c4f09c5023441b808fe14d43735e562ac59eeb27
[dragonfly.git] / sys / vm / vm_object.c
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  */
68
69 /*
70  *      Virtual memory object module.
71  */
72
73 #include <sys/param.h>
74 #include <sys/systm.h>
75 #include <sys/proc.h>           /* for curproc, pageproc */
76 #include <sys/thread.h>
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 #include <sys/refcount.h>
84
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_pager.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_zone.h>
97
98 #define EASY_SCAN_FACTOR        8
99
100 static void     vm_object_qcollapse(vm_object_t object,
101                                     vm_object_t backing_object);
102 static int      vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
103                                              int pagerflags);
104 static void     vm_object_lock_init(vm_object_t);
105
106
107 /*
108  *      Virtual memory objects maintain the actual data
109  *      associated with allocated virtual memory.  A given
110  *      page of memory exists within exactly one object.
111  *
112  *      An object is only deallocated when all "references"
113  *      are given up.  Only one "reference" to a given
114  *      region of an object should be writeable.
115  *
116  *      Associated with each object is a list of all resident
117  *      memory pages belonging to that object; this list is
118  *      maintained by the "vm_page" module, and locked by the object's
119  *      lock.
120  *
121  *      Each object also records a "pager" routine which is
122  *      used to retrieve (and store) pages to the proper backing
123  *      storage.  In addition, objects may be backed by other
124  *      objects from which they were virtual-copied.
125  *
126  *      The only items within the object structure which are
127  *      modified after time of creation are:
128  *              reference count         locked by object's lock
129  *              pager routine           locked by object's lock
130  *
131  */
132
133 struct object_q vm_object_list;         /* locked by vmobj_token */
134 struct vm_object kernel_object;
135
136 static long vm_object_count;            /* locked by vmobj_token */
137 extern int vm_pageout_page_count;
138
139 static long object_collapses;
140 static long object_bypasses;
141 static int next_index;
142 static vm_zone_t obj_zone;
143 static struct vm_zone obj_zone_store;
144 #define VM_OBJECTS_INIT 256
145 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
146
147 /*
148  * Misc low level routines
149  */
150 static void
151 vm_object_lock_init(vm_object_t obj)
152 {
153 #if defined(DEBUG_LOCKS)
154         int i;
155
156         obj->debug_hold_bitmap = 0;
157         obj->debug_hold_ovfl = 0;
158         for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) {
159                 obj->debug_hold_thrs[i] = NULL;
160                 obj->debug_hold_file[i] = NULL;
161                 obj->debug_hold_line[i] = 0;
162         }
163 #endif
164 }
165
166 void
167 vm_object_lock_swap(void)
168 {
169         lwkt_token_swap();
170 }
171
172 void
173 vm_object_lock(vm_object_t obj)
174 {
175         lwkt_gettoken(&obj->token);
176 }
177
178 /*
179  * Returns TRUE on sucesss
180  */
181 static int
182 vm_object_lock_try(vm_object_t obj)
183 {
184         return(lwkt_trytoken(&obj->token));
185 }
186
187 void
188 vm_object_lock_shared(vm_object_t obj)
189 {
190         lwkt_gettoken_shared(&obj->token);
191 }
192
193 void
194 vm_object_unlock(vm_object_t obj)
195 {
196         lwkt_reltoken(&obj->token);
197 }
198
199 static __inline void
200 vm_object_assert_held(vm_object_t obj)
201 {
202         ASSERT_LWKT_TOKEN_HELD(&obj->token);
203 }
204
205 void
206 #ifndef DEBUG_LOCKS
207 vm_object_hold(vm_object_t obj)
208 #else
209 debugvm_object_hold(vm_object_t obj, char *file, int line)
210 #endif
211 {
212         KKASSERT(obj != NULL);
213
214         /*
215          * Object must be held (object allocation is stable due to callers
216          * context, typically already holding the token on a parent object)
217          * prior to potentially blocking on the lock, otherwise the object
218          * can get ripped away from us.
219          */
220         refcount_acquire(&obj->hold_count);
221         vm_object_lock(obj);
222
223 #if defined(DEBUG_LOCKS)
224         int i;
225         u_int mask;
226
227         for (;;) {
228                 mask = ~obj->debug_hold_bitmap;
229                 cpu_ccfence();
230                 if (mask == 0xFFFFFFFFU) {
231                         if (obj->debug_hold_ovfl == 0)
232                                 obj->debug_hold_ovfl = 1;
233                         break;
234                 }
235                 i = ffs(mask) - 1;
236                 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask,
237                                       ~mask | (1 << i))) {
238                         obj->debug_hold_bitmap |= (1 << i);
239                         obj->debug_hold_thrs[i] = curthread;
240                         obj->debug_hold_file[i] = file;
241                         obj->debug_hold_line[i] = line;
242                         break;
243                 }
244         }
245 #endif
246 }
247
248 int
249 #ifndef DEBUG_LOCKS
250 vm_object_hold_try(vm_object_t obj)
251 #else
252 debugvm_object_hold_try(vm_object_t obj, char *file, int line)
253 #endif
254 {
255         KKASSERT(obj != NULL);
256
257         /*
258          * Object must be held (object allocation is stable due to callers
259          * context, typically already holding the token on a parent object)
260          * prior to potentially blocking on the lock, otherwise the object
261          * can get ripped away from us.
262          */
263         refcount_acquire(&obj->hold_count);
264         if (vm_object_lock_try(obj) == 0) {
265                 if (refcount_release(&obj->hold_count)) {
266                         if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD))
267                                 zfree(obj_zone, obj);
268                 }
269                 return(0);
270         }
271
272 #if defined(DEBUG_LOCKS)
273         int i;
274         u_int mask;
275
276         for (;;) {
277                 mask = ~obj->debug_hold_bitmap;
278                 cpu_ccfence();
279                 if (mask == 0xFFFFFFFFU) {
280                         if (obj->debug_hold_ovfl == 0)
281                                 obj->debug_hold_ovfl = 1;
282                         break;
283                 }
284                 i = ffs(mask) - 1;
285                 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask,
286                                       ~mask | (1 << i))) {
287                         obj->debug_hold_bitmap |= (1 << i);
288                         obj->debug_hold_thrs[i] = curthread;
289                         obj->debug_hold_file[i] = file;
290                         obj->debug_hold_line[i] = line;
291                         break;
292                 }
293         }
294 #endif
295         return(1);
296 }
297
298 void
299 #ifndef DEBUG_LOCKS
300 vm_object_hold_shared(vm_object_t obj)
301 #else
302 debugvm_object_hold_shared(vm_object_t obj, char *file, int line)
303 #endif
304 {
305         KKASSERT(obj != NULL);
306
307         /*
308          * Object must be held (object allocation is stable due to callers
309          * context, typically already holding the token on a parent object)
310          * prior to potentially blocking on the lock, otherwise the object
311          * can get ripped away from us.
312          */
313         refcount_acquire(&obj->hold_count);
314         vm_object_lock_shared(obj);
315
316 #if defined(DEBUG_LOCKS)
317         int i;
318         u_int mask;
319
320         for (;;) {
321                 mask = ~obj->debug_hold_bitmap;
322                 cpu_ccfence();
323                 if (mask == 0xFFFFFFFFU) {
324                         if (obj->debug_hold_ovfl == 0)
325                                 obj->debug_hold_ovfl = 1;
326                         break;
327                 }
328                 i = ffs(mask) - 1;
329                 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask,
330                                       ~mask | (1 << i))) {
331                         obj->debug_hold_bitmap |= (1 << i);
332                         obj->debug_hold_thrs[i] = curthread;
333                         obj->debug_hold_file[i] = file;
334                         obj->debug_hold_line[i] = line;
335                         break;
336                 }
337         }
338 #endif
339 }
340
341 /*
342  * Drop the token and hold_count on the object.
343  */
344 void
345 vm_object_drop(vm_object_t obj)
346 {
347         if (obj == NULL)
348                 return;
349
350 #if defined(DEBUG_LOCKS)
351         int found = 0;
352         int i;
353
354         for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) {
355                 if ((obj->debug_hold_bitmap & (1 << i)) &&
356                     (obj->debug_hold_thrs[i] == curthread)) {
357                         obj->debug_hold_bitmap &= ~(1 << i);
358                         obj->debug_hold_thrs[i] = NULL;
359                         obj->debug_hold_file[i] = NULL;
360                         obj->debug_hold_line[i] = 0;
361                         found = 1;
362                         break;
363                 }
364         }
365
366         if (found == 0 && obj->debug_hold_ovfl == 0)
367                 panic("vm_object: attempt to drop hold on non-self-held obj");
368 #endif
369
370         /*
371          * No new holders should be possible once we drop hold_count 1->0 as
372          * there is no longer any way to reference the object.
373          */
374         KKASSERT(obj->hold_count > 0);
375         if (refcount_release(&obj->hold_count)) {
376                 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD)) {
377                         vm_object_unlock(obj);
378                         zfree(obj_zone, obj);
379                 } else {
380                         vm_object_unlock(obj);
381                 }
382         } else {
383                 vm_object_unlock(obj);
384         }
385 }
386
387 /*
388  * Initialize a freshly allocated object, returning a held object.
389  *
390  * Used only by vm_object_allocate() and zinitna().
391  *
392  * No requirements.
393  */
394 void
395 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
396 {
397         int incr;
398
399         RB_INIT(&object->rb_memq);
400         LIST_INIT(&object->shadow_head);
401         lwkt_token_init(&object->token, "vmobj");
402
403         object->type = type;
404         object->size = size;
405         object->ref_count = 1;
406         object->hold_count = 0;
407         object->flags = 0;
408         if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
409                 vm_object_set_flag(object, OBJ_ONEMAPPING);
410         object->paging_in_progress = 0;
411         object->resident_page_count = 0;
412         object->agg_pv_list_count = 0;
413         object->shadow_count = 0;
414 #ifdef SMP
415         /* cpu localization twist */
416         object->pg_color = (int)(intptr_t)curthread;
417 #else
418         object->pg_color = next_index;
419 #endif
420         if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
421                 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
422         else
423                 incr = size;
424         next_index = (next_index + incr) & PQ_L2_MASK;
425         object->handle = NULL;
426         object->backing_object = NULL;
427         object->backing_object_offset = (vm_ooffset_t)0;
428
429         object->generation++;
430         object->swblock_count = 0;
431         RB_INIT(&object->swblock_root);
432         vm_object_lock_init(object);
433
434         vm_object_hold(object);
435         lwkt_gettoken(&vmobj_token);
436         TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
437         vm_object_count++;
438         lwkt_reltoken(&vmobj_token);
439 }
440
441 /*
442  * Initialize the VM objects module.
443  *
444  * Called from the low level boot code only.
445  */
446 void
447 vm_object_init(void)
448 {
449         TAILQ_INIT(&vm_object_list);
450         
451         _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd),
452                             &kernel_object);
453         vm_object_drop(&kernel_object);
454
455         obj_zone = &obj_zone_store;
456         zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
457                 vm_objects_init, VM_OBJECTS_INIT);
458 }
459
460 void
461 vm_object_init2(void)
462 {
463         zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1);
464 }
465
466 /*
467  * Allocate and return a new object of the specified type and size.
