xref: /freebsd/sys/vm/vm_object.c (revision 830940567b49bb0c08dfaed40418999e76616909)
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->type != OBJT_VNODE ||
777 		(object->flags & OBJ_MIGHTBEDIRTY) == 0)
778 		return;
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 		struct vnode *vp;
880 
881 		vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
882 		if (object->type == OBJT_VNODE &&
883 		    (vp = (struct vnode *)object->handle) != NULL) {
884 			VI_LOCK(vp);
885 			if (vp->v_iflag & VI_OBJDIRTY)
886 				vp->v_iflag &= ~VI_OBJDIRTY;
887 			VI_UNLOCK(vp);
888 		}
889 	}
890 
891 rescan:
892 	curgeneration = object->generation;
893 
894 	for (p = TAILQ_FIRST(&object->memq); p; p = np) {
895 		int n;
896 
897 		np = TAILQ_NEXT(p, listq);
898 
899 again:
900 		pi = p->pindex;
901 		if ((p->oflags & VPO_CLEANCHK) == 0 ||
902 			(pi < tstart) || (pi >= tend) ||
903 		    p->valid == 0) {
904 			p->oflags &= ~VPO_CLEANCHK;
905 			continue;
906 		}
907 
908 		vm_page_test_dirty(p);
909 		if (p->dirty == 0) {
910 			p->oflags &= ~VPO_CLEANCHK;
911 			continue;
912 		}
913 
914 		/*
915 		 * If we have been asked to skip nosync pages and this is a
916 		 * nosync page, skip it.  Note that the object flags were
917 		 * not cleared in this case so we do not have to set them.
918 		 */
919 		if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
920 			p->oflags &= ~VPO_CLEANCHK;
921 			continue;
922 		}
923 
924 		n = vm_object_page_collect_flush(object, p,
925 			curgeneration, pagerflags);
926 		if (n == 0)
927 			goto rescan;
928 
929 		if (object->generation != curgeneration)
930 			goto rescan;
931 
932 		/*
933 		 * Try to optimize the next page.  If we can't we pick up
934 		 * our (random) scan where we left off.
935 		 */
936 		if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
937 			if ((p = vm_page_lookup(object, pi + n)) != NULL)
938 				goto again;
939 		}
940 	}
941 	vm_page_unlock_queues();
942 #if 0
943 	VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
944 #endif
945 
946 	vm_object_clear_flag(object, OBJ_CLEANING);
947 	return;
948 }
949 
950 static int
951 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
952 {
953 	int runlen;
954 	int maxf;
955 	int chkb;
956 	int maxb;
957 	int i;
958 	vm_pindex_t pi;
959 	vm_page_t maf[vm_pageout_page_count];
960 	vm_page_t mab[vm_pageout_page_count];
961 	vm_page_t ma[vm_pageout_page_count];
962 
963 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
964 	pi = p->pindex;
965 	while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
966 		vm_page_lock_queues();
967 		if (object->generation != curgeneration) {
968 			return(0);
969 		}
970 	}
971 	maxf = 0;
972 	for(i = 1; i < vm_pageout_page_count; i++) {
973 		vm_page_t tp;
974 
975 		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
976 			if ((tp->oflags & VPO_BUSY) ||
977 				((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
978 				 (tp->oflags & VPO_CLEANCHK) == 0) ||
979 				(tp->busy != 0))
980 				break;
981 			vm_page_test_dirty(tp);
982 			if (tp->dirty == 0) {
983 				tp->oflags &= ~VPO_CLEANCHK;
984 				break;
985 			}
986 			maf[ i - 1 ] = tp;
987 			maxf++;
988 			continue;
989 		}
990 		break;
991 	}
992 
993 	maxb = 0;
994 	chkb = vm_pageout_page_count -  maxf;
995 	if (chkb) {
996 		for(i = 1; i < chkb;i++) {
997 			vm_page_t tp;
998 
999 			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
1000 				if ((tp->oflags & VPO_BUSY) ||
1001 					((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1002 					 (tp->oflags & VPO_CLEANCHK) == 0) ||
1003 					(tp->busy != 0))
1004 					break;
1005 				vm_page_test_dirty(tp);
1006 				if (tp->dirty == 0) {
1007 					tp->oflags &= ~VPO_CLEANCHK;
1008 					break;
1009 				}
1010 				mab[ i - 1 ] = tp;
1011 				maxb++;
1012 				continue;
1013 			}
1014 			break;
1015 		}
1016 	}
1017 
1018 	for(i = 0; i < maxb; i++) {
1019 		int index = (maxb - i) - 1;
1020 		ma[index] = mab[i];
1021 		ma[index]->oflags &= ~VPO_CLEANCHK;
1022 	}
1023 	p->oflags &= ~VPO_CLEANCHK;
1024 	ma[maxb] = p;
1025 	for(i = 0; i < maxf; i++) {
1026 		int index = (maxb + i) + 1;
1027 		ma[index] = maf[i];
1028 		ma[index]->oflags &= ~VPO_CLEANCHK;
1029 	}
1030 	runlen = maxb + maxf + 1;
1031 
1032 	vm_pageout_flush(ma, runlen, pagerflags);
1033 	for (i = 0; i < runlen; i++) {
1034 		if (ma[i]->dirty) {
1035 			pmap_remove_write(ma[i]);
1036 			ma[i]->oflags |= VPO_CLEANCHK;
1037 
1038 			/*
1039 			 * maxf will end up being the actual number of pages
1040 			 * we wrote out contiguously, non-inclusive of the
1041 			 * first page.  We do not count look-behind pages.
1042 			 */
1043 			if (i >= maxb + 1 && (maxf > i - maxb - 1))
1044 				maxf = i - maxb - 1;
1045 		}
1046 	}
1047 	return(maxf + 1);
1048 }
1049 
1050 /*
1051  * Note that there is absolutely no sense in writing out
1052  * anonymous objects, so we track down the vnode object
1053  * to write out.
1054  * We invalidate (remove) all pages from the address space
1055  * for semantic correctness.
1056  *
1057  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1058  * may start out with a NULL object.
1059  */
1060 void
1061 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1062     boolean_t syncio, boolean_t invalidate)
1063 {
1064 	vm_object_t backing_object;
1065 	struct vnode *vp;
1066 	struct mount *mp;
1067 	int flags;
1068 
1069 	if (object == NULL)
1070 		return;
1071 	VM_OBJECT_LOCK(object);
1072 	while ((backing_object = object->backing_object) != NULL) {
1073 		VM_OBJECT_LOCK(backing_object);
1074 		offset += object->backing_object_offset;
1075 		VM_OBJECT_UNLOCK(object);
1076 		object = backing_object;
1077 		if (object->size < OFF_TO_IDX(offset + size))
1078 			size = IDX_TO_OFF(object->size) - offset;
1079 	}
1080 	/*
1081 	 * Flush pages if writing is allowed, invalidate them
1082 	 * if invalidation requested.  Pages undergoing I/O
1083 	 * will be ignored by vm_object_page_remove().
