e14b7ba9acb9871055f3a9a099b5e6f49051c4f3
[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 1.25 2006/05/25 07:36:37 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 static long vm_object_count;            /* count of all objects */
135 vm_object_t kernel_object;
136 vm_object_t kmem_object;
137 static struct vm_object kernel_object_store;
138 static struct vm_object kmem_object_store;
139 extern int vm_pageout_page_count;
140
141 static long object_collapses;
142 static long object_bypasses;
143 static int next_index;
144 static vm_zone_t obj_zone;
145 static struct vm_zone obj_zone_store;
146 static int object_hash_rand;
147 #define VM_OBJECTS_INIT 256
148 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
149
150 void
151 _vm_object_allocate(objtype_t type, vm_size_t size, vm_object_t object)
152 {
153         int incr;
154         TAILQ_INIT(&object->memq);
155         LIST_INIT(&object->shadow_head);
156
157         object->type = type;
158         object->size = size;
159         object->ref_count = 1;
160         object->flags = 0;
161         if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
162                 vm_object_set_flag(object, OBJ_ONEMAPPING);
163         object->paging_in_progress = 0;
164         object->resident_page_count = 0;
165         object->shadow_count = 0;
166         object->pg_color = next_index;
167         if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
168                 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
169         else
170                 incr = size;
171         next_index = (next_index + incr) & PQ_L2_MASK;
172         object->handle = NULL;
173         object->backing_object = NULL;
174         object->backing_object_offset = (vm_ooffset_t) 0;
175         /*
176          * Try to generate a number that will spread objects out in the
177          * hash table.  We 'wipe' new objects across the hash in 128 page
178          * increments plus 1 more to offset it a little more by the time
179          * it wraps around.
180          */
181         object->hash_rand = object_hash_rand - 129;
182
183         object->generation++;
184
185         crit_enter();
186         TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
187         vm_object_count++;
188         object_hash_rand = object->hash_rand;
189         crit_exit();
190 }
191
192 /*
193  *      vm_object_init:
194  *
195  *      Initialize the VM objects module.
196  */
197 void
198 vm_object_init(void)
199 {
200         TAILQ_INIT(&vm_object_list);
201         
202         kernel_object = &kernel_object_store;
203         _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
204             kernel_object);
205
206         kmem_object = &kmem_object_store;
207         _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
208             kmem_object);
209
210         obj_zone = &obj_zone_store;
211         zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
212                 vm_objects_init, VM_OBJECTS_INIT);
213 }
214
215 void
216 vm_object_init2(void)
217 {
218         zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1);
219 }
220
221 /*
222  *      vm_object_allocate:
223  *
224  *      Returns a new object with the given size.
225  */
226
227 vm_object_t
228 vm_object_allocate(objtype_t type, vm_size_t size)
229 {
230         vm_object_t result;
231
232         result = (vm_object_t) zalloc(obj_zone);
233
234         _vm_object_allocate(type, size, result);
235
236         return (result);
237 }
238
239
240 /*
241  *      vm_object_reference:
242  *
243  *      Gets another reference to the given object.
244  */
245 void
246 vm_object_reference(vm_object_t object)
247 {
248         if (object == NULL)
249                 return;
250
251         object->ref_count++;
252         if (object->type == OBJT_VNODE) {
253                 vref(object->handle);
254                 /* XXX what if the vnode is being destroyed? */
255         }
256 }
257
258 static void
259 vm_object_vndeallocate(vm_object_t object)
260 {
261         struct vnode *vp = (struct vnode *) object->handle;
262
263         KASSERT(object->type == OBJT_VNODE,
264             ("vm_object_vndeallocate: not a vnode object"));
265         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
266 #ifdef INVARIANTS
267         if (object->ref_count == 0) {
268                 vprint("vm_object_vndeallocate", vp);
269                 panic("vm_object_vndeallocate: bad object reference count");
270         }
271 #endif
272
273         object->ref_count--;
274         if (object->ref_count == 0)
275                 vp->v_flag &= ~VTEXT;
276         vrele(vp);
277 }
278
279 /*
280  *      vm_object_deallocate:
281  *
282  *      Release a reference to the specified object,
283  *      gained either through a vm_object_allocate
284  *      or a vm_object_reference call.  When all references
285  *      are gone, storage associated with this object
286  *      may be relinquished.
287  *
288  *      No object may be locked.
289  */
290 void
291 vm_object_deallocate(vm_object_t object)
292 {
293         vm_object_t temp;
294
295         while (object != NULL) {
296                 if (object->type == OBJT_VNODE) {
297                         vm_object_vndeallocate(object);
298                         return;
299                 }
300
301                 if (object->ref_count == 0) {
302                         panic("vm_object_deallocate: object deallocated too many times: %d", object->type);
303                 } else if (object->ref_count > 2) {
304                         object->ref_count--;
305                         return;
306                 }
307
308                 /*
309                  * Here on ref_count of one or two, which are special cases for
310                  * objects.
