xref: /dragonfly/sys/vm/vm_object.c (revision 7d2302ac)
1 /*
2  * Copyright (c) 1991, 1993, 2013
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  * 3. 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  * $FreeBSD: src/sys/vm/vm_object.c,v 1.171.2.8 2003/05/26 19:17:56 alc Exp $
61  */
62 
63 /*
64  *	Virtual memory object module.
65  */
66 
67 #include <sys/param.h>
68 #include <sys/systm.h>
69 #include <sys/proc.h>		/* for curproc, pageproc */
70 #include <sys/thread.h>
71 #include <sys/vnode.h>
72 #include <sys/vmmeter.h>
73 #include <sys/mman.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/refcount.h>
78 
79 #include <vm/vm.h>
80 #include <vm/vm_param.h>
81 #include <vm/pmap.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_pager.h>
87 #include <vm/swap_pager.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_extern.h>
90 #include <vm/vm_zone.h>
91 
92 #include <vm/vm_page2.h>
93 
94 #include <machine/specialreg.h>
95 
96 #define EASY_SCAN_FACTOR	8
97 
98 static void	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
99 					     int pagerflags);
100 static void	vm_object_lock_init(vm_object_t);
101 
102 /*
103  *	Virtual memory objects maintain the actual data
104  *	associated with allocated virtual memory.  A given
105  *	page of memory exists within exactly one object.
106  *
107  *	An object is only deallocated when all "references"
108  *	are given up.  Only one "reference" to a given
109  *	region of an object should be writeable.
110  *
111  *	Associated with each object is a list of all resident
112  *	memory pages belonging to that object; this list is
113  *	maintained by the "vm_page" module, and locked by the object's
114  *	lock.
115  *
116  *	Each object also records a "pager" routine which is
117  *	used to retrieve (and store) pages to the proper backing
118  *	storage.  In addition, objects may be backed by other
119  *	objects from which they were virtual-copied.
120  *
121  *	The only items within the object structure which are
122  *	modified after time of creation are:
123  *		reference count		locked by object's lock
124  *		pager routine		locked by object's lock
125  *
126  */
127 
128 struct vm_object kernel_object;
129 
130 struct vm_object_hash vm_object_hash[VMOBJ_HSIZE];
131 
132 MALLOC_DEFINE(M_VM_OBJECT, "vm_object", "vm_object structures");
133 
134 #define VMOBJ_HASH_PRIME1	66555444443333333ULL
135 #define VMOBJ_HASH_PRIME2	989042931893ULL
136 
137 int vm_object_debug;
138 SYSCTL_INT(_vm, OID_AUTO, object_debug, CTLFLAG_RW, &vm_object_debug, 0, "");
139 
140 static __inline
141 struct vm_object_hash *
142 vmobj_hash(vm_object_t obj)
143 {
144 	uintptr_t hash1;
145 	uintptr_t hash2;
146 
147 	hash1 = (uintptr_t)obj + ((uintptr_t)obj >> 18);
148 	hash1 %= VMOBJ_HASH_PRIME1;
149 	hash2 = ((uintptr_t)obj >> 8) + ((uintptr_t)obj >> 24);
150 	hash2 %= VMOBJ_HASH_PRIME2;
151 	return (&vm_object_hash[(hash1 ^ hash2) & VMOBJ_HMASK]);
152 }
153 
154 #if defined(DEBUG_LOCKS)
155 
156 #define vm_object_vndeallocate(obj, vpp)	\
157                 debugvm_object_vndeallocate(obj, vpp, __FILE__, __LINE__)
158 
159 /*
160  * Debug helper to track hold/drop/ref/deallocate calls.
161  */
162 static void
163 debugvm_object_add(vm_object_t obj, char *file, int line, int addrem)
164 {
165 	int i;
166 
167 	i = atomic_fetchadd_int(&obj->debug_index, 1);
168 	i = i & (VMOBJ_DEBUG_ARRAY_SIZE - 1);
169 	ksnprintf(obj->debug_hold_thrs[i],
170 		  sizeof(obj->debug_hold_thrs[i]),
171 		  "%c%d:(%d):%s",
172 		  (addrem == -1 ? '-' : (addrem == 1 ? '+' : '=')),
173 		  (curthread->td_proc ? curthread->td_proc->p_pid : -1),
174 		  obj->ref_count,
175 		  curthread->td_comm);
176 	obj->debug_hold_file[i] = file;
177 	obj->debug_hold_line[i] = line;
178 #if 0
179 	/* Uncomment for debugging obj refs/derefs in reproducable cases */
180 	if (strcmp(curthread->td_comm, "sshd") == 0) {
181 		kprintf("%d %p refs=%d ar=%d file: %s/%d\n",
182 			(curthread->td_proc ? curthread->td_proc->p_pid : -1),
183 			obj, obj->ref_count, addrem, file, line);
184 	}
185 #endif
186 }
187 
188 #endif
189 
190 /*
191  * Misc low level routines
192  */
193 static void
194 vm_object_lock_init(vm_object_t obj)
195 {
196 #if defined(DEBUG_LOCKS)
197 	int i;
198 
199 	obj->debug_index = 0;
200 	for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) {
201 		obj->debug_hold_thrs[i][0] = 0;
202 		obj->debug_hold_file[i] = NULL;
203 		obj->debug_hold_line[i] = 0;
204 	}
205 #endif
206 }
207 
208 void
209 vm_object_lock_swap(void)
210 {
211 	lwkt_token_swap();
212 }
213 
214 void
215 vm_object_lock(vm_object_t obj)
216 {
217 	lwkt_gettoken(&obj->token);
218 }
219 
220 /*
221  * Returns TRUE on sucesss
222  */
223 static int
224 vm_object_lock_try(vm_object_t obj)
225 {
226 	return(lwkt_trytoken(&obj->token));
227 }
228 
229 void
230 vm_object_lock_shared(vm_object_t obj)
231 {
232 	lwkt_gettoken_shared(&obj->token);
233 }
234 
235 void
236 vm_object_unlock(vm_object_t obj)
237 {
238 	lwkt_reltoken(&obj->token);
239 }
240 
241 void
242 vm_object_upgrade(vm_object_t obj)
243 {
244 	lwkt_reltoken(&obj->token);
245 	lwkt_gettoken(&obj->token);
246 }
247 
248 void
249 vm_object_downgrade(vm_object_t obj)
250 {
251 	lwkt_reltoken(&obj->token);
252 	lwkt_gettoken_shared(&obj->token);
253 }
254 
255 static __inline void
256 vm_object_assert_held(vm_object_t obj)
257 {
258 	ASSERT_LWKT_TOKEN_HELD(&obj->token);
259 }
260 
261 int
262 vm_quickcolor(void)
263 {
264 	globaldata_t gd = mycpu;
265 	int pg_color;
266 
267 	pg_color = (int)(intptr_t)gd->gd_curthread >> 10;
268 	pg_color += gd->gd_quick_color;
269 	gd->gd_quick_color += PQ_PRIME2;
270 
271 	return pg_color;
272 }
273 
274 void
275 VMOBJDEBUG(vm_object_hold)(vm_object_t obj VMOBJDBARGS)
276 {
277 	KKASSERT(obj != NULL);
278 
279 	/*
280 	 * Object must be held (object allocation is stable due to callers
281 	 * context, typically already holding the token on a parent object)
282 	 * prior to potentially blocking on the lock, otherwise the object
283 	 * can get ripped away from us.
284 	 */
285 	refcount_acquire(&obj->hold_count);
286 	vm_object_lock(obj);
287 
288 #if defined(DEBUG_LOCKS)
289 	debugvm_object_add(obj, file, line, 1);
290 #endif
291 }
292 
293 int
294 VMOBJDEBUG(vm_object_hold_try)(vm_object_t obj VMOBJDBARGS)
295 {
296 	KKASSERT(obj != NULL);
297 
298 	/*
299 	 * Object must be held (object allocation is stable due to callers
300 	 * context, typically already holding the token on a parent object)
301 	 * prior to potentially blocking on the lock, otherwise the object
302 	 * can get ripped away from us.