1084 	 *
1085 	 * We cannot lock the vnode and then wait for paging
1086 	 * to complete without deadlocking against vm_fault.
1087 	 * Instead we simply call vm_object_page_remove() and
1088 	 * allow it to block internally on a page-by-page
1089 	 * basis when it encounters pages undergoing async
1090 	 * I/O.
1091 	 */
1092 	if (object->type == OBJT_VNODE &&
1093 	    (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1094 		int vfslocked;
1095 		vp = object->handle;
1096 		VM_OBJECT_UNLOCK(object);
1097 		(void) vn_start_write(vp, &mp, V_WAIT);
1098 		vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1099 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1100 		flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1101 		flags |= invalidate ? OBJPC_INVAL : 0;
1102 		VM_OBJECT_LOCK(object);
1103 		vm_object_page_clean(object,
1104 		    OFF_TO_IDX(offset),
1105 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1106 		    flags);
1107 		VM_OBJECT_UNLOCK(object);
1108 		VOP_UNLOCK(vp, 0);
1109 		VFS_UNLOCK_GIANT(vfslocked);
1110 		vn_finished_write(mp);
1111 		VM_OBJECT_LOCK(object);
1112 	}
1113 	if ((object->type == OBJT_VNODE ||
1114 	     object->type == OBJT_DEVICE) && invalidate) {
1115 		boolean_t purge;
1116 		purge = old_msync || (object->type == OBJT_DEVICE);
1117 		vm_object_page_remove(object,
1118 		    OFF_TO_IDX(offset),
1119 		    OFF_TO_IDX(offset + size + PAGE_MASK),
1120 		    purge ? FALSE : TRUE);
1121 	}
1122 	VM_OBJECT_UNLOCK(object);
1123 }
1124 
1125 /*
1126  *	vm_object_madvise:
1127  *
1128  *	Implements the madvise function at the object/page level.
1129  *
1130  *	MADV_WILLNEED	(any object)
1131  *
1132  *	    Activate the specified pages if they are resident.
1133  *
1134  *	MADV_DONTNEED	(any object)
1135  *
1136  *	    Deactivate the specified pages if they are resident.
1137  *
1138  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
1139  *			 OBJ_ONEMAPPING only)
1140  *
1141  *	    Deactivate and clean the specified pages if they are
1142  *	    resident.  This permits the process to reuse the pages
1143  *	    without faulting or the kernel to reclaim the pages
1144  *	    without I/O.
1145  */
1146 void
1147 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1148 {
1149 	vm_pindex_t end, tpindex;
1150 	vm_object_t backing_object, tobject;
1151 	vm_page_t m;
1152 
1153 	if (object == NULL)
1154 		return;
1155 	VM_OBJECT_LOCK(object);
1156 	end = pindex + count;
1157 	/*
1158 	 * Locate and adjust resident pages
1159 	 */
1160 	for (; pindex < end; pindex += 1) {
1161 relookup:
1162 		tobject = object;
1163 		tpindex = pindex;
1164 shadowlookup:
1165 		/*
1166 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1167 		 * and those pages must be OBJ_ONEMAPPING.
1168 		 */
1169 		if (advise == MADV_FREE) {
1170 			if ((tobject->type != OBJT_DEFAULT &&
1171 			     tobject->type != OBJT_SWAP) ||
1172 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
1173 				goto unlock_tobject;
1174 			}
1175 		}
1176 		m = vm_page_lookup(tobject, tpindex);
1177 		if (m == NULL && advise == MADV_WILLNEED) {
1178 			/*
1179 			 * If the page is cached, reactivate it.
1180 			 */
1181 			m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1182 			    VM_ALLOC_NOBUSY);
1183 		}
1184 		if (m == NULL) {
1185 			/*
1186 			 * There may be swap even if there is no backing page
1187 			 */
1188 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1189 				swap_pager_freespace(tobject, tpindex, 1);
1190 			/*
1191 			 * next object
1192 			 */
1193 			backing_object = tobject->backing_object;
1194 			if (backing_object == NULL)
1195 				goto unlock_tobject;
1196 			VM_OBJECT_LOCK(backing_object);
1197 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1198 			if (tobject != object)
1199 				VM_OBJECT_UNLOCK(tobject);
1200 			tobject = backing_object;
1201 			goto shadowlookup;
1202 		}
1203 		/*
1204 		 * If the page is busy or not in a normal active state,
1205 		 * we skip it.  If the page is not managed there are no
1206 		 * page queues to mess with.  Things can break if we mess
1207 		 * with pages in any of the below states.
1208 		 */
1209 		vm_page_lock_queues();
1210 		if (m->hold_count ||
1211 		    m->wire_count ||
1212 		    (m->flags & PG_UNMANAGED) ||
1213 		    m->valid != VM_PAGE_BITS_ALL) {
1214 			vm_page_unlock_queues();
1215 			goto unlock_tobject;
1216 		}
1217 		if ((m->oflags & VPO_BUSY) || m->busy) {
1218 			vm_page_flag_set(m, PG_REFERENCED);
1219 			vm_page_unlock_queues();
1220 			if (object != tobject)
1221 				VM_OBJECT_UNLOCK(object);
1222 			m->oflags |= VPO_WANTED;
1223 			msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo", 0);
1224 			VM_OBJECT_LOCK(object);
1225   			goto relookup;
1226 		}
1227 		if (advise == MADV_WILLNEED) {
1228 			vm_page_activate(m);
1229 		} else if (advise == MADV_DONTNEED) {
1230 			vm_page_dontneed(m);
1231 		} else if (advise == MADV_FREE) {
1232 			/*
1233 			 * Mark the page clean.  This will allow the page
1234 			 * to be freed up by the system.  However, such pages
1235 			 * are often reused quickly by malloc()/free()
1236 			 * so we do not do anything that would cause
1237 			 * a page fault if we can help it.
1238 			 *
1239 			 * Specifically, we do not try to actually free
1240 			 * the page now nor do we try to put it in the
1241 			 * cache (which would cause a page fault on reuse).
1242 			 *
1243 			 * But we do make the page is freeable as we
1244 			 * can without actually taking the step of unmapping
1245 			 * it.