311                  */
312                 if ((object->ref_count == 2) && (object->shadow_count == 0)) {
313                         vm_object_set_flag(object, OBJ_ONEMAPPING);
314                         object->ref_count--;
315                         return;
316                 } else if ((object->ref_count == 2) && (object->shadow_count == 1)) {
317                         object->ref_count--;
318                         if ((object->handle == NULL) &&
319                             (object->type == OBJT_DEFAULT ||
320                              object->type == OBJT_SWAP)) {
321                                 vm_object_t robject;
322
323                                 robject = LIST_FIRST(&object->shadow_head);
324                                 KASSERT(robject != NULL,
325                                     ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
326                                          object->ref_count,
327                                          object->shadow_count));
328                                 if ((robject->handle == NULL) &&
329                                     (robject->type == OBJT_DEFAULT ||
330                                      robject->type == OBJT_SWAP)) {
331
332                                         robject->ref_count++;
333
334                                         while (
335                                                 robject->paging_in_progress ||
336                                                 object->paging_in_progress
337                                         ) {
338                                                 vm_object_pip_sleep(robject, "objde1");
339                                                 vm_object_pip_sleep(object, "objde2");
340                                         }
341
342                                         if (robject->ref_count == 1) {
343                                                 robject->ref_count--;
344                                                 object = robject;
345                                                 goto doterm;
346                                         }
347
348                                         object = robject;
349                                         vm_object_collapse(object);
350                                         continue;
351                                 }
352                         }
353
354                         return;
355
356                 } else {
357                         object->ref_count--;
358                         if (object->ref_count != 0)
359                                 return;
360                 }
361
362 doterm:
363
364                 temp = object->backing_object;
365                 if (temp) {
366                         LIST_REMOVE(object, shadow_list);
367                         temp->shadow_count--;
368                         temp->generation++;
369                         object->backing_object = NULL;
370                 }
371
372                 /*
373                  * Don't double-terminate, we could be in a termination
374                  * recursion due to the terminate having to sync data
375                  * to disk.
376                  */
377                 if ((object->flags & OBJ_DEAD) == 0)
378                         vm_object_terminate(object);
379                 object = temp;
380         }
381 }
382
383 /*
384  *      vm_object_terminate actually destroys the specified object, freeing
385  *      up all previously used resources.
386  *
387  *      The object must be locked.
388  *      This routine may block.
389  */
390 void
391 vm_object_terminate(vm_object_t object)
392 {
393         vm_page_t p;
394
395         /*
396          * Make sure no one uses us.
397          */
398         vm_object_set_flag(object, OBJ_DEAD);
399
400         /*
401          * wait for the pageout daemon to be done with the object
402          */
403         vm_object_pip_wait(object, "objtrm");
404
405         KASSERT(!object->paging_in_progress,
406                 ("vm_object_terminate: pageout in progress"));
407
408         /*
409          * Clean and free the pages, as appropriate. All references to the
410          * object are gone, so we don't need to lock it.
411          */
412         if (object->type == OBJT_VNODE) {
413                 struct vnode *vp;
414
415                 /*
416                  * Clean pages and flush buffers.
417                  */
418                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
419
420                 vp = (struct vnode *) object->handle;
421                 vinvalbuf(vp, V_SAVE, 0, 0);
422         }
423
424         /*
425          * Wait for any I/O to complete, after which there had better not
426          * be any references left on the object.
427          */
428         vm_object_pip_wait(object, "objtrm");
429
430         if (object->ref_count != 0)
431                 panic("vm_object_terminate: object with references, ref_count=%d", object->ref_count);
432
433         /*
434          * Now free any remaining pages. For internal objects, this also
435          * removes them from paging queues. Don't free wired pages, just
436          * remove them from the object. 
437          */
438         crit_enter();
439         while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
440                 if (p->busy || (p->flags & PG_BUSY))
441                         panic("vm_object_terminate: freeing busy page %p", p);
442                 if (p->wire_count == 0) {
443                         vm_page_busy(p);
444                         vm_page_free(p);
445                         mycpu->gd_cnt.v_pfree++;
446                 } else {
447                         vm_page_busy(p);
448                         vm_page_remove(p);
449                         vm_page_wakeup(p);
450                 }
451         }
452         crit_exit();
453
454         /*
455          * Let the pager know object is dead.
456          */
457         vm_pager_deallocate(object);
458
459         /*
460          * Remove the object from the global object list.
