xref: /freebsd/sys/vm/vm_object.c (revision a10cee30c94cf5944826d2a495e9cdf339dfbcc8)
1 /*-
2  * Copyright (c) 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * The Mach Operating System project at Carnegie-Mellon University.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 4. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	from: @(#)vm_object.c	8.5 (Berkeley) 3/22/94
33  *
34  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39  *
40  * Permission to use, copy, modify and distribute this software and
41  * its documentation is hereby granted, provided that both the copyright
42  * notice and this permission notice appear in all copies of the
43  * software, derivative works or modified versions, and any portions
44  * thereof, and that both notices appear in supporting documentation.
45  *
46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49  *
50  * Carnegie Mellon requests users of this software to return to
51  *
52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
53  *  School of Computer Science
54  *  Carnegie Mellon University
55  *  Pittsburgh PA 15213-3890
56  *
57  * any improvements or extensions that they make and grant Carnegie the
58  * rights to redistribute these changes.
59  */
60 
61 /*
62  *	Virtual memory object module.
63  */
64 
65 #include <sys/cdefs.h>
66 __FBSDID("$FreeBSD$");
67 
68 #include "opt_vm.h"
69 
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/lock.h>
73 #include <sys/mman.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/mutex.h>
78 #include <sys/proc.h>		/* for curproc, pageproc */
79 #include <sys/socket.h>
80 #include <sys/resourcevar.h>
81 #include <sys/vnode.h>
82 #include <sys/vmmeter.h>
83 #include <sys/sx.h>
84 
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_pager.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_reserv.h>
97 #include <vm/uma.h>
98 
99 static int old_msync;
100 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
101     "Use old (insecure) msync behavior");
102 
103 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
104 		    int pagerflags, int flags, boolean_t *clearobjflags,
105 		    boolean_t *eio);
106 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
107 		    boolean_t *clearobjflags);
108 static void	vm_object_qcollapse(vm_object_t object);
109 static void	vm_object_vndeallocate(vm_object_t object);
110 
111 /*
112  *	Virtual memory objects maintain the actual data
113  *	associated with allocated virtual memory.  A given
114  *	page of memory exists within exactly one object.
115  *
116  *	An object is only deallocated when all "references"
117  *	are given up.  Only one "reference" to a given
118  *	region of an object should be writeable.
119  *
120  *	Associated with each object is a list of all resident
121  *	memory pages belonging to that object; this list is
122  *	maintained by the "vm_page" module, and locked by the object's
123  *	lock.
124  *
125  *	Each object also records a "pager" routine which is
126  *	used to retrieve (and store) pages to the proper backing
127  *	storage.  In addition, objects may be backed by other
128  *	objects from which they were virtual-copied.
129  *
130  *	The only items within the object structure which are
131  *	modified after time of creation are:
132  *		reference count		locked by object's lock
133  *		pager routine		locked by object's lock
134  *
135  */
136 
137 struct object_q vm_object_list;
138 struct mtx vm_object_list_mtx;	/* lock for object list and count */
139 
140 struct vm_object kernel_object_store;
141 struct vm_object kmem_object_store;
142 
143 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
144     "VM object stats");
145 
146 static long object_collapses;
147 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
148     &object_collapses, 0, "VM object collapses");
149 
150 static long object_bypasses;
151 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
152     &object_bypasses, 0, "VM object bypasses");
153 
154 static uma_zone_t obj_zone;
155 
156 static int vm_object_zinit(void *mem, int size, int flags);
157 
158 #ifdef INVARIANTS
159 static void vm_object_zdtor(void *mem, int size, void *arg);
160 
161 static void
162 vm_object_zdtor(void *mem, int size, void *arg)
163 {
164 	vm_object_t object;
165 
166 	object = (vm_object_t)mem;
167 	KASSERT(TAILQ_EMPTY(&object->memq),
168 	    ("object %p has resident pages",
169 	    object));
170 #if VM_NRESERVLEVEL > 0
171 	KASSERT(LIST_EMPTY(&object->rvq),
172 	    ("object %p has reservations",
173 	    object));
174 #endif
175 	KASSERT(object->cache == NULL,
176 	    ("object %p has cached pages",
177 	    object));
178 	KASSERT(object->paging_in_progress == 0,
179 	    ("object %p paging_in_progress = %d",
180 	    object, object->paging_in_progress));
181 	KASSERT(object->resident_page_count == 0,
182 	    ("object %p resident_page_count = %d",
183 	    object, object->resident_page_count));
184 	KASSERT(object->shadow_count == 0,
185 	    ("object %p shadow_count = %d",
186 	    object, object->shadow_count));
187 }
188 #endif
189 
190 static int
191 vm_object_zinit(void *mem, int size, int flags)
192 {
193 	vm_object_t object;
194 
195 	object = (vm_object_t)mem;
196 	bzero(&object->mtx, sizeof(object->mtx));
197 	VM_OBJECT_LOCK_INIT(object, "standard object");
198 
199 	/* These are true for any object that has been freed */
200 	object->paging_in_progress = 0;
201 	object->resident_page_count = 0;
202 	object->shadow_count = 0;
203 	return (0);
204 }
205 
206 void
207 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
208 {
209 
210 	TAILQ_INIT(&object->memq);
211 	LIST_INIT(&object->shadow_head);
212 
213 	object->root = NULL;
214 	object->type = type;
215 	object->size = size;
216 	object->generation = 1;
217 	object->ref_count = 1;
218 	object->memattr = VM_MEMATTR_DEFAULT;
219 	object->flags = 0;
220 	object->cred = NULL;
221 	object->charge = 0;
222 	if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
223 		object->flags = OBJ_ONEMAPPING;
224 	object->pg_color = 0;
225 	object->handle = NULL;
226 	object->backing_object = NULL;
227 	object->backing_object_offset = (vm_ooffset_t) 0;
228 #if VM_NRESERVLEVEL > 0
229 	LIST_INIT(&object->rvq);
230 #endif
231 	object->cache = NULL;
232 
233 	mtx_lock(&vm_object_list_mtx);
234 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
235 	mtx_unlock(&vm_object_list_mtx);
236 }
237 
238 /*
239  *	vm_object_init:
240  *
241  *	Initialize the VM objects module.
242  */
243 void
244 vm_object_init(void)
245 {
246 	TAILQ_INIT(&vm_object_list);
247 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
248 
249 	VM_OBJECT_LOCK_INIT(kernel_object, "kernel object");
250 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
251 	    kernel_object);
252 #if VM_NRESERVLEVEL > 0
253 	kernel_object->flags |= OBJ_COLORED;
254 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
255 #endif
256 
257 	VM_OBJECT_LOCK_INIT(kmem_object, "kmem object");
258 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
259 	    kmem_object);
260 #if VM_NRESERVLEVEL > 0
261 	kmem_object->flags |= OBJ_COLORED;
262 	kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
263 #endif
264 
265 	/*
266 	 * The lock portion of struct vm_object must be type stable due
267 	 * to vm_pageout_fallback_object_lock locking a vm object
268 	 * without holding any references to it.
269 	 */
270 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
271 #ifdef INVARIANTS
272 	    vm_object_zdtor,
273 #else
274 	    NULL,
275 #endif
276 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
277 }
278 
279 void
280 vm_object_clear_flag(vm_object_t object, u_short bits)
281 {
282 
283 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
284 	object->flags &= ~bits;
285 }
286 
287 /*
288  *	Sets the default memory attribute for the specified object.  Pages
289  *	that are allocated to this object are by default assigned this memory
290  *	attribute.
291  *
292  *	Presently, this function must be called before any pages are allocated
293  *	to the object.  In the future, this requirement may be relaxed for
294  *	"default" and "swap" objects.
295  */
296 int
297 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
298 {
299 
300 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
301 	switch (object->type) {
302 	case OBJT_DEFAULT:
303 	case OBJT_DEVICE:
304 	case OBJT_PHYS:
305 	case OBJT_SG:
306 	case OBJT_SWAP:
307 	case OBJT_VNODE:
308 		if (!TAILQ_EMPTY(&object->memq))
309 			return (KERN_FAILURE);
310 		break;
311 	case OBJT_DEAD:
312 		return (KERN_INVALID_ARGUMENT);
313 	}
314 	object->memattr = memattr;
315 	return (KERN_SUCCESS);
316 }
317 
318 void
319 vm_object_pip_add(vm_object_t object, short i)
320 {
321 
322 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
323 	object->paging_in_progress += i;
324 }
325 
326 void
327 vm_object_pip_subtract(vm_object_t object, short i)
328 {
329 
330 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
331 	object->paging_in_progress -= i;
332 }
333 
334 void
335 vm_object_pip_wakeup(vm_object_t object)
336 {
337 
338 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
339 	object->paging_in_progress--;
340 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
341 		vm_object_clear_flag(object, OBJ_PIPWNT);
342 		wakeup(object);
343 	}
344 }
345 
346 void
347 vm_object_pip_wakeupn(vm_object_t object, short i)
348 {
349 
350 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
351 	if (i)
352 		object->paging_in_progress -= i;
353 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
354 		vm_object_clear_flag(object, OBJ_PIPWNT);
355 		wakeup(object);
356 	}
357 }
358 
359 void
360 vm_object_pip_wait(vm_object_t object, char *waitid)
361 {
362 
363 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
364 	while (object->paging_in_progress) {
365 		object->flags |= OBJ_PIPWNT;
366 		msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
367 	}
368 }
369 
370 /*
371  *	vm_object_allocate:
372  *
373  *	Returns a new object with the given size.