468  *
469  * No requirements.
470  */
471 vm_object_t
472 vm_object_allocate(objtype_t type, vm_pindex_t size)
473 {
474         vm_object_t result;
475
476         result = (vm_object_t) zalloc(obj_zone);
477
478         _vm_object_allocate(type, size, result);
479         vm_object_drop(result);
480
481         return (result);
482 }
483
484 /*
485  * This version returns a held object, allowing further atomic initialization
486  * of the object.
487  */
488 vm_object_t
489 vm_object_allocate_hold(objtype_t type, vm_pindex_t size)
490 {
491         vm_object_t result;
492
493         result = (vm_object_t) zalloc(obj_zone);
494
495         _vm_object_allocate(type, size, result);
496
497         return (result);
498 }
499
500 /*
501  * Add an additional reference to a vm_object.  The object must already be
502  * held.  The original non-lock version is no longer supported.  The object
503  * must NOT be chain locked by anyone at the time the reference is added.
504  *
505  * Referencing a chain-locked object can blow up the fairly sensitive
506  * ref_count and shadow_count tests in the deallocator.  Most callers
507  * will call vm_object_chain_wait() prior to calling
508  * vm_object_reference_locked() to avoid the case.
509  *
510  * The object must be held.
511  */
512 void
513 vm_object_reference_locked(vm_object_t object)
514 {
515         KKASSERT(object != NULL);
516         ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
517         KKASSERT((object->flags & OBJ_CHAINLOCK) == 0);
518         object->ref_count++;
519         if (object->type == OBJT_VNODE) {
520                 vref(object->handle);
521                 /* XXX what if the vnode is being destroyed? */
522         }
523 }
524
525 /*
526  * Object OBJ_CHAINLOCK lock handling.
527  *
528  * The caller can chain-lock backing objects recursively and then
529  * use vm_object_chain_release_all() to undo the whole chain.
530  *
531  * Chain locks are used to prevent collapses and are only applicable
532  * to OBJT_DEFAULT and OBJT_SWAP objects.  Chain locking operations
533  * on other object types are ignored.  This is also important because
534  * it allows e.g. the vnode underlying a memory mapping to take concurrent
535  * faults.
536  *
537  * The object must usually be held on entry, though intermediate
538  * objects need not be held on release.
539  */
540 void
541 vm_object_chain_wait(vm_object_t object)
542 {
543         ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
544         while (object->flags & OBJ_CHAINLOCK) {
545                 vm_object_set_flag(object, OBJ_CHAINWANT);
546                 tsleep(object, 0, "objchain", 0);
547         }
548 }
549
550 void
551 vm_object_chain_acquire(vm_object_t object)
552 {
553         if (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) {
554                 vm_object_chain_wait(object);
555                 vm_object_set_flag(object, OBJ_CHAINLOCK);
556         }
557 }
558
559 void
560 vm_object_chain_release(vm_object_t object)
561 {
562         ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
563         if (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) {
564                 KKASSERT(object->flags & OBJ_CHAINLOCK);
565                 if (object->flags & OBJ_CHAINWANT) {
566                         vm_object_clear_flag(object,
567                                              OBJ_CHAINLOCK | OBJ_CHAINWANT);
568                         wakeup(object);
569                 } else {
570                         vm_object_clear_flag(object, OBJ_CHAINLOCK);
571                 }
572         }
573 }
574
575 /*
576  * This releases the entire chain of objects from first_object to and
577  * including stopobj, flowing through object->backing_object.
578  *
579  * We release stopobj first as an optimization as this object is most
580  * likely to be shared across multiple processes.
581  */
582 void
583 vm_object_chain_release_all(vm_object_t first_object, vm_object_t stopobj)
584 {
585         vm_object_t backing_object;
586         vm_object_t object;
587
588         vm_object_chain_release(stopobj);
589         object = first_object;
590
591         while (object != stopobj) {
592                 KKASSERT(object);
593                 if (object != first_object)
594                         vm_object_hold(object);
595                 backing_object = object->backing_object;
596                 vm_object_chain_release(object);
597                 if (object != first_object)
598                         vm_object_drop(object);
599                 object = backing_object;
600         }
601 }
602
603 /*
604  * Dereference an object and its underlying vnode.
605  *
606  * The object must be held and will be held on return.
607  */
608 static void
609 vm_object_vndeallocate(vm_object_t object)
610 {
611         struct vnode *vp = (struct vnode *) object->handle;
612
613         KASSERT(object->type == OBJT_VNODE,
614             ("vm_object_vndeallocate: not a vnode object"));
615         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
616         ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
617 #ifdef INVARIANTS
618         if (object->ref_count == 0) {
619                 vprint("vm_object_vndeallocate", vp);
620                 panic("vm_object_vndeallocate: bad object reference count");
621         }
622 #endif
623         object->ref_count--;
624         if (object->ref_count == 0)
625                 vclrflags(vp, VTEXT);
626         vrele(vp);
627 }
628
629 /*
630  * Release a reference to the specified object, gained either through a
631  * vm_object_allocate or a vm_object_reference call.  When all references
632  * are gone, storage associated with this object may be relinquished.
633  *
634  * The caller does not have to hold the object locked but must have control
635  * over the reference in question in order to guarantee that the object
636  * does not get ripped out from under us.
637  */
638 void
639 vm_object_deallocate(vm_object_t object)
640 {
641         if (object) {
642                 vm_object_hold(object);
643                 vm_object_deallocate_locked(object);
644                 vm_object_drop(object);
645         }
646 }
647
648 void
649 vm_object_deallocate_locked(vm_object_t object)
650 {
651         struct vm_object_dealloc_list *dlist = NULL;
652         struct vm_object_dealloc_list *dtmp;
653         vm_object_t temp;
654         int must_drop = 0;
655
656         /*
657          * We may chain deallocate object, but additional objects may
658          * collect on the dlist which also have to be deallocated.  We
659          * must avoid a recursion, vm_object chains can get deep.
660          */
661 again:
662         while (object != NULL) {
663 #if 0
664                 /*
665                  * Don't rip a ref_count out from under an object undergoing
666                  * collapse, it will confuse the collapse code.
667                  */
668                 vm_object_chain_wait(object);
669 #endif
670                 if (object->type == OBJT_VNODE) {
671                         vm_object_vndeallocate(object);
672                         break;
673                 }
674
675                 if (object->ref_count == 0) {
676                         panic("vm_object_deallocate: object deallocated "
677                               "too many times: %d", object->type);
678                 }
679                 if (object->ref_count > 2) {
680                         object->ref_count--;
681                         break;
682                 }
683
684                 /*
685                  * Here on ref_count of one or two, which are special cases for
686                  * objects.
687                  *
688                  * Nominal ref_count > 1 case if the second ref is not from
689                  * a shadow.
690                  */
691                 if (object->ref_count == 2 && object->shadow_count == 0) {
692                         vm_object_set_flag(object, OBJ_ONEMAPPING);
693                         object->ref_count--;
694                         break;
695                 }
696
697                 /*
698                  * If the second ref is from a shadow we chain along it
699                  * upwards if object's handle is exhausted.
700                  *
701                  * We have to decrement object->ref_count before potentially
702                  * collapsing the first shadow object or the collapse code
703                  * will not be able to handle the degenerate case to remove
704                  * object.  However, if we do it too early the object can
705                  * get ripped out from under us.
706                  */
707                 if (object->ref_count == 2 && object->shadow_count == 1 &&
708                     object->handle == NULL && (object->type == OBJT_DEFAULT ||
709                                                object->type == OBJT_SWAP)) {
710                         temp = LIST_FIRST(&object->shadow_head);
711                         KKASSERT(temp != NULL);
712                         vm_object_hold(temp);
713
714                         /*
715                          * Wait for any paging to complete so the collapse
716                          * doesn't (or isn't likely to) qcollapse.  pip
717                          * waiting must occur before we acquire the
718                          * chainlock.
719                          */
720                         while (
721                                 temp->paging_in_progress ||
722                                 object->paging_in_progress
723                         ) {
724                                 vm_object_pip_wait(temp, "objde1");
725                                 vm_object_pip_wait(object, "objde2");
726                         }
727
728                         /*
729                          * If the parent is locked we have to give up, as
730                          * otherwise we would be acquiring locks in the
731                          * wrong order and potentially deadlock.
732                          */
733                         if (temp->flags & OBJ_CHAINLOCK) {
734                                 vm_object_drop(temp);
735                                 goto skip;
736                         }
737                         vm_object_chain_acquire(temp);
738
739                         /*
740                          * Recheck/retry after the hold and the paging
741                          * wait, both of which can block us.
742                          */
743                         if (object->ref_count != 2 ||
744                             object->shadow_count != 1 ||
745                             object->handle ||
746                             LIST_FIRST(&object->shadow_head) != temp ||
747                             (object->type != OBJT_DEFAULT &&
748                              object->type != OBJT_SWAP)) {
749                                 vm_object_chain_release(temp);
750                                 vm_object_drop(temp);
751                                 continue;
752                         }
753
754                         /*
755                          * We can safely drop object's ref_count now.
756                          */
757                         KKASSERT(object->ref_count == 2);
758                         object->ref_count--;
759
760                         /*
761                          * If our single parent is not collapseable just
762                          * decrement ref_count (2->1) and stop.
763                          */
764                         if (temp->handle || (temp->type != OBJT_DEFAULT &&
765                                              temp->type != OBJT_SWAP)) {
766                                 vm_object_chain_release(temp);
767                                 vm_object_drop(temp);
768                                 break;
769                         }
770
771                         /*
772                          * At this point we have already dropped object's
773                          * ref_count so it is possible for a race to
774                          * deallocate obj out from under us.  Any collapse
775                          * will re-check the situation.  We must not block
776                          * until we are able to collapse.
777                          *
778                          * Bump temp's ref_count to avoid an unwanted
779                          * degenerate recursion (can't call
780                          * vm_object_reference_locked() because it asserts
781                          * that CHAINLOCK is not set).
782                          */
783                         temp->ref_count++;
784                         KKASSERT(temp->ref_count > 1);
785
786                         /*
787                          * Collapse temp, then deallocate the extra ref
788                          * formally.
789                          */
790                         vm_object_collapse(temp, &dlist);
791                         vm_object_chain_release(temp);
792                         if (must_drop) {
793                                 vm_object_lock_swap();
794                                 vm_object_drop(object);
795                         }
796                         object = temp;
797                         must_drop = 1;
798                         continue;
799                 }
800
801                 /*
802                  * Drop the ref and handle termination on the 1->0 transition.
803                  * We may have blocked above so we have to recheck.
804                  */
805 skip:
806                 KKASSERT(object->ref_count != 0);
807                 if (object->ref_count >= 2) {
808                         object->ref_count--;
809                         break;
810                 }
811                 KKASSERT(object->ref_count == 1);
812
813                 /*
814                  * 1->0 transition.  Chain through the backing_object.
815                  * Maintain the ref until we've located the backing object,
816                  * then re-check.
