xref: /freebsd/sys/vm/vm_object.c (revision d8b878873e7aa8df1972cc6a642804b17eb61087)
1 /*-
2  * Copyright (c) 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * The Mach Operating System project at Carnegie-Mellon University.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 4. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *	from: @(#)vm_object.c	8.5 (Berkeley) 3/22/94
33  *
34  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39  *
40  * Permission to use, copy, modify and distribute this software and
41  * its documentation is hereby granted, provided that both the copyright
42  * notice and this permission notice appear in all copies of the
43  * software, derivative works or modified versions, and any portions
44  * thereof, and that both notices appear in supporting documentation.
45  *
46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49  *
50  * Carnegie Mellon requests users of this software to return to
51  *
52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
53  *  School of Computer Science
54  *  Carnegie Mellon University
55  *  Pittsburgh PA 15213-3890
56  *
57  * any improvements or extensions that they make and grant Carnegie the
58  * rights to redistribute these changes.
59  */
60 
61 /*
62  *	Virtual memory object module.
63  */
64 
65 #include <sys/cdefs.h>
66 __FBSDID("$FreeBSD$");
67 
68 #include "opt_vm.h"
69 
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/lock.h>
73 #include <sys/mman.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/mutex.h>
78 #include <sys/proc.h>		/* for curproc, pageproc */
79 #include <sys/socket.h>
80 #include <sys/resourcevar.h>
81 #include <sys/vnode.h>
82 #include <sys/vmmeter.h>
83 #include <sys/sx.h>
84 
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_pager.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_reserv.h>
97 #include <vm/uma.h>
98 
99 #define EASY_SCAN_FACTOR       8
100 
101 #define MSYNC_FLUSH_HARDSEQ	0x01
102 #define MSYNC_FLUSH_SOFTSEQ	0x02
103 
104 /*
105  * msync / VM object flushing optimizations
106  */
107 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ;
108 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags, CTLFLAG_RW, &msync_flush_flags, 0,
109     "Enable sequential iteration optimization");
110 
111 static int old_msync;
112 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
113     "Use old (insecure) msync behavior");
114 
115 static void	vm_object_qcollapse(vm_object_t object);
116 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
117 static void	vm_object_vndeallocate(vm_object_t object);
118 
119 /*
120  *	Virtual memory objects maintain the actual data
121  *	associated with allocated virtual memory.  A given
122  *	page of memory exists within exactly one object.
123  *
124  *	An object is only deallocated when all "references"
125  *	are given up.  Only one "reference" to a given
126  *	region of an object should be writeable.
127  *
128  *	Associated with each object is a list of all resident
129  *	memory pages belonging to that object; this list is
130  *	maintained by the "vm_page" module, and locked by the object's
131  *	lock.
132  *
133  *	Each object also records a "pager" routine which is
134  *	used to retrieve (and store) pages to the proper backing
135  *	storage.  In addition, objects may be backed by other
136  *	objects from which they were virtual-copied.
137  *
138  *	The only items within the object structure which are
139  *	modified after time of creation are:
140  *		reference count		locked by object's lock
141  *		pager routine		locked by object's lock
142  *
143  */
144 
145 struct object_q vm_object_list;
146 struct mtx vm_object_list_mtx;	/* lock for object list and count */
147 
148 struct vm_object kernel_object_store;
149 struct vm_object kmem_object_store;
150 
151 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats");
152 
153 static long object_collapses;
154 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
155     &object_collapses, 0, "VM object collapses");
156 
157 static long object_bypasses;
158 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
159     &object_bypasses, 0, "VM object bypasses");
160 
161 static uma_zone_t obj_zone;
162 
163 static int vm_object_zinit(void *mem, int size, int flags);
164 
165 #ifdef INVARIANTS
166 static void vm_object_zdtor(void *mem, int size, void *arg);
167 
168 static void
169 vm_object_zdtor(void *mem, int size, void *arg)
170 {
171 	vm_object_t object;
172 
173 	object = (vm_object_t)mem;
174 	KASSERT(TAILQ_EMPTY(&object->memq),
175 	    ("object %p has resident pages",
176 	    object));
177 #if VM_NRESERVLEVEL > 0
178 	KASSERT(LIST_EMPTY(&object->rvq),
179 	    ("object %p has reservations",
180 	    object));
181 #endif
182 	KASSERT(object->cache == NULL,
183 	    ("object %p has cached pages",
184 	    object));
185 	KASSERT(object->paging_in_progress == 0,
186 	    ("object %p paging_in_progress = %d",
187 	    object, object->paging_in_progress));
188 	KASSERT(object->resident_page_count == 0,
189 	    ("object %p resident_page_count = %d",
190 	    object, object->resident_page_count));
191 	KASSERT(object->shadow_count == 0,
192 	    ("object %p shadow_count = %d",
193 	    object, object->shadow_count));
194 }
195 #endif
196 
197 static int
198 vm_object_zinit(void *mem, int size, int flags)
199 {
200 	vm_object_t object;
201 
202 	object = (vm_object_t)mem;
203 	bzero(&object->mtx, sizeof(object->mtx));
204 	VM_OBJECT_LOCK_INIT(object, "standard object");
205 
206 	/* These are true for any object that has been freed */
207 	object->paging_in_progress = 0;
208 	object->resident_page_count = 0;
209 	object->shadow_count = 0;
210 	return (0);
211 }
212 
213 void
214 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
215 {
216 
217 	TAILQ_INIT(&object->memq);
218 	LIST_INIT(&object->shadow_head);
219 
220 	object->root = NULL;
221 	object->type = type;
222 	object->size = size;
223 	object->generation = 1;
224 	object->ref_count = 1;
225 	object->memattr = VM_MEMATTR_DEFAULT;
226 	object->flags = 0;
227 	object->uip = NULL;
228 	object->charge = 0;
229 	if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
230 		object->flags = OBJ_ONEMAPPING;
231 	object->pg_color = 0;
232 	object->handle = NULL;
233 	object->backing_object = NULL;
234 	object->backing_object_offset = (vm_ooffset_t) 0;
235 #if VM_NRESERVLEVEL > 0
236 	LIST_INIT(&object->rvq);
237 #endif
238 	object->cache = NULL;
239 
240 	mtx_lock(&vm_object_list_mtx);
241 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
242 	mtx_unlock(&vm_object_list_mtx);
243 }
244 
245 /*
246  *	vm_object_init:
247  *
248  *	Initialize the VM objects module.
249  */
250 void
251 vm_object_init(void)
252 {
253 	TAILQ_INIT(&vm_object_list);
254 	mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
255 
256 	VM_OBJECT_LOCK_INIT(&kernel_object_store, "kernel object");
257 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
258 	    kernel_object);
259 #if VM_NRESERVLEVEL > 0
260 	kernel_object->flags |= OBJ_COLORED;
261 	kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
262 #endif
263 
264 	VM_OBJECT_LOCK_INIT(&kmem_object_store, "kmem object");
265 	_vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
266 	    kmem_object);
267 #if VM_NRESERVLEVEL > 0
268 	kmem_object->flags |= OBJ_COLORED;
269 	kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
270 #endif
271 
272 	/*
273 	 * The lock portion of struct vm_object must be type stable due
274 	 * to vm_pageout_fallback_object_lock locking a vm object
275 	 * without holding any references to it.
276 	 */
277 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
278 #ifdef INVARIANTS
279 	    vm_object_zdtor,
280 #else
281 	    NULL,
282 #endif
283 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
284 }
285 
286 void
287 vm_object_clear_flag(vm_object_t object, u_short bits)
288 {
289 
290 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
291 	object->flags &= ~bits;
292 }
293 
294 /*
295  *	Sets the default memory attribute for the specified object.  Pages
296  *	that are allocated to this object are by default assigned this memory
297  *	attribute.
298  *
299  *	Presently, this function must be called before any pages are allocated
300  *	to the object.  In the future, this requirement may be relaxed for
301  *	"default" and "swap" objects.
302  */
303 int
304 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
305 {
306 
307 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
308 	switch (object->type) {
309 	case OBJT_DEFAULT:
310 	case OBJT_DEVICE:
311 	case OBJT_PHYS:
312 	case OBJT_SG:
313 	case OBJT_SWAP:
314 	case OBJT_VNODE:
315 		if (!TAILQ_EMPTY(&object->memq))
316 			return (KERN_FAILURE);
317 		break;
318 	case OBJT_DEAD:
319 		return (KERN_INVALID_ARGUMENT);
320 	}
321 	object->memattr = memattr;
322 	return (KERN_SUCCESS);
323 }
324 
325 void
326 vm_object_pip_add(vm_object_t object, short i)
327 {
328 
329 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
330 	object->paging_in_progress += i;
331 }
332 
333 void
334 vm_object_pip_subtract(vm_object_t object, short i)
335 {
336 
337 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
338 	object->paging_in_progress -= i;
339 }
340 
341 void
342 vm_object_pip_wakeup(vm_object_t object)
343 {
344 
345 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
346 	object->paging_in_progress--;
347 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
348 		vm_object_clear_flag(object, OBJ_PIPWNT);
349 		wakeup(object);
350 	}
351 }
352 
353 void
354 vm_object_pip_wakeupn(vm_object_t object, short i)
355 {
356 
357 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
358 	if (i)
359 		object->paging_in_progress -= i;
360 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
361 		vm_object_clear_flag(object, OBJ_PIPWNT);
362 		wakeup(object);
363 	}
364 }
365 
366 void
367 vm_object_pip_wait(vm_object_t object, char *waitid)
368 {
369 
370 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
371 	while (object->paging_in_progress) {
372 		object->flags |= OBJ_PIPWNT;
373 		msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
374 	}
375 }
376 
377 /*
378  *	vm_object_allocate:
379  *
380  *	Returns a new object with the given size.