461          */
462         crit_enter();
463         TAILQ_REMOVE(&vm_object_list, object, object_list);
464         vm_object_count--;
465         crit_exit();
466
467         wakeup(object);
468         if (object->ref_count != 0)
469                 panic("vm_object_terminate2: object with references, ref_count=%d", object->ref_count);
470
471         /*
472          * Free the space for the object.
473          */
474         zfree(obj_zone, object);
475 }
476
477 /*
478  *      vm_object_page_clean
479  *
480  *      Clean all dirty pages in the specified range of object.  Leaves page 
481  *      on whatever queue it is currently on.   If NOSYNC is set then do not
482  *      write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
483  *      leaving the object dirty.
484  *
485  *      When stuffing pages asynchronously, allow clustering.  XXX we need a
486  *      synchronous clustering mode implementation.
487  *
488  *      Odd semantics: if start == end, we clean everything.
489  */
490
491 void
492 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
493     int flags)
494 {
495         vm_page_t p, np;
496         vm_offset_t tstart, tend;
497         vm_pindex_t pi;
498         struct vnode *vp;
499         int clearobjflags;
500         int pagerflags;
501         int curgeneration;
502
503         if (object->type != OBJT_VNODE ||
504                 (object->flags & OBJ_MIGHTBEDIRTY) == 0)
505                 return;
506
507         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
508         pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
509
510         vp = object->handle;
511
512         vm_object_set_flag(object, OBJ_CLEANING);
513
514         /*
515          * Handle 'entire object' case
516          */
517         tstart = start;
518         if (end == 0) {
519                 tend = object->size;
520         } else {
521                 tend = end;
522         }
523
524         /*
525          * If the caller is smart and only msync()s a range he knows is
526          * dirty, we may be able to avoid an object scan.  This results in
527          * a phenominal improvement in performance.  We cannot do this
528          * as a matter of course because the object may be huge - e.g.
529          * the size might be in the gigabytes or terrabytes.
530          */
531         if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
532                 vm_offset_t tscan;
533                 int scanlimit;
534                 int scanreset;
535
536                 scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
537                 if (scanreset < 16)
538                         scanreset = 16;
539                 pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
540
541                 scanlimit = scanreset;
542                 tscan = tstart;
543
544                 /*
545                  * spl protection is required despite the obj generation
546                  * tracking because we cannot safely call vm_page_test_dirty()
547                  * or avoid page field tests against an interrupt unbusy/free
548                  * race that might occur prior to the busy check in
549                  * vm_object_page_collect_flush().
550                  */
551                 crit_enter();
552                 while (tscan < tend) {
553                         curgeneration = object->generation;
554                         p = vm_page_lookup(object, tscan);
555                         if (p == NULL || p->valid == 0 ||
556                             (p->queue - p->pc) == PQ_CACHE) {
557                                 if (--scanlimit == 0)
558                                         break;
559                                 ++tscan;
560                                 continue;
561                         }
562                         vm_page_test_dirty(p);
563                         if ((p->dirty & p->valid) == 0) {
564                                 if (--scanlimit == 0)
565                                         break;
566                                 ++tscan;
567                                 continue;
568                         }
569                         /*
570                          * If we have been asked to skip nosync pages and 
571                          * this is a nosync page, we can't continue.
572                          */
573                         if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
574                                 if (--scanlimit == 0)
575                                         break;
576                                 ++tscan;
577                                 continue;
578                         }
579                         scanlimit = scanreset;
580
581                         /*
582                          * This returns 0 if it was unable to busy the first
583                          * page (i.e. had to sleep).
584                          */
585                         tscan += vm_object_page_collect_flush(object, p, 
586                                                 curgeneration, pagerflags);
587                 }
588                 crit_exit();
589
590                 /*
591                  * If everything was dirty and we flushed it successfully,
592                  * and the requested range is not the entire object, we
593                  * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
594                  * return immediately.
595                  */
596                 if (tscan >= tend && (tstart || tend < object->size)) {
597                         vm_object_clear_flag(object, OBJ_CLEANING);
598                         return;
599                 }
600                 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
601         }
602
603         /*
604          * Generally set CLEANCHK interlock and make the page read-only so
605          * we can then clear the object flags.
606          *
607          * However, if this is a nosync mmap then the object is likely to 
608          * stay dirty so do not mess with the page and do not clear the
609          * object flags.
610          *
611          * spl protection is required because an interrupt can remove page
612          * from the object.
