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