374  */
375 vm_object_t
376 vm_object_allocate(objtype_t type, vm_pindex_t size)
377 {
378 	vm_object_t object;
379 
380 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
381 	_vm_object_allocate(type, size, object);
382 	return (object);
383 }
384 
385 
386 /*
387  *	vm_object_reference:
388  *
389  *	Gets another reference to the given object.  Note: OBJ_DEAD
390  *	objects can be referenced during final cleaning.
391  */
392 void
393 vm_object_reference(vm_object_t object)
394 {
395 	if (object == NULL)
396 		return;
397 	VM_OBJECT_LOCK(object);
398 	vm_object_reference_locked(object);
399 	VM_OBJECT_UNLOCK(object);
400 }
401 
402 /*
403  *	vm_object_reference_locked:
404  *
405  *	Gets another reference to the given object.
406  *
407  *	The object must be locked.
408  */
409 void
410 vm_object_reference_locked(vm_object_t object)
411 {
412 	struct vnode *vp;
413 
414 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
415 	object->ref_count++;
416 	if (object->type == OBJT_VNODE) {
417 		vp = object->handle;
418 		vref(vp);
419 	}
420 }
421 
422 /*
423  * Handle deallocating an object of type OBJT_VNODE.
424  */
425 static void
426 vm_object_vndeallocate(vm_object_t object)
427 {
428 	struct vnode *vp = (struct vnode *) object->handle;
429 
430 	VFS_ASSERT_GIANT(vp->v_mount);
431 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
432 	KASSERT(object->type == OBJT_VNODE,
433 	    ("vm_object_vndeallocate: not a vnode object"));
434 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
435 #ifdef INVARIANTS
436 	if (object->ref_count == 0) {
437 		vprint("vm_object_vndeallocate", vp);
438 		panic("vm_object_vndeallocate: bad object reference count");
439 	}
440 #endif
441 
442 	if (object->ref_count > 1) {
443 		object->ref_count--;
444 		VM_OBJECT_UNLOCK(object);
445 		/* vrele may need the vnode lock. */
446 		vrele(vp);
447 	} else {
448 		vhold(vp);
449 		VM_OBJECT_UNLOCK(object);
450 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
451 		vdrop(vp);
452 		VM_OBJECT_LOCK(object);
453 		object->ref_count--;
454 		if (object->type == OBJT_DEAD) {
455 			VM_OBJECT_UNLOCK(object);
456 			VOP_UNLOCK(vp, 0);
457 		} else {
458 			if (object->ref_count == 0)
459 				VOP_UNSET_TEXT(vp);
460 			VM_OBJECT_UNLOCK(object);
461 			vput(vp);
462 		}
463 	}
464 }
465 
466 /*
467  *	vm_object_deallocate:
468  *
469  *	Release a reference to the specified object,
470  *	gained either through a vm_object_allocate
471  *	or a vm_object_reference call.  When all references
472  *	are gone, storage associated with this object
473  *	may be relinquished.
474  *
475  *	No object may be locked.
476  */
477 void
478 vm_object_deallocate(vm_object_t object)
479 {
480 	vm_object_t temp;
481 
482 	while (object != NULL) {
483 		int vfslocked;
484 
485 		vfslocked = 0;
486 	restart:
487 		VM_OBJECT_LOCK(object);
488 		if (object->type == OBJT_VNODE) {
489 			struct vnode *vp = (struct vnode *) object->handle;
490 
491 			/*
492 			 * Conditionally acquire Giant for a vnode-backed
493 			 * object.  We have to be careful since the type of
494 			 * a vnode object can change while the object is
495 			 * unlocked.
496 			 */
497 			if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
498 				vfslocked = 1;
499 				if (!mtx_trylock(&Giant)) {
500 					VM_OBJECT_UNLOCK(object);
501 					mtx_lock(&Giant);
502 					goto restart;
503 				}
504 			}
505 			vm_object_vndeallocate(object);
506 			VFS_UNLOCK_GIANT(vfslocked);
507 			return;
508 		} else
509 			/*
510 			 * This is to handle the case that the object
511 			 * changed type while we dropped its lock to
512 			 * obtain Giant.
513 			 */
514 			VFS_UNLOCK_GIANT(vfslocked);
515 
516 		KASSERT(object->ref_count != 0,
517 			("vm_object_deallocate: object deallocated too many times: %d", object->type));
518 
519 		/*
520 		 * If the reference count goes to 0 we start calling
521 		 * vm_object_terminate() on the object chain.
522 		 * A ref count of 1 may be a special case depending on the
523 		 * shadow count being 0 or 1.
524 		 */
525 		object->ref_count--;
526 		if (object->ref_count > 1) {
527 			VM_OBJECT_UNLOCK(object);
528 			return;
529 		} else if (object->ref_count == 1) {
530 			if (object->shadow_count == 0 &&
531 			    object->handle == NULL &&
532 			    (object->type == OBJT_DEFAULT ||
533 			     object->type == OBJT_SWAP)) {
534 				vm_object_set_flag(object, OBJ_ONEMAPPING);
535 			} else if ((object->shadow_count == 1) &&
536 			    (object->handle == NULL) &&
537 			    (object->type == OBJT_DEFAULT ||
538 			     object->type == OBJT_SWAP)) {
539 				vm_object_t robject;
540 
541 				robject = LIST_FIRST(&object->shadow_head);
542 				KASSERT(robject != NULL,
543 				    ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
544 					 object->ref_count,
545 					 object->shadow_count));
546 				if (!VM_OBJECT_TRYLOCK(robject)) {
547 					/*
548 					 * Avoid a potential deadlock.
549 					 */
550 					object->ref_count++;
551 					VM_OBJECT_UNLOCK(object);
552 					/*
553 					 * More likely than not the thread
554 					 * holding robject's lock has lower
555 					 * priority than the current thread.
556 					 * Let the lower priority thread run.
557 					 */
558 					pause("vmo_de", 1);
559 					continue;
560 				}
561 				/*
562 				 * Collapse object into its shadow unless its
563 				 * shadow is dead.  In that case, object will
564 				 * be deallocated by the thread that is
565 				 * deallocating its shadow.
566 				 */
567 				if ((robject->flags & OBJ_DEAD) == 0 &&
568 				    (robject->handle == NULL) &&
569 				    (robject->type == OBJT_DEFAULT ||
570 				     robject->type == OBJT_SWAP)) {
571 
572 					robject->ref_count++;
573 retry:
574 					if (robject->paging_in_progress) {
575 						VM_OBJECT_UNLOCK(object);
576 						vm_object_pip_wait(robject,
577 						    "objde1");
578 						temp = robject->backing_object;
579 						if (object == temp) {
580 							VM_OBJECT_LOCK(object);
581 							goto retry;
582 						}
583 					} else if (object->paging_in_progress) {
584 						VM_OBJECT_UNLOCK(robject);
585 						object->flags |= OBJ_PIPWNT;
586 						msleep(object,
587 						    VM_OBJECT_MTX(object),
588 						    PDROP | PVM, "objde2", 0);
589 						VM_OBJECT_LOCK(robject);
590 						temp = robject->backing_object;
591 						if (object == temp) {
592 							VM_OBJECT_LOCK(object);
593 							goto retry;
594 						}
595 					} else
596 						VM_OBJECT_UNLOCK(object);
597 
598 					if (robject->ref_count == 1) {
599 						robject->ref_count--;
600 						object = robject;
601 						goto doterm;
602 					}
603 					object = robject;
604 					vm_object_collapse(object);
605 					VM_OBJECT_UNLOCK(object);
606 					continue;
607 				}
608 				VM_OBJECT_UNLOCK(robject);
609 			}
610 			VM_OBJECT_UNLOCK(object);
611 			return;
612 		}
613 doterm:
614 		temp = object->backing_object;
615 		if (temp != NULL) {
616 			VM_OBJECT_LOCK(temp);
617 			LIST_REMOVE(object, shadow_list);
618 			temp->shadow_count--;
619 			VM_OBJECT_UNLOCK(temp);
620 			object->backing_object = NULL;
621 		}
622 		/*
623 		 * Don't double-terminate, we could be in a termination
624 		 * recursion due to the terminate having to sync data
625 		 * to disk.
626 		 */
627 		if ((object->flags & OBJ_DEAD) == 0)
628 			vm_object_terminate(object);
629 		else
630 			VM_OBJECT_UNLOCK(object);
631 		object = temp;
632 	}
633 }
634 
635 /*
636  *	vm_object_destroy removes the object from the global object list
637  *      and frees the space for the object.
638  */
639 void
640 vm_object_destroy(vm_object_t object)
641 {
642 
643 	/*
644 	 * Remove the object from the global object list.
645 	 */
646 	mtx_lock(&vm_object_list_mtx);
647 	TAILQ_REMOVE(&vm_object_list, object, object_list);
648 	mtx_unlock(&vm_object_list_mtx);
649 
650 	/*
651 	 * Release the allocation charge.
652 	 */
653 	if (object->cred != NULL) {
654 		KASSERT(object->type == OBJT_DEFAULT ||
655 		    object->type == OBJT_SWAP,
656 		    ("vm_object_terminate: non-swap obj %p has cred",
657 		     object));
658 		swap_release_by_cred(object->charge, object->cred);
659 		object->charge = 0;
660 		crfree(object->cred);
661 		object->cred = NULL;
662 	}
663 
664 	/*
665 	 * Free the space for the object.
666 	 */
667 	uma_zfree(obj_zone, object);
668 }
669 
670 /*
671  *	vm_object_terminate actually destroys the specified object, freeing
672  *	up all previously used resources.
673  *
674  *	The object must be locked.
675  *	This routine may block.
676  */
677 void
678 vm_object_terminate(vm_object_t object)
679 {
680 	vm_page_t p, p_next;
681 
682 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
683 
684 	/*
685 	 * Make sure no one uses us.