817                  */
818                 while ((temp = object->backing_object) != NULL) {
819                         vm_object_hold(temp);
820                         if (temp == object->backing_object)
821                                 break;
822                         vm_object_drop(temp);
823                 }
824
825                 /*
826                  * 1->0 transition verified, retry if ref_count is no longer
827                  * 1.  Otherwise disconnect the backing_object (temp) and
828                  * clean up.
829                  */
830                 if (object->ref_count != 1) {
831                         vm_object_drop(temp);
832                         continue;
833                 }
834
835                 /*
836                  * It shouldn't be possible for the object to be chain locked
837                  * if we're removing the last ref on it.
838                  */
839                 KKASSERT((object->flags & OBJ_CHAINLOCK) == 0);
840
841                 if (temp) {
842                         LIST_REMOVE(object, shadow_list);
843                         temp->shadow_count--;
844                         temp->generation++;
845                         object->backing_object = NULL;
846                 }
847
848                 --object->ref_count;
849                 if ((object->flags & OBJ_DEAD) == 0)
850                         vm_object_terminate(object);
851                 if (must_drop && temp)
852                         vm_object_lock_swap();
853                 if (must_drop)
854                         vm_object_drop(object);
855                 object = temp;
856                 must_drop = 1;
857         }
858         if (must_drop && object)
859                 vm_object_drop(object);
860
861         /*
862          * Additional tail recursion on dlist.  Avoid a recursion.  Objects
863          * on the dlist have a hold count but are not locked.
864          */
865         if ((dtmp = dlist) != NULL) {
866                 dlist = dtmp->next;
867                 object = dtmp->object;
868                 kfree(dtmp, M_TEMP);
869
870                 vm_object_lock(object); /* already held, add lock */
871                 must_drop = 1;          /* and we're responsible for it */
872                 goto again;
873         }
874 }
875
876 /*
877  * Destroy the specified object, freeing up related resources.
878  *
879  * The object must have zero references.
880  *
881  * The object must held.  The caller is responsible for dropping the object
882  * after terminate returns.  Terminate does NOT drop the object.
883  */
884 static int vm_object_terminate_callback(vm_page_t p, void *data);
885
886 void
887 vm_object_terminate(vm_object_t object)
888 {
889         /*
890          * Make sure no one uses us.  Once we set OBJ_DEAD we should be
891          * able to safely block.
892          */
893         ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
894         KKASSERT((object->flags & OBJ_DEAD) == 0);
895         vm_object_set_flag(object, OBJ_DEAD);
896
897         /*
898          * Wait for the pageout daemon to be done with the object
899          */
900         vm_object_pip_wait(object, "objtrm1");
901
902         KASSERT(!object->paging_in_progress,
903                 ("vm_object_terminate: pageout in progress"));
904
905         /*
906          * Clean and free the pages, as appropriate. All references to the
907          * object are gone, so we don't need to lock it.
908          */
909         if (object->type == OBJT_VNODE) {
910                 struct vnode *vp;
911
912                 /*
913                  * Clean pages and flush buffers.
914                  */
915                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
916
917                 vp = (struct vnode *) object->handle;
918                 vinvalbuf(vp, V_SAVE, 0, 0);
919         }
920
921         /*
922          * Wait for any I/O to complete, after which there had better not
923          * be any references left on the object.
924          */
925         vm_object_pip_wait(object, "objtrm2");
926
927         if (object->ref_count != 0) {
928                 panic("vm_object_terminate: object with references, "
929                       "ref_count=%d", object->ref_count);
930         }
931
932         /*
933          * Now free any remaining pages. For internal objects, this also
934          * removes them from paging queues. Don't free wired pages, just
935          * remove them from the object. 
936          */
937         vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
938                                 vm_object_terminate_callback, NULL);
939
940         /*
941          * Let the pager know object is dead.
942          */
943         vm_pager_deallocate(object);
944
945         /*
946          * Wait for the object hold count to hit 1, clean out pages as
947          * we go.  vmobj_token interlocks any race conditions that might
948          * pick the object up from the vm_object_list after we have cleared
949          * rb_memq.
950          */
951         for (;;) {
952                 if (RB_ROOT(&object->rb_memq) == NULL)
953                         break;
954                 kprintf("vm_object_terminate: Warning, object %p "
955                         "still has %d pages\n",
956                         object, object->resident_page_count);
957                 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
958                                         vm_object_terminate_callback, NULL);
959         }
960
961         /*
962          * There had better not be any pages left
963          */
964         KKASSERT(object->resident_page_count == 0);
965
966         /*
967          * Remove the object from the global object list.
968          */
969         lwkt_gettoken(&vmobj_token);
970         TAILQ_REMOVE(&vm_object_list, object, object_list);
971         vm_object_count--;
972         lwkt_reltoken(&vmobj_token);
973         vm_object_dead_wakeup(object);
974
975         if (object->ref_count != 0) {
976                 panic("vm_object_terminate2: object with references, "
977                       "ref_count=%d", object->ref_count);
978         }
979
980         /*
981          * NOTE: The object hold_count is at least 1, so we cannot zfree()
982          *       the object here.  See vm_object_drop().
983          */
984 }
985
986 /*
987  * The caller must hold the object.
988  */
989 static int
990 vm_object_terminate_callback(vm_page_t p, void *data __unused)
991 {
992         vm_object_t object;
993
994         object = p->object;
995         vm_page_busy_wait(p, TRUE, "vmpgtrm");
996         if (object != p->object) {
997                 kprintf("vm_object_terminate: Warning: Encountered "
998                         "busied page %p on queue %d\n", p, p->queue);
999                 vm_page_wakeup(p);
1000         } else if (p->wire_count == 0) {
1001                 /*
1002                  * NOTE: PG_NEED_COMMIT is ignored.
1003                  */
1004                 vm_page_free(p);
1005                 mycpu->gd_cnt.v_pfree++;
1006         } else {
1007                 if (p->queue != PQ_NONE)
1008                         kprintf("vm_object_terminate: Warning: Encountered "
1009                                 "wired page %p on queue %d\n", p, p->queue);
1010                 vm_page_remove(p);
1011                 vm_page_wakeup(p);
1012         }
1013         lwkt_yield();
1014         return(0);
1015 }
1016
1017 /*
1018  * The object is dead but still has an object<->pager association.  Sleep
1019  * and return.  The caller typically retests the association in a loop.
1020  *
1021  * The caller must hold the object.
1022  */
1023 void
1024 vm_object_dead_sleep(vm_object_t object, const char *wmesg)
1025 {
1026         ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1027         if (object->handle) {
1028                 vm_object_set_flag(object, OBJ_DEADWNT);
1029                 tsleep(object, 0, wmesg, 0);
1030                 /* object may be invalid after this point */
1031         }
1032 }
1033
1034 /*
1035  * Wakeup anyone waiting for the object<->pager disassociation on
1036  * a dead object.
1037  *
1038  * The caller must hold the object.
1039  */
1040 void
1041 vm_object_dead_wakeup(vm_object_t object)
1042 {
1043         ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1044         if (object->flags & OBJ_DEADWNT) {
1045                 vm_object_clear_flag(object, OBJ_DEADWNT);
1046                 wakeup(object);
1047         }
1048 }
1049
1050 /*
1051  * Clean all dirty pages in the specified range of object.  Leaves page
1052  * on whatever queue it is currently on.   If NOSYNC is set then do not
1053  * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
1054  * leaving the object dirty.
1055  *
1056  * When stuffing pages asynchronously, allow clustering.  XXX we need a
1057  * synchronous clustering mode implementation.
1058  *
1059  * Odd semantics: if start == end, we clean everything.
1060  *
1061  * The object must be locked? XXX
1062  */
1063 static int vm_object_page_clean_pass1(struct vm_page *p, void *data);
1064 static int vm_object_page_clean_pass2(struct vm_page *p, void *data);
1065
1066 void
1067 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1068                      int flags)
1069 {
1070         struct rb_vm_page_scan_info info;
1071         struct vnode *vp;
1072         int wholescan;
1073         int pagerflags;
1074         int generation;
1075
1076         vm_object_hold(object);
1077         if (object->type != OBJT_VNODE ||
1078             (object->flags & OBJ_MIGHTBEDIRTY) == 0) {
1079                 vm_object_drop(object);
1080                 return;
1081         }
1082
1083         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? 
1084                         VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1085         pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
1086
1087         vp = object->handle;
1088
1089         /*
1090          * Interlock other major object operations.  This allows us to 
1091          * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY.
1092          */
1093         vm_object_set_flag(object, OBJ_CLEANING);
1094
1095         /*
1096          * Handle 'entire object' case
1097          */
1098         info.start_pindex = start;
1099         if (end == 0) {
1100                 info.end_pindex = object->size - 1;
1101         } else {
1102                 info.end_pindex = end - 1;
1103         }
1104         wholescan = (start == 0 && info.end_pindex == object->size - 1);
1105         info.limit = flags;
1106         info.pagerflags = pagerflags;
1107         info.object = object;
1108
1109         /*
1110          * If cleaning the entire object do a pass to mark the pages read-only.
1111          * If everything worked out ok, clear OBJ_WRITEABLE and
1112          * OBJ_MIGHTBEDIRTY.
1113          */
1114         if (wholescan) {
1115                 info.error = 0;
1116                 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1117                                         vm_object_page_clean_pass1, &info);
1118                 if (info.error == 0) {
1119                         vm_object_clear_flag(object,
1120                                              OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1121                         if (object->type == OBJT_VNODE &&
1122                             (vp = (struct vnode *)object->handle) != NULL) {
1123                                 if (vp->v_flag & VOBJDIRTY) 
1124                                         vclrflags(vp, VOBJDIRTY);
1125                         }
1126                 }
1127         }
1128
1129         /*
1130          * Do a pass to clean all the dirty pages we find.
1131          */
1132         do {
1133                 info.error = 0;
1134                 generation = object->generation;
1135                 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1136                                         vm_object_page_clean_pass2, &info);
1137         } while (info.error || generation != object->generation);
1138
1139         vm_object_clear_flag(object, OBJ_CLEANING);
1140         vm_object_drop(object);
1141 }
1142
1143 /*
1144  * The caller must hold the object.
1145  */
1146 static 
1147 int
1148 vm_object_page_clean_pass1(struct vm_page *p, void *data)
1149 {
1150         struct rb_vm_page_scan_info *info = data;
1151
1152         vm_page_flag_set(p, PG_CLEANCHK);
1153         if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1154                 info->error = 1;
1155         } else if (vm_page_busy_try(p, FALSE) == 0) {
1156                 vm_page_protect(p, VM_PROT_READ);       /* must not block */
1157                 vm_page_wakeup(p);
1158         } else {
1159                 info->error = 1;
1160         }
1161         lwkt_yield();
1162         return(0);
1163 }
1164
1165 /*
1166  * The caller must hold the object
1167  */
1168 static 
1169 int
1170 vm_object_page_clean_pass2(struct vm_page *p, void *data)
1171 {
1172         struct rb_vm_page_scan_info *info = data;
1173         int generation;
1174
1175         /*
1176          * Do not mess with pages that were inserted after we started
1177          * the cleaning pass.