303 	 */
304 	refcount_acquire(&obj->hold_count);
305 	if (vm_object_lock_try(obj) == 0) {
306 		if (refcount_release(&obj->hold_count)) {
307 			if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD))
308 				kfree(obj, M_VM_OBJECT);
309 		}
310 		return(0);
311 	}
312 
313 #if defined(DEBUG_LOCKS)
314 	debugvm_object_add(obj, file, line, 1);
315 #endif
316 	return(1);
317 }
318 
319 void
320 VMOBJDEBUG(vm_object_hold_shared)(vm_object_t obj VMOBJDBARGS)
321 {
322 	KKASSERT(obj != NULL);
323 
324 	/*
325 	 * Object must be held (object allocation is stable due to callers
326 	 * context, typically already holding the token on a parent object)
327 	 * prior to potentially blocking on the lock, otherwise the object
328 	 * can get ripped away from us.
329 	 */
330 	refcount_acquire(&obj->hold_count);
331 	vm_object_lock_shared(obj);
332 
333 #if defined(DEBUG_LOCKS)
334 	debugvm_object_add(obj, file, line, 1);
335 #endif
336 }
337 
338 /*
339  * Drop the token and hold_count on the object.
340  *
341  * WARNING! Token might be shared.
342  */
343 void
344 VMOBJDEBUG(vm_object_drop)(vm_object_t obj VMOBJDBARGS)
345 {
346 	if (obj == NULL)
347 		return;
348 
349 	/*
350 	 * No new holders should be possible once we drop hold_count 1->0 as
351 	 * there is no longer any way to reference the object.
352 	 */
353 	KKASSERT(obj->hold_count > 0);
354 	if (refcount_release(&obj->hold_count)) {
355 #if defined(DEBUG_LOCKS)
356 		debugvm_object_add(obj, file, line, -1);
357 #endif
358 
359 		if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD)) {
360 			vm_object_unlock(obj);
361 			kfree(obj, M_VM_OBJECT);
362 		} else {
363 			vm_object_unlock(obj);
364 		}
365 	} else {
366 #if defined(DEBUG_LOCKS)
367 		debugvm_object_add(obj, file, line, -1);
368 #endif
369 		vm_object_unlock(obj);
370 	}
371 }
372 
373 /*
374  * Initialize a freshly allocated object, returning a held object.
375  *
376  * Used only by vm_object_allocate(), zinitna() and vm_object_init().
377  *
378  * No requirements.
379  */
380 void
381 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
382 {
383 	struct vm_object_hash *hash;
384 
385 	RB_INIT(&object->rb_memq);
386 	lwkt_token_init(&object->token, "vmobj");
387 
388 	TAILQ_INIT(&object->backing_list);
389 	lockinit(&object->backing_lk, "baclk", 0, 0);
390 
391 	object->type = type;
392 	object->size = size;
393 	object->ref_count = 1;
394 	object->memattr = VM_MEMATTR_DEFAULT;
395 	object->hold_count = 0;
396 	object->flags = 0;
397 	if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
398 		vm_object_set_flag(object, OBJ_ONEMAPPING);
399 	object->paging_in_progress = 0;
400 	object->resident_page_count = 0;
401 	/* cpu localization twist */
402 	object->pg_color = vm_quickcolor();
403 	object->handle = NULL;
404 
405 	atomic_add_int(&object->generation, 1);
406 	object->swblock_count = 0;
407 	RB_INIT(&object->swblock_root);
408 	vm_object_lock_init(object);
409 	pmap_object_init(object);
410 
411 	vm_object_hold(object);
412 
413 	hash = vmobj_hash(object);
414 	lwkt_gettoken(&hash->token);
415 	TAILQ_INSERT_TAIL(&hash->list, object, object_entry);
416 	lwkt_reltoken(&hash->token);
417 }
418 
419 /*
420  * Initialize a VM object.
421  */
422 void
423 vm_object_init(vm_object_t object, vm_pindex_t size)
424 {
425 	_vm_object_allocate(OBJT_DEFAULT, size, object);
426 	vm_object_drop(object);
427 }
428 
429 /*
430  * Initialize the VM objects module.
431  *
432  * Called from the low level boot code only.  Note that this occurs before
433  * kmalloc is initialized so we cannot allocate any VM objects.
434  */
435 void
436 vm_object_init1(void)
437 {
438 	int i;
439 
440 	for (i = 0; i < VMOBJ_HSIZE; ++i) {
441 		TAILQ_INIT(&vm_object_hash[i].list);
442 		lwkt_token_init(&vm_object_hash[i].token, "vmobjlst");
443 	}
444 
445 	_vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd),
446 			    &kernel_object);
447 	vm_object_drop(&kernel_object);
448 }
449 
450 void
451 vm_object_init2(void)
452 {
453 	kmalloc_set_unlimited(M_VM_OBJECT);
454 }
455 
456 /*
457  * Allocate and return a new object of the specified type and size.
458  *
459  * No requirements.
460  */
461 vm_object_t
462 vm_object_allocate(objtype_t type, vm_pindex_t size)
463 {
464 	vm_object_t obj;
465 
466 	obj = kmalloc(sizeof(*obj), M_VM_OBJECT, M_INTWAIT|M_ZERO);
467 	_vm_object_allocate(type, size, obj);
468 	vm_object_drop(obj);
469 
470 	return (obj);
471 }
472 
473 /*
474  * This version returns a held object, allowing further atomic initialization
475  * of the object.
476  */
477 vm_object_t
478 vm_object_allocate_hold(objtype_t type, vm_pindex_t size)
479 {
480 	vm_object_t obj;
481 
482 	obj = kmalloc(sizeof(*obj), M_VM_OBJECT, M_INTWAIT|M_ZERO);
483 	_vm_object_allocate(type, size, obj);
484 
485 	return (obj);
486 }
487 
488 /*
489  * Add an additional reference to a vm_object.  The object must already be
490  * held.  The original non-lock version is no longer supported.  The object
491  * must NOT be chain locked by anyone at the time the reference is added.
492  *
493  * The object must be held, but may be held shared if desired (hence why
494  * we use an atomic op).
495  */
496 void
497 VMOBJDEBUG(vm_object_reference_locked)(vm_object_t object VMOBJDBARGS)
498 {
499 	KKASSERT(object != NULL);
500 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
501 	atomic_add_int(&object->ref_count, 1);
502 	if (object->type == OBJT_VNODE) {
503 		vref(object->handle);
504 		/* XXX what if the vnode is being destroyed? */
505 	}
506 #if defined(DEBUG_LOCKS)
507 	debugvm_object_add(object, file, line, 1);
508 #endif
509 }
510 
511 /*
512  * This version is only allowed in situations where the caller
513  * already knows that the object is deterministically referenced
514  * (usually because its taken from a ref'd vnode, or during a map_entry
515  * replication).
516  */
517 void
518 VMOBJDEBUG(vm_object_reference_quick)(vm_object_t object VMOBJDBARGS)
519 {
520 	KKASSERT(object->type == OBJT_VNODE || object->ref_count > 0);
521 	atomic_add_int(&object->ref_count, 1);
522 	if (object->type == OBJT_VNODE)
523 		vref(object->handle);
524 #if defined(DEBUG_LOCKS)
525 	debugvm_object_add(object, file, line, 1);
526 #endif
527 }
528 
529 /*
530  * Dereference an object and its underlying vnode.  The object may be
531  * held shared.  On return the object will remain held.
532  *
533  * This function may return a vnode in *vpp which the caller must release
534  * after the caller drops its own lock.  If vpp is NULL, we assume that
535  * the caller was holding an exclusive lock on the object and we vrele()
536  * the vp ourselves.