381  */
382 vm_object_t
383 vm_object_allocate(objtype_t type, vm_pindex_t size)
384 {
385 	vm_object_t object;
386 
387 	object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
388 	_vm_object_allocate(type, size, object);
389 	return (object);
390 }
391 
392 
393 /*
394  *	vm_object_reference:
395  *
396  *	Gets another reference to the given object.  Note: OBJ_DEAD
397  *	objects can be referenced during final cleaning.
398  */
399 void
400 vm_object_reference(vm_object_t object)
401 {
402 	if (object == NULL)
403 		return;
404 	VM_OBJECT_LOCK(object);
405 	vm_object_reference_locked(object);
406 	VM_OBJECT_UNLOCK(object);
407 }
408 
409 /*
410  *	vm_object_reference_locked:
411  *
412  *	Gets another reference to the given object.
413  *
414  *	The object must be locked.
415  */
416 void
417 vm_object_reference_locked(vm_object_t object)
418 {
419 	struct vnode *vp;
420 
421 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
422 	object->ref_count++;
423 	if (object->type == OBJT_VNODE) {
424 		vp = object->handle;
425 		vref(vp);
426 	}
427 }
428 
429 /*
430  * Handle deallocating an object of type OBJT_VNODE.
431  */
432 static void
433 vm_object_vndeallocate(vm_object_t object)
434 {
435 	struct vnode *vp = (struct vnode *) object->handle;
436 
437 	VFS_ASSERT_GIANT(vp->v_mount);
438 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
439 	KASSERT(object->type == OBJT_VNODE,
440 	    ("vm_object_vndeallocate: not a vnode object"));
441 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
442 #ifdef INVARIANTS
443 	if (object->ref_count == 0) {
444 		vprint("vm_object_vndeallocate", vp);
445 		panic("vm_object_vndeallocate: bad object reference count");
446 	}
447 #endif
448 
449 	object->ref_count--;
450 	if (object->ref_count == 0) {
451 		mp_fixme("Unlocked vflag access.");
452 		vp->v_vflag &= ~VV_TEXT;
453 	}
454 	VM_OBJECT_UNLOCK(object);
455 	/*
456 	 * vrele may need a vop lock
457 	 */
458 	vrele(vp);
459 }
460 
461 /*
462  *	vm_object_deallocate:
463  *
464  *	Release a reference to the specified object,
465  *	gained either through a vm_object_allocate
466  *	or a vm_object_reference call.  When all references
467  *	are gone, storage associated with this object
468  *	may be relinquished.
469  *
470  *	No object may be locked.
471  */
472 void
473 vm_object_deallocate(vm_object_t object)
474 {
475 	vm_object_t temp;
476 
477 	while (object != NULL) {
478 		int vfslocked;
479 
480 		vfslocked = 0;
481 	restart:
482 		VM_OBJECT_LOCK(object);
483 		if (object->type == OBJT_VNODE) {
484 			struct vnode *vp = (struct vnode *) object->handle;
485 
486 			/*
487 			 * Conditionally acquire Giant for a vnode-backed
488 			 * object.  We have to be careful since the type of
489 			 * a vnode object can change while the object is
490 			 * unlocked.
491 			 */
492 			if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
493 				vfslocked = 1;
494 				if (!mtx_trylock(&Giant)) {
495 					VM_OBJECT_UNLOCK(object);
496 					mtx_lock(&Giant);
497 					goto restart;
498 				}
499 			}
500 			vm_object_vndeallocate(object);
501 			VFS_UNLOCK_GIANT(vfslocked);
502 			return;
503 		} else
504 			/*
505 			 * This is to handle the case that the object
506 			 * changed type while we dropped its lock to
507 			 * obtain Giant.
508 			 */
509 			VFS_UNLOCK_GIANT(vfslocked);
510 
511 		KASSERT(object->ref_count != 0,
512 			("vm_object_deallocate: object deallocated too many times: %d", object->type));
513 
514 		/*
515 		 * If the reference count goes to 0 we start calling
516 		 * vm_object_terminate() on the object chain.
517 		 * A ref count of 1 may be a special case depending on the
518 		 * shadow count being 0 or 1.
519 		 */
520 		object->ref_count--;
521 		if (object->ref_count > 1) {
522 			VM_OBJECT_UNLOCK(object);
523 			return;
524 		} else if (object->ref_count == 1) {
525 			if (object->shadow_count == 0 &&
526 			    object->handle == NULL &&
527 			    (object->type == OBJT_DEFAULT ||
528 			     object->type == OBJT_SWAP)) {
529 				vm_object_set_flag(object, OBJ_ONEMAPPING);
530 			} else if ((object->shadow_count == 1) &&
531 			    (object->handle == NULL) &&
532 			    (object->type == OBJT_DEFAULT ||
533 			     object->type == OBJT_SWAP)) {
534 				vm_object_t robject;
535 
536 				robject = LIST_FIRST(&object->shadow_head);
537 				KASSERT(robject != NULL,
538 				    ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
539 					 object->ref_count,
540 					 object->shadow_count));
541 				if (!VM_OBJECT_TRYLOCK(robject)) {
542 					/*
543 					 * Avoid a potential deadlock.
544 					 */
545 					object->ref_count++;
546 					VM_OBJECT_UNLOCK(object);
547 					/*
548 					 * More likely than not the thread
549 					 * holding robject's lock has lower
550 					 * priority than the current thread.
551 					 * Let the lower priority thread run.
552 					 */
553 					pause("vmo_de", 1);
554 					continue;
555 				}
556 				/*
557 				 * Collapse object into its shadow unless its
558 				 * shadow is dead.  In that case, object will
559 				 * be deallocated by the thread that is
560 				 * deallocating its shadow.
561 				 */
562 				if ((robject->flags & OBJ_DEAD) == 0 &&
563 				    (robject->handle == NULL) &&
564 				    (robject->type == OBJT_DEFAULT ||
565 				     robject->type == OBJT_SWAP)) {
566 
567 					robject->ref_count++;
568 retry:
569 					if (robject->paging_in_progress) {
570 						VM_OBJECT_UNLOCK(object);
571 						vm_object_pip_wait(robject,
572 						    "objde1");
573 						temp = robject->backing_object;
574 						if (object == temp) {
575 							VM_OBJECT_LOCK(object);
576 							goto retry;
577 						}
578 					} else if (object->paging_in_progress) {
579 						VM_OBJECT_UNLOCK(robject);
580 						object->flags |= OBJ_PIPWNT;
581 						msleep(object,
582 						    VM_OBJECT_MTX(object),
583 						    PDROP | PVM, "objde2", 0);
584 						VM_OBJECT_LOCK(robject);
585 						temp = robject->backing_object;
586 						if (object == temp) {
587 							VM_OBJECT_LOCK(object);
588 							goto retry;
589 						}
590 					} else
591 						VM_OBJECT_UNLOCK(object);
592 
593 					if (robject->ref_count == 1) {
594 						robject->ref_count--;
595 						object = robject;
596 						goto doterm;
597 					}
598 					object = robject;
599 					vm_object_collapse(object);
600 					VM_OBJECT_UNLOCK(object);
601 					continue;
602 				}
603 				VM_OBJECT_UNLOCK(robject);
604 			}
605 			VM_OBJECT_UNLOCK(object);
606 			return;
607 		}
608 doterm:
609 		temp = object->backing_object;
610 		if (temp != NULL) {
611 			VM_OBJECT_LOCK(temp);
612 			LIST_REMOVE(object, shadow_list);
613 			temp->shadow_count--;
614 			temp->generation++;
615 			VM_OBJECT_UNLOCK(temp);
616 			object->backing_object = NULL;
617 		}
618 		/*
619 		 * Don't double-terminate, we could be in a termination
620 		 * recursion due to the terminate having to sync data
621 		 * to disk.
622 		 */
623 		if ((object->flags & OBJ_DEAD) == 0)
624 			vm_object_terminate(object);
625 		else
626 			VM_OBJECT_UNLOCK(object);
627 		object = temp;
628 	}
629 }
630 
631 /*
632  *	vm_object_destroy removes the object from the global object list
633  *      and frees the space for the object.
634  */
635 void
636 vm_object_destroy(vm_object_t object)
637 {
638 
639 	/*
640 	 * Remove the object from the global object list.
641 	 */
642 	mtx_lock(&vm_object_list_mtx);
643 	TAILQ_REMOVE(&vm_object_list, object, object_list);
644 	mtx_unlock(&vm_object_list_mtx);
645 
646 	/*
647 	 * Release the allocation charge.
648 	 */
649 	if (object->uip != NULL) {
650 		KASSERT(object->type == OBJT_DEFAULT ||
651 		    object->type == OBJT_SWAP,
652 		    ("vm_object_terminate: non-swap obj %p has uip",
653 		     object));
654 		swap_release_by_uid(object->charge, object->uip);
655 		object->charge = 0;
656 		uifree(object->uip);
657 		object->uip = NULL;
658 	}
659 
660 	/*
661 	 * Free the space for the object.
662 	 */
663 	uma_zfree(obj_zone, object);
664 }
665 
666 /*
667  *	vm_object_terminate actually destroys the specified object, freeing
668  *	up all previously used resources.
669  *
670  *	The object must be locked.
671  *	This routine may block.
