proc->thread stage 4: rework the VFS and DEVICE subsystems to take thread
[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 1.3 2003/06/25 03:56:13 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 #define MSYNC_FLUSH_HARDSEQ     0x01
98 #define MSYNC_FLUSH_SOFTSEQ     0x02
99
100 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ;
101 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags,
102         CTLFLAG_RW, &msync_flush_flags, 0, "");
103
104 static void     vm_object_qcollapse (vm_object_t object);
105 static int      vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
106
107 /*
108  *      Virtual memory objects maintain the actual data
109  *      associated with allocated virtual memory.  A given
110  *      page of memory exists within exactly one object.
111  *
112  *      An object is only deallocated when all "references"
113  *      are given up.  Only one "reference" to a given
114  *      region of an object should be writeable.
115  *
116  *      Associated with each object is a list of all resident
117  *      memory pages belonging to that object; this list is
118  *      maintained by the "vm_page" module, and locked by the object's
119  *      lock.
120  *
121  *      Each object also records a "pager" routine which is
122  *      used to retrieve (and store) pages to the proper backing
123  *      storage.  In addition, objects may be backed by other
124  *      objects from which they were virtual-copied.
125  *
126  *      The only items within the object structure which are
127  *      modified after time of creation are:
128  *              reference count         locked by object's lock
129  *              pager routine           locked by object's lock
130  *
131  */
132
133 struct object_q vm_object_list;
134 #ifndef NULL_SIMPLELOCKS
135 static struct simplelock vm_object_list_lock;
136 #endif
137 static long vm_object_count;            /* count of all objects */
138 vm_object_t kernel_object;
139 vm_object_t kmem_object;
140 static struct vm_object kernel_object_store;
141 static struct vm_object kmem_object_store;
142 extern int vm_pageout_page_count;
143
144 static long object_collapses;
145 static long object_bypasses;
146 static int next_index;
147 static vm_zone_t obj_zone;
148 static struct vm_zone obj_zone_store;
149 static int object_hash_rand;
150 #define VM_OBJECTS_INIT 256
151 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
152
153 void
154 _vm_object_allocate(type, size, object)
155         objtype_t type;
156         vm_size_t size;
157         vm_object_t object;
158 {
159         int incr;
160         TAILQ_INIT(&object->memq);
161         LIST_INIT(&object->shadow_head);
162
163         object->type = type;
164         object->size = size;
165         object->ref_count = 1;
166         object->flags = 0;
167         if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
168                 vm_object_set_flag(object, OBJ_ONEMAPPING);
169         object->paging_in_progress = 0;
170         object->resident_page_count = 0;
171         object->shadow_count = 0;
172         object->pg_color = next_index;
173         if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
174                 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
175         else
176                 incr = size;
177         next_index = (next_index + incr) & PQ_L2_MASK;
178         object->handle = NULL;
179         object->backing_object = NULL;
180         object->backing_object_offset = (vm_ooffset_t) 0;
181         /*
182          * Try to generate a number that will spread objects out in the
183          * hash table.  We 'wipe' new objects across the hash in 128 page
184          * increments plus 1 more to offset it a little more by the time
185          * it wraps around.
186          */
187         object->hash_rand = object_hash_rand - 129;
188
189         object->generation++;
190
191         TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
192         vm_object_count++;
193         object_hash_rand = object->hash_rand;
194 }
195
196 /*
197  *      vm_object_init:
198  *
199  *      Initialize the VM objects module.
200  */
201 void
202 vm_object_init()
203 {
204         TAILQ_INIT(&vm_object_list);
205         simple_lock_init(&vm_object_list_lock);
206         vm_object_count = 0;
207         
208         kernel_object = &kernel_object_store;
209         _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
210             kernel_object);
211
212         kmem_object = &kmem_object_store;
213         _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
214             kmem_object);
215
216         obj_zone = &obj_zone_store;
217         zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
218                 vm_objects_init, VM_OBJECTS_INIT);
219 }
220
221 void
222 vm_object_init2() {
223         zinitna(obj_zone, NULL, NULL, 0, 0, 0, 1);
224 }
225
226 /*
227  *      vm_object_allocate:
228  *
229  *      Returns a new object with the given size.
230  */
231
232 vm_object_t
233 vm_object_allocate(type, size)
234         objtype_t type;
235         vm_size_t size;
236 {
237         vm_object_t result;
238
239         result = (vm_object_t) zalloc(obj_zone);
240
241         _vm_object_allocate(type, size, result);
242
243         return (result);
244 }
245
246
247 /*
248  *      vm_object_reference:
249  *
250  *      Gets another reference to the given object.
251  */
252 void
253 vm_object_reference(object)
254         vm_object_t object;
255 {
256         if (object == NULL)
257                 return;
258
259 #if 0
260         /* object can be re-referenced during final cleaning */
261         KASSERT(!(object->flags & OBJ_DEAD),
262             ("vm_object_reference: attempting to reference dead obj"));
263 #endif
264
265         object->ref_count++;
266         if (object->type == OBJT_VNODE) {
267                 while (vget((struct vnode *) object->handle, LK_RETRY|LK_NOOBJ, curthread)) {
268                         printf("vm_object_reference: delay in getting object\n");
269                 }
270         }
271 }
272
273 void
274 vm_object_vndeallocate(object)
275         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 #ifdef INVARIANTS
283         if (object->ref_count == 0) {
284                 vprint("vm_object_vndeallocate", vp);
285                 panic("vm_object_vndeallocate: bad object reference count");
286         }
287 #endif
288
289         object->ref_count--;
290         if (object->ref_count == 0) {
291                 vp->v_flag &= ~VTEXT;
292                 vm_object_clear_flag(object, OBJ_OPT);
293         }
294         vrele(vp);
295 }
296
297 /*
298  *      vm_object_deallocate:
299  *
300  *      Release a reference to the specified object,
301  *      gained either through a vm_object_allocate
302  *      or a vm_object_reference call.  When all references
303  *      are gone, storage associated with this object
304  *      may be relinquished.
305  *
306  *      No object may be locked.
307  */
308 void
309 vm_object_deallocate(object)
310         vm_object_t object;
311 {
312         vm_object_t temp;
313
314         while (object != NULL) {
315
316                 if (object->type == OBJT_VNODE) {
317                         vm_object_vndeallocate(object);
318                         return;
319                 }
320
321                 if (object->ref_count == 0) {
322                         panic("vm_object_deallocate: object deallocated too many times: %d", object->type);
323                 } else if (object->ref_count > 2) {
324                         object->ref_count--;
325                         return;
326                 }
327
328                 /*
329                  * Here on ref_count of one or two, which are special cases for
330                  * objects.
