xref: /freebsd/sys/vm/vm_object.c (revision c17d43407fe04133a94055b0dbc7ea8965654a9f)
1 /*
2  * Copyright (c) 1991, 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * This code is derived from software contributed to Berkeley by
6  * The Mach Operating System project at Carnegie-Mellon University.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *	This product includes software developed by the University of
19  *	California, Berkeley and its contributors.
20  * 4. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	from: @(#)vm_object.c	8.5 (Berkeley) 3/22/94
37  *
38  *
39  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40  * All rights reserved.
41  *
42  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
43  *
44  * Permission to use, copy, modify and distribute this software and
45  * its documentation is hereby granted, provided that both the copyright
46  * notice and this permission notice appear in all copies of the
47  * software, derivative works or modified versions, and any portions
48  * thereof, and that both notices appear in supporting documentation.
49  *
50  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
53  *
54  * Carnegie Mellon requests users of this software to return to
55  *
56  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
57  *  School of Computer Science
58  *  Carnegie Mellon University
59  *  Pittsburgh PA 15213-3890
60  *
61  * any improvements or extensions that they make and grant Carnegie the
62  * rights to redistribute these changes.
63  *
64  * $FreeBSD$
65  */
66 
67 /*
68  *	Virtual memory object module.
69  */
70 
71 #include <sys/param.h>
72 #include <sys/systm.h>
73 #include <sys/lock.h>
74 #include <sys/mman.h>
75 #include <sys/mount.h>
76 #include <sys/kernel.h>
77 #include <sys/sysctl.h>
78 #include <sys/mutex.h>
79 #include <sys/proc.h>		/* for curproc, pageproc */
80 #include <sys/socket.h>
81 #include <sys/vnode.h>
82 #include <sys/vmmeter.h>
83 #include <sys/sx.h>
84 
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_pager.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/uma.h>
97 
98 #define EASY_SCAN_FACTOR       8
99 
100 #define MSYNC_FLUSH_HARDSEQ	0x01
101 #define MSYNC_FLUSH_SOFTSEQ	0x02
102 
103 /*
104  * msync / VM object flushing optimizations
105  */
106 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ;
107 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags,
108         CTLFLAG_RW, &msync_flush_flags, 0, "");
109 
110 static void	vm_object_qcollapse(vm_object_t object);
111 static int	vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
112 
113 /*
114  *	Virtual memory objects maintain the actual data
115  *	associated with allocated virtual memory.  A given
116  *	page of memory exists within exactly one object.
117  *
118  *	An object is only deallocated when all "references"
119  *	are given up.  Only one "reference" to a given
120  *	region of an object should be writeable.
121  *
122  *	Associated with each object is a list of all resident
123  *	memory pages belonging to that object; this list is
124  *	maintained by the "vm_page" module, and locked by the object's
125  *	lock.
126  *
127  *	Each object also records a "pager" routine which is
128  *	used to retrieve (and store) pages to the proper backing
129  *	storage.  In addition, objects may be backed by other
130  *	objects from which they were virtual-copied.
131  *
132  *	The only items within the object structure which are
133  *	modified after time of creation are:
134  *		reference count		locked by object's lock
135  *		pager routine		locked by object's lock
136  *
137  */
138 
139 struct object_q vm_object_list;
140 static struct mtx vm_object_list_mtx;	/* lock for object list and count */
141 vm_object_t kernel_object;
142 vm_object_t kmem_object;
143 static struct vm_object kernel_object_store;
144 static struct vm_object kmem_object_store;
145 extern int vm_pageout_page_count;
146 
147 static long object_collapses;
148 static long object_bypasses;
149 static int next_index;
150 static int object_hash_rand;
151 static uma_zone_t obj_zone;
152 #define VM_OBJECTS_INIT 256
153 
154 static void vm_object_zinit(void *mem, int size);
155 
156 #ifdef INVARIANTS
157 static void vm_object_zdtor(void *mem, int size, void *arg);
158 
159 static void
160 vm_object_zdtor(void *mem, int size, void *arg)
161 {
162 	vm_object_t object;
163 
164 	object = (vm_object_t)mem;
165 	KASSERT(object->paging_in_progress == 0,
166 	    ("object %p paging_in_progress = %d",
167 	    object, object->paging_in_progress));
168 	KASSERT(object->resident_page_count == 0,
169 	    ("object %p resident_page_count = %d",
170 	    object, object->resident_page_count));
171 	KASSERT(object->shadow_count == 0,
172 	    ("object %p shadow_count = %d",
173 	    object, object->shadow_count));
174 }
175 #endif
176 
177 static void
178 vm_object_zinit(void *mem, int size)
179 {
180 	vm_object_t object;
181 
182 	object = (vm_object_t)mem;
183 
184 	/* These are true for any object that has been freed */
185 	object->paging_in_progress = 0;
186 	object->resident_page_count = 0;
187 	object->shadow_count = 0;
188 }
189 
190 void
191 _vm_object_allocate(objtype_t type, vm_size_t size, vm_object_t object)
192 {
193 	int incr;
194 
195 	GIANT_REQUIRED;
196 
197 	TAILQ_INIT(&object->memq);
198 	TAILQ_INIT(&object->shadow_head);
199 
200 	object->type = type;
201 	object->size = size;
202 	object->ref_count = 1;
203 	object->flags = 0;
204 	if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
205 		vm_object_set_flag(object, OBJ_ONEMAPPING);
206 	object->pg_color = next_index;
207 	if (size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
208 		incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
209 	else
210 		incr = size;
211 	next_index = (next_index + incr) & PQ_L2_MASK;
212 	object->handle = NULL;
213 	object->backing_object = NULL;
214 	object->backing_object_offset = (vm_ooffset_t) 0;
215 	/*
216 	 * Try to generate a number that will spread objects out in the
217 	 * hash table.  We 'wipe' new objects across the hash in 128 page
218 	 * increments plus 1 more to offset it a little more by the time
219 	 * it wraps around.
220 	 */
221 	object->hash_rand = object_hash_rand - 129;
222 
223 	object->generation++;
224 
225 	TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
226 
227 	object_hash_rand = object->hash_rand;
228 }
229 
230 /*
231  *	vm_object_init:
232  *
233  *	Initialize the VM objects module.
234  */
235 void
236 vm_object_init(void)
237 {
238 	GIANT_REQUIRED;
239 
240 	TAILQ_INIT(&vm_object_list);
241 	mtx_init(&vm_object_list_mtx, "vm object_list", MTX_DEF);
242 
243 	kernel_object = &kernel_object_store;
244 	_vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
245 	    kernel_object);
246 
247 	kmem_object = &kmem_object_store;
248 	_vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
249 	    kmem_object);
250 	obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
251 #ifdef INVARIANTS
252 	    vm_object_zdtor,
253 #else
254 	    NULL,
255 #endif
256 	    vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
257 	uma_prealloc(obj_zone, VM_OBJECTS_INIT);
258 }
259 
260 void
261 vm_object_init2(void)
262 {
263 }
264 
265 void
266 vm_object_set_flag(vm_object_t object, u_short bits)
267 {
268 	GIANT_REQUIRED;
269 	object->flags |= bits;
270 }
271 
272 void
273 vm_object_clear_flag(vm_object_t object, u_short bits)
274 {
275 	GIANT_REQUIRED;
276 	object->flags &= ~bits;
277 }
278 
279 void
280 vm_object_pip_add(vm_object_t object, short i)
281 {
282 	GIANT_REQUIRED;
283 	object->paging_in_progress += i;
284 }
285 
286 void
287 vm_object_pip_subtract(vm_object_t object, short i)
288 {
289 	GIANT_REQUIRED;
290 	object->paging_in_progress -= i;
291 }
292 
293 void
294 vm_object_pip_wakeup(vm_object_t object)
295 {
296 	GIANT_REQUIRED;
297 	object->paging_in_progress--;
298 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
299 		vm_object_clear_flag(object, OBJ_PIPWNT);
300 		wakeup(object);
301 	}
302 }
303 
304 void
305 vm_object_pip_wakeupn(vm_object_t object, short i)
306 {
307 	GIANT_REQUIRED;
308 	if (i)
309 		object->paging_in_progress -= i;
310 	if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
311 		vm_object_clear_flag(object, OBJ_PIPWNT);
312 		wakeup(object);
313 	}
314 }
315 
316 void
317 vm_object_pip_sleep(vm_object_t object, char *waitid)
318 {
319 	GIANT_REQUIRED;
320 	if (object->paging_in_progress) {
321 		int s = splvm();
322 		if (object->paging_in_progress) {
323 			vm_object_set_flag(object, OBJ_PIPWNT);
324 			tsleep(object, PVM, waitid, 0);
325 		}
326 		splx(s);
327 	}
328 }
329 
330 void
331 vm_object_pip_wait(vm_object_t object, char *waitid)
332 {
333 	GIANT_REQUIRED;
334 	while (object->paging_in_progress)
335 		vm_object_pip_sleep(object, waitid);
336 }
337 
338 /*
339  *	vm_object_allocate:
340  *
341  *	Returns a new object with the given size.