1178          */
1179         if ((p->flags & PG_CLEANCHK) == 0)
1180                 goto done;
1181
1182         generation = info->object->generation;
1183         vm_page_busy_wait(p, TRUE, "vpcwai");
1184         if (p->object != info->object ||
1185             info->object->generation != generation) {
1186                 info->error = 1;
1187                 vm_page_wakeup(p);
1188                 goto done;
1189         }
1190
1191         /*
1192          * Before wasting time traversing the pmaps, check for trivial
1193          * cases where the page cannot be dirty.
1194          */
1195         if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) {
1196                 KKASSERT((p->dirty & p->valid) == 0);
1197                 vm_page_wakeup(p);
1198                 goto done;
1199         }
1200
1201         /*
1202          * Check whether the page is dirty or not.  The page has been set
1203          * to be read-only so the check will not race a user dirtying the
1204          * page.
1205          */
1206         vm_page_test_dirty(p);
1207         if ((p->dirty & p->valid) == 0) {
1208                 vm_page_flag_clear(p, PG_CLEANCHK);
1209                 vm_page_wakeup(p);
1210                 goto done;
1211         }
1212
1213         /*
1214          * If we have been asked to skip nosync pages and this is a
1215          * nosync page, skip it.  Note that the object flags were
1216          * not cleared in this case (because pass1 will have returned an
1217          * error), so we do not have to set them.
1218          */
1219         if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1220                 vm_page_flag_clear(p, PG_CLEANCHK);
1221                 vm_page_wakeup(p);
1222                 goto done;
1223         }
1224
1225         /*
1226          * Flush as many pages as we can.  PG_CLEANCHK will be cleared on
1227          * the pages that get successfully flushed.  Set info->error if
1228          * we raced an object modification.
1229          */
1230         vm_object_page_collect_flush(info->object, p, info->pagerflags);
1231         vm_wait_nominal();
1232 done:
1233         lwkt_yield();
1234         return(0);
1235 }
1236
1237 /*
1238  * Collect the specified page and nearby pages and flush them out.
1239  * The number of pages flushed is returned.  The passed page is busied
1240  * by the caller and we are responsible for its disposition.
1241  *
1242  * The caller must hold the object.
1243  */
1244 static int
1245 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags)
1246 {
1247         int runlen;
1248         int error;
1249         int maxf;
1250         int chkb;
1251         int maxb;
1252         int i;
1253         vm_pindex_t pi;
1254         vm_page_t maf[vm_pageout_page_count];
1255         vm_page_t mab[vm_pageout_page_count];
1256         vm_page_t ma[vm_pageout_page_count];
1257
1258         ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1259
1260         pi = p->pindex;
1261
1262         maxf = 0;
1263         for(i = 1; i < vm_pageout_page_count; i++) {
1264                 vm_page_t tp;
1265
1266                 tp = vm_page_lookup_busy_try(object, pi + i, TRUE, &error);
1267                 if (error)
1268                         break;
1269                 if (tp == NULL)
1270                         break;
1271                 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1272                     (tp->flags & PG_CLEANCHK) == 0) {
1273                         vm_page_wakeup(tp);
1274                         break;
1275                 }
1276                 if ((tp->queue - tp->pc) == PQ_CACHE) {
1277                         vm_page_flag_clear(tp, PG_CLEANCHK);
1278                         vm_page_wakeup(tp);
1279                         break;
1280                 }
1281                 vm_page_test_dirty(tp);
1282                 if ((tp->dirty & tp->valid) == 0) {
1283                         vm_page_flag_clear(tp, PG_CLEANCHK);
1284                         vm_page_wakeup(tp);
1285                         break;
1286                 }
1287                 maf[i - 1] = tp;
1288                 maxf++;
1289         }
1290
1291         maxb = 0;
1292         chkb = vm_pageout_page_count -  maxf;
1293         /*
1294          * NOTE: chkb can be 0
1295          */
1296         for(i = 1; chkb && i < chkb; i++) {
1297                 vm_page_t tp;
1298
1299                 tp = vm_page_lookup_busy_try(object, pi - i, TRUE, &error);
1300                 if (error)
1301                         break;
1302                 if (tp == NULL)
1303                         break;
1304                 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1305                     (tp->flags & PG_CLEANCHK) == 0) {
1306                         vm_page_wakeup(tp);
1307                         break;
1308                 }
1309                 if ((tp->queue - tp->pc) == PQ_CACHE) {
1310                         vm_page_flag_clear(tp, PG_CLEANCHK);
1311                         vm_page_wakeup(tp);
1312                         break;
1313                 }
1314                 vm_page_test_dirty(tp);
1315                 if ((tp->dirty & tp->valid) == 0) {
1316                         vm_page_flag_clear(tp, PG_CLEANCHK);
1317                         vm_page_wakeup(tp);
1318                         break;
1319                 }
1320                 mab[i - 1] = tp;
1321                 maxb++;
1322         }
1323
1324         /*
1325          * All pages in the maf[] and mab[] array are busied.
1326          */
1327         for (i = 0; i < maxb; i++) {
1328                 int index = (maxb - i) - 1;
1329                 ma[index] = mab[i];
1330                 vm_page_flag_clear(ma[index], PG_CLEANCHK);
1331         }
1332         vm_page_flag_clear(p, PG_CLEANCHK);
1333         ma[maxb] = p;
1334         for(i = 0; i < maxf; i++) {
1335                 int index = (maxb + i) + 1;
1336                 ma[index] = maf[i];
1337                 vm_page_flag_clear(ma[index], PG_CLEANCHK);
1338         }
1339         runlen = maxb + maxf + 1;
1340
1341         for (i = 0; i < runlen; i++)
1342                 vm_page_hold(ma[i]);
1343
1344         vm_pageout_flush(ma, runlen, pagerflags);
1345
1346         for (i = 0; i < runlen; i++) {
1347                 if (ma[i]->valid & ma[i]->dirty) {
1348                         vm_page_protect(ma[i], VM_PROT_READ);
1349                         vm_page_flag_set(ma[i], PG_CLEANCHK);
1350
1351                         /*
1352                          * maxf will end up being the actual number of pages
1353                          * we wrote out contiguously, non-inclusive of the
1354                          * first page.  We do not count look-behind pages.
1355                          */
1356                         if (i >= maxb + 1 && (maxf > i - maxb - 1))
1357                                 maxf = i - maxb - 1;
1358                 }
1359                 vm_page_unhold(ma[i]);
1360         }
1361         return(maxf + 1);
1362 }
1363
1364 /*
1365  * Same as vm_object_pmap_copy, except range checking really
1366  * works, and is meant for small sections of an object.
1367  *
1368  * This code protects resident pages by making them read-only
1369  * and is typically called on a fork or split when a page
1370  * is converted to copy-on-write.  
1371  *
1372  * NOTE: If the page is already at VM_PROT_NONE, calling
1373  * vm_page_protect will have no effect.
1374  */
1375 void
1376 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1377 {
1378         vm_pindex_t idx;
1379         vm_page_t p;
1380
1381         if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
1382                 return;
1383
1384         vm_object_hold(object);
1385         for (idx = start; idx < end; idx++) {
1386                 p = vm_page_lookup(object, idx);
1387                 if (p == NULL)
1388                         continue;
1389                 vm_page_protect(p, VM_PROT_READ);
1390         }
1391         vm_object_drop(object);
1392 }
1393
1394 /*
1395  * Removes all physical pages in the specified object range from all
1396  * physical maps.
1397  *
1398  * The object must *not* be locked.
1399  */
1400
1401 static int vm_object_pmap_remove_callback(vm_page_t p, void *data);
1402
1403 void
1404 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1405 {
1406         struct rb_vm_page_scan_info info;
1407
1408         if (object == NULL)
1409                 return;
1410         info.start_pindex = start;
1411         info.end_pindex = end - 1;
1412
1413         vm_object_hold(object);
1414         vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1415                                 vm_object_pmap_remove_callback, &info);
1416         if (start == 0 && end == object->size)
1417                 vm_object_clear_flag(object, OBJ_WRITEABLE);
1418         vm_object_drop(object);
1419 }
1420
1421 /*
1422  * The caller must hold the object
1423  */
1424 static int
1425 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused)
1426 {
1427         vm_page_protect(p, VM_PROT_NONE);
1428         return(0);
1429 }
1430
1431 /*
1432  * Implements the madvise function at the object/page level.
1433  *
1434  * MADV_WILLNEED        (any object)
1435  *
1436  *      Activate the specified pages if they are resident.
1437  *
1438  * MADV_DONTNEED        (any object)
1439  *
1440  *      Deactivate the specified pages if they are resident.
1441  *
1442  * MADV_FREE    (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only)
1443  *
1444  *      Deactivate and clean the specified pages if they are
1445  *      resident.  This permits the process to reuse the pages
1446  *      without faulting or the kernel to reclaim the pages
1447  *      without I/O.
1448  *
1449  * No requirements.
1450  */
1451 void
1452 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1453 {
1454         vm_pindex_t end, tpindex;
1455         vm_object_t tobject;
1456         vm_object_t xobj;
1457         vm_page_t m;
1458         int error;
1459
1460         if (object == NULL)
1461                 return;
1462
1463         end = pindex + count;
1464
1465         vm_object_hold(object);
1466         tobject = object;
1467
1468         /*
1469          * Locate and adjust resident pages
1470          */
1471         for (; pindex < end; pindex += 1) {
1472 relookup:
1473                 if (tobject != object)
1474                         vm_object_drop(tobject);
1475                 tobject = object;
1476                 tpindex = pindex;
1477 shadowlookup:
1478                 /*
1479                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1480                  * and those pages must be OBJ_ONEMAPPING.
1481                  */
1482                 if (advise == MADV_FREE) {
1483                         if ((tobject->type != OBJT_DEFAULT &&
1484                              tobject->type != OBJT_SWAP) ||
1485                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
1486                                 continue;
1487                         }
1488                 }
1489
1490                 m = vm_page_lookup_busy_try(tobject, tpindex, TRUE, &error);
1491
1492                 if (error) {
1493                         vm_page_sleep_busy(m, TRUE, "madvpo");
1494                         goto relookup;
1495                 }
1496                 if (m == NULL) {
1497                         /*
1498                          * There may be swap even if there is no backing page
1499                          */
1500                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1501                                 swap_pager_freespace(tobject, tpindex, 1);
1502
1503                         /*
1504                          * next object
1505                          */
1506                         while ((xobj = tobject->backing_object) != NULL) {
1507                                 KKASSERT(xobj != object);
1508                                 vm_object_hold(xobj);
1509                                 if (xobj == tobject->backing_object)
1510                                         break;
1511                                 vm_object_drop(xobj);
1512                         }
1513                         if (xobj == NULL)
1514                                 continue;
1515                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1516                         if (tobject != object) {
1517                                 vm_object_lock_swap();
1518                                 vm_object_drop(tobject);
1519                         }
1520                         tobject = xobj;
1521                         goto shadowlookup;
1522                 }
1523
1524                 /*
1525                  * If the page is not in a normal active state, we skip it.