331                  */
332                 if ((object->ref_count == 2) && (object->shadow_count == 0)) {
333                         vm_object_set_flag(object, OBJ_ONEMAPPING);
334                         object->ref_count--;
335                         return;
336                 } else if ((object->ref_count == 2) && (object->shadow_count == 1)) {
337                         object->ref_count--;
338                         if ((object->handle == NULL) &&
339                             (object->type == OBJT_DEFAULT ||
340                              object->type == OBJT_SWAP)) {
341                                 vm_object_t robject;
342
343                                 robject = LIST_FIRST(&object->shadow_head);
344                                 KASSERT(robject != NULL,
345                                     ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
346                                          object->ref_count,
347                                          object->shadow_count));
348                                 if ((robject->handle == NULL) &&
349                                     (robject->type == OBJT_DEFAULT ||
350                                      robject->type == OBJT_SWAP)) {
351
352                                         robject->ref_count++;
353
354                                         while (
355                                                 robject->paging_in_progress ||
356                                                 object->paging_in_progress
357                                         ) {
358                                                 vm_object_pip_sleep(robject, "objde1");
359                                                 vm_object_pip_sleep(object, "objde2");
360                                         }
361
362                                         if (robject->ref_count == 1) {
363                                                 robject->ref_count--;
364                                                 object = robject;
365                                                 goto doterm;
366                                         }
367
368                                         object = robject;
369                                         vm_object_collapse(object);
370                                         continue;
371                                 }
372                         }
373
374                         return;
375
376                 } else {
377                         object->ref_count--;
378                         if (object->ref_count != 0)
379                                 return;
380                 }
381
382 doterm:
383
384                 temp = object->backing_object;
385                 if (temp) {
386                         LIST_REMOVE(object, shadow_list);
387                         temp->shadow_count--;
388                         if (temp->ref_count == 0)
389                                 vm_object_clear_flag(temp, OBJ_OPT);
390                         temp->generation++;
391                         object->backing_object = NULL;
392                 }
393
394                 /*
395                  * Don't double-terminate, we could be in a termination
396                  * recursion due to the terminate having to sync data
397                  * to disk.
398                  */
399                 if ((object->flags & OBJ_DEAD) == 0)
400                         vm_object_terminate(object);
401                 object = temp;
402         }
403 }
404
405 /*
406  *      vm_object_terminate actually destroys the specified object, freeing
407  *      up all previously used resources.
408  *
409  *      The object must be locked.
410  *      This routine may block.
411  */
412 void
413 vm_object_terminate(object)
414         vm_object_t object;
415 {
416         vm_page_t p;
417         int s;
418
419         /*
420          * Make sure no one uses us.
421          */
422         vm_object_set_flag(object, OBJ_DEAD);
423
424         /*
425          * wait for the pageout daemon to be done with the object
426          */
427         vm_object_pip_wait(object, "objtrm");
428
429         KASSERT(!object->paging_in_progress,
430                 ("vm_object_terminate: pageout in progress"));
431
432         /*
433          * Clean and free the pages, as appropriate. All references to the
434          * object are gone, so we don't need to lock it.
435          */
436         if (object->type == OBJT_VNODE) {
437                 struct vnode *vp;
438
439                 /*
440                  * Freeze optimized copies.
441                  */
442                 vm_freeze_copyopts(object, 0, object->size);
443
444                 /*
445                  * Clean pages and flush buffers.
446                  */
447                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
448
449                 vp = (struct vnode *) object->handle;
450                 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0);
451         }
452
453         /*
454          * Wait for any I/O to complete, after which there had better not
455          * be any references left on the object.
456          */
457         vm_object_pip_wait(object, "objtrm");
458
459         if (object->ref_count != 0)
460                 panic("vm_object_terminate: object with references, ref_count=%d", object->ref_count);
461
462         /*
463          * Now free any remaining pages. For internal objects, this also
464          * removes them from paging queues. Don't free wired pages, just
465          * remove them from the object. 
466          */
467         s = splvm();
468         while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
469                 if (p->busy || (p->flags & PG_BUSY))
470                         panic("vm_object_terminate: freeing busy page %p\n", p);
471                 if (p->wire_count == 0) {
472                         vm_page_busy(p);
473                         vm_page_free(p);
474                         cnt.v_pfree++;
475                 } else {
476                         vm_page_busy(p);
477                         vm_page_remove(p);
478                 }
479         }
480         splx(s);
481
482         /*
483          * Let the pager know object is dead.
484          */
485         vm_pager_deallocate(object);
486
487         /*
488          * Remove the object from the global object list.
489          */
490         simple_lock(&vm_object_list_lock);
491         TAILQ_REMOVE(&vm_object_list, object, object_list);
492         simple_unlock(&vm_object_list_lock);
493
494         wakeup(object);
495
496         /*
497          * Free the space for the object.
498          */
499         zfree(obj_zone, object);
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  *      The object must be locked.
516  */
517
518 void
519 vm_object_page_clean(object, start, end, flags)
520         vm_object_t object;
521         vm_pindex_t start;
522         vm_pindex_t end;
523         int flags;
524 {
525         vm_page_t p, np;
526         vm_offset_t tstart, tend;
527         vm_pindex_t pi;
528         struct vnode *vp;
529         int clearobjflags;
530         int pagerflags;
531         int curgeneration;
532
533         if (object->type != OBJT_VNODE ||
534                 (object->flags & OBJ_MIGHTBEDIRTY) == 0)
535                 return;
536
537         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
538         pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
539
540         vp = object->handle;
541
542         vm_object_set_flag(object, OBJ_CLEANING);
543
544         /*
545          * Handle 'entire object' case
546          */
547         tstart = start;
548         if (end == 0) {
549                 tend = object->size;
550         } else {
551                 tend = end;
552         }
553
554         /*
555          * If the caller is smart and only msync()s a range he knows is
556          * dirty, we may be able to avoid an object scan.  This results in
557          * a phenominal improvement in performance.  We cannot do this
558          * as a matter of course because the object may be huge - e.g.
559          * the size might be in the gigabytes or terrabytes.
560          */
561         if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
562                 vm_offset_t tscan;
563                 int scanlimit;
564                 int scanreset;
565
566                 scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
567                 if (scanreset < 16)
568                         scanreset = 16;
569                 pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
570
571                 scanlimit = scanreset;
572                 tscan = tstart;
573                 while (tscan < tend) {
574                         curgeneration = object->generation;
575                         p = vm_page_lookup(object, tscan);
576                         if (p == NULL || p->valid == 0 ||
577                             (p->queue - p->pc) == PQ_CACHE) {
578                                 if (--scanlimit == 0)
579                                         break;
580                                 ++tscan;
581                                 continue;
582                         }
583                         vm_page_test_dirty(p);
584                         if ((p->dirty & p->valid) == 0) {
585                                 if (--scanlimit == 0)
586                                         break;
587                                 ++tscan;
588                                 continue;
589                         }
590                         /*
591                          * If we have been asked to skip nosync pages and 
592                          * this is a nosync page, we can't continue.
593                          */
594                         if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
595                                 if (--scanlimit == 0)
596                                         break;
597                                 ++tscan;
598                                 continue;
599                         }
600                         scanlimit = scanreset;
601
602                         /*
603                          * This returns 0 if it was unable to busy the first
604                          * page (i.e. had to sleep).