342  */
343 vm_object_t
344 vm_object_allocate(objtype_t type, vm_size_t size)
345 {
346 	vm_object_t result;
347 
348 	GIANT_REQUIRED;
349 
350 	result = (vm_object_t) uma_zalloc(obj_zone, M_WAITOK);
351 	_vm_object_allocate(type, size, result);
352 
353 	return (result);
354 }
355 
356 
357 /*
358  *	vm_object_reference:
359  *
360  *	Gets another reference to the given object.
361  */
362 void
363 vm_object_reference(vm_object_t object)
364 {
365 	GIANT_REQUIRED;
366 
367 	if (object == NULL)
368 		return;
369 
370 #if 0
371 	/* object can be re-referenced during final cleaning */
372 	KASSERT(!(object->flags & OBJ_DEAD),
373 	    ("vm_object_reference: attempting to reference dead obj"));
374 #endif
375 
376 	object->ref_count++;
377 	if (object->type == OBJT_VNODE) {
378 		while (vget((struct vnode *) object->handle, LK_RETRY|LK_NOOBJ, curthread)) {
379 			printf("vm_object_reference: delay in getting object\n");
380 		}
381 	}
382 }
383 
384 /*
385  * handle deallocating a object of type OBJT_VNODE
386  */
387 void
388 vm_object_vndeallocate(vm_object_t object)
389 {
390 	struct vnode *vp = (struct vnode *) object->handle;
391 
392 	GIANT_REQUIRED;
393 	KASSERT(object->type == OBJT_VNODE,
394 	    ("vm_object_vndeallocate: not a vnode object"));
395 	KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
396 #ifdef INVARIANTS
397 	if (object->ref_count == 0) {
398 		vprint("vm_object_vndeallocate", vp);
399 		panic("vm_object_vndeallocate: bad object reference count");
400 	}
401 #endif
402 
403 	object->ref_count--;
404 	if (object->ref_count == 0) {
405 		vp->v_flag &= ~VTEXT;
406 		vm_object_clear_flag(object, OBJ_OPT);
407 	}
408 	/*
409 	 * vrele may need a vop lock
410 	 */
411 	vrele(vp);
412 }
413 
414 /*
415  *	vm_object_deallocate:
416  *
417  *	Release a reference to the specified object,
418  *	gained either through a vm_object_allocate
419  *	or a vm_object_reference call.  When all references
420  *	are gone, storage associated with this object
421  *	may be relinquished.
422  *
423  *	No object may be locked.
424  */
425 void
426 vm_object_deallocate(vm_object_t object)
427 {
428 	vm_object_t temp;
429 
430 	GIANT_REQUIRED;
431 
432 	while (object != NULL) {
433 
434 		if (object->type == OBJT_VNODE) {
435 			vm_object_vndeallocate(object);
436 			return;
437 		}
438 
439 		KASSERT(object->ref_count != 0,
440 			("vm_object_deallocate: object deallocated too many times: %d", object->type));
441 
442 		/*
443 		 * If the reference count goes to 0 we start calling
444 		 * vm_object_terminate() on the object chain.
445 		 * A ref count of 1 may be a special case depending on the
446 		 * shadow count being 0 or 1.
447 		 */
448 		object->ref_count--;
449 		if (object->ref_count > 1) {
450 			return;
451 		} else if (object->ref_count == 1) {
452 			if (object->shadow_count == 0) {
453 				vm_object_set_flag(object, OBJ_ONEMAPPING);
454 			} else if ((object->shadow_count == 1) &&
455 			    (object->handle == NULL) &&
456 			    (object->type == OBJT_DEFAULT ||
457 			     object->type == OBJT_SWAP)) {
458 				vm_object_t robject;
459 
460 				robject = TAILQ_FIRST(&object->shadow_head);
461 				KASSERT(robject != NULL,
462 				    ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
463 					 object->ref_count,
464 					 object->shadow_count));
465 				if ((robject->handle == NULL) &&
466 				    (robject->type == OBJT_DEFAULT ||
467 				     robject->type == OBJT_SWAP)) {
468 
469 					robject->ref_count++;
470 
471 					while (
472 						robject->paging_in_progress ||
473 						object->paging_in_progress
474 					) {
475 						vm_object_pip_sleep(robject, "objde1");
476 						vm_object_pip_sleep(object, "objde2");
477 					}
478 
479 					if (robject->ref_count == 1) {
480 						robject->ref_count--;
481 						object = robject;
482 						goto doterm;
483 					}
484 
485 					object = robject;
486 					vm_object_collapse(object);
487 					continue;
488 				}
489 			}
490 
491 			return;
492 
493 		}
494 
495 doterm:
496 
497 		temp = object->backing_object;
498 		if (temp) {
499 			TAILQ_REMOVE(&temp->shadow_head, object, shadow_list);
500 			temp->shadow_count--;
501 			if (temp->ref_count == 0)
502 				vm_object_clear_flag(temp, OBJ_OPT);
503 			temp->generation++;
504 			object->backing_object = NULL;
505 		}
506 		/*
507 		 * Don't double-terminate, we could be in a termination
508 		 * recursion due to the terminate having to sync data
509 		 * to disk.
510 		 */
511 		if ((object->flags & OBJ_DEAD) == 0)
512 			vm_object_terminate(object);
513 		object = temp;
514 	}
515 }
516 
517 /*
518  *	vm_object_terminate actually destroys the specified object, freeing
519  *	up all previously used resources.
520  *
521  *	The object must be locked.
522  *	This routine may block.
523  */
524 void
525 vm_object_terminate(vm_object_t object)
526 {
527 	vm_page_t p;
528 	int s;
529 
530 	GIANT_REQUIRED;
531 
532 	/*
533 	 * Make sure no one uses us.
534 	 */
535 	vm_object_set_flag(object, OBJ_DEAD);
536 
537 	/*
538 	 * wait for the pageout daemon to be done with the object
539 	 */
540 	vm_object_pip_wait(object, "objtrm");
541 
542 	KASSERT(!object->paging_in_progress,
543 		("vm_object_terminate: pageout in progress"));
544 
545 	/*
546 	 * Clean and free the pages, as appropriate. All references to the
547 	 * object are gone, so we don't need to lock it.
548 	 */
549 	if (object->type == OBJT_VNODE) {
550 		struct vnode *vp;
551 
552 		/*
553 		 * Freeze optimized copies.