537  */
538 static void
539 VMOBJDEBUG(vm_object_vndeallocate)(vm_object_t object, struct vnode **vpp
540 				   VMOBJDBARGS)
541 {
542 	struct vnode *vp = (struct vnode *) object->handle;
543 
544 	KASSERT(object->type == OBJT_VNODE,
545 	    ("vm_object_vndeallocate: not a vnode object"));
546 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
547 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
548 #ifdef INVARIANTS
549 	if (object->ref_count == 0) {
550 		vprint("vm_object_vndeallocate", vp);
551 		panic("vm_object_vndeallocate: bad object reference count");
552 	}
553 #endif
554 	for (;;) {
555 		int count = object->ref_count;
556 		cpu_ccfence();
557 		if (count == 1) {
558 			vm_object_upgrade(object);
559 			if (atomic_cmpset_int(&object->ref_count, count, 0)) {
560 				vclrflags(vp, VTEXT);
561 				break;
562 			}
563 		} else {
564 			if (atomic_cmpset_int(&object->ref_count,
565 					      count, count - 1)) {
566 				break;
567 			}
568 		}
569 		/* retry */
570 	}
571 #if defined(DEBUG_LOCKS)
572 	debugvm_object_add(object, file, line, -1);
573 #endif
574 
575 	/*
576 	 * vrele or return the vp to vrele.  We can only safely vrele(vp)
577 	 * if the object was locked exclusively.  But there are two races
578 	 * here.
579 	 *
580 	 * We had to upgrade the object above to safely clear VTEXT
581 	 * but the alternative path where the shared lock is retained
582 	 * can STILL race to 0 in other paths and cause our own vrele()
583 	 * to terminate the vnode.  We can't allow that if the VM object
584 	 * is still locked shared.
585 	 */
586 	if (vpp)
587 		*vpp = vp;
588 	else
589 		vrele(vp);
590 }
591 
592 /*
593  * Release a reference to the specified object, gained either through a
594  * vm_object_allocate or a vm_object_reference call.  When all references
595  * are gone, storage associated with this object may be relinquished.
596  *
597  * The caller does not have to hold the object locked but must have control
598  * over the reference in question in order to guarantee that the object
599  * does not get ripped out from under us.
600  *
601  * XXX Currently all deallocations require an exclusive lock.
602  */
603 void
604 VMOBJDEBUG(vm_object_deallocate)(vm_object_t object VMOBJDBARGS)
605 {
606 	struct vnode *vp;
607 	int count;
608 
609 	if (object == NULL)
610 		return;
611 
612 	for (;;) {
613 		count = object->ref_count;
614 		cpu_ccfence();
615 
616 		/*
617 		 * If decrementing the count enters into special handling
618 		 * territory (0, 1, or 2) we have to do it the hard way.
619 		 * Fortunate though, objects with only a few refs like this
620 		 * are not likely to be heavily contended anyway.
621 		 *
622 		 * For vnode objects we only care about 1->0 transitions.
623 		 */
624 		if (count <= 3 || (object->type == OBJT_VNODE && count <= 1)) {
625 #if defined(DEBUG_LOCKS)
626 			debugvm_object_add(object, file, line, 0);
627 #endif
628 			vm_object_hold(object);
629 			vm_object_deallocate_locked(object);
630 			vm_object_drop(object);
631 			break;
632 		}
633 
634 		/*
635 		 * Try to decrement ref_count without acquiring a hold on
636 		 * the object.  This is particularly important for the exec*()
637 		 * and exit*() code paths because the program binary may
638 		 * have a great deal of sharing and an exclusive lock will
639 		 * crowbar performance in those circumstances.
640 		 */
641 		if (object->type == OBJT_VNODE) {
642 			vp = (struct vnode *)object->handle;
643 			if (atomic_cmpset_int(&object->ref_count,
644 					      count, count - 1)) {
645 #if defined(DEBUG_LOCKS)
646 				debugvm_object_add(object, file, line, -1);
647 #endif
648 
649 				vrele(vp);
650 				break;
651 			}
652 			/* retry */
653 		} else {
654 			if (atomic_cmpset_int(&object->ref_count,
655 					      count, count - 1)) {
656 #if defined(DEBUG_LOCKS)
657 				debugvm_object_add(object, file, line, -1);
658 #endif
659 				break;
660 			}
661 			/* retry */
662 		}
663 		/* retry */
664 	}
665 }
666 
667 void
668 VMOBJDEBUG(vm_object_deallocate_locked)(vm_object_t object VMOBJDBARGS)
669 {
670 	/*
671 	 * Degenerate case
672 	 */
673 	if (object == NULL)
674 		return;
675 
676 	/*
677 	 * vnode case, caller either locked the object exclusively
678 	 * or this is a recursion with must_drop != 0 and the vnode
679 	 * object will be locked shared.
680 	 *
681 	 * If locked shared we have to drop the object before we can
682 	 * call vrele() or risk a shared/exclusive livelock.
683 	 */
684 	if (object->type == OBJT_VNODE) {
685 		ASSERT_LWKT_TOKEN_HELD(&object->token);
686 		vm_object_vndeallocate(object, NULL);
687 		return;
688 	}
689 	ASSERT_LWKT_TOKEN_HELD_EXCL(&object->token);
690 
691 	/*
692 	 * Normal case (object is locked exclusively)
693 	 */
694 	if (object->ref_count == 0) {
695 		panic("vm_object_deallocate: object deallocated "
696 		      "too many times: %d", object->type);
697 	}
698 	if (object->ref_count > 2) {
699 		atomic_add_int(&object->ref_count, -1);
700 #if defined(DEBUG_LOCKS)
701 		debugvm_object_add(object, file, line, -1);
702 #endif
703 		return;
704 	}
705 
706 	/*
707 	 * Drop the ref and handle termination on the 1->0 transition.
708 	 * We may have blocked above so we have to recheck.
709 	 */
710 	KKASSERT(object->ref_count != 0);
711 	if (object->ref_count >= 2) {
712 		atomic_add_int(&object->ref_count, -1);
713 #if defined(DEBUG_LOCKS)
714 		debugvm_object_add(object, file, line, -1);
715 #endif
716 		return;
717 	}
718 
719 	atomic_add_int(&object->ref_count, -1);
720 	if ((object->flags & OBJ_DEAD) == 0)
721 		vm_object_terminate(object);
722 }
723 
724 /*
725  * Destroy the specified object, freeing up related resources.
726  *
727  * The object must have zero references.
728  *
729  * The object must held.  The caller is responsible for dropping the object
730  * after terminate returns.  Terminate does NOT drop the object.
731  */
732 static int vm_object_terminate_callback(vm_page_t p, void *data);
733 
734 void
735 vm_object_terminate(vm_object_t object)
736 {
737 	struct rb_vm_page_scan_info info;
738 	struct vm_object_hash *hash;
739 
740 	/*
741 	 * Make sure no one uses us.  Once we set OBJ_DEAD we should be
742 	 * able to safely block.
743 	 */
744 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
745 	KKASSERT((object->flags & OBJ_DEAD) == 0);
746 	vm_object_set_flag(object, OBJ_DEAD);
747 
748 	/*
749 	 * Wait for the pageout daemon to be done with the object
750 	 */
751 	vm_object_pip_wait(object, "objtrm1");
752 
753 	KASSERT(!object->paging_in_progress,
754 		("vm_object_terminate: pageout in progress"));
755 
756 	/*
757 	 * Clean and free the pages, as appropriate. All references to the
758 	 * object are gone, so we don't need to lock it.
759 	 */
760 	if (object->type == OBJT_VNODE) {
761 		struct vnode *vp;
762 
763 		/*
764 		 * Clean pages and flush buffers.
765 		 *
766 		 * NOTE!  TMPFS buffer flushes do not typically flush the
767 		 *	  actual page to swap as this would be highly
768 		 *	  inefficient, and normal filesystems usually wrap
769 		 *	  page flushes with buffer cache buffers.