613          */
614         clearobjflags = 1;
615
616         crit_enter();
617         for (p = TAILQ_FIRST(&object->memq); p; p = TAILQ_NEXT(p, listq)) {
618                 vm_page_flag_set(p, PG_CLEANCHK);
619                 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
620                         clearobjflags = 0;
621                 else
622                         vm_page_protect(p, VM_PROT_READ);
623         }
624         crit_exit();
625
626         if (clearobjflags && (tstart == 0) && (tend == object->size)) {
627                 struct vnode *vp;
628
629                 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
630                 if (object->type == OBJT_VNODE &&
631                     (vp = (struct vnode *)object->handle) != NULL) {
632                         if (vp->v_flag & VOBJDIRTY) 
633                                 vclrflags(vp, VOBJDIRTY);
634                 }
635         }
636
637         /*
638          * spl protection is required both to avoid an interrupt unbusy/free
639          * race against a vm_page_lookup(), and also to ensure that the
640          * memq is consistent.  We do not want a busy page to be ripped out
641          * from under us.
642          */
643         crit_enter();
644 rescan:
645         crit_exit();
646         crit_enter();
647         curgeneration = object->generation;
648
649         for (p = TAILQ_FIRST(&object->memq); p; p = np) {
650                 int n;
651
652                 np = TAILQ_NEXT(p, listq);
653
654 again:
655                 pi = p->pindex;
656                 if (((p->flags & PG_CLEANCHK) == 0) ||
657                         (pi < tstart) || (pi >= tend) ||
658                         (p->valid == 0) ||
659                         ((p->queue - p->pc) == PQ_CACHE)) {
660                         vm_page_flag_clear(p, PG_CLEANCHK);
661                         continue;
662                 }
663
664                 vm_page_test_dirty(p);
665                 if ((p->dirty & p->valid) == 0) {
666                         vm_page_flag_clear(p, PG_CLEANCHK);
667                         continue;
668                 }
669
670                 /*
671                  * If we have been asked to skip nosync pages and this is a
672                  * nosync page, skip it.  Note that the object flags were
673                  * not cleared in this case so we do not have to set them.
674                  */
675                 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
676                         vm_page_flag_clear(p, PG_CLEANCHK);
677                         continue;
678                 }
679
680                 n = vm_object_page_collect_flush(object, p,
681                         curgeneration, pagerflags);
682                 if (n == 0)
683                         goto rescan;
684                 if (object->generation != curgeneration)
685                         goto rescan;
686
687                 /*
688                  * Try to optimize the next page.  If we can't we pick up
689                  * our (random) scan where we left off.
690                  */
691                 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
692                         if ((p = vm_page_lookup(object, pi + n)) != NULL)
693                                 goto again;
694                 }
695         }
696         crit_exit();
697
698         vm_object_clear_flag(object, OBJ_CLEANING);
699         return;
700 }
701
702 /*
703  * This routine must be called within a critical section to properly avoid
704  * an interrupt unbusy/free race that can occur prior to the busy check.
705  *
706  * Using the object generation number here to detect page ripout is not
707  * the best idea in the world. XXX
708  *
709  * NOTE: we operate under the assumption that a page found to not be busy
710  * will not be ripped out from under us by an interrupt.  XXX we should
711  * recode this to explicitly busy the pages.
712  */
713 static int
714 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
715 {
716         int runlen;
717         int maxf;
718         int chkb;
719         int maxb;
720         int i;
721         vm_pindex_t pi;
722         vm_page_t maf[vm_pageout_page_count];
723         vm_page_t mab[vm_pageout_page_count];
724         vm_page_t ma[vm_pageout_page_count];
725
726         pi = p->pindex;
727         while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
728                 if (object->generation != curgeneration) {
729                         return(0);
730                 }
731         }
732
733         maxf = 0;
734         for(i = 1; i < vm_pageout_page_count; i++) {
735                 vm_page_t tp;
736
737                 if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
738                         if ((tp->flags & PG_BUSY) ||
739                                 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 
740                                  (tp->flags & PG_CLEANCHK) == 0) ||
741                                 (tp->busy != 0))
742                                 break;
743                         if((tp->queue - tp->pc) == PQ_CACHE) {
744                                 vm_page_flag_clear(tp, PG_CLEANCHK);
745                                 break;
746                         }
747                         vm_page_test_dirty(tp);
748                         if ((tp->dirty & tp->valid) == 0) {
749                                 vm_page_flag_clear(tp, PG_CLEANCHK);
750                                 break;
751                         }
752                         maf[ i - 1 ] = tp;
753                         maxf++;
754                         continue;
755                 }
756                 break;
757         }
758
759         maxb = 0;
760         chkb = vm_pageout_page_count -  maxf;