554 		 */
555 		vm_freeze_copyopts(object, 0, object->size);
556 
557 		/*
558 		 * Clean pages and flush buffers.
559 		 */
560 		vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
561 
562 		vp = (struct vnode *) object->handle;
563 		vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0);
564 	}
565 
566 	KASSERT(object->ref_count == 0,
567 		("vm_object_terminate: object with references, ref_count=%d",
568 		object->ref_count));
569 
570 	/*
571 	 * Now free any remaining pages. For internal objects, this also
572 	 * removes them from paging queues. Don't free wired pages, just
573 	 * remove them from the object.
574 	 */
575 	s = splvm();
576 	while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
577 		KASSERT(!p->busy && (p->flags & PG_BUSY) == 0,
578 			("vm_object_terminate: freeing busy page %p "
579 			"p->busy = %d, p->flags %x\n", p, p->busy, p->flags));
580 		if (p->wire_count == 0) {
581 			vm_page_busy(p);
582 			vm_page_free(p);
583 			cnt.v_pfree++;
584 		} else {
585 			vm_page_busy(p);
586 			vm_page_remove(p);
587 		}
588 	}
589 	splx(s);
590 
591 	/*
592 	 * Let the pager know object is dead.
593 	 */
594 	vm_pager_deallocate(object);
595 
596 	/*
597 	 * Remove the object from the global object list.
598 	 */
599 	mtx_lock(&vm_object_list_mtx);
600 	TAILQ_REMOVE(&vm_object_list, object, object_list);
601 	mtx_unlock(&vm_object_list_mtx);
602 
603 	wakeup(object);
604 
605 	/*
606 	 * Free the space for the object.
607 	 */
608 	uma_zfree(obj_zone, object);
609 }
610 
611 /*
612  *	vm_object_page_clean
613  *
614  *	Clean all dirty pages in the specified range of object.  Leaves page
615  * 	on whatever queue it is currently on.   If NOSYNC is set then do not
616  *	write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
617  *	leaving the object dirty.
618  *
619  *	Odd semantics: if start == end, we clean everything.
620  *
621  *	The object must be locked.
622  */
623 void
624 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags)
625 {
626 	vm_page_t p, np;
627 	vm_offset_t tstart, tend;
628 	vm_pindex_t pi;
629 	struct vnode *vp;
630 	int clearobjflags;
631 	int pagerflags;
632 	int curgeneration;
633 
634 	GIANT_REQUIRED;
635 
636 	if (object->type != OBJT_VNODE ||
637 		(object->flags & OBJ_MIGHTBEDIRTY) == 0)
638 		return;
639 
640 	pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : 0;
641 	pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
642 
643 	vp = object->handle;
644 
645 	vm_object_set_flag(object, OBJ_CLEANING);
646 
647 	tstart = start;
648 	if (end == 0) {
649 		tend = object->size;
650 	} else {
651 		tend = end;
652 	}
653 
654 	/*
655 	 * If the caller is smart and only msync()s a range he knows is
656 	 * dirty, we may be able to avoid an object scan.  This results in
657 	 * a phenominal improvement in performance.  We cannot do this
658 	 * as a matter of course because the object may be huge - e.g.
659 	 * the size might be in the gigabytes or terrabytes.
660 	 */
661 	if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
662 		vm_offset_t tscan;
663 		int scanlimit;
664 		int scanreset;
665 
666 		scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
667 		if (scanreset < 16)
668 			scanreset = 16;
669 
670 		scanlimit = scanreset;
671 		tscan = tstart;
672 		while (tscan < tend) {
673 			curgeneration = object->generation;
674 			p = vm_page_lookup(object, tscan);
675 			if (p == NULL || p->valid == 0 ||
676 			    (p->queue - p->pc) == PQ_CACHE) {
677 				if (--scanlimit == 0)
678 					break;
679 				++tscan;
680 				continue;
681 			}
682 			vm_page_test_dirty(p);
683 			if ((p->dirty & p->valid) == 0) {
684 				if (--scanlimit == 0)
685 					break;
686 				++tscan;
687 				continue;
688 			}
689 			/*
690 			 * If we have been asked to skip nosync pages and
691 			 * this is a nosync page, we can't continue.
692 			 */
693 			if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
694 				if (--scanlimit == 0)
695 					break;
696 				++tscan;
697 				continue;
698 			}
699 			scanlimit = scanreset;
700 
701 			/*
702 			 * This returns 0 if it was unable to busy the first
703 			 * page (i.e. had to sleep).
704 			 */
705 			tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
706 		}
707 
708 		/*
709 		 * If everything was dirty and we flushed it successfully,
710 		 * and the requested range is not the entire object, we
711 		 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
712 		 * return immediately.
713 		 */
714 		if (tscan >= tend && (tstart || tend < object->size)) {
715 			vm_object_clear_flag(object, OBJ_CLEANING);
716 			return;
717 		}
718 	}
719 
720 	/*
721 	 * Generally set CLEANCHK interlock and make the page read-only so
722 	 * we can then clear the object flags.
723 	 *
724 	 * However, if this is a nosync mmap then the object is likely to
725 	 * stay dirty so do not mess with the page and do not clear the
726 	 * object flags.
727 	 */
728 	clearobjflags = 1;
729 
730 	TAILQ_FOREACH(p, &object->memq, listq) {
731 		vm_page_flag_set(p, PG_CLEANCHK);
732 		if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
733 			clearobjflags = 0;
734 		else
735 			vm_page_protect(p, VM_PROT_READ);
736 	}
737 
738 	if (clearobjflags && (tstart == 0) && (tend == object->size)) {
739 		struct vnode *vp;
740 
741 		vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
742 		if (object->type == OBJT_VNODE &&
743 		    (vp = (struct vnode *)object->handle) != NULL) {
744 			if (vp->v_flag & VOBJDIRTY) {
745 				mtx_lock(&vp->v_interlock);
746 				vp->v_flag &= ~VOBJDIRTY;
747 				mtx_unlock(&vp->v_interlock);
748 			}
749 		}
750 	}
751 
752 rescan:
753 	curgeneration = object->generation;
754 
755 	for (p = TAILQ_FIRST(&object->memq); p; p = np) {
756 		int n;
757 
758 		np = TAILQ_NEXT(p, listq);
759 
760 again:
761 		pi = p->pindex;
762 		if (((p->flags & PG_CLEANCHK) == 0) ||
763 			(pi < tstart) || (pi >= tend) ||
764 			(p->valid == 0) ||
765 			((p->queue - p->pc) == PQ_CACHE)) {
766 			vm_page_flag_clear(p, PG_CLEANCHK);
767 			continue;
768 		}
769 
770 		vm_page_test_dirty(p);
771 		if ((p->dirty & p->valid) == 0) {
772 			vm_page_flag_clear(p, PG_CLEANCHK);
773 			continue;
774 		}
775 
776 		/*
777 		 * If we have been asked to skip nosync pages and this is a
778 		 * nosync page, skip it.  Note that the object flags were
779 		 * not cleared in this case so we do not have to set them.
780 		 */
781 		if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
782 			vm_page_flag_clear(p, PG_CLEANCHK);
783 			continue;
784 		}
785 
786 		n = vm_object_page_collect_flush(object, p,
787 			curgeneration, pagerflags);
788 		if (n == 0)
789 			goto rescan;
790 
791 		if (object->generation != curgeneration)
792 			goto rescan;
793 
794 		/*
795 		 * Try to optimize the next page.  If we can't we pick up
796 		 * our (random) scan where we left off.
