xref: /dflybsd-src/sys/vm/vm_object.c (revision d2d1103f52e6fb116ee65a9940477c5449933f28)
1 /*
2  * (MPSAFE)
3  *
4  * Copyright (c) 1991, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *	This product includes software developed by the University of
21  *	California, Berkeley and its contributors.
22  * 4. Neither the name of the University nor the names of its contributors
23  *    may be used to endorse or promote products derived from this software
24  *    without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	from: @(#)vm_object.c	8.5 (Berkeley) 3/22/94
39  *
40  *
41  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42  * All rights reserved.
43  *
44  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
45  *
46  * Permission to use, copy, modify and distribute this software and
47  * its documentation is hereby granted, provided that both the copyright
48  * notice and this permission notice appear in all copies of the
49  * software, derivative works or modified versions, and any portions
50  * thereof, and that both notices appear in supporting documentation.
51  *
52  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
55  *
56  * Carnegie Mellon requests users of this software to return to
57  *
58  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
59  *  School of Computer Science
60  *  Carnegie Mellon University
61  *  Pittsburgh PA 15213-3890
62  *
63  * any improvements or extensions that they make and grant Carnegie the
64  * rights to redistribute these changes.
65  *
66  * $FreeBSD: src/sys/vm/vm_object.c,v 1.171.2.8 2003/05/26 19:17:56 alc Exp $
67  * $DragonFly: src/sys/vm/vm_object.c,v 1.33 2008/05/09 07:24:48 dillon Exp $
68  */
69 
70 /*
71  *	Virtual memory object module.
72  */
73 
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/proc.h>		/* for curproc, pageproc */
77 #include <sys/vnode.h>
78 #include <sys/vmmeter.h>
79 #include <sys/mman.h>
80 #include <sys/mount.h>
81 #include <sys/kernel.h>
82 #include <sys/sysctl.h>
83 
84 #include <vm/vm.h>
85 #include <vm/vm_param.h>
86 #include <vm/pmap.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_pager.h>
92 #include <vm/swap_pager.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
95 #include <vm/vm_zone.h>
96 
97 #define EASY_SCAN_FACTOR	8
98 
99 static void	vm_object_qcollapse(vm_object_t object);
100 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
101 					     int pagerflags);
102 
103 /*
104  *	Virtual memory objects maintain the actual data
105  *	associated with allocated virtual memory.  A given
106  *	page of memory exists within exactly one object.
107  *
108  *	An object is only deallocated when all "references"
109  *	are given up.  Only one "reference" to a given
110  *	region of an object should be writeable.
111  *
112  *	Associated with each object is a list of all resident
113  *	memory pages belonging to that object; this list is
114  *	maintained by the "vm_page" module, and locked by the object's
115  *	lock.
116  *
117  *	Each object also records a "pager" routine which is
118  *	used to retrieve (and store) pages to the proper backing
119  *	storage.  In addition, objects may be backed by other
120  *	objects from which they were virtual-copied.
121  *
122  *	The only items within the object structure which are
123  *	modified after time of creation are:
124  *		reference count		locked by object's lock
125  *		pager routine		locked by object's lock
126  *
127  */
128 
129 struct object_q vm_object_list;		/* locked by vmobj_token */
130 struct vm_object kernel_object;
131 
132 static long vm_object_count;		/* locked by vmobj_token */
133 extern int vm_pageout_page_count;
134 
135 static long object_collapses;
136 static long object_bypasses;
137 static int next_index;
138 static vm_zone_t obj_zone;
139 static struct vm_zone obj_zone_store;
140 static int object_hash_rand;
141 #define VM_OBJECTS_INIT 256
142 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
143 
144 /*
145  * Initialize a freshly allocated object
146  *
147  * Used only by vm_object_allocate() and zinitna().
148  *
149  * No requirements.
150  */
151 void
152 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
153 {
154 	int incr;
155 
156 	RB_INIT(&object->rb_memq);
157 	LIST_INIT(&object->shadow_head);
158 
159 	object->type = type;
160 	object->size = size;
161 	object->ref_count = 1;
162 	object->flags = 0;
163 	if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
164 		vm_object_set_flag(object, OBJ_ONEMAPPING);
165 	object->paging_in_progress = 0;
166 	object->resident_page_count = 0;
167 	object->shadow_count = 0;
168 	object->pg_color = next_index;
169 	if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
170 		incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
171 	else
172 		incr = size;
173 	next_index = (next_index + incr) & PQ_L2_MASK;
174 	object->handle = NULL;
175 	object->backing_object = NULL;
176 	object->backing_object_offset = (vm_ooffset_t) 0;
177 	/*
178 	 * Try to generate a number that will spread objects out in the
179 	 * hash table.  We 'wipe' new objects across the hash in 128 page
180 	 * increments plus 1 more to offset it a little more by the time
181 	 * it wraps around.
182 	 */
183 	object->hash_rand = object_hash_rand - 129;
184 
185 	object->generation++;
186 	object->swblock_count = 0;
187 	RB_INIT(&object->swblock_root);
188 
189 	lwkt_gettoken(&vmobj_token);
190 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
191 	vm_object_count++;
192 	object_hash_rand = object->hash_rand;
193 	lwkt_reltoken(&vmobj_token);
194 }
195 
196 /*
197  * Initialize the VM objects module.
198  *
199  * Called from the low level boot code only.
200  */
201 void
202 vm_object_init(void)
203 {
204 	TAILQ_INIT(&vm_object_list);
205 
206 	_vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd),
207 			    &kernel_object);
208 
209 	obj_zone = &obj_zone_store;
210 	zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
211 		vm_objects_init, VM_OBJECTS_INIT);
212 }
213 
214 void
215 vm_object_init2(void)
216 {
217 	zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1);
218 }
219 
220 /*
221  * Allocate and return a new object of the specified type and size.
222  *
223  * No requirements.
224  */
225 vm_object_t
226 vm_object_allocate(objtype_t type, vm_pindex_t size)
227 {
228 	vm_object_t result;
229 
230 	result = (vm_object_t) zalloc(obj_zone);
231 
232 	_vm_object_allocate(type, size, result);
233 
234 	return (result);
235 }
236 
237 /*
238  * Add an additional reference to a vm_object.
239  *
240  * Object passed by caller must be stable or caller must already
241  * hold vmobj_token to avoid races.
242  */
243 void
244 vm_object_reference(vm_object_t object)
245 {
246 	if (object) {
247 		lwkt_gettoken(&vmobj_token);
248 		object->ref_count++;
249 		if (object->type == OBJT_VNODE) {
250 			vref(object->handle);
251 			/* XXX what if the vnode is being destroyed? */
252 		}
253 		lwkt_reltoken(&vmobj_token);
254 	}
255 }
256 
257 void
258 vm_object_reference_locked(vm_object_t object)
259 {
260 	if (object) {
261 		ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
262 		object->ref_count++;
263 		if (object->type == OBJT_VNODE) {
264 			vref(object->handle);
265 			/* XXX what if the vnode is being destroyed? */
266 		}
267 	}
268 }
269 
270 /*
271  * Dereference an object and its underlying vnode.
272  *
273  * The caller must hold vmobj_token.
274  */
275 static void
276 vm_object_vndeallocate(vm_object_t object)
277 {
278 	struct vnode *vp = (struct vnode *) object->handle;
279 
280 	KASSERT(object->type == OBJT_VNODE,
281 	    ("vm_object_vndeallocate: not a vnode object"));
282 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
283 	ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
284 #ifdef INVARIANTS
285 	if (object->ref_count == 0) {
286 		vprint("vm_object_vndeallocate", vp);
287 		panic("vm_object_vndeallocate: bad object reference count");
288 	}
289 #endif
290 
291 	object->ref_count--;
292 	if (object->ref_count == 0)
293 		vclrflags(vp, VTEXT);
294 	vrele(vp);
295 }
296 
297 /*
298  * Release a reference to the specified object, gained either through a
299  * vm_object_allocate or a vm_object_reference call.  When all references
300  * are gone, storage associated with this object may be relinquished.