672  */
673 void
674 vm_object_terminate(vm_object_t object)
675 {
676 	vm_page_t p;
677 
678 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
679 
680 	/*
681 	 * Make sure no one uses us.
682 	 */
683 	vm_object_set_flag(object, OBJ_DEAD);
684 
685 	/*
686 	 * wait for the pageout daemon to be done with the object
687 	 */
688 	vm_object_pip_wait(object, "objtrm");
689 
690 	KASSERT(!object->paging_in_progress,
691 		("vm_object_terminate: pageout in progress"));
692 
693 	/*
694 	 * Clean and free the pages, as appropriate. All references to the
695 	 * object are gone, so we don't need to lock it.
696 	 */
697 	if (object->type == OBJT_VNODE) {
698 		struct vnode *vp = (struct vnode *)object->handle;
699 
700 		/*
701 		 * Clean pages and flush buffers.
702 		 */
703 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
704 		VM_OBJECT_UNLOCK(object);
705 
706 		vinvalbuf(vp, V_SAVE, 0, 0);
707 
708 		VM_OBJECT_LOCK(object);
709 	}
710 
711 	KASSERT(object->ref_count == 0,
712 		("vm_object_terminate: object with references, ref_count=%d",
713 		object->ref_count));
714 
715 	/*
716 	 * Now free any remaining pages. For internal objects, this also
717 	 * removes them from paging queues. Don't free wired pages, just
718 	 * remove them from the object.
719 	 */
720 	vm_page_lock_queues();
721 	while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
722 		KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
723 			("vm_object_terminate: freeing busy page %p "
724 			"p->busy = %d, p->oflags %x\n", p, p->busy, p->oflags));
725 		if (p->wire_count == 0) {
726 			vm_page_free(p);
727 			cnt.v_pfree++;
728 		} else {
729 			vm_page_remove(p);
730 		}
731 	}
732 	vm_page_unlock_queues();
733 
734 #if VM_NRESERVLEVEL > 0
735 	if (__predict_false(!LIST_EMPTY(&object->rvq)))
736 		vm_reserv_break_all(object);
737 #endif
738 	if (__predict_false(object->cache != NULL))
739 		vm_page_cache_free(object, 0, 0);
740 
741 	/*
742 	 * Let the pager know object is dead.
743 	 */
744 	vm_pager_deallocate(object);
745 	VM_OBJECT_UNLOCK(object);
746 
747 	vm_object_destroy(object);
748 }
749 
750 /*
751  *	vm_object_page_clean
752  *
753  *	Clean all dirty pages in the specified range of object.  Leaves page
754  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
755  *	write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
756  *	leaving the object dirty.
757  *
758  *	When stuffing pages asynchronously, allow clustering.  XXX we need a
759  *	synchronous clustering mode implementation.
760  *
761  *	Odd semantics: if start == end, we clean everything.
762  *
763  *	The object must be locked.
764  */
765 void
766 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags)
767 {
768 	vm_page_t p, np;
769 	vm_pindex_t tstart, tend;
770 	vm_pindex_t pi;
771 	int clearobjflags;
772 	int pagerflags;
773 	int curgeneration;
774 
775 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
776 	if ((object->flags & OBJ_MIGHTBEDIRTY) == 0)
777 		return;
778 	KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
779 
780 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
781 	pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
782 
783 	vm_object_set_flag(object, OBJ_CLEANING);
784 
785 	tstart = start;
786 	if (end == 0) {
787 		tend = object->size;
788 	} else {
789 		tend = end;
790 	}
791 
792 	vm_page_lock_queues();
793 	/*
794 	 * If the caller is smart and only msync()s a range he knows is
795 	 * dirty, we may be able to avoid an object scan.  This results in
796 	 * a phenominal improvement in performance.  We cannot do this
797 	 * as a matter of course because the object may be huge - e.g.
798 	 * the size might be in the gigabytes or terrabytes.
799 	 */
800 	if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
801 		vm_pindex_t tscan;
802 		int scanlimit;
803 		int scanreset;
804 
805 		scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
806 		if (scanreset < 16)
807 			scanreset = 16;
808 		pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
809 
810 		scanlimit = scanreset;
811 		tscan = tstart;
812 		while (tscan < tend) {
813 			curgeneration = object->generation;
814 			p = vm_page_lookup(object, tscan);
815 			if (p == NULL || p->valid == 0) {
816 				if (--scanlimit == 0)
817 					break;
818 				++tscan;
819 				continue;
820 			}
821 			vm_page_test_dirty(p);
822 			if (p->dirty == 0) {
823 				if (--scanlimit == 0)
824 					break;
825 				++tscan;
826 				continue;
827 			}
828 			/*
829 			 * If we have been asked to skip nosync pages and
830 			 * this is a nosync page, we can't continue.
831 			 */
832 			if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
833 				if (--scanlimit == 0)
834 					break;
835 				++tscan;
836 				continue;
837 			}
838 			scanlimit = scanreset;
839 
840 			/*
841 			 * This returns 0 if it was unable to busy the first
842 			 * page (i.e. had to sleep).
843 			 */
844 			tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
845 		}
846 
847 		/*
848 		 * If everything was dirty and we flushed it successfully,
849 		 * and the requested range is not the entire object, we
850 		 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
851 		 * return immediately.
852 		 */
853 		if (tscan >= tend && (tstart || tend < object->size)) {
854 			vm_page_unlock_queues();
855 			vm_object_clear_flag(object, OBJ_CLEANING);
856 			return;
857 		}
858 		pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
859 	}
860 
861 	/*
862 	 * Generally set CLEANCHK interlock and make the page read-only so
863 	 * we can then clear the object flags.
864 	 *
865 	 * However, if this is a nosync mmap then the object is likely to
866 	 * stay dirty so do not mess with the page and do not clear the
867 	 * object flags.
868 	 */
869 	clearobjflags = 1;
870 	TAILQ_FOREACH(p, &object->memq, listq) {
871 		p->oflags |= VPO_CLEANCHK;
872 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC))
873 			clearobjflags = 0;
874 		else
875 			pmap_remove_write(p);
876 	}
877 
878 	if (clearobjflags && (tstart == 0) && (tend == object->size))
879 		vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
880 
881 rescan:
882 	curgeneration = object->generation;
883 
884 	for (p = TAILQ_FIRST(&object->memq); p; p = np) {
885 		int n;
886 
887 		np = TAILQ_NEXT(p, listq);
888 
889 again:
890 		pi = p->pindex;
891 		if ((p->oflags & VPO_CLEANCHK) == 0 ||
892 			(pi < tstart) || (pi >= tend) ||
893 		    p->valid == 0) {
894 			p->oflags &= ~VPO_CLEANCHK;
895 			continue;
896 		}
897 
898 		vm_page_test_dirty(p);
899 		if (p->dirty == 0) {
900 			p->oflags &= ~VPO_CLEANCHK;
901 			continue;
902 		}
903 
904 		/*
905 		 * If we have been asked to skip nosync pages and this is a
906 		 * nosync page, skip it.  Note that the object flags were
907 		 * not cleared in this case so we do not have to set them.
908 		 */
909 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
910 			p->oflags &= ~VPO_CLEANCHK;
911 			continue;
912 		}
913 
914 		n = vm_object_page_collect_flush(object, p,
915 			curgeneration, pagerflags);
916 		if (n == 0)
917 			goto rescan;
918 
919 		if (object->generation != curgeneration)
920 			goto rescan;
921 
922 		/*
923 		 * Try to optimize the next page.  If we can't we pick up
924 		 * our (random) scan where we left off.
925 		 */
926 		if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
927 			if ((p = vm_page_lookup(object, pi + n)) != NULL)
928 				goto again;
929 		}
930 	}
931 	vm_page_unlock_queues();
932 #if 0
933 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
934 #endif
935 
936 	vm_object_clear_flag(object, OBJ_CLEANING);
937 	return;
938 }
939 
940 static int
941 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
942 {
943 	int runlen;
944 	int maxf;
945 	int chkb;
946 	int maxb;
947 	int i;
948 	vm_pindex_t pi;
949 	vm_page_t maf[vm_pageout_page_count];
950 	vm_page_t mab[vm_pageout_page_count];
951 	vm_page_t ma[vm_pageout_page_count];
952 
953 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
954 	pi = p->pindex;
955 	while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
956 		vm_page_lock_queues();
957 		if (object->generation != curgeneration) {
958 			return(0);
959 		}
960 	}
961 	maxf = 0;
962 	for(i = 1; i < vm_pageout_page_count; i++) {
963 		vm_page_t tp;
964 
965 		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
966 			if ((tp->oflags & VPO_BUSY) ||
967 				((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
968 				 (tp->oflags & VPO_CLEANCHK) == 0) ||
969 				(tp->busy != 0))
970 				break;
971 			vm_page_test_dirty(tp);
972 			if (tp->dirty == 0) {
973 				tp->oflags &= ~VPO_CLEANCHK;
974 				break;
975 			}
976 			maf[ i - 1 ] = tp;
977 			maxf++;
978 			continue;
979 		}
980 		break;
981 	}
982 
983 	maxb = 0;
984 	chkb = vm_pageout_page_count -  maxf;
985 	if (chkb) {
986 		for(i = 1; i < chkb;i++) {
987 			vm_page_t tp;
988 
989 			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
990 				if ((tp->oflags & VPO_BUSY) ||
991 					((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
992 					 (tp->oflags & VPO_CLEANCHK) == 0) ||
993 					(tp->busy != 0))
994 					break;
995 				vm_page_test_dirty(tp);
996 				if (tp->dirty == 0) {
997 					tp->oflags &= ~VPO_CLEANCHK;
998 					break;
999 				}
1000 				mab[ i - 1 ] = tp;
1001 				maxb++;
1002 				continue;
1003 			}
1004 			break;
1005 		}
1006 	}
1007 
1008 	for(i = 0; i < maxb; i++) {
1009 		int index = (maxb - i) - 1;
1010 		ma[index] = mab[i];
1011 		ma[index]->oflags &= ~VPO_CLEANCHK;
1012 	}
1013 	p->oflags &= ~VPO_CLEANCHK;
1014 	ma[maxb] = p;
1015 	for(i = 0; i < maxf; i++) {
1016 		int index = (maxb + i) + 1;
1017 		ma[index] = maf[i];
1018 		ma[index]->oflags &= ~VPO_CLEANCHK;
1019 	}
1020 	runlen = maxb + maxf + 1;
1021 
1022 	vm_pageout_flush(ma, runlen, pagerflags);
1023 	for (i = 0; i < runlen; i++) {
1024 		if (ma[i]->dirty) {
1025 			pmap_remove_write(ma[i]);
1026 			ma[i]->oflags |= VPO_CLEANCHK;
1027 
1028 			/*
1029 			 * maxf will end up being the actual number of pages
1030 			 * we wrote out contiguously, non-inclusive of the
1031 			 * first page.  We do not count look-behind pages.