1246 			 */
1247 			pmap_clear_modify(m);
1248 			m->dirty = 0;
1249 			m->act_count = 0;
1250 			vm_page_dontneed(m);
1251 		}
1252 		vm_page_unlock_queues();
1253 		if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1254 			swap_pager_freespace(tobject, tpindex, 1);
1255 unlock_tobject:
1256 		if (tobject != object)
1257 			VM_OBJECT_UNLOCK(tobject);
1258 	}
1259 	VM_OBJECT_UNLOCK(object);
1260 }
1261 
1262 /*
1263  *	vm_object_shadow:
1264  *
1265  *	Create a new object which is backed by the
1266  *	specified existing object range.  The source
1267  *	object reference is deallocated.
1268  *
1269  *	The new object and offset into that object
1270  *	are returned in the source parameters.
1271  */
1272 void
1273 vm_object_shadow(
1274 	vm_object_t *object,	/* IN/OUT */
1275 	vm_ooffset_t *offset,	/* IN/OUT */
1276 	vm_size_t length)
1277 {
1278 	vm_object_t source;
1279 	vm_object_t result;
1280 
1281 	source = *object;
1282 
1283 	/*
1284 	 * Don't create the new object if the old object isn't shared.
1285 	 */
1286 	if (source != NULL) {
1287 		VM_OBJECT_LOCK(source);
1288 		if (source->ref_count == 1 &&
1289 		    source->handle == NULL &&
1290 		    (source->type == OBJT_DEFAULT ||
1291 		     source->type == OBJT_SWAP)) {
1292 			VM_OBJECT_UNLOCK(source);
1293 			return;
1294 		}
1295 		VM_OBJECT_UNLOCK(source);
1296 	}
1297 
1298 	/*
1299 	 * Allocate a new object with the given length.
1300 	 */
1301 	result = vm_object_allocate(OBJT_DEFAULT, length);
1302 
1303 	/*
1304 	 * The new object shadows the source object, adding a reference to it.
1305 	 * Our caller changes his reference to point to the new object,
1306 	 * removing a reference to the source object.  Net result: no change
1307 	 * of reference count.
1308 	 *
1309 	 * Try to optimize the result object's page color when shadowing
1310 	 * in order to maintain page coloring consistency in the combined
1311 	 * shadowed object.
1312 	 */
1313 	result->backing_object = source;
1314 	/*
1315 	 * Store the offset into the source object, and fix up the offset into
1316 	 * the new object.
1317 	 */
1318 	result->backing_object_offset = *offset;
1319 	if (source != NULL) {
1320 		VM_OBJECT_LOCK(source);
1321 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1322 		source->shadow_count++;
1323 		source->generation++;
1324 #if VM_NRESERVLEVEL > 0
1325 		result->flags |= source->flags & OBJ_COLORED;
1326 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1327 		    ((1 << (VM_NFREEORDER - 1)) - 1);
1328 #endif
1329 		VM_OBJECT_UNLOCK(source);
1330 	}
1331 
1332 
1333 	/*
1334 	 * Return the new things
1335 	 */
1336 	*offset = 0;
1337 	*object = result;
1338 }
1339 
1340 /*
1341  *	vm_object_split:
1342  *
1343  * Split the pages in a map entry into a new object.  This affords
1344  * easier removal of unused pages, and keeps object inheritance from
1345  * being a negative impact on memory usage.
1346  */
1347 void
1348 vm_object_split(vm_map_entry_t entry)
1349 {
1350 	vm_page_t m, m_next;
1351 	vm_object_t orig_object, new_object, source;
1352 	vm_pindex_t idx, offidxstart;
1353 	vm_size_t size;
1354 
1355 	orig_object = entry->object.vm_object;
1356 	if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1357 		return;
1358 	if (orig_object->ref_count <= 1)
1359 		return;
1360 	VM_OBJECT_UNLOCK(orig_object);
1361 
1362 	offidxstart = OFF_TO_IDX(entry->offset);
1363 	size = atop(entry->end - entry->start);
1364 
1365 	/*
1366 	 * If swap_pager_copy() is later called, it will convert new_object
1367 	 * into a swap object.
1368 	 */
1369 	new_object = vm_object_allocate(OBJT_DEFAULT, size);
1370 
1371 	/*
1372 	 * At this point, the new object is still private, so the order in
1373 	 * which the original and new objects are locked does not matter.
1374 	 */
1375 	VM_OBJECT_LOCK(new_object);
1376 	VM_OBJECT_LOCK(orig_object);
1377 	source = orig_object->backing_object;
1378 	if (source != NULL) {
1379 		VM_OBJECT_LOCK(source);
1380 		if ((source->flags & OBJ_DEAD) != 0) {
1381 			VM_OBJECT_UNLOCK(source);
1382 			VM_OBJECT_UNLOCK(orig_object);
1383 			VM_OBJECT_UNLOCK(new_object);
1384 			vm_object_deallocate(new_object);
1385 			VM_OBJECT_LOCK(orig_object);
1386 			return;
1387 		}
1388 		LIST_INSERT_HEAD(&source->shadow_head,
1389 				  new_object, shadow_list);
1390 		source->shadow_count++;
1391 		source->generation++;
1392 		vm_object_reference_locked(source);	/* for new_object */
1393 		vm_object_clear_flag(source, OBJ_ONEMAPPING);
1394 		VM_OBJECT_UNLOCK(source);
1395 		new_object->backing_object_offset =
1396 			orig_object->backing_object_offset + entry->offset;
1397 		new_object->backing_object = source;
1398 	}
1399 	if (orig_object->uip != NULL) {
1400 		new_object->uip = orig_object->uip;
1401 		uihold(orig_object->uip);
1402 		new_object->charge = ptoa(size);
1403 		KASSERT(orig_object->charge >= ptoa(size),
1404 		    ("orig_object->charge < 0"));
1405 		orig_object->charge -= ptoa(size);
1406 	}
1407 retry:
1408 	if ((m = TAILQ_FIRST(&orig_object->memq)) != NULL) {
1409 		if (m->pindex < offidxstart) {
1410 			m = vm_page_splay(offidxstart, orig_object->root);
1411 			if ((orig_object->root = m)->pindex < offidxstart)
1412 				m = TAILQ_NEXT(m, listq);
1413 		}
1414 	}
1415 	vm_page_lock_queues();
1416 	for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1417 	    m = m_next) {
1418 		m_next = TAILQ_NEXT(m, listq);
1419 
1420 		/*
1421 		 * We must wait for pending I/O to complete before we can
1422 		 * rename the page.
1423 		 *
1424 		 * We do not have to VM_PROT_NONE the page as mappings should
1425 		 * not be changed by this operation.
