kernel -- vm locking: Add vm object locking to vm_object_referenced.
[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  * $DragonFly: src/sys/vm/vm_object.c,v 1.33 2008/05/09 07:24:48 dillon Exp $
68  */
69
70 /*
71  *      Virtual memory object module.
72  */
73
74 #include <sys/param.h>
75 #include <sys/systm.h>
76 #include <sys/proc.h>           /* for curproc, pageproc */
77 #include <sys/thread.h>
78 #include <sys/vnode.h>
79 #include <sys/vmmeter.h>
80 #include <sys/mman.h>
81 #include <sys/mount.h>
82 #include <sys/kernel.h>
83 #include <sys/sysctl.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 static int      vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
102                                              int pagerflags);
103
104 /*
105  *      Virtual memory objects maintain the actual data
106  *      associated with allocated virtual memory.  A given
107  *      page of memory exists within exactly one object.
108  *
109  *      An object is only deallocated when all "references"
110  *      are given up.  Only one "reference" to a given
111  *      region of an object should be writeable.
112  *
113  *      Associated with each object is a list of all resident
114  *      memory pages belonging to that object; this list is
115  *      maintained by the "vm_page" module, and locked by the object's
116  *      lock.
117  *
118  *      Each object also records a "pager" routine which is
119  *      used to retrieve (and store) pages to the proper backing
120  *      storage.  In addition, objects may be backed by other
121  *      objects from which they were virtual-copied.
122  *
123  *      The only items within the object structure which are
124  *      modified after time of creation are:
125  *              reference count         locked by object's lock
126  *              pager routine           locked by object's lock
127  *
128  */
129
130 struct object_q vm_object_list;         /* locked by vmobj_token */
131 struct vm_object kernel_object;
132
133 static long vm_object_count;            /* locked by vmobj_token */
134 extern int vm_pageout_page_count;
135
136 static long object_collapses;
137 static long object_bypasses;
138 static int next_index;
139 static vm_zone_t obj_zone;
140 static struct vm_zone obj_zone_store;
141 static int object_hash_rand;
142 #define VM_OBJECTS_INIT 256
143 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
144
145 /*
146  * Initialize a freshly allocated object
147  *
148  * Used only by vm_object_allocate() and zinitna().
149  *
150  * No requirements.
151  */
152 void
153 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
154 {
155         int incr;
156
157         RB_INIT(&object->rb_memq);
158         LIST_INIT(&object->shadow_head);
159
160         object->type = type;
161         object->size = size;
162         object->ref_count = 1;
163         object->flags = 0;
164         if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
165                 vm_object_set_flag(object, OBJ_ONEMAPPING);
166         object->paging_in_progress = 0;
167         object->resident_page_count = 0;
168         object->agg_pv_list_count = 0;
169         object->shadow_count = 0;
170         object->pg_color = next_index;
171         if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
172                 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
173         else
174                 incr = size;
175         next_index = (next_index + incr) & PQ_L2_MASK;
176         object->handle = NULL;
177         object->backing_object = NULL;
178         object->backing_object_offset = (vm_ooffset_t) 0;
179         /*
180          * Try to generate a number that will spread objects out in the
181          * hash table.  We 'wipe' new objects across the hash in 128 page
182          * increments plus 1 more to offset it a little more by the time
183          * it wraps around.
184          */
185         object->hash_rand = object_hash_rand - 129;
186
187         object->generation++;
188         object->swblock_count = 0;
189         RB_INIT(&object->swblock_root);
190         lwkt_token_init(&object->tok, "vmobjtk");
191
192         lwkt_gettoken(&vmobj_token);
193         TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
194         vm_object_count++;
195         object_hash_rand = object->hash_rand;
196         lwkt_reltoken(&vmobj_token);
197 }
198
199 /*
200  * Initialize the VM objects module.
201  *
202  * Called from the low level boot code only.
203  */
204 void
205 vm_object_init(void)
206 {
207         TAILQ_INIT(&vm_object_list);
208         
209         _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd),
210                             &kernel_object);
211
212         obj_zone = &obj_zone_store;
213         zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
214                 vm_objects_init, VM_OBJECTS_INIT);
215 }
216
217 void
218 vm_object_init2(void)
219 {
220         zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1);
221 }
222
223 /*
224  * Allocate and return a new object of the specified type and size.
225  *
226  * No requirements.
227  */
228 vm_object_t
229 vm_object_allocate(objtype_t type, vm_pindex_t size)
230 {
231         vm_object_t result;
232
233         result = (vm_object_t) zalloc(obj_zone);
234
235         _vm_object_allocate(type, size, result);
236
237         return (result);
238 }
239
240 /*
241  * Add an additional reference to a vm_object.
242  *
243  * Object passed by caller must be stable or caller must already
244  * hold vmobj_token to avoid races.
245  */
246 void
247 vm_object_reference(vm_object_t object)
248 {
249         lwkt_gettoken(&vmobj_token);
250         vm_object_reference_locked(object);
251         lwkt_reltoken(&vmobj_token);
252 }
253
254 void
255 vm_object_reference_locked(vm_object_t object)
256 {
257         if (object) {
258                 ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
259                 vm_object_lock(object);
260                 object->ref_count++;
261                 if (object->type == OBJT_VNODE) {
262                         vref(object->handle);
263                         /* XXX what if the vnode is being destroyed? */
264                 }
265                 vm_object_unlock(object);
266         }
267 }
268
269 /*
270  * Dereference an object and its underlying vnode.
271  *
272  * The caller must hold vmobj_token.
273  */
274 static void
275 vm_object_vndeallocate(vm_object_t object)
276 {
277         struct vnode *vp = (struct vnode *) object->handle;
278
279         KASSERT(object->type == OBJT_VNODE,
280             ("vm_object_vndeallocate: not a vnode object"));
281         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
282         ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
283 #ifdef INVARIANTS
284         if (object->ref_count == 0) {
285                 vprint("vm_object_vndeallocate", vp);
286                 panic("vm_object_vndeallocate: bad object reference count");
287         }
288 #endif
289
290         object->ref_count--;
291         if (object->ref_count == 0)
292                 vclrflags(vp, VTEXT);
293         vrele(vp);
294 }
295
296 /*
297  * Release a reference to the specified object, gained either through a
298  * vm_object_allocate or a vm_object_reference call.  When all references
299  * are gone, storage associated with this object may be relinquished.
300  */
301 void
302 vm_object_deallocate(vm_object_t object)
303 {
304         lwkt_gettoken(&vmobj_token);
305         vm_object_deallocate_locked(object);
306         lwkt_reltoken(&vmobj_token);
307 }
308
309 void
310 vm_object_deallocate_locked(vm_object_t object)
311 {
312         vm_object_t temp;
313
314         ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
315
316         while (object != NULL) {
317                 if (object->type == OBJT_VNODE) {
318                         vm_object_vndeallocate(object);
319                         break;
320                 }
321
322                 if (object->ref_count == 0) {
323                         panic("vm_object_deallocate: object deallocated "
324                               "too many times: %d", object->type);
325                 }
326                 if (object->ref_count > 2) {
327                         object->ref_count--;
328                         break;
329                 }
330
331                 /*
332                  * We currently need the vm_token from this point on, and
333                  * we must recheck ref_count after acquiring it.
334                  */
335                 lwkt_gettoken(&vm_token);
336
337                 if (object->ref_count > 2) {
338                         object->ref_count--;
339                         lwkt_reltoken(&vm_token);
340                         break;
341                 }
342
343                 /*
344                  * Here on ref_count of one or two, which are special cases for
345                  * objects.