797 		 */
798 		if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
799 			if ((p = vm_page_lookup(object, pi + n)) != NULL)
800 				goto again;
801 		}
802 	}
803 
804 #if 0
805 	VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
806 #endif
807 
808 	vm_object_clear_flag(object, OBJ_CLEANING);
809 	return;
810 }
811 
812 static int
813 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
814 {
815 	int runlen;
816 	int s;
817 	int maxf;
818 	int chkb;
819 	int maxb;
820 	int i;
821 	vm_pindex_t pi;
822 	vm_page_t maf[vm_pageout_page_count];
823 	vm_page_t mab[vm_pageout_page_count];
824 	vm_page_t ma[vm_pageout_page_count];
825 
826 	s = splvm();
827 	pi = p->pindex;
828 	while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
829 		if (object->generation != curgeneration) {
830 			splx(s);
831 			return(0);
832 		}
833 	}
834 
835 	maxf = 0;
836 	for(i = 1; i < vm_pageout_page_count; i++) {
837 		vm_page_t tp;
838 
839 		if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
840 			if ((tp->flags & PG_BUSY) ||
841 				(tp->flags & PG_CLEANCHK) == 0 ||
842 				(tp->busy != 0))
843 				break;
844 			if((tp->queue - tp->pc) == PQ_CACHE) {
845 				vm_page_flag_clear(tp, PG_CLEANCHK);
846 				break;
847 			}
848 			vm_page_test_dirty(tp);
849 			if ((tp->dirty & tp->valid) == 0) {
850 				vm_page_flag_clear(tp, PG_CLEANCHK);
851 				break;
852 			}
853 			maf[ i - 1 ] = tp;
854 			maxf++;
855 			continue;
856 		}
857 		break;
858 	}
859 
860 	maxb = 0;
861 	chkb = vm_pageout_page_count -  maxf;
862 	if (chkb) {
863 		for(i = 1; i < chkb;i++) {
864 			vm_page_t tp;
865 
866 			if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
867 				if ((tp->flags & PG_BUSY) ||
868 					(tp->flags & PG_CLEANCHK) == 0 ||
869 					(tp->busy != 0))
870 					break;
871 				if ((tp->queue - tp->pc) == PQ_CACHE) {
872 					vm_page_flag_clear(tp, PG_CLEANCHK);
873 					break;
874 				}
875 				vm_page_test_dirty(tp);
876 				if ((tp->dirty & tp->valid) == 0) {
877 					vm_page_flag_clear(tp, PG_CLEANCHK);
878 					break;
879 				}
880 				mab[ i - 1 ] = tp;
881 				maxb++;
882 				continue;
883 			}
884 			break;
885 		}
886 	}
887 
888 	for(i = 0; i < maxb; i++) {
889 		int index = (maxb - i) - 1;
890 		ma[index] = mab[i];
891 		vm_page_flag_clear(ma[index], PG_CLEANCHK);
892 	}
893 	vm_page_flag_clear(p, PG_CLEANCHK);
894 	ma[maxb] = p;
895 	for(i = 0; i < maxf; i++) {
896 		int index = (maxb + i) + 1;
897 		ma[index] = maf[i];
898 		vm_page_flag_clear(ma[index], PG_CLEANCHK);
899 	}
900 	runlen = maxb + maxf + 1;
901 
902 	splx(s);
903 	vm_pageout_flush(ma, runlen, pagerflags);
904 	for (i = 0; i < runlen; i++) {
905 		if (ma[i]->valid & ma[i]->dirty) {
906 			vm_page_protect(ma[i], VM_PROT_READ);
907 			vm_page_flag_set(ma[i], PG_CLEANCHK);
908 
909 			/*
910 			 * maxf will end up being the actual number of pages
911 			 * we wrote out contiguously, non-inclusive of the
912 			 * first page.  We do not count look-behind pages.
913 			 */
914 			if (i >= maxb + 1 && (maxf > i - maxb - 1))
915 				maxf = i - maxb - 1;
916 		}
917 	}
918 	return(maxf + 1);
919 }
920 
921 /*
922  * Same as vm_object_pmap_copy, except range checking really
923  * works, and is meant for small sections of an object.
924  *
925  * This code protects resident pages by making them read-only
926  * and is typically called on a fork or split when a page
927  * is converted to copy-on-write.
928  *
929  * NOTE: If the page is already at VM_PROT_NONE, calling
930  * vm_page_protect will have no effect.
931  */
932 void
933 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
934 {
935 	vm_pindex_t idx;
936 	vm_page_t p;
937 
938 	GIANT_REQUIRED;
939 
940 	if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
941 		return;
942 
943 	for (idx = start; idx < end; idx++) {
944 		p = vm_page_lookup(object, idx);
945 		if (p == NULL)
946 			continue;
947 		vm_page_protect(p, VM_PROT_READ);
948 	}
949 }
950 
951 /*
952  *	vm_object_pmap_remove:
953  *
954  *	Removes all physical pages in the specified
955  *	object range from all physical maps.
956  *
957  *	The object must *not* be locked.
958  */
959 void
960 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
961 {
962 	vm_page_t p;
963 
964 	GIANT_REQUIRED;
965 	if (object == NULL)
966 		return;
967 	TAILQ_FOREACH(p, &object->memq, listq) {
968 		if (p->pindex >= start && p->pindex < end)
969 			vm_page_protect(p, VM_PROT_NONE);
970 	}
971 	if ((start == 0) && (object->size == end))
972 		vm_object_clear_flag(object, OBJ_WRITEABLE);
973 }
974 
975 /*
976  *	vm_object_madvise:
977  *
978  *	Implements the madvise function at the object/page level.
979  *
980  *	MADV_WILLNEED	(any object)
981  *
982  *	    Activate the specified pages if they are resident.
983  *
984  *	MADV_DONTNEED	(any object)
985  *
986  *	    Deactivate the specified pages if they are resident.
987  *
988  *	MADV_FREE	(OBJT_DEFAULT/OBJT_SWAP objects,
989  *			 OBJ_ONEMAPPING only)
990  *
991  *	    Deactivate and clean the specified pages if they are
992  *	    resident.  This permits the process to reuse the pages
993  *	    without faulting or the kernel to reclaim the pages
994  *	    without I/O.
995  */
996 void
997 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
998 {
999 	vm_pindex_t end, tpindex;
1000 	vm_object_t tobject;
1001 	vm_page_t m;
1002 
1003 	GIANT_REQUIRED;
1004 	if (object == NULL)
1005 		return;
1006 
1007 	end = pindex + count;
1008 
1009 	/*
1010 	 * Locate and adjust resident pages
1011 	 */
1012 	for (; pindex < end; pindex += 1) {
1013 relookup:
1014 		tobject = object;
1015 		tpindex = pindex;
1016 shadowlookup:
1017 		/*
1018 		 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1019 		 * and those pages must be OBJ_ONEMAPPING.
1020 		 */
1021 		if (advise == MADV_FREE) {
1022 			if ((tobject->type != OBJT_DEFAULT &&
1023 			     tobject->type != OBJT_SWAP) ||
1024 			    (tobject->flags & OBJ_ONEMAPPING) == 0) {
1025 				continue;
1026 			}
1027 		}
1028 
1029 		m = vm_page_lookup(tobject, tpindex);
1030 
1031 		if (m == NULL) {
1032 			/*
1033 			 * There may be swap even if there is no backing page
1034 			 */
1035 			if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1036 				swap_pager_freespace(tobject, tpindex, 1);
1037 
1038 			/*
1039 			 * next object
1040 			 */
1041 			tobject = tobject->backing_object;
1042 			if (tobject == NULL)
1043 				continue;
1044 			tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1045 			goto shadowlookup;
1046 		}
1047 
1048 		/*
1049 		 * If the page is busy or not in a normal active state,
1050 		 * we skip it.  If the page is not managed there are no
1051 		 * page queues to mess with.  Things can break if we mess
1052 		 * with pages in any of the below states.