1032 			 */
1033 			if (i >= maxb + 1 && (maxf > i - maxb - 1))
1034 				maxf = i - maxb - 1;
1035 		}
1036 	}
1037 	return(maxf + 1);
1038 }
1039 
1040 /*
1041  * Note that there is absolutely no sense in writing out
1042  * anonymous objects, so we track down the vnode object
1043  * to write out.
1044  * We invalidate (remove) all pages from the address space
1045  * for semantic correctness.
1046  *
1047  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1048  * may start out with a NULL object.
1049  */
1050 void
1051 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1052     boolean_t syncio, boolean_t invalidate)
1053 {
1054 	vm_object_t backing_object;
1055 	struct vnode *vp;
1056 	struct mount *mp;
1057 	int flags;
1058 
1059 	if (object == NULL)
1060 		return;
1061 	VM_OBJECT_LOCK(object);
1062 	while ((backing_object = object->backing_object) != NULL) {
1063 		VM_OBJECT_LOCK(backing_object);
1064 		offset += object->backing_object_offset;
1065 		VM_OBJECT_UNLOCK(object);
1066 		object = backing_object;
1067 		if (object->size < OFF_TO_IDX(offset + size))
1068 			size = IDX_TO_OFF(object->size) - offset;
1069 	}
1070 	/*
1071 	 * Flush pages if writing is allowed, invalidate them
1072 	 * if invalidation requested.  Pages undergoing I/O
1073 	 * will be ignored by vm_object_page_remove().
1074 	 *
1075 	 * We cannot lock the vnode and then wait for paging
1076 	 * to complete without deadlocking against vm_fault.
1077 	 * Instead we simply call vm_object_page_remove() and
1078 	 * allow it to block internally on a page-by-page
1079 	 * basis when it encounters pages undergoing async
1080 	 * I/O.
1081 	 */
1082 	if (object->type == OBJT_VNODE &&
1083 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1084 		int vfslocked;
1085 		vp = object->handle;
1086 		VM_OBJECT_UNLOCK(object);
1087 		(void) vn_start_write(vp, &mp, V_WAIT);
1088 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1089 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1090 		flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1091 		flags |= invalidate ? OBJPC_INVAL : 0;
1092 		VM_OBJECT_LOCK(object);
1093 		vm_object_page_clean(object,
1094 		    OFF_TO_IDX(offset),
1095 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1096 		    flags);
1097 		VM_OBJECT_UNLOCK(object);
1098 		VOP_UNLOCK(vp, 0);
1099 		VFS_UNLOCK_GIANT(vfslocked);
1100 		vn_finished_write(mp);
1101 		VM_OBJECT_LOCK(object);
1102 	}
1103 	if ((object->type == OBJT_VNODE ||
1104 	     object->type == OBJT_DEVICE) && invalidate) {
1105 		boolean_t purge;
1106 		purge = old_msync || (object->type == OBJT_DEVICE);
1107 		vm_object_page_remove(object,
1108 		    OFF_TO_IDX(offset),
1109 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1110 		    purge ? FALSE : TRUE);
1111 	}
1112 	VM_OBJECT_UNLOCK(object);
1113 }
1114 
1115 /*
1116  *	vm_object_madvise:
1117  *
1118  *	Implements the madvise function at the object/page level.
1119  *
1120  *	MADV_WILLNEED	(any object)
1121  *
1122  *	    Activate the specified pages if they are resident.
1123  *
1124  *	MADV_DONTNEED	(any object)
1125  *
1126  *	    Deactivate the specified pages if they are resident.
1127  *
1128  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1129  *			 OBJ_ONEMAPPING only)
1130  *
1131  *	    Deactivate and clean the specified pages if they are
1132  *	    resident.  This permits the process to reuse the pages
1133  *	    without faulting or the kernel to reclaim the pages
1134  *	    without I/O.
1135  */
1136 void
1137 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1138 {
1139 	vm_pindex_t end, tpindex;
1140 	vm_object_t backing_object, tobject;
1141 	vm_page_t m;
1142 
1143 	if (object == NULL)
1144 		return;
1145 	VM_OBJECT_LOCK(object);
1146 	end = pindex + count;
1147 	/*
1148 	 * Locate and adjust resident pages
1149 	 */
1150 	for (; pindex < end; pindex += 1) {
1151 relookup:
1152 		tobject = object;
1153 		tpindex = pindex;
1154 shadowlookup:
1155 		/*
1156 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1157 		 * and those pages must be OBJ_ONEMAPPING.
1158 		 */
1159 		if (advise == MADV_FREE) {
1160 			if ((tobject->type != OBJT_DEFAULT &&
1161 			     tobject->type != OBJT_SWAP) ||
1162 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
1163 				goto unlock_tobject;
1164 			}
1165 		} else if (tobject->type == OBJT_PHYS)
1166 			goto unlock_tobject;
1167 		m = vm_page_lookup(tobject, tpindex);
1168 		if (m == NULL && advise == MADV_WILLNEED) {
1169 			/*
1170 			 * If the page is cached, reactivate it.
1171 			 */
1172 			m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1173 			    VM_ALLOC_NOBUSY);
1174 		}
1175 		if (m == NULL) {
1176 			/*
1177 			 * There may be swap even if there is no backing page
1178 			 */
1179 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1180 				swap_pager_freespace(tobject, tpindex, 1);
1181 			/*
1182 			 * next object
1183 			 */
1184 			backing_object = tobject->backing_object;
1185 			if (backing_object == NULL)
1186 				goto unlock_tobject;
1187 			VM_OBJECT_LOCK(backing_object);
1188 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1189 			if (tobject != object)
1190 				VM_OBJECT_UNLOCK(tobject);
1191 			tobject = backing_object;
1192 			goto shadowlookup;
1193 		} else if (m->valid != VM_PAGE_BITS_ALL)
1194 			goto unlock_tobject;
1195 		/*
1196 		 * If the page is not in a normal state, skip it.
1197 		 */
1198 		vm_page_lock_queues();
1199 		if (m->hold_count != 0 || m->wire_count != 0) {
1200 			vm_page_unlock_queues();
1201 			goto unlock_tobject;
1202 		}
1203 		if ((m->oflags & VPO_BUSY) || m->busy) {
1204 			if (advise == MADV_WILLNEED)
1205 				/*
1206 				 * Reference the page before unlocking and
1207 				 * sleeping so that the page daemon is less
1208 				 * likely to reclaim it.
1209 				 */
1210 				vm_page_flag_set(m, PG_REFERENCED);
1211 			vm_page_unlock_queues();
1212 			if (object != tobject)
1213 				VM_OBJECT_UNLOCK(object);
1214 			m->oflags |= VPO_WANTED;
1215 			msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1216 			    0);
1217 			VM_OBJECT_LOCK(object);
1218   			goto relookup;
1219 		}
1220 		if (advise == MADV_WILLNEED) {
1221 			vm_page_activate(m);
1222 		} else if (advise == MADV_DONTNEED) {
1223 			vm_page_dontneed(m);
1224 		} else if (advise == MADV_FREE) {
1225 			/*
1226 			 * Mark the page clean.  This will allow the page
1227 			 * to be freed up by the system.  However, such pages
1228 			 * are often reused quickly by malloc()/free()
1229 			 * so we do not do anything that would cause
1230 			 * a page fault if we can help it.
1231 			 *
1232 			 * Specifically, we do not try to actually free
1233 			 * the page now nor do we try to put it in the
1234 			 * cache (which would cause a page fault on reuse).
1235 			 *
1236 			 * But we do make the page is freeable as we
1237 			 * can without actually taking the step of unmapping
1238 			 * it.
1239 			 */
1240 			pmap_clear_modify(m);
1241 			m->dirty = 0;
1242 			m->act_count = 0;
1243 			vm_page_dontneed(m);
1244 		}
1245 		vm_page_unlock_queues();
1246 		if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1247 			swap_pager_freespace(tobject, tpindex, 1);
1248 unlock_tobject:
1249 		if (tobject != object)
1250 			VM_OBJECT_UNLOCK(tobject);
1251 	}
1252 	VM_OBJECT_UNLOCK(object);
1253 }
1254 
1255 /*
1256  *	vm_object_shadow:
1257  *
1258  *	Create a new object which is backed by the
1259  *	specified existing object range.  The source
1260  *	object reference is deallocated.