346                  */
347                 if ((object->ref_count == 2) && (object->shadow_count == 0)) {
348                         vm_object_set_flag(object, OBJ_ONEMAPPING);
349                         object->ref_count--;
350                         lwkt_reltoken(&vm_token);
351                         break;
352                 }
353                 if ((object->ref_count == 2) && (object->shadow_count == 1)) {
354                         object->ref_count--;
355                         if ((object->handle == NULL) &&
356                             (object->type == OBJT_DEFAULT ||
357                              object->type == OBJT_SWAP)) {
358                                 vm_object_t robject;
359
360                                 robject = LIST_FIRST(&object->shadow_head);
361                                 KASSERT(robject != NULL,
362                                         ("vm_object_deallocate: ref_count: "
363                                         "%d, shadow_count: %d",
364                                         object->ref_count,
365                                         object->shadow_count));
366
367                                 if ((robject->handle == NULL) &&
368                                     (robject->type == OBJT_DEFAULT ||
369                                      robject->type == OBJT_SWAP)) {
370
371                                         robject->ref_count++;
372
373                                         while (
374                                                 robject->paging_in_progress ||
375                                                 object->paging_in_progress
376                                         ) {
377                                                 vm_object_pip_sleep(robject, "objde1");
378                                                 vm_object_pip_sleep(object, "objde2");
379                                         }
380
381                                         if (robject->ref_count == 1) {
382                                                 robject->ref_count--;
383                                                 object = robject;
384                                                 goto doterm;
385                                         }
386
387                                         object = robject;
388                                         vm_object_collapse(object);
389                                         lwkt_reltoken(&vm_token);
390                                         continue;
391                                 }
392                         }
393                         lwkt_reltoken(&vm_token);
394                         break;
395                 }
396
397                 /*
398                  * Normal dereferencing path
399                  */
400                 object->ref_count--;
401                 if (object->ref_count != 0) {
402                         lwkt_reltoken(&vm_token);
403                         break;
404                 }
405
406                 /*
407                  * Termination path
408                  */
409 doterm:
410                 temp = object->backing_object;
411                 if (temp) {
412                         LIST_REMOVE(object, shadow_list);
413                         temp->shadow_count--;
414                         temp->generation++;
415                         object->backing_object = NULL;
416                 }
417                 lwkt_reltoken(&vm_token);
418
419                 /*
420                  * Don't double-terminate, we could be in a termination
421                  * recursion due to the terminate having to sync data
422                  * to disk.
423                  */
424                 if ((object->flags & OBJ_DEAD) == 0)
425                         vm_object_terminate(object);
426                 object = temp;
427         }
428 }
429
430 /*
431  * Destroy the specified object, freeing up related resources.
432  *
433  * The object must have zero references.
434  *
435  * The caller must be holding vmobj_token and properly interlock with
436  * OBJ_DEAD.
437  */
438 static int vm_object_terminate_callback(vm_page_t p, void *data);
439
440 void
441 vm_object_terminate(vm_object_t object)
442 {
443         /*
444          * Make sure no one uses us.  Once we set OBJ_DEAD we should be
445          * able to safely block.
446          */
447         KKASSERT((object->flags & OBJ_DEAD) == 0);
448         ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
449         vm_object_set_flag(object, OBJ_DEAD);
450
451         /*
452          * Wait for the pageout daemon to be done with the object
453          */
454         vm_object_pip_wait(object, "objtrm");
455
456         KASSERT(!object->paging_in_progress,
457                 ("vm_object_terminate: pageout in progress"));
458
459         /*
460          * Clean and free the pages, as appropriate. All references to the
461          * object are gone, so we don't need to lock it.
462          */
463         if (object->type == OBJT_VNODE) {
464                 struct vnode *vp;
465
466                 /*
467                  * Clean pages and flush buffers.
468                  */
469                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
470
471                 vp = (struct vnode *) object->handle;
472                 vinvalbuf(vp, V_SAVE, 0, 0);
473         }
474
475         /*
476          * Wait for any I/O to complete, after which there had better not
477          * be any references left on the object.
478          */
479         vm_object_pip_wait(object, "objtrm");
480
481         if (object->ref_count != 0) {
482                 panic("vm_object_terminate: object with references, "
483                       "ref_count=%d", object->ref_count);
484         }
485
486         /*
487          * Now free any remaining pages. For internal objects, this also
488          * removes them from paging queues. Don't free wired pages, just
489          * remove them from the object. 
490          */
491         lwkt_gettoken(&vm_token);
492         vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
493                                 vm_object_terminate_callback, NULL);
494         lwkt_reltoken(&vm_token);
495
496         /*
497          * Let the pager know object is dead.
498          */
499         vm_pager_deallocate(object);
500
501         /*
502          * Remove the object from the global object list.
503          *
504          * (we are holding vmobj_token)
505          */
506         TAILQ_REMOVE(&vm_object_list, object, object_list);
507         vm_object_count--;
508         vm_object_dead_wakeup(object);
509
510         if (object->ref_count != 0) {
511                 panic("vm_object_terminate2: object with references, "
512                       "ref_count=%d", object->ref_count);
513         }
514
515         /*
516          * Free the space for the object.
517          */
518         zfree(obj_zone, object);
519 }
520
521 /*
522  * The caller must hold vm_token.
523  */
524 static int
525 vm_object_terminate_callback(vm_page_t p, void *data __unused)
526 {
527         if (p->busy || (p->flags & PG_BUSY))
528                 panic("vm_object_terminate: freeing busy page %p", p);
529         if (p->wire_count == 0) {
530                 vm_page_busy(p);
531                 vm_page_free(p);
532                 mycpu->gd_cnt.v_pfree++;
533         } else {
534                 if (p->queue != PQ_NONE)
535                         kprintf("vm_object_terminate: Warning: Encountered wired page %p on queue %d\n", p, p->queue);
536                 vm_page_busy(p);
537                 vm_page_remove(p);
538                 vm_page_wakeup(p);
539         }
540         return(0);
541 }
542
543 /*
544  * The object is dead but still has an object<->pager association.  Sleep
545  * and return.  The caller typically retests the association in a loop.
546  *
547  * Must be called with the vmobj_token held.
548  */
549 void
550 vm_object_dead_sleep(vm_object_t object, const char *wmesg)
551 {
552         ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
553         if (object->handle) {
554                 vm_object_set_flag(object, OBJ_DEADWNT);
555                 tsleep(object, 0, wmesg, 0);
556                 /* object may be invalid after this point */
557         }
558 }
559
560 /*
561  * Wakeup anyone waiting for the object<->pager disassociation on
562  * a dead object.
563  *
564  * Must be called with the vmobj_token held.
565  */
566 void
567 vm_object_dead_wakeup(vm_object_t object)
568 {
569         ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
570         if (object->flags & OBJ_DEADWNT) {
571                 vm_object_clear_flag(object, OBJ_DEADWNT);
572                 wakeup(object);
573         }
574 }
575
576 /*
577  * Clean all dirty pages in the specified range of object.  Leaves page
578  * on whatever queue it is currently on.   If NOSYNC is set then do not
579  * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
580  * leaving the object dirty.
581  *
582  * When stuffing pages asynchronously, allow clustering.  XXX we need a
583  * synchronous clustering mode implementation.
584  *
585  * Odd semantics: if start == end, we clean everything.
586  *
587  * The object must be locked? XXX
588  */
589 static int vm_object_page_clean_pass1(struct vm_page *p, void *data);
590 static int vm_object_page_clean_pass2(struct vm_page *p, void *data);
591
592 void
593 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
594                      int flags)
595 {
596         struct rb_vm_page_scan_info info;
597         struct vnode *vp;
598         int wholescan;
599         int pagerflags;
600         int curgeneration;
601
602         lwkt_gettoken(&vm_token);
603         if (object->type != OBJT_VNODE ||
604             (object->flags & OBJ_MIGHTBEDIRTY) == 0) {
605                 lwkt_reltoken(&vm_token);
606                 return;
607         }
608
609         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? 
610                         VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
611         pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
612
613         vp = object->handle;
614
615         /*
616          * Interlock other major object operations.  This allows us to 
617          * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY.