686 	 */
687 	vm_object_set_flag(object, OBJ_DEAD);
688 
689 	/*
690 	 * wait for the pageout daemon to be done with the object
691 	 */
692 	vm_object_pip_wait(object, "objtrm");
693 
694 	KASSERT(!object->paging_in_progress,
695 		("vm_object_terminate: pageout in progress"));
696 
697 	/*
698 	 * Clean and free the pages, as appropriate. All references to the
699 	 * object are gone, so we don't need to lock it.
700 	 */
701 	if (object->type == OBJT_VNODE) {
702 		struct vnode *vp = (struct vnode *)object->handle;
703 
704 		/*
705 		 * Clean pages and flush buffers.
706 		 */
707 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
708 		VM_OBJECT_UNLOCK(object);
709 
710 		vinvalbuf(vp, V_SAVE, 0, 0);
711 
712 		VM_OBJECT_LOCK(object);
713 	}
714 
715 	KASSERT(object->ref_count == 0,
716 		("vm_object_terminate: object with references, ref_count=%d",
717 		object->ref_count));
718 
719 	/*
720 	 * Free any remaining pageable pages.  This also removes them from the
721 	 * paging queues.  However, don't free wired pages, just remove them
722 	 * from the object.  Rather than incrementally removing each page from
723 	 * the object, the page and object are reset to any empty state.
724 	 */
725 	TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
726 		KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
727 		    ("vm_object_terminate: freeing busy page %p", p));
728 		vm_page_lock(p);
729 		/*
730 		 * Optimize the page's removal from the object by resetting
731 		 * its "object" field.  Specifically, if the page is not
732 		 * wired, then the effect of this assignment is that
733 		 * vm_page_free()'s call to vm_page_remove() will return
734 		 * immediately without modifying the page or the object.
735 		 */
736 		p->object = NULL;
737 		if (p->wire_count == 0) {
738 			vm_page_free(p);
739 			PCPU_INC(cnt.v_pfree);
740 		}
741 		vm_page_unlock(p);
742 	}
743 	/*
744 	 * If the object contained any pages, then reset it to an empty state.
745 	 * None of the object's fields, including "resident_page_count", were
746 	 * modified by the preceding loop.
747 	 */
748 	if (object->resident_page_count != 0) {
749 		object->root = NULL;
750 		TAILQ_INIT(&object->memq);
751 		object->resident_page_count = 0;
752 		if (object->type == OBJT_VNODE)
753 			vdrop(object->handle);
754 	}
755 
756 #if VM_NRESERVLEVEL > 0
757 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
758 		vm_reserv_break_all(object);
759 #endif
760 	if (__predict_false(object->cache != NULL))
761 		vm_page_cache_free(object, 0, 0);
762 
763 	/*
764 	 * Let the pager know object is dead.
765 	 */
766 	vm_pager_deallocate(object);
767 	VM_OBJECT_UNLOCK(object);
768 
769 	vm_object_destroy(object);
770 }
771 
772 /*
773  * Make the page read-only so that we can clear the object flags.  However, if
774  * this is a nosync mmap then the object is likely to stay dirty so do not
775  * mess with the page and do not clear the object flags.  Returns TRUE if the
776  * page should be flushed, and FALSE otherwise.
777  */
778 static boolean_t
779 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
780 {
781 
782 	/*
783 	 * If we have been asked to skip nosync pages and this is a
784 	 * nosync page, skip it.  Note that the object flags were not
785 	 * cleared in this case so we do not have to set them.
786 	 */
787 	if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
788 		*clearobjflags = FALSE;
789 		return (FALSE);
790 	} else {
791 		pmap_remove_write(p);
792 		return (p->dirty != 0);
793 	}
794 }
795 
796 /*
797  *	vm_object_page_clean
798  *
799  *	Clean all dirty pages in the specified range of object.  Leaves page
800  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
801  *	write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
802  *	leaving the object dirty.
803  *
804  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
805  *	synchronous clustering mode implementation.
806  *
807  *	Odd semantics: if start == end, we clean everything.
808  *
809  *	The object must be locked.
810  *
811  *	Returns FALSE if some page from the range was not written, as
812  *	reported by the pager, and TRUE otherwise.
813  */
814 boolean_t
815 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
816     int flags)
817 {
818 	vm_page_t np, p;
819 	vm_pindex_t pi, tend, tstart;
820 	int curgeneration, n, pagerflags;
821 	boolean_t clearobjflags, eio, res;
822 
823 	mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
824 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
825 	KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
826 	if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
827 	    object->resident_page_count == 0)
828 		return (TRUE);
829 
830 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
831 	    VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
832 	pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
833 
834 	tstart = OFF_TO_IDX(start);
835 	tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
836 	clearobjflags = tstart == 0 && tend >= object->size;
837 	res = TRUE;
838 
839 rescan:
840 	curgeneration = object->generation;
841 
842 	for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
843 		pi = p->pindex;
844 		if (pi >= tend)
845 			break;
846 		np = TAILQ_NEXT(p, listq);
847 		if (p->valid == 0)
848 			continue;
849 		if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
850 			if (object->generation != curgeneration) {
851 				if ((flags & OBJPC_SYNC) != 0)
852 					goto rescan;
853 				else
854 					clearobjflags = FALSE;
855 			}
856 			np = vm_page_find_least(object, pi);
857 			continue;
858 		}
859 		if (!vm_object_page_remove_write(p, flags, &clearobjflags))
860 			continue;
861 
862 		n = vm_object_page_collect_flush(object, p, pagerflags,
863 		    flags, &clearobjflags, &eio);
864 		if (eio) {
865 			res = FALSE;
866 			clearobjflags = FALSE;
867 		}
868 		if (object->generation != curgeneration) {
869 			if ((flags & OBJPC_SYNC) != 0)
870 				goto rescan;
871 			else
872 				clearobjflags = FALSE;
873 		}
874 
875 		/*
876 		 * If the VOP_PUTPAGES() did a truncated write, so
877 		 * that even the first page of the run is not fully
878 		 * written, vm_pageout_flush() returns 0 as the run
879 		 * length.  Since the condition that caused truncated
880 		 * write may be permanent, e.g. exhausted free space,
881 		 * accepting n == 0 would cause an infinite loop.
882 		 *
883 		 * Forwarding the iterator leaves the unwritten page
884 		 * behind, but there is not much we can do there if
885 		 * filesystem refuses to write it.
886 		 */
887 		if (n == 0) {
888 			n = 1;
889 			clearobjflags = FALSE;
890 		}
891 		np = vm_page_find_least(object, pi + n);
892 	}
893 #if 0
894 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
895 #endif
896 
897 	if (clearobjflags)
898 		vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
899 	return (res);
900 }
901 
902 static int
903 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
904     int flags, boolean_t *clearobjflags, boolean_t *eio)
905 {
906 	vm_page_t ma[vm_pageout_page_count], p_first, tp;
907 	int count, i, mreq, runlen;
908 
909 	mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
910 	vm_page_lock_assert(p, MA_NOTOWNED);
911 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
912 
913 	count = 1;
914 	mreq = 0;
915 
916 	for (tp = p; count < vm_pageout_page_count; count++) {
917 		tp = vm_page_next(tp);
918 		if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
919 			break;
920 		if (!vm_object_page_remove_write(tp, flags, clearobjflags))
921 			break;
922 	}
923 
924 	for (p_first = p; count < vm_pageout_page_count; count++) {
925 		tp = vm_page_prev(p_first);
926 		if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
927 			break;
928 		if (!vm_object_page_remove_write(tp, flags, clearobjflags))
929 			break;
930 		p_first = tp;
931 		mreq++;
932 	}
933 
934 	for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
935 		ma[i] = tp;
936 
937 	vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
938 	return (runlen);
939 }
940 
941 /*
942  * Note that there is absolutely no sense in writing out
943  * anonymous objects, so we track down the vnode object
944  * to write out.
945  * We invalidate (remove) all pages from the address space
946  * for semantic correctness.
947  *
948  * If the backing object is a device object with unmanaged pages, then any
949  * mappings to the specified range of pages must be removed before this
950  * function is called.
951  *
952  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
953  * may start out with a NULL object.
954  */
955 boolean_t
956 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
957     boolean_t syncio, boolean_t invalidate)
958 {
959 	vm_object_t backing_object;
960 	struct vnode *vp;
961 	struct mount *mp;
962 	int error, flags, fsync_after;
963 	boolean_t res;
964 
965 	if (object == NULL)
966 		return (TRUE);
967 	res = TRUE;
968 	error = 0;
969 	VM_OBJECT_LOCK(object);
970 	while ((backing_object = object->backing_object) != NULL) {
971 		VM_OBJECT_LOCK(backing_object);
972 		offset += object->backing_object_offset;
973 		VM_OBJECT_UNLOCK(object);
974 		object = backing_object;
975 		if (object->size < OFF_TO_IDX(offset + size))
976 			size = IDX_TO_OFF(object->size) - offset;
977 	}
978 	/*
979 	 * Flush pages if writing is allowed, invalidate them
980 	 * if invalidation requested.  Pages undergoing I/O
981 	 * will be ignored by vm_object_page_remove().
982 	 *
983 	 * We cannot lock the vnode and then wait for paging
984 	 * to complete without deadlocking against vm_fault.
985 	 * Instead we simply call vm_object_page_remove() and
986 	 * allow it to block internally on a page-by-page
987 	 * basis when it encounters pages undergoing async
988 	 * I/O.