1526                  * If the page is not managed there are no page queues to
1527                  * mess with.  Things can break if we mess with pages in
1528                  * any of the below states.
1529                  */
1530                 if (m->wire_count ||
1531                     (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
1532                     m->valid != VM_PAGE_BITS_ALL
1533                 ) {
1534                         vm_page_wakeup(m);
1535                         continue;
1536                 }
1537
1538                 /*
1539                  * Theoretically once a page is known not to be busy, an
1540                  * interrupt cannot come along and rip it out from under us.
1541                  */
1542
1543                 if (advise == MADV_WILLNEED) {
1544                         vm_page_activate(m);
1545                 } else if (advise == MADV_DONTNEED) {
1546                         vm_page_dontneed(m);
1547                 } else if (advise == MADV_FREE) {
1548                         /*
1549                          * Mark the page clean.  This will allow the page
1550                          * to be freed up by the system.  However, such pages
1551                          * are often reused quickly by malloc()/free()
1552                          * so we do not do anything that would cause
1553                          * a page fault if we can help it.
1554                          *
1555                          * Specifically, we do not try to actually free
1556                          * the page now nor do we try to put it in the
1557                          * cache (which would cause a page fault on reuse).
1558                          *
1559                          * But we do make the page is freeable as we
1560                          * can without actually taking the step of unmapping
1561                          * it.
1562                          */
1563                         pmap_clear_modify(m);
1564                         m->dirty = 0;
1565                         m->act_count = 0;
1566                         vm_page_dontneed(m);
1567                         if (tobject->type == OBJT_SWAP)
1568                                 swap_pager_freespace(tobject, tpindex, 1);
1569                 }
1570                 vm_page_wakeup(m);
1571         }       
1572         if (tobject != object)
1573                 vm_object_drop(tobject);
1574         vm_object_drop(object);
1575 }
1576
1577 /*
1578  * Create a new object which is backed by the specified existing object
1579  * range.  Replace the pointer and offset that was pointing at the existing
1580  * object with the pointer/offset for the new object.
1581  *
1582  * No other requirements.
1583  */
1584 void
1585 vm_object_shadow(vm_object_t *objectp, vm_ooffset_t *offset, vm_size_t length,
1586                  int addref)
1587 {
1588         vm_object_t source;
1589         vm_object_t result;
1590
1591         source = *objectp;
1592
1593         /*
1594          * Don't create the new object if the old object isn't shared.
1595          * We have to chain wait before adding the reference to avoid
1596          * racing a collapse or deallocation.
1597          *
1598          * Add the additional ref to source here to avoid racing a later
1599          * collapse or deallocation. Clear the ONEMAPPING flag whether
1600          * addref is TRUE or not in this case because the original object
1601          * will be shadowed.
1602          */
1603         if (source) {
1604                 vm_object_hold(source);
1605                 vm_object_chain_wait(source);
1606                 if (source->ref_count == 1 &&
1607                     source->handle == NULL &&
1608                     (source->type == OBJT_DEFAULT ||
1609                      source->type == OBJT_SWAP)) {
1610                         vm_object_drop(source);
1611                         if (addref) {
1612                                 vm_object_reference_locked(source);
1613                                 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1614                         }
1615                         return;
1616                 }
1617                 vm_object_reference_locked(source);
1618                 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1619         }
1620
1621         /*
1622          * Allocate a new object with the given length.  The new object
1623          * is returned referenced but we may have to add another one.
1624          * If we are adding a second reference we must clear OBJ_ONEMAPPING.
1625          * (typically because the caller is about to clone a vm_map_entry).
1626          *
1627          * The source object currently has an extra reference to prevent
1628          * collapses into it while we mess with its shadow list, which
1629          * we will remove later in this routine.
1630          */
1631         if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1632                 panic("vm_object_shadow: no object for shadowing");
1633         vm_object_hold(result);
1634         if (addref) {
1635                 vm_object_reference_locked(result);
1636                 vm_object_clear_flag(result, OBJ_ONEMAPPING);
1637         }
1638
1639         /*
1640          * The new object shadows the source object.  Chain wait before
1641          * adjusting shadow_count or the shadow list to avoid races.
1642          *
1643          * Try to optimize the result object's page color when shadowing
1644          * in order to maintain page coloring consistency in the combined 
1645          * shadowed object.
1646          */
1647         KKASSERT(result->backing_object == NULL);
1648         result->backing_object = source;
1649         if (source) {
1650                 vm_object_chain_wait(source);
1651                 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1652                 source->shadow_count++;
1653                 source->generation++;
1654 #ifdef SMP
1655                 /* cpu localization twist */
1656                 result->pg_color = (int)(intptr_t)curthread;
1657 #else
1658                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1659                                    PQ_L2_MASK;
1660 #endif
1661         }
1662
1663         /*
1664          * Adjust the return storage.  Drop the ref on source before
1665          * returning.
1666          */
1667         result->backing_object_offset = *offset;
1668         vm_object_drop(result);
1669         *offset = 0;
1670         if (source) {
1671                 vm_object_deallocate_locked(source);
1672                 vm_object_drop(source);
1673         }
1674
1675         /*
1676          * Return the new things
1677          */
1678         *objectp = result;
1679 }
1680
1681 #define OBSC_TEST_ALL_SHADOWED  0x0001
1682 #define OBSC_COLLAPSE_NOWAIT    0x0002
1683 #define OBSC_COLLAPSE_WAIT      0x0004
1684
1685 static int vm_object_backing_scan_callback(vm_page_t p, void *data);
1686
1687 /*
1688  * The caller must hold the object.
1689  */
1690 static __inline int
1691 vm_object_backing_scan(vm_object_t object, vm_object_t backing_object, int op)
1692 {
1693         struct rb_vm_page_scan_info info;
1694
1695         vm_object_assert_held(object);
1696         vm_object_assert_held(backing_object);
1697
1698         KKASSERT(backing_object == object->backing_object);
1699         info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1700
1701         /*
1702          * Initial conditions
1703          */
1704         if (op & OBSC_TEST_ALL_SHADOWED) {
1705                 /*
1706                  * We do not want to have to test for the existence of
1707                  * swap pages in the backing object.  XXX but with the
1708                  * new swapper this would be pretty easy to do.
1709                  *
1710                  * XXX what about anonymous MAP_SHARED memory that hasn't
1711                  * been ZFOD faulted yet?  If we do not test for this, the
1712                  * shadow test may succeed! XXX
1713                  */
1714                 if (backing_object->type != OBJT_DEFAULT)
1715                         return(0);
1716         }
1717         if (op & OBSC_COLLAPSE_WAIT) {
1718                 KKASSERT((backing_object->flags & OBJ_DEAD) == 0);
1719                 vm_object_set_flag(backing_object, OBJ_DEAD);
1720                 lwkt_gettoken(&vmobj_token);
1721                 TAILQ_REMOVE(&vm_object_list, backing_object, object_list);
1722                 vm_object_count--;
1723                 lwkt_reltoken(&vmobj_token);
1724                 vm_object_dead_wakeup(backing_object);
1725         }
1726
1727         /*
1728          * Our scan.   We have to retry if a negative error code is returned,
1729          * otherwise 0 or 1 will be returned in info.error.  0 Indicates that
1730          * the scan had to be stopped because the parent does not completely
1731          * shadow the child.
1732          */
1733         info.object = object;
1734         info.backing_object = backing_object;
1735         info.limit = op;
1736         do {
1737                 info.error = 1;
1738                 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL,
1739                                         vm_object_backing_scan_callback,
1740                                         &info);
1741         } while (info.error < 0);
1742
1743         return(info.error);
1744 }
1745
1746 /*
1747  * The caller must hold the object.
1748  */
1749 static int
1750 vm_object_backing_scan_callback(vm_page_t p, void *data)
1751 {
1752         struct rb_vm_page_scan_info *info = data;
1753         vm_object_t backing_object;
1754         vm_object_t object;
1755         vm_pindex_t pindex;
1756         vm_pindex_t new_pindex;
1757         vm_pindex_t backing_offset_index;
1758         int op;
1759
1760         pindex = p->pindex;
1761         new_pindex = pindex - info->backing_offset_index;
1762         op = info->limit;
1763         object = info->object;
1764         backing_object = info->backing_object;
1765         backing_offset_index = info->backing_offset_index;
1766
1767         if (op & OBSC_TEST_ALL_SHADOWED) {
1768                 vm_page_t pp;
1769
1770                 /*
1771                  * Ignore pages outside the parent object's range
1772                  * and outside the parent object's mapping of the 
1773                  * backing object.
1774                  *
1775                  * note that we do not busy the backing object's
1776                  * page.
1777                  */
1778                 if (pindex < backing_offset_index ||
1779                     new_pindex >= object->size
1780                 ) {
1781                         return(0);
1782                 }
1783
1784                 /*
1785                  * See if the parent has the page or if the parent's
1786                  * object pager has the page.  If the parent has the
1787                  * page but the page is not valid, the parent's
1788                  * object pager must have the page.
1789                  *
1790                  * If this fails, the parent does not completely shadow
1791                  * the object and we might as well give up now.
1792                  */
1793                 pp = vm_page_lookup(object, new_pindex);
1794                 if ((pp == NULL || pp->valid == 0) &&
1795                     !vm_pager_has_page(object, new_pindex)
1796                 ) {
1797                         info->error = 0;        /* problemo */
1798                         return(-1);             /* stop the scan */
1799                 }
1800         }
1801
1802         /*
1803          * Check for busy page.  Note that we may have lost (p) when we
1804          * possibly blocked above.
1805          */
1806         if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1807                 vm_page_t pp;
1808
1809                 if (vm_page_busy_try(p, TRUE)) {
1810                         if (op & OBSC_COLLAPSE_NOWAIT) {
1811                                 return(0);
1812                         } else {
1813                                 /*
1814                                  * If we slept, anything could have
1815                                  * happened.   Ask that the scan be restarted.
1816                                  *
1817                                  * Since the object is marked dead, the
1818                                  * backing offset should not have changed.  
1819                                  */
1820                                 vm_page_sleep_busy(p, TRUE, "vmocol");
1821                                 info->error = -1;
1822                                 return(-1);
1823                         }
1824                 }
1825
1826                 /*
1827                  * If (p) is no longer valid restart the scan.
1828                  */
1829                 if (p->object != backing_object || p->pindex != pindex) {
1830                         kprintf("vm_object_backing_scan: Warning: page "
1831                                 "%p ripped out from under us\n", p);
1832                         vm_page_wakeup(p);
1833                         info->error = -1;
1834                         return(-1);
1835                 }
1836
1837                 if (op & OBSC_COLLAPSE_NOWAIT) {
1838                         if (p->valid == 0 ||
1839                             p->wire_count ||
1840                             (p->flags & PG_NEED_COMMIT)) {
1841                                 vm_page_wakeup(p);
1842                                 return(0);
1843                         }
1844                 } else {
1845                         /* XXX what if p->valid == 0 , hold_count, etc? */
1846                 }
1847
1848                 KASSERT(
1849                     p->object == backing_object,
1850                     ("vm_object_qcollapse(): object mismatch")
1851                 );
1852
1853                 /*
1854                  * Destroy any associated swap
1855                  */
1856                 if (backing_object->type == OBJT_SWAP)
1857                         swap_pager_freespace(backing_object, p->pindex, 1);
1858
1859                 if (
1860                     p->pindex < backing_offset_index ||
1861                     new_pindex >= object->size
1862                 ) {
1863                         /*
1864                          * Page is out of the parent object's range, we 
1865                          * can simply destroy it. 