301  */
302 void
303 vm_object_deallocate(vm_object_t object)
304 {
305 	lwkt_gettoken(&vmobj_token);
306 	vm_object_deallocate_locked(object);
307 	lwkt_reltoken(&vmobj_token);
308 }
309 
310 void
311 vm_object_deallocate_locked(vm_object_t object)
312 {
313 	vm_object_t temp;
314 
315 	ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
316 
317 	while (object != NULL) {
318 		if (object->type == OBJT_VNODE) {
319 			vm_object_vndeallocate(object);
320 			break;
321 		}
322 
323 		if (object->ref_count == 0) {
324 			panic("vm_object_deallocate: object deallocated "
325 			      "too many times: %d", object->type);
326 		}
327 		if (object->ref_count > 2) {
328 			object->ref_count--;
329 			break;
330 		}
331 
332 		/*
333 		 * We currently need the vm_token from this point on, and
334 		 * we must recheck ref_count after acquiring it.
335 		 */
336 		lwkt_gettoken(&vm_token);
337 
338 		if (object->ref_count > 2) {
339 			object->ref_count--;
340 			lwkt_reltoken(&vm_token);
341 			break;
342 		}
343 
344 		/*
345 		 * Here on ref_count of one or two, which are special cases for
346 		 * objects.
347 		 */
348 		if ((object->ref_count == 2) && (object->shadow_count == 0)) {
349 			vm_object_set_flag(object, OBJ_ONEMAPPING);
350 			object->ref_count--;
351 			lwkt_reltoken(&vm_token);
352 			break;
353 		}
354 		if ((object->ref_count == 2) && (object->shadow_count == 1)) {
355 			object->ref_count--;
356 			if ((object->handle == NULL) &&
357 			    (object->type == OBJT_DEFAULT ||
358 			     object->type == OBJT_SWAP)) {
359 				vm_object_t robject;
360 
361 				robject = LIST_FIRST(&object->shadow_head);
362 				KASSERT(robject != NULL,
363 					("vm_object_deallocate: ref_count: "
364 					"%d, shadow_count: %d",
365 					object->ref_count,
366 					object->shadow_count));
367 
368 				if ((robject->handle == NULL) &&
369 				    (robject->type == OBJT_DEFAULT ||
370 				     robject->type == OBJT_SWAP)) {
371 
372 					robject->ref_count++;
373 
374 					while (
375 						robject->paging_in_progress ||
376 						object->paging_in_progress
377 					) {
378 						vm_object_pip_sleep(robject, "objde1");
379 						vm_object_pip_sleep(object, "objde2");
380 					}
381 
382 					if (robject->ref_count == 1) {
383 						robject->ref_count--;
384 						object = robject;
385 						goto doterm;
386 					}
387 
388 					object = robject;
389 					vm_object_collapse(object);
390 					lwkt_reltoken(&vm_token);
391 					continue;
392 				}
393 			}
394 			lwkt_reltoken(&vm_token);
395 			break;
396 		}
397 
398 		/*
399 		 * Normal dereferencing path
400 		 */
401 		object->ref_count--;
402 		if (object->ref_count != 0) {
403 			lwkt_reltoken(&vm_token);
404 			break;
405 		}
406 
407 		/*
408 		 * Termination path
409 		 */
410 doterm:
411 		temp = object->backing_object;
412 		if (temp) {
413 			LIST_REMOVE(object, shadow_list);
414 			temp->shadow_count--;
415 			temp->generation++;
416 			object->backing_object = NULL;
417 		}
418 		lwkt_reltoken(&vm_token);
419 
420 		/*
421 		 * Don't double-terminate, we could be in a termination
422 		 * recursion due to the terminate having to sync data
423 		 * to disk.
424 		 */
425 		if ((object->flags & OBJ_DEAD) == 0)
426 			vm_object_terminate(object);
427 		object = temp;
428 	}
429 }
430 
431 /*
432  * Destroy the specified object, freeing up related resources.
433  *
434  * The object must have zero references.
435  *
436  * The caller must be holding vmobj_token and properly interlock with
437  * OBJ_DEAD.
438  */
439 static int vm_object_terminate_callback(vm_page_t p, void *data);
440 
441 void
442 vm_object_terminate(vm_object_t object)
443 {
444 	/*
445 	 * Make sure no one uses us.  Once we set OBJ_DEAD we should be
446 	 * able to safely block.
447 	 */
448 	KKASSERT((object->flags & OBJ_DEAD) == 0);
449 	ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
450 	vm_object_set_flag(object, OBJ_DEAD);
451 
452 	/*
453 	 * Wait for the pageout daemon to be done with the object
454 	 */
455 	vm_object_pip_wait(object, "objtrm");
456 
457 	KASSERT(!object->paging_in_progress,
458 		("vm_object_terminate: pageout in progress"));
459 
460 	/*
461 	 * Clean and free the pages, as appropriate. All references to the
462 	 * object are gone, so we don't need to lock it.
463 	 */
464 	if (object->type == OBJT_VNODE) {
465 		struct vnode *vp;
466 
467 		/*
468 		 * Clean pages and flush buffers.
469 		 */
470 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
471 
472 		vp = (struct vnode *) object->handle;
473 		vinvalbuf(vp, V_SAVE, 0, 0);
474 	}
475 
476 	/*
477 	 * Wait for any I/O to complete, after which there had better not
478 	 * be any references left on the object.
479 	 */
480 	vm_object_pip_wait(object, "objtrm");
481 
482 	if (object->ref_count != 0) {
483 		panic("vm_object_terminate: object with references, "
484 		      "ref_count=%d", object->ref_count);
485 	}
486 
487 	/*
488 	 * Now free any remaining pages. For internal objects, this also
489 	 * removes them from paging queues. Don't free wired pages, just
490 	 * remove them from the object.
491 	 */
492 	lwkt_gettoken(&vm_token);
493 	vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
494 				vm_object_terminate_callback, NULL);
495 	lwkt_reltoken(&vm_token);
496 
497 	/*
498 	 * Let the pager know object is dead.
499 	 */
500 	vm_pager_deallocate(object);
501 
502 	/*
503 	 * Remove the object from the global object list.
504 	 *
505 	 * (we are holding vmobj_token)
506 	 */
507 	TAILQ_REMOVE(&vm_object_list, object, object_list);
508 	vm_object_count--;
509 	vm_object_dead_wakeup(object);
510 
511 	if (object->ref_count != 0) {
512 		panic("vm_object_terminate2: object with references, "
513 		      "ref_count=%d", object->ref_count);
514 	}
515 
516 	/*
517 	 * Free the space for the object.
518 	 */
519 	zfree(obj_zone, object);
520 }
521 
522 /*
523  * The caller must hold vm_token.
524  */
525 static int
526 vm_object_terminate_callback(vm_page_t p, void *data __unused)
527 {
528 	if (p->busy || (p->flags & PG_BUSY))
529 		panic("vm_object_terminate: freeing busy page %p", p);
530 	if (p->wire_count == 0) {
531 		vm_page_busy(p);
532 		vm_page_free(p);
533 		mycpu->gd_cnt.v_pfree++;
534 	} else {
535 		if (p->queue != PQ_NONE)
536 			kprintf("vm_object_terminate: Warning: Encountered wired page %p on queue %d\n", p, p->queue);
537 		vm_page_busy(p);
538 		vm_page_remove(p);
539 		vm_page_wakeup(p);
540 	}
541 	return(0);
542 }
543 
544 /*
545  * The object is dead but still has an object<->pager association.  Sleep
546  * and return.  The caller typically retests the association in a loop.
547  *
548  * Must be called with the vmobj_token held.
549  */
550 void
551 vm_object_dead_sleep(vm_object_t object, const char *wmesg)
552 {
553 	ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
554 	if (object->handle) {
555 		vm_object_set_flag(object, OBJ_DEADWNT);
556 		tsleep(object, 0, wmesg, 0);
557 		/* object may be invalid after this point */
558 	}
559 }
560 
561 /*
562  * Wakeup anyone waiting for the object<->pager disassociation on
563  * a dead object.
564  *
565  * Must be called with the vmobj_token held.