989 	 */
990 	if (object->type == OBJT_VNODE &&
991 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
992 		int vfslocked;
993 		vp = object->handle;
994 		VM_OBJECT_UNLOCK(object);
995 		(void) vn_start_write(vp, &mp, V_WAIT);
996 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
997 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
998 		if (syncio && !invalidate && offset == 0 &&
999 		    OFF_TO_IDX(size) == object->size) {
1000 			/*
1001 			 * If syncing the whole mapping of the file,
1002 			 * it is faster to schedule all the writes in
1003 			 * async mode, also allowing the clustering,
1004 			 * and then wait for i/o to complete.
1005 			 */
1006 			flags = 0;
1007 			fsync_after = TRUE;
1008 		} else {
1009 			flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1010 			flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1011 			fsync_after = FALSE;
1012 		}
1013 		VM_OBJECT_LOCK(object);
1014 		res = vm_object_page_clean(object, offset, offset + size,
1015 		    flags);
1016 		VM_OBJECT_UNLOCK(object);
1017 		if (fsync_after)
1018 			error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1019 		VOP_UNLOCK(vp, 0);
1020 		VFS_UNLOCK_GIANT(vfslocked);
1021 		vn_finished_write(mp);
1022 		if (error != 0)
1023 			res = FALSE;
1024 		VM_OBJECT_LOCK(object);
1025 	}
1026 	if ((object->type == OBJT_VNODE ||
1027 	     object->type == OBJT_DEVICE) && invalidate) {
1028 		if (object->type == OBJT_DEVICE)
1029 			/*
1030 			 * The option OBJPR_NOTMAPPED must be passed here
1031 			 * because vm_object_page_remove() cannot remove
1032 			 * unmanaged mappings.
1033 			 */
1034 			flags = OBJPR_NOTMAPPED;
1035 		else if (old_msync)
1036 			flags = 0;
1037 		else
1038 			flags = OBJPR_CLEANONLY;
1039 		vm_object_page_remove(object, OFF_TO_IDX(offset),
1040 		    OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1041 	}
1042 	VM_OBJECT_UNLOCK(object);
1043 	return (res);
1044 }
1045 
1046 /*
1047  *	vm_object_madvise:
1048  *
1049  *	Implements the madvise function at the object/page level.
1050  *
1051  *	MADV_WILLNEED	(any object)
1052  *
1053  *	    Activate the specified pages if they are resident.
1054  *
1055  *	MADV_DONTNEED	(any object)
1056  *
1057  *	    Deactivate the specified pages if they are resident.
1058  *
1059  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1060  *			 OBJ_ONEMAPPING only)
1061  *
1062  *	    Deactivate and clean the specified pages if they are
1063  *	    resident.  This permits the process to reuse the pages
1064  *	    without faulting or the kernel to reclaim the pages
1065  *	    without I/O.
1066  */
1067 void
1068 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1069     int advise)
1070 {
1071 	vm_pindex_t tpindex;
1072 	vm_object_t backing_object, tobject;
1073 	vm_page_t m;
1074 
1075 	if (object == NULL)
1076 		return;
1077 	VM_OBJECT_LOCK(object);
1078 	/*
1079 	 * Locate and adjust resident pages
1080 	 */
1081 	for (; pindex < end; pindex += 1) {
1082 relookup:
1083 		tobject = object;
1084 		tpindex = pindex;
1085 shadowlookup:
1086 		/*
1087 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1088 		 * and those pages must be OBJ_ONEMAPPING.
1089 		 */
1090 		if (advise == MADV_FREE) {
1091 			if ((tobject->type != OBJT_DEFAULT &&
1092 			     tobject->type != OBJT_SWAP) ||
1093 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
1094 				goto unlock_tobject;
1095 			}
1096 		} else if (tobject->type == OBJT_PHYS)
1097 			goto unlock_tobject;
1098 		m = vm_page_lookup(tobject, tpindex);
1099 		if (m == NULL && advise == MADV_WILLNEED) {
1100 			/*
1101 			 * If the page is cached, reactivate it.
1102 			 */
1103 			m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1104 			    VM_ALLOC_NOBUSY);
1105 		}
1106 		if (m == NULL) {
1107 			/*
1108 			 * There may be swap even if there is no backing page
1109 			 */
1110 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1111 				swap_pager_freespace(tobject, tpindex, 1);
1112 			/*
1113 			 * next object
1114 			 */
1115 			backing_object = tobject->backing_object;
1116 			if (backing_object == NULL)
1117 				goto unlock_tobject;
1118 			VM_OBJECT_LOCK(backing_object);
1119 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1120 			if (tobject != object)
1121 				VM_OBJECT_UNLOCK(tobject);
1122 			tobject = backing_object;
1123 			goto shadowlookup;
1124 		} else if (m->valid != VM_PAGE_BITS_ALL)
1125 			goto unlock_tobject;
1126 		/*
1127 		 * If the page is not in a normal state, skip it.
1128 		 */
1129 		vm_page_lock(m);
1130 		if (m->hold_count != 0 || m->wire_count != 0) {
1131 			vm_page_unlock(m);
1132 			goto unlock_tobject;
1133 		}
1134 		KASSERT((m->flags & PG_FICTITIOUS) == 0,
1135 		    ("vm_object_madvise: page %p is fictitious", m));
1136 		KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1137 		    ("vm_object_madvise: page %p is not managed", m));
1138 		if ((m->oflags & VPO_BUSY) || m->busy) {
1139 			if (advise == MADV_WILLNEED) {
1140 				/*
1141 				 * Reference the page before unlocking and
1142 				 * sleeping so that the page daemon is less
1143 				 * likely to reclaim it.
1144 				 */
1145 				vm_page_aflag_set(m, PGA_REFERENCED);
1146 			}
1147 			vm_page_unlock(m);
1148 			if (object != tobject)
1149 				VM_OBJECT_UNLOCK(object);
1150 			m->oflags |= VPO_WANTED;
1151 			msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1152 			    0);
1153 			VM_OBJECT_LOCK(object);
1154   			goto relookup;
1155 		}
1156 		if (advise == MADV_WILLNEED) {
1157 			vm_page_activate(m);
1158 		} else if (advise == MADV_DONTNEED) {
1159 			vm_page_dontneed(m);
1160 		} else if (advise == MADV_FREE) {
1161 			/*
1162 			 * Mark the page clean.  This will allow the page
1163 			 * to be freed up by the system.  However, such pages
1164 			 * are often reused quickly by malloc()/free()
1165 			 * so we do not do anything that would cause
1166 			 * a page fault if we can help it.
1167 			 *
1168 			 * Specifically, we do not try to actually free
1169 			 * the page now nor do we try to put it in the
1170 			 * cache (which would cause a page fault on reuse).
1171 			 *
1172 			 * But we do make the page is freeable as we
1173 			 * can without actually taking the step of unmapping
1174 			 * it.
1175 			 */
1176 			pmap_clear_modify(m);
1177 			m->dirty = 0;
1178 			m->act_count = 0;
1179 			vm_page_dontneed(m);
1180 		}
1181 		vm_page_unlock(m);
1182 		if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1183 			swap_pager_freespace(tobject, tpindex, 1);
1184 unlock_tobject:
1185 		if (tobject != object)
1186 			VM_OBJECT_UNLOCK(tobject);
1187 	}
1188 	VM_OBJECT_UNLOCK(object);
1189 }
1190 
1191 /*
1192  *	vm_object_shadow:
1193  *
1194  *	Create a new object which is backed by the
1195  *	specified existing object range.  The source
1196  *	object reference is deallocated.
1197  *
1198  *	The new object and offset into that object
1199  *	are returned in the source parameters.
1200  */
1201 void
1202 vm_object_shadow(
1203 	vm_object_t *object,	/* IN/OUT */
1204 	vm_ooffset_t *offset,	/* IN/OUT */
1205 	vm_size_t length)
1206 {
1207 	vm_object_t source;
1208 	vm_object_t result;
1209 
1210 	source = *object;
1211 
1212 	/*
1213 	 * Don't create the new object if the old object isn't shared.
1214 	 */
1215 	if (source != NULL) {
1216 		VM_OBJECT_LOCK(source);
1217 		if (source->ref_count == 1 &&
1218 		    source->handle == NULL &&
1219 		    (source->type == OBJT_DEFAULT ||
1220 		     source->type == OBJT_SWAP)) {
1221 			VM_OBJECT_UNLOCK(source);
1222 			return;
1223 		}
1224 		VM_OBJECT_UNLOCK(source);
1225 	}
1226 
1227 	/*
1228 	 * Allocate a new object with the given length.
1229 	 */
1230 	result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1231 
1232 	/*
1233 	 * The new object shadows the source object, adding a reference to it.
1234 	 * Our caller changes his reference to point to the new object,
1235 	 * removing a reference to the source object.  Net result: no change
1236 	 * of reference count.
1237 	 *
1238 	 * Try to optimize the result object's page color when shadowing
1239 	 * in order to maintain page coloring consistency in the combined
1240 	 * shadowed object.
1241 	 */
1242 	result->backing_object = source;
1243 	/*
1244 	 * Store the offset into the source object, and fix up the offset into
1245 	 * the new object.
1246 	 */
1247 	result->backing_object_offset = *offset;
1248 	if (source != NULL) {
1249 		VM_OBJECT_LOCK(source);
1250 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1251 		source->shadow_count++;
1252 #if VM_NRESERVLEVEL > 0
1253 		result->flags |= source->flags & OBJ_COLORED;
1254 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1255 		    ((1 << (VM_NFREEORDER - 1)) - 1);
1256 #endif
1257 		VM_OBJECT_UNLOCK(source);
1258 	}
1259 
1260 
1261 	/*
1262 	 * Return the new things
1263 	 */
1264 	*offset = 0;
1265 	*object = result;
1266 }
1267 
1268 /*
1269  *	vm_object_split:
1270  *
1271  * Split the pages in a map entry into a new object.  This affords
1272  * easier removal of unused pages, and keeps object inheritance from
1273  * being a negative impact on memory usage.