1053 		 */
1054 		if (
1055 		    m->hold_count ||
1056 		    m->wire_count ||
1057 		    (m->flags & PG_UNMANAGED) ||
1058 		    m->valid != VM_PAGE_BITS_ALL
1059 		) {
1060 			continue;
1061 		}
1062 
1063  		if (vm_page_sleep_busy(m, TRUE, "madvpo"))
1064   			goto relookup;
1065 
1066 		if (advise == MADV_WILLNEED) {
1067 			vm_page_activate(m);
1068 		} else if (advise == MADV_DONTNEED) {
1069 			vm_page_dontneed(m);
1070 		} else if (advise == MADV_FREE) {
1071 			/*
1072 			 * Mark the page clean.  This will allow the page
1073 			 * to be freed up by the system.  However, such pages
1074 			 * are often reused quickly by malloc()/free()
1075 			 * so we do not do anything that would cause
1076 			 * a page fault if we can help it.
1077 			 *
1078 			 * Specifically, we do not try to actually free
1079 			 * the page now nor do we try to put it in the
1080 			 * cache (which would cause a page fault on reuse).
1081 			 *
1082 			 * But we do make the page is freeable as we
1083 			 * can without actually taking the step of unmapping
1084 			 * it.
1085 			 */
1086 			pmap_clear_modify(m);
1087 			m->dirty = 0;
1088 			m->act_count = 0;
1089 			vm_page_dontneed(m);
1090 			if (tobject->type == OBJT_SWAP)
1091 				swap_pager_freespace(tobject, tpindex, 1);
1092 		}
1093 	}
1094 }
1095 
1096 /*
1097  *	vm_object_shadow:
1098  *
1099  *	Create a new object which is backed by the
1100  *	specified existing object range.  The source
1101  *	object reference is deallocated.
1102  *
1103  *	The new object and offset into that object
1104  *	are returned in the source parameters.
1105  */
1106 void
1107 vm_object_shadow(
1108 	vm_object_t *object,	/* IN/OUT */
1109 	vm_ooffset_t *offset,	/* IN/OUT */
1110 	vm_size_t length)
1111 {
1112 	vm_object_t source;
1113 	vm_object_t result;
1114 
1115 	GIANT_REQUIRED;
1116 	source = *object;
1117 
1118 	/*
1119 	 * Don't create the new object if the old object isn't shared.
1120 	 */
1121 	if (source != NULL &&
1122 	    source->ref_count == 1 &&
1123 	    source->handle == NULL &&
1124 	    (source->type == OBJT_DEFAULT ||
1125 	     source->type == OBJT_SWAP))
1126 		return;
1127 
1128 	/*
1129 	 * Allocate a new object with the given length
1130 	 */
1131 	result = vm_object_allocate(OBJT_DEFAULT, length);
1132 	KASSERT(result != NULL, ("vm_object_shadow: no object for shadowing"));
1133 
1134 	/*
1135 	 * The new object shadows the source object, adding a reference to it.
1136 	 * Our caller changes his reference to point to the new object,
1137 	 * removing a reference to the source object.  Net result: no change
1138 	 * of reference count.
1139 	 *
1140 	 * Try to optimize the result object's page color when shadowing
1141 	 * in order to maintain page coloring consistency in the combined
1142 	 * shadowed object.
1143 	 */
1144 	result->backing_object = source;
1145 	if (source) {
1146 		TAILQ_INSERT_TAIL(&source->shadow_head, result, shadow_list);
1147 		source->shadow_count++;
1148 		source->generation++;
1149 		result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1150 	}
1151 
1152 	/*
1153 	 * Store the offset into the source object, and fix up the offset into
1154 	 * the new object.
1155 	 */
1156 	result->backing_object_offset = *offset;
1157 
1158 	/*
1159 	 * Return the new things
1160 	 */
1161 	*offset = 0;
1162 	*object = result;
1163 }
1164 
1165 #define	OBSC_TEST_ALL_SHADOWED	0x0001
1166 #define	OBSC_COLLAPSE_NOWAIT	0x0002
1167 #define	OBSC_COLLAPSE_WAIT	0x0004
1168 
1169 static __inline int
1170 vm_object_backing_scan(vm_object_t object, int op)
1171 {
1172 	int s;
1173 	int r = 1;
1174 	vm_page_t p;
1175 	vm_object_t backing_object;
1176 	vm_pindex_t backing_offset_index;
1177 
1178 	s = splvm();
1179 	GIANT_REQUIRED;
1180 
1181 	backing_object = object->backing_object;
1182 	backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1183 
1184 	/*
1185 	 * Initial conditions
1186 	 */
1187 	if (op & OBSC_TEST_ALL_SHADOWED) {
1188 		/*
1189 		 * We do not want to have to test for the existence of
1190 		 * swap pages in the backing object.  XXX but with the
1191 		 * new swapper this would be pretty easy to do.
1192 		 *
1193 		 * XXX what about anonymous MAP_SHARED memory that hasn't
1194 		 * been ZFOD faulted yet?  If we do not test for this, the
1195 		 * shadow test may succeed! XXX
1196 		 */
1197 		if (backing_object->type != OBJT_DEFAULT) {
1198 			splx(s);
1199 			return (0);
1200 		}
1201 	}
1202 	if (op & OBSC_COLLAPSE_WAIT) {
1203 		vm_object_set_flag(backing_object, OBJ_DEAD);
1204 	}
1205 
1206 	/*
1207 	 * Our scan
1208 	 */
1209 	p = TAILQ_FIRST(&backing_object->memq);
1210 	while (p) {
1211 		vm_page_t next = TAILQ_NEXT(p, listq);
1212 		vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1213 
1214 		if (op & OBSC_TEST_ALL_SHADOWED) {
1215 			vm_page_t pp;
1216 
1217 			/*
1218 			 * Ignore pages outside the parent object's range
1219 			 * and outside the parent object's mapping of the
1220 			 * backing object.
1221 			 *
1222 			 * note that we do not busy the backing object's
1223 			 * page.
1224 			 */
1225 			if (
1226 			    p->pindex < backing_offset_index ||
1227 			    new_pindex >= object->size
1228 			) {
1229 				p = next;
1230 				continue;
1231 			}
1232 
1233 			/*
1234 			 * See if the parent has the page or if the parent's
1235 			 * object pager has the page.  If the parent has the
1236 			 * page but the page is not valid, the parent's
1237 			 * object pager must have the page.
1238 			 *
1239 			 * If this fails, the parent does not completely shadow
1240 			 * the object and we might as well give up now.
1241 			 */
1242 
1243 			pp = vm_page_lookup(object, new_pindex);
1244 			if (
1245 			    (pp == NULL || pp->valid == 0) &&
1246 			    !vm_pager_has_page(object, new_pindex, NULL, NULL)
1247 			) {
1248 				r = 0;
1249 				break;
1250 			}
1251 		}
1252 
1253 		/*
1254 		 * Check for busy page
1255 		 */
1256 		if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1257 			vm_page_t pp;
1258 
1259 			if (op & OBSC_COLLAPSE_NOWAIT) {
1260 				if (
1261 				    (p->flags & PG_BUSY) ||
1262 				    !p->valid ||
1263 				    p->hold_count ||
1264 				    p->wire_count ||
1265 				    p->busy
1266 				) {
1267 					p = next;
1268 					continue;
1269 				}
1270 			} else if (op & OBSC_COLLAPSE_WAIT) {
1271 				if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1272 					/*
1273 					 * If we slept, anything could have
1274 					 * happened.  Since the object is
1275 					 * marked dead, the backing offset
1276 					 * should not have changed so we
1277 					 * just restart our scan.