770 		 *
771 		 *	  To deal with this we have to call vinvalbuf() both
772 		 *	  before and after the vm_object_page_clean().
773 		 */
774 		vp = (struct vnode *) object->handle;
775 		vinvalbuf(vp, V_SAVE, 0, 0);
776 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
777 		vinvalbuf(vp, V_SAVE, 0, 0);
778 	}
779 
780 	/*
781 	 * Wait for any I/O to complete, after which there had better not
782 	 * be any references left on the object.
783 	 */
784 	vm_object_pip_wait(object, "objtrm2");
785 
786 	if (object->ref_count != 0) {
787 		panic("vm_object_terminate: object with references, "
788 		      "ref_count=%d", object->ref_count);
789 	}
790 
791 	/*
792 	 * Cleanup any shared pmaps associated with this object.
793 	 */
794 	pmap_object_free(object);
795 
796 	/*
797 	 * Now free any remaining pages. For internal objects, this also
798 	 * removes them from paging queues. Don't free wired pages, just
799 	 * remove them from the object.
800 	 */
801 	info.count = 0;
802 	info.object = object;
803 	do {
804 		info.error = 0;
805 		vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
806 					vm_object_terminate_callback, &info);
807 	} while (info.error);
808 
809 	/*
810 	 * Let the pager know object is dead.
811 	 */
812 	vm_pager_deallocate(object);
813 
814 	/*
815 	 * Wait for the object hold count to hit 1, clean out pages as
816 	 * we go.  vmobj_token interlocks any race conditions that might
817 	 * pick the object up from the vm_object_list after we have cleared
818 	 * rb_memq.
819 	 */
820 	for (;;) {
821 		if (RB_ROOT(&object->rb_memq) == NULL)
822 			break;
823 		kprintf("vm_object_terminate: Warning, object %p "
824 			"still has %ld pages\n",
825 			object, object->resident_page_count);
826 		vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
827 					vm_object_terminate_callback, &info);
828 	}
829 
830 	/*
831 	 * There had better not be any pages left
832 	 */
833 	KKASSERT(object->resident_page_count == 0);
834 
835 	/*
836 	 * Remove the object from the global object list.
837 	 */
838 	hash = vmobj_hash(object);
839 	lwkt_gettoken(&hash->token);
840 	TAILQ_REMOVE(&hash->list, object, object_entry);
841 	lwkt_reltoken(&hash->token);
842 
843 	if (object->ref_count != 0) {
844 		panic("vm_object_terminate2: object with references, "
845 		      "ref_count=%d", object->ref_count);
846 	}
847 
848 	/*
849 	 * NOTE: The object hold_count is at least 1, so we cannot kfree()
850 	 *	 the object here.  See vm_object_drop().
851 	 */
852 }
853 
854 /*
855  * The caller must hold the object.
856  */
857 static int
858 vm_object_terminate_callback(vm_page_t p, void *data)
859 {
860 	struct rb_vm_page_scan_info *info = data;
861 	vm_object_t object;
862 
863 	object = p->object;
864 	KKASSERT(object == info->object);
865 	if (vm_page_busy_try(p, TRUE)) {
866 		vm_page_sleep_busy(p, TRUE, "vmotrm");
867 		info->error = 1;
868 		return 0;
869 	}
870 	if (object != p->object) {
871 		/* XXX remove once we determine it can't happen */
872 		kprintf("vm_object_terminate: Warning: Encountered "
873 			"busied page %p on queue %d\n", p, p->queue);
874 		vm_page_wakeup(p);
875 		info->error = 1;
876 	} else if (p->wire_count == 0) {
877 		/*
878 		 * NOTE: p->dirty and PG_NEED_COMMIT are ignored.
879 		 */
880 		vm_page_free(p);
881 		mycpu->gd_cnt.v_pfree++;
882 	} else {
883 		if (p->queue != PQ_NONE) {
884 			kprintf("vm_object_terminate: Warning: Encountered "
885 				"wired page %p on queue %d\n", p, p->queue);
886 			if (vm_object_debug > 0) {
887 				--vm_object_debug;
888 				print_backtrace(10);
889 			}
890 		}
891 		vm_page_remove(p);
892 		vm_page_wakeup(p);
893 	}
894 
895 	/*
896 	 * Must be at end to avoid SMP races, caller holds object token
897 	 */
898 	if ((++info->count & 63) == 0)
899 		lwkt_user_yield();
900 	return(0);
901 }
902 
903 /*
904  * Clean all dirty pages in the specified range of object.  Leaves page
905  * on whatever queue it is currently on.   If NOSYNC is set then do not
906  * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
907  * leaving the object dirty.
908  *
909  * When stuffing pages asynchronously, allow clustering.  XXX we need a
910  * synchronous clustering mode implementation.
911  *
912  * Odd semantics: if start == end, we clean everything.
913  *
914  * The object must be locked? XXX
915  */
916 static int vm_object_page_clean_pass1(struct vm_page *p, void *data);
917 static int vm_object_page_clean_pass2(struct vm_page *p, void *data);
918 
919 void
920 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
921 		     int flags)
922 {
923 	struct rb_vm_page_scan_info info;
924 	struct vnode *vp;
925 	int wholescan;
926 	int pagerflags;
927 	int generation;
928 
929 	vm_object_hold(object);
930 	if (object->type != OBJT_VNODE ||
931 	    (object->flags & OBJ_MIGHTBEDIRTY) == 0) {
932 		vm_object_drop(object);
933 		return;
934 	}
935 
936 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ?
937 			VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
938 	pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
939 
940 	vp = object->handle;
941 
942 	/*
943 	 * Interlock other major object operations.  This allows us to
944 	 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY.
945 	 */
946 	vm_object_set_flag(object, OBJ_CLEANING);
947 
948 	/*
949 	 * Handle 'entire object' case
950 	 */
951 	info.start_pindex = start;
952 	if (end == 0) {
953 		info.end_pindex = object->size - 1;
954 	} else {
955 		info.end_pindex = end - 1;
956 	}
957 	wholescan = (start == 0 && info.end_pindex == object->size - 1);
958 	info.limit = flags;
959 	info.pagerflags = pagerflags;
960 	info.object = object;
961 
962 	/*
963 	 * If cleaning the entire object do a pass to mark the pages read-only.
964 	 * If everything worked out ok, clear OBJ_WRITEABLE and
965 	 * OBJ_MIGHTBEDIRTY.
966 	 */
967 	if (wholescan) {
968 		info.error = 0;
969 		info.count = 0;
970 		vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
971 					vm_object_page_clean_pass1, &info);
972 		if (info.error == 0) {
973 			vm_object_clear_flag(object,
974 					     OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
975 			if (object->type == OBJT_VNODE &&
976 			    (vp = (struct vnode *)object->handle) != NULL) {
977 				/*
978 				 * Use new-style interface to clear VISDIRTY
979 				 * because the vnode is not necessarily removed
980 				 * from the syncer list(s) as often as it was
981 				 * under the old interface, which can leave
982 				 * the vnode on the syncer list after reclaim.
983 				 */
984 				vclrobjdirty(vp);
985 			}
986 		}
987 	}
988 
989 	/*
990 	 * Do a pass to clean all the dirty pages we find.
991 	 */
992 	do {
993 		info.error = 0;
994 		info.count = 0;
995 		generation = object->generation;
996 		vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
997 					vm_object_page_clean_pass2, &info);
998 	} while (info.error || generation != object->generation);
999 
1000 	vm_object_clear_flag(object, OBJ_CLEANING);
1001 	vm_object_drop(object);
1002 }
1003 
1004 /*
1005  * The caller must hold the object.