618          */
619         crit_enter();
620         vm_object_set_flag(object, OBJ_CLEANING);
621
622         /*
623          * Handle 'entire object' case
624          */
625         info.start_pindex = start;
626         if (end == 0) {
627                 info.end_pindex = object->size - 1;
628         } else {
629                 info.end_pindex = end - 1;
630         }
631         wholescan = (start == 0 && info.end_pindex == object->size - 1);
632         info.limit = flags;
633         info.pagerflags = pagerflags;
634         info.object = object;
635
636         /*
637          * If cleaning the entire object do a pass to mark the pages read-only.
638          * If everything worked out ok, clear OBJ_WRITEABLE and
639          * OBJ_MIGHTBEDIRTY.
640          */
641         if (wholescan) {
642                 info.error = 0;
643                 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
644                                         vm_object_page_clean_pass1, &info);
645                 if (info.error == 0) {
646                         vm_object_clear_flag(object,
647                                              OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
648                         if (object->type == OBJT_VNODE &&
649                             (vp = (struct vnode *)object->handle) != NULL) {
650                                 if (vp->v_flag & VOBJDIRTY) 
651                                         vclrflags(vp, VOBJDIRTY);
652                         }
653                 }
654         }
655
656         /*
657          * Do a pass to clean all the dirty pages we find.
658          */
659         do {
660                 info.error = 0;
661                 curgeneration = object->generation;
662                 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
663                                         vm_object_page_clean_pass2, &info);
664         } while (info.error || curgeneration != object->generation);
665
666         vm_object_clear_flag(object, OBJ_CLEANING);
667         crit_exit();
668         lwkt_reltoken(&vm_token);
669 }
670
671 /*
672  * The caller must hold vm_token.
673  */
674 static 
675 int
676 vm_object_page_clean_pass1(struct vm_page *p, void *data)
677 {
678         struct rb_vm_page_scan_info *info = data;
679
680         vm_page_flag_set(p, PG_CLEANCHK);
681         if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
682                 info->error = 1;
683         else
684                 vm_page_protect(p, VM_PROT_READ);       /* must not block */
685         return(0);
686 }
687
688 /*
689  * The caller must hold vm_token.
690  */
691 static 
692 int
693 vm_object_page_clean_pass2(struct vm_page *p, void *data)
694 {
695         struct rb_vm_page_scan_info *info = data;
696         int n;
697
698         /*
699          * Do not mess with pages that were inserted after we started
700          * the cleaning pass.
701          */
702         if ((p->flags & PG_CLEANCHK) == 0)
703                 return(0);
704
705         /*
706          * Before wasting time traversing the pmaps, check for trivial
707          * cases where the page cannot be dirty.
708          */
709         if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) {
710                 KKASSERT((p->dirty & p->valid) == 0);
711                 return(0);
712         }
713
714         /*
715          * Check whether the page is dirty or not.  The page has been set
716          * to be read-only so the check will not race a user dirtying the
717          * page.
718          */
719         vm_page_test_dirty(p);
720         if ((p->dirty & p->valid) == 0) {
721                 vm_page_flag_clear(p, PG_CLEANCHK);
722                 return(0);
723         }
724
725         /*
726          * If we have been asked to skip nosync pages and this is a
727          * nosync page, skip it.  Note that the object flags were
728          * not cleared in this case (because pass1 will have returned an
729          * error), so we do not have to set them.
730          */
731         if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
732                 vm_page_flag_clear(p, PG_CLEANCHK);
733                 return(0);
734         }
735
736         /*
737          * Flush as many pages as we can.  PG_CLEANCHK will be cleared on
738          * the pages that get successfully flushed.  Set info->error if
739          * we raced an object modification.
740          */
741         n = vm_object_page_collect_flush(info->object, p, info->pagerflags);
742         if (n == 0)
743                 info->error = 1;
744         return(0);
745 }
746
747 /*
748  * Collect the specified page and nearby pages and flush them out.
749  * The number of pages flushed is returned.
750  *
751  * The caller must hold vm_token.
752  */
753 static int
754 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags)
755 {
756         int runlen;
757         int maxf;
758         int chkb;
759         int maxb;
760         int i;
761         int curgeneration;
762         vm_pindex_t pi;
763         vm_page_t maf[vm_pageout_page_count];
764         vm_page_t mab[vm_pageout_page_count];
765         vm_page_t ma[vm_pageout_page_count];
766
767         curgeneration = object->generation;
768
769         pi = p->pindex;
770         while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
771                 if (object->generation != curgeneration) {
772                         return(0);
773                 }
774         }
775         KKASSERT(p->object == object && p->pindex == pi);
776
777         maxf = 0;
778         for(i = 1; i < vm_pageout_page_count; i++) {
779                 vm_page_t tp;
780
781                 if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
782                         if ((tp->flags & PG_BUSY) ||
783                                 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 
784                                  (tp->flags & PG_CLEANCHK) == 0) ||
785                                 (tp->busy != 0))
786                                 break;
787                         if((tp->queue - tp->pc) == PQ_CACHE) {
788                                 vm_page_flag_clear(tp, PG_CLEANCHK);
789                                 break;
790                         }
791                         vm_page_test_dirty(tp);
792                         if ((tp->dirty & tp->valid) == 0) {
793                                 vm_page_flag_clear(tp, PG_CLEANCHK);
794                                 break;
795                         }
796                         maf[ i - 1 ] = tp;
797                         maxf++;
798                         continue;
799                 }
800                 break;
801         }
802
803         maxb = 0;
804         chkb = vm_pageout_page_count -  maxf;
805         if (chkb) {
806                 for(i = 1; i < chkb;i++) {
807                         vm_page_t tp;
808
809                         if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
810                                 if ((tp->flags & PG_BUSY) ||
811                                         ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 
812                                          (tp->flags & PG_CLEANCHK) == 0) ||
813                                         (tp->busy != 0))
814                                         break;
815                                 if((tp->queue - tp->pc) == PQ_CACHE) {
816                                         vm_page_flag_clear(tp, PG_CLEANCHK);
817                                         break;
818                                 }
819                                 vm_page_test_dirty(tp);
820                                 if ((tp->dirty & tp->valid) == 0) {
821                                         vm_page_flag_clear(tp, PG_CLEANCHK);
822                                         break;
823                                 }
824                                 mab[ i - 1 ] = tp;
825                                 maxb++;
826                                 continue;
827                         }
828                         break;
829                 }
830         }
831
832         for(i = 0; i < maxb; i++) {
833                 int index = (maxb - i) - 1;
834                 ma[index] = mab[i];
835                 vm_page_flag_clear(ma[index], PG_CLEANCHK);
836         }
837         vm_page_flag_clear(p, PG_CLEANCHK);
838         ma[maxb] = p;
839         for(i = 0; i < maxf; i++) {
840                 int index = (maxb + i) + 1;
841                 ma[index] = maf[i];
842                 vm_page_flag_clear(ma[index], PG_CLEANCHK);
843         }
844         runlen = maxb + maxf + 1;
845
846         vm_pageout_flush(ma, runlen, pagerflags);
847         for (i = 0; i < runlen; i++) {
848                 if (ma[i]->valid & ma[i]->dirty) {
849                         vm_page_protect(ma[i], VM_PROT_READ);
850                         vm_page_flag_set(ma[i], PG_CLEANCHK);
851
852                         /*
853                          * maxf will end up being the actual number of pages
854                          * we wrote out contiguously, non-inclusive of the
855                          * first page.  We do not count look-behind pages.
856                          */
857                         if (i >= maxb + 1 && (maxf > i - maxb - 1))
858                                 maxf = i - maxb - 1;
859                 }
860         }
861         return(maxf + 1);
862 }
863
864 /*
865  * Same as vm_object_pmap_copy, except range checking really
866  * works, and is meant for small sections of an object.