1261  *
1262  *	The new object and offset into that object
1263  *	are returned in the source parameters.
1264  */
1265 void
1266 vm_object_shadow(
1267 	vm_object_t *object,	/* IN/OUT */
1268 	vm_ooffset_t *offset,	/* IN/OUT */
1269 	vm_size_t length)
1270 {
1271 	vm_object_t source;
1272 	vm_object_t result;
1273 
1274 	source = *object;
1275 
1276 	/*
1277 	 * Don't create the new object if the old object isn't shared.
1278 	 */
1279 	if (source != NULL) {
1280 		VM_OBJECT_LOCK(source);
1281 		if (source->ref_count == 1 &&
1282 		    source->handle == NULL &&
1283 		    (source->type == OBJT_DEFAULT ||
1284 		     source->type == OBJT_SWAP)) {
1285 			VM_OBJECT_UNLOCK(source);
1286 			return;
1287 		}
1288 		VM_OBJECT_UNLOCK(source);
1289 	}
1290 
1291 	/*
1292 	 * Allocate a new object with the given length.
1293 	 */
1294 	result = vm_object_allocate(OBJT_DEFAULT, length);
1295 
1296 	/*
1297 	 * The new object shadows the source object, adding a reference to it.
1298 	 * Our caller changes his reference to point to the new object,
1299 	 * removing a reference to the source object.  Net result: no change
1300 	 * of reference count.
1301 	 *
1302 	 * Try to optimize the result object's page color when shadowing
1303 	 * in order to maintain page coloring consistency in the combined
1304 	 * shadowed object.
1305 	 */
1306 	result->backing_object = source;
1307 	/*
1308 	 * Store the offset into the source object, and fix up the offset into
1309 	 * the new object.
1310 	 */
1311 	result->backing_object_offset = *offset;
1312 	if (source != NULL) {
1313 		VM_OBJECT_LOCK(source);
1314 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1315 		source->shadow_count++;
1316 		source->generation++;
1317 #if VM_NRESERVLEVEL > 0
1318 		result->flags |= source->flags & OBJ_COLORED;
1319 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1320 		    ((1 << (VM_NFREEORDER - 1)) - 1);
1321 #endif
1322 		VM_OBJECT_UNLOCK(source);
1323 	}
1324 
1325 
1326 	/*
1327 	 * Return the new things
1328 	 */
1329 	*offset = 0;
1330 	*object = result;
1331 }
1332 
1333 /*
1334  *	vm_object_split:
1335  *
1336  * Split the pages in a map entry into a new object.  This affords
1337  * easier removal of unused pages, and keeps object inheritance from
1338  * being a negative impact on memory usage.
1339  */
1340 void
1341 vm_object_split(vm_map_entry_t entry)
1342 {
1343 	vm_page_t m, m_next;
1344 	vm_object_t orig_object, new_object, source;
1345 	vm_pindex_t idx, offidxstart;
1346 	vm_size_t size;
1347 
1348 	orig_object = entry->object.vm_object;
1349 	if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1350 		return;
1351 	if (orig_object->ref_count <= 1)
1352 		return;
1353 	VM_OBJECT_UNLOCK(orig_object);
1354 
1355 	offidxstart = OFF_TO_IDX(entry->offset);
1356 	size = atop(entry->end - entry->start);
1357 
1358 	/*
1359 	 * If swap_pager_copy() is later called, it will convert new_object
1360 	 * into a swap object.
1361 	 */
1362 	new_object = vm_object_allocate(OBJT_DEFAULT, size);
1363 
1364 	/*
1365 	 * At this point, the new object is still private, so the order in
1366 	 * which the original and new objects are locked does not matter.
1367 	 */
1368 	VM_OBJECT_LOCK(new_object);
1369 	VM_OBJECT_LOCK(orig_object);
1370 	source = orig_object->backing_object;
1371 	if (source != NULL) {
1372 		VM_OBJECT_LOCK(source);
1373 		if ((source->flags & OBJ_DEAD) != 0) {
1374 			VM_OBJECT_UNLOCK(source);
1375 			VM_OBJECT_UNLOCK(orig_object);
1376 			VM_OBJECT_UNLOCK(new_object);
1377 			vm_object_deallocate(new_object);
1378 			VM_OBJECT_LOCK(orig_object);
1379 			return;
1380 		}
1381 		LIST_INSERT_HEAD(&source->shadow_head,
1382 				  new_object, shadow_list);
1383 		source->shadow_count++;
1384 		source->generation++;
1385 		vm_object_reference_locked(source);	/* for new_object */
1386 		vm_object_clear_flag(source, OBJ_ONEMAPPING);
1387 		VM_OBJECT_UNLOCK(source);
1388 		new_object->backing_object_offset =
1389 			orig_object->backing_object_offset + entry->offset;
1390 		new_object->backing_object = source;
1391 	}
1392 	if (orig_object->uip != NULL) {
1393 		new_object->uip = orig_object->uip;
1394 		uihold(orig_object->uip);
1395 		new_object->charge = ptoa(size);
1396 		KASSERT(orig_object->charge >= ptoa(size),
1397 		    ("orig_object->charge < 0"));
1398 		orig_object->charge -= ptoa(size);
1399 	}
1400 retry:
1401 	if ((m = TAILQ_FIRST(&orig_object->memq)) != NULL) {
1402 		if (m->pindex < offidxstart) {
1403 			m = vm_page_splay(offidxstart, orig_object->root);
1404 			if ((orig_object->root = m)->pindex < offidxstart)
1405 				m = TAILQ_NEXT(m, listq);
1406 		}
1407 	}
1408 	vm_page_lock_queues();
1409 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1410 	    m = m_next) {
1411 		m_next = TAILQ_NEXT(m, listq);
1412 
1413 		/*
1414 		 * We must wait for pending I/O to complete before we can
1415 		 * rename the page.
1416 		 *
1417 		 * We do not have to VM_PROT_NONE the page as mappings should
1418 		 * not be changed by this operation.
1419 		 */
1420 		if ((m->oflags & VPO_BUSY) || m->busy) {
1421 			vm_page_unlock_queues();
1422 			VM_OBJECT_UNLOCK(new_object);
1423 			m->oflags |= VPO_WANTED;
1424 			msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1425 			VM_OBJECT_LOCK(new_object);
1426 			goto retry;
1427 		}
1428 		vm_page_rename(m, new_object, idx);
1429 		/* page automatically made dirty by rename and cache handled */
1430 		vm_page_busy(m);
1431 	}
1432 	vm_page_unlock_queues();
1433 	if (orig_object->type == OBJT_SWAP) {
1434 		/*
1435 		 * swap_pager_copy() can sleep, in which case the orig_object's
1436 		 * and new_object's locks are released and reacquired.
1437 		 */
1438 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1439 
1440 		/*
1441 		 * Transfer any cached pages from orig_object to new_object.
1442 		 */
1443 		if (__predict_false(orig_object->cache != NULL))
1444 			vm_page_cache_transfer(orig_object, offidxstart,
1445 			    new_object);
1446 	}
1447 	VM_OBJECT_UNLOCK(orig_object);
1448 	TAILQ_FOREACH(m, &new_object->memq, listq)
1449 		vm_page_wakeup(m);
1450 	VM_OBJECT_UNLOCK(new_object);
1451 	entry->object.vm_object = new_object;
1452 	entry->offset = 0LL;
1453 	vm_object_deallocate(orig_object);
1454 	VM_OBJECT_LOCK(new_object);
1455 }
1456 
1457 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1458 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1459 #define	OBSC_COLLAPSE_WAIT	0x0004
1460 
1461 static int
1462 vm_object_backing_scan(vm_object_t object, int op)
1463 {
1464 	int r = 1;
1465 	vm_page_t p;
1466 	vm_object_t backing_object;
1467 	vm_pindex_t backing_offset_index;
1468 
1469 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1470 	VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1471 
1472 	backing_object = object->backing_object;
1473 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1474 
1475 	/*
1476 	 * Initial conditions
1477 	 */
1478 	if (op & OBSC_TEST_ALL_SHADOWED) {
1479 		/*
1480 		 * We do not want to have to test for the existence of cache
1481 		 * or swap pages in the backing object.  XXX but with the
1482 		 * new swapper this would be pretty easy to do.
1483 		 *
1484 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1485 		 * been ZFOD faulted yet?  If we do not test for this, the
1486 		 * shadow test may succeed! XXX
1487 		 */
1488 		if (backing_object->type != OBJT_DEFAULT) {
1489 			return (0);
1490 		}
1491 	}
1492 	if (op & OBSC_COLLAPSE_WAIT) {
1493 		vm_object_set_flag(backing_object, OBJ_DEAD);
1494 	}
1495 
1496 	/*
1497 	 * Our scan
1498 	 */
1499 	p = TAILQ_FIRST(&backing_object->memq);
1500 	while (p) {
1501 		vm_page_t next = TAILQ_NEXT(p, listq);
1502 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1503 
1504 		if (op & OBSC_TEST_ALL_SHADOWED) {
1505 			vm_page_t pp;
1506 
1507 			/*
1508 			 * Ignore pages outside the parent object's range
1509 			 * and outside the parent object's mapping of the
1510 			 * backing object.
1511 			 *
1512 			 * note that we do not busy the backing object's
1513 			 * page.
1514 			 */
1515 			if (
1516 			    p->pindex < backing_offset_index ||
1517 			    new_pindex >= object->size
1518 			) {
1519 				p = next;
1520 				continue;
1521 			}
1522 
1523 			/*
1524 			 * See if the parent has the page or if the parent's
1525 			 * object pager has the page.  If the parent has the
1526 			 * page but the page is not valid, the parent's
1527 			 * object pager must have the page.