1278 					 */
1279 					p = TAILQ_FIRST(&backing_object->memq);
1280 					continue;
1281 				}
1282 			}
1283 
1284 			/*
1285 			 * Busy the page
1286 			 */
1287 			vm_page_busy(p);
1288 
1289 			KASSERT(
1290 			    p->object == backing_object,
1291 			    ("vm_object_qcollapse(): object mismatch")
1292 			);
1293 
1294 			/*
1295 			 * Destroy any associated swap
1296 			 */
1297 			if (backing_object->type == OBJT_SWAP) {
1298 				swap_pager_freespace(
1299 				    backing_object,
1300 				    p->pindex,
1301 				    1
1302 				);
1303 			}
1304 
1305 			if (
1306 			    p->pindex < backing_offset_index ||
1307 			    new_pindex >= object->size
1308 			) {
1309 				/*
1310 				 * Page is out of the parent object's range, we
1311 				 * can simply destroy it.
1312 				 */
1313 				vm_page_protect(p, VM_PROT_NONE);
1314 				vm_page_free(p);
1315 				p = next;
1316 				continue;
1317 			}
1318 
1319 			pp = vm_page_lookup(object, new_pindex);
1320 			if (
1321 			    pp != NULL ||
1322 			    vm_pager_has_page(object, new_pindex, NULL, NULL)
1323 			) {
1324 				/*
1325 				 * page already exists in parent OR swap exists
1326 				 * for this location in the parent.  Destroy
1327 				 * the original page from the backing object.
1328 				 *
1329 				 * Leave the parent's page alone
1330 				 */
1331 				vm_page_protect(p, VM_PROT_NONE);
1332 				vm_page_free(p);
1333 				p = next;
1334 				continue;
1335 			}
1336 
1337 			/*
1338 			 * Page does not exist in parent, rename the
1339 			 * page from the backing object to the main object.
1340 			 *
1341 			 * If the page was mapped to a process, it can remain
1342 			 * mapped through the rename.
1343 			 */
1344 			if ((p->queue - p->pc) == PQ_CACHE)
1345 				vm_page_deactivate(p);
1346 
1347 			vm_page_rename(p, object, new_pindex);
1348 			/* page automatically made dirty by rename */
1349 		}
1350 		p = next;
1351 	}
1352 	splx(s);
1353 	return (r);
1354 }
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 static void
1363 vm_object_qcollapse(vm_object_t object)
1364 {
1365 	vm_object_t backing_object = object->backing_object;
1366 
1367 	GIANT_REQUIRED;
1368 
1369 	if (backing_object->ref_count != 1)
1370 		return;
1371 
1372 	backing_object->ref_count += 2;
1373 
1374 	vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1375 
1376 	backing_object->ref_count -= 2;
1377 }
1378 
1379 /*
1380  *	vm_object_collapse:
1381  *
1382  *	Collapse an object with the object backing it.
1383  *	Pages in the backing object are moved into the
1384  *	parent, and the backing object is deallocated.
1385  */
1386 void
1387 vm_object_collapse(vm_object_t object)
1388 {
1389 	GIANT_REQUIRED;
1390 
1391 	while (TRUE) {
1392 		vm_object_t backing_object;
1393 
1394 		/*
1395 		 * Verify that the conditions are right for collapse:
1396 		 *
1397 		 * The object exists and the backing object exists.
1398 		 */
1399 		if (object == NULL)
1400 			break;
1401 
1402 		if ((backing_object = object->backing_object) == NULL)
1403 			break;
1404 
1405 		/*
1406 		 * we check the backing object first, because it is most likely
1407 		 * not collapsable.
1408 		 */
1409 		if (backing_object->handle != NULL ||
1410 		    (backing_object->type != OBJT_DEFAULT &&
1411 		     backing_object->type != OBJT_SWAP) ||
1412 		    (backing_object->flags & OBJ_DEAD) ||
1413 		    object->handle != NULL ||
1414 		    (object->type != OBJT_DEFAULT &&
1415 		     object->type != OBJT_SWAP) ||
1416 		    (object->flags & OBJ_DEAD)) {
1417 			break;
1418 		}
1419 
1420 		if (
1421 		    object->paging_in_progress != 0 ||
1422 		    backing_object->paging_in_progress != 0
1423 		) {
1424 			vm_object_qcollapse(object);
1425 			break;
1426 		}
1427 
1428 		/*
1429 		 * We know that we can either collapse the backing object (if
1430 		 * the parent is the only reference to it) or (perhaps) have
1431 		 * the parent bypass the object if the parent happens to shadow
1432 		 * all the resident pages in the entire backing object.
1433 		 *
1434 		 * This is ignoring pager-backed pages such as swap pages.
1435 		 * vm_object_backing_scan fails the shadowing test in this
1436 		 * case.
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 			if (backing_object->type == OBJT_SWAP) {
1449 				vm_object_pip_add(backing_object, 1);
1450 
1451 				/*
1452 				 * scrap the paging_offset junk and do a
1453 				 * discrete copy.  This also removes major
1454 				 * assumptions about how the swap-pager
1455 				 * works from where it doesn't belong.  The
1456 				 * new swapper is able to optimize the
1457 				 * destroy-source case.
1458 				 */
1459 				vm_object_pip_add(object, 1);
1460 				swap_pager_copy(
1461 				    backing_object,
1462 				    object,
1463 				    OFF_TO_IDX(object->backing_object_offset), TRUE);
1464 				vm_object_pip_wakeup(object);
1465 
1466 				vm_object_pip_wakeup(backing_object);
1467 			}
1468 			/*
1469 			 * Object now shadows whatever backing_object did.
1470 			 * Note that the reference to
1471 			 * backing_object->backing_object moves from within
1472 			 * backing_object to within object.
1473 			 */
1474 			TAILQ_REMOVE(
1475 			    &object->backing_object->shadow_head,
1476 			    object,
1477 			    shadow_list
1478 			);
1479 			object->backing_object->shadow_count--;
1480 			object->backing_object->generation++;
1481 			if (backing_object->backing_object) {
1482 				TAILQ_REMOVE(
1483 				    &backing_object->backing_object->shadow_head,
1484 				    backing_object,
1485 				    shadow_list
1486 				);
1487 				backing_object->backing_object->shadow_count--;
1488 				backing_object->backing_object->generation++;
1489 			}
1490 			object->backing_object = backing_object->backing_object;
1491 			if (object->backing_object) {
1492 				TAILQ_INSERT_TAIL(
1493 				    &object->backing_object->shadow_head,
1494 				    object,
1495 				    shadow_list
1496 				);
1497 				object->backing_object->shadow_count++;
1498 				object->backing_object->generation++;
1499 			}
1500 
1501 			object->backing_object_offset +=
1502 			    backing_object->backing_object_offset;
1503 
1504 			/*
1505 			 * Discard backing_object.
1506 			 *
1507 			 * Since the backing object has no pages, no pager left,
1508 			 * and no object references within it, all that is
1509 			 * necessary is to dispose of it.
1510 			 */
1511 			KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1512 			KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object));
1513 
1514 			TAILQ_REMOVE(
1515 			    &vm_object_list,
1516 			    backing_object,
1517 			    object_list
1518 			);
1519 
1520 			uma_zfree(obj_zone, backing_object);
1521 
1522 			object_collapses++;
1523 		} else {
1524 			vm_object_t new_backing_object;
1525 
1526 			/*
1527 			 * If we do not entirely shadow the backing object,
1528 			 * there is nothing we can do so we give up.
1529 			 */
1530 			if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1531 				break;
1532 			}
1533 
1534 			/*
1535 			 * Make the parent shadow the next object in the
1536 			 * chain.  Deallocating backing_object will not remove
1537 			 * it, since its reference count is at least 2.