867  *
868  * This code protects resident pages by making them read-only
869  * and is typically called on a fork or split when a page
870  * is converted to copy-on-write.  
871  *
872  * NOTE: If the page is already at VM_PROT_NONE, calling
873  * vm_page_protect will have no effect.
874  */
875 void
876 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
877 {
878         vm_pindex_t idx;
879         vm_page_t p;
880
881         if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
882                 return;
883
884         /*
885          * spl protection needed to prevent races between the lookup,
886          * an interrupt unbusy/free, and our protect call.
887          */
888         crit_enter();
889         lwkt_gettoken(&vm_token);
890         for (idx = start; idx < end; idx++) {
891                 p = vm_page_lookup(object, idx);
892                 if (p == NULL)
893                         continue;
894                 vm_page_protect(p, VM_PROT_READ);
895         }
896         lwkt_reltoken(&vm_token);
897         crit_exit();
898 }
899
900 /*
901  * Removes all physical pages in the specified object range from all
902  * physical maps.
903  *
904  * The object must *not* be locked.
905  */
906
907 static int vm_object_pmap_remove_callback(vm_page_t p, void *data);
908
909 void
910 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
911 {
912         struct rb_vm_page_scan_info info;
913
914         if (object == NULL)
915                 return;
916         info.start_pindex = start;
917         info.end_pindex = end - 1;
918
919         crit_enter();
920         lwkt_gettoken(&vm_token);
921         vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
922                                 vm_object_pmap_remove_callback, &info);
923         if (start == 0 && end == object->size)
924                 vm_object_clear_flag(object, OBJ_WRITEABLE);
925         lwkt_reltoken(&vm_token);
926         crit_exit();
927 }
928
929 /*
930  * The caller must hold vm_token.
931  */
932 static int
933 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused)
934 {
935         vm_page_protect(p, VM_PROT_NONE);
936         return(0);
937 }
938
939 /*
940  * Implements the madvise function at the object/page level.
941  *
942  * MADV_WILLNEED        (any object)
943  *
944  *      Activate the specified pages if they are resident.
945  *
946  * MADV_DONTNEED        (any object)
947  *
948  *      Deactivate the specified pages if they are resident.
949  *
950  * MADV_FREE    (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only)
951  *
952  *      Deactivate and clean the specified pages if they are
953  *      resident.  This permits the process to reuse the pages
954  *      without faulting or the kernel to reclaim the pages
955  *      without I/O.
956  *
957  * No requirements.
958  */
959 void
960 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
961 {
962         vm_pindex_t end, tpindex;
963         vm_object_t tobject;
964         vm_page_t m;
965
966         if (object == NULL)
967                 return;
968
969         end = pindex + count;
970
971         lwkt_gettoken(&vm_token);
972
973         /*
974          * Locate and adjust resident pages
975          */
976         for (; pindex < end; pindex += 1) {
977 relookup:
978                 tobject = object;
979                 tpindex = pindex;
980 shadowlookup:
981                 /*
982                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
983                  * and those pages must be OBJ_ONEMAPPING.
984                  */
985                 if (advise == MADV_FREE) {
986                         if ((tobject->type != OBJT_DEFAULT &&
987                              tobject->type != OBJT_SWAP) ||
988                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
989                                 continue;
990                         }
991                 }
992
993                 /*
994                  * spl protection is required to avoid a race between the
995                  * lookup, an interrupt unbusy/free, and our busy check.
996                  */
997
998                 crit_enter();
999                 m = vm_page_lookup(tobject, tpindex);
1000
1001                 if (m == NULL) {
1002                         /*
1003                          * There may be swap even if there is no backing page
1004                          */
1005                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1006                                 swap_pager_freespace(tobject, tpindex, 1);
1007
1008                         /*
1009                          * next object
1010                          */
1011                         crit_exit();
1012                         if (tobject->backing_object == NULL)
1013                                 continue;
1014                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1015                         tobject = tobject->backing_object;
1016                         goto shadowlookup;
1017                 }
1018
1019                 /*
1020                  * If the page is busy or not in a normal active state,
1021                  * we skip it.  If the page is not managed there are no
1022                  * page queues to mess with.  Things can break if we mess
1023                  * with pages in any of the below states.
1024                  */
1025                 if (
1026                     m->hold_count ||
1027                     m->wire_count ||
1028                     (m->flags & PG_UNMANAGED) ||
1029                     m->valid != VM_PAGE_BITS_ALL
1030                 ) {
1031                         crit_exit();
1032                         continue;
1033                 }
1034
1035                 if (vm_page_sleep_busy(m, TRUE, "madvpo")) {
1036                         crit_exit();
1037                         goto relookup;
1038                 }
1039                 vm_page_busy(m);
1040                 crit_exit();
1041
1042                 /*
1043                  * Theoretically once a page is known not to be busy, an
1044                  * interrupt cannot come along and rip it out from under us.
1045                  */
1046
1047                 if (advise == MADV_WILLNEED) {
1048                         vm_page_activate(m);
1049                 } else if (advise == MADV_DONTNEED) {
1050                         vm_page_dontneed(m);
1051                 } else if (advise == MADV_FREE) {
1052                         /*
1053                          * Mark the page clean.  This will allow the page
1054                          * to be freed up by the system.  However, such pages
1055                          * are often reused quickly by malloc()/free()
1056                          * so we do not do anything that would cause
1057                          * a page fault if we can help it.
1058                          *
1059                          * Specifically, we do not try to actually free
1060                          * the page now nor do we try to put it in the
1061                          * cache (which would cause a page fault on reuse).
1062                          *
1063                          * But we do make the page is freeable as we
1064                          * can without actually taking the step of unmapping
1065                          * it.
1066                          */
1067                         pmap_clear_modify(m);
1068                         m->dirty = 0;
1069                         m->act_count = 0;
1070                         vm_page_dontneed(m);
1071                         if (tobject->type == OBJT_SWAP)
1072                                 swap_pager_freespace(tobject, tpindex, 1);
1073                 }
1074                 vm_page_wakeup(m);
1075         }       
1076         lwkt_reltoken(&vm_token);
1077 }
1078
1079 /*
1080  * Create a new object which is backed by the specified existing object
1081  * range.  The source object reference is deallocated.
1082  *
1083  * The new object and offset into that object are returned in the source
1084  * parameters.
1085  *
1086  * No other requirements.
1087  */
1088 void
1089 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length)
1090 {
1091         vm_object_t source;
1092         vm_object_t result;
1093
1094         source = *object;
1095
1096         /*
1097          * Don't create the new object if the old object isn't shared.
1098          */
1099         lwkt_gettoken(&vm_token);
1100
1101         if (source != NULL &&
1102             source->ref_count == 1 &&
1103             source->handle == NULL &&
1104             (source->type == OBJT_DEFAULT ||
1105              source->type == OBJT_SWAP)) {
1106                 lwkt_reltoken(&vm_token);
1107                 return;
1108         }
1109
1110         /*
1111          * Allocate a new object with the given length
1112          */
1113
1114         if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1115                 panic("vm_object_shadow: no object for shadowing");
1116
1117         /*
1118          * The new object shadows the source object, adding a reference to it.
1119          * Our caller changes his reference to point to the new object,
1120          * removing a reference to the source object.  Net result: no change
1121          * of reference count.
1122          *
1123          * Try to optimize the result object's page color when shadowing
1124          * in order to maintain page coloring consistency in the combined 
1125          * shadowed object.
1126          */
1127         result->backing_object = source;
1128         if (source) {
1129                 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1130                 source->shadow_count++;
1131                 source->generation++;
1132                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1133         }
1134
1135         /*
1136          * Store the offset into the source object, and fix up the offset into
1137          * the new object.