1274  */
1275 void
1276 vm_object_split(vm_map_entry_t entry)
1277 {
1278 	vm_page_t m, m_next;
1279 	vm_object_t orig_object, new_object, source;
1280 	vm_pindex_t idx, offidxstart;
1281 	vm_size_t size;
1282 
1283 	orig_object = entry->object.vm_object;
1284 	if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1285 		return;
1286 	if (orig_object->ref_count <= 1)
1287 		return;
1288 	VM_OBJECT_UNLOCK(orig_object);
1289 
1290 	offidxstart = OFF_TO_IDX(entry->offset);
1291 	size = atop(entry->end - entry->start);
1292 
1293 	/*
1294 	 * If swap_pager_copy() is later called, it will convert new_object
1295 	 * into a swap object.
1296 	 */
1297 	new_object = vm_object_allocate(OBJT_DEFAULT, size);
1298 
1299 	/*
1300 	 * At this point, the new object is still private, so the order in
1301 	 * which the original and new objects are locked does not matter.
1302 	 */
1303 	VM_OBJECT_LOCK(new_object);
1304 	VM_OBJECT_LOCK(orig_object);
1305 	source = orig_object->backing_object;
1306 	if (source != NULL) {
1307 		VM_OBJECT_LOCK(source);
1308 		if ((source->flags & OBJ_DEAD) != 0) {
1309 			VM_OBJECT_UNLOCK(source);
1310 			VM_OBJECT_UNLOCK(orig_object);
1311 			VM_OBJECT_UNLOCK(new_object);
1312 			vm_object_deallocate(new_object);
1313 			VM_OBJECT_LOCK(orig_object);
1314 			return;
1315 		}
1316 		LIST_INSERT_HEAD(&source->shadow_head,
1317 				  new_object, shadow_list);
1318 		source->shadow_count++;
1319 		vm_object_reference_locked(source);	/* for new_object */
1320 		vm_object_clear_flag(source, OBJ_ONEMAPPING);
1321 		VM_OBJECT_UNLOCK(source);
1322 		new_object->backing_object_offset =
1323 			orig_object->backing_object_offset + entry->offset;
1324 		new_object->backing_object = source;
1325 	}
1326 	if (orig_object->cred != NULL) {
1327 		new_object->cred = orig_object->cred;
1328 		crhold(orig_object->cred);
1329 		new_object->charge = ptoa(size);
1330 		KASSERT(orig_object->charge >= ptoa(size),
1331 		    ("orig_object->charge < 0"));
1332 		orig_object->charge -= ptoa(size);
1333 	}
1334 retry:
1335 	m = vm_page_find_least(orig_object, offidxstart);
1336 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1337 	    m = m_next) {
1338 		m_next = TAILQ_NEXT(m, listq);
1339 
1340 		/*
1341 		 * We must wait for pending I/O to complete before we can
1342 		 * rename the page.
1343 		 *
1344 		 * We do not have to VM_PROT_NONE the page as mappings should
1345 		 * not be changed by this operation.
1346 		 */
1347 		if ((m->oflags & VPO_BUSY) || m->busy) {
1348 			VM_OBJECT_UNLOCK(new_object);
1349 			m->oflags |= VPO_WANTED;
1350 			msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1351 			VM_OBJECT_LOCK(new_object);
1352 			goto retry;
1353 		}
1354 #if VM_NRESERVLEVEL > 0
1355 		/*
1356 		 * If some of the reservation's allocated pages remain with
1357 		 * the original object, then transferring the reservation to
1358 		 * the new object is neither particularly beneficial nor
1359 		 * particularly harmful as compared to leaving the reservation
1360 		 * with the original object.  If, however, all of the
1361 		 * reservation's allocated pages are transferred to the new
1362 		 * object, then transferring the reservation is typically
1363 		 * beneficial.  Determining which of these two cases applies
1364 		 * would be more costly than unconditionally renaming the
1365 		 * reservation.
1366 		 */
1367 		vm_reserv_rename(m, new_object, orig_object, offidxstart);
1368 #endif
1369 		vm_page_lock(m);
1370 		vm_page_rename(m, new_object, idx);
1371 		vm_page_unlock(m);
1372 		/* page automatically made dirty by rename and cache handled */
1373 		vm_page_busy(m);
1374 	}
1375 	if (orig_object->type == OBJT_SWAP) {
1376 		/*
1377 		 * swap_pager_copy() can sleep, in which case the orig_object's
1378 		 * and new_object's locks are released and reacquired.
1379 		 */
1380 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1381 
1382 		/*
1383 		 * Transfer any cached pages from orig_object to new_object.
1384 		 * If swap_pager_copy() found swapped out pages within the
1385 		 * specified range of orig_object, then it changed
1386 		 * new_object's type to OBJT_SWAP when it transferred those
1387 		 * pages to new_object.  Otherwise, new_object's type
1388 		 * should still be OBJT_DEFAULT and orig_object should not
1389 		 * contain any cached pages within the specified range.
1390 		 */
1391 		if (__predict_false(orig_object->cache != NULL))
1392 			vm_page_cache_transfer(orig_object, offidxstart,
1393 			    new_object);
1394 	}
1395 	VM_OBJECT_UNLOCK(orig_object);
1396 	TAILQ_FOREACH(m, &new_object->memq, listq)
1397 		vm_page_wakeup(m);
1398 	VM_OBJECT_UNLOCK(new_object);
1399 	entry->object.vm_object = new_object;
1400 	entry->offset = 0LL;
1401 	vm_object_deallocate(orig_object);
1402 	VM_OBJECT_LOCK(new_object);
1403 }
1404 
1405 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1406 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1407 #define	OBSC_COLLAPSE_WAIT	0x0004
1408 
1409 static int
1410 vm_object_backing_scan(vm_object_t object, int op)
1411 {
1412 	int r = 1;
1413 	vm_page_t p;
1414 	vm_object_t backing_object;
1415 	vm_pindex_t backing_offset_index;
1416 
1417 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1418 	VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1419 
1420 	backing_object = object->backing_object;
1421 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1422 
1423 	/*
1424 	 * Initial conditions
1425 	 */
1426 	if (op & OBSC_TEST_ALL_SHADOWED) {
1427 		/*
1428 		 * We do not want to have to test for the existence of cache
1429 		 * or swap pages in the backing object.  XXX but with the
1430 		 * new swapper this would be pretty easy to do.
1431 		 *
1432 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1433 		 * been ZFOD faulted yet?  If we do not test for this, the
1434 		 * shadow test may succeed! XXX
1435 		 */
1436 		if (backing_object->type != OBJT_DEFAULT) {
1437 			return (0);
1438 		}
1439 	}
1440 	if (op & OBSC_COLLAPSE_WAIT) {
1441 		vm_object_set_flag(backing_object, OBJ_DEAD);
1442 	}
1443 
1444 	/*
1445 	 * Our scan
1446 	 */
1447 	p = TAILQ_FIRST(&backing_object->memq);
1448 	while (p) {
1449 		vm_page_t next = TAILQ_NEXT(p, listq);
1450 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1451 
1452 		if (op & OBSC_TEST_ALL_SHADOWED) {
1453 			vm_page_t pp;
1454 
1455 			/*
1456 			 * Ignore pages outside the parent object's range
1457 			 * and outside the parent object's mapping of the
1458 			 * backing object.
1459 			 *
1460 			 * note that we do not busy the backing object's
1461 			 * page.
1462 			 */
1463 			if (
1464 			    p->pindex < backing_offset_index ||
1465 			    new_pindex >= object->size
1466 			) {
1467 				p = next;
1468 				continue;
1469 			}
1470 
1471 			/*
1472 			 * See if the parent has the page or if the parent's
1473 			 * object pager has the page.  If the parent has the
1474 			 * page but the page is not valid, the parent's
1475 			 * object pager must have the page.
1476 			 *
1477 			 * If this fails, the parent does not completely shadow
1478 			 * the object and we might as well give up now.
1479 			 */
1480 
1481 			pp = vm_page_lookup(object, new_pindex);
1482 			if (
1483 			    (pp == NULL || pp->valid == 0) &&
1484 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1485 			) {
1486 				r = 0;
1487 				break;
1488 			}
1489 		}
1490 
1491 		/*
1492 		 * Check for busy page
1493 		 */
1494 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1495 			vm_page_t pp;
1496 
1497 			if (op & OBSC_COLLAPSE_NOWAIT) {
1498 				if ((p->oflags & VPO_BUSY) ||
1499 				    !p->valid ||
1500 				    p->busy) {
1501 					p = next;
1502 					continue;
1503 				}
1504 			} else if (op & OBSC_COLLAPSE_WAIT) {
1505 				if ((p->oflags & VPO_BUSY) || p->busy) {
1506 					VM_OBJECT_UNLOCK(object);
1507 					p->oflags |= VPO_WANTED;
1508 					msleep(p, VM_OBJECT_MTX(backing_object),
1509 					    PDROP | PVM, "vmocol", 0);
1510 					VM_OBJECT_LOCK(object);
1511 					VM_OBJECT_LOCK(backing_object);
1512 					/*
1513 					 * If we slept, anything could have
1514 					 * happened.  Since the object is
1515 					 * marked dead, the backing offset
1516 					 * should not have changed so we
1517 					 * just restart our scan.