1538 			 */
1539 			TAILQ_REMOVE(
1540 			    &backing_object->shadow_head,
1541 			    object,
1542 			    shadow_list
1543 			);
1544 			backing_object->shadow_count--;
1545 			backing_object->generation++;
1546 
1547 			new_backing_object = backing_object->backing_object;
1548 			if ((object->backing_object = new_backing_object) != NULL) {
1549 				vm_object_reference(new_backing_object);
1550 				TAILQ_INSERT_TAIL(
1551 				    &new_backing_object->shadow_head,
1552 				    object,
1553 				    shadow_list
1554 				);
1555 				new_backing_object->shadow_count++;
1556 				new_backing_object->generation++;
1557 				object->backing_object_offset +=
1558 					backing_object->backing_object_offset;
1559 			}
1560 
1561 			/*
1562 			 * Drop the reference count on backing_object. Since
1563 			 * its ref_count was at least 2, it will not vanish;
1564 			 * so we don't need to call vm_object_deallocate, but
1565 			 * we do anyway.
1566 			 */
1567 			vm_object_deallocate(backing_object);
1568 			object_bypasses++;
1569 		}
1570 
1571 		/*
1572 		 * Try again with this object's new backing object.
1573 		 */
1574 	}
1575 }
1576 
1577 /*
1578  *	vm_object_page_remove: [internal]
1579  *
1580  *	Removes all physical pages in the specified
1581  *	object range from the object's list of pages.
1582  *
1583  *	The object must be locked.
1584  */
1585 void
1586 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, boolean_t clean_only)
1587 {
1588 	vm_page_t p, next;
1589 	unsigned int size;
1590 	int all;
1591 
1592 	GIANT_REQUIRED;
1593 
1594 	if (object == NULL ||
1595 	    object->resident_page_count == 0)
1596 		return;
1597 
1598 	all = ((end == 0) && (start == 0));
1599 
1600 	/*
1601 	 * Since physically-backed objects do not use managed pages, we can't
1602 	 * remove pages from the object (we must instead remove the page
1603 	 * references, and then destroy the object).
1604 	 */
1605 	KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object"));
1606 
1607 	vm_object_pip_add(object, 1);
1608 again:
1609 	size = end - start;
1610 	if (all || size > object->resident_page_count / 4) {
1611 		for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
1612 			next = TAILQ_NEXT(p, listq);
1613 			if (all || ((start <= p->pindex) && (p->pindex < end))) {
1614 				if (p->wire_count != 0) {
1615 					vm_page_protect(p, VM_PROT_NONE);
1616 					if (!clean_only)
1617 						p->valid = 0;
1618 					continue;
1619 				}
1620 
1621 				/*
1622 				 * The busy flags are only cleared at
1623 				 * interrupt -- minimize the spl transitions
1624 				 */
1625  				if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1626  					goto again;
1627 
1628 				if (clean_only && p->valid) {
1629 					vm_page_test_dirty(p);
1630 					if (p->valid & p->dirty)
1631 						continue;
1632 				}
1633 
1634 				vm_page_busy(p);
1635 				vm_page_protect(p, VM_PROT_NONE);
1636 				vm_page_free(p);
1637 			}
1638 		}
1639 	} else {
1640 		while (size > 0) {
1641 			if ((p = vm_page_lookup(object, start)) != 0) {
1642 
1643 				if (p->wire_count != 0) {
1644 					vm_page_protect(p, VM_PROT_NONE);
1645 					if (!clean_only)
1646 						p->valid = 0;
1647 					start += 1;
1648 					size -= 1;
1649 					continue;
1650 				}
1651 
1652 				/*
1653 				 * The busy flags are only cleared at
1654 				 * interrupt -- minimize the spl transitions
1655 				 */
1656  				if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1657 					goto again;
1658 
1659 				if (clean_only && p->valid) {
1660 					vm_page_test_dirty(p);
1661 					if (p->valid & p->dirty) {
1662 						start += 1;
1663 						size -= 1;
1664 						continue;
1665 					}
1666 				}
1667 
1668 				vm_page_busy(p);
1669 				vm_page_protect(p, VM_PROT_NONE);
1670 				vm_page_free(p);
1671 			}
1672 			start += 1;
1673 			size -= 1;
1674 		}
1675 	}
1676 	vm_object_pip_wakeup(object);
1677 }
1678 
1679 /*
1680  *	Routine:	vm_object_coalesce
1681  *	Function:	Coalesces two objects backing up adjoining
1682  *			regions of memory into a single object.
1683  *
1684  *	returns TRUE if objects were combined.
1685  *
1686  *	NOTE:	Only works at the moment if the second object is NULL -
1687  *		if it's not, which object do we lock first?
1688  *
1689  *	Parameters:
1690  *		prev_object	First object to coalesce
1691  *		prev_offset	Offset into prev_object
1692  *		next_object	Second object into coalesce
1693  *		next_offset	Offset into next_object
1694  *
1695  *		prev_size	Size of reference to prev_object
1696  *		next_size	Size of reference to next_object
1697  *
1698  *	Conditions:
1699  *	The object must *not* be locked.
1700  */
1701 boolean_t
1702 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, vm_size_t prev_size, vm_size_t next_size)
1703 {
1704 	vm_pindex_t next_pindex;
1705 
1706 	GIANT_REQUIRED;
1707 
1708 	if (prev_object == NULL) {
1709 		return (TRUE);
1710 	}
1711 
1712 	if (prev_object->type != OBJT_DEFAULT &&
1713 	    prev_object->type != OBJT_SWAP) {
1714 		return (FALSE);
1715 	}
1716 
1717 	/*
1718 	 * Try to collapse the object first
1719 	 */
1720 	vm_object_collapse(prev_object);
1721 
1722 	/*
1723 	 * Can't coalesce if: . more than one reference . paged out . shadows
1724 	 * another object . has a copy elsewhere (any of which mean that the
1725 	 * pages not mapped to prev_entry may be in use anyway)
1726 	 */
1727 	if (prev_object->backing_object != NULL) {
1728 		return (FALSE);
1729 	}
1730 
1731 	prev_size >>= PAGE_SHIFT;
1732 	next_size >>= PAGE_SHIFT;
1733 	next_pindex = prev_pindex + prev_size;
1734 
1735 	if ((prev_object->ref_count > 1) &&
1736 	    (prev_object->size != next_pindex)) {
1737 		return (FALSE);
1738 	}
1739 
1740 	/*
1741 	 * Remove any pages that may still be in the object from a previous
1742 	 * deallocation.
1743 	 */
1744 	if (next_pindex < prev_object->size) {
1745 		vm_object_page_remove(prev_object,
1746 				      next_pindex,
1747 				      next_pindex + next_size, FALSE);
1748 		if (prev_object->type == OBJT_SWAP)
1749 			swap_pager_freespace(prev_object,
1750 					     next_pindex, next_size);
1751 	}
1752 
1753 	/*
1754 	 * Extend the object if necessary.