1138          */
1139         result->backing_object_offset = *offset;
1140         lwkt_reltoken(&vm_token);
1141
1142         /*
1143          * Return the new things
1144          */
1145         *offset = 0;
1146         *object = result;
1147 }
1148
1149 #define OBSC_TEST_ALL_SHADOWED  0x0001
1150 #define OBSC_COLLAPSE_NOWAIT    0x0002
1151 #define OBSC_COLLAPSE_WAIT      0x0004
1152
1153 static int vm_object_backing_scan_callback(vm_page_t p, void *data);
1154
1155 /*
1156  * The caller must hold vm_token.
1157  */
1158 static __inline int
1159 vm_object_backing_scan(vm_object_t object, int op)
1160 {
1161         struct rb_vm_page_scan_info info;
1162         vm_object_t backing_object;
1163
1164         crit_enter();
1165
1166         backing_object = object->backing_object;
1167         info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1168
1169         /*
1170          * Initial conditions
1171          */
1172
1173         if (op & OBSC_TEST_ALL_SHADOWED) {
1174                 /*
1175                  * We do not want to have to test for the existence of
1176                  * swap pages in the backing object.  XXX but with the
1177                  * new swapper this would be pretty easy to do.
1178                  *
1179                  * XXX what about anonymous MAP_SHARED memory that hasn't
1180                  * been ZFOD faulted yet?  If we do not test for this, the
1181                  * shadow test may succeed! XXX
1182                  */
1183                 if (backing_object->type != OBJT_DEFAULT) {
1184                         crit_exit();
1185                         return(0);
1186                 }
1187         }
1188         if (op & OBSC_COLLAPSE_WAIT) {
1189                 KKASSERT((backing_object->flags & OBJ_DEAD) == 0);
1190                 vm_object_set_flag(backing_object, OBJ_DEAD);
1191         }
1192
1193         /*
1194          * Our scan.   We have to retry if a negative error code is returned,
1195          * otherwise 0 or 1 will be returned in info.error.  0 Indicates that
1196          * the scan had to be stopped because the parent does not completely
1197          * shadow the child.
1198          */
1199         info.object = object;
1200         info.backing_object = backing_object;
1201         info.limit = op;
1202         do {
1203                 info.error = 1;
1204                 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL,
1205                                         vm_object_backing_scan_callback,
1206                                         &info);
1207         } while (info.error < 0);
1208         crit_exit();
1209         return(info.error);
1210 }
1211
1212 /*
1213  * The caller must hold vm_token.
1214  */
1215 static int
1216 vm_object_backing_scan_callback(vm_page_t p, void *data)
1217 {
1218         struct rb_vm_page_scan_info *info = data;
1219         vm_object_t backing_object;
1220         vm_object_t object;
1221         vm_pindex_t new_pindex;
1222         vm_pindex_t backing_offset_index;
1223         int op;
1224
1225         new_pindex = p->pindex - info->backing_offset_index;
1226         op = info->limit;
1227         object = info->object;
1228         backing_object = info->backing_object;
1229         backing_offset_index = info->backing_offset_index;
1230
1231         if (op & OBSC_TEST_ALL_SHADOWED) {
1232                 vm_page_t pp;
1233
1234                 /*
1235                  * Ignore pages outside the parent object's range
1236                  * and outside the parent object's mapping of the 
1237                  * backing object.
1238                  *
1239                  * note that we do not busy the backing object's
1240                  * page.
1241                  */
1242                 if (
1243                     p->pindex < backing_offset_index ||
1244                     new_pindex >= object->size
1245                 ) {
1246                         return(0);
1247                 }
1248
1249                 /*
1250                  * See if the parent has the page or if the parent's
1251                  * object pager has the page.  If the parent has the
1252                  * page but the page is not valid, the parent's
1253                  * object pager must have the page.
1254                  *
1255                  * If this fails, the parent does not completely shadow
1256                  * the object and we might as well give up now.
1257                  */
1258
1259                 pp = vm_page_lookup(object, new_pindex);
1260                 if ((pp == NULL || pp->valid == 0) &&
1261                     !vm_pager_has_page(object, new_pindex)
1262                 ) {
1263                         info->error = 0;        /* problemo */
1264                         return(-1);             /* stop the scan */
1265                 }
1266         }
1267
1268         /*
1269          * Check for busy page
1270          */
1271
1272         if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1273                 vm_page_t pp;
1274
1275                 if (op & OBSC_COLLAPSE_NOWAIT) {
1276                         if (
1277                             (p->flags & PG_BUSY) ||
1278                             !p->valid || 
1279                             p->hold_count || 
1280                             p->wire_count ||
1281                             p->busy
1282                         ) {
1283                                 return(0);
1284                         }
1285                 } else if (op & OBSC_COLLAPSE_WAIT) {
1286                         if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1287                                 /*
1288                                  * If we slept, anything could have
1289                                  * happened.   Ask that the scan be restarted.
1290                                  *
1291                                  * Since the object is marked dead, the
1292                                  * backing offset should not have changed.  
1293                                  */
1294                                 info->error = -1;
1295                                 return(-1);
1296                         }
1297                 }
1298
1299                 /* 
1300                  * Busy the page
1301                  */
1302                 vm_page_busy(p);
1303
1304                 KASSERT(
1305                     p->object == backing_object,
1306                     ("vm_object_qcollapse(): object mismatch")
1307                 );
1308
1309                 /*
1310                  * Destroy any associated swap
1311                  */
1312                 if (backing_object->type == OBJT_SWAP)
1313                         swap_pager_freespace(backing_object, p->pindex, 1);
1314
1315                 if (
1316                     p->pindex < backing_offset_index ||
1317                     new_pindex >= object->size
1318                 ) {
1319                         /*
1320                          * Page is out of the parent object's range, we 
1321                          * can simply destroy it. 
1322                          */
1323                         vm_page_protect(p, VM_PROT_NONE);
1324                         vm_page_free(p);
1325                         return(0);
1326                 }
1327
1328                 pp = vm_page_lookup(object, new_pindex);
1329                 if (pp != NULL || vm_pager_has_page(object, new_pindex)) {
1330                         /*
1331                          * page already exists in parent OR swap exists
1332                          * for this location in the parent.  Destroy 
1333                          * the original page from the backing object.
1334                          *
1335                          * Leave the parent's page alone
1336                          */
1337                         vm_page_protect(p, VM_PROT_NONE);
1338                         vm_page_free(p);
1339                         return(0);
1340                 }
1341
1342                 /*
1343                  * Page does not exist in parent, rename the
1344                  * page from the backing object to the main object. 
1345                  *
1346                  * If the page was mapped to a process, it can remain 
1347                  * mapped through the rename.
1348                  */
1349                 if ((p->queue - p->pc) == PQ_CACHE)
1350                         vm_page_deactivate(p);
1351
1352                 vm_page_rename(p, object, new_pindex);
1353                 /* page automatically made dirty by rename */
1354         }
1355         return(0);
1356 }
1357
1358 /*
1359  * This version of collapse allows the operation to occur earlier and
1360  * when paging_in_progress is true for an object...  This is not a complete
1361  * operation, but should plug 99.9% of the rest of the leaks.
1362  *
1363  * The caller must hold vm_token and vmobj_token.
1364  * (only called from vm_object_collapse)
1365  */
1366 static void
1367 vm_object_qcollapse(vm_object_t object)
1368 {
1369         vm_object_t backing_object = object->backing_object;
1370
1371         if (backing_object->ref_count != 1)
1372                 return;
1373
1374         backing_object->ref_count += 2;
1375
1376         vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1377
1378         backing_object->ref_count -= 2;
1379 }
1380
1381 /*
1382  * Collapse an object with the object backing it.  Pages in the backing
1383  * object are moved into the parent, and the backing object is deallocated.