605                          */
606                         tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
607                 }
608
609                 /*
610                  * If everything was dirty and we flushed it successfully,
611                  * and the requested range is not the entire object, we
612                  * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
613                  * return immediately.
614                  */
615                 if (tscan >= tend && (tstart || tend < object->size)) {
616                         vm_object_clear_flag(object, OBJ_CLEANING);
617                         return;
618                 }
619                 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
620         }
621
622         /*
623          * Generally set CLEANCHK interlock and make the page read-only so
624          * we can then clear the object flags.
625          *
626          * However, if this is a nosync mmap then the object is likely to 
627          * stay dirty so do not mess with the page and do not clear the
628          * object flags.
629          */
630
631         clearobjflags = 1;
632
633         for(p = TAILQ_FIRST(&object->memq); p; p = TAILQ_NEXT(p, listq)) {
634                 vm_page_flag_set(p, PG_CLEANCHK);
635                 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
636                         clearobjflags = 0;
637                 else
638                         vm_page_protect(p, VM_PROT_READ);
639         }
640
641         if (clearobjflags && (tstart == 0) && (tend == object->size)) {
642                 struct vnode *vp;
643
644                 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
645                 if (object->type == OBJT_VNODE &&
646                     (vp = (struct vnode *)object->handle) != NULL) {
647                         if (vp->v_flag & VOBJDIRTY) {
648                                 simple_lock(&vp->v_interlock);
649                                 vp->v_flag &= ~VOBJDIRTY;
650                                 simple_unlock(&vp->v_interlock);
651                         }
652                 }
653         }
654
655 rescan:
656         curgeneration = object->generation;
657
658         for(p = TAILQ_FIRST(&object->memq); p; p = np) {
659                 int n;
660
661                 np = TAILQ_NEXT(p, listq);
662
663 again:
664                 pi = p->pindex;
665                 if (((p->flags & PG_CLEANCHK) == 0) ||
666                         (pi < tstart) || (pi >= tend) ||
667                         (p->valid == 0) ||
668                         ((p->queue - p->pc) == PQ_CACHE)) {
669                         vm_page_flag_clear(p, PG_CLEANCHK);
670                         continue;
671                 }
672
673                 vm_page_test_dirty(p);
674                 if ((p->dirty & p->valid) == 0) {
675                         vm_page_flag_clear(p, PG_CLEANCHK);
676                         continue;
677                 }
678
679                 /*
680                  * If we have been asked to skip nosync pages and this is a
681                  * nosync page, skip it.  Note that the object flags were
682                  * not cleared in this case so we do not have to set them.
683                  */
684                 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
685                         vm_page_flag_clear(p, PG_CLEANCHK);
686                         continue;
687                 }
688
689                 n = vm_object_page_collect_flush(object, p,
690                         curgeneration, pagerflags);
691                 if (n == 0)
692                         goto rescan;
693                 if (object->generation != curgeneration)
694                         goto rescan;
695
696                 /*
697                  * Try to optimize the next page.  If we can't we pick up
698                  * our (random) scan where we left off.
699                  */
700                 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
701                         if ((p = vm_page_lookup(object, pi + n)) != NULL)
702                                 goto again;
703                 }
704         }
705
706 #if 0
707         VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
708 #endif
709
710         vm_object_clear_flag(object, OBJ_CLEANING);
711         return;
712 }
713
714 static int
715 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
716 {
717         int runlen;
718         int s;
719         int maxf;
720         int chkb;
721         int maxb;
722         int i;
723         vm_pindex_t pi;
724         vm_page_t maf[vm_pageout_page_count];
725         vm_page_t mab[vm_pageout_page_count];
726         vm_page_t ma[vm_pageout_page_count];
727
728         s = splvm();
729         pi = p->pindex;
730         while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
731                 if (object->generation != curgeneration) {
732                         splx(s);
733                         return(0);
734                 }
735         }
736
737         maxf = 0;
738         for(i = 1; i < vm_pageout_page_count; i++) {
739                 vm_page_t tp;
740
741                 if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
742                         if ((tp->flags & PG_BUSY) ||
743                                 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 
744                                  (tp->flags & PG_CLEANCHK) == 0) ||
745                                 (tp->busy != 0))
746                                 break;
747                         if((tp->queue - tp->pc) == PQ_CACHE) {
748                                 vm_page_flag_clear(tp, PG_CLEANCHK);
749                                 break;
750                         }
751                         vm_page_test_dirty(tp);
752                         if ((tp->dirty & tp->valid) == 0) {
753                                 vm_page_flag_clear(tp, PG_CLEANCHK);
754                                 break;
755                         }
756                         maf[ i - 1 ] = tp;
757                         maxf++;
758                         continue;
759                 }
760                 break;
761         }
762
763         maxb = 0;
764         chkb = vm_pageout_page_count -  maxf;
765         if (chkb) {
766                 for(i = 1; i < chkb;i++) {
767                         vm_page_t tp;
768
769                         if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
770                                 if ((tp->flags & PG_BUSY) ||
771                                         ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 
772                                          (tp->flags & PG_CLEANCHK) == 0) ||
773                                         (tp->busy != 0))
774                                         break;
775                                 if((tp->queue - tp->pc) == PQ_CACHE) {
776                                         vm_page_flag_clear(tp, PG_CLEANCHK);
777                                         break;
778                                 }
779                                 vm_page_test_dirty(tp);
780                                 if ((tp->dirty & tp->valid) == 0) {
781                                         vm_page_flag_clear(tp, PG_CLEANCHK);
782                                         break;
783                                 }
784                                 mab[ i - 1 ] = tp;
785                                 maxb++;
786                                 continue;
787                         }
788                         break;
789                 }
790         }
791
792         for(i = 0; i < maxb; i++) {
793                 int index = (maxb - i) - 1;
794                 ma[index] = mab[i];
795                 vm_page_flag_clear(ma[index], PG_CLEANCHK);
796         }
797         vm_page_flag_clear(p, PG_CLEANCHK);
798         ma[maxb] = p;
799         for(i = 0; i < maxf; i++) {
800                 int index = (maxb + i) + 1;
801                 ma[index] = maf[i];
802                 vm_page_flag_clear(ma[index], PG_CLEANCHK);
803         }
804         runlen = maxb + maxf + 1;
805
806         splx(s);
807         vm_pageout_flush(ma, runlen, pagerflags);
808         for (i = 0; i < runlen; i++) {
809                 if (ma[i]->valid & ma[i]->dirty) {
810                         vm_page_protect(ma[i], VM_PROT_READ);
811                         vm_page_flag_set(ma[i], PG_CLEANCHK);
812
813                         /*
814                          * maxf will end up being the actual number of pages
815                          * we wrote out contiguously, non-inclusive of the
816                          * first page.  We do not count look-behind pages.
817                          */
818                         if (i >= maxb + 1 && (maxf > i - maxb - 1))
819                                 maxf = i - maxb - 1;
820                 }
821         }
822         return(maxf + 1);
823 }
824
825 #ifdef not_used
826 /* XXX I cannot tell if this should be an exported symbol */
827 /*
828  *      vm_object_deactivate_pages
829  *
830  *      Deactivate all pages in the specified object.  (Keep its pages
831  *      in memory even though it is no longer referenced.)
