindex, next_size);
2169 #if 0
2170 		if (prev_object->uip != NULL) {
2171 			KASSERT(prev_object->charge >=
2172 			    ptoa(prev_object->size - next_pindex),
2173 			    ("object %p overcharged 1 %jx %jx", prev_object,
2174 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2175 			prev_object->charge -= ptoa(prev_object->size -
2176 			    next_pindex);
2177 		}
2178 #endif
2179 	}
2180 
2181 	/*
2182 	 * Extend the object if necessary.
2183 	 */
2184 	if (next_pindex + next_size > prev_object->size)
2185 		prev_object->size = next_pindex + next_size;
2186 
2187 	VM_OBJECT_UNLOCK(prev_object);
2188 	return (TRUE);
2189 }
2190 
2191 void
2192 vm_object_set_writeable_dirty(vm_object_t object)
2193 {
2194 
2195 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2196 	if (object->type != OBJT_VNODE ||
2197 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0)
2198 		return;
2199 	vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2200 }
2201 
2202 #include "opt_ddb.h"
2203 #ifdef DDB
2204 #include <sys/kernel.h>
2205 
2206 #include <sys/cons.h>
2207 
2208 #include <ddb/ddb.h>
2209 
2210 static int
2211 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2212 {
2213 	vm_map_t tmpm;
2214 	vm_map_entry_t tmpe;
2215 	vm_object_t obj;
2216 	int entcount;
2217 
2218 	if (map == 0)
2219 		return 0;
2220 
2221 	if (entry == 0) {
2222 		tmpe = map->header.next;
2223 		entcount = map->nentries;
2224 		while (entcount-- && (tmpe != &map->header)) {
2225 			if (_vm_object_in_map(map, object, tmpe)) {
2226 				return 1;
2227 			}
2228 			tmpe = tmpe->next;
2229 		}
2230 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2231 		tmpm = entry->object.sub_map;
2232 		tmpe = tmpm->header.next;
2233 		entcount = tmpm->nentries;
2234 		while (entcount-- && tmpe != &tmpm->header) {
2235 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2236 				return 1;
2237 			}
2238 			tmpe = tmpe->next;
2239 		}
2240 	} else if ((obj = entry->object.vm_object) != NULL) {
2241 		for (; obj; obj = obj->backing_object)
2242 			if (obj == object) {
2243 				return 1;
2244 			}
2245 	}
2246 	return 0;
2247 }
2248 
2249 static int
2250 vm_object_in_map(vm_object_t object)
2251 {
2252 	struct proc *p;
2253 
2254 	/* sx_slock(&allproc_lock); */
2255 	FOREACH_PROC_IN_SYSTEM(p) {
2256 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2257 			continue;
2258 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2259 			/* sx_sunlock(&allproc_lock); */
2260 			return 1;
2261 		}
2262 	}
2263 	/* sx_sunlock(&allproc_lock); */
2264 	if (_vm_object_in_map(kernel_map, object, 0))
2265 		return 1;
2266 	if (_vm_object_in_map(kmem_map, object, 0))
2267 		return 1;
2268 	if (_vm_object_in_map(pager_map, object, 0))
2269 		return 1;
2270 	if (_vm_object_in_map(buffer_map, object, 0))
2271 		return 1;
2272 	return 0;
2273 }
2274 
2275 DB_SHOW_COMMAND(vmochk, vm_object_check)
2276 {
2277 	vm_object_t object;
2278 
2279 	/*
2280 	 * make sure that internal objs are in a map somewhere
2281 	 * and none have zero ref counts.