1426 		 */
1427 		if ((m->oflags & VPO_BUSY) || m->busy) {
1428 			vm_page_flag_set(m, PG_REFERENCED);
1429 			vm_page_unlock_queues();
1430 			VM_OBJECT_UNLOCK(new_object);
1431 			m->oflags |= VPO_WANTED;
1432 			msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1433 			VM_OBJECT_LOCK(new_object);
1434 			goto retry;
1435 		}
1436 		vm_page_rename(m, new_object, idx);
1437 		/* page automatically made dirty by rename and cache handled */
1438 		vm_page_busy(m);
1439 	}
1440 	vm_page_unlock_queues();
1441 	if (orig_object->type == OBJT_SWAP) {
1442 		/*
1443 		 * swap_pager_copy() can sleep, in which case the orig_object's
1444 		 * and new_object's locks are released and reacquired.
1445 		 */
1446 		swap_pager_copy(orig_object, new_object, offidxstart, 0);
1447 
1448 		/*
1449 		 * Transfer any cached pages from orig_object to new_object.
1450 		 */
1451 		if (__predict_false(orig_object->cache != NULL))
1452 			vm_page_cache_transfer(orig_object, offidxstart,
1453 			    new_object);
1454 	}
1455 	VM_OBJECT_UNLOCK(orig_object);
1456 	TAILQ_FOREACH(m, &new_object->memq, listq)
1457 		vm_page_wakeup(m);
1458 	VM_OBJECT_UNLOCK(new_object);
1459 	entry->object.vm_object = new_object;
1460 	entry->offset = 0LL;
1461 	vm_object_deallocate(orig_object);
1462 	VM_OBJECT_LOCK(new_object);
1463 }
1464 
1465 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1466 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1467 #define	OBSC_COLLAPSE_WAIT	0x0004
1468 
1469 static int
1470 vm_object_backing_scan(vm_object_t object, int op)
1471 {
1472 	int r = 1;
1473 	vm_page_t p;
1474 	vm_object_t backing_object;
1475 	vm_pindex_t backing_offset_index;
1476 
1477 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1478 	VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1479 
1480 	backing_object = object->backing_object;
1481 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1482 
1483 	/*
1484 	 * Initial conditions
1485 	 */
1486 	if (op & OBSC_TEST_ALL_SHADOWED) {
1487 		/*
1488 		 * We do not want to have to test for the existence of cache
1489 		 * or swap pages in the backing object.  XXX but with the
1490 		 * new swapper this would be pretty easy to do.
1491 		 *
1492 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1493 		 * been ZFOD faulted yet?  If we do not test for this, the
1494 		 * shadow test may succeed! XXX
1495 		 */
1496 		if (backing_object->type != OBJT_DEFAULT) {
1497 			return (0);
1498 		}
1499 	}
1500 	if (op & OBSC_COLLAPSE_WAIT) {
1501 		vm_object_set_flag(backing_object, OBJ_DEAD);
1502 	}
1503 
1504 	/*
1505 	 * Our scan
1506 	 */
1507 	p = TAILQ_FIRST(&backing_object->memq);
1508 	while (p) {
1509 		vm_page_t next = TAILQ_NEXT(p, listq);
1510 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1511 
1512 		if (op & OBSC_TEST_ALL_SHADOWED) {
1513 			vm_page_t pp;
1514 
1515 			/*
1516 			 * Ignore pages outside the parent object's range
1517 			 * and outside the parent object's mapping of the
1518 			 * backing object.
1519 			 *
1520 			 * note that we do not busy the backing object's
1521 			 * page.
1522 			 */
1523 			if (
1524 			    p->pindex < backing_offset_index ||
1525 			    new_pindex >= object->size
1526 			) {
1527 				p = next;
1528 				continue;
1529 			}
1530 
1531 			/*
1532 			 * See if the parent has the page or if the parent's
1533 			 * object pager has the page.  If the parent has the
1534 			 * page but the page is not valid, the parent's
1535 			 * object pager must have the page.
1536 			 *
1537 			 * If this fails, the parent does not completely shadow
1538 			 * the object and we might as well give up now.
1539 			 */
1540 
1541 			pp = vm_page_lookup(object, new_pindex);
1542 			if (
1543 			    (pp == NULL || pp->valid == 0) &&
1544 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1545 			) {
1546 				r = 0;
1547 				break;
1548 			}
1549 		}
1550 
1551 		/*
1552 		 * Check for busy page
1553 		 */
1554 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1555 			vm_page_t pp;
1556 
1557 			if (op & OBSC_COLLAPSE_NOWAIT) {
1558 				if ((p->oflags & VPO_BUSY) ||
1559 				    !p->valid ||
1560 				    p->busy) {
1561 					p = next;
1562 					continue;
1563 				}
1564 			} else if (op & OBSC_COLLAPSE_WAIT) {
1565 				if ((p->oflags & VPO_BUSY) || p->busy) {
1566 					vm_page_lock_queues();
1567 					vm_page_flag_set(p, PG_REFERENCED);
1568 					vm_page_unlock_queues();
1569 					VM_OBJECT_UNLOCK(object);
1570 					p->oflags |= VPO_WANTED;
1571 					msleep(p, VM_OBJECT_MTX(backing_object),
1572 					    PDROP | PVM, "vmocol", 0);
1573 					VM_OBJECT_LOCK(object);
1574 					VM_OBJECT_LOCK(backing_object);
1575 					/*
1576 					 * If we slept, anything could have
1577 					 * happened.  Since the object is
1578 					 * marked dead, the backing offset
1579 					 * should not have changed so we
1580 					 * just restart our scan.
1581 					 */
1582 					p = TAILQ_FIRST(&backing_object->memq);
1583 					continue;
1584 				}
1585 			}
1586 
1587 			KASSERT(
1588 			    p->object == backing_object,
1589 			    ("vm_object_backing_scan: object mismatch")
1590 			);
1591 
1592 			/*
1593 			 * Destroy any associated swap
1594 			 */
1595 			if (backing_object->type == OBJT_SWAP) {
1596 				swap_pager_freespace(
1597 				    backing_object,
1598 				    p->pindex,
1599 				    1
1600 				);
1601 			}
1602 
1603 			if (
1604 			    p->pindex < backing_offset_index ||
1605 			    new_pindex >= object->size
1606 			) {
1607 				/*
1608 				 * Page is out of the parent object's range, we
1609 				 * can simply destroy it.