1384  */
1385 void
1386 vm_object_collapse(vm_object_t object)
1387 {
1388         ASSERT_LWKT_TOKEN_HELD(&vm_token);
1389         ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
1390
1391         while (TRUE) {
1392                 vm_object_t backing_object;
1393
1394                 /*
1395                  * Verify that the conditions are right for collapse:
1396                  *
1397                  * The object exists and the backing object exists.
1398                  */
1399                 if (object == NULL)
1400                         break;
1401
1402                 if ((backing_object = object->backing_object) == NULL)
1403                         break;
1404
1405                 /*
1406                  * we check the backing object first, because it is most likely
1407                  * not collapsable.
1408                  */
1409                 if (backing_object->handle != NULL ||
1410                     (backing_object->type != OBJT_DEFAULT &&
1411                      backing_object->type != OBJT_SWAP) ||
1412                     (backing_object->flags & OBJ_DEAD) ||
1413                     object->handle != NULL ||
1414                     (object->type != OBJT_DEFAULT &&
1415                      object->type != OBJT_SWAP) ||
1416                     (object->flags & OBJ_DEAD)) {
1417                         break;
1418                 }
1419
1420                 if (
1421                     object->paging_in_progress != 0 ||
1422                     backing_object->paging_in_progress != 0
1423                 ) {
1424                         vm_object_qcollapse(object);
1425                         break;
1426                 }
1427
1428                 /*
1429                  * We know that we can either collapse the backing object (if
1430                  * the parent is the only reference to it) or (perhaps) have
1431                  * the parent bypass the object if the parent happens to shadow
1432                  * all the resident pages in the entire backing object.
1433                  *
1434                  * This is ignoring pager-backed pages such as swap pages.
1435                  * vm_object_backing_scan fails the shadowing test in this
1436                  * case.
1437                  */
1438
1439                 if (backing_object->ref_count == 1) {
1440                         /*
1441                          * If there is exactly one reference to the backing
1442                          * object, we can collapse it into the parent.  
1443                          */
1444                         vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1445
1446                         /*
1447                          * Move the pager from backing_object to object.
1448                          */
1449
1450                         if (backing_object->type == OBJT_SWAP) {
1451                                 vm_object_pip_add(backing_object, 1);
1452
1453                                 /*
1454                                  * scrap the paging_offset junk and do a 
1455                                  * discrete copy.  This also removes major 
1456                                  * assumptions about how the swap-pager 
1457                                  * works from where it doesn't belong.  The
1458                                  * new swapper is able to optimize the
1459                                  * destroy-source case.
1460                                  */
1461
1462                                 vm_object_pip_add(object, 1);
1463                                 swap_pager_copy(
1464                                     backing_object,
1465                                     object,
1466                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1467                                 vm_object_pip_wakeup(object);
1468
1469                                 vm_object_pip_wakeup(backing_object);
1470                         }
1471                         /*
1472                          * Object now shadows whatever backing_object did.
1473                          * Note that the reference to 
1474                          * backing_object->backing_object moves from within 
1475                          * backing_object to within object.
1476                          */
1477
1478                         LIST_REMOVE(object, shadow_list);
1479                         object->backing_object->shadow_count--;
1480                         object->backing_object->generation++;
1481                         if (backing_object->backing_object) {
1482                                 LIST_REMOVE(backing_object, shadow_list);
1483                                 backing_object->backing_object->shadow_count--;
1484                                 backing_object->backing_object->generation++;
1485                         }
1486                         object->backing_object = backing_object->backing_object;
1487                         if (object->backing_object) {
1488                                 LIST_INSERT_HEAD(
1489                                     &object->backing_object->shadow_head,
1490                                     object, 
1491                                     shadow_list
1492                                 );
1493                                 object->backing_object->shadow_count++;
1494                                 object->backing_object->generation++;
1495                         }
1496
1497                         object->backing_object_offset +=
1498                             backing_object->backing_object_offset;
1499
1500                         /*
1501                          * Discard backing_object.
1502                          *
1503                          * Since the backing object has no pages, no pager left,
1504                          * and no object references within it, all that is
1505                          * necessary is to dispose of it.
1506                          */
1507
1508                         KASSERT(backing_object->ref_count == 1,
1509                                 ("backing_object %p was somehow "
1510                                  "re-referenced during collapse!",
1511                                  backing_object));
1512                         KASSERT(RB_EMPTY(&backing_object->rb_memq),
1513                                 ("backing_object %p somehow has left "
1514                                  "over pages during collapse!",
1515                                  backing_object));
1516
1517                         /* (we are holding vmobj_token) */
1518                         TAILQ_REMOVE(&vm_object_list, backing_object,
1519                                      object_list);
1520                         vm_object_count--;
1521
1522                         zfree(obj_zone, backing_object);
1523
1524                         object_collapses++;
1525                 } else {
1526                         vm_object_t new_backing_object;
1527
1528                         /*
1529                          * If we do not entirely shadow the backing object,
1530                          * there is nothing we can do so we give up.
1531                          */
1532
1533                         if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1534                                 break;
1535                         }
1536
1537                         /*
1538                          * Make the parent shadow the next object in the
1539                          * chain.  Deallocating backing_object will not remove
1540                          * it, since its reference count is at least 2.
1541                          */
1542
1543                         LIST_REMOVE(object, shadow_list);
1544                         backing_object->shadow_count--;
1545                         backing_object->generation++;
1546
1547                         new_backing_object = backing_object->backing_object;
1548                         if ((object->backing_object = new_backing_object) != NULL) {
1549                                 vm_object_reference(new_backing_object);
1550                                 LIST_INSERT_HEAD(
1551                                     &new_backing_object->shadow_head,
1552                                     object,
1553                                     shadow_list
1554                                 );
1555                                 new_backing_object->shadow_count++;
1556                                 new_backing_object->generation++;
1557                                 object->backing_object_offset +=
1558                                         backing_object->backing_object_offset;
1559                         }
1560
1561                         /*
1562                          * Drop the reference count on backing_object. Since
1563                          * its ref_count was at least 2, it will not vanish;
1564                          * so we don't need to call vm_object_deallocate, but
1565                          * we do anyway.
1566                          */
1567                         vm_object_deallocate_locked(backing_object);
1568                         object_bypasses++;
1569                 }
1570
1571                 /*
1572                  * Try again with this object's new backing object.
1573                  */
1574         }
1575 }
1576
1577 /*
1578  * Removes all physical pages in the specified object range from the
1579  * object's list of pages.
1580  *
1581  * No requirements.
1582  */
1583 static int vm_object_page_remove_callback(vm_page_t p, void *data);
1584
1585 void
1586 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1587                       boolean_t clean_only)
1588 {
1589         struct rb_vm_page_scan_info info;
1590         int all;
1591
1592         /*
1593          * Degenerate cases and assertions
1594          */
1595         lwkt_gettoken(&vm_token);
1596         if (object == NULL ||
1597             (object->resident_page_count == 0 && object->swblock_count == 0)) {
1598                 lwkt_reltoken(&vm_token);
1599                 return;
1600         }
1601         KASSERT(object->type != OBJT_PHYS, 
1602                 ("attempt to remove pages from a physical object"));
1603
1604         /*
1605          * Indicate that paging is occuring on the object
1606          */
1607         crit_enter();
1608         vm_object_pip_add(object, 1);
1609
1610         /*
1611          * Figure out the actual removal range and whether we are removing
1612          * the entire contents of the object or not.  If removing the entire
1613          * contents, be sure to get all pages, even those that might be 
1614          * beyond the end of the object.
1615          */
1616         info.start_pindex = start;
1617         if (end == 0)
1618                 info.end_pindex = (vm_pindex_t)-1;
1619         else
1620                 info.end_pindex = end - 1;
1621         info.limit = clean_only;
1622         all = (start == 0 && info.end_pindex >= object->size - 1);
1623
1624         /*
1625          * Loop until we are sure we have gotten them all.