1866                          */
1867                         vm_page_protect(p, VM_PROT_NONE);
1868                         vm_page_free(p);
1869                         return(0);
1870                 }
1871
1872                 pp = vm_page_lookup(object, new_pindex);
1873                 if (pp != NULL || vm_pager_has_page(object, new_pindex)) {
1874                         /*
1875                          * page already exists in parent OR swap exists
1876                          * for this location in the parent.  Destroy 
1877                          * the original page from the backing object.
1878                          *
1879                          * Leave the parent's page alone
1880                          */
1881                         vm_page_protect(p, VM_PROT_NONE);
1882                         vm_page_free(p);
1883                         return(0);
1884                 }
1885
1886                 /*
1887                  * Page does not exist in parent, rename the
1888                  * page from the backing object to the main object. 
1889                  *
1890                  * If the page was mapped to a process, it can remain 
1891                  * mapped through the rename.
1892                  */
1893                 if ((p->queue - p->pc) == PQ_CACHE)
1894                         vm_page_deactivate(p);
1895
1896                 vm_page_rename(p, object, new_pindex);
1897                 vm_page_wakeup(p);
1898                 /* page automatically made dirty by rename */
1899         }
1900         return(0);
1901 }
1902
1903 /*
1904  * This version of collapse allows the operation to occur earlier and
1905  * when paging_in_progress is true for an object...  This is not a complete
1906  * operation, but should plug 99.9% of the rest of the leaks.
1907  *
1908  * The caller must hold the object and backing_object and both must be
1909  * chainlocked.
1910  *
1911  * (only called from vm_object_collapse)
1912  */
1913 static void
1914 vm_object_qcollapse(vm_object_t object, vm_object_t backing_object)
1915 {
1916         if (backing_object->ref_count == 1) {
1917                 backing_object->ref_count += 2;
1918                 vm_object_backing_scan(object, backing_object,
1919                                        OBSC_COLLAPSE_NOWAIT);
1920                 backing_object->ref_count -= 2;
1921         }
1922 }
1923
1924 /*
1925  * Collapse an object with the object backing it.  Pages in the backing
1926  * object are moved into the parent, and the backing object is deallocated.
1927  * Any conflict is resolved in favor of the parent's existing pages.
1928  *
1929  * object must be held and chain-locked on call.
1930  *
1931  * The caller must have an extra ref on object to prevent a race from
1932  * destroying it during the collapse.
1933  */
1934 void
1935 vm_object_collapse(vm_object_t object, struct vm_object_dealloc_list **dlistp)
1936 {
1937         struct vm_object_dealloc_list *dlist = NULL;
1938         vm_object_t backing_object;
1939
1940         /*
1941          * Only one thread is attempting a collapse at any given moment.
1942          * There are few restrictions for (object) that callers of this
1943          * function check so reentrancy is likely.
1944          */
1945         KKASSERT(object != NULL);
1946         vm_object_assert_held(object);
1947         KKASSERT(object->flags & OBJ_CHAINLOCK);
1948
1949         for (;;) {
1950                 vm_object_t bbobj;
1951                 int dodealloc;
1952
1953                 /*
1954                  * We have to hold the backing object, check races.
1955                  */
1956                 while ((backing_object = object->backing_object) != NULL) {
1957                         vm_object_hold(backing_object);
1958                         if (backing_object == object->backing_object)
1959                                 break;
1960                         vm_object_drop(backing_object);
1961                 }
1962
1963                 /*
1964                  * No backing object?  Nothing to collapse then.
1965                  */
1966                 if (backing_object == NULL)
1967                         break;
1968
1969                 /*
1970                  * You can't collapse with a non-default/non-swap object.
1971                  */
1972                 if (backing_object->type != OBJT_DEFAULT &&
1973                     backing_object->type != OBJT_SWAP) {
1974                         vm_object_drop(backing_object);
1975                         backing_object = NULL;
1976                         break;
1977                 }
1978
1979                 /*
1980                  * Chain-lock the backing object too because if we
1981                  * successfully merge its pages into the top object we
1982                  * will collapse backing_object->backing_object as the
1983                  * new backing_object.  Re-check that it is still our
1984                  * backing object.
1985                  */
1986                 vm_object_chain_acquire(backing_object);
1987                 if (backing_object != object->backing_object) {
1988                         vm_object_chain_release(backing_object);
1989                         vm_object_drop(backing_object);
1990                         continue;
1991                 }
1992
1993                 /*
1994                  * we check the backing object first, because it is most likely
1995                  * not collapsable.
1996                  */
1997                 if (backing_object->handle != NULL ||
1998                     (backing_object->type != OBJT_DEFAULT &&
1999                      backing_object->type != OBJT_SWAP) ||
2000                     (backing_object->flags & OBJ_DEAD) ||
2001                     object->handle != NULL ||
2002                     (object->type != OBJT_DEFAULT &&
2003                      object->type != OBJT_SWAP) ||
2004                     (object->flags & OBJ_DEAD)) {
2005                         break;
2006                 }
2007
2008                 /*
2009                  * If paging is in progress we can't do a normal collapse.
2010                  */
2011                 if (
2012                     object->paging_in_progress != 0 ||
2013                     backing_object->paging_in_progress != 0
2014                 ) {
2015                         vm_object_qcollapse(object, backing_object);
2016                         break;
2017                 }
2018
2019                 /*
2020                  * We know that we can either collapse the backing object (if
2021                  * the parent is the only reference to it) or (perhaps) have
2022                  * the parent bypass the object if the parent happens to shadow
2023                  * all the resident pages in the entire backing object.
2024                  *
2025                  * This is ignoring pager-backed pages such as swap pages.
2026                  * vm_object_backing_scan fails the shadowing test in this
2027                  * case.
2028                  */
2029                 if (backing_object->ref_count == 1) {
2030                         /*
2031                          * If there is exactly one reference to the backing
2032                          * object, we can collapse it into the parent.  
2033                          */
2034                         KKASSERT(object->backing_object == backing_object);
2035                         vm_object_backing_scan(object, backing_object,
2036                                                OBSC_COLLAPSE_WAIT);
2037
2038                         /*
2039                          * Move the pager from backing_object to object.
2040                          */
2041                         if (backing_object->type == OBJT_SWAP) {
2042                                 vm_object_pip_add(backing_object, 1);
2043
2044                                 /*
2045                                  * scrap the paging_offset junk and do a 
2046                                  * discrete copy.  This also removes major 
2047                                  * assumptions about how the swap-pager 
2048                                  * works from where it doesn't belong.  The
2049                                  * new swapper is able to optimize the
2050                                  * destroy-source case.
2051                                  */
2052                                 vm_object_pip_add(object, 1);
2053                                 swap_pager_copy(backing_object, object,
2054                                     OFF_TO_IDX(object->backing_object_offset),
2055                                     TRUE);
2056                                 vm_object_pip_wakeup(object);
2057                                 vm_object_pip_wakeup(backing_object);
2058                         }
2059
2060                         /*
2061                          * Object now shadows whatever backing_object did.
2062                          * Remove object from backing_object's shadow_list.
2063                          */
2064                         LIST_REMOVE(object, shadow_list);
2065                         KKASSERT(object->backing_object == backing_object);
2066                         backing_object->shadow_count--;
2067                         backing_object->generation++;
2068
2069                         /*
2070                          * backing_object->backing_object moves from within
2071                          * backing_object to within object.
2072                          */
2073                         while ((bbobj = backing_object->backing_object) != NULL) {
2074                                 vm_object_hold(bbobj);
2075                                 if (bbobj == backing_object->backing_object)
2076                                         break;
2077                                 vm_object_drop(bbobj);
2078                         }
2079                         if (bbobj) {
2080                                 LIST_REMOVE(backing_object, shadow_list);
2081                                 bbobj->shadow_count--;
2082                                 bbobj->generation++;
2083                                 backing_object->backing_object = NULL;
2084                         }
2085                         object->backing_object = bbobj;
2086                         if (bbobj) {
2087                                 LIST_INSERT_HEAD(&bbobj->shadow_head,
2088                                                  object, shadow_list);
2089                                 bbobj->shadow_count++;
2090                                 bbobj->generation++;
2091                         }
2092
2093                         object->backing_object_offset +=
2094                                 backing_object->backing_object_offset;
2095
2096                         vm_object_drop(bbobj);
2097
2098                         /*
2099                          * Discard the old backing_object.  Nothing should be
2100                          * able to ref it, other than a vm_map_split(),
2101                          * and vm_map_split() will stall on our chain lock.
2102                          * And we control the parent so it shouldn't be
2103                          * possible for it to go away either.
2104                          *
2105                          * Since the backing object has no pages, no pager
2106                          * left, and no object references within it, all
2107                          * that is necessary is to dispose of it.
2108                          */
2109                         KASSERT(backing_object->ref_count == 1,
2110                                 ("backing_object %p was somehow "
2111                                  "re-referenced during collapse!",
2112                                  backing_object));
2113                         KASSERT(RB_EMPTY(&backing_object->rb_memq),
2114                                 ("backing_object %p somehow has left "
2115                                  "over pages during collapse!",
2116                                  backing_object));
2117
2118                         /*
2119                          * The object can be destroyed.
2120                          *
2121                          * XXX just fall through and dodealloc instead
2122                          *     of forcing destruction?
2123                          */
2124                         --backing_object->ref_count;
2125                         if ((backing_object->flags & OBJ_DEAD) == 0)
2126                                 vm_object_terminate(backing_object);
2127                         object_collapses++;
2128                         dodealloc = 0;
2129                 } else {
2130                         /*
2131                          * If we do not entirely shadow the backing object,
2132                          * there is nothing we can do so we give up.
2133                          */
2134                         if (vm_object_backing_scan(object, backing_object,
2135                                                 OBSC_TEST_ALL_SHADOWED) == 0) {
2136                                 break;
2137                         }
2138
2139                         /*
2140                          * bbobj is backing_object->backing_object.  Since
2141                          * object completely shadows backing_object we can
2142                          * bypass it and become backed by bbobj instead.
2143                          */
2144                         while ((bbobj = backing_object->backing_object) != NULL) {
2145                                 vm_object_hold(bbobj);
2146                                 if (bbobj == backing_object->backing_object)
2147                                         break;
2148                                 vm_object_drop(bbobj);
2149                         }
2150
2151                         /*
2152                          * Make object shadow bbobj instead of backing_object.
2153                          * Remove object from backing_object's shadow list.
2154                          *
2155                          * Deallocating backing_object will not remove
2156                          * it, since its reference count is at least 2.