1610 				 */
1611 				vm_page_lock_queues();
1612 				KASSERT(!pmap_page_is_mapped(p),
1613 				    ("freeing mapped page %p", p));
1614 				if (p->wire_count == 0)
1615 					vm_page_free(p);
1616 				else
1617 					vm_page_remove(p);
1618 				vm_page_unlock_queues();
1619 				p = next;
1620 				continue;
1621 			}
1622 
1623 			pp = vm_page_lookup(object, new_pindex);
1624 			if (
1625 			    pp != NULL ||
1626 			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1627 			) {
1628 				/*
1629 				 * page already exists in parent OR swap exists
1630 				 * for this location in the parent.  Destroy
1631 				 * the original page from the backing object.
1632 				 *
1633 				 * Leave the parent's page alone
1634 				 */
1635 				vm_page_lock_queues();
1636 				KASSERT(!pmap_page_is_mapped(p),
1637 				    ("freeing mapped page %p", p));
1638 				if (p->wire_count == 0)
1639 					vm_page_free(p);
1640 				else
1641 					vm_page_remove(p);
1642 				vm_page_unlock_queues();
1643 				p = next;
1644 				continue;
1645 			}
1646 
1647 #if VM_NRESERVLEVEL > 0
1648 			/*
1649 			 * Rename the reservation.
1650 			 */
1651 			vm_reserv_rename(p, object, backing_object,
1652 			    backing_offset_index);
1653 #endif
1654 
1655 			/*
1656 			 * Page does not exist in parent, rename the
1657 			 * page from the backing object to the main object.
1658 			 *
1659 			 * If the page was mapped to a process, it can remain
1660 			 * mapped through the rename.
1661 			 */
1662 			vm_page_lock_queues();
1663 			vm_page_rename(p, object, new_pindex);
1664 			vm_page_unlock_queues();
1665 			/* page automatically made dirty by rename */
1666 		}
1667 		p = next;
1668 	}
1669 	return (r);
1670 }
1671 
1672 
1673 /*
1674  * this version of collapse allows the operation to occur earlier and
1675  * when paging_in_progress is true for an object...  This is not a complete
1676  * operation, but should plug 99.9% of the rest of the leaks.
1677  */
1678 static void
1679 vm_object_qcollapse(vm_object_t object)
1680 {
1681 	vm_object_t backing_object = object->backing_object;
1682 
1683 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1684 	VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1685 
1686 	if (backing_object->ref_count != 1)
1687 		return;
1688 
1689 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1690 }
1691 
1692 /*
1693  *	vm_object_collapse:
1694  *
1695  *	Collapse an object with the object backing it.
1696  *	Pages in the backing object are moved into the
1697  *	parent, and the backing object is deallocated.
1698  */
1699 void
1700 vm_object_collapse(vm_object_t object)
1701 {
1702 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1703 
1704 	while (TRUE) {
1705 		vm_object_t backing_object;
1706 
1707 		/*
1708 		 * Verify that the conditions are right for collapse:
1709 		 *
1710 		 * The object exists and the backing object exists.
1711 		 */
1712 		if ((backing_object = object->backing_object) == NULL)
1713 			break;
1714 
1715 		/*
1716 		 * we check the backing object first, because it is most likely
1717 		 * not collapsable.
1718 		 */
1719 		VM_OBJECT_LOCK(backing_object);
1720 		if (backing_object->handle != NULL ||
1721 		    (backing_object->type != OBJT_DEFAULT &&
1722 		     backing_object->type != OBJT_SWAP) ||
1723 		    (backing_object->flags & OBJ_DEAD) ||
1724 		    object->handle != NULL ||
1725 		    (object->type != OBJT_DEFAULT &&
1726 		     object->type != OBJT_SWAP) ||
1727 		    (object->flags & OBJ_DEAD)) {
1728 			VM_OBJECT_UNLOCK(backing_object);
1729 			break;
1730 		}
1731 
1732 		if (
1733 		    object->paging_in_progress != 0 ||
1734 		    backing_object->paging_in_progress != 0
1735 		) {
1736 			vm_object_qcollapse(object);
1737 			VM_OBJECT_UNLOCK(backing_object);
1738 			break;
1739 		}
1740 		/*
1741 		 * We know that we can either collapse the backing object (if
1742 		 * the parent is the only reference to it) or (perhaps) have
1743 		 * the parent bypass the object if the parent happens to shadow
1744 		 * all the resident pages in the entire backing object.
1745 		 *
1746 		 * This is ignoring pager-backed pages such as swap pages.
1747 		 * vm_object_backing_scan fails the shadowing test in this
1748 		 * case.
1749 		 */
1750 		if (backing_object->ref_count == 1) {
1751 			/*
1752 			 * If there is exactly one reference to the backing
1753 			 * object, we can collapse it into the parent.
1754 			 */
1755 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1756 
1757 #if VM_NRESERVLEVEL > 0
1758 			/*
1759 			 * Break any reservations from backing_object.
1760 			 */
1761 			if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1762 				vm_reserv_break_all(backing_object);
1763 #endif
1764 
1765 			/*
1766 			 * Move the pager from backing_object to object.
1767 			 */
1768 			if (backing_object->type == OBJT_SWAP) {
1769 				/*
1770 				 * swap_pager_copy() can sleep, in which case
1771 				 * the backing_object's and object's locks are
1772 				 * released and reacquired.
1773 				 */
1774 				swap_pager_copy(
1775 				    backing_object,
1776 				    object,
1777 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1778 
1779 				/*
1780 				 * Free any cached pages from backing_object.
1781 				 */
1782 				if (__predict_false(backing_object->cache != NULL))
1783 					vm_page_cache_free(backing_object, 0, 0);
1784 			}
1785 			/*
1786 			 * Object now shadows whatever backing_object did.
1787 			 * Note that the reference to
1788 			 * backing_object->backing_object moves from within
1789 			 * backing_object to within object.
1790 			 */
1791 			LIST_REMOVE(object, shadow_list);
1792 			backing_object->shadow_count--;
1793 			backing_object->generation++;
1794 			if (backing_object->backing_object) {
1795 				VM_OBJECT_LOCK(backing_object->backing_object);
1796 				LIST_REMOVE(backing_object, shadow_list);
1797 				LIST_INSERT_HEAD(
1798 				    &backing_object->backing_object->shadow_head,
1799 				    object, shadow_list);
1800 				/*
1801 				 * The shadow_count has not changed.
1802 				 */
1803 				backing_object->backing_object->generation++;
1804 				VM_OBJECT_UNLOCK(backing_object->backing_object);
1805 			}
1806 			object->backing_object = backing_object->backing_object;
1807 			object->backing_object_offset +=
1808 			    backing_object->backing_object_offset;
1809 
1810 			/*
1811 			 * Discard backing_object.
1812 			 *
1813 			 * Since the backing object has no pages, no pager left,
1814 			 * and no object references within it, all that is
1815 			 * necessary is to dispose of it.