1626          */
1627         do {
1628                 info.error = 0;
1629                 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1630                                         vm_object_page_remove_callback, &info);
1631         } while (info.error);
1632
1633         /*
1634          * Remove any related swap if throwing away pages, or for
1635          * non-swap objects (the swap is a clean copy in that case).
1636          */
1637         if (object->type != OBJT_SWAP || clean_only == FALSE) {
1638                 if (all)
1639                         swap_pager_freespace_all(object);
1640                 else
1641                         swap_pager_freespace(object, info.start_pindex,
1642                              info.end_pindex - info.start_pindex + 1);
1643         }
1644
1645         /*
1646          * Cleanup
1647          */
1648         vm_object_pip_wakeup(object);
1649         crit_exit();
1650         lwkt_reltoken(&vm_token);
1651 }
1652
1653 /*
1654  * The caller must hold vm_token.
1655  */
1656 static int
1657 vm_object_page_remove_callback(vm_page_t p, void *data)
1658 {
1659         struct rb_vm_page_scan_info *info = data;
1660
1661         /*
1662          * Wired pages cannot be destroyed, but they can be invalidated
1663          * and we do so if clean_only (limit) is not set.
1664          *
1665          * WARNING!  The page may be wired due to being part of a buffer
1666          *           cache buffer, and the buffer might be marked B_CACHE.
1667          *           This is fine as part of a truncation but VFSs must be
1668          *           sure to fix the buffer up when re-extending the file.
1669          */
1670         if (p->wire_count != 0) {
1671                 vm_page_protect(p, VM_PROT_NONE);
1672                 if (info->limit == 0)
1673                         p->valid = 0;
1674                 return(0);
1675         }
1676
1677         /*
1678          * The busy flags are only cleared at
1679          * interrupt -- minimize the spl transitions
1680          */
1681
1682         if (vm_page_sleep_busy(p, TRUE, "vmopar")) {
1683                 info->error = 1;
1684                 return(0);
1685         }
1686
1687         /*
1688          * limit is our clean_only flag.  If set and the page is dirty, do
1689          * not free it.  If set and the page is being held by someone, do
1690          * not free it.
1691          */
1692         if (info->limit && p->valid) {
1693                 vm_page_test_dirty(p);
1694                 if (p->valid & p->dirty)
1695                         return(0);
1696                 if (p->hold_count)
1697                         return(0);
1698         }
1699
1700         /*
1701          * Destroy the page
1702          */
1703         vm_page_busy(p);
1704         vm_page_protect(p, VM_PROT_NONE);
1705         vm_page_free(p);
1706         return(0);
1707 }
1708
1709 /*
1710  * Coalesces two objects backing up adjoining regions of memory into a
1711  * single object.
1712  *
1713  * returns TRUE if objects were combined.
1714  *
1715  * NOTE: Only works at the moment if the second object is NULL -
1716  *       if it's not, which object do we lock first?
1717  *
1718  * Parameters:
1719  *      prev_object     First object to coalesce
1720  *      prev_offset     Offset into prev_object
1721  *      next_object     Second object into coalesce
1722  *      next_offset     Offset into next_object
1723  *
1724  *      prev_size       Size of reference to prev_object
1725  *      next_size       Size of reference to next_object
1726  *
1727  * The object must not be locked.
1728  * The caller must hold vm_token and vmobj_token.
1729  */
1730 boolean_t
1731 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
1732                    vm_size_t prev_size, vm_size_t next_size)
1733 {
1734         vm_pindex_t next_pindex;
1735
1736         ASSERT_LWKT_TOKEN_HELD(&vm_token);
1737         ASSERT_LWKT_TOKEN_HELD(&vmobj_token);
1738
1739         if (prev_object == NULL) {
1740                 return (TRUE);
1741         }
1742
1743         vm_object_lock(prev_object);
1744         if (prev_object->type != OBJT_DEFAULT &&
1745             prev_object->type != OBJT_SWAP) {
1746                 vm_object_unlock(prev_object);
1747                 return (FALSE);
1748         }
1749
1750         /*
1751          * Try to collapse the object first
1752          */
1753         vm_object_collapse(prev_object);
1754
1755         /*
1756          * Can't coalesce if: . more than one reference . paged out . shadows
1757          * another object . has a copy elsewhere (any of which mean that the
1758          * pages not mapped to prev_entry may be in use anyway)
1759          */
1760
1761         if (prev_object->backing_object != NULL) {
1762                 vm_object_unlock(prev_object);
1763                 return (FALSE);
1764         }
1765
1766         prev_size >>= PAGE_SHIFT;
1767         next_size >>= PAGE_SHIFT;
1768         next_pindex = prev_pindex + prev_size;
1769
1770         if ((prev_object->ref_count > 1) &&
1771             (prev_object->size != next_pindex)) {
1772                 vm_object_unlock(prev_object);
1773                 return (FALSE);
1774         }
1775
1776         /*
1777          * Remove any pages that may still be in the object from a previous
1778          * deallocation.
1779          */
1780         if (next_pindex < prev_object->size) {
1781                 vm_object_page_remove(prev_object,
1782                                       next_pindex,
1783                                       next_pindex + next_size, FALSE);
1784                 if (prev_object->type == OBJT_SWAP)
1785                         swap_pager_freespace(prev_object,
1786                                              next_pindex, next_size);
1787         }
1788
1789         /*
1790          * Extend the object if necessary.
1791          */
1792         if (next_pindex + next_size > prev_object->size)
1793                 prev_object->size = next_pindex + next_size;
1794
1795         vm_object_unlock(prev_object);
1796         return (TRUE);
1797 }
1798
1799 /*
1800  * Make the object writable and flag is being possibly dirty.
1801  *
1802  * No requirements.
1803  */
1804 void
1805 vm_object_set_writeable_dirty(vm_object_t object)
1806 {
1807         struct vnode *vp;
1808
1809         lwkt_gettoken(&vm_token);
1810         vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1811         if (object->type == OBJT_VNODE &&
1812             (vp = (struct vnode *)object->handle) != NULL) {
1813                 if ((vp->v_flag & VOBJDIRTY) == 0) {
1814                         vsetflags(vp, VOBJDIRTY);
1815                 }
1816         }
1817         lwkt_reltoken(&vm_token);
1818 }
1819
1820 void
1821 vm_object_lock(vm_object_t object)
1822 {
1823         lwkt_gettoken(&object->tok);
1824 }
1825
1826 void
1827 vm_object_unlock(vm_object_t object)
1828 {
1829         lwkt_reltoken(&object->tok);
1830 }
1831
1832 #include "opt_ddb.h"
1833 #ifdef DDB
1834 #include <sys/kernel.h>
1835
1836 #include <sys/cons.h>
1837
1838 #include <ddb/ddb.h>
1839
1840 static int      _vm_object_in_map (vm_map_t map, vm_object_t object,
1841                                        vm_map_entry_t entry);
1842 static int      vm_object_in_map (vm_object_t object);
1843
1844 /*
1845  * The caller must hold vm_token.