832  *
833  *      The object must be locked.
834  */
835 static void
836 vm_object_deactivate_pages(object)
837         vm_object_t object;
838 {
839         vm_page_t p, next;
840
841         for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
842                 next = TAILQ_NEXT(p, listq);
843                 vm_page_deactivate(p);
844         }
845 }
846 #endif
847
848 /*
849  * Same as vm_object_pmap_copy, except range checking really
850  * works, and is meant for small sections of an object.
851  *
852  * This code protects resident pages by making them read-only
853  * and is typically called on a fork or split when a page
854  * is converted to copy-on-write.  
855  *
856  * NOTE: If the page is already at VM_PROT_NONE, calling
857  * vm_page_protect will have no effect.
858  */
859
860 void
861 vm_object_pmap_copy_1(object, start, end)
862         vm_object_t object;
863         vm_pindex_t start;
864         vm_pindex_t end;
865 {
866         vm_pindex_t idx;
867         vm_page_t p;
868
869         if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
870                 return;
871
872         for (idx = start; idx < end; idx++) {
873                 p = vm_page_lookup(object, idx);
874                 if (p == NULL)
875                         continue;
876                 vm_page_protect(p, VM_PROT_READ);
877         }
878 }
879
880 /*
881  *      vm_object_pmap_remove:
882  *
883  *      Removes all physical pages in the specified
884  *      object range from all physical maps.
885  *
886  *      The object must *not* be locked.
887  */
888 void
889 vm_object_pmap_remove(object, start, end)
890         vm_object_t object;
891         vm_pindex_t start;
892         vm_pindex_t end;
893 {
894         vm_page_t p;
895
896         if (object == NULL)
897                 return;
898         for (p = TAILQ_FIRST(&object->memq);
899                 p != NULL;
900                 p = TAILQ_NEXT(p, listq)) {
901                 if (p->pindex >= start && p->pindex < end)
902                         vm_page_protect(p, VM_PROT_NONE);
903         }
904         if ((start == 0) && (object->size == end))
905                 vm_object_clear_flag(object, OBJ_WRITEABLE);
906 }
907
908 /*
909  *      vm_object_madvise:
910  *
911  *      Implements the madvise function at the object/page level.
912  *
913  *      MADV_WILLNEED   (any object)
914  *
915  *          Activate the specified pages if they are resident.
916  *
917  *      MADV_DONTNEED   (any object)
918  *
919  *          Deactivate the specified pages if they are resident.
920  *
921  *      MADV_FREE       (OBJT_DEFAULT/OBJT_SWAP objects,
922  *                       OBJ_ONEMAPPING only)
923  *
924  *          Deactivate and clean the specified pages if they are
925  *          resident.  This permits the process to reuse the pages
926  *          without faulting or the kernel to reclaim the pages
927  *          without I/O.
928  */
929 void
930 vm_object_madvise(object, pindex, count, advise)
931         vm_object_t object;
932         vm_pindex_t pindex;
933         int count;
934         int advise;
935 {
936         vm_pindex_t end, tpindex;
937         vm_object_t tobject;
938         vm_page_t m;
939
940         if (object == NULL)
941                 return;
942
943         end = pindex + count;
944
945         /*
946          * Locate and adjust resident pages
947          */
948
949         for (; pindex < end; pindex += 1) {
950 relookup:
951                 tobject = object;
952                 tpindex = pindex;
953 shadowlookup:
954                 /*
955                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
956                  * and those pages must be OBJ_ONEMAPPING.
957                  */
958                 if (advise == MADV_FREE) {
959                         if ((tobject->type != OBJT_DEFAULT &&
960                              tobject->type != OBJT_SWAP) ||
961                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
962                                 continue;
963                         }
964                 }
965
966                 m = vm_page_lookup(tobject, tpindex);
967
968                 if (m == NULL) {
969                         /*
970                          * There may be swap even if there is no backing page
971                          */
972                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
973                                 swap_pager_freespace(tobject, tpindex, 1);
974
975                         /*
976                          * next object
977                          */
978                         tobject = tobject->backing_object;
979                         if (tobject == NULL)
980                                 continue;
981                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
982                         goto shadowlookup;
983                 }
984
985                 /*
986                  * If the page is busy or not in a normal active state,
987                  * we skip it.  If the page is not managed there are no
988                  * page queues to mess with.  Things can break if we mess
989                  * with pages in any of the below states.
990                  */
991                 if (
992                     m->hold_count ||
993                     m->wire_count ||
994                     (m->flags & PG_UNMANAGED) ||
995                     m->valid != VM_PAGE_BITS_ALL
996                 ) {
997                         continue;
998                 }
999
1000                 if (vm_page_sleep_busy(m, TRUE, "madvpo"))
1001                         goto relookup;
1002
1003                 if (advise == MADV_WILLNEED) {
1004                         vm_page_activate(m);
1005                 } else if (advise == MADV_DONTNEED) {
1006                         vm_page_dontneed(m);
1007                 } else if (advise == MADV_FREE) {
1008                         /*
1009                          * Mark the page clean.  This will allow the page
1010                          * to be freed up by the system.  However, such pages
1011                          * are often reused quickly by malloc()/free()
1012                          * so we do not do anything that would cause
1013                          * a page fault if we can help it.
1014                          *
1015                          * Specifically, we do not try to actually free
1016                          * the page now nor do we try to put it in the
1017                          * cache (which would cause a page fault on reuse).
1018                          *
1019                          * But we do make the page is freeable as we
1020                          * can without actually taking the step of unmapping
1021                          * it.
1022                          */
1023                         pmap_clear_modify(m);
1024                         m->dirty = 0;
1025                         m->act_count = 0;
1026                         vm_page_dontneed(m);
1027                         if (tobject->type == OBJT_SWAP)
1028                                 swap_pager_freespace(tobject, tpindex, 1);
1029                 }
1030         }       
1031 }
1032
1033 /*
1034  *      vm_object_shadow:
1035  *
1036  *      Create a new object which is backed by the
1037  *      specified existing object range.  The source
1038  *      object reference is deallocated.
1039  *
1040  *      The new object and offset into that object
1041  *      are returned in the source parameters.
1042  */
1043
1044 void
1045 vm_object_shadow(object, offset, length)
1046         vm_object_t *object;    /* IN/OUT */
1047         vm_ooffset_t *offset;   /* IN/OUT */
1048         vm_size_t length;
1049 {
1050         vm_object_t source;
1051         vm_object_t result;
1052
1053         source = *object;
1054
1055         /*
1056          * Don't create the new object if the old object isn't shared.
1057          */
1058
1059         if (source != NULL &&
1060             source->ref_count == 1 &&
1061             source->handle == NULL &&
1062             (source->type == OBJT_DEFAULT ||
1063              source->type == OBJT_SWAP))
1064                 return;
1065
1066         /*
1067          * Allocate a new object with the given length
1068          */
1069
1070         if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1071                 panic("vm_object_shadow: no object for shadowing");
1072
1073         /*
1074          * The new object shadows the source object, adding a reference to it.