1816 			 */
1817 			KASSERT(backing_object->ref_count == 1, (
1818 "backing_object %p was somehow re-referenced during collapse!",
1819 			    backing_object));
1820 			VM_OBJECT_UNLOCK(backing_object);
1821 			vm_object_destroy(backing_object);
1822 
1823 			object_collapses++;
1824 		} else {
1825 			vm_object_t new_backing_object;
1826 
1827 			/*
1828 			 * If we do not entirely shadow the backing object,
1829 			 * there is nothing we can do so we give up.
1830 			 */
1831 			if (object->resident_page_count != object->size &&
1832 			    vm_object_backing_scan(object,
1833 			    OBSC_TEST_ALL_SHADOWED) == 0) {
1834 				VM_OBJECT_UNLOCK(backing_object);
1835 				break;
1836 			}
1837 
1838 			/*
1839 			 * Make the parent shadow the next object in the
1840 			 * chain.  Deallocating backing_object will not remove
1841 			 * it, since its reference count is at least 2.
1842 			 */
1843 			LIST_REMOVE(object, shadow_list);
1844 			backing_object->shadow_count--;
1845 			backing_object->generation++;
1846 
1847 			new_backing_object = backing_object->backing_object;
1848 			if ((object->backing_object = new_backing_object) != NULL) {
1849 				VM_OBJECT_LOCK(new_backing_object);
1850 				LIST_INSERT_HEAD(
1851 				    &new_backing_object->shadow_head,
1852 				    object,
1853 				    shadow_list
1854 				);
1855 				new_backing_object->shadow_count++;
1856 				new_backing_object->generation++;
1857 				vm_object_reference_locked(new_backing_object);
1858 				VM_OBJECT_UNLOCK(new_backing_object);
1859 				object->backing_object_offset +=
1860 					backing_object->backing_object_offset;
1861 			}
1862 
1863 			/*
1864 			 * Drop the reference count on backing_object. Since
1865 			 * its ref_count was at least 2, it will not vanish.
1866 			 */
1867 			backing_object->ref_count--;
1868 			VM_OBJECT_UNLOCK(backing_object);
1869 			object_bypasses++;
1870 		}
1871 
1872 		/*
1873 		 * Try again with this object's new backing object.
1874 		 */
1875 	}
1876 }
1877 
1878 /*
1879  *	vm_object_page_remove:
1880  *
1881  *	For the given object, either frees or invalidates each of the
1882  *	specified pages.  In general, a page is freed.  However, if a
1883  *	page is wired for any reason other than the existence of a
1884  *	managed, wired mapping, then it may be invalidated but not
1885  *	removed from the object.  Pages are specified by the given
1886  *	range ["start", "end") and Boolean "clean_only".  As a
1887  *	special case, if "end" is zero, then the range extends from
1888  *	"start" to the end of the object.  If "clean_only" is TRUE,
1889  *	then only the non-dirty pages within the specified range are
1890  *	affected.
1891  *
1892  *	In general, this operation should only be performed on objects
1893  *	that contain managed pages.  There are two exceptions.  First,
1894  *	it may be performed on the kernel and kmem objects.  Second,
1895  *	it may be used by msync(..., MS_INVALIDATE) to invalidate
1896  *	device-backed pages.
1897  *
1898  *	The object must be locked.
1899  */
1900 void
1901 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1902     boolean_t clean_only)
1903 {
1904 	vm_page_t p, next;
1905 	int wirings;
1906 
1907 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1908 	if (object->resident_page_count == 0)
1909 		goto skipmemq;
1910 
1911 	/*
1912 	 * Since physically-backed objects do not use managed pages, we can't
1913 	 * remove pages from the object (we must instead remove the page
1914 	 * references, and then destroy the object).
1915 	 */
1916 	KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1917 	    object == kmem_object,
1918 	    ("attempt to remove pages from a physical object"));
1919 
1920 	vm_object_pip_add(object, 1);
1921 again:
1922 	if ((p = TAILQ_FIRST(&object->memq)) != NULL) {
1923 		if (p->pindex < start) {
1924 			p = vm_page_splay(start, object->root);
1925 			if ((object->root = p)->pindex < start)
1926 				p = TAILQ_NEXT(p, listq);
1927 		}
1928 	}
1929 	vm_page_lock_queues();
1930 	/*
1931 	 * Assert: the variable p is either (1) the page with the
1932 	 * least pindex greater than or equal to the parameter pindex
1933 	 * or (2) NULL.
1934 	 */
1935 	for (;
1936 	     p != NULL && (p->pindex < end || end == 0);
1937 	     p = next) {
1938 		next = TAILQ_NEXT(p, listq);
1939 
1940 		/*
1941 		 * If the page is wired for any reason besides the
1942 		 * existence of managed, wired mappings, then it cannot
1943 		 * be freed.  For example, fictitious pages, which
1944 		 * represent device memory, are inherently wired and
1945 		 * cannot be freed.  They can, however, be invalidated
1946 		 * if "clean_only" is FALSE.
1947 		 */
1948 		if ((wirings = p->wire_count) != 0 &&
1949 		    (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1950 			/* Fictitious pages do not have managed mappings. */
1951 			if ((p->flags & PG_FICTITIOUS) == 0)
1952 				pmap_remove_all(p);
1953 			/* Account for removal of managed, wired mappings. */
1954 			p->wire_count -= wirings;
1955 			if (!clean_only) {
1956 				p->valid = 0;
1957 				vm_page_undirty(p);
1958 			}
1959 			continue;
1960 		}
1961 		if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1962 			goto again;
1963 		KASSERT((p->flags & PG_FICTITIOUS) == 0,
1964 		    ("vm_object_page_remove: page %p is fictitious", p));
1965 		if (clean_only && p->valid) {
1966 			pmap_remove_write(p);
1967 			if (p->dirty)
1968 				continue;
1969 		}
1970 		pmap_remove_all(p);
1971 		/* Account for removal of managed, wired mappings. */
1972 		if (wirings != 0)
1973 			p->wire_count -= wirings;
1974 		vm_page_free(p);
1975 	}
1976 	vm_page_unlock_queues();
1977 	vm_object_pip_wakeup(object);
1978 skipmemq:
1979 	if (__predict_false(object->cache != NULL))
1980 		vm_page_cache_free(object, start, end);
1981 }
1982 
1983 /*
1984  *	Populate the specified range of the object with valid pages.  Returns
1985  *	TRUE if the range is successfully populated and FALSE otherwise.
1986  *
1987  *	Note: This function should be optimized to pass a larger array of
1988  *	pages to vm_pager_get_pages() before it is applied to a non-
1989  *	OBJT_DEVICE object.
1990  *
1991  *	The object must be locked.