2282 	 */
2283 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2284 		if (object->handle == NULL &&
2285 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2286 			if (object->ref_count == 0) {
2287 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2288 					(long)object->size);
2289 			}
2290 			if (!vm_object_in_map(object)) {
2291 				db_printf(
2292 			"vmochk: internal obj is not in a map: "
2293 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2294 				    object->ref_count, (u_long)object->size,
2295 				    (u_long)object->size,
2296 				    (void *)object->backing_object);
2297 			}
2298 		}
2299 	}
2300 }
2301 
2302 /*
2303  *	vm_object_print:	[ debug ]
2304  */
2305 DB_SHOW_COMMAND(object, vm_object_print_static)
2306 {
2307 	/* XXX convert args. */
2308 	vm_object_t object = (vm_object_t)addr;
2309 	boolean_t full = have_addr;
2310 
2311 	vm_page_t p;
2312 
2313 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2314 #define	count	was_count
2315 
2316 	int count;
2317 
2318 	if (object == NULL)
2319 		return;
2320 
2321 	db_iprintf(
2322 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x uip %d charge %jx\n",
2323 	    object, (int)object->type, (uintmax_t)object->size,
2324 	    object->resident_page_count, object->ref_count, object->flags,
2325 	    object->uip ? object->uip->ui_uid : -1, (uintmax_t)object->charge);
2326 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2327 	    object->shadow_count,
2328 	    object->backing_object ? object->backing_object->ref_count : 0,
2329 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2330 
2331 	if (!full)
2332 		return;
2333 
2334 	db_indent += 2;
2335 	count = 0;
2336 	TAILQ_FOREACH(p, &object->memq, listq) {
2337 		if (count == 0)
2338 			db_iprintf("memory:=");
2339 		else if (count == 6) {
2340 			db_printf("\n");
2341 			db_iprintf(" ...");
2342 			count = 0;
2343 		} else
2344 			db_printf(",");
2345 		count++;
2346 
2347 		db_printf("(off=0x%jx,page=0x%jx)",
2348 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2349 	}
2350 	if (count != 0)
2351 		db_printf("\n");
2352 	db_indent -= 2;
2353 }
2354 
2355 /* XXX. */
2356 #undef count
2357 
2358 /* XXX need this non-static entry for calling from vm_map_print. */
2359 void
2360 vm_object_print(
2361         /* db_expr_t */ long addr,
2362 	boolean_t have_addr,
2363 	/* db_expr_t */ long count,
2364 	char *modif)
2365 {
2366 	vm_object_print_static(addr, have_addr, count, modif);
2367 }
2368 
2369 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2370 {
2371 	vm_object_t object;
2372 	vm_pindex_t fidx;
2373 	vm_paddr_t pa;
2374 	vm_page_t m, prev_m;
2375 	int rcount, nl, c;
2376 
2377 	nl = 0;
2378 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2379 		db_printf("new object: %p\n", (void *)object);
2380 		if (nl > 18) {
2381 			c = cngetc();
2382 			if (c != ' ')
2383 				return;
2384 			nl = 0;
2385 		}
2386 		nl++;
2387 		rcount = 0;
2388 		fidx = 0;
2389 		pa = -1;
2390 		TAILQ_FOREACH(m, &object->memq, listq) {
2391 			if (m->pindex > 128)
2392 				break;
2393 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2394 			    prev_m->pindex + 1 != m->pindex) {
2395 				if (rcount) {
2396 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2397 						(long)fidx, rcount, (long)pa);
2398 					if (nl > 18) {
2399 						c = cngetc();
2400 						if (c != ' ')
2401 							return;
2402 						nl = 0;
2403 					}
2404 					nl++;
2405 					rcount = 0;
2406 				}
2407 			}
2408 			if (rcount &&
2409 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2410 				++rcount;
2411 				continue;
2412 			}
2413 			if (rcount) {
2414 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2415 					(long)fidx, rcount, (long)pa);
2416 				if (nl > 18) {
2417 					c = cngetc();
2418 					if (c != ' ')
2419 						return;
2420 					nl = 0;
2421 				}
2422 				nl++;
2423 			}
2424 			fidx = m->pindex;
2425 			pa = VM_PAGE_TO_PHYS(m);
2426 			rcount = 1;
2427 		}
2428 		if (rcount) {
2429 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2430 				(long)fidx, rcount, (long)pa);
2431 			if (nl > 18) {
2432 				c = cngetc();
2433 				if (c != ' ')
2434 					return;
2435 				nl = 0;
2436 			}
2437 			nl++;
2438 		}
2439 	}
2440 }
2441 #endif /* DDB */
2442