1075          * Our caller changes his reference to point to the new object,
1076          * removing a reference to the source object.  Net result: no change
1077          * of reference count.
1078          *
1079          * Try to optimize the result object's page color when shadowing
1080          * in order to maintain page coloring consistency in the combined 
1081          * shadowed object.
1082          */
1083         result->backing_object = source;
1084         if (source) {
1085                 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1086                 source->shadow_count++;
1087                 source->generation++;
1088                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1089         }
1090
1091         /*
1092          * Store the offset into the source object, and fix up the offset into
1093          * the new object.
1094          */
1095
1096         result->backing_object_offset = *offset;
1097
1098         /*
1099          * Return the new things
1100          */
1101
1102         *offset = 0;
1103         *object = result;
1104 }
1105
1106 #define OBSC_TEST_ALL_SHADOWED  0x0001
1107 #define OBSC_COLLAPSE_NOWAIT    0x0002
1108 #define OBSC_COLLAPSE_WAIT      0x0004
1109
1110 static __inline int
1111 vm_object_backing_scan(vm_object_t object, int op)
1112 {
1113         int s;
1114         int r = 1;
1115         vm_page_t p;
1116         vm_object_t backing_object;
1117         vm_pindex_t backing_offset_index;
1118
1119         s = splvm();
1120
1121         backing_object = object->backing_object;
1122         backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1123
1124         /*
1125          * Initial conditions
1126          */
1127
1128         if (op & OBSC_TEST_ALL_SHADOWED) {
1129                 /*
1130                  * We do not want to have to test for the existence of
1131                  * swap pages in the backing object.  XXX but with the
1132                  * new swapper this would be pretty easy to do.
1133                  *
1134                  * XXX what about anonymous MAP_SHARED memory that hasn't
1135                  * been ZFOD faulted yet?  If we do not test for this, the
1136                  * shadow test may succeed! XXX
1137                  */
1138                 if (backing_object->type != OBJT_DEFAULT) {
1139                         splx(s);
1140                         return(0);
1141                 }
1142         }
1143         if (op & OBSC_COLLAPSE_WAIT) {
1144                 vm_object_set_flag(backing_object, OBJ_DEAD);
1145         }
1146
1147         /*
1148          * Our scan
1149          */
1150
1151         p = TAILQ_FIRST(&backing_object->memq);
1152         while (p) {
1153                 vm_page_t next = TAILQ_NEXT(p, listq);
1154                 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1155
1156                 if (op & OBSC_TEST_ALL_SHADOWED) {
1157                         vm_page_t pp;
1158
1159                         /*
1160                          * Ignore pages outside the parent object's range
1161                          * and outside the parent object's mapping of the 
1162                          * backing object.
1163                          *
1164                          * note that we do not busy the backing object's
1165                          * page.
1166                          */
1167
1168                         if (
1169                             p->pindex < backing_offset_index ||
1170                             new_pindex >= object->size
1171                         ) {
1172                                 p = next;
1173                                 continue;
1174                         }
1175
1176                         /*
1177                          * See if the parent has the page or if the parent's
1178                          * object pager has the page.  If the parent has the
1179                          * page but the page is not valid, the parent's
1180                          * object pager must have the page.
1181                          *
1182                          * If this fails, the parent does not completely shadow
1183                          * the object and we might as well give up now.
1184                          */
1185
1186                         pp = vm_page_lookup(object, new_pindex);
1187                         if (
1188                             (pp == NULL || pp->valid == 0) &&
1189                             !vm_pager_has_page(object, new_pindex, NULL, NULL)
1190                         ) {
1191                                 r = 0;
1192                                 break;
1193                         }
1194                 }
1195
1196                 /*
1197                  * Check for busy page
1198                  */
1199
1200                 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1201                         vm_page_t pp;
1202
1203                         if (op & OBSC_COLLAPSE_NOWAIT) {
1204                                 if (
1205                                     (p->flags & PG_BUSY) ||
1206                                     !p->valid || 
1207                                     p->hold_count || 
1208                                     p->wire_count ||
1209                                     p->busy
1210                                 ) {
1211                                         p = next;
1212                                         continue;
1213                                 }
1214                         } else if (op & OBSC_COLLAPSE_WAIT) {
1215                                 if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1216                                         /*
1217                                          * If we slept, anything could have
1218                                          * happened.  Since the object is
1219                                          * marked dead, the backing offset
1220                                          * should not have changed so we
1221                                          * just restart our scan.
1222                                          */
1223                                         p = TAILQ_FIRST(&backing_object->memq);
1224                                         continue;
1225                                 }
1226                         }
1227
1228                         /* 
1229                          * Busy the page
1230                          */
1231                         vm_page_busy(p);
1232
1233                         KASSERT(
1234                             p->object == backing_object,
1235                             ("vm_object_qcollapse(): object mismatch")
1236                         );
1237
1238                         /*
1239                          * Destroy any associated swap
1240                          */
1241                         if (backing_object->type == OBJT_SWAP) {
1242                                 swap_pager_freespace(
1243                                     backing_object, 
1244                                     p->pindex,
1245                                     1
1246                                 );
1247                         }
1248
1249                         if (
1250                             p->pindex < backing_offset_index ||
1251                             new_pindex >= object->size
1252                         ) {
1253                                 /*
1254                                  * Page is out of the parent object's range, we 
1255                                  * can simply destroy it. 
1256                                  */
1257                                 vm_page_protect(p, VM_PROT_NONE);
1258                                 vm_page_free(p);
1259                                 p = next;
1260                                 continue;
1261                         }
1262
1263                         pp = vm_page_lookup(object, new_pindex);
1264                         if (
1265                             pp != NULL ||
1266                             vm_pager_has_page(object, new_pindex, NULL, NULL)
1267                         ) {
1268                                 /*
1269                                  * page already exists in parent OR swap exists
1270                                  * for this location in the parent.  Destroy 
1271                                  * the original page from the backing object.
1272                                  *
1273                                  * Leave the parent's page alone
1274                                  */
1275                                 vm_page_protect(p, VM_PROT_NONE);
1276                                 vm_page_free(p);
1277                                 p = next;
1278                                 continue;
1279                         }
1280
1281                         /*
1282                          * Page does not exist in parent, rename the
1283                          * page from the backing object to the main object. 
1284                          *
1285                          * If the page was mapped to a process, it can remain 
1286                          * mapped through the rename.
1287                          */
1288                         if ((p->queue - p->pc) == PQ_CACHE)
1289                                 vm_page_deactivate(p);
1290
1291                         vm_page_rename(p, object, new_pindex);
1292                         /* page automatically made dirty by rename */
1293                 }
1294                 p = next;
1295         }
1296         splx(s);
1297         return(r);
1298 }
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(object)
1308         vm_object_t object;
1309 {
1310         vm_object_t backing_object = object->backing_object;
1311
1312         if (backing_object->ref_count != 1)
1313                 return;
1314
1315         backing_object->ref_count += 2;
1316
1317         vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1318
1319         backing_object->ref_count -= 2;
1320 }
1321
1322 /*
1323  *      vm_object_collapse:
1324  *
1325  *      Collapse an object with the object backing it.