1518 					 */
1519 					p = TAILQ_FIRST(&backing_object->memq);
1520 					continue;
1521 				}
1522 			}
1523 
1524 			KASSERT(
1525 			    p->object == backing_object,
1526 			    ("vm_object_backing_scan: object mismatch")
1527 			);
1528 
1529 			/*
1530 			 * Destroy any associated swap
1531 			 */
1532 			if (backing_object->type == OBJT_SWAP) {
1533 				swap_pager_freespace(
1534 				    backing_object,
1535 				    p->pindex,
1536 				    1
1537 				);
1538 			}
1539 
1540 			if (
1541 			    p->pindex < backing_offset_index ||
1542 			    new_pindex >= object->size
1543 			) {
1544 				/*
1545 				 * Page is out of the parent object's range, we
1546 				 * can simply destroy it.
1547 				 */
1548 				vm_page_lock(p);
1549 				KASSERT(!pmap_page_is_mapped(p),
1550 				    ("freeing mapped page %p", p));
1551 				if (p->wire_count == 0)
1552 					vm_page_free(p);
1553 				else
1554 					vm_page_remove(p);
1555 				vm_page_unlock(p);
1556 				p = next;
1557 				continue;
1558 			}
1559 
1560 			pp = vm_page_lookup(object, new_pindex);
1561 			if (
1562 			    (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1563 			    (pp != NULL && pp->valid == 0)
1564 			) {
1565 				/*
1566 				 * The page in the parent is not (yet) valid.
1567 				 * We don't know anything about the state of
1568 				 * the original page.  It might be mapped,
1569 				 * so we must avoid the next if here.
1570 				 *
1571 				 * This is due to a race in vm_fault() where
1572 				 * we must unbusy the original (backing_obj)
1573 				 * page before we can (re)lock the parent.
1574 				 * Hence we can get here.
1575 				 */
1576 				p = next;
1577 				continue;
1578 			}
1579 			if (
1580 			    pp != NULL ||
1581 			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1582 			) {
1583 				/*
1584 				 * page already exists in parent OR swap exists
1585 				 * for this location in the parent.  Destroy
1586 				 * the original page from the backing object.
1587 				 *
1588 				 * Leave the parent's page alone
1589 				 */
1590 				vm_page_lock(p);
1591 				KASSERT(!pmap_page_is_mapped(p),
1592 				    ("freeing mapped page %p", p));
1593 				if (p->wire_count == 0)
1594 					vm_page_free(p);
1595 				else
1596 					vm_page_remove(p);
1597 				vm_page_unlock(p);
1598 				p = next;
1599 				continue;
1600 			}
1601 
1602 #if VM_NRESERVLEVEL > 0
1603 			/*
1604 			 * Rename the reservation.
1605 			 */
1606 			vm_reserv_rename(p, object, backing_object,
1607 			    backing_offset_index);
1608 #endif
1609 
1610 			/*
1611 			 * Page does not exist in parent, rename the
1612 			 * page from the backing object to the main object.
1613 			 *
1614 			 * If the page was mapped to a process, it can remain
1615 			 * mapped through the rename.
1616 			 */
1617 			vm_page_lock(p);
1618 			vm_page_rename(p, object, new_pindex);
1619 			vm_page_unlock(p);
1620 			/* page automatically made dirty by rename */
1621 		}
1622 		p = next;
1623 	}
1624 	return (r);
1625 }
1626 
1627 
1628 /*
1629  * this version of collapse allows the operation to occur earlier and
1630  * when paging_in_progress is true for an object...  This is not a complete
1631  * operation, but should plug 99.9% of the rest of the leaks.
1632  */
1633 static void
1634 vm_object_qcollapse(vm_object_t object)
1635 {
1636 	vm_object_t backing_object = object->backing_object;
1637 
1638 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1639 	VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1640 
1641 	if (backing_object->ref_count != 1)
1642 		return;
1643 
1644 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1645 }
1646 
1647 /*
1648  *	vm_object_collapse:
1649  *
1650  *	Collapse an object with the object backing it.
1651  *	Pages in the backing object are moved into the
1652  *	parent, and the backing object is deallocated.
1653  */
1654 void
1655 vm_object_collapse(vm_object_t object)
1656 {
1657 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1658 
1659 	while (TRUE) {
1660 		vm_object_t backing_object;
1661 
1662 		/*
1663 		 * Verify that the conditions are right for collapse:
1664 		 *
1665 		 * The object exists and the backing object exists.
1666 		 */
1667 		if ((backing_object = object->backing_object) == NULL)
1668 			break;
1669 
1670 		/*
1671 		 * we check the backing object first, because it is most likely
1672 		 * not collapsable.
1673 		 */
1674 		VM_OBJECT_LOCK(backing_object);
1675 		if (backing_object->handle != NULL ||
1676 		    (backing_object->type != OBJT_DEFAULT &&
1677 		     backing_object->type != OBJT_SWAP) ||
1678 		    (backing_object->flags & OBJ_DEAD) ||
1679 		    object->handle != NULL ||
1680 		    (object->type != OBJT_DEFAULT &&
1681 		     object->type != OBJT_SWAP) ||
1682 		    (object->flags & OBJ_DEAD)) {
1683 			VM_OBJECT_UNLOCK(backing_object);
1684 			break;
1685 		}
1686 
1687 		if (
1688 		    object->paging_in_progress != 0 ||
1689 		    backing_object->paging_in_progress != 0
1690 		) {
1691 			vm_object_qcollapse(object);
1692 			VM_OBJECT_UNLOCK(backing_object);
1693 			break;
1694 		}
1695 		/*
1696 		 * We know that we can either collapse the backing object (if
1697 		 * the parent is the only reference to it) or (perhaps) have
1698 		 * the parent bypass the object if the parent happens to shadow
1699 		 * all the resident pages in the entire backing object.
1700 		 *
1701 		 * This is ignoring pager-backed pages such as swap pages.
1702 		 * vm_object_backing_scan fails the shadowing test in this
1703 		 * case.
1704 		 */
1705 		if (backing_object->ref_count == 1) {
1706 			/*
1707 			 * If there is exactly one reference to the backing
1708 			 * object, we can collapse it into the parent.
1709 			 */
1710 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1711 
1712 #if VM_NRESERVLEVEL > 0
1713 			/*
1714 			 * Break any reservations from backing_object.
1715 			 */
1716 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1717 				vm_reserv_break_all(backing_object);
1718 #endif
1719 
1720 			/*
1721 			 * Move the pager from backing_object to object.
1722 			 */
1723 			if (backing_object->type == OBJT_SWAP) {
1724 				/*
1725 				 * swap_pager_copy() can sleep, in which case
1726 				 * the backing_object's and object's locks are
1727 				 * released and reacquired.
1728 				 * Since swap_pager_copy() is being asked to
1729 				 * destroy the source, it will change the
1730 				 * backing_object's type to OBJT_DEFAULT.
1731 				 */
1732 				swap_pager_copy(
1733 				    backing_object,
1734 				    object,
1735 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1736 
1737 				/*
1738 				 * Free any cached pages from backing_object.
1739 				 */
1740 				if (__predict_false(backing_object->cache != NULL))
1741 					vm_page_cache_free(backing_object, 0, 0);
1742 			}
1743 			/*
1744 			 * Object now shadows whatever backing_object did.
1745 			 * Note that the reference to
1746 			 * backing_object->backing_object moves from within
1747 			 * backing_object to within object.
1748 			 */
1749 			LIST_REMOVE(object, shadow_list);
1750 			backing_object->shadow_count--;
1751 			if (backing_object->backing_object) {
1752 				VM_OBJECT_LOCK(backing_object->backing_object);
1753 				LIST_REMOVE(backing_object, shadow_list);
1754 				LIST_INSERT_HEAD(
1755 				    &backing_object->backing_object->shadow_head,
1756 				    object, shadow_list);
1757 				/*
1758 				 * The shadow_count has not changed.
1759 				 */
1760 				VM_OBJECT_UNLOCK(backing_object->backing_object);
1761 			}
1762 			object->backing_object = backing_object->backing_object;
1763 			object->backing_object_offset +=
1764 			    backing_object->backing_object_offset;
1765 
1766 			/*
1767 			 * Discard backing_object.
1768 			 *
1769 			 * Since the backing object has no pages, no pager left,
1770 			 * and no object references within it, all that is
1771 			 * necessary is to dispose of it.
1772 			 */
1773 			KASSERT(backing_object->ref_count == 1, (
1774 "backing_object %p was somehow re-referenced during collapse!",
1775 			    backing_object));
1776 			VM_OBJECT_UNLOCK(backing_object);
1777 			vm_object_destroy(backing_object);
1778 
1779 			object_collapses++;
1780 		} else {
1781 			vm_object_t new_backing_object;
1782 
1783 			/*
1784 			 * If we do not entirely shadow the backing object,
1785 			 * there is nothing we can do so we give up.
1786 			 */
1787 			if (object->resident_page_count != object->size &&
1788 			    vm_object_backing_scan(object,
1789 			    OBSC_TEST_ALL_SHADOWED) == 0) {
1790 				VM_OBJECT_UNLOCK(backing_object);
1791 				break;
1792 			}
1793 
1794 			/*
1795 			 * Make the parent shadow the next object in the
1796 			 * chain.  Deallocating backing_object will not remove
1797 			 * it, since its reference count is at least 2.