566  */
567 void
568 vm_object_dead_wakeup(vm_object_t object)
569 {
570 	ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
571 	if (object->flags & OBJ_DEADWNT) {
572 		vm_object_clear_flag(object, OBJ_DEADWNT);
573 		wakeup(object);
574 	}
575 }
576 
577 /*
578  * Clean all dirty pages in the specified range of object.  Leaves page
579  * on whatever queue it is currently on.   If NOSYNC is set then do not
580  * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
581  * leaving the object dirty.
582  *
583  * When stuffing pages asynchronously, allow clustering.  XXX we need a
584  * synchronous clustering mode implementation.
585  *
586  * Odd semantics: if start == end, we clean everything.
587  *
588  * The object must be locked? XXX
589  */
590 static int vm_object_page_clean_pass1(struct vm_page *p, void *data);
591 static int vm_object_page_clean_pass2(struct vm_page *p, void *data);
592 
593 void
594 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
595 		     int flags)
596 {
597 	struct rb_vm_page_scan_info info;
598 	struct vnode *vp;
599 	int wholescan;
600 	int pagerflags;
601 	int curgeneration;
602 
603 	lwkt_gettoken(&vm_token);
604 	if (object->type != OBJT_VNODE ||
605 	    (object->flags & OBJ_MIGHTBEDIRTY) == 0) {
606 		lwkt_reltoken(&vm_token);
607 		return;
608 	}
609 
610 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ?
611 			VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
612 	pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
613 
614 	vp = object->handle;
615 
616 	/*
617 	 * Interlock other major object operations.  This allows us to
618 	 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY.
619 	 */
620 	crit_enter();
621 	vm_object_set_flag(object, OBJ_CLEANING);
622 
623 	/*
624 	 * Handle 'entire object' case
625 	 */
626 	info.start_pindex = start;
627 	if (end == 0) {
628 		info.end_pindex = object->size - 1;
629 	} else {
630 		info.end_pindex = end - 1;
631 	}
632 	wholescan = (start == 0 && info.end_pindex == object->size - 1);
633 	info.limit = flags;
634 	info.pagerflags = pagerflags;
635 	info.object = object;
636 
637 	/*
638 	 * If cleaning the entire object do a pass to mark the pages read-only.
639 	 * If everything worked out ok, clear OBJ_WRITEABLE and
640 	 * OBJ_MIGHTBEDIRTY.
641 	 */
642 	if (wholescan) {
643 		info.error = 0;
644 		vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
645 					vm_object_page_clean_pass1, &info);
646 		if (info.error == 0) {
647 			vm_object_clear_flag(object,
648 					     OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
649 			if (object->type == OBJT_VNODE &&
650 			    (vp = (struct vnode *)object->handle) != NULL) {
651 				if (vp->v_flag & VOBJDIRTY)
652 					vclrflags(vp, VOBJDIRTY);
653 			}
654 		}
655 	}
656 
657 	/*
658 	 * Do a pass to clean all the dirty pages we find.
659 	 */
660 	do {
661 		info.error = 0;
662 		curgeneration = object->generation;
663 		vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
664 					vm_object_page_clean_pass2, &info);
665 	} while (info.error || curgeneration != object->generation);
666 
667 	vm_object_clear_flag(object, OBJ_CLEANING);
668 	crit_exit();
669 	lwkt_reltoken(&vm_token);
670 }
671 
672 /*
673  * The caller must hold vm_token.
674  */
675 static
676 int
677 vm_object_page_clean_pass1(struct vm_page *p, void *data)
678 {
679 	struct rb_vm_page_scan_info *info = data;
680 
681 	vm_page_flag_set(p, PG_CLEANCHK);
682 	if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
683 		info->error = 1;
684 	else
685 		vm_page_protect(p, VM_PROT_READ);	/* must not block */
686 	return(0);
687 }
688 
689 /*
690  * The caller must hold vm_token.
691  */
692 static
693 int
694 vm_object_page_clean_pass2(struct vm_page *p, void *data)
695 {
696 	struct rb_vm_page_scan_info *info = data;
697 	int n;
698 
699 	/*
700 	 * Do not mess with pages that were inserted after we started
701 	 * the cleaning pass.
702 	 */
703 	if ((p->flags & PG_CLEANCHK) == 0)
704 		return(0);
705 
706 	/*
707 	 * Before wasting time traversing the pmaps, check for trivial
708 	 * cases where the page cannot be dirty.
709 	 */
710 	if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) {
711 		KKASSERT((p->dirty & p->valid) == 0);
712 		return(0);
713 	}
714 
715 	/*
716 	 * Check whether the page is dirty or not.  The page has been set
717 	 * to be read-only so the check will not race a user dirtying the
718 	 * page.
719 	 */
720 	vm_page_test_dirty(p);
721 	if ((p->dirty & p->valid) == 0) {
722 		vm_page_flag_clear(p, PG_CLEANCHK);
723 		return(0);
724 	}
725 
726 	/*
727 	 * If we have been asked to skip nosync pages and this is a
728 	 * nosync page, skip it.  Note that the object flags were
729 	 * not cleared in this case (because pass1 will have returned an
730 	 * error), so we do not have to set them.
731 	 */
732 	if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
733 		vm_page_flag_clear(p, PG_CLEANCHK);
734 		return(0);
735 	}
736 
737 	/*
738 	 * Flush as many pages as we can.  PG_CLEANCHK will be cleared on
739 	 * the pages that get successfully flushed.  Set info->error if
740 	 * we raced an object modification.
741 	 */
742 	n = vm_object_page_collect_flush(info->object, p, info->pagerflags);
743 	if (n == 0)
744 		info->error = 1;
745 	return(0);
746 }
747 
748 /*
749  * Collect the specified page and nearby pages and flush them out.
750  * The number of pages flushed is returned.
751  *
752  * The caller must hold vm_token.
753  */
754 static int
755 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags)
756 {
757 	int runlen;
758 	int maxf;
759 	int chkb;
760 	int maxb;
761 	int i;
762 	int curgeneration;
763 	vm_pindex_t pi;
764 	vm_page_t maf[vm_pageout_page_count];
765 	vm_page_t mab[vm_pageout_page_count];
766 	vm_page_t ma[vm_pageout_page_count];
767 
768 	curgeneration = object->generation;
769 
770 	pi = p->pindex;
771 	while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
772 		if (object->generation != curgeneration) {
773 			return(0);
774 		}
775 	}
776 	KKASSERT(p->object == object && p->pindex == pi);
777 
778 	maxf = 0;
779 	for(i = 1; i < vm_pageout_page_count; i++) {
780 		vm_page_t tp;
781 
782 		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
783 			if ((tp->flags & PG_BUSY) ||
784 				((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
785 				 (tp->flags & PG_CLEANCHK) == 0) ||
786 				(tp->busy != 0))
787 				break;
788 			if((tp->queue - tp->pc) == PQ_CACHE) {
789 				vm_page_flag_clear(tp, PG_CLEANCHK);
790 				break;
791 			}
792 			vm_page_test_dirty(tp);
793 			if ((tp->dirty & tp->valid) == 0) {
794 				vm_page_flag_clear(tp, PG_CLEANCHK);
795 				break;
796 			}
797 			maf[ i - 1 ] = tp;
798 			maxf++;
799 			continue;
800 		}
801 		break;
802 	}
803 
804 	maxb = 0;
805 	chkb = vm_pageout_page_count -  maxf;
806 	if (chkb) {
807 		for(i = 1; i < chkb;i++) {
808 			vm_page_t tp;
809 
810 			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
811 				if ((tp->flags & PG_BUSY) ||
812 					((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
813 					 (tp->flags & PG_CLEANCHK) == 0) ||
814 					(tp->busy != 0))
815 					break;
816 				if((tp->queue - tp->pc) == PQ_CACHE) {
817 					vm_page_flag_clear(tp, PG_CLEANCHK);
818 					break;
819 				}
820 				vm_page_test_dirty(tp);
821 				if ((tp->dirty & tp->valid) == 0) {
822 					vm_page_flag_clear(tp, PG_CLEANCHK);
823 					break;
824 				}
825 				mab[ i - 1 ] = tp;
826 				maxb++;
827 				continue;
828 			}
829 			break;
830 		}
831 	}
832 
833 	for(i = 0; i < maxb; i++) {
834 		int index = (maxb - i) - 1;
835 		ma[index] = mab[i];
836 		vm_page_flag_clear(ma[index], PG_CLEANCHK);
837 	}
838 	vm_page_flag_clear(p, PG_CLEANCHK);
839 	ma[maxb] = p;
840 	for(i = 0; i < maxf; i++) {
841 		int index = (maxb + i) + 1;
842 		ma[index] = maf[i];
843 		vm_page_flag_clear(ma[index], PG_CLEANCHK);
844 	}
845 	runlen = maxb + maxf + 1;
846 
847 	vm_pageout_flush(ma, runlen, pagerflags);
848 	for (i = 0; i < runlen; i++) {
849 		if (ma[i]->valid & ma[i]->dirty) {
850 			vm_page_protect(ma[i], VM_PROT_READ);
851 			vm_page_flag_set(ma[i], PG_CLEANCHK);
852 
853 			/*
854 			 * maxf will end up being the actual number of pages
855 			 * we wrote out contiguously, non-inclusive of the
856 			 * first page.  We do not count look-behind pages.