1326  *      Pages in the backing object are moved into the
1327  *      parent, and the backing object is deallocated.
1328  */
1329 void
1330 vm_object_collapse(object)
1331         vm_object_t object;
1332 {
1333         while (TRUE) {
1334                 vm_object_t backing_object;
1335
1336                 /*
1337                  * Verify that the conditions are right for collapse:
1338                  *
1339                  * The object exists and the backing object exists.
1340                  */
1341                 if (object == NULL)
1342                         break;
1343
1344                 if ((backing_object = object->backing_object) == NULL)
1345                         break;
1346
1347                 /*
1348                  * we check the backing object first, because it is most likely
1349                  * not collapsable.
1350                  */
1351                 if (backing_object->handle != NULL ||
1352                     (backing_object->type != OBJT_DEFAULT &&
1353                      backing_object->type != OBJT_SWAP) ||
1354                     (backing_object->flags & OBJ_DEAD) ||
1355                     object->handle != NULL ||
1356                     (object->type != OBJT_DEFAULT &&
1357                      object->type != OBJT_SWAP) ||
1358                     (object->flags & OBJ_DEAD)) {
1359                         break;
1360                 }
1361
1362                 if (
1363                     object->paging_in_progress != 0 ||
1364                     backing_object->paging_in_progress != 0
1365                 ) {
1366                         vm_object_qcollapse(object);
1367                         break;
1368                 }
1369
1370                 /*
1371                  * We know that we can either collapse the backing object (if
1372                  * the parent is the only reference to it) or (perhaps) have
1373                  * the parent bypass the object if the parent happens to shadow
1374                  * all the resident pages in the entire backing object.
1375                  *
1376                  * This is ignoring pager-backed pages such as swap pages.
1377                  * vm_object_backing_scan fails the shadowing test in this
1378                  * case.
1379                  */
1380
1381                 if (backing_object->ref_count == 1) {
1382                         /*
1383                          * If there is exactly one reference to the backing
1384                          * object, we can collapse it into the parent.  
1385                          */
1386
1387                         vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1388
1389                         /*
1390                          * Move the pager from backing_object to object.
1391                          */
1392
1393                         if (backing_object->type == OBJT_SWAP) {
1394                                 vm_object_pip_add(backing_object, 1);
1395
1396                                 /*
1397                                  * scrap the paging_offset junk and do a 
1398                                  * discrete copy.  This also removes major 
1399                                  * assumptions about how the swap-pager 
1400                                  * works from where it doesn't belong.  The
1401                                  * new swapper is able to optimize the
1402                                  * destroy-source case.
1403                                  */
1404
1405                                 vm_object_pip_add(object, 1);
1406                                 swap_pager_copy(
1407                                     backing_object,
1408                                     object,
1409                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1410                                 vm_object_pip_wakeup(object);
1411
1412                                 vm_object_pip_wakeup(backing_object);
1413                         }
1414                         /*
1415                          * Object now shadows whatever backing_object did.
1416                          * Note that the reference to 
1417                          * backing_object->backing_object moves from within 
1418                          * backing_object to within object.
1419                          */
1420
1421                         LIST_REMOVE(object, shadow_list);
1422                         object->backing_object->shadow_count--;
1423                         object->backing_object->generation++;
1424                         if (backing_object->backing_object) {
1425                                 LIST_REMOVE(backing_object, shadow_list);
1426                                 backing_object->backing_object->shadow_count--;
1427                                 backing_object->backing_object->generation++;
1428                         }
1429                         object->backing_object = backing_object->backing_object;
1430                         if (object->backing_object) {
1431                                 LIST_INSERT_HEAD(
1432                                     &object->backing_object->shadow_head,
1433                                     object, 
1434                                     shadow_list
1435                                 );
1436                                 object->backing_object->shadow_count++;
1437                                 object->backing_object->generation++;
1438                         }
1439
1440                         object->backing_object_offset +=
1441                             backing_object->backing_object_offset;
1442
1443                         /*
1444                          * Discard backing_object.
1445                          *
1446                          * Since the backing object has no pages, no pager left,
1447                          * and no object references within it, all that is
1448                          * necessary is to dispose of it.
1449                          */
1450
1451                         KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1452                         KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object));
1453                         TAILQ_REMOVE(
1454                             &vm_object_list, 
1455                             backing_object,
1456                             object_list
1457                         );
1458                         vm_object_count--;
1459
1460                         zfree(obj_zone, backing_object);
1461
1462                         object_collapses++;
1463                 } else {
1464                         vm_object_t new_backing_object;
1465
1466                         /*
1467                          * If we do not entirely shadow the backing object,
1468                          * there is nothing we can do so we give up.
1469                          */
1470
1471                         if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1472                                 break;
1473                         }
1474
1475                         /*
1476                          * Make the parent shadow the next object in the
1477                          * chain.  Deallocating backing_object will not remove
1478                          * it, since its reference count is at least 2.
1479                          */
1480
1481                         LIST_REMOVE(object, shadow_list);
1482                         backing_object->shadow_count--;
1483                         backing_object->generation++;
1484
1485                         new_backing_object = backing_object->backing_object;
1486                         if ((object->backing_object = new_backing_object) != NULL) {
1487                                 vm_object_reference(new_backing_object);
1488                                 LIST_INSERT_HEAD(
1489                                     &new_backing_object->shadow_head,
1490                                     object,
1491                                     shadow_list
1492                                 );
1493                                 new_backing_object->shadow_count++;
1494                                 new_backing_object->generation++;
1495                                 object->backing_object_offset +=
1496                                         backing_object->backing_object_offset;
1497                         }
1498
1499                         /*
1500                          * Drop the reference count on backing_object. Since
1501                          * its ref_count was at least 2, it will not vanish;
1502                          * so we don't need to call vm_object_deallocate, but
1503                          * we do anyway.
1504                          */
1505                         vm_object_deallocate(backing_object);
1506                         object_bypasses++;
1507                 }
1508
1509                 /*
1510                  * Try again with this object's new backing object.
1511                  */
1512         }
1513 }
1514
1515 /*
1516  *      vm_object_page_remove: [internal]
1517  *
1518  *      Removes all physical pages in the specified
1519  *      object range from the object's list of pages.
1520  *
1521  *      The object must be locked.
1522  */
1523 void
1524 vm_object_page_remove(object, start, end, clean_only)
1525         vm_object_t object;
1526         vm_pindex_t start;
1527         vm_pindex_t end;
1528         boolean_t clean_only;
1529 {
1530         vm_page_t p, next;
1531         unsigned int size;
1532         int all;
1533
1534         if (object == NULL ||
1535             object->resident_page_count == 0)
1536                 return;
1537
1538         all = ((end == 0) && (start == 0));
1539
1540         /*
1541          * Since physically-backed objects do not use managed pages, we can't
1542          * remove pages from the object (we must instead remove the page
1543          * references, and then destroy the object).