1798 			 */
1799 			LIST_REMOVE(object, shadow_list);
1800 			backing_object->shadow_count--;
1801 
1802 			new_backing_object = backing_object->backing_object;
1803 			if ((object->backing_object = new_backing_object) != NULL) {
1804 				VM_OBJECT_LOCK(new_backing_object);
1805 				LIST_INSERT_HEAD(
1806 				    &new_backing_object->shadow_head,
1807 				    object,
1808 				    shadow_list
1809 				);
1810 				new_backing_object->shadow_count++;
1811 				vm_object_reference_locked(new_backing_object);
1812 				VM_OBJECT_UNLOCK(new_backing_object);
1813 				object->backing_object_offset +=
1814 					backing_object->backing_object_offset;
1815 			}
1816 
1817 			/*
1818 			 * Drop the reference count on backing_object. Since
1819 			 * its ref_count was at least 2, it will not vanish.
1820 			 */
1821 			backing_object->ref_count--;
1822 			VM_OBJECT_UNLOCK(backing_object);
1823 			object_bypasses++;
1824 		}
1825 
1826 		/*
1827 		 * Try again with this object's new backing object.
1828 		 */
1829 	}
1830 }
1831 
1832 /*
1833  *	vm_object_page_remove:
1834  *
1835  *	For the given object, either frees or invalidates each of the
1836  *	specified pages.  In general, a page is freed.  However, if a page is
1837  *	wired for any reason other than the existence of a managed, wired
1838  *	mapping, then it may be invalidated but not removed from the object.
1839  *	Pages are specified by the given range ["start", "end") and the option
1840  *	OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
1841  *	extends from "start" to the end of the object.  If the option
1842  *	OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1843  *	specified range are affected.  If the option OBJPR_NOTMAPPED is
1844  *	specified, then the pages within the specified range must have no
1845  *	mappings.  Otherwise, if this option is not specified, any mappings to
1846  *	the specified pages are removed before the pages are freed or
1847  *	invalidated.
1848  *
1849  *	In general, this operation should only be performed on objects that
1850  *	contain managed pages.  There are, however, two exceptions.  First, it
1851  *	is performed on the kernel and kmem objects by vm_map_entry_delete().
1852  *	Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1853  *	backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
1854  *	not be specified and the option OBJPR_NOTMAPPED must be specified.
1855  *
1856  *	The object must be locked.
1857  */
1858 void
1859 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1860     int options)
1861 {
1862 	vm_page_t p, next;
1863 	int wirings;
1864 
1865 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1866 	KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_PHYS) ||
1867 	    (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1868 	    ("vm_object_page_remove: illegal options for object %p", object));
1869 	if (object->resident_page_count == 0)
1870 		goto skipmemq;
1871 	vm_object_pip_add(object, 1);
1872 again:
1873 	p = vm_page_find_least(object, start);
1874 
1875 	/*
1876 	 * Here, the variable "p" is either (1) the page with the least pindex
1877 	 * greater than or equal to the parameter "start" or (2) NULL.
1878 	 */
1879 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1880 		next = TAILQ_NEXT(p, listq);
1881 
1882 		/*
1883 		 * If the page is wired for any reason besides the existence
1884 		 * of managed, wired mappings, then it cannot be freed.  For
1885 		 * example, fictitious pages, which represent device memory,
1886 		 * are inherently wired and cannot be freed.  They can,
1887 		 * however, be invalidated if the option OBJPR_CLEANONLY is
1888 		 * not specified.
1889 		 */
1890 		vm_page_lock(p);
1891 		if ((wirings = p->wire_count) != 0 &&
1892 		    (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1893 			if ((options & OBJPR_NOTMAPPED) == 0) {
1894 				pmap_remove_all(p);
1895 				/* Account for removal of wired mappings. */
1896 				if (wirings != 0)
1897 					p->wire_count -= wirings;
1898 			}
1899 			if ((options & OBJPR_CLEANONLY) == 0) {
1900 				p->valid = 0;
1901 				vm_page_undirty(p);
1902 			}
1903 			vm_page_unlock(p);
1904 			continue;
1905 		}
1906 		if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1907 			goto again;
1908 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
1909 		    ("vm_object_page_remove: page %p is fictitious", p));
1910 		if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1911 			if ((options & OBJPR_NOTMAPPED) == 0)
1912 				pmap_remove_write(p);
1913 			if (p->dirty) {
1914 				vm_page_unlock(p);
1915 				continue;
1916 			}
1917 		}
1918 		if ((options & OBJPR_NOTMAPPED) == 0) {
1919 			pmap_remove_all(p);
1920 			/* Account for removal of wired mappings. */
1921 			if (wirings != 0) {
1922 				KASSERT(p->wire_count == wirings,
1923 				    ("inconsistent wire count %d %d %p",
1924 				    p->wire_count, wirings, p));
1925 				p->wire_count = 0;
1926 				atomic_subtract_int(&cnt.v_wire_count, 1);
1927 			}
1928 		}
1929 		vm_page_free(p);
1930 		vm_page_unlock(p);
1931 	}
1932 	vm_object_pip_wakeup(object);
1933 skipmemq:
1934 	if (__predict_false(object->cache != NULL))
1935 		vm_page_cache_free(object, start, end);
1936 }
1937 
1938 /*
1939  *	vm_object_page_cache:
1940  *
1941  *	For the given object, attempt to move the specified clean
1942  *	pages to the cache queue.  If a page is wired for any reason,
1943  *	then it will not be changed.  Pages are specified by the given
1944  *	range ["start", "end").  As a special case, if "end" is zero,
1945  *	then the range extends from "start" to the end of the object.
1946  *	Any mappings to the specified pages are removed before the
1947  *	pages are moved to the cache queue.
1948  *
1949  *	This operation should only be performed on objects that
1950  *	contain managed pages.
1951  *
1952  *	The object must be locked.
1953  */
1954 void
1955 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1956 {
1957 	struct mtx *mtx, *new_mtx;
1958 	vm_page_t p, next;
1959 
1960 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1961 	KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_SG &&
1962 	    object->type != OBJT_PHYS),
1963 	    ("vm_object_page_cache: illegal object %p", object));
1964 	if (object->resident_page_count == 0)
1965 		return;
1966 	p = vm_page_find_least(object, start);
1967 
1968 	/*
1969 	 * Here, the variable "p" is either (1) the page with the least pindex
1970 	 * greater than or equal to the parameter "start" or (2) NULL.
1971 	 */
1972 	mtx = NULL;
1973 	for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1974 		next = TAILQ_NEXT(p, listq);
1975 
1976 		/*
1977 		 * Avoid releasing and reacquiring the same page lock.
1978 		 */
1979 		new_mtx = vm_page_lockptr(p);
1980 		if (mtx != new_mtx) {
1981 			if (mtx != NULL)
1982 				mtx_unlock(mtx);
1983 			mtx = new_mtx;
1984 			mtx_lock(mtx);
1985 		}
1986 		vm_page_try_to_cache(p);
1987 	}
1988 	if (mtx != NULL)
1989 		mtx_unlock(mtx);
1990 }
1991 
1992 /*
1993  *	Populate the specified range of the object with valid pages.  Returns
1994  *	TRUE if the range is successfully populated and FALSE otherwise.
1995  *
1996  *	Note: This function should be optimized to pass a larger array of
1997  *	pages to vm_pager_get_pages() before it is applied to a non-
1998  *	OBJT_DEVICE object.
1999  *
2000  *	The object must be locked.
2001  */
2002 boolean_t
2003 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2004 {
2005 	vm_page_t m, ma[1];
2006 	vm_pindex_t pindex;
2007 	int rv;
2008 
2009 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2010 	for (pindex = start; pindex < end; pindex++) {
2011 		m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
2012 		    VM_ALLOC_RETRY);
2013 		if (m->valid != VM_PAGE_BITS_ALL) {
2014 			ma[0] = m;
2015 			rv = vm_pager_get_pages(object, ma, 1, 0);
2016 			m = vm_page_lookup(object, pindex);
2017 			if (m == NULL)
2018 				break;
2019 			if (rv != VM_PAGER_OK) {
2020 				vm_page_lock(m);
2021 				vm_page_free(m);
2022 				vm_page_unlock(m);
2023 				break;
2024 			}
2025 		}
2026 		/*
2027 		 * Keep "m" busy because a subsequent iteration may unlock
2028 		 * the object.
2029 		 */
2030 	}
2031 	if (pindex > start) {
2032 		m = vm_page_lookup(object, start);
2033 		while (m != NULL && m->pindex < pindex) {
2034 			vm_page_wakeup(m);
2035 			m = TAILQ_NEXT(m, listq);
2036 		}
2037 	}
2038 	return (pindex == end);
2039 }
2040 
2041 /*
2042  *	Routine:	vm_object_coalesce
2043  *	Function:	Coalesces two objects backing up adjoining
2044  *			regions of memory into a single object.
2045  *
2046  *	returns TRUE if objects were combined.
2047  *
2048  *	NOTE:	Only works at the moment if the second object is NULL -
2049  *		if it's not, which object do we lock first?
2050  *
2051  *	Parameters:
2052  *		prev_object	First object to coalesce
2053  *		prev_offset	Offset into prev_object
2054  *		prev_size	Size of reference to prev_object
2055  *		next_size	Size of reference to the second object
2056  *		reserved	Indicator that extension region has
2057  *				swap accounted for
2058  *
2059  *	Conditions:
2060  *	The object must *not* be locked.