1528 			 *
1529 			 * If this fails, the parent does not completely shadow
1530 			 * the object and we might as well give up now.
1531 			 */
1532 
1533 			pp = vm_page_lookup(object, new_pindex);
1534 			if (
1535 			    (pp == NULL || pp->valid == 0) &&
1536 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1537 			) {
1538 				r = 0;
1539 				break;
1540 			}
1541 		}
1542 
1543 		/*
1544 		 * Check for busy page
1545 		 */
1546 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1547 			vm_page_t pp;
1548 
1549 			if (op & OBSC_COLLAPSE_NOWAIT) {
1550 				if ((p->oflags & VPO_BUSY) ||
1551 				    !p->valid ||
1552 				    p->busy) {
1553 					p = next;
1554 					continue;
1555 				}
1556 			} else if (op & OBSC_COLLAPSE_WAIT) {
1557 				if ((p->oflags & VPO_BUSY) || p->busy) {
1558 					VM_OBJECT_UNLOCK(object);
1559 					p->oflags |= VPO_WANTED;
1560 					msleep(p, VM_OBJECT_MTX(backing_object),
1561 					    PDROP | PVM, "vmocol", 0);
1562 					VM_OBJECT_LOCK(object);
1563 					VM_OBJECT_LOCK(backing_object);
1564 					/*
1565 					 * If we slept, anything could have
1566 					 * happened.  Since the object is
1567 					 * marked dead, the backing offset
1568 					 * should not have changed so we
1569 					 * just restart our scan.
1570 					 */
1571 					p = TAILQ_FIRST(&backing_object->memq);
1572 					continue;
1573 				}
1574 			}
1575 
1576 			KASSERT(
1577 			    p->object == backing_object,
1578 			    ("vm_object_backing_scan: object mismatch")
1579 			);
1580 
1581 			/*
1582 			 * Destroy any associated swap
1583 			 */
1584 			if (backing_object->type == OBJT_SWAP) {
1585 				swap_pager_freespace(
1586 				    backing_object,
1587 				    p->pindex,
1588 				    1
1589 				);
1590 			}
1591 
1592 			if (
1593 			    p->pindex < backing_offset_index ||
1594 			    new_pindex >= object->size
1595 			) {
1596 				/*
1597 				 * Page is out of the parent object's range, we
1598 				 * can simply destroy it.
1599 				 */
1600 				vm_page_lock_queues();
1601 				KASSERT(!pmap_page_is_mapped(p),
1602 				    ("freeing mapped page %p", p));
1603 				if (p->wire_count == 0)
1604 					vm_page_free(p);
1605 				else
1606 					vm_page_remove(p);
1607 				vm_page_unlock_queues();
1608 				p = next;
1609 				continue;
1610 			}
1611 
1612 			pp = vm_page_lookup(object, new_pindex);
1613 			if (
1614 			    pp != NULL ||
1615 			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1616 			) {
1617 				/*
1618 				 * page already exists in parent OR swap exists
1619 				 * for this location in the parent.  Destroy
1620 				 * the original page from the backing object.
1621 				 *
1622 				 * Leave the parent's page alone
1623 				 */
1624 				vm_page_lock_queues();
1625 				KASSERT(!pmap_page_is_mapped(p),
1626 				    ("freeing mapped page %p", p));
1627 				if (p->wire_count == 0)
1628 					vm_page_free(p);
1629 				else
1630 					vm_page_remove(p);
1631 				vm_page_unlock_queues();
1632 				p = next;
1633 				continue;
1634 			}
1635 
1636 #if VM_NRESERVLEVEL > 0
1637 			/*
1638 			 * Rename the reservation.
1639 			 */
1640 			vm_reserv_rename(p, object, backing_object,
1641 			    backing_offset_index);
1642 #endif
1643 
1644 			/*
1645 			 * Page does not exist in parent, rename the
1646 			 * page from the backing object to the main object.
1647 			 *
1648 			 * If the page was mapped to a process, it can remain
1649 			 * mapped through the rename.
1650 			 */
1651 			vm_page_lock_queues();
1652 			vm_page_rename(p, object, new_pindex);
1653 			vm_page_unlock_queues();
1654 			/* page automatically made dirty by rename */
1655 		}
1656 		p = next;
1657 	}
1658 	return (r);
1659 }
1660 
1661 
1662 /*
1663  * this version of collapse allows the operation to occur earlier and
1664  * when paging_in_progress is true for an object...  This is not a complete
1665  * operation, but should plug 99.9% of the rest of the leaks.
1666  */
1667 static void
1668 vm_object_qcollapse(vm_object_t object)
1669 {
1670 	vm_object_t backing_object = object->backing_object;
1671 
1672 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1673 	VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1674 
1675 	if (backing_object->ref_count != 1)
1676 		return;
1677 
1678 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1679 }
1680 
1681 /*
1682  *	vm_object_collapse:
1683  *
1684  *	Collapse an object with the object backing it.
1685  *	Pages in the backing object are moved into the
1686  *	parent, and the backing object is deallocated.
1687  */
1688 void
1689 vm_object_collapse(vm_object_t object)
1690 {
1691 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1692 
1693 	while (TRUE) {
1694 		vm_object_t backing_object;
1695 
1696 		/*
1697 		 * Verify that the conditions are right for collapse:
1698 		 *
1699 		 * The object exists and the backing object exists.
1700 		 */
1701 		if ((backing_object = object->backing_object) == NULL)
1702 			break;
1703 
1704 		/*
1705 		 * we check the backing object first, because it is most likely
1706 		 * not collapsable.
1707 		 */
1708 		VM_OBJECT_LOCK(backing_object);
1709 		if (backing_object->handle != NULL ||
1710 		    (backing_object->type != OBJT_DEFAULT &&
1711 		     backing_object->type != OBJT_SWAP) ||
1712 		    (backing_object->flags & OBJ_DEAD) ||
1713 		    object->handle != NULL ||
1714 		    (object->type != OBJT_DEFAULT &&
1715 		     object->type != OBJT_SWAP) ||
1716 		    (object->flags & OBJ_DEAD)) {
1717 			VM_OBJECT_UNLOCK(backing_object);
1718 			break;
1719 		}
1720 
1721 		if (
1722 		    object->paging_in_progress != 0 ||
1723 		    backing_object->paging_in_progress != 0
1724 		) {
1725 			vm_object_qcollapse(object);
1726 			VM_OBJECT_UNLOCK(backing_object);
1727 			break;
1728 		}
1729 		/*
1730 		 * We know that we can either collapse the backing object (if
1731 		 * the parent is the only reference to it) or (perhaps) have
1732 		 * the parent bypass the object if the parent happens to shadow
1733 		 * all the resident pages in the entire backing object.
1734 		 *
1735 		 * This is ignoring pager-backed pages such as swap pages.
1736 		 * vm_object_backing_scan fails the shadowing test in this
1737 		 * case.
1738 		 */
1739 		if (backing_object->ref_count == 1) {
1740 			/*
1741 			 * If there is exactly one reference to the backing
1742 			 * object, we can collapse it into the parent.
1743 			 */
1744 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1745 
1746 #if VM_NRESERVLEVEL > 0
1747 			/*
1748 			 * Break any reservations from backing_object.
1749 			 */
1750 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1751 				vm_reserv_break_all(backing_object);
1752 #endif
1753 
1754 			/*
1755 			 * Move the pager from backing_object to object.
1756 			 */
1757 			if (backing_object->type == OBJT_SWAP) {
1758 				/*
1759 				 * swap_pager_copy() can sleep, in which case
1760 				 * the backing_object's and object's locks are
1761 				 * released and reacquired.
1762 				 */
1763 				swap_pager_copy(
1764 				    backing_object,
1765 				    object,
1766 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1767 
1768 				/*
1769 				 * Free any cached pages from backing_object.
1770 				 */
1771 				if (__predict_false(backing_object->cache != NULL))
1772 					vm_page_cache_free(backing_object, 0, 0);
1773 			}
1774 			/*
1775 			 * Object now shadows whatever backing_object did.
1776 			 * Note that the reference to
1777 			 * backing_object->backing_object moves from within
1778 			 * backing_object to within object.
1779 			 */
1780 			LIST_REMOVE(object, shadow_list);
1781 			backing_object->shadow_count--;
1782 			backing_object->generation++;
1783 			if (backing_object->backing_object) {
1784 				VM_OBJECT_LOCK(backing_object->backing_object);
1785 				LIST_REMOVE(backing_object, shadow_list);
1786 				LIST_INSERT_HEAD(
1787 				    &backing_object->backing_object->shadow_head,
1788 				    object, shadow_list);
1789 				/*
1790 				 * The shadow_count has not changed.
1791 				 */
1792 				backing_object->backing_object->generation++;
1793 				VM_OBJECT_UNLOCK(backing_object->backing_object);
1794 			}
1795 			object->backing_object = backing_object->backing_object;
1796 			object->backing_object_offset +=
1797 			    backing_object->backing_object_offset;
1798 
1799 			/*
1800 			 * Discard backing_object.
1801 			 *
1802 			 * Since the backing object has no pages, no pager left,
1803 			 * and no object references within it, all that is
1804 			 * necessary is to dispose of it.