2157                          */
2158                         KKASSERT(object->backing_object == backing_object);
2159                         LIST_REMOVE(object, shadow_list);
2160                         backing_object->shadow_count--;
2161                         backing_object->generation++;
2162
2163                         /*
2164                          * Add a ref to bbobj, bbobj now shadows object.
2165                          *
2166                          * NOTE: backing_object->backing_object still points
2167                          *       to bbobj.  That relationship remains intact
2168                          *       because backing_object has > 1 ref, so
2169                          *       someone else is pointing to it (hence why
2170                          *       we can't collapse it into object and can
2171                          *       only handle the all-shadowed bypass case).
2172                          */
2173                         if (bbobj) {
2174                                 vm_object_chain_wait(bbobj);
2175                                 vm_object_reference_locked(bbobj);
2176                                 LIST_INSERT_HEAD(&bbobj->shadow_head,
2177                                                  object, shadow_list);
2178                                 bbobj->shadow_count++;
2179                                 bbobj->generation++;
2180                                 object->backing_object_offset +=
2181                                         backing_object->backing_object_offset;
2182                                 object->backing_object = bbobj;
2183                                 vm_object_drop(bbobj);
2184                         } else {
2185                                 object->backing_object = NULL;
2186                         }
2187
2188                         /*
2189                          * Drop the reference count on backing_object.  To
2190                          * handle ref_count races properly we can't assume
2191                          * that the ref_count is still at least 2 so we
2192                          * have to actually call vm_object_deallocate()
2193                          * (after clearing the chainlock).
2194                          */
2195                         object_bypasses++;
2196                         dodealloc = 1;
2197                 }
2198
2199                 /*
2200                  * Ok, we want to loop on the new object->bbobj association,
2201                  * possibly collapsing it further.  However if dodealloc is
2202                  * non-zero we have to deallocate the backing_object which
2203                  * itself can potentially undergo a collapse, creating a
2204                  * recursion depth issue with the LWKT token subsystem.
2205                  *
2206                  * In the case where we must deallocate the backing_object
2207                  * it is possible now that the backing_object has a single
2208                  * shadow count on some other object (not represented here
2209                  * as yet), since it no longer shadows us.  Thus when we
2210                  * call vm_object_deallocate() it may attempt to collapse
2211                  * itself into its remaining parent.
2212                  */
2213                 if (dodealloc) {
2214                         struct vm_object_dealloc_list *dtmp;
2215
2216                         vm_object_chain_release(backing_object);
2217                         vm_object_unlock(backing_object);
2218                         /* backing_object remains held */
2219
2220                         /*
2221                          * Auto-deallocation list for caller convenience.
2222                          */
2223                         if (dlistp == NULL)
2224                                 dlistp = &dlist;
2225
2226                         dtmp = kmalloc(sizeof(*dtmp), M_TEMP, M_WAITOK);
2227                         dtmp->object = backing_object;
2228                         dtmp->next = *dlistp;
2229                         *dlistp = dtmp;
2230                 } else {
2231                         vm_object_chain_release(backing_object);
2232                         vm_object_drop(backing_object);
2233                 }
2234                 /* backing_object = NULL; not needed */
2235                 /* loop */
2236         }
2237
2238         /*
2239          * Clean up any left over backing_object
2240          */
2241         if (backing_object) {
2242                 vm_object_chain_release(backing_object);
2243                 vm_object_drop(backing_object);
2244         }
2245
2246         /*
2247          * Clean up any auto-deallocation list.  This is a convenience
2248          * for top-level callers so they don't have to pass &dlist.
2249          * Do not clean up any caller-passed dlistp, the caller will
2250          * do that.
2251          */
2252         if (dlist)
2253                 vm_object_deallocate_list(&dlist);
2254
2255 }
2256
2257 /*
2258  * vm_object_collapse() may collect additional objects in need of
2259  * deallocation.  This routine deallocates these objects.  The
2260  * deallocation itself can trigger additional collapses (which the
2261  * deallocate function takes care of).  This procedure is used to
2262  * reduce procedural recursion since these vm_object shadow chains
2263  * can become quite long.
2264  */
2265 void
2266 vm_object_deallocate_list(struct vm_object_dealloc_list **dlistp)
2267 {
2268         struct vm_object_dealloc_list *dlist;
2269
2270         while ((dlist = *dlistp) != NULL) {
2271                 *dlistp = dlist->next;
2272                 vm_object_lock(dlist->object);
2273                 vm_object_deallocate_locked(dlist->object);
2274                 vm_object_drop(dlist->object);
2275                 kfree(dlist, M_TEMP);
2276         }
2277 }
2278
2279 /*
2280  * Removes all physical pages in the specified object range from the
2281  * object's list of pages.
2282  *
2283  * No requirements.
2284  */
2285 static int vm_object_page_remove_callback(vm_page_t p, void *data);
2286
2287 void
2288 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2289                       boolean_t clean_only)
2290 {
2291         struct rb_vm_page_scan_info info;
2292         int all;
2293
2294         /*
2295          * Degenerate cases and assertions
2296          */
2297         vm_object_hold(object);
2298         if (object == NULL ||
2299             (object->resident_page_count == 0 && object->swblock_count == 0)) {
2300                 vm_object_drop(object);
2301                 return;
2302         }
2303         KASSERT(object->type != OBJT_PHYS, 
2304                 ("attempt to remove pages from a physical object"));
2305
2306         /*
2307          * Indicate that paging is occuring on the object
2308          */
2309         vm_object_pip_add(object, 1);
2310
2311         /*
2312          * Figure out the actual removal range and whether we are removing
2313          * the entire contents of the object or not.  If removing the entire
2314          * contents, be sure to get all pages, even those that might be 
2315          * beyond the end of the object.
2316          */
2317         info.start_pindex = start;
2318         if (end == 0)
2319                 info.end_pindex = (vm_pindex_t)-1;
2320         else
2321                 info.end_pindex = end - 1;
2322         info.limit = clean_only;
2323         all = (start == 0 && info.end_pindex >= object->size - 1);
2324
2325         /*
2326          * Loop until we are sure we have gotten them all.
2327          */
2328         do {
2329                 info.error = 0;
2330                 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2331                                         vm_object_page_remove_callback, &info);
2332         } while (info.error);
2333
2334         /*
2335          * Remove any related swap if throwing away pages, or for
2336          * non-swap objects (the swap is a clean copy in that case).
2337          */
2338         if (object->type != OBJT_SWAP || clean_only == FALSE) {
2339                 if (all)
2340                         swap_pager_freespace_all(object);
2341                 else
2342                         swap_pager_freespace(object, info.start_pindex,
2343                              info.end_pindex - info.start_pindex + 1);
2344         }
2345
2346         /*
2347          * Cleanup
2348          */
2349         vm_object_pip_wakeup(object);
2350         vm_object_drop(object);
2351 }
2352
2353 /*
2354  * The caller must hold the object
2355  */
2356 static int
2357 vm_object_page_remove_callback(vm_page_t p, void *data)
2358 {
2359         struct rb_vm_page_scan_info *info = data;
2360
2361         if (vm_page_busy_try(p, TRUE)) {
2362                 vm_page_sleep_busy(p, TRUE, "vmopar");
2363                 info->error = 1;
2364                 return(0);
2365         }
2366
2367         /*
2368          * Wired pages cannot be destroyed, but they can be invalidated
2369          * and we do so if clean_only (limit) is not set.
2370          *
2371          * WARNING!  The page may be wired due to being part of a buffer
2372          *           cache buffer, and the buffer might be marked B_CACHE.
2373          *           This is fine as part of a truncation but VFSs must be
2374          *           sure to fix the buffer up when re-extending the file.
2375          *
2376          * NOTE!     PG_NEED_COMMIT is ignored.
2377          */
2378         if (p->wire_count != 0) {
2379                 vm_page_protect(p, VM_PROT_NONE);
2380                 if (info->limit == 0)
2381                         p->valid = 0;
2382                 vm_page_wakeup(p);
2383                 return(0);
2384         }
2385
2386         /*
2387          * limit is our clean_only flag.  If set and the page is dirty, do
2388          * not free it.  If set and the page is being held by someone, do
2389          * not free it.
2390          */
2391         if (info->limit && p->valid) {
2392                 vm_page_test_dirty(p);
2393                 if (p->valid & p->dirty) {
2394                         vm_page_wakeup(p);
2395                         return(0);
2396                 }
2397 #if 0
2398                 if (p->hold_count) {
2399                         vm_page_wakeup(p);
2400                         return(0);
2401                 }
2402 #endif
2403         }
2404
2405         /*
2406          * Destroy the page
2407          */
2408         vm_page_protect(p, VM_PROT_NONE);
2409         vm_page_free(p);
2410         return(0);
2411 }
2412
2413 /*
2414  * Coalesces two objects backing up adjoining regions of memory into a
2415  * single object.
2416  *
2417  * returns TRUE if objects were combined.
2418  *
2419  * NOTE: Only works at the moment if the second object is NULL -
2420  *       if it's not, which object do we lock first?
2421  *
2422  * Parameters:
2423  *      prev_object     First object to coalesce
2424  *      prev_offset     Offset into prev_object
2425  *      next_object     Second object into coalesce
2426  *      next_offset     Offset into next_object
2427  *
2428  *      prev_size       Size of reference to prev_object
2429  *      next_size       Size of reference to next_object
2430  *
2431  * The caller does not need to hold (prev_object) but must have a stable
2432  * pointer to it (typically by holding the vm_map locked).
2433  */
2434 boolean_t
2435 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
2436                    vm_size_t prev_size, vm_size_t next_size)
2437 {
2438         vm_pindex_t next_pindex;
2439
2440         if (prev_object == NULL)
2441                 return (TRUE);
2442
2443         vm_object_hold(prev_object);
2444
2445         if (prev_object->type != OBJT_DEFAULT &&
2446             prev_object->type != OBJT_SWAP) {
2447                 vm_object_drop(prev_object);
2448                 return (FALSE);
2449         }
2450
2451         /*
2452          * Try to collapse the object first
2453          */
2454         vm_object_chain_acquire(prev_object);
2455         vm_object_collapse(prev_object, NULL);
2456
2457         /*
2458          * Can't coalesce if: . more than one reference . paged out . shadows
2459          * another object . has a copy elsewhere (any of which mean that the
2460          * pages not mapped to prev_entry may be in use anyway)
2461          */
2462
2463         if (prev_object->backing_object != NULL) {
2464                 vm_object_chain_release(prev_object);
2465                 vm_object_drop(prev_object);
2466                 return (FALSE);
2467         }
2468
2469         prev_size >>= PAGE_SHIFT;
2470         next_size >>= PAGE_SHIFT;
2471         next_pindex = prev_pindex + prev_size;
2472
2473         if ((prev_object->ref_count > 1) &&
2474             (prev_object->size != next_pindex)) {
2475                 vm_object_chain_release(prev_object);
2476                 vm_object_drop(prev_object);
2477                 return (FALSE);
2478         }
2479
2480         /*
2481          * Remove any pages that may still be in the object from a previous
2482          * deallocation.