1846  */
1847 static int
1848 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
1849 {
1850         vm_map_t tmpm;
1851         vm_map_entry_t tmpe;
1852         vm_object_t obj;
1853         int entcount;
1854
1855         if (map == 0)
1856                 return 0;
1857         if (entry == 0) {
1858                 tmpe = map->header.next;
1859                 entcount = map->nentries;
1860                 while (entcount-- && (tmpe != &map->header)) {
1861                         if( _vm_object_in_map(map, object, tmpe)) {
1862                                 return 1;
1863                         }
1864                         tmpe = tmpe->next;
1865                 }
1866                 return (0);
1867         }
1868         switch(entry->maptype) {
1869         case VM_MAPTYPE_SUBMAP:
1870                 tmpm = entry->object.sub_map;
1871                 tmpe = tmpm->header.next;
1872                 entcount = tmpm->nentries;
1873                 while (entcount-- && tmpe != &tmpm->header) {
1874                         if( _vm_object_in_map(tmpm, object, tmpe)) {
1875                                 return 1;
1876                         }
1877                         tmpe = tmpe->next;
1878                 }
1879                 break;
1880         case VM_MAPTYPE_NORMAL:
1881         case VM_MAPTYPE_VPAGETABLE:
1882                 obj = entry->object.vm_object;
1883                 while (obj) {
1884                         if (obj == object)
1885                                 return 1;
1886                         obj = obj->backing_object;
1887                 }
1888                 break;
1889         default:
1890                 break;
1891         }
1892         return 0;
1893 }
1894
1895 static int vm_object_in_map_callback(struct proc *p, void *data);
1896
1897 struct vm_object_in_map_info {
1898         vm_object_t object;
1899         int rv;
1900 };
1901
1902 /*
1903  * Debugging only
1904  */
1905 static int
1906 vm_object_in_map(vm_object_t object)
1907 {
1908         struct vm_object_in_map_info info;
1909
1910         info.rv = 0;
1911         info.object = object;
1912
1913         allproc_scan(vm_object_in_map_callback, &info);
1914         if (info.rv)
1915                 return 1;
1916         if( _vm_object_in_map(&kernel_map, object, 0))
1917                 return 1;
1918         if( _vm_object_in_map(&pager_map, object, 0))
1919                 return 1;
1920         if( _vm_object_in_map(&buffer_map, object, 0))
1921                 return 1;
1922         return 0;
1923 }
1924
1925 /*
1926  * Debugging only
1927  */
1928 static int
1929 vm_object_in_map_callback(struct proc *p, void *data)
1930 {
1931         struct vm_object_in_map_info *info = data;
1932
1933         if (p->p_vmspace) {
1934                 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) {
1935                         info->rv = 1;
1936                         return -1;
1937                 }
1938         }
1939         return (0);
1940 }
1941
1942 DB_SHOW_COMMAND(vmochk, vm_object_check)
1943 {
1944         vm_object_t object;
1945
1946         /*
1947          * make sure that internal objs are in a map somewhere
1948          * and none have zero ref counts.
1949          */
1950         for (object = TAILQ_FIRST(&vm_object_list);
1951                         object != NULL;
1952                         object = TAILQ_NEXT(object, object_list)) {
1953                 if (object->type == OBJT_MARKER)
1954                         continue;
1955                 if (object->handle == NULL &&
1956                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1957                         if (object->ref_count == 0) {
1958                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
1959                                         (long)object->size);
1960                         }
1961                         if (!vm_object_in_map(object)) {
1962                                 db_printf(
1963                         "vmochk: internal obj is not in a map: "
1964                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1965                                     object->ref_count, (u_long)object->size, 
1966                                     (u_long)object->size,
1967                                     (void *)object->backing_object);
1968                         }
1969                 }
1970         }
1971 }
1972
1973 /*
1974  * Debugging only
1975  */
1976 DB_SHOW_COMMAND(object, vm_object_print_static)
1977 {
1978         /* XXX convert args. */
1979         vm_object_t object = (vm_object_t)addr;
1980         boolean_t full = have_addr;
1981
1982         vm_page_t p;
1983
1984         /* XXX count is an (unused) arg.  Avoid shadowing it. */
1985 #define count   was_count
1986
1987         int count;
1988
1989         if (object == NULL)
1990                 return;
1991
1992         db_iprintf(
1993             "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1994             object, (int)object->type, (u_long)object->size,
1995             object->resident_page_count, object->ref_count, object->flags);
1996         /*
1997          * XXX no %qd in kernel.  Truncate object->backing_object_offset.
1998          */
1999         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
2000             object->shadow_count, 
2001             object->backing_object ? object->backing_object->ref_count : 0,
2002             object->backing_object, (long)object->backing_object_offset);
2003
2004         if (!full)
2005                 return;
2006
2007         db_indent += 2;
2008         count = 0;
2009         RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) {
2010                 if (count == 0)
2011                         db_iprintf("memory:=");
2012                 else if (count == 6) {
2013                         db_printf("\n");
2014                         db_iprintf(" ...");
2015                         count = 0;
2016                 } else
2017                         db_printf(",");
2018                 count++;
2019
2020                 db_printf("(off=0x%lx,page=0x%lx)",
2021                     (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
2022         }
2023         if (count != 0)
2024                 db_printf("\n");
2025         db_indent -= 2;
2026 }
2027
2028 /* XXX. */
2029 #undef count
2030
2031 /*
2032  * XXX need this non-static entry for calling from vm_map_print.
2033  *
2034  * Debugging only
2035  */
2036 void
2037 vm_object_print(/* db_expr_t */ long addr,
2038                 boolean_t have_addr,
2039                 /* db_expr_t */ long count,
2040                 char *modif)
2041 {
2042         vm_object_print_static(addr, have_addr, count, modif);
2043 }
2044
2045 /*
2046  * Debugging only
2047  */
2048 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2049 {
2050         vm_object_t object;
2051         int nl = 0;
2052         int c;
2053         for (object = TAILQ_FIRST(&vm_object_list);
2054                         object != NULL;
2055                         object = TAILQ_NEXT(object, object_list)) {
2056                 vm_pindex_t idx, fidx;
2057                 vm_pindex_t osize;
2058                 vm_paddr_t pa = -1, padiff;
2059                 int rcount;
2060                 vm_page_t m;
2061
2062                 if (object->type == OBJT_MARKER)
2063                         continue;
2064                 db_printf("new object: %p\n", (void *)object);
2065                 if ( nl > 18) {
2066                         c = cngetc();
2067                         if (c != ' ')
2068                                 return;
2069                         nl = 0;
2070                 }
2071                 nl++;
2072                 rcount = 0;
2073                 fidx = 0;
2074                 osize = object->size;
2075                 if (osize > 128)
2076                         osize = 128;
2077                 for (idx = 0; idx < osize; idx++) {
2078                         m = vm_page_lookup(object, idx);
2079                         if (m == NULL) {
2080                                 if (rcount) {
2081                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2082                                                 (long)fidx, rcount, (long)pa);
2083                                         if ( nl > 18) {
2084                                                 c = cngetc();
2085                                                 if (c != ' ')
2086                                                         return;
2087                                                 nl = 0;
2088                                         }
2089                                         nl++;
2090                                         rcount = 0;
2091                                 }
2092                                 continue;
2093                         }
2094
2095                                 
2096                         if (rcount &&
2097                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2098                                 ++rcount;
2099                                 continue;
2100                         }
2101                         if (rcount) {
2102                                 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
2103                                 padiff >>= PAGE_SHIFT;
2104                                 padiff &= PQ_L2_MASK;
2105                                 if (padiff == 0) {
2106                                         pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
2107                                         ++rcount;
2108                                         continue;
2109                                 }
2110                                 db_printf(" index(%ld)run(%d)pa(0x%lx)",
2111                                         (long)fidx, rcount, (long)pa);
2112                                 db_printf("pd(%ld)\n", (long)padiff);
2113                                 if ( nl > 18) {
2114                                         c = cngetc();
2115                                         if (c != ' ')
2116                                                 return;
2117                                         nl = 0;
2118                                 }
2119                                 nl++;
2120                         }
2121                         fidx = idx;
2122                         pa = VM_PAGE_TO_PHYS(m);
2123                         rcount = 1;
2124                 }
2125                 if (rcount) {
2126                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2127                                 (long)fidx, rcount, (long)pa);
2128                         if ( nl > 18) {
2129                                 c = cngetc();
2130                                 if (c != ' ')
2131                                         return;
2132                                 nl = 0;
2133                         }
2134                         nl++;
2135                 }
2136         }
2137 }
2138 #endif /* DDB */