1805 			 */
1806 			KASSERT(backing_object->ref_count == 1, (
1807 "backing_object %p was somehow re-referenced during collapse!",
1808 			    backing_object));
1809 			VM_OBJECT_UNLOCK(backing_object);
1810 			vm_object_destroy(backing_object);
1811 
1812 			object_collapses++;
1813 		} else {
1814 			vm_object_t new_backing_object;
1815 
1816 			/*
1817 			 * If we do not entirely shadow the backing object,
1818 			 * there is nothing we can do so we give up.
1819 			 */
1820 			if (object->resident_page_count != object->size &&
1821 			    vm_object_backing_scan(object,
1822 			    OBSC_TEST_ALL_SHADOWED) == 0) {
1823 				VM_OBJECT_UNLOCK(backing_object);
1824 				break;
1825 			}
1826 
1827 			/*
1828 			 * Make the parent shadow the next object in the
1829 			 * chain.  Deallocating backing_object will not remove
1830 			 * it, since its reference count is at least 2.
1831 			 */
1832 			LIST_REMOVE(object, shadow_list);
1833 			backing_object->shadow_count--;
1834 			backing_object->generation++;
1835 
1836 			new_backing_object = backing_object->backing_object;
1837 			if ((object->backing_object = new_backing_object) != NULL) {
1838 				VM_OBJECT_LOCK(new_backing_object);
1839 				LIST_INSERT_HEAD(
1840 				    &new_backing_object->shadow_head,
1841 				    object,
1842 				    shadow_list
1843 				);
1844 				new_backing_object->shadow_count++;
1845 				new_backing_object->generation++;
1846 				vm_object_reference_locked(new_backing_object);
1847 				VM_OBJECT_UNLOCK(new_backing_object);
1848 				object->backing_object_offset +=
1849 					backing_object->backing_object_offset;
1850 			}
1851 
1852 			/*
1853 			 * Drop the reference count on backing_object. Since
1854 			 * its ref_count was at least 2, it will not vanish.
1855 			 */
1856 			backing_object->ref_count--;
1857 			VM_OBJECT_UNLOCK(backing_object);
1858 			object_bypasses++;
1859 		}
1860 
1861 		/*
1862 		 * Try again with this object's new backing object.
1863 		 */
1864 	}
1865 }
1866 
1867 /*
1868  *	vm_object_page_remove:
1869  *
1870  *	For the given object, either frees or invalidates each of the
1871  *	specified pages.  In general, a page is freed.  However, if a
1872  *	page is wired for any reason other than the existence of a
1873  *	managed, wired mapping, then it may be invalidated but not
1874  *	removed from the object.  Pages are specified by the given
1875  *	range ["start", "end") and Boolean "clean_only".  As a
1876  *	special case, if "end" is zero, then the range extends from
1877  *	"start" to the end of the object.  If "clean_only" is TRUE,
1878  *	then only the non-dirty pages within the specified range are
1879  *	affected.
1880  *
1881  *	In general, this operation should only be performed on objects
1882  *	that contain managed pages.  There are two exceptions.  First,
1883  *	it may be performed on the kernel and kmem objects.  Second,
1884  *	it may be used by msync(..., MS_INVALIDATE) to invalidate
1885  *	device-backed pages.
1886  *
1887  *	The object must be locked.
1888  */
1889 void
1890 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1891     boolean_t clean_only)
1892 {
1893 	vm_page_t p, next;
1894 	int wirings;
1895 
1896 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1897 	if (object->resident_page_count == 0)
1898 		goto skipmemq;
1899 
1900 	/*
1901 	 * Since physically-backed objects do not use managed pages, we can't
1902 	 * remove pages from the object (we must instead remove the page
1903 	 * references, and then destroy the object).
1904 	 */
1905 	KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1906 	    object == kmem_object,
1907 	    ("attempt to remove pages from a physical object"));
1908 
1909 	vm_object_pip_add(object, 1);
1910 again:
1911 	if ((p = TAILQ_FIRST(&object->memq)) != NULL) {
1912 		if (p->pindex < start) {
1913 			p = vm_page_splay(start, object->root);
1914 			if ((object->root = p)->pindex < start)
1915 				p = TAILQ_NEXT(p, listq);
1916 		}
1917 	}
1918 	vm_page_lock_queues();
1919 	/*
1920 	 * Assert: the variable p is either (1) the page with the
1921 	 * least pindex greater than or equal to the parameter pindex
1922 	 * or (2) NULL.
1923 	 */
1924 	for (;
1925 	     p != NULL && (p->pindex < end || end == 0);
1926 	     p = next) {
1927 		next = TAILQ_NEXT(p, listq);
1928 
1929 		/*
1930 		 * If the page is wired for any reason besides the
1931 		 * existence of managed, wired mappings, then it cannot
1932 		 * be freed.  For example, fictitious pages, which
1933 		 * represent device memory, are inherently wired and
1934 		 * cannot be freed.  They can, however, be invalidated
1935 		 * if "clean_only" is FALSE.
1936 		 */
1937 		if ((wirings = p->wire_count) != 0 &&
1938 		    (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1939 			/* Fictitious pages do not have managed mappings. */
1940 			if ((p->flags & PG_FICTITIOUS) == 0)
1941 				pmap_remove_all(p);
1942 			/* Account for removal of managed, wired mappings. */
1943 			p->wire_count -= wirings;
1944 			if (!clean_only) {
1945 				p->valid = 0;
1946 				vm_page_undirty(p);
1947 			}
1948 			continue;
1949 		}
1950 		if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1951 			goto again;
1952 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
1953 		    ("vm_object_page_remove: page %p is fictitious", p));
1954 		if (clean_only && p->valid) {
1955 			pmap_remove_write(p);
1956 			if (p->dirty)
1957 				continue;
1958 		}
1959 		pmap_remove_all(p);
1960 		/* Account for removal of managed, wired mappings. */
1961 		if (wirings != 0)
1962 			p->wire_count -= wirings;
1963 		vm_page_free(p);
1964 	}
1965 	vm_page_unlock_queues();
1966 	vm_object_pip_wakeup(object);
1967 skipmemq:
1968 	if (__predict_false(object->cache != NULL))
1969 		vm_page_cache_free(object, start, end);
1970 }
1971 
1972 /*
1973  *	Populate the specified range of the object with valid pages.  Returns
1974  *	TRUE if the range is successfully populated and FALSE otherwise.
1975  *
1976  *	Note: This function should be optimized to pass a larger array of
1977  *	pages to vm_pager_get_pages() before it is applied to a non-
1978  *	OBJT_DEVICE object.
1979  *
1980  *	The object must be locked.
1981  */
1982 boolean_t
1983 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1984 {
1985 	vm_page_t m, ma[1];
1986 	vm_pindex_t pindex;
1987 	int rv;
1988 
1989 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1990 	for (pindex = start; pindex < end; pindex++) {
1991 		m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1992 		    VM_ALLOC_RETRY);
1993 		if (m->valid != VM_PAGE_BITS_ALL) {
1994 			ma[0] = m;
1995 			rv = vm_pager_get_pages(object, ma, 1, 0);
1996 			m = vm_page_lookup(object, pindex);
1997 			if (m == NULL)
1998 				break;
1999 			if (rv != VM_PAGER_OK) {
2000 				vm_page_lock_queues();
2001 				vm_page_free(m);
2002 				vm_page_unlock_queues();
2003 				break;
2004 			}
2005 		}
2006 		/*
2007 		 * Keep "m" busy because a subsequent iteration may unlock
2008 		 * the object.
2009 		 */
2010 	}
2011 	if (pindex > start) {
2012 		m = vm_page_lookup(object, start);
2013 		while (m != NULL && m->pindex < pindex) {
2014 			vm_page_wakeup(m);
2015 			m = TAILQ_NEXT(m, listq);
2016 		}
2017 	}
2018 	return (pindex == end);
2019 }
2020 
2021 /*
2022  *	Routine:	vm_object_coalesce
2023  *	Function:	Coalesces two objects backing up adjoining
2024  *			regions of memory into a single object.
2025  *
2026  *	returns TRUE if objects were combined.
2027  *
2028  *	NOTE:	Only works at the moment if the second object is NULL -
2029  *		if it's not, which object do we lock first?
2030  *
2031  *	Parameters:
2032  *		prev_object	First object to coalesce
2033  *		prev_offset	Offset into prev_object
2034  *		prev_size	Size of reference to prev_object
2035  *		next_size	Size of reference to the second object
2036  *		reserved	Indicator that extension region has
2037  *				swap accounted for
2038  *
2039  *	Conditions:
2040  *	The object must *not* be locked.
2041  */
2042 boolean_t
2043 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2044     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2045 {
2046 	vm_pindex_t next_pindex;
2047 
2048 	if (prev_object == NULL)
2049 		return (TRUE);
2050 	VM_OBJECT_LOCK(prev_object);
2051 	if (prev_object->type != OBJT_DEFAULT &&
2052 	    prev_object->type != OBJT_SWAP) {
2053 		VM_OBJECT_UNLOCK(prev_object);
2054 		return (FALSE);
2055 	}
2056 
2057 	/*
2058 	 * Try to collapse the object first
2059 	 */
2060 	vm_object_collapse(prev_object);
2061 
2062 	/*
2063 	 * Can't coalesce if: . more than one reference . paged out . shadows
2064 	 * another object . has a copy elsewhere (any of which mean that the
2065 	 * pages not mapped to prev_entry may be in use anyway)
2066 	 */
2067 	if (prev_object->backing_object != NULL) {
2068 		VM_OBJECT_UNLOCK(prev_object);
2069 		return (FALSE);
2070 	}
2071 
2072 	prev_size >>= PAGE_SHIFT;
2073 	next_size >>= PAGE_SHIFT;
2074 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2075 
2076 	if ((prev_object->ref_count > 1) &&
2077 	    (prev_object->size != next_pindex)) {
2078 		VM_OBJECT_UNLOCK(prev_object);
2079 		return (FALSE);
2080 	}
2081 
2082 	/*
2083 	 * Account for the charge.