1755 	 */
1756 	if (next_pindex + next_size > prev_object->size)
1757 		prev_object->size = next_pindex + next_size;
1758 
1759 	return (TRUE);
1760 }
1761 
1762 void
1763 vm_object_set_writeable_dirty(vm_object_t object)
1764 {
1765 	struct vnode *vp;
1766 
1767 	vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1768 	if (object->type == OBJT_VNODE &&
1769 	    (vp = (struct vnode *)object->handle) != NULL) {
1770 		if ((vp->v_flag & VOBJDIRTY) == 0) {
1771 			mtx_lock(&vp->v_interlock);
1772 			vp->v_flag |= VOBJDIRTY;
1773 			mtx_unlock(&vp->v_interlock);
1774 		}
1775 	}
1776 }
1777 
1778 #include "opt_ddb.h"
1779 #ifdef DDB
1780 #include <sys/kernel.h>
1781 
1782 #include <sys/cons.h>
1783 
1784 #include <ddb/ddb.h>
1785 
1786 static int
1787 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
1788 {
1789 	vm_map_t tmpm;
1790 	vm_map_entry_t tmpe;
1791 	vm_object_t obj;
1792 	int entcount;
1793 
1794 	if (map == 0)
1795 		return 0;
1796 
1797 	if (entry == 0) {
1798 		tmpe = map->header.next;
1799 		entcount = map->nentries;
1800 		while (entcount-- && (tmpe != &map->header)) {
1801 			if (_vm_object_in_map(map, object, tmpe)) {
1802 				return 1;
1803 			}
1804 			tmpe = tmpe->next;
1805 		}
1806 	} else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
1807 		tmpm = entry->object.sub_map;
1808 		tmpe = tmpm->header.next;
1809 		entcount = tmpm->nentries;
1810 		while (entcount-- && tmpe != &tmpm->header) {
1811 			if (_vm_object_in_map(tmpm, object, tmpe)) {
1812 				return 1;
1813 			}
1814 			tmpe = tmpe->next;
1815 		}
1816 	} else if ((obj = entry->object.vm_object) != NULL) {
1817 		for (; obj; obj = obj->backing_object)
1818 			if (obj == object) {
1819 				return 1;
1820 			}
1821 	}
1822 	return 0;
1823 }
1824 
1825 static int
1826 vm_object_in_map(vm_object_t object)
1827 {
1828 	struct proc *p;
1829 
1830 	/* sx_slock(&allproc_lock); */
1831 	LIST_FOREACH(p, &allproc, p_list) {
1832 		if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1833 			continue;
1834 		if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
1835 			/* sx_sunlock(&allproc_lock); */
1836 			return 1;
1837 		}
1838 	}
1839 	/* sx_sunlock(&allproc_lock); */
1840 	if (_vm_object_in_map(kernel_map, object, 0))
1841 		return 1;
1842 	if (_vm_object_in_map(kmem_map, object, 0))
1843 		return 1;
1844 	if (_vm_object_in_map(pager_map, object, 0))
1845 		return 1;
1846 	if (_vm_object_in_map(buffer_map, object, 0))
1847 		return 1;
1848 	return 0;
1849 }
1850 
1851 DB_SHOW_COMMAND(vmochk, vm_object_check)
1852 {
1853 	vm_object_t object;
1854 
1855 	/*
1856 	 * make sure that internal objs are in a map somewhere
1857 	 * and none have zero ref counts.
1858 	 */
1859 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
1860 		if (object->handle == NULL &&
1861 		    (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1862 			if (object->ref_count == 0) {
1863 				db_printf("vmochk: internal obj has zero ref count: %ld\n",
1864 					(long)object->size);
1865 			}
1866 			if (!vm_object_in_map(object)) {
1867 				db_printf(
1868 			"vmochk: internal obj is not in a map: "
1869 			"ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1870 				    object->ref_count, (u_long)object->size,
1871 				    (u_long)object->size,
1872 				    (void *)object->backing_object);
1873 			}
1874 		}
1875 	}
1876 }
1877 
1878 /*
1879  *	vm_object_print:	[ debug ]
1880  */
1881 DB_SHOW_COMMAND(object, vm_object_print_static)
1882 {
1883 	/* XXX convert args. */
1884 	vm_object_t object = (vm_object_t)addr;
1885 	boolean_t full = have_addr;
1886 
1887 	vm_page_t p;
1888 
1889 	/* XXX count is an (unused) arg.  Avoid shadowing it. */
1890 #define	count	was_count
1891 
1892 	int count;
1893 
1894 	if (object == NULL)
1895 		return;
1896 
1897 	db_iprintf(
1898 	    "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1899 	    object, (int)object->type, (u_long)object->size,
1900 	    object->resident_page_count, object->ref_count, object->flags);
1901 	/*
1902 	 * XXX no %qd in kernel.  Truncate object->backing_object_offset.
1903 	 */
1904 	db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1905 	    object->shadow_count,
1906 	    object->backing_object ? object->backing_object->ref_count : 0,
1907 	    object->backing_object, (long)object->backing_object_offset);
1908 
1909 	if (!full)
1910 		return;
1911 
1912 	db_indent += 2;
1913 	count = 0;
1914 	TAILQ_FOREACH(p, &object->memq, listq) {
1915 		if (count == 0)
1916 			db_iprintf("memory:=");
1917 		else if (count == 6) {
1918 			db_printf("\n");
1919 			db_iprintf(" ...");
1920 			count = 0;
1921 		} else
1922 			db_printf(",");
1923 		count++;
1924 
1925 		db_printf("(off=0x%lx,page=0x%lx)",
1926 		    (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1927 	}
1928 	if (count != 0)
1929 		db_printf("\n");
1930 	db_indent -= 2;
1931 }
1932 
1933 /* XXX. */
1934 #undef count
1935 
1936 /* XXX need this non-static entry for calling from vm_map_print. */
1937 void
1938 vm_object_print(
1939         /* db_expr_t */ long addr,
1940 	boolean_t have_addr,
1941 	/* db_expr_t */ long count,
1942 	char *modif)
1943 {
1944 	vm_object_print_static(addr, have_addr, count, modif);
1945 }
1946 
1947 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1948 {
1949 	vm_object_t object;
1950 	int nl = 0;
1951 	int c;
1952 
1953 	TAILQ_FOREACH(object, &vm_object_list, object_list) {
1954 		vm_pindex_t idx, fidx;
1955 		vm_pindex_t osize;
1956 		vm_offset_t pa = -1, padiff;
1957 		int rcount;
1958 		vm_page_t m;
1959 
1960 		db_printf("new object: %p\n", (void *)object);
1961 		if (nl > 18) {
1962 			c = cngetc();
1963 			if (c != ' ')
1964 				return;
1965 			nl = 0;
1966 		}
1967 		nl++;
1968 		rcount = 0;
1969 		fidx = 0;
1970 		osize = object->size;
1971 		if (osize > 128)
1972 			osize = 128;
1973 		for (idx = 0; idx < osize; idx++) {
1974 			m = vm_page_lookup(object, idx);
1975 			if (m == NULL) {
1976 				if (rcount) {
1977 					db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1978 						(long)fidx, rcount, (long)pa);
1979 					if (nl > 18) {
1980 						c = cngetc();
1981 						if (c != ' ')
1982 							return;
1983 						nl = 0;
1984 					}
1985 					nl++;
1986 					rcount = 0;
1987 				}
1988 				continue;
1989 			}
1990 
1991 
1992 			if (rcount &&
1993 				(VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1994 				++rcount;
1995 				continue;
1996 			}
1997 			if (rcount) {
1998 				padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1999 				padiff >>= PAGE_SHIFT;
2000 				padiff &= PQ_L2_MASK;
2001 				if (padiff == 0) {
2002 					pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
2003 					++rcount;
2004 					continue;
2005 				}
2006 				db_printf(" index(%ld)run(%d)pa(0x%lx)",
2007 					(long)fidx, rcount, (long)pa);
2008 				db_printf("pd(%ld)\n", (long)padiff);
2009 				if (nl > 18) {
2010 					c = cngetc();
2011 					if (c != ' ')
2012 						return;
2013 					nl = 0;
2014 				}
2015 				nl++;
2016 			}
2017 			fidx = idx;
2018 			pa = VM_PAGE_TO_PHYS(m);
2019 			rcount = 1;
2020 		}
2021 		if (rcount) {
2022 			db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2023 				(long)fidx, rcount, (long)pa);
2024 			if (nl > 18) {
2025 				c = cngetc();
2026 				if (c != ' ')
2027 					return;
2028 				nl = 0;
2029 			}
2030 			nl++;
2031 		}
2032 	}
2033 }
2034 #endif /* DDB */
2035