1006  */
1007 static
1008 int
1009 vm_object_page_clean_pass1(struct vm_page *p, void *data)
1010 {
1011 	struct rb_vm_page_scan_info *info = data;
1012 
1013 	KKASSERT(p->object == info->object);
1014 
1015 	vm_page_flag_set(p, PG_CLEANCHK);
1016 	if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1017 		info->error = 1;
1018 	} else if (vm_page_busy_try(p, FALSE)) {
1019 		info->error = 1;
1020 	} else {
1021 		KKASSERT(p->object == info->object);
1022 		vm_page_protect(p, VM_PROT_READ);
1023 		vm_page_wakeup(p);
1024 	}
1025 
1026 	/*
1027 	 * Must be at end to avoid SMP races, caller holds object token
1028 	 */
1029 	if ((++info->count & 63) == 0)
1030 		lwkt_user_yield();
1031 	return(0);
1032 }
1033 
1034 /*
1035  * The caller must hold the object
1036  */
1037 static
1038 int
1039 vm_object_page_clean_pass2(struct vm_page *p, void *data)
1040 {
1041 	struct rb_vm_page_scan_info *info = data;
1042 	int generation;
1043 
1044 	KKASSERT(p->object == info->object);
1045 
1046 	/*
1047 	 * Do not mess with pages that were inserted after we started
1048 	 * the cleaning pass.
1049 	 */
1050 	if ((p->flags & PG_CLEANCHK) == 0)
1051 		goto done;
1052 
1053 	generation = info->object->generation;
1054 
1055 	if (vm_page_busy_try(p, TRUE)) {
1056 		vm_page_sleep_busy(p, TRUE, "vpcwai");
1057 		info->error = 1;
1058 		goto done;
1059 	}
1060 
1061 	KKASSERT(p->object == info->object &&
1062 		 info->object->generation == generation);
1063 
1064 	/*
1065 	 * Before wasting time traversing the pmaps, check for trivial
1066 	 * cases where the page cannot be dirty.
1067 	 */
1068 	if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) {
1069 		KKASSERT((p->dirty & p->valid) == 0 &&
1070 			 (p->flags & PG_NEED_COMMIT) == 0);
1071 		vm_page_wakeup(p);
1072 		goto done;
1073 	}
1074 
1075 	/*
1076 	 * Check whether the page is dirty or not.  The page has been set
1077 	 * to be read-only so the check will not race a user dirtying the
1078 	 * page.
1079 	 */
1080 	vm_page_test_dirty(p);
1081 	if ((p->dirty & p->valid) == 0 && (p->flags & PG_NEED_COMMIT) == 0) {
1082 		vm_page_flag_clear(p, PG_CLEANCHK);
1083 		vm_page_wakeup(p);
1084 		goto done;
1085 	}
1086 
1087 	/*
1088 	 * If we have been asked to skip nosync pages and this is a
1089 	 * nosync page, skip it.  Note that the object flags were
1090 	 * not cleared in this case (because pass1 will have returned an
1091 	 * error), so we do not have to set them.
1092 	 */
1093 	if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1094 		vm_page_flag_clear(p, PG_CLEANCHK);
1095 		vm_page_wakeup(p);
1096 		goto done;
1097 	}
1098 
1099 	/*
1100 	 * Flush as many pages as we can.  PG_CLEANCHK will be cleared on
1101 	 * the pages that get successfully flushed.  Set info->error if
1102 	 * we raced an object modification.
1103 	 */
1104 	vm_object_page_collect_flush(info->object, p, info->pagerflags);
1105 	/* vm_wait_nominal(); this can deadlock the system in syncer/pageout */
1106 
1107 	/*
1108 	 * Must be at end to avoid SMP races, caller holds object token
1109 	 */
1110 done:
1111 	if ((++info->count & 63) == 0)
1112 		lwkt_user_yield();
1113 	return(0);
1114 }
1115 
1116 /*
1117  * Collect the specified page and nearby pages and flush them out.
1118  * The number of pages flushed is returned.  The passed page is busied
1119  * by the caller and we are responsible for its disposition.
1120  *
1121  * The caller must hold the object.
1122  */
1123 static void
1124 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags)
1125 {
1126 	int error;
1127 	int is;
1128 	int ib;
1129 	int i;
1130 	int page_base;
1131 	vm_pindex_t pi;
1132 	vm_page_t ma[BLIST_MAX_ALLOC];
1133 
1134 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1135 
1136 	pi = p->pindex;
1137 	page_base = pi % BLIST_MAX_ALLOC;
1138 	ma[page_base] = p;
1139 	ib = page_base - 1;
1140 	is = page_base + 1;
1141 
1142 	while (ib >= 0) {
1143 		vm_page_t tp;
1144 
1145 		tp = vm_page_lookup_busy_try(object, pi - page_base + ib,
1146 					     TRUE, &error);
1147 		if (error)
1148 			break;
1149 		if (tp == NULL)
1150 			break;
1151 		if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1152 		    (tp->flags & PG_CLEANCHK) == 0) {
1153 			vm_page_wakeup(tp);
1154 			break;
1155 		}
1156 		if ((tp->queue - tp->pc) == PQ_CACHE) {
1157 			vm_page_flag_clear(tp, PG_CLEANCHK);
1158 			vm_page_wakeup(tp);
1159 			break;
1160 		}
1161 		vm_page_test_dirty(tp);
1162 		if ((tp->dirty & tp->valid) == 0 &&
1163 		    (tp->flags & PG_NEED_COMMIT) == 0) {
1164 			vm_page_flag_clear(tp, PG_CLEANCHK);
1165 			vm_page_wakeup(tp);
1166 			break;
1167 		}
1168 		ma[ib] = tp;
1169 		--ib;
1170 	}
1171 	++ib;	/* fixup */
1172 
1173 	while (is < BLIST_MAX_ALLOC &&
1174 	       pi - page_base + is < object->size) {
1175 		vm_page_t tp;
1176 
1177 		tp = vm_page_lookup_busy_try(object, pi - page_base + is,
1178 					     TRUE, &error);
1179 		if (error)
1180 			break;
1181 		if (tp == NULL)
1182 			break;
1183 		if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1184 		    (tp->flags & PG_CLEANCHK) == 0) {
1185 			vm_page_wakeup(tp);
1186 			break;
1187 		}
1188 		if ((tp->queue - tp->pc) == PQ_CACHE) {
1189 			vm_page_flag_clear(tp, PG_CLEANCHK);
1190 			vm_page_wakeup(tp);
1191 			break;
1192 		}
1193 		vm_page_test_dirty(tp);
1194 		if ((tp->dirty & tp->valid) == 0 &&
1195 		    (tp->flags & PG_NEED_COMMIT) == 0) {
1196 			vm_page_flag_clear(tp, PG_CLEANCHK);
1197 			vm_page_wakeup(tp);
1198 			break;
1199 		}
1200 		ma[is] = tp;
1201 		++is;
1202 	}
1203 
1204 	/*
1205 	 * All pages in the ma[] array are busied now
1206 	 */
1207 	for (i = ib; i < is; ++i) {
1208 		vm_page_flag_clear(ma[i], PG_CLEANCHK);
1209 		vm_page_hold(ma[i]);	/* XXX need this any more? */
1210 	}
1211 	vm_pageout_flush(&ma[ib], is - ib, pagerflags);
1212 	for (i = ib; i < is; ++i)	/* XXX need this any more? */
1213 		vm_page_unhold(ma[i]);
1214 }
1215 
1216 /*
1217  * Implements the madvise function at the object/page level.
1218  *
1219  * MADV_WILLNEED	(any object)
1220  *
1221  *	Activate the specified pages if they are resident.
1222  *
1223  * MADV_DONTNEED	(any object)
1224  *
1225  *	Deactivate the specified pages if they are resident.
1226  *
1227  * MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only)
1228  *
1229  *	Deactivate and clean the specified pages if they are
1230  *	resident.  This permits the process to reuse the pages
1231  *	without faulting or the kernel to reclaim the pages
1232  *	without I/O.
1233  *
1234  * No requirements.