857 			 */
858 			if (i >= maxb + 1 && (maxf > i - maxb - 1))
859 				maxf = i - maxb - 1;
860 		}
861 	}
862 	return(maxf + 1);
863 }
864 
865 /*
866  * Same as vm_object_pmap_copy, except range checking really
867  * works, and is meant for small sections of an object.
868  *
869  * This code protects resident pages by making them read-only
870  * and is typically called on a fork or split when a page
871  * is converted to copy-on-write.
872  *
873  * NOTE: If the page is already at VM_PROT_NONE, calling
874  * vm_page_protect will have no effect.
875  */
876 void
877 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
878 {
879 	vm_pindex_t idx;
880 	vm_page_t p;
881 
882 	if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
883 		return;
884 
885 	/*
886 	 * spl protection needed to prevent races between the lookup,
887 	 * an interrupt unbusy/free, and our protect call.
888 	 */
889 	crit_enter();
890 	lwkt_gettoken(&vm_token);
891 	for (idx = start; idx < end; idx++) {
892 		p = vm_page_lookup(object, idx);
893 		if (p == NULL)
894 			continue;
895 		vm_page_protect(p, VM_PROT_READ);
896 	}
897 	lwkt_reltoken(&vm_token);
898 	crit_exit();
899 }
900 
901 /*
902  * Removes all physical pages in the specified object range from all
903  * physical maps.
904  *
905  * The object must *not* be locked.
906  */
907 
908 static int vm_object_pmap_remove_callback(vm_page_t p, void *data);
909 
910 void
911 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
912 {
913 	struct rb_vm_page_scan_info info;
914 
915 	if (object == NULL)
916 		return;
917 	info.start_pindex = start;
918 	info.end_pindex = end - 1;
919 
920 	crit_enter();
921 	lwkt_gettoken(&vm_token);
922 	vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
923 				vm_object_pmap_remove_callback, &info);
924 	if (start == 0 && end == object->size)
925 		vm_object_clear_flag(object, OBJ_WRITEABLE);
926 	lwkt_reltoken(&vm_token);
927 	crit_exit();
928 }
929 
930 /*
931  * The caller must hold vm_token.
932  */
933 static int
934 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused)
935 {
936 	vm_page_protect(p, VM_PROT_NONE);
937 	return(0);
938 }
939 
940 /*
941  * Implements the madvise function at the object/page level.
942  *
943  * MADV_WILLNEED	(any object)
944  *
945  *	Activate the specified pages if they are resident.
946  *
947  * MADV_DONTNEED	(any object)
948  *
949  *	Deactivate the specified pages if they are resident.
950  *
951  * MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only)
952  *
953  *	Deactivate and clean the specified pages if they are
954  *	resident.  This permits the process to reuse the pages
955  *	without faulting or the kernel to reclaim the pages
956  *	without I/O.
957  *
958  * No requirements.
959  */
960 void
961 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
962 {
963 	vm_pindex_t end, tpindex;
964 	vm_object_t tobject;
965 	vm_page_t m;
966 
967 	if (object == NULL)
968 		return;
969 
970 	end = pindex + count;
971 
972 	lwkt_gettoken(&vm_token);
973 
974 	/*
975 	 * Locate and adjust resident pages
976 	 */
977 	for (; pindex < end; pindex += 1) {
978 relookup:
979 		tobject = object;
980 		tpindex = pindex;
981 shadowlookup:
982 		/*
983 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
984 		 * and those pages must be OBJ_ONEMAPPING.
985 		 */
986 		if (advise == MADV_FREE) {
987 			if ((tobject->type != OBJT_DEFAULT &&
988 			     tobject->type != OBJT_SWAP) ||
989 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
990 				continue;
991 			}
992 		}
993 
994 		/*
995 		 * spl protection is required to avoid a race between the
996 		 * lookup, an interrupt unbusy/free, and our busy check.
997 		 */
998 
999 		crit_enter();
1000 		m = vm_page_lookup(tobject, tpindex);
1001 
1002 		if (m == NULL) {
1003 			/*
1004 			 * There may be swap even if there is no backing page
1005 			 */
1006 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1007 				swap_pager_freespace(tobject, tpindex, 1);
1008 
1009 			/*
1010 			 * next object
1011 			 */
1012 			crit_exit();
1013 			if (tobject->backing_object == NULL)
1014 				continue;
1015 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1016 			tobject = tobject->backing_object;
1017 			goto shadowlookup;
1018 		}
1019 
1020 		/*
1021 		 * If the page is busy or not in a normal active state,
1022 		 * we skip it.  If the page is not managed there are no
1023 		 * page queues to mess with.  Things can break if we mess
1024 		 * with pages in any of the below states.
1025 		 */
1026 		if (
1027 		    m->hold_count ||
1028 		    m->wire_count ||
1029 		    (m->flags & PG_UNMANAGED) ||
1030 		    m->valid != VM_PAGE_BITS_ALL
1031 		) {
1032 			crit_exit();
1033 			continue;
1034 		}
1035 
1036  		if (vm_page_sleep_busy(m, TRUE, "madvpo")) {
1037 			crit_exit();
1038   			goto relookup;
1039 		}
1040 		crit_exit();
1041 
1042 		/*
1043 		 * Theoretically once a page is known not to be busy, an
1044 		 * interrupt cannot come along and rip it out from under us.
1045 		 */
1046 
1047 		if (advise == MADV_WILLNEED) {
1048 			vm_page_activate(m);
1049 		} else if (advise == MADV_DONTNEED) {
1050 			vm_page_dontneed(m);
1051 		} else if (advise == MADV_FREE) {
1052 			/*
1053 			 * Mark the page clean.  This will allow the page
1054 			 * to be freed up by the system.  However, such pages
1055 			 * are often reused quickly by malloc()/free()
1056 			 * so we do not do anything that would cause
1057 			 * a page fault if we can help it.
1058 			 *
1059 			 * Specifically, we do not try to actually free
1060 			 * the page now nor do we try to put it in the
1061 			 * cache (which would cause a page fault on reuse).
1062 			 *
1063 			 * But we do make the page is freeable as we
1064 			 * can without actually taking the step of unmapping
1065 			 * it.
1066 			 */
1067 			pmap_clear_modify(m);
1068 			m->dirty = 0;
1069 			m->act_count = 0;
1070 			vm_page_dontneed(m);
1071 			if (tobject->type == OBJT_SWAP)
1072 				swap_pager_freespace(tobject, tpindex, 1);
1073 		}
1074 	}
1075 	lwkt_reltoken(&vm_token);
1076 }
1077 
1078 /*
1079  * Create a new object which is backed by the specified existing object
1080  * range.  The source object reference is deallocated.
1081  *
1082  * The new object and offset into that object are returned in the source
1083  * parameters.
1084  *
1085  * No other requirements.
1086  */
1087 void
1088 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length)
1089 {
1090 	vm_object_t source;
1091 	vm_object_t result;
1092 
1093 	source = *object;
1094 
1095 	/*
1096 	 * Don't create the new object if the old object isn't shared.