761         if (chkb) {
762                 for(i = 1; i < chkb;i++) {
763                         vm_page_t tp;
764
765                         if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
766                                 if ((tp->flags & PG_BUSY) ||
767                                         ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 && 
768                                          (tp->flags & PG_CLEANCHK) == 0) ||
769                                         (tp->busy != 0))
770                                         break;
771                                 if((tp->queue - tp->pc) == PQ_CACHE) {
772                                         vm_page_flag_clear(tp, PG_CLEANCHK);
773                                         break;
774                                 }
775                                 vm_page_test_dirty(tp);
776                                 if ((tp->dirty & tp->valid) == 0) {
777                                         vm_page_flag_clear(tp, PG_CLEANCHK);
778                                         break;
779                                 }
780                                 mab[ i - 1 ] = tp;
781                                 maxb++;
782                                 continue;
783                         }
784                         break;
785                 }
786         }
787
788         for(i = 0; i < maxb; i++) {
789                 int index = (maxb - i) - 1;
790                 ma[index] = mab[i];
791                 vm_page_flag_clear(ma[index], PG_CLEANCHK);
792         }
793         vm_page_flag_clear(p, PG_CLEANCHK);
794         ma[maxb] = p;
795         for(i = 0; i < maxf; i++) {
796                 int index = (maxb + i) + 1;
797                 ma[index] = maf[i];
798                 vm_page_flag_clear(ma[index], PG_CLEANCHK);
799         }
800         runlen = maxb + maxf + 1;
801
802         vm_pageout_flush(ma, runlen, pagerflags);
803         for (i = 0; i < runlen; i++) {
804                 if (ma[i]->valid & ma[i]->dirty) {
805                         vm_page_protect(ma[i], VM_PROT_READ);
806                         vm_page_flag_set(ma[i], PG_CLEANCHK);
807
808                         /*
809                          * maxf will end up being the actual number of pages
810                          * we wrote out contiguously, non-inclusive of the
811                          * first page.  We do not count look-behind pages.
812                          */
813                         if (i >= maxb + 1 && (maxf > i - maxb - 1))
814                                 maxf = i - maxb - 1;
815                 }
816         }
817         return(maxf + 1);
818 }
819
820 #ifdef not_used
821 /* XXX I cannot tell if this should be an exported symbol */
822 /*
823  *      vm_object_deactivate_pages
824  *
825  *      Deactivate all pages in the specified object.  (Keep its pages
826  *      in memory even though it is no longer referenced.)
827  *
828  *      The object must be locked.
829  */
830 static void
831 vm_object_deactivate_pages(vm_object_t object)
832 {
833         vm_page_t p, next;
834
835         crit_enter();
836         for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
837                 next = TAILQ_NEXT(p, listq);
838                 vm_page_deactivate(p);
839         }
840         crit_exit();
841 }
842 #endif
843
844 /*
845  * Same as vm_object_pmap_copy, except range checking really
846  * works, and is meant for small sections of an object.
847  *
848  * This code protects resident pages by making them read-only
849  * and is typically called on a fork or split when a page
850  * is converted to copy-on-write.  
851  *
852  * NOTE: If the page is already at VM_PROT_NONE, calling
853  * vm_page_protect will have no effect.
854  */
855 void
856 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
857 {
858         vm_pindex_t idx;
859         vm_page_t p;
860
861         if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
862                 return;
863
864         /*
865          * spl protection needed to prevent races between the lookup,
866          * an interrupt unbusy/free, and our protect call.
867          */
868         crit_enter();
869         for (idx = start; idx < end; idx++) {
870                 p = vm_page_lookup(object, idx);
871                 if (p == NULL)
872                         continue;
873                 vm_page_protect(p, VM_PROT_READ);
874         }
875         crit_exit();
876 }
877
878 /*
879  *      vm_object_pmap_remove:
880  *
881  *      Removes all physical pages in the specified
882  *      object range from all physical maps.
883  *
884  *      The object must *not* be locked.
885  */
886 void
887 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
888 {
889         vm_page_t p;
890
891         if (object == NULL)
892                 return;
893
894         /*
895          * spl protection is required because an interrupt can unbusy/free
896          * a page.
897          */
898         crit_enter();
899         for (p = TAILQ_FIRST(&object->memq);
900             p != NULL;
901             p = TAILQ_NEXT(p, listq)
902         ) {
903                 if (p->pindex >= start && p->pindex < end)
904                         vm_page_protect(p, VM_PROT_NONE);
905         }
906         crit_exit();
907         if ((start == 0) && (object->size == end))
908                 vm_object_clear_flag(object, OBJ_WRITEABLE);
909 }
910
911 /*
912  *      vm_object_madvise:
913  *
914  *      Implements the madvise function at the object/page level.
915  *
916  *      MADV_WILLNEED   (any object)
917  *
918  *          Activate the specified pages if they are resident.
919  *
920  *      MADV_DONTNEED   (any object)
921  *
922  *          Deactivate the specified pages if they are resident.