2061  */
2062 boolean_t
2063 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2064     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2065 {
2066 	vm_pindex_t next_pindex;
2067 
2068 	if (prev_object == NULL)
2069 		return (TRUE);
2070 	VM_OBJECT_LOCK(prev_object);
2071 	if (prev_object->type != OBJT_DEFAULT &&
2072 	    prev_object->type != OBJT_SWAP) {
2073 		VM_OBJECT_UNLOCK(prev_object);
2074 		return (FALSE);
2075 	}
2076 
2077 	/*
2078 	 * Try to collapse the object first
2079 	 */
2080 	vm_object_collapse(prev_object);
2081 
2082 	/*
2083 	 * Can't coalesce if: . more than one reference . paged out . shadows
2084 	 * another object . has a copy elsewhere (any of which mean that the
2085 	 * pages not mapped to prev_entry may be in use anyway)
2086 	 */
2087 	if (prev_object->backing_object != NULL) {
2088 		VM_OBJECT_UNLOCK(prev_object);
2089 		return (FALSE);
2090 	}
2091 
2092 	prev_size >>= PAGE_SHIFT;
2093 	next_size >>= PAGE_SHIFT;
2094 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2095 
2096 	if ((prev_object->ref_count > 1) &&
2097 	    (prev_object->size != next_pindex)) {
2098 		VM_OBJECT_UNLOCK(prev_object);
2099 		return (FALSE);
2100 	}
2101 
2102 	/*
2103 	 * Account for the charge.
2104 	 */
2105 	if (prev_object->cred != NULL) {
2106 
2107 		/*
2108 		 * If prev_object was charged, then this mapping,
2109 		 * althought not charged now, may become writable
2110 		 * later. Non-NULL cred in the object would prevent
2111 		 * swap reservation during enabling of the write
2112 		 * access, so reserve swap now. Failed reservation
2113 		 * cause allocation of the separate object for the map
2114 		 * entry, and swap reservation for this entry is
2115 		 * managed in appropriate time.
2116 		 */
2117 		if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2118 		    prev_object->cred)) {
2119 			return (FALSE);
2120 		}
2121 		prev_object->charge += ptoa(next_size);
2122 	}
2123 
2124 	/*
2125 	 * Remove any pages that may still be in the object from a previous
2126 	 * deallocation.
2127 	 */
2128 	if (next_pindex < prev_object->size) {
2129 		vm_object_page_remove(prev_object, next_pindex, next_pindex +
2130 		    next_size, 0);
2131 		if (prev_object->type == OBJT_SWAP)
2132 			swap_pager_freespace(prev_object,
2133 					     next_pindex, next_size);
2134 #if 0
2135 		if (prev_object->cred != NULL) {
2136 			KASSERT(prev_object->charge >=
2137 			    ptoa(prev_object->size - next_pindex),
2138 			    ("object %p overcharged 1 %jx %jx", prev_object,
2139 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2140 			prev_object->charge -= ptoa(prev_object->size -
2141 			    next_pindex);
2142 		}
2143 #endif
2144 	}
2145 
2146 	/*
2147 	 * Extend the object if necessary.
2148 	 */
2149 	if (next_pindex + next_size > prev_object->size)
2150 		prev_object->size = next_pindex + next_size;
2151 
2152 	VM_OBJECT_UNLOCK(prev_object);
2153 	return (TRUE);
2154 }
2155 
2156 void
2157 vm_object_set_writeable_dirty(vm_object_t object)
2158 {
2159 
2160 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2161 	if (object->type != OBJT_VNODE)
2162 		return;
2163 	object->generation++;
2164 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2165 		return;
2166 	vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2167 }
2168 
2169 #include "opt_ddb.h"
2170 #ifdef DDB
2171 #include <sys/kernel.h>
2172 
2173 #include <sys/cons.h>
2174 
2175 #include <ddb/ddb.h>
2176 
2177 static int
2178 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2179 {
2180 	vm_map_t tmpm;
2181 	vm_map_entry_t tmpe;
2182 	vm_object_t obj;
2183 	int entcount;
2184 
2185 	if (map == 0)
2186 		return 0;
2187 
2188 	if (entry == 0) {
2189 		tmpe = map->header.next;
2190 		entcount = map->nentries;
2191 		while (entcount-- && (tmpe != &map->header)) {
2192 			if (_vm_object_in_map(map, object, tmpe)) {
2193 				return 1;
2194 			}
2195 			tmpe = tmpe->next;
2196 		}
2197 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2198 		tmpm = entry->object.sub_map;
2199 		tmpe = tmpm->header.next;
2200 		entcount = tmpm->nentries;
2201 		while (entcount-- && tmpe != &tmpm->header) {
2202 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2203 				return 1;
2204 			}
2205 			tmpe = tmpe->next;
2206 		}
2207 	} else if ((obj = entry->object.vm_object) != NULL) {
2208 		for (; obj; obj = obj->backing_object)
2209 			if (obj == object) {
2210 				return 1;
2211 			}
2212 	}
2213 	return 0;
2214 }
2215 
2216 static int
2217 vm_object_in_map(vm_object_t object)
2218 {
2219 	struct proc *p;
2220 
2221 	/* sx_slock(&allproc_lock); */
2222 	FOREACH_PROC_IN_SYSTEM(p) {
2223 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2224 			continue;
2225 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2226 			/* sx_sunlock(&allproc_lock); */
2227 			return 1;
2228 		}
2229 	}
2230 	/* sx_sunlock(&allproc_lock); */
2231 	if (_vm_object_in_map(kernel_map, object, 0))
2232 		return 1;
2233 	if (_vm_object_in_map(kmem_map, object, 0))
2234 		return 1;
2235 	if (_vm_object_in_map(pager_map, object, 0))
2236 		return 1;
2237 	if (_vm_object_in_map(buffer_map, object, 0))
2238 		return 1;
2239 	return 0;
2240 }
2241 
2242 DB_SHOW_COMMAND(vmochk, vm_object_check)
2243 {
2244 	vm_object_t object;
2245 
2246 	/*
2247 	 * make sure that internal objs are in a map somewhere
2248 	 * and none have zero ref counts.
2249 	 */
2250 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2251 		if (object->handle == NULL &&
2252 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2253 			if (object->ref_count == 0) {
2254 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2255 					(long)object->size);
2256 			}
2257 			if (!vm_object_in_map(object)) {
2258 				db_printf(
2259 			"vmochk: internal obj is not in a map: "
2260 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2261 				    object->ref_count, (u_long)object->size,
2262 				    (u_long)object->size,
2263 				    (void *)object->backing_object);
2264 			}
2265 		}
2266 	}
2267 }
2268 
2269 /*
2270  *	vm_object_print:	[ debug ]
2271  */
2272 DB_SHOW_COMMAND(object, vm_object_print_static)
2273 {
2274 	/* XXX convert args. */
2275 	vm_object_t object = (vm_object_t)addr;
2276 	boolean_t full = have_addr;
2277 
2278 	vm_page_t p;
2279 
2280 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2281 #define	count	was_count
2282 
2283 	int count;
2284 
2285 	if (object == NULL)
2286 		return;
2287 
2288 	db_iprintf(
2289 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2290 	    object, (int)object->type, (uintmax_t)object->size,
2291 	    object->resident_page_count, object->ref_count, object->flags,
2292 	    object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2293 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2294 	    object->shadow_count,
2295 	    object->backing_object ? object->backing_object->ref_count : 0,
2296 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2297 
2298 	if (!full)
2299 		return;
2300 
2301 	db_indent += 2;
2302 	count = 0;
2303 	TAILQ_FOREACH(p, &object->memq, listq) {
2304 		if (count == 0)
2305 			db_iprintf("memory:=");
2306 		else if (count == 6) {
2307 			db_printf("\n");
2308 			db_iprintf(" ...");
2309 			count = 0;
2310 		} else
2311 			db_printf(",");
2312 		count++;
2313 
2314 		db_printf("(off=0x%jx,page=0x%jx)",
2315 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2316 	}
2317 	if (count != 0)
2318 		db_printf("\n");
2319 	db_indent -= 2;
2320 }
2321 
2322 /* XXX. */
2323 #undef count
2324 
2325 /* XXX need this non-static entry for calling from vm_map_print. */
2326 void
2327 vm_object_print(
2328         /* db_expr_t */ long addr,
2329 	boolean_t have_addr,
2330 	/* db_expr_t */ long count,
2331 	char *modif)
2332 {
2333 	vm_object_print_static(addr, have_addr, count, modif);
2334 }
2335 
2336 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2337 {
2338 	vm_object_t object;
2339 	vm_pindex_t fidx;
2340 	vm_paddr_t pa;
2341 	vm_page_t m, prev_m;
2342 	int rcount, nl, c;
2343 
2344 	nl = 0;
2345 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2346 		db_printf("new object: %p\n", (void *)object);
2347 		if (nl > 18) {
2348 			c = cngetc();
2349 			if (c != ' ')
2350 				return;
2351 			nl = 0;
2352 		}
2353 		nl++;
2354 		rcount = 0;
2355 		fidx = 0;
2356 		pa = -1;
2357 		TAILQ_FOREACH(m, &object->memq, listq) {
2358 			if (m->pindex > 128)
2359 				break;
2360 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2361 			    prev_m->pindex + 1 != m->pindex) {
2362 				if (rcount) {
2363 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2364 						(long)fidx, rcount, (long)pa);
2365 					if (nl > 18) {
2366 						c = cngetc();
2367 						if (c != ' ')
2368 							return;
2369 						nl = 0;
2370 					}
2371 					nl++;
2372 					rcount = 0;
2373 				}
2374 			}
2375 			if (rcount &&
2376 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2377 				++rcount;
2378 				continue;
2379 			}
2380 			if (rcount) {
2381 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2382 					(long)fidx, rcount, (long)pa);
2383 				if (nl > 18) {
2384 					c = cngetc();
2385 					if (c != ' ')
2386 						return;
2387 					nl = 0;
2388 				}
2389 				nl++;
2390 			}
2391 			fidx = m->pindex;
2392 			pa = VM_PAGE_TO_PHYS(m);
2393 			rcount = 1;
2394 		}
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 		}
2406 	}
2407 }
2408 #endif /* DDB */
2409