1097 	 */
1098 	lwkt_gettoken(&vm_token);
1099 
1100 	if (source != NULL &&
1101 	    source->ref_count == 1 &&
1102 	    source->handle == NULL &&
1103 	    (source->type == OBJT_DEFAULT ||
1104 	     source->type == OBJT_SWAP)) {
1105 		lwkt_reltoken(&vm_token);
1106 		return;
1107 	}
1108 
1109 	/*
1110 	 * Allocate a new object with the given length
1111 	 */
1112 
1113 	if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1114 		panic("vm_object_shadow: no object for shadowing");
1115 
1116 	/*
1117 	 * The new object shadows the source object, adding a reference to it.
1118 	 * Our caller changes his reference to point to the new object,
1119 	 * removing a reference to the source object.  Net result: no change
1120 	 * of reference count.
1121 	 *
1122 	 * Try to optimize the result object's page color when shadowing
1123 	 * in order to maintain page coloring consistency in the combined
1124 	 * shadowed object.
1125 	 */
1126 	result->backing_object = source;
1127 	if (source) {
1128 		LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1129 		source->shadow_count++;
1130 		source->generation++;
1131 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1132 	}
1133 
1134 	/*
1135 	 * Store the offset into the source object, and fix up the offset into
1136 	 * the new object.
1137 	 */
1138 	result->backing_object_offset = *offset;
1139 	lwkt_reltoken(&vm_token);
1140 
1141 	/*
1142 	 * Return the new things
1143 	 */
1144 	*offset = 0;
1145 	*object = result;
1146 }
1147 
1148 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1149 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1150 #define	OBSC_COLLAPSE_WAIT	0x0004
1151 
1152 static int vm_object_backing_scan_callback(vm_page_t p, void *data);
1153 
1154 /*
1155  * The caller must hold vm_token.
1156  */
1157 static __inline int
1158 vm_object_backing_scan(vm_object_t object, int op)
1159 {
1160 	struct rb_vm_page_scan_info info;
1161 	vm_object_t backing_object;
1162 
1163 	crit_enter();
1164 
1165 	backing_object = object->backing_object;
1166 	info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1167 
1168 	/*
1169 	 * Initial conditions
1170 	 */
1171 
1172 	if (op & OBSC_TEST_ALL_SHADOWED) {
1173 		/*
1174 		 * We do not want to have to test for the existence of
1175 		 * swap pages in the backing object.  XXX but with the
1176 		 * new swapper this would be pretty easy to do.
1177 		 *
1178 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1179 		 * been ZFOD faulted yet?  If we do not test for this, the
1180 		 * shadow test may succeed! XXX
1181 		 */
1182 		if (backing_object->type != OBJT_DEFAULT) {
1183 			crit_exit();
1184 			return(0);
1185 		}
1186 	}
1187 	if (op & OBSC_COLLAPSE_WAIT) {
1188 		KKASSERT((backing_object->flags & OBJ_DEAD) == 0);
1189 		vm_object_set_flag(backing_object, OBJ_DEAD);
1190 	}
1191 
1192 	/*
1193 	 * Our scan.   We have to retry if a negative error code is returned,
1194 	 * otherwise 0 or 1 will be returned in info.error.  0 Indicates that
1195 	 * the scan had to be stopped because the parent does not completely
1196 	 * shadow the child.
1197 	 */
1198 	info.object = object;
1199 	info.backing_object = backing_object;
1200 	info.limit = op;
1201 	do {
1202 		info.error = 1;
1203 		vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL,
1204 					vm_object_backing_scan_callback,
1205 					&info);
1206 	} while (info.error < 0);
1207 	crit_exit();
1208 	return(info.error);
1209 }
1210 
1211 /*
1212  * The caller must hold vm_token.
1213  */
1214 static int
1215 vm_object_backing_scan_callback(vm_page_t p, void *data)
1216 {
1217 	struct rb_vm_page_scan_info *info = data;
1218 	vm_object_t backing_object;
1219 	vm_object_t object;
1220 	vm_pindex_t new_pindex;
1221 	vm_pindex_t backing_offset_index;
1222 	int op;
1223 
1224 	new_pindex = p->pindex - info->backing_offset_index;
1225 	op = info->limit;
1226 	object = info->object;
1227 	backing_object = info->backing_object;
1228 	backing_offset_index = info->backing_offset_index;
1229 
1230 	if (op & OBSC_TEST_ALL_SHADOWED) {
1231 		vm_page_t pp;
1232 
1233 		/*
1234 		 * Ignore pages outside the parent object's range
1235 		 * and outside the parent object's mapping of the
1236 		 * backing object.
1237 		 *
1238 		 * note that we do not busy the backing object's
1239 		 * page.
1240 		 */
1241 		if (
1242 		    p->pindex < backing_offset_index ||
1243 		    new_pindex >= object->size
1244 		) {
1245 			return(0);
1246 		}
1247 
1248 		/*
1249 		 * See if the parent has the page or if the parent's
1250 		 * object pager has the page.  If the parent has the
1251 		 * page but the page is not valid, the parent's
1252 		 * object pager must have the page.
1253 		 *
1254 		 * If this fails, the parent does not completely shadow
1255 		 * the object and we might as well give up now.
1256 		 */
1257 
1258 		pp = vm_page_lookup(object, new_pindex);
1259 		if ((pp == NULL || pp->valid == 0) &&
1260 		    !vm_pager_has_page(object, new_pindex)
1261 		) {
1262 			info->error = 0;	/* problemo */
1263 			return(-1);		/* stop the scan */
1264 		}
1265 	}
1266 
1267 	/*
1268 	 * Check for busy page
1269 	 */
1270 
1271 	if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1272 		vm_page_t pp;
1273 
1274 		if (op & OBSC_COLLAPSE_NOWAIT) {
1275 			if (
1276 			    (p->flags & PG_BUSY) ||
1277 			    !p->valid ||
1278 			    p->hold_count ||
1279 			    p->wire_count ||
1280 			    p->busy
1281 			) {
1282 				return(0);
1283 			}
1284 		} else if (op & OBSC_COLLAPSE_WAIT) {
1285 			if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1286 				/*
1287 				 * If we slept, anything could have
1288 				 * happened.   Ask that the scan be restarted.
1289 				 *
1290 				 * Since the object is marked dead, the
1291 				 * backing offset should not have changed.
1292 				 */
1293 				info->error = -1;
1294 				return(-1);
1295 			}
1296 		}
1297 
1298 		/*
1299 		 * Busy the page
1300 		 */
1301 		vm_page_busy(p);
1302 
1303 		KASSERT(
1304 		    p->object == backing_object,
1305 		    ("vm_object_qcollapse(): object mismatch")
1306 		);
1307 
1308 		/*
1309 		 * Destroy any associated swap
1310 		 */
1311 		if (backing_object->type == OBJT_SWAP)
1312 			swap_pager_freespace(backing_object, p->pindex, 1);
1313 
1314 		if (
1315 		    p->pindex < backing_offset_index ||
1316 		    new_pindex >= object->size
1317 		) {
1318 			/*
1319 			 * Page is out of the parent object's range, we
1320 			 * can simply destroy it.
1321 			 */
1322 			vm_page_protect(p, VM_PROT_NONE);
1323 			vm_page_free(p);
1324 			return(0);
1325 		}
1326 
1327 		pp = vm_page_lookup(object, new_pindex);
1328 		if (pp != NULL || vm_pager_has_page(object, new_pindex)) {
1329 			/*
1330 			 * page already exists in parent OR swap exists
1331 			 * for this location in the parent.  Destroy
1332 			 * the original page from the backing object.
1333 			 *
1334 			 * Leave the parent's page alone
1335 			 */
1336 			vm_page_protect(p, VM_PROT_NONE);
1337 			vm_page_free(p);
1338 			return(0);
1339 		}
1340 
1341 		/*
1342 		 * Page does not exist in parent, rename the
1343 		 * page from the backing object to the main object.
1344 		 *
1345 		 * If the page was mapped to a process, it can remain
1346 		 * mapped through the rename.