923  *
924  *      MADV_FREE       (OBJT_DEFAULT/OBJT_SWAP objects,
925  *                       OBJ_ONEMAPPING only)
926  *
927  *          Deactivate and clean the specified pages if they are
928  *          resident.  This permits the process to reuse the pages
929  *          without faulting or the kernel to reclaim the pages
930  *          without I/O.
931  */
932 void
933 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
934 {
935         vm_pindex_t end, tpindex;
936         vm_object_t tobject;
937         vm_page_t m;
938
939         if (object == NULL)
940                 return;
941
942         end = pindex + count;
943
944         /*
945          * Locate and adjust resident pages
946          */
947
948         for (; pindex < end; pindex += 1) {
949 relookup:
950                 tobject = object;
951                 tpindex = pindex;
952 shadowlookup:
953                 /*
954                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
955                  * and those pages must be OBJ_ONEMAPPING.
956                  */
957                 if (advise == MADV_FREE) {
958                         if ((tobject->type != OBJT_DEFAULT &&
959                              tobject->type != OBJT_SWAP) ||
960                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
961                                 continue;
962                         }
963                 }
964
965                 /*
966                  * spl protection is required to avoid a race between the
967                  * lookup, an interrupt unbusy/free, and our busy check.
968                  */
969
970                 crit_enter();
971                 m = vm_page_lookup(tobject, tpindex);
972
973                 if (m == NULL) {
974                         /*
975                          * There may be swap even if there is no backing page
976                          */
977                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
978                                 swap_pager_freespace(tobject, tpindex, 1);
979
980                         /*
981                          * next object
982                          */
983                         crit_exit();
984                         if (tobject->backing_object == NULL)
985                                 continue;
986                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
987                         tobject = tobject->backing_object;
988                         goto shadowlookup;
989                 }
990
991                 /*
992                  * If the page is busy or not in a normal active state,
993                  * we skip it.  If the page is not managed there are no
994                  * page queues to mess with.  Things can break if we mess
995                  * with pages in any of the below states.
996                  */
997                 if (
998                     m->hold_count ||
999                     m->wire_count ||
1000                     (m->flags & PG_UNMANAGED) ||
1001                     m->valid != VM_PAGE_BITS_ALL
1002                 ) {
1003                         crit_exit();
1004                         continue;
1005                 }
1006
1007                 if (vm_page_sleep_busy(m, TRUE, "madvpo")) {
1008                         crit_exit();
1009                         goto relookup;
1010                 }
1011                 crit_exit();
1012
1013                 /*
1014                  * Theoretically once a page is known not to be busy, an
1015                  * interrupt cannot come along and rip it out from under us.
1016                  */
1017
1018                 if (advise == MADV_WILLNEED) {
1019                         vm_page_activate(m);
1020                 } else if (advise == MADV_DONTNEED) {
1021                         vm_page_dontneed(m);
1022                 } else if (advise == MADV_FREE) {
1023                         /*
1024                          * Mark the page clean.  This will allow the page
1025                          * to be freed up by the system.  However, such pages
1026                          * are often reused quickly by malloc()/free()
1027                          * so we do not do anything that would cause
1028                          * a page fault if we can help it.
1029                          *
1030                          * Specifically, we do not try to actually free
1031                          * the page now nor do we try to put it in the
1032                          * cache (which would cause a page fault on reuse).
1033                          *
1034                          * But we do make the page is freeable as we
1035                          * can without actually taking the step of unmapping
1036                          * it.
1037                          */
1038                         pmap_clear_modify(m);
1039                         m->dirty = 0;
1040                         m->act_count = 0;
1041                         vm_page_dontneed(m);
1042                         if (tobject->type == OBJT_SWAP)
1043                                 swap_pager_freespace(tobject, tpindex, 1);
1044                 }
1045         }       
1046 }
1047
1048 /*
1049  *      vm_object_shadow:
1050  *
1051  *      Create a new object which is backed by the
1052  *      specified existing object range.  The source
1053  *      object reference is deallocated.
1054  *
1055  *      The new object and offset into that object
1056  *      are returned in the source parameters.
1057  */
1058
1059 void
1060 vm_object_shadow(vm_object_t *object,   /* IN/OUT */
1061                  vm_ooffset_t *offset,  /* IN/OUT */
1062                  vm_size_t length)
1063 {
1064         vm_object_t source;
1065         vm_object_t result;
1066
1067         source = *object;
1068
1069         /*
1070          * Don't create the new object if the old object isn't shared.
1071          */
1072
1073         if (source != NULL &&
1074             source->ref_count == 1 &&
1075             source->handle == NULL &&
1076             (source->type == OBJT_DEFAULT ||
1077              source->type == OBJT_SWAP))
1078                 return;
1079
1080         /*
1081          * Allocate a new object with the given length
1082          */
1083
1084         if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1085                 panic("vm_object_shadow: no object for shadowing");
1086
1087         /*
1088          * The new object shadows the source object, adding a reference to it.