1992  */
1993 boolean_t
1994 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1995 {
1996 	vm_page_t m, ma[1];
1997 	vm_pindex_t pindex;
1998 	int rv;
1999 
2000 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2001 	for (pindex = start; pindex < end; pindex++) {
2002 		m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
2003 		    VM_ALLOC_RETRY);
2004 		if (m->valid != VM_PAGE_BITS_ALL) {
2005 			ma[0] = m;
2006 			rv = vm_pager_get_pages(object, ma, 1, 0);
2007 			m = vm_page_lookup(object, pindex);
2008 			if (m == NULL)
2009 				break;
2010 			if (rv != VM_PAGER_OK) {
2011 				vm_page_lock_queues();
2012 				vm_page_free(m);
2013 				vm_page_unlock_queues();
2014 				break;
2015 			}
2016 		}
2017 		/*
2018 		 * Keep "m" busy because a subsequent iteration may unlock
2019 		 * the object.
2020 		 */
2021 	}
2022 	if (pindex > start) {
2023 		m = vm_page_lookup(object, start);
2024 		while (m != NULL && m->pindex < pindex) {
2025 			vm_page_wakeup(m);
2026 			m = TAILQ_NEXT(m, listq);
2027 		}
2028 	}
2029 	return (pindex == end);
2030 }
2031 
2032 /*
2033  *	Routine:	vm_object_coalesce
2034  *	Function:	Coalesces two objects backing up adjoining
2035  *			regions of memory into a single object.
2036  *
2037  *	returns TRUE if objects were combined.
2038  *
2039  *	NOTE:	Only works at the moment if the second object is NULL -
2040  *		if it's not, which object do we lock first?
2041  *
2042  *	Parameters:
2043  *		prev_object	First object to coalesce
2044  *		prev_offset	Offset into prev_object
2045  *		prev_size	Size of reference to prev_object
2046  *		next_size	Size of reference to the second object
2047  *		reserved	Indicator that extension region has
2048  *				swap accounted for
2049  *
2050  *	Conditions:
2051  *	The object must *not* be locked.
2052  */
2053 boolean_t
2054 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2055     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2056 {
2057 	vm_pindex_t next_pindex;
2058 
2059 	if (prev_object == NULL)
2060 		return (TRUE);
2061 	VM_OBJECT_LOCK(prev_object);
2062 	if (prev_object->type != OBJT_DEFAULT &&
2063 	    prev_object->type != OBJT_SWAP) {
2064 		VM_OBJECT_UNLOCK(prev_object);
2065 		return (FALSE);
2066 	}
2067 
2068 	/*
2069 	 * Try to collapse the object first
2070 	 */
2071 	vm_object_collapse(prev_object);
2072 
2073 	/*
2074 	 * Can't coalesce if: . more than one reference . paged out . shadows
2075 	 * another object . has a copy elsewhere (any of which mean that the
2076 	 * pages not mapped to prev_entry may be in use anyway)
2077 	 */
2078 	if (prev_object->backing_object != NULL) {
2079 		VM_OBJECT_UNLOCK(prev_object);
2080 		return (FALSE);
2081 	}
2082 
2083 	prev_size >>= PAGE_SHIFT;
2084 	next_size >>= PAGE_SHIFT;
2085 	next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2086 
2087 	if ((prev_object->ref_count > 1) &&
2088 	    (prev_object->size != next_pindex)) {
2089 		VM_OBJECT_UNLOCK(prev_object);
2090 		return (FALSE);
2091 	}
2092 
2093 	/*
2094 	 * Account for the charge.
2095 	 */
2096 	if (prev_object->uip != NULL) {
2097 
2098 		/*
2099 		 * If prev_object was charged, then this mapping,
2100 		 * althought not charged now, may become writable
2101 		 * later. Non-NULL uip in the object would prevent
2102 		 * swap reservation during enabling of the write
2103 		 * access, so reserve swap now. Failed reservation
2104 		 * cause allocation of the separate object for the map
2105 		 * entry, and swap reservation for this entry is
2106 		 * managed in appropriate time.
2107 		 */
2108 		if (!reserved && !swap_reserve_by_uid(ptoa(next_size),
2109 		    prev_object->uip)) {
2110 			return (FALSE);
2111 		}
2112 		prev_object->charge += ptoa(next_size);
2113 	}
2114 
2115 	/*
2116 	 * Remove any pages that may still be in the object from a previous
2117 	 * deallocation.
2118 	 */
2119 	if (next_pindex < prev_object->size) {
2120 		vm_object_page_remove(prev_object,
2121 				      next_pindex,
2122 				      next_pindex + next_size, FALSE);
2123 		if (prev_object->type == OBJT_SWAP)
2124 			swap_pager_freespace(prev_object,
2125 					     next_pindex, next_size);
2126 #if 0
2127 		if (prev_object->uip != NULL) {
2128 			KASSERT(prev_object->charge >=
2129 			    ptoa(prev_object->size - next_pindex),
2130 			    ("object %p overcharged 1 %jx %jx", prev_object,
2131 				(uintmax_t)next_pindex, (uintmax_t)next_size));
2132 			prev_object->charge -= ptoa(prev_object->size -
2133 			    next_pindex);
2134 		}
2135 #endif
2136 	}
2137 
2138 	/*
2139 	 * Extend the object if necessary.