1347 		 */
1348 		if ((p->queue - p->pc) == PQ_CACHE)
1349 			vm_page_deactivate(p);
1350 
1351 		vm_page_rename(p, object, new_pindex);
1352 		/* page automatically made dirty by rename */
1353 	}
1354 	return(0);
1355 }
1356 
1357 /*
1358  * This version of collapse allows the operation to occur earlier and
1359  * when paging_in_progress is true for an object...  This is not a complete
1360  * operation, but should plug 99.9% of the rest of the leaks.
1361  *
1362  * The caller must hold vm_token and vmobj_token.
1363  * (only called from vm_object_collapse)
1364  */
1365 static void
1366 vm_object_qcollapse(vm_object_t object)
1367 {
1368 	vm_object_t backing_object = object->backing_object;
1369 
1370 	if (backing_object->ref_count != 1)
1371 		return;
1372 
1373 	backing_object->ref_count += 2;
1374 
1375 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1376 
1377 	backing_object->ref_count -= 2;
1378 }
1379 
1380 /*
1381  * Collapse an object with the object backing it.  Pages in the backing
1382  * object are moved into the parent, and the backing object is deallocated.
1383  */
1384 void
1385 vm_object_collapse(vm_object_t object)
1386 {
1387 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
1388 	ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
1389 
1390 	while (TRUE) {
1391 		vm_object_t backing_object;
1392 
1393 		/*
1394 		 * Verify that the conditions are right for collapse:
1395 		 *
1396 		 * The object exists and the backing object exists.
1397 		 */
1398 		if (object == NULL)
1399 			break;
1400 
1401 		if ((backing_object = object->backing_object) == NULL)
1402 			break;
1403 
1404 		/*
1405 		 * we check the backing object first, because it is most likely
1406 		 * not collapsable.
1407 		 */
1408 		if (backing_object->handle != NULL ||
1409 		    (backing_object->type != OBJT_DEFAULT &&
1410 		     backing_object->type != OBJT_SWAP) ||
1411 		    (backing_object->flags & OBJ_DEAD) ||
1412 		    object->handle != NULL ||
1413 		    (object->type != OBJT_DEFAULT &&
1414 		     object->type != OBJT_SWAP) ||
1415 		    (object->flags & OBJ_DEAD)) {
1416 			break;
1417 		}
1418 
1419 		if (
1420 		    object->paging_in_progress != 0 ||
1421 		    backing_object->paging_in_progress != 0
1422 		) {
1423 			vm_object_qcollapse(object);
1424 			break;
1425 		}
1426 
1427 		/*
1428 		 * We know that we can either collapse the backing object (if
1429 		 * the parent is the only reference to it) or (perhaps) have
1430 		 * the parent bypass the object if the parent happens to shadow
1431 		 * all the resident pages in the entire backing object.
1432 		 *
1433 		 * This is ignoring pager-backed pages such as swap pages.
1434 		 * vm_object_backing_scan fails the shadowing test in this
1435 		 * case.
1436 		 */
1437 
1438 		if (backing_object->ref_count == 1) {
1439 			/*
1440 			 * If there is exactly one reference to the backing
1441 			 * object, we can collapse it into the parent.
1442 			 */
1443 			vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1444 
1445 			/*
1446 			 * Move the pager from backing_object to object.
1447 			 */
1448 
1449 			if (backing_object->type == OBJT_SWAP) {
1450 				vm_object_pip_add(backing_object, 1);
1451 
1452 				/*
1453 				 * scrap the paging_offset junk and do a
1454 				 * discrete copy.  This also removes major
1455 				 * assumptions about how the swap-pager
1456 				 * works from where it doesn't belong.  The
1457 				 * new swapper is able to optimize the
1458 				 * destroy-source case.
1459 				 */
1460 
1461 				vm_object_pip_add(object, 1);
1462 				swap_pager_copy(
1463 				    backing_object,
1464 				    object,
1465 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1466 				vm_object_pip_wakeup(object);
1467 
1468 				vm_object_pip_wakeup(backing_object);
1469 			}
1470 			/*
1471 			 * Object now shadows whatever backing_object did.
1472 			 * Note that the reference to
1473 			 * backing_object->backing_object moves from within
1474 			 * backing_object to within object.
1475 			 */
1476 
1477 			LIST_REMOVE(object, shadow_list);
1478 			object->backing_object->shadow_count--;
1479 			object->backing_object->generation++;
1480 			if (backing_object->backing_object) {
1481 				LIST_REMOVE(backing_object, shadow_list);
1482 				backing_object->backing_object->shadow_count--;
1483 				backing_object->backing_object->generation++;
1484 			}
1485 			object->backing_object = backing_object->backing_object;
1486 			if (object->backing_object) {
1487 				LIST_INSERT_HEAD(
1488 				    &object->backing_object->shadow_head,
1489 				    object,
1490 				    shadow_list
1491 				);
1492 				object->backing_object->shadow_count++;
1493 				object->backing_object->generation++;
1494 			}
1495 
1496 			object->backing_object_offset +=
1497 			    backing_object->backing_object_offset;
1498 
1499 			/*
1500 			 * Discard backing_object.
1501 			 *
1502 			 * Since the backing object has no pages, no pager left,
1503 			 * and no object references within it, all that is
1504 			 * necessary is to dispose of it.
1505 			 */
1506 
1507 			KASSERT(backing_object->ref_count == 1,
1508 				("backing_object %p was somehow "
1509 				 "re-referenced during collapse!",
1510 				 backing_object));
1511 			KASSERT(RB_EMPTY(&backing_object->rb_memq),
1512 				("backing_object %p somehow has left "
1513 				 "over pages during collapse!",
1514 				 backing_object));
1515 
1516 			/* (we are holding vmobj_token) */
1517 			TAILQ_REMOVE(&vm_object_list, backing_object,
1518 				     object_list);
1519 			vm_object_count--;
1520 
1521 			zfree(obj_zone, backing_object);
1522 
1523 			object_collapses++;
1524 		} else {
1525 			vm_object_t new_backing_object;
1526 
1527 			/*
1528 			 * If we do not entirely shadow the backing object,
1529 			 * there is nothing we can do so we give up.
1530 			 */
1531 
1532 			if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1533 				break;
1534 			}
1535 
1536 			/*
1537 			 * Make the parent shadow the next object in the
1538 			 * chain.  Deallocating backing_object will not remove
1539 			 * it, since its reference count is at least 2.
1540 			 */
1541 
1542 			LIST_REMOVE(object, shadow_list);
1543 			backing_object->shadow_count--;
1544 			backing_object->generation++;
1545 
1546 			new_backing_object = backing_object->backing_object;
1547 			if ((object->backing_object = new_backing_object) != NULL) {
1548 				vm_object_reference(new_backing_object);
1549 				LIST_INSERT_HEAD(
1550 				    &new_backing_object->shadow_head,
1551 				    object,
1552 				    shadow_list
1553 				);
1554 				new_backing_object->shadow_count++;
1555 				new_backing_object->generation++;
1556 				object->backing_object_offset +=
1557 					backing_object->backing_object_offset;
1558 			}
1559 
1560 			/*
1561 			 * Drop the reference count on backing_object. Since
1562 			 * its ref_count was at least 2, it will not vanish;
1563 			 * so we don't need to call vm_object_deallocate, but
1564 			 * we do anyway.
1565 			 */
1566 			vm_object_deallocate_locked(backing_object);
1567 			object_bypasses++;
1568 		}
1569 
1570 		/*
1571 		 * Try again with this object's new backing object.
1572 		 */
1573 	}
1574 }
1575 
1576 /*
1577  * Removes all physical pages in the specified object range from the
1578  * object's list of pages.
1579  *
1580  * No requirements.
1581  */
1582 static int vm_object_page_remove_callback(vm_page_t p, void *data);
1583 
1584 void
1585 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1586 		      boolean_t clean_only)
1587 {
1588 	struct rb_vm_page_scan_info info;
1589 	int all;
1590 
1591 	/*
1592 	 * Degenerate cases and assertions
1593 	 */
1594 	lwkt_gettoken(&vm_token);
1595 	if (object == NULL ||
1596 	    (object->resident_page_count == 0 && object->swblock_count == 0)) {
1597 		lwkt_reltoken(&vm_token);
1598 		return;
1599 	}
1600 	KASSERT(object->type != OBJT_PHYS,
1601 		("attempt to remove pages from a physical object"));
1602 
1603 	/*
1604 	 * Indicate that paging is occuring on the object
1605 	 */
1606 	crit_enter();
1607 	vm_object_pip_add(object, 1);
1608 
1609 	/*
1610 	 * Figure out the actual removal range and whether we are removing
1611 	 * the entire contents of the object or not.  If removing the entire
1612 	 * contents, be sure to get all pages, even those that might be
1613 	 * beyond the end of the object.