2084 	 */
2085 	if (prev_object->uip != NULL) {
2086 
2087 		/*
2088 		 * If prev_object was charged, then this mapping,
2089 		 * althought not charged now, may become writable
2090 		 * later. Non-NULL uip in the object would prevent
2091 		 * swap reservation during enabling of the write
2092 		 * access, so reserve swap now. Failed reservation
2093 		 * cause allocation of the separate object for the map
2094 		 * entry, and swap reservation for this entry is
2095 		 * managed in appropriate time.
2096 		 */
2097 		if (!reserved && !swap_reserve_by_uid(ptoa(next_size),
2098 		    prev_object->uip)) {
2099 			return (FALSE);
2100 		}
2101 		prev_object->charge += ptoa(next_size);
2102 	}
2103 
2104 	/*
2105 	 * Remove any pages that may still be in the object from a previous
2106 	 * deallocation.
2107 	 */
2108 	if (next_pindex < prev_object->size) {
2109 		vm_object_page_remove(prev_object,
2110 				      next_pindex,
2111 				      next_pindex + next_size, FALSE);
2112 		if (prev_object->type == OBJT_SWAP)
2113 			swap_pager_freespace(prev_object,
2114 					     next_pindex, next_size);
2115 #if 0
2116 		if (prev_object->uip != NULL) {
2117 			KASSERT(prev_object->charge >=
2118 			    ptoa(prev_object->size - next_pindex),
2119 			    ("object %p overcharged 1 %jx %jx", prev_object,
2120 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2121 			prev_object->charge -= ptoa(prev_object->size -
2122 			    next_pindex);
2123 		}
2124 #endif
2125 	}
2126 
2127 	/*
2128 	 * Extend the object if necessary.
2129 	 */
2130 	if (next_pindex + next_size > prev_object->size)
2131 		prev_object->size = next_pindex + next_size;
2132 
2133 	VM_OBJECT_UNLOCK(prev_object);
2134 	return (TRUE);
2135 }
2136 
2137 void
2138 vm_object_set_writeable_dirty(vm_object_t object)
2139 {
2140 
2141 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2142 	if (object->type != OBJT_VNODE ||
2143 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0)
2144 		return;
2145 	vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2146 }
2147 
2148 #include "opt_ddb.h"
2149 #ifdef DDB
2150 #include <sys/kernel.h>
2151 
2152 #include <sys/cons.h>
2153 
2154 #include <ddb/ddb.h>
2155 
2156 static int
2157 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2158 {
2159 	vm_map_t tmpm;
2160 	vm_map_entry_t tmpe;
2161 	vm_object_t obj;
2162 	int entcount;
2163 
2164 	if (map == 0)
2165 		return 0;
2166 
2167 	if (entry == 0) {
2168 		tmpe = map->header.next;
2169 		entcount = map->nentries;
2170 		while (entcount-- && (tmpe != &map->header)) {
2171 			if (_vm_object_in_map(map, object, tmpe)) {
2172 				return 1;
2173 			}
2174 			tmpe = tmpe->next;
2175 		}
2176 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2177 		tmpm = entry->object.sub_map;
2178 		tmpe = tmpm->header.next;
2179 		entcount = tmpm->nentries;
2180 		while (entcount-- && tmpe != &tmpm->header) {
2181 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2182 				return 1;
2183 			}
2184 			tmpe = tmpe->next;
2185 		}
2186 	} else if ((obj = entry->object.vm_object) != NULL) {
2187 		for (; obj; obj = obj->backing_object)
2188 			if (obj == object) {
2189 				return 1;
2190 			}
2191 	}
2192 	return 0;
2193 }
2194 
2195 static int
2196 vm_object_in_map(vm_object_t object)
2197 {
2198 	struct proc *p;
2199 
2200 	/* sx_slock(&allproc_lock); */
2201 	FOREACH_PROC_IN_SYSTEM(p) {
2202 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2203 			continue;
2204 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2205 			/* sx_sunlock(&allproc_lock); */
2206 			return 1;
2207 		}
2208 	}
2209 	/* sx_sunlock(&allproc_lock); */
2210 	if (_vm_object_in_map(kernel_map, object, 0))
2211 		return 1;
2212 	if (_vm_object_in_map(kmem_map, object, 0))
2213 		return 1;
2214 	if (_vm_object_in_map(pager_map, object, 0))
2215 		return 1;
2216 	if (_vm_object_in_map(buffer_map, object, 0))
2217 		return 1;
2218 	return 0;
2219 }
2220 
2221 DB_SHOW_COMMAND(vmochk, vm_object_check)
2222 {
2223 	vm_object_t object;
2224 
2225 	/*
2226 	 * make sure that internal objs are in a map somewhere
2227 	 * and none have zero ref counts.
2228 	 */
2229 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2230 		if (object->handle == NULL &&
2231 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2232 			if (object->ref_count == 0) {
2233 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2234 					(long)object->size);
2235 			}
2236 			if (!vm_object_in_map(object)) {
2237 				db_printf(
2238 			"vmochk: internal obj is not in a map: "
2239 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2240 				    object->ref_count, (u_long)object->size,
2241 				    (u_long)object->size,
2242 				    (void *)object->backing_object);
2243 			}
2244 		}
2245 	}
2246 }
2247 
2248 /*
2249  *	vm_object_print:	[ debug ]
2250  */
2251 DB_SHOW_COMMAND(object, vm_object_print_static)
2252 {
2253 	/* XXX convert args. */
2254 	vm_object_t object = (vm_object_t)addr;
2255 	boolean_t full = have_addr;
2256 
2257 	vm_page_t p;
2258 
2259 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2260 #define	count	was_count
2261 
2262 	int count;
2263 
2264 	if (object == NULL)
2265 		return;
2266 
2267 	db_iprintf(
2268 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x uip %d charge %jx\n",
2269 	    object, (int)object->type, (uintmax_t)object->size,
2270 	    object->resident_page_count, object->ref_count, object->flags,
2271 	    object->uip ? object->uip->ui_uid : -1, (uintmax_t)object->charge);
2272 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2273 	    object->shadow_count,
2274 	    object->backing_object ? object->backing_object->ref_count : 0,
2275 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2276 
2277 	if (!full)
2278 		return;
2279 
2280 	db_indent += 2;
2281 	count = 0;
2282 	TAILQ_FOREACH(p, &object->memq, listq) {
2283 		if (count == 0)
2284 			db_iprintf("memory:=");
2285 		else if (count == 6) {
2286 			db_printf("\n");
2287 			db_iprintf(" ...");
2288 			count = 0;
2289 		} else
2290 			db_printf(",");
2291 		count++;
2292 
2293 		db_printf("(off=0x%jx,page=0x%jx)",
2294 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2295 	}
2296 	if (count != 0)
2297 		db_printf("\n");
2298 	db_indent -= 2;
2299 }
2300 
2301 /* XXX. */
2302 #undef count
2303 
2304 /* XXX need this non-static entry for calling from vm_map_print. */
2305 void
2306 vm_object_print(
2307         /* db_expr_t */ long addr,
2308 	boolean_t have_addr,
2309 	/* db_expr_t */ long count,
2310 	char *modif)
2311 {
2312 	vm_object_print_static(addr, have_addr, count, modif);
2313 }
2314 
2315 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2316 {
2317 	vm_object_t object;
2318 	vm_pindex_t fidx;
2319 	vm_paddr_t pa;
2320 	vm_page_t m, prev_m;
2321 	int rcount, nl, c;
2322 
2323 	nl = 0;
2324 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2325 		db_printf("new object: %p\n", (void *)object);
2326 		if (nl > 18) {
2327 			c = cngetc();
2328 			if (c != ' ')
2329 				return;
2330 			nl = 0;
2331 		}
2332 		nl++;
2333 		rcount = 0;
2334 		fidx = 0;
2335 		pa = -1;
2336 		TAILQ_FOREACH(m, &object->memq, listq) {
2337 			if (m->pindex > 128)
2338 				break;
2339 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2340 			    prev_m->pindex + 1 != m->pindex) {
2341 				if (rcount) {
2342 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2343 						(long)fidx, rcount, (long)pa);
2344 					if (nl > 18) {
2345 						c = cngetc();
2346 						if (c != ' ')
2347 							return;
2348 						nl = 0;
2349 					}
2350 					nl++;
2351 					rcount = 0;
2352 				}
2353 			}
2354 			if (rcount &&
2355 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2356 				++rcount;
2357 				continue;
2358 			}
2359 			if (rcount) {
2360 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2361 					(long)fidx, rcount, (long)pa);
2362 				if (nl > 18) {
2363 					c = cngetc();
2364 					if (c != ' ')
2365 						return;
2366 					nl = 0;
2367 				}
2368 				nl++;
2369 			}
2370 			fidx = m->pindex;
2371 			pa = VM_PAGE_TO_PHYS(m);
2372 			rcount = 1;
2373 		}
2374 		if (rcount) {
2375 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2376 				(long)fidx, rcount, (long)pa);
2377 			if (nl > 18) {
2378 				c = cngetc();
2379 				if (c != ' ')
2380 					return;
2381 				nl = 0;
2382 			}
2383 			nl++;
2384 		}
2385 	}
2386 }
2387 #endif /* DDB */
2388