1544          */
1545         KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object"));
1546
1547         vm_object_pip_add(object, 1);
1548 again:
1549         size = end - start;
1550         if (all || size > object->resident_page_count / 4) {
1551                 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
1552                         next = TAILQ_NEXT(p, listq);
1553                         if (all || ((start <= p->pindex) && (p->pindex < end))) {
1554                                 if (p->wire_count != 0) {
1555                                         vm_page_protect(p, VM_PROT_NONE);
1556                                         if (!clean_only)
1557                                                 p->valid = 0;
1558                                         continue;
1559                                 }
1560
1561                                 /*
1562                                  * The busy flags are only cleared at
1563                                  * interrupt -- minimize the spl transitions
1564                                  */
1565
1566                                 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1567                                         goto again;
1568
1569                                 if (clean_only && p->valid) {
1570                                         vm_page_test_dirty(p);
1571                                         if (p->valid & p->dirty)
1572                                                 continue;
1573                                 }
1574
1575                                 vm_page_busy(p);
1576                                 vm_page_protect(p, VM_PROT_NONE);
1577                                 vm_page_free(p);
1578                         }
1579                 }
1580         } else {
1581                 while (size > 0) {
1582                         if ((p = vm_page_lookup(object, start)) != 0) {
1583
1584                                 if (p->wire_count != 0) {
1585                                         vm_page_protect(p, VM_PROT_NONE);
1586                                         if (!clean_only)
1587                                                 p->valid = 0;
1588                                         start += 1;
1589                                         size -= 1;
1590                                         continue;
1591                                 }
1592
1593                                 /*
1594                                  * The busy flags are only cleared at
1595                                  * interrupt -- minimize the spl transitions
1596                                  */
1597                                 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1598                                         goto again;
1599
1600                                 if (clean_only && p->valid) {
1601                                         vm_page_test_dirty(p);
1602                                         if (p->valid & p->dirty) {
1603                                                 start += 1;
1604                                                 size -= 1;
1605                                                 continue;
1606                                         }
1607                                 }
1608
1609                                 vm_page_busy(p);
1610                                 vm_page_protect(p, VM_PROT_NONE);
1611                                 vm_page_free(p);
1612                         }
1613                         start += 1;
1614                         size -= 1;
1615                 }
1616         }
1617         vm_object_pip_wakeup(object);
1618 }
1619
1620 /*
1621  *      Routine:        vm_object_coalesce
1622  *      Function:       Coalesces two objects backing up adjoining
1623  *                      regions of memory into a single object.
1624  *
1625  *      returns TRUE if objects were combined.
1626  *
1627  *      NOTE:   Only works at the moment if the second object is NULL -
1628  *              if it's not, which object do we lock first?
1629  *
1630  *      Parameters:
1631  *              prev_object     First object to coalesce
1632  *              prev_offset     Offset into prev_object
1633  *              next_object     Second object into coalesce
1634  *              next_offset     Offset into next_object
1635  *
1636  *              prev_size       Size of reference to prev_object
1637  *              next_size       Size of reference to next_object
1638  *
1639  *      Conditions:
1640  *      The object must *not* be locked.
1641  */
1642 boolean_t
1643 vm_object_coalesce(prev_object, prev_pindex, prev_size, next_size)
1644         vm_object_t prev_object;
1645         vm_pindex_t prev_pindex;
1646         vm_size_t prev_size, next_size;
1647 {
1648         vm_pindex_t next_pindex;
1649
1650         if (prev_object == NULL) {
1651                 return (TRUE);
1652         }
1653
1654         if (prev_object->type != OBJT_DEFAULT &&
1655             prev_object->type != OBJT_SWAP) {
1656                 return (FALSE);
1657         }
1658
1659         /*
1660          * Try to collapse the object first
1661          */
1662         vm_object_collapse(prev_object);
1663
1664         /*
1665          * Can't coalesce if: . more than one reference . paged out . shadows
1666          * another object . has a copy elsewhere (any of which mean that the
1667          * pages not mapped to prev_entry may be in use anyway)
1668          */
1669
1670         if (prev_object->backing_object != NULL) {
1671                 return (FALSE);
1672         }
1673
1674         prev_size >>= PAGE_SHIFT;
1675         next_size >>= PAGE_SHIFT;
1676         next_pindex = prev_pindex + prev_size;
1677
1678         if ((prev_object->ref_count > 1) &&
1679             (prev_object->size != next_pindex)) {
1680                 return (FALSE);
1681         }
1682
1683         /*
1684          * Remove any pages that may still be in the object from a previous
1685          * deallocation.
1686          */
1687         if (next_pindex < prev_object->size) {
1688                 vm_object_page_remove(prev_object,
1689                                       next_pindex,
1690                                       next_pindex + next_size, FALSE);
1691                 if (prev_object->type == OBJT_SWAP)
1692                         swap_pager_freespace(prev_object,
1693                                              next_pindex, next_size);
1694         }
1695
1696         /*
1697          * Extend the object if necessary.
1698          */
1699         if (next_pindex + next_size > prev_object->size)
1700                 prev_object->size = next_pindex + next_size;
1701
1702         return (TRUE);
1703 }
1704
1705 void
1706 vm_object_set_writeable_dirty(vm_object_t object)
1707 {
1708         struct vnode *vp;
1709
1710         vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1711         if (object->type == OBJT_VNODE &&
1712             (vp = (struct vnode *)object->handle) != NULL) {
1713                 if ((vp->v_flag & VOBJDIRTY) == 0) {
1714                         simple_lock(&vp->v_interlock);
1715                         vp->v_flag |= VOBJDIRTY;
1716                         simple_unlock(&vp->v_interlock);
1717                 }
1718         }
1719 }
1720
1721
1722
1723 #include "opt_ddb.h"
1724 #ifdef DDB
1725 #include <sys/kernel.h>
1726
1727 #include <sys/cons.h>
1728
1729 #include <ddb/ddb.h>
1730
1731 static int      _vm_object_in_map __P((vm_map_t map, vm_object_t object,
1732                                        vm_map_entry_t entry));
1733 static int      vm_object_in_map __P((vm_object_t object));
1734
1735 static int
1736 _vm_object_in_map(map, object, entry)
1737         vm_map_t map;
1738         vm_object_t object;
1739         vm_map_entry_t entry;
1740 {
1741         vm_map_t tmpm;
1742         vm_map_entry_t tmpe;
1743         vm_object_t obj;
1744         int entcount;
1745
1746         if (map == 0)
1747                 return 0;
1748
1749         if (entry == 0) {
1750                 tmpe = map->header.next;
1751                 entcount = map->nentries;
1752                 while (entcount-- && (tmpe != &map->header)) {
1753                         if( _vm_object_in_map(map, object, tmpe)) {
1754                                 return 1;
1755                         }
1756                         tmpe = tmpe->next;
1757                 }
1758         } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
1759                 tmpm = entry->object.sub_map;
1760                 tmpe = tmpm->header.next;
1761                 entcount = tmpm->nentries;
1762                 while (entcount-- && tmpe != &tmpm->header) {
1763                         if( _vm_object_in_map(tmpm, object, tmpe)) {
1764                                 return 1;
1765                         }
1766                         tmpe = tmpe->next;
1767                 }
1768         } else if ((obj = entry->object.vm_object) != NULL) {
1769                 for(; obj; obj=obj->backing_object)
1770                         if( obj == object) {
1771                                 return 1;
1772                         }
1773         }
1774         return 0;
1775 }
1776
1777 static int
1778 vm_object_in_map( object)
1779         vm_object_t object;
1780 {
1781         struct proc *p;
1782         for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1783                 if( !p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1784                         continue;
1785                 if( _vm_object_in_map(&p->p_vmspace->vm_map, object, 0))
1786                         return 1;
1787         }
1788         if( _vm_object_in_map( kernel_map, object, 0))
1789                 return 1;
1790         if( _vm_object_in_map( kmem_map, object, 0))
1791                 return 1;
1792         if( _vm_object_in_map( pager_map, object, 0))
1793                 return 1;
1794         if( _vm_object_in_map( buffer_map, object, 0))
1795                 return 1;
1796         if( _vm_object_in_map( mb_map, object, 0))
1797                 return 1;
1798         return 0;
1799 }
1800
1801 DB_SHOW_COMMAND(vmochk, vm_object_check)
1802 {
1803         vm_object_t object;
1804
1805         /*
1806          * make sure that internal objs are in a map somewhere
1807          * and none have zero ref counts.