1614 	 */
1615 	info.start_pindex = start;
1616 	if (end == 0)
1617 		info.end_pindex = (vm_pindex_t)-1;
1618 	else
1619 		info.end_pindex = end - 1;
1620 	info.limit = clean_only;
1621 	all = (start == 0 && info.end_pindex >= object->size - 1);
1622 
1623 	/*
1624 	 * Loop until we are sure we have gotten them all.
1625 	 */
1626 	do {
1627 		info.error = 0;
1628 		vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1629 					vm_object_page_remove_callback, &info);
1630 	} while (info.error);
1631 
1632 	/*
1633 	 * Remove any related swap if throwing away pages, or for
1634 	 * non-swap objects (the swap is a clean copy in that case).
1635 	 */
1636 	if (object->type != OBJT_SWAP || clean_only == FALSE) {
1637 		if (all)
1638 			swap_pager_freespace_all(object);
1639 		else
1640 			swap_pager_freespace(object, info.start_pindex,
1641 			     info.end_pindex - info.start_pindex + 1);
1642 	}
1643 
1644 	/*
1645 	 * Cleanup
1646 	 */
1647 	vm_object_pip_wakeup(object);
1648 	crit_exit();
1649 	lwkt_reltoken(&vm_token);
1650 }
1651 
1652 /*
1653  * The caller must hold vm_token.
1654  */
1655 static int
1656 vm_object_page_remove_callback(vm_page_t p, void *data)
1657 {
1658 	struct rb_vm_page_scan_info *info = data;
1659 
1660 	/*
1661 	 * Wired pages cannot be destroyed, but they can be invalidated
1662 	 * and we do so if clean_only (limit) is not set.
1663 	 *
1664 	 * WARNING!  The page may be wired due to being part of a buffer
1665 	 *	     cache buffer, and the buffer might be marked B_CACHE.
1666 	 *	     This is fine as part of a truncation but VFSs must be
1667 	 *	     sure to fix the buffer up when re-extending the file.
1668 	 */
1669 	if (p->wire_count != 0) {
1670 		vm_page_protect(p, VM_PROT_NONE);
1671 		if (info->limit == 0)
1672 			p->valid = 0;
1673 		return(0);
1674 	}
1675 
1676 	/*
1677 	 * The busy flags are only cleared at
1678 	 * interrupt -- minimize the spl transitions
1679 	 */
1680 
1681 	if (vm_page_sleep_busy(p, TRUE, "vmopar")) {
1682 		info->error = 1;
1683 		return(0);
1684 	}
1685 
1686 	/*
1687 	 * limit is our clean_only flag.  If set and the page is dirty, do
1688 	 * not free it.  If set and the page is being held by someone, do
1689 	 * not free it.
1690 	 */
1691 	if (info->limit && p->valid) {
1692 		vm_page_test_dirty(p);
1693 		if (p->valid & p->dirty)
1694 			return(0);
1695 		if (p->hold_count)
1696 			return(0);
1697 	}
1698 
1699 	/*
1700 	 * Destroy the page
1701 	 */
1702 	vm_page_busy(p);
1703 	vm_page_protect(p, VM_PROT_NONE);
1704 	vm_page_free(p);
1705 	return(0);
1706 }
1707 
1708 /*
1709  * Coalesces two objects backing up adjoining regions of memory into a
1710  * single object.
1711  *
1712  * returns TRUE if objects were combined.
1713  *
1714  * NOTE: Only works at the moment if the second object is NULL -
1715  *	 if it's not, which object do we lock first?
1716  *
1717  * Parameters:
1718  *	prev_object	First object to coalesce
1719  *	prev_offset	Offset into prev_object
1720  *	next_object	Second object into coalesce
1721  *	next_offset	Offset into next_object
1722  *
1723  *	prev_size	Size of reference to prev_object
1724  *	next_size	Size of reference to next_object
1725  *
1726  * The object must not be locked.
1727  * The caller must hold vm_token and vmobj_token.
1728  */
1729 boolean_t
1730 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
1731 		   vm_size_t prev_size, vm_size_t next_size)
1732 {
1733 	vm_pindex_t next_pindex;
1734 
1735 	ASSERT_LWKT_TOKEN_HELD(&vm_token);
1736 	ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
1737 
1738 	if (prev_object == NULL) {
1739 		return (TRUE);
1740 	}
1741 
1742 	if (prev_object->type != OBJT_DEFAULT &&
1743 	    prev_object->type != OBJT_SWAP) {
1744 		return (FALSE);
1745 	}
1746 
1747 	/*
1748 	 * Try to collapse the object first
1749 	 */
1750 	vm_object_collapse(prev_object);
1751 
1752 	/*
1753 	 * Can't coalesce if: . more than one reference . paged out . shadows
1754 	 * another object . has a copy elsewhere (any of which mean that the
1755 	 * pages not mapped to prev_entry may be in use anyway)
1756 	 */
1757 
1758 	if (prev_object->backing_object != NULL)
1759 		return (FALSE);
1760 
1761 	prev_size >>= PAGE_SHIFT;
1762 	next_size >>= PAGE_SHIFT;
1763 	next_pindex = prev_pindex + prev_size;
1764 
1765 	if ((prev_object->ref_count > 1) &&
1766 	    (prev_object->size != next_pindex)) {
1767 		return (FALSE);
1768 	}
1769 
1770 	/*
1771 	 * Remove any pages that may still be in the object from a previous
1772 	 * deallocation.
1773 	 */
1774 	if (next_pindex < prev_object->size) {
1775 		vm_object_page_remove(prev_object,
1776 				      next_pindex,
1777 				      next_pindex + next_size, FALSE);
1778 		if (prev_object->type == OBJT_SWAP)
1779 			swap_pager_freespace(prev_object,
1780 					     next_pindex, next_size);
1781 	}
1782 
1783 	/*
1784 	 * Extend the object if necessary.
1785 	 */
1786 	if (next_pindex + next_size > prev_object->size)
1787 		prev_object->size = next_pindex + next_size;
1788 	return (TRUE);
1789 }
1790 
1791 /*
1792  * Make the object writable and flag is being possibly dirty.
1793  *
1794  * No requirements.
1795  */
1796 void
1797 vm_object_set_writeable_dirty(vm_object_t object)
1798 {
1799 	struct vnode *vp;
1800 
1801 	lwkt_gettoken(&vm_token);
1802 	vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1803 	if (object->type == OBJT_VNODE &&
1804 	    (vp = (struct vnode *)object->handle) != NULL) {
1805 		if ((vp->v_flag & VOBJDIRTY) == 0) {
1806 			vsetflags(vp, VOBJDIRTY);
1807 		}
1808 	}
1809 	lwkt_reltoken(&vm_token);
1810 }
1811 
1812 #include "opt_ddb.h"
1813 #ifdef DDB
1814 #include <sys/kernel.h>
1815 
1816 #include <sys/cons.h>
1817 
1818 #include <ddb/ddb.h>
1819 
1820 static int	_vm_object_in_map (vm_map_t map, vm_object_t object,
1821 				       vm_map_entry_t entry);
1822 static int	vm_object_in_map (vm_object_t object);
1823 
1824 /*
1825  * The caller must hold vm_token.