2140 	 */
2141 	if (next_pindex + next_size > prev_object->size)
2142 		prev_object->size = next_pindex + next_size;
2143 
2144 	VM_OBJECT_UNLOCK(prev_object);
2145 	return (TRUE);
2146 }
2147 
2148 void
2149 vm_object_set_writeable_dirty(vm_object_t object)
2150 {
2151 	struct vnode *vp;
2152 
2153 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2154 	if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2155 		return;
2156 	vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2157 	if (object->type == OBJT_VNODE &&
2158 	    (vp = (struct vnode *)object->handle) != NULL) {
2159 		VI_LOCK(vp);
2160 		vp->v_iflag |= VI_OBJDIRTY;
2161 		VI_UNLOCK(vp);
2162 	}
2163 }
2164 
2165 #include "opt_ddb.h"
2166 #ifdef DDB
2167 #include <sys/kernel.h>
2168 
2169 #include <sys/cons.h>
2170 
2171 #include <ddb/ddb.h>
2172 
2173 static int
2174 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2175 {
2176 	vm_map_t tmpm;
2177 	vm_map_entry_t tmpe;
2178 	vm_object_t obj;
2179 	int entcount;
2180 
2181 	if (map == 0)
2182 		return 0;
2183 
2184 	if (entry == 0) {
2185 		tmpe = map->header.next;
2186 		entcount = map->nentries;
2187 		while (entcount-- && (tmpe != &map->header)) {
2188 			if (_vm_object_in_map(map, object, tmpe)) {
2189 				return 1;
2190 			}
2191 			tmpe = tmpe->next;
2192 		}
2193 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2194 		tmpm = entry->object.sub_map;
2195 		tmpe = tmpm->header.next;
2196 		entcount = tmpm->nentries;
2197 		while (entcount-- && tmpe != &tmpm->header) {
2198 			if (_vm_object_in_map(tmpm, object, tmpe)) {
2199 				return 1;
2200 			}
2201 			tmpe = tmpe->next;
2202 		}
2203 	} else if ((obj = entry->object.vm_object) != NULL) {
2204 		for (; obj; obj = obj->backing_object)
2205 			if (obj == object) {
2206 				return 1;
2207 			}
2208 	}
2209 	return 0;
2210 }
2211 
2212 static int
2213 vm_object_in_map(vm_object_t object)
2214 {
2215 	struct proc *p;
2216 
2217 	/* sx_slock(&allproc_lock); */
2218 	FOREACH_PROC_IN_SYSTEM(p) {
2219 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2220 			continue;
2221 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2222 			/* sx_sunlock(&allproc_lock); */
2223 			return 1;
2224 		}
2225 	}
2226 	/* sx_sunlock(&allproc_lock); */
2227 	if (_vm_object_in_map(kernel_map, object, 0))
2228 		return 1;
2229 	if (_vm_object_in_map(kmem_map, object, 0))
2230 		return 1;
2231 	if (_vm_object_in_map(pager_map, object, 0))
2232 		return 1;
2233 	if (_vm_object_in_map(buffer_map, object, 0))
2234 		return 1;
2235 	return 0;
2236 }
2237 
2238 DB_SHOW_COMMAND(vmochk, vm_object_check)
2239 {
2240 	vm_object_t object;
2241 
2242 	/*
2243 	 * make sure that internal objs are in a map somewhere
2244 	 * and none have zero ref counts.
2245 	 */
2246 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2247 		if (object->handle == NULL &&
2248 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2249 			if (object->ref_count == 0) {
2250 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
2251 					(long)object->size);
2252 			}
2253 			if (!vm_object_in_map(object)) {
2254 				db_printf(
2255 			"vmochk: internal obj is not in a map: "
2256 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2257 				    object->ref_count, (u_long)object->size,
2258 				    (u_long)object->size,
2259 				    (void *)object->backing_object);
2260 			}
2261 		}
2262 	}
2263 }
2264 
2265 /*
2266  *	vm_object_print:	[ debug ]
2267  */
2268 DB_SHOW_COMMAND(object, vm_object_print_static)
2269 {
2270 	/* XXX convert args. */
2271 	vm_object_t object = (vm_object_t)addr;
2272 	boolean_t full = have_addr;
2273 
2274 	vm_page_t p;
2275 
2276 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
2277 #define	count	was_count
2278 
2279 	int count;
2280 
2281 	if (object == NULL)
2282 		return;
2283 
2284 	db_iprintf(
2285 	    "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x uip %d charge %jx\n",
2286 	    object, (int)object->type, (uintmax_t)object->size,
2287 	    object->resident_page_count, object->ref_count, object->flags,
2288 	    object->uip ? object->uip->ui_uid : -1, (uintmax_t)object->charge);
2289 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2290 	    object->shadow_count,
2291 	    object->backing_object ? object->backing_object->ref_count : 0,
2292 	    object->backing_object, (uintmax_t)object->backing_object_offset);
2293 
2294 	if (!full)
2295 		return;
2296 
2297 	db_indent += 2;
2298 	count = 0;
2299 	TAILQ_FOREACH(p, &object->memq, listq) {
2300 		if (count == 0)
2301 			db_iprintf("memory:=");
2302 		else if (count == 6) {
2303 			db_printf("\n");
2304 			db_iprintf(" ...");
2305 			count = 0;
2306 		} else
2307 			db_printf(",");
2308 		count++;
2309 
2310 		db_printf("(off=0x%jx,page=0x%jx)",
2311 		    (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2312 	}
2313 	if (count != 0)
2314 		db_printf("\n");
2315 	db_indent -= 2;
2316 }
2317 
2318 /* XXX. */
2319 #undef count
2320 
2321 /* XXX need this non-static entry for calling from vm_map_print. */
2322 void
2323 vm_object_print(
2324         /* db_expr_t */ long addr,
2325 	boolean_t have_addr,
2326 	/* db_expr_t */ long count,
2327 	char *modif)
2328 {
2329 	vm_object_print_static(addr, have_addr, count, modif);
2330 }
2331 
2332 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2333 {
2334 	vm_object_t object;
2335 	vm_pindex_t fidx;
2336 	vm_paddr_t pa;
2337 	vm_page_t m, prev_m;
2338 	int rcount, nl, c;
2339 
2340 	nl = 0;
2341 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
2342 		db_printf("new object: %p\n", (void *)object);
2343 		if (nl > 18) {
2344 			c = cngetc();
2345 			if (c != ' ')
2346 				return;
2347 			nl = 0;
2348 		}
2349 		nl++;
2350 		rcount = 0;
2351 		fidx = 0;
2352 		pa = -1;
2353 		TAILQ_FOREACH(m, &object->memq, listq) {
2354 			if (m->pindex > 128)
2355 				break;
2356 			if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2357 			    prev_m->pindex + 1 != m->pindex) {
2358 				if (rcount) {
2359 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2360 						(long)fidx, rcount, (long)pa);
2361 					if (nl > 18) {
2362 						c = cngetc();
2363 						if (c != ' ')
2364 							return;
2365 						nl = 0;
2366 					}
2367 					nl++;
2368 					rcount = 0;
2369 				}
2370 			}
2371 			if (rcount &&
2372 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2373 				++rcount;
2374 				continue;
2375 			}
2376 			if (rcount) {
2377 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2378 					(long)fidx, rcount, (long)pa);
2379 				if (nl > 18) {
2380 					c = cngetc();
2381 					if (c != ' ')
2382 						return;
2383 					nl = 0;
2384 				}
2385 				nl++;
2386 			}
2387 			fidx = m->pindex;
2388 			pa = VM_PAGE_TO_PHYS(m);
2389 			rcount = 1;
2390 		}
2391 		if (rcount) {
2392 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2393 				(long)fidx, rcount, (long)pa);
2394 			if (nl > 18) {
2395 				c = cngetc();
2396 				if (c != ' ')
2397 					return;
2398 				nl = 0;
2399 			}
2400 			nl++;
2401 		}
2402 	}
2403 }
2404 #endif /* DDB */
2405