1808          */
1809         for (object = TAILQ_FIRST(&vm_object_list);
1810                         object != NULL;
1811                         object = TAILQ_NEXT(object, object_list)) {
1812                 if (object->handle == NULL &&
1813                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1814                         if (object->ref_count == 0) {
1815                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
1816                                         (long)object->size);
1817                         }
1818                         if (!vm_object_in_map(object)) {
1819                                 db_printf(
1820                         "vmochk: internal obj is not in a map: "
1821                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1822                                     object->ref_count, (u_long)object->size, 
1823                                     (u_long)object->size,
1824                                     (void *)object->backing_object);
1825                         }
1826                 }
1827         }
1828 }
1829
1830 /*
1831  *      vm_object_print:        [ debug ]
1832  */
1833 DB_SHOW_COMMAND(object, vm_object_print_static)
1834 {
1835         /* XXX convert args. */
1836         vm_object_t object = (vm_object_t)addr;
1837         boolean_t full = have_addr;
1838
1839         vm_page_t p;
1840
1841         /* XXX count is an (unused) arg.  Avoid shadowing it. */
1842 #define count   was_count
1843
1844         int count;
1845
1846         if (object == NULL)
1847                 return;
1848
1849         db_iprintf(
1850             "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1851             object, (int)object->type, (u_long)object->size,
1852             object->resident_page_count, object->ref_count, object->flags);
1853         /*
1854          * XXX no %qd in kernel.  Truncate object->backing_object_offset.
1855          */
1856         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1857             object->shadow_count, 
1858             object->backing_object ? object->backing_object->ref_count : 0,
1859             object->backing_object, (long)object->backing_object_offset);
1860
1861         if (!full)
1862                 return;
1863
1864         db_indent += 2;
1865         count = 0;
1866         for (p = TAILQ_FIRST(&object->memq); p != NULL; p = TAILQ_NEXT(p, listq)) {
1867                 if (count == 0)
1868                         db_iprintf("memory:=");
1869                 else if (count == 6) {
1870                         db_printf("\n");
1871                         db_iprintf(" ...");
1872                         count = 0;
1873                 } else
1874                         db_printf(",");
1875                 count++;
1876
1877                 db_printf("(off=0x%lx,page=0x%lx)",
1878                     (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1879         }
1880         if (count != 0)
1881                 db_printf("\n");
1882         db_indent -= 2;
1883 }
1884
1885 /* XXX. */
1886 #undef count
1887
1888 /* XXX need this non-static entry for calling from vm_map_print. */
1889 void
1890 vm_object_print(addr, have_addr, count, modif)
1891         /* db_expr_t */ long addr;
1892         boolean_t have_addr;
1893         /* db_expr_t */ long count;
1894         char *modif;
1895 {
1896         vm_object_print_static(addr, have_addr, count, modif);
1897 }
1898
1899 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1900 {
1901         vm_object_t object;
1902         int nl = 0;
1903         int c;
1904         for (object = TAILQ_FIRST(&vm_object_list);
1905                         object != NULL;
1906                         object = TAILQ_NEXT(object, object_list)) {
1907                 vm_pindex_t idx, fidx;
1908                 vm_pindex_t osize;
1909                 vm_offset_t pa = -1, padiff;
1910                 int rcount;
1911                 vm_page_t m;
1912
1913                 db_printf("new object: %p\n", (void *)object);
1914                 if ( nl > 18) {
1915                         c = cngetc();
1916                         if (c != ' ')
1917                                 return;
1918                         nl = 0;
1919                 }
1920                 nl++;
1921                 rcount = 0;
1922                 fidx = 0;
1923                 osize = object->size;
1924                 if (osize > 128)
1925                         osize = 128;
1926                 for(idx=0;idx<osize;idx++) {
1927                         m = vm_page_lookup(object, idx);
1928                         if (m == NULL) {
1929                                 if (rcount) {
1930                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1931                                                 (long)fidx, rcount, (long)pa);
1932                                         if ( nl > 18) {
1933                                                 c = cngetc();
1934                                                 if (c != ' ')
1935                                                         return;
1936                                                 nl = 0;
1937                                         }
1938                                         nl++;
1939                                         rcount = 0;
1940                                 }
1941                                 continue;
1942                         }
1943
1944                                 
1945                         if (rcount &&
1946                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1947                                 ++rcount;
1948                                 continue;
1949                         }
1950                         if (rcount) {
1951                                 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1952                                 padiff >>= PAGE_SHIFT;
1953                                 padiff &= PQ_L2_MASK;
1954                                 if (padiff == 0) {
1955                                         pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1956                                         ++rcount;
1957                                         continue;
1958                                 }
1959                                 db_printf(" index(%ld)run(%d)pa(0x%lx)",
1960                                         (long)fidx, rcount, (long)pa);
1961                                 db_printf("pd(%ld)\n", (long)padiff);
1962                                 if ( nl > 18) {
1963                                         c = cngetc();
1964                                         if (c != ' ')
1965                                                 return;
1966                                         nl = 0;
1967                                 }
1968                                 nl++;
1969                         }
1970                         fidx = idx;
1971                         pa = VM_PAGE_TO_PHYS(m);
1972                         rcount = 1;
1973                 }
1974                 if (rcount) {
1975                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1976                                 (long)fidx, rcount, (long)pa);
1977                         if ( nl > 18) {
1978                                 c = cngetc();
1979                                 if (c != ' ')
1980                                         return;
1981                                 nl = 0;
1982                         }
1983                         nl++;
1984                 }
1985         }
1986 }
1987 #endif /* DDB */