1089          * Our caller changes his reference to point to the new object,
1090          * removing a reference to the source object.  Net result: no change
1091          * of reference count.
1092          *
1093          * Try to optimize the result object's page color when shadowing
1094          * in order to maintain page coloring consistency in the combined 
1095          * shadowed object.
1096          */
1097         result->backing_object = source;
1098         if (source) {
1099                 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1100                 source->shadow_count++;
1101                 source->generation++;
1102                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1103         }
1104
1105         /*
1106          * Store the offset into the source object, and fix up the offset into
1107          * the new object.
1108          */
1109
1110         result->backing_object_offset = *offset;
1111
1112         /*
1113          * Return the new things
1114          */
1115
1116         *offset = 0;
1117         *object = result;
1118 }
1119
1120 #define OBSC_TEST_ALL_SHADOWED  0x0001
1121 #define OBSC_COLLAPSE_NOWAIT    0x0002
1122 #define OBSC_COLLAPSE_WAIT      0x0004
1123
1124 static __inline int
1125 vm_object_backing_scan(vm_object_t object, int op)
1126 {
1127         int r = 1;
1128         vm_page_t p;
1129         vm_object_t backing_object;
1130         vm_pindex_t backing_offset_index;
1131
1132         /*
1133          * spl protection is required to avoid races between the memq/lookup,
1134          * an interrupt doing an unbusy/free, and our busy check.  Amoung
1135          * other things.
1136          */
1137         crit_enter();
1138
1139         backing_object = object->backing_object;
1140         backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1141
1142         /*
1143          * Initial conditions
1144          */
1145
1146         if (op & OBSC_TEST_ALL_SHADOWED) {
1147                 /*
1148                  * We do not want to have to test for the existence of
1149                  * swap pages in the backing object.  XXX but with the
1150                  * new swapper this would be pretty easy to do.
1151                  *
1152                  * XXX what about anonymous MAP_SHARED memory that hasn't
1153                  * been ZFOD faulted yet?  If we do not test for this, the
1154                  * shadow test may succeed! XXX
1155                  */
1156                 if (backing_object->type != OBJT_DEFAULT) {
1157                         crit_exit();
1158                         return(0);
1159                 }
1160         }
1161         if (op & OBSC_COLLAPSE_WAIT) {
1162                 KKASSERT((backing_object->flags & OBJ_DEAD) == 0);
1163                 vm_object_set_flag(backing_object, OBJ_DEAD);
1164         }
1165
1166         /*
1167          * Our scan
1168          */
1169
1170         p = TAILQ_FIRST(&backing_object->memq);
1171         while (p) {
1172                 vm_page_t next = TAILQ_NEXT(p, listq);
1173                 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1174
1175                 if (op & OBSC_TEST_ALL_SHADOWED) {
1176                         vm_page_t pp;
1177
1178                         /*
1179                          * Ignore pages outside the parent object's range
1180                          * and outside the parent object's mapping of the 
1181                          * backing object.
1182                          *
1183                          * note that we do not busy the backing object's
1184                          * page.
1185                          */
1186
1187                         if (
1188                             p->pindex < backing_offset_index ||
1189                             new_pindex >= object->size
1190                         ) {
1191                                 p = next;
1192                                 continue;
1193                         }
1194
1195                         /*
1196                          * See if the parent has the page or if the parent's
1197                          * object pager has the page.  If the parent has the
1198                          * page but the page is not valid, the parent's
1199                          * object pager must have the page.
1200                          *
1201                          * If this fails, the parent does not completely shadow
1202                          * the object and we might as well give up now.
1203                          */
1204
1205                         pp = vm_page_lookup(object, new_pindex);
1206                         if (
1207                             (pp == NULL || pp->valid == 0) &&
1208                             !vm_pager_has_page(object, new_pindex, NULL, NULL)
1209                         ) {
1210                                 r = 0;
1211                                 break;
1212                         }
1213                 }
1214
1215                 /*
1216                  * Check for busy page
1217                  */
1218
1219                 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1220                         vm_page_t pp;
1221
1222                         if (op & OBSC_COLLAPSE_NOWAIT) {
1223                                 if (
1224                                     (p->flags & PG_BUSY) ||
1225                                     !p->valid || 
1226                                     p->hold_count || 
1227                                     p->wire_count ||
1228                                     p->busy
1229                                 ) {
1230                                         p = next;
1231                                         continue;
1232                                 }
1233                         } else if (op & OBSC_COLLAPSE_WAIT) {
1234                                 if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1235                                         /*
1236                                          * If we slept, anything could have
1237                                          * happened.  Since the object is
1238                                          * marked dead, the backing offset
1239                                          * should not have changed so we
1240                                          * just restart our scan.