2483          */
2484         if (next_pindex < prev_object->size) {
2485                 vm_object_page_remove(prev_object,
2486                                       next_pindex,
2487                                       next_pindex + next_size, FALSE);
2488                 if (prev_object->type == OBJT_SWAP)
2489                         swap_pager_freespace(prev_object,
2490                                              next_pindex, next_size);
2491         }
2492
2493         /*
2494          * Extend the object if necessary.
2495          */
2496         if (next_pindex + next_size > prev_object->size)
2497                 prev_object->size = next_pindex + next_size;
2498
2499         vm_object_chain_release(prev_object);
2500         vm_object_drop(prev_object);
2501         return (TRUE);
2502 }
2503
2504 /*
2505  * Make the object writable and flag is being possibly dirty.
2506  *
2507  * The caller must hold the object. XXX called from vm_page_dirty(),
2508  * There is currently no requirement to hold the object.
2509  */
2510 void
2511 vm_object_set_writeable_dirty(vm_object_t object)
2512 {
2513         struct vnode *vp;
2514
2515         /*vm_object_assert_held(object);*/
2516         /*
2517          * Avoid contention in vm fault path by checking the state before
2518          * issuing an atomic op on it.
2519          */
2520         if ((object->flags & (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) !=
2521             (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) {
2522                 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
2523         }
2524         if (object->type == OBJT_VNODE &&
2525             (vp = (struct vnode *)object->handle) != NULL) {
2526                 if ((vp->v_flag & VOBJDIRTY) == 0) {
2527                         vsetflags(vp, VOBJDIRTY);
2528                 }
2529         }
2530 }
2531
2532 #include "opt_ddb.h"
2533 #ifdef DDB
2534 #include <sys/kernel.h>
2535
2536 #include <sys/cons.h>
2537
2538 #include <ddb/ddb.h>
2539
2540 static int      _vm_object_in_map (vm_map_t map, vm_object_t object,
2541                                        vm_map_entry_t entry);
2542 static int      vm_object_in_map (vm_object_t object);
2543
2544 /*
2545  * The caller must hold the object.
2546  */
2547 static int
2548 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2549 {
2550         vm_map_t tmpm;
2551         vm_map_entry_t tmpe;
2552         vm_object_t obj, nobj;
2553         int entcount;
2554
2555         if (map == 0)
2556                 return 0;
2557         if (entry == 0) {
2558                 tmpe = map->header.next;
2559                 entcount = map->nentries;
2560                 while (entcount-- && (tmpe != &map->header)) {
2561                         if( _vm_object_in_map(map, object, tmpe)) {
2562                                 return 1;
2563                         }
2564                         tmpe = tmpe->next;
2565                 }
2566                 return (0);
2567         }
2568         switch(entry->maptype) {
2569         case VM_MAPTYPE_SUBMAP:
2570                 tmpm = entry->object.sub_map;
2571                 tmpe = tmpm->header.next;
2572                 entcount = tmpm->nentries;
2573                 while (entcount-- && tmpe != &tmpm->header) {
2574                         if( _vm_object_in_map(tmpm, object, tmpe)) {
2575                                 return 1;
2576                         }
2577                         tmpe = tmpe->next;
2578                 }
2579                 break;
2580         case VM_MAPTYPE_NORMAL:
2581         case VM_MAPTYPE_VPAGETABLE:
2582                 obj = entry->object.vm_object;
2583                 while (obj) {
2584                         if (obj == object) {
2585                                 if (obj != entry->object.vm_object)
2586                                         vm_object_drop(obj);
2587                                 return 1;
2588                         }
2589                         while ((nobj = obj->backing_object) != NULL) {
2590                                 vm_object_hold(nobj);
2591                                 if (nobj == obj->backing_object)
2592                                         break;
2593                                 vm_object_drop(nobj);
2594                         }
2595                         if (obj != entry->object.vm_object) {
2596                                 if (nobj)
2597                                         vm_object_lock_swap();
2598                                 vm_object_drop(obj);
2599                         }
2600                         obj = nobj;
2601                 }
2602                 break;
2603         default:
2604                 break;
2605         }
2606         return 0;
2607 }
2608
2609 static int vm_object_in_map_callback(struct proc *p, void *data);
2610
2611 struct vm_object_in_map_info {
2612         vm_object_t object;
2613         int rv;
2614 };
2615
2616 /*
2617  * Debugging only
2618  */
2619 static int
2620 vm_object_in_map(vm_object_t object)
2621 {
2622         struct vm_object_in_map_info info;
2623
2624         info.rv = 0;
2625         info.object = object;
2626
2627         allproc_scan(vm_object_in_map_callback, &info);
2628         if (info.rv)
2629                 return 1;
2630         if( _vm_object_in_map(&kernel_map, object, 0))
2631                 return 1;
2632         if( _vm_object_in_map(&pager_map, object, 0))
2633                 return 1;
2634         if( _vm_object_in_map(&buffer_map, object, 0))
2635                 return 1;
2636         return 0;
2637 }
2638
2639 /*
2640  * Debugging only
2641  */
2642 static int
2643 vm_object_in_map_callback(struct proc *p, void *data)
2644 {
2645         struct vm_object_in_map_info *info = data;
2646
2647         if (p->p_vmspace) {
2648                 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) {
2649                         info->rv = 1;
2650                         return -1;
2651                 }
2652         }
2653         return (0);
2654 }
2655
2656 DB_SHOW_COMMAND(vmochk, vm_object_check)
2657 {
2658         vm_object_t object;
2659
2660         /*
2661          * make sure that internal objs are in a map somewhere
2662          * and none have zero ref counts.
2663          */
2664         for (object = TAILQ_FIRST(&vm_object_list);
2665                         object != NULL;
2666                         object = TAILQ_NEXT(object, object_list)) {
2667                 if (object->type == OBJT_MARKER)
2668                         continue;
2669                 if (object->handle == NULL &&
2670                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2671                         if (object->ref_count == 0) {
2672                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2673                                         (long)object->size);
2674                         }
2675                         if (!vm_object_in_map(object)) {
2676                                 db_printf(
2677                         "vmochk: internal obj is not in a map: "
2678                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2679                                     object->ref_count, (u_long)object->size, 
2680                                     (u_long)object->size,
2681                                     (void *)object->backing_object);
2682                         }
2683                 }
2684         }
2685 }
2686
2687 /*
2688  * Debugging only
2689  */
2690 DB_SHOW_COMMAND(object, vm_object_print_static)
2691 {
2692         /* XXX convert args. */
2693         vm_object_t object = (vm_object_t)addr;
2694         boolean_t full = have_addr;
2695
2696         vm_page_t p;
2697
2698         /* XXX count is an (unused) arg.  Avoid shadowing it. */
2699 #define count   was_count
2700
2701         int count;
2702
2703         if (object == NULL)
2704                 return;
2705
2706         db_iprintf(
2707             "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
2708             object, (int)object->type, (u_long)object->size,
2709             object->resident_page_count, object->ref_count, object->flags);
2710         /*
2711          * XXX no %qd in kernel.  Truncate object->backing_object_offset.
2712          */
2713         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
2714             object->shadow_count, 
2715             object->backing_object ? object->backing_object->ref_count : 0,
2716             object->backing_object, (long)object->backing_object_offset);
2717
2718         if (!full)
2719                 return;
2720
2721         db_indent += 2;
2722         count = 0;
2723         RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) {
2724                 if (count == 0)
2725                         db_iprintf("memory:=");
2726                 else if (count == 6) {
2727                         db_printf("\n");
2728                         db_iprintf(" ...");
2729                         count = 0;
2730                 } else
2731                         db_printf(",");
2732                 count++;
2733
2734                 db_printf("(off=0x%lx,page=0x%lx)",
2735                     (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
2736         }
2737         if (count != 0)
2738                 db_printf("\n");
2739         db_indent -= 2;
2740 }
2741
2742 /* XXX. */
2743 #undef count
2744
2745 /*
2746  * XXX need this non-static entry for calling from vm_map_print.
2747  *
2748  * Debugging only
2749  */
2750 void
2751 vm_object_print(/* db_expr_t */ long addr,
2752                 boolean_t have_addr,
2753                 /* db_expr_t */ long count,
2754                 char *modif)
2755 {
2756         vm_object_print_static(addr, have_addr, count, modif);
2757 }
2758
2759 /*
2760  * Debugging only
2761  */
2762 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2763 {
2764         vm_object_t object;
2765         int nl = 0;
2766         int c;
2767         for (object = TAILQ_FIRST(&vm_object_list);
2768                         object != NULL;
2769                         object = TAILQ_NEXT(object, object_list)) {
2770                 vm_pindex_t idx, fidx;
2771                 vm_pindex_t osize;
2772                 vm_paddr_t pa = -1, padiff;
2773                 int rcount;
2774                 vm_page_t m;
2775
2776                 if (object->type == OBJT_MARKER)
2777                         continue;
2778                 db_printf("new object: %p\n", (void *)object);
2779                 if ( nl > 18) {
2780                         c = cngetc();
2781                         if (c != ' ')
2782                                 return;
2783                         nl = 0;
2784                 }
2785                 nl++;
2786                 rcount = 0;
2787                 fidx = 0;
2788                 osize = object->size;
2789                 if (osize > 128)
2790                         osize = 128;
2791                 for (idx = 0; idx < osize; idx++) {
2792                         m = vm_page_lookup(object, idx);
2793                         if (m == NULL) {
2794                                 if (rcount) {
2795                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2796                                                 (long)fidx, rcount, (long)pa);
2797                                         if ( nl > 18) {
2798                                                 c = cngetc();
2799                                                 if (c != ' ')
2800                                                         return;
2801                                                 nl = 0;
2802                                         }
2803                                         nl++;
2804                                         rcount = 0;
2805                                 }
2806                                 continue;
2807                         }
2808
2809                                 
2810                         if (rcount &&
2811                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2812                                 ++rcount;
2813                                 continue;
2814                         }
2815                         if (rcount) {
2816                                 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
2817                                 padiff >>= PAGE_SHIFT;
2818                                 padiff &= PQ_L2_MASK;
2819                                 if (padiff == 0) {
2820                                         pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
2821                                         ++rcount;
2822                                         continue;
2823                                 }
2824                                 db_printf(" index(%ld)run(%d)pa(0x%lx)",
2825                                         (long)fidx, rcount, (long)pa);
2826                                 db_printf("pd(%ld)\n", (long)padiff);
2827                                 if ( nl > 18) {
2828                                         c = cngetc();
2829                                         if (c != ' ')
2830                                                 return;
2831                                         nl = 0;
2832                                 }
2833                                 nl++;
2834                         }
2835                         fidx = idx;
2836                         pa = VM_PAGE_TO_PHYS(m);
2837                         rcount = 1;
2838                 }
2839                 if (rcount) {
2840                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2841                                 (long)fidx, rcount, (long)pa);
2842                         if ( nl > 18) {
2843                                 c = cngetc();
2844                                 if (c != ' ')
2845                                         return;
2846                                 nl = 0;
2847                         }
2848                         nl++;
2849                 }
2850         }
2851 }
2852 #endif /* DDB */