1235  */
1236 void
1237 vm_object_madvise(vm_object_t object, vm_pindex_t pindex,
1238 		  vm_pindex_t count, int advise)
1239 {
1240 	vm_pindex_t end;
1241 	vm_page_t m;
1242 	int error;
1243 
1244 	if (object == NULL)
1245 		return;
1246 
1247 	end = pindex + count;
1248 
1249 	vm_object_hold(object);
1250 
1251 	/*
1252 	 * Locate and adjust resident pages.  This only applies to the
1253 	 * primary object in the mapping.
1254 	 */
1255 	for (; pindex < end; pindex += 1) {
1256 relookup:
1257 		/*
1258 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1259 		 * and those pages must be OBJ_ONEMAPPING.
1260 		 */
1261 		if (advise == MADV_FREE) {
1262 			if ((object->type != OBJT_DEFAULT &&
1263 			     object->type != OBJT_SWAP) ||
1264 			    (object->flags & OBJ_ONEMAPPING) == 0) {
1265 				continue;
1266 			}
1267 		}
1268 
1269 		m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
1270 
1271 		if (error) {
1272 			vm_page_sleep_busy(m, TRUE, "madvpo");
1273 			goto relookup;
1274 		}
1275 		if (m == NULL) {
1276 			/*
1277 			 * There may be swap even if there is no backing page
1278 			 */
1279 			if (advise == MADV_FREE && object->type == OBJT_SWAP)
1280 				swap_pager_freespace(object, pindex, 1);
1281 			continue;
1282 		}
1283 
1284 		/*
1285 		 * If the page is not in a normal active state, we skip it.
1286 		 * If the page is not managed there are no page queues to
1287 		 * mess with.  Things can break if we mess with pages in
1288 		 * any of the below states.
1289 		 */
1290 		if (m->wire_count ||
1291 		    (m->flags & (PG_FICTITIOUS | PG_UNQUEUED |
1292 				 PG_NEED_COMMIT)) ||
1293 		    m->valid != VM_PAGE_BITS_ALL
1294 		) {
1295 			vm_page_wakeup(m);
1296 			continue;
1297 		}
1298 
1299 		/*
1300 		 * Theoretically once a page is known not to be busy, an
1301 		 * interrupt cannot come along and rip it out from under us.
1302 		 */
1303 		if (advise == MADV_WILLNEED) {
1304 			vm_page_activate(m);
1305 		} else if (advise == MADV_DONTNEED) {
1306 			vm_page_dontneed(m);
1307 		} else if (advise == MADV_FREE) {
1308 			/*
1309 			 * Mark the page clean.  This will allow the page
1310 			 * to be freed up by the system.  However, such pages
1311 			 * are often reused quickly by malloc()/free()
1312 			 * so we do not do anything that would cause
1313 			 * a page fault if we can help it.
1314 			 *
1315 			 * Specifically, we do not try to actually free
1316 			 * the page now nor do we try to put it in the
1317 			 * cache (which would cause a page fault on reuse).
1318 			 *
1319 			 * But we do make the page is freeable as we
1320 			 * can without actually taking the step of unmapping
1321 			 * it.
1322 			 */
1323 			pmap_clear_modify(m);
1324 			m->dirty = 0;
1325 			m->act_count = 0;
1326 			vm_page_dontneed(m);
1327 			if (object->type == OBJT_SWAP)
1328 				swap_pager_freespace(object, pindex, 1);
1329 		}
1330 		vm_page_wakeup(m);
1331 	}
1332 	vm_object_drop(object);
1333 }
1334 
1335 /*
1336  * Removes all physical pages in the specified object range from the
1337  * object's list of pages.
1338  *
1339  * No requirements.
1340  */
1341 static int vm_object_page_remove_callback(vm_page_t p, void *data);
1342 
1343 void
1344 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1345 		      boolean_t clean_only)
1346 {
1347 	struct rb_vm_page_scan_info info;
1348 	int all;
1349 
1350 	/*
1351 	 * Degenerate cases and assertions
1352 	 */
1353 	vm_object_hold(object);
1354 	if (object == NULL ||
1355 	    (object->resident_page_count == 0 && object->swblock_count == 0)) {
1356 		vm_object_drop(object);
1357 		return;
1358 	}
1359 	KASSERT(object->type != OBJT_PHYS,
1360 		("attempt to remove pages from a physical object"));
1361 
1362 	/*
1363 	 * Indicate that paging is occuring on the object
1364 	 */
1365 	vm_object_pip_add(object, 1);
1366 
1367 	/*
1368 	 * Figure out the actual removal range and whether we are removing
1369 	 * the entire contents of the object or not.  If removing the entire
1370 	 * contents, be sure to get all pages, even those that might be
1371 	 * beyond the end of the object.
1372 	 */
1373 	info.object = object;
1374 	info.start_pindex = start;
1375 	if (end == 0)
1376 		info.end_pindex = (vm_pindex_t)-1;
1377 	else
1378 		info.end_pindex = end - 1;
1379 	info.limit = clean_only;
1380 	info.count = 0;
1381 	all = (start == 0 && info.end_pindex >= object->size - 1);
1382 
1383 	/*
1384 	 * Loop until we are sure we have gotten them all.
1385 	 */
1386 	do {
1387 		info.error = 0;
1388 		vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1389 					vm_object_page_remove_callback, &info);
1390 	} while (info.error);
1391 
1392 	/*
1393 	 * Remove any related swap if throwing away pages, or for
1394 	 * non-swap objects (the swap is a clean copy in that case).
1395 	 */
1396 	if (object->type != OBJT_SWAP || clean_only == FALSE) {
1397 		if (all)
1398 			swap_pager_freespace_all(object);
1399 		else
1400 			swap_pager_freespace(object, info.start_pindex,
1401 			     info.end_pindex - info.start_pindex + 1);
1402 	}
1403 
1404 	/*
1405 	 * Cleanup
1406 	 */
1407 	vm_object_pip_wakeup(object);
1408 	vm_object_drop(object);
1409 }
1410 
1411 /*
1412  * The caller must hold the object.
1413  *
1414  * NOTE: User yields are allowed when removing more than one page, but not
1415  *	 allowed if only removing one page (the path for single page removals
1416  *	 might hold a spinlock).
1417  */
1418 static int
1419 vm_object_page_remove_callback(vm_page_t p, void *data)
1420 {
1421 	struct rb_vm_page_scan_info *info = data;
1422 
1423 	if (info->object != p->object ||
1424 	    p->pindex < info->start_pindex ||
1425 	    p->pindex > info->end_pindex) {
1426 		kprintf("vm_object_page_remove_callbackA: obj/pg race %p/%p\n",
1427 			info->object, p);
1428 		return(0);
1429 	}
1430 	if (vm_page_busy_try(p, TRUE)) {
1431 		vm_page_sleep_busy(p, TRUE, "vmopar");
1432 		info->error = 1;
1433 		return(0);
1434 	}
1435 	if (info->object != p->object) {
1436 		/* this should never happen */
1437 		kprintf("vm_object_page_remove_callbackB: obj/pg race %p/%p\n",
1438 			info->object, p);
1439 		vm_page_wakeup(p);
1440 		return(0);
1441 	}
1442 
1443 	/*
1444 	 * Wired pages cannot be destroyed, but they can be invalidated
1445 	 * and we do so if clean_only (limit) is not set.
1446 	 *
1447 	 * WARNING!  The page may be wired due to being part of a buffer
1448 	 *	     cache buffer, and the buffer might be marked B_CACHE.
1449 	 *	     This is fine as part of a truncation but VFSs must be
1450 	 *	     sure to fix the buffer up when re-extending the file.
1451 	 *
1452 	 * NOTE!     PG_NEED_COMMIT is ignored.
1453 	 */
1454 	if (p->wire_count != 0) {
1455 		vm_page_protect(p, VM_PROT_NONE);
1456 		if (info->limit == 0)
1457 			p->valid = 0;
1458 		vm_page_wakeup(p);
1459 		goto done;
1460 	}
1461 
1462 	/*
1463 	 * limit is our clean_only flag.  If set and the page is dirty or
1464 	 * requires a commit, do not free it.  If set and the page is being
1465 	 * held by someone, do not free it.