1826  */
1827 static int
1828 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
1829 {
1830 	vm_map_t tmpm;
1831 	vm_map_entry_t tmpe;
1832 	vm_object_t obj;
1833 	int entcount;
1834 
1835 	if (map == 0)
1836 		return 0;
1837 	if (entry == 0) {
1838 		tmpe = map->header.next;
1839 		entcount = map->nentries;
1840 		while (entcount-- && (tmpe != &map->header)) {
1841 			if( _vm_object_in_map(map, object, tmpe)) {
1842 				return 1;
1843 			}
1844 			tmpe = tmpe->next;
1845 		}
1846 		return (0);
1847 	}
1848 	switch(entry->maptype) {
1849 	case VM_MAPTYPE_SUBMAP:
1850 		tmpm = entry->object.sub_map;
1851 		tmpe = tmpm->header.next;
1852 		entcount = tmpm->nentries;
1853 		while (entcount-- && tmpe != &tmpm->header) {
1854 			if( _vm_object_in_map(tmpm, object, tmpe)) {
1855 				return 1;
1856 			}
1857 			tmpe = tmpe->next;
1858 		}
1859 		break;
1860 	case VM_MAPTYPE_NORMAL:
1861 	case VM_MAPTYPE_VPAGETABLE:
1862 		obj = entry->object.vm_object;
1863 		while (obj) {
1864 			if (obj == object)
1865 				return 1;
1866 			obj = obj->backing_object;
1867 		}
1868 		break;
1869 	default:
1870 		break;
1871 	}
1872 	return 0;
1873 }
1874 
1875 static int vm_object_in_map_callback(struct proc *p, void *data);
1876 
1877 struct vm_object_in_map_info {
1878 	vm_object_t object;
1879 	int rv;
1880 };
1881 
1882 /*
1883  * Debugging only
1884  */
1885 static int
1886 vm_object_in_map(vm_object_t object)
1887 {
1888 	struct vm_object_in_map_info info;
1889 
1890 	info.rv = 0;
1891 	info.object = object;
1892 
1893 	allproc_scan(vm_object_in_map_callback, &info);
1894 	if (info.rv)
1895 		return 1;
1896 	if( _vm_object_in_map(&kernel_map, object, 0))
1897 		return 1;
1898 	if( _vm_object_in_map(&pager_map, object, 0))
1899 		return 1;
1900 	if( _vm_object_in_map(&buffer_map, object, 0))
1901 		return 1;
1902 	return 0;
1903 }
1904 
1905 /*
1906  * Debugging only
1907  */
1908 static int
1909 vm_object_in_map_callback(struct proc *p, void *data)
1910 {
1911 	struct vm_object_in_map_info *info = data;
1912 
1913 	if (p->p_vmspace) {
1914 		if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) {
1915 			info->rv = 1;
1916 			return -1;
1917 		}
1918 	}
1919 	return (0);
1920 }
1921 
1922 DB_SHOW_COMMAND(vmochk, vm_object_check)
1923 {
1924 	vm_object_t object;
1925 
1926 	/*
1927 	 * make sure that internal objs are in a map somewhere
1928 	 * and none have zero ref counts.
1929 	 */
1930 	for (object = TAILQ_FIRST(&vm_object_list);
1931 			object != NULL;
1932 			object = TAILQ_NEXT(object, object_list)) {
1933 		if (object->type == OBJT_MARKER)
1934 			continue;
1935 		if (object->handle == NULL &&
1936 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1937 			if (object->ref_count == 0) {
1938 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
1939 					(long)object->size);
1940 			}
1941 			if (!vm_object_in_map(object)) {
1942 				db_printf(
1943 			"vmochk: internal obj is not in a map: "
1944 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1945 				    object->ref_count, (u_long)object->size,
1946 				    (u_long)object->size,
1947 				    (void *)object->backing_object);
1948 			}
1949 		}
1950 	}
1951 }
1952 
1953 /*
1954  * Debugging only
1955  */
1956 DB_SHOW_COMMAND(object, vm_object_print_static)
1957 {
1958 	/* XXX convert args. */
1959 	vm_object_t object = (vm_object_t)addr;
1960 	boolean_t full = have_addr;
1961 
1962 	vm_page_t p;
1963 
1964 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
1965 #define	count	was_count
1966 
1967 	int count;
1968 
1969 	if (object == NULL)
1970 		return;
1971 
1972 	db_iprintf(
1973 	    "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1974 	    object, (int)object->type, (u_long)object->size,
1975 	    object->resident_page_count, object->ref_count, object->flags);
1976 	/*
1977 	 * XXX no %qd in kernel.  Truncate object->backing_object_offset.
1978 	 */
1979 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1980 	    object->shadow_count,
1981 	    object->backing_object ? object->backing_object->ref_count : 0,
1982 	    object->backing_object, (long)object->backing_object_offset);
1983 
1984 	if (!full)
1985 		return;
1986 
1987 	db_indent += 2;
1988 	count = 0;
1989 	RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) {
1990 		if (count == 0)
1991 			db_iprintf("memory:=");
1992 		else if (count == 6) {
1993 			db_printf("\n");
1994 			db_iprintf(" ...");
1995 			count = 0;
1996 		} else
1997 			db_printf(",");
1998 		count++;
1999 
2000 		db_printf("(off=0x%lx,page=0x%lx)",
2001 		    (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
2002 	}
2003 	if (count != 0)
2004 		db_printf("\n");
2005 	db_indent -= 2;
2006 }
2007 
2008 /* XXX. */
2009 #undef count
2010 
2011 /*
2012  * XXX need this non-static entry for calling from vm_map_print.
2013  *
2014  * Debugging only
2015  */
2016 void
2017 vm_object_print(/* db_expr_t */ long addr,
2018 		boolean_t have_addr,
2019 		/* db_expr_t */ long count,
2020 		char *modif)
2021 {
2022 	vm_object_print_static(addr, have_addr, count, modif);
2023 }
2024 
2025 /*
2026  * Debugging only
2027  */
2028 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2029 {
2030 	vm_object_t object;
2031 	int nl = 0;
2032 	int c;
2033 	for (object = TAILQ_FIRST(&vm_object_list);
2034 			object != NULL;
2035 			object = TAILQ_NEXT(object, object_list)) {
2036 		vm_pindex_t idx, fidx;
2037 		vm_pindex_t osize;
2038 		vm_paddr_t pa = -1, padiff;
2039 		int rcount;
2040 		vm_page_t m;
2041 
2042 		if (object->type == OBJT_MARKER)
2043 			continue;
2044 		db_printf("new object: %p\n", (void *)object);
2045 		if ( nl > 18) {
2046 			c = cngetc();
2047 			if (c != ' ')
2048 				return;
2049 			nl = 0;
2050 		}
2051 		nl++;
2052 		rcount = 0;
2053 		fidx = 0;
2054 		osize = object->size;
2055 		if (osize > 128)
2056 			osize = 128;
2057 		for (idx = 0; idx < osize; idx++) {
2058 			m = vm_page_lookup(object, idx);
2059 			if (m == NULL) {
2060 				if (rcount) {
2061 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2062 						(long)fidx, rcount, (long)pa);
2063 					if ( nl > 18) {
2064 						c = cngetc();
2065 						if (c != ' ')
2066 							return;
2067 						nl = 0;
2068 					}
2069 					nl++;
2070 					rcount = 0;
2071 				}
2072 				continue;
2073 			}
2074 
2075 
2076 			if (rcount &&
2077 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2078 				++rcount;
2079 				continue;
2080 			}
2081 			if (rcount) {
2082 				padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
2083 				padiff >>= PAGE_SHIFT;
2084 				padiff &= PQ_L2_MASK;
2085 				if (padiff == 0) {
2086 					pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
2087 					++rcount;
2088 					continue;
2089 				}
2090 				db_printf(" index(%ld)run(%d)pa(0x%lx)",
2091 					(long)fidx, rcount, (long)pa);
2092 				db_printf("pd(%ld)\n", (long)padiff);
2093 				if ( nl > 18) {
2094 					c = cngetc();
2095 					if (c != ' ')
2096 						return;
2097 					nl = 0;
2098 				}
2099 				nl++;
2100 			}
2101 			fidx = idx;
2102 			pa = VM_PAGE_TO_PHYS(m);
2103 			rcount = 1;
2104 		}
2105 		if (rcount) {
2106 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2107 				(long)fidx, rcount, (long)pa);
2108 			if ( nl > 18) {
2109 				c = cngetc();
2110 				if (c != ' ')
2111 					return;
2112 				nl = 0;
2113 			}
2114 			nl++;
2115 		}
2116 	}
2117 }
2118 #endif /* DDB */
2119