1466 	 */
1467 	if (info->limit && p->valid) {
1468 		vm_page_test_dirty(p);
1469 		if ((p->valid & p->dirty) || (p->flags & PG_NEED_COMMIT)) {
1470 			vm_page_wakeup(p);
1471 			goto done;
1472 		}
1473 	}
1474 
1475 	/*
1476 	 * Destroy the page.  But we have to re-test whether its dirty after
1477 	 * removing it from its pmaps.
1478 	 */
1479 	vm_page_protect(p, VM_PROT_NONE);
1480 	if (info->limit && p->valid) {
1481 		vm_page_test_dirty(p);
1482 		if ((p->valid & p->dirty) || (p->flags & PG_NEED_COMMIT)) {
1483 			vm_page_wakeup(p);
1484 			goto done;
1485 		}
1486 	}
1487 	vm_page_free(p);
1488 
1489 	/*
1490 	 * Must be at end to avoid SMP races, caller holds object token
1491 	 */
1492 done:
1493 	if ((++info->count & 63) == 0)
1494 		lwkt_user_yield();
1495 
1496 	return(0);
1497 }
1498 
1499 /*
1500  * Try to extend prev_object into an adjoining region of virtual
1501  * memory, return TRUE on success.
1502  *
1503  * The caller does not need to hold (prev_object) but must have a stable
1504  * pointer to it (typically by holding the vm_map locked).
1505  *
1506  * This function only works for anonymous memory objects which either
1507  * have (a) one reference or (b) we are extending the object's size.
1508  * Otherwise the related VM pages we want to use for the object might
1509  * be in use by another mapping.
1510  */
1511 boolean_t
1512 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
1513 		   vm_size_t prev_size, vm_size_t next_size)
1514 {
1515 	vm_pindex_t next_pindex;
1516 
1517 	if (prev_object == NULL)
1518 		return (TRUE);
1519 
1520 	vm_object_hold(prev_object);
1521 
1522 	if (prev_object->type != OBJT_DEFAULT &&
1523 	    prev_object->type != OBJT_SWAP) {
1524 		vm_object_drop(prev_object);
1525 		return (FALSE);
1526 	}
1527 
1528 #if 0
1529 	/* caller now checks this */
1530 	/*
1531 	 * Try to collapse the object first
1532 	 */
1533 	vm_object_collapse(prev_object, NULL);
1534 #endif
1535 
1536 #if 0
1537 	/* caller now checks this */
1538 	/*
1539 	 * We can't coalesce if we shadow another object (figuring out the
1540 	 * relationships become too complex).
1541 	 */
1542 	if (prev_object->backing_object != NULL) {
1543 		vm_object_chain_release(prev_object);
1544 		vm_object_drop(prev_object);
1545 		return (FALSE);
1546 	}
1547 #endif
1548 
1549 	prev_size >>= PAGE_SHIFT;
1550 	next_size >>= PAGE_SHIFT;
1551 	next_pindex = prev_pindex + prev_size;
1552 
1553 	/*
1554 	 * We can't if the object has more than one ref count unless we
1555 	 * are extending it into newly minted space.
1556 	 */
1557 	if (prev_object->ref_count > 1 &&
1558 	    prev_object->size != next_pindex) {
1559 		vm_object_drop(prev_object);
1560 		return (FALSE);
1561 	}
1562 
1563 	/*
1564 	 * Remove any pages that may still be in the object from a previous
1565 	 * deallocation.
1566 	 */
1567 	if (next_pindex < prev_object->size) {
1568 		vm_object_page_remove(prev_object,
1569 				      next_pindex,
1570 				      next_pindex + next_size, FALSE);
1571 		if (prev_object->type == OBJT_SWAP)
1572 			swap_pager_freespace(prev_object,
1573 					     next_pindex, next_size);
1574 	}
1575 
1576 	/*
1577 	 * Extend the object if necessary.
1578 	 */
1579 	if (next_pindex + next_size > prev_object->size)
1580 		prev_object->size = next_pindex + next_size;
1581 	vm_object_drop(prev_object);
1582 
1583 	return (TRUE);
1584 }
1585 
1586 /*
1587  * Make the object writable and flag is being possibly dirty.
1588  *
1589  * The object might not be held (or might be held but held shared),
1590  * the related vnode is probably not held either.  Object and vnode are
1591  * stable by virtue of the vm_page busied by the caller preventing
1592  * destruction.
1593  *
1594  * If the related mount is flagged MNTK_THR_SYNC we need to call
1595  * vsetobjdirty().  Filesystems using this option usually shortcut
1596  * synchronization by only scanning the syncer list.
1597  */
1598 void
1599 vm_object_set_writeable_dirty(vm_object_t object)
1600 {
1601 	struct vnode *vp;
1602 
1603 	/*vm_object_assert_held(object);*/
1604 	/*
1605 	 * Avoid contention in vm fault path by checking the state before
1606 	 * issuing an atomic op on it.
1607 	 */
1608 	if ((object->flags & (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) !=
1609 	    (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) {
1610 		vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1611 	}
1612 	if (object->type == OBJT_VNODE &&
1613 	    (vp = (struct vnode *)object->handle) != NULL) {
1614 		if ((vp->v_flag & VOBJDIRTY) == 0) {
1615 			if (vp->v_mount &&
1616 			    (vp->v_mount->mnt_kern_flag & MNTK_THR_SYNC)) {
1617 				/*
1618 				 * New style THR_SYNC places vnodes on the
1619 				 * syncer list more deterministically.
1620 				 */
1621 				vsetobjdirty(vp);
1622 			} else {
1623 				/*
1624 				 * Old style scan would not necessarily place
1625 				 * a vnode on the syncer list when possibly
1626 				 * modified via mmap.
1627 				 */
1628 				vsetflags(vp, VOBJDIRTY);
1629 			}
1630 		}
1631 	}
1632 }
1633 
1634 #include "opt_ddb.h"
1635 #ifdef DDB
1636 #include <sys/cons.h>
1637 
1638 #include <ddb/ddb.h>
1639 
1640 static int	_vm_object_in_map (vm_map_t map, vm_object_t object,
1641 				       vm_map_entry_t entry);
1642 static int	vm_object_in_map (vm_object_t object);
1643 
1644 /*
1645  * The caller must hold the object.
1646  */
1647 static int
1648 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
1649 {
1650 	vm_map_backing_t ba;
1651 	vm_map_t tmpm;
1652 	vm_map_entry_t tmpe;
1653 	int entcount;
1654 
1655 	if (map == NULL)
1656 		return 0;
1657 	if (entry == NULL) {
1658 		tmpe = RB_MIN(vm_map_rb_tree, &map->rb_root);
1659 		entcount = map->nentries;
1660 		while (entcount-- && tmpe) {
1661 			if( _vm_object_in_map(map, object, tmpe)) {
1662 				return 1;
1663 			}
1664 			tmpe = vm_map_rb_tree_RB_NEXT(tmpe);
1665 		}
1666 		return (0);
1667 	}
1668 	switch(entry->maptype) {
1669 	case VM_MAPTYPE_SUBMAP:
1670 		tmpm = entry->ba.sub_map;
1671 		tmpe = RB_MIN(vm_map_rb_tree, &tmpm->rb_root);
1672 		entcount = tmpm->nentries;
1673 		while (entcount-- && tmpe) {
1674 			if( _vm_object_in_map(tmpm, object, tmpe)) {
1675 				return 1;
1676 			}
1677 			tmpe = vm_map_rb_tree_RB_NEXT(tmpe);
1678 		}
1679 		break;
1680 	case VM_MAPTYPE_NORMAL:
1681 	case VM_MAPTYPE_VPAGETABLE:
1682 		ba = &entry->ba;
1683 		while (ba) {
1684 			if (ba->object == object)
1685 				return TRUE;
1686 			ba = ba->backing_ba;
1687 		}
1688 		break;
1689 	default:
1690 		break;
1691 	}
1692 	return 0;
1693 }
1694 
1695 static int vm_object_in_map_callback(struct proc *p, void *data);
1696 
1697 struct vm_object_in_map_info {
1698 	vm_object_t object;
1699 	int rv;
1700 };
1701 
1702 /*
1703  * Debugging only
1704  */
1705 static int
1706 vm_object_in_map(vm_object_t object)
1707 {
1708 	struct vm_object_in_map_info info;
1709 
1710 	info.rv = 0;
1711 	info.object = object;
1712 
1713 	allproc_scan(vm_object_in_map_callback, &info, 0);
1714 	if (info.rv)
1715 		return 1;
1716 	if( _vm_object_in_map(&kernel_map, object, 0))
1717 		return 1;
1718 	if( _vm_object_in_map(&pager_map, object, 0))
1719 		return 1;
1720 	if( _vm_object_in_map(&buffer_map, object, 0))
1721 		return 1;
1722 	return 0;
1723 }
1724 
1725 /*
1726  * Debugging only
1727  */
1728 static int
1729 vm_object_in_map_callback(struct proc *p, void *data)
1730 {
1731 	struct vm_object_in_map_info *info = data;
1732 
1733 	if (p->p_vmspace) {
1734 		if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) {
1735 			info->rv = 1;
1736 			return -1;
1737 		}
1738 	}
1739 	return (0);
1740 }
1741 
1742 DB_SHOW_COMMAND(vmochk, vm_object_check)
1743 {
1744 	struct vm_object_hash *hash;
1745 	vm_object_t object;
1746 	int n;
1747 
1748 	/*
1749 	 * make sure that internal objs are in a map somewhere
1750 	 * and none have zero ref counts.
1751 	 */
1752 	for (n = 0; n < VMOBJ_HSIZE; ++n) {
1753 		hash = &vm_object_hash[n];
1754 		for (object = TAILQ_FIRST(&hash->list);
1755 				object != NULL;
1756 				object = TAILQ_NEXT(object, object_entry)) {
1757 			if (object->type == OBJT_MARKER)
1758 				continue;
1759 			if (object->handle != NULL ||
1760 			    (object->type != OBJT_DEFAULT &&
1761 			     object->type != OBJT_SWAP)) {
1762 				continue;
1763 			}
1764 			if (object->ref_count == 0) {
1765 				db_printf("vmochk: internal obj has "
1766 					  "zero ref count: %ld\n",
1767 					  (long)object->size);
1768 			}
1769 			if (vm_object_in_map(object))
1770 				continue;
1771 			db_printf("vmochk: internal obj is not in a map: "
1772 				  "ref: %d, size: %lu: 0x%lx\n",
1773 				  object->ref_count, (u_long)object->size,
1774 				  (u_long)object->size);
1775 		}
1776 	}
1777 }
1778 
1779 /*
1780  * Debugging only
1781  */
1782 DB_SHOW_COMMAND(object, vm_object_print_static)
1783 {
1784 	/* XXX convert args. */
1785 	vm_object_t object = (vm_object_t)addr;
1786 	boolean_t full = have_addr;
1787 
1788 	vm_page_t p;
1789 
1790 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
1791 #define	count	was_count
1792 
1793 	int count;
1794 
1795 	if (object == NULL)
1796 		return;
1797 
1798 	db_iprintf(
1799 	    "Object %p: type=%d, size=0x%lx, res=%ld, ref=%d, flags=0x%x\n",
1800 	    object, (int)object->type, (u_long)object->size,
1801 	    object->resident_page_count, object->ref_count, object->flags);
1802 	/*
1803 	 * XXX no %qd in kernel.  Truncate object->backing_object_offset.
1804 	 */
1805 	db_iprintf("\n");
1806 
1807 	if (!full)
1808 		return;
1809 
1810 	db_indent += 2;
1811 	count = 0;
1812 	RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) {
1813 		if (count == 0)
1814 			db_iprintf("memory:=");
1815 		else if (count == 6) {
1816 			db_printf("\n");
1817 			db_iprintf(" ...");
1818 			count = 0;
1819 		} else
1820 			db_printf(",");
1821 		count++;
1822 
1823 		db_printf("(off=0x%lx,page=0x%lx)",
1824 		    (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1825 	}
1826 	if (count != 0)
1827 		db_printf("\n");
1828 	db_indent -= 2;
1829 }
1830 
1831 /* XXX. */
1832 #undef count
1833 
1834 /*
1835  * XXX need this non-static entry for calling from vm_map_print.
1836  *
1837  * Debugging only
1838  */
1839 void
1840 vm_object_print(/* db_expr_t */ long addr,
1841 		boolean_t have_addr,
1842 		/* db_expr_t */ long count,
1843 		char *modif)
1844 {
1845 	vm_object_print_static(addr, have_addr, count, modif);
1846 }
1847 
1848 /*
1849  * Debugging only
1850  */
1851 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1852 {
1853 	struct vm_object_hash *hash;
1854 	vm_object_t object;
1855 	int nl = 0;
1856 	int c;
1857 	int n;
1858 
1859 	for (n = 0; n < VMOBJ_HSIZE; ++n) {
1860 		hash = &vm_object_hash[n];
1861 		for (object = TAILQ_FIRST(&hash->list);
1862 				object != NULL;
1863 				object = TAILQ_NEXT(object, object_entry)) {
1864 			vm_pindex_t idx, fidx;
1865 			vm_pindex_t osize;
1866 			vm_paddr_t pa = -1, padiff;
1867 			int rcount;
1868 			vm_page_t m;
1869 
1870 			if (object->type == OBJT_MARKER)
1871 				continue;
1872 			db_printf("new object: %p\n", (void *)object);
1873 			if ( nl > 18) {
1874 				c = cngetc();
1875 				if (c != ' ')
1876 					return;
1877 				nl = 0;
1878 			}
1879 			nl++;
1880 			rcount = 0;
1881 			fidx = 0;
1882 			osize = object->size;
1883 			if (osize > 128)
1884 				osize = 128;
1885 			for (idx = 0; idx < osize; idx++) {
1886 				m = vm_page_lookup(object, idx);
1887 				if (m == NULL) {
1888 					if (rcount) {
1889 						db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1890 							(long)fidx, rcount, (long)pa);
1891 						if ( nl > 18) {
1892 							c = cngetc();
1893 							if (c != ' ')
1894 								return;
1895 							nl = 0;
1896 						}
1897 						nl++;
1898 						rcount = 0;
1899 					}
1900 					continue;
1901 				}
1902 
1903 				if (rcount &&
1904 					(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1905 					++rcount;
1906 					continue;
1907 				}
1908 				if (rcount) {
1909 					padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1910 					padiff >>= PAGE_SHIFT;
1911 					padiff &= PQ_L2_MASK;
1912 					if (padiff == 0) {
1913 						pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1914 						++rcount;
1915 						continue;
1916 					}
1917 					db_printf(" index(%ld)run(%d)pa(0x%lx)",
1918 						(long)fidx, rcount, (long)pa);
1919 					db_printf("pd(%ld)\n", (long)padiff);
1920 					if ( nl > 18) {
1921 						c = cngetc();
1922 						if (c != ' ')
1923 							return;
1924 						nl = 0;
1925 					}
1926 					nl++;
1927 				}
1928 				fidx = idx;
1929 				pa = VM_PAGE_TO_PHYS(m);
1930 				rcount = 1;
1931 			}
1932 			if (rcount) {
1933 				db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1934 					(long)fidx, rcount, (long)pa);
1935 				if ( nl > 18) {
1936 					c = cngetc();
1937 					if (c != ' ')
1938 						return;
1939 					nl = 0;
1940 				}
1941 				nl++;
1942 			}
1943 		}
1944 	}
1945 }
1946 #endif /* DDB */
1947