1241                                          */
1242                                         p = TAILQ_FIRST(&backing_object->memq);
1243                                         continue;
1244                                 }
1245                         }
1246
1247                         /* 
1248                          * Busy the page
1249                          */
1250                         vm_page_busy(p);
1251
1252                         KASSERT(
1253                             p->object == backing_object,
1254                             ("vm_object_qcollapse(): object mismatch")
1255                         );
1256
1257                         /*
1258                          * Destroy any associated swap
1259                          */
1260                         if (backing_object->type == OBJT_SWAP) {
1261                                 swap_pager_freespace(
1262                                     backing_object, 
1263                                     p->pindex,
1264                                     1
1265                                 );
1266                         }
1267
1268                         if (
1269                             p->pindex < backing_offset_index ||
1270                             new_pindex >= object->size
1271                         ) {
1272                                 /*
1273                                  * Page is out of the parent object's range, we 
1274                                  * can simply destroy it. 
1275                                  */
1276                                 vm_page_protect(p, VM_PROT_NONE);
1277                                 vm_page_free(p);
1278                                 p = next;
1279                                 continue;
1280                         }
1281
1282                         pp = vm_page_lookup(object, new_pindex);
1283                         if (
1284                             pp != NULL ||
1285                             vm_pager_has_page(object, new_pindex, NULL, NULL)
1286                         ) {
1287                                 /*
1288                                  * page already exists in parent OR swap exists
1289                                  * for this location in the parent.  Destroy 
1290                                  * the original page from the backing object.
1291                                  *
1292                                  * Leave the parent's page alone
1293                                  */
1294                                 vm_page_protect(p, VM_PROT_NONE);
1295                                 vm_page_free(p);
1296                                 p = next;
1297                                 continue;
1298                         }
1299
1300                         /*
1301                          * Page does not exist in parent, rename the
1302                          * page from the backing object to the main object. 
1303                          *
1304                          * If the page was mapped to a process, it can remain 
1305                          * mapped through the rename.
1306                          */
1307                         if ((p->queue - p->pc) == PQ_CACHE)
1308                                 vm_page_deactivate(p);
1309
1310                         vm_page_rename(p, object, new_pindex);
1311                         /* page automatically made dirty by rename */
1312                 }
1313                 p = next;
1314         }
1315         crit_exit();
1316         return(r);
1317 }
1318
1319
1320 /*
1321  * this version of collapse allows the operation to occur earlier and
1322  * when paging_in_progress is true for an object...  This is not a complete
1323  * operation, but should plug 99.9% of the rest of the leaks.
1324  */
1325 static void
1326 vm_object_qcollapse(vm_object_t object)
1327 {
1328         vm_object_t backing_object = object->backing_object;
1329
1330         if (backing_object->ref_count != 1)
1331                 return;
1332
1333         backing_object->ref_count += 2;
1334
1335         vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1336
1337         backing_object->ref_count -= 2;
1338 }
1339
1340 /*
1341  *      vm_object_collapse:
1342  *
1343  *      Collapse an object with the object backing it.
1344  *      Pages in the backing object are moved into the
1345  *      parent, and the backing object is deallocated.
1346  */
1347 void
1348 vm_object_collapse(vm_object_t object)
1349 {
1350         while (TRUE) {
1351                 vm_object_t backing_object;
1352
1353                 /*
1354                  * Verify that the conditions are right for collapse:
1355                  *
1356                  * The object exists and the backing object exists.
1357                  */
1358                 if (object == NULL)
1359                         break;
1360
1361                 if ((backing_object = object->backing_object) == NULL)
1362                         break;
1363
1364                 /*
1365                  * we check the backing object first, because it is most likely
1366                  * not collapsable.
1367                  */
1368                 if (backing_object->handle != NULL ||
1369                     (backing_object->type != OBJT_DEFAULT &&
1370                      backing_object->type != OBJT_SWAP) ||
1371                     (backing_object->flags & OBJ_DEAD) ||
1372                     object->handle != NULL ||
1373                     (object->type != OBJT_DEFAULT &&
1374                      object->type != OBJT_SWAP) ||
1375                     (object->flags & OBJ_DEAD)) {
1376                         break;
1377                 }
1378
1379                 if (
1380                     object->paging_in_progress != 0 ||
1381                     backing_object->paging_in_progress != 0
1382                 ) {
1383                         vm_object_qcollapse(object);
1384                         break;
1385                 }
1386
1387                 /*
1388                  * We know that we can either collapse the backing object (if
1389                  * the parent is the only reference to it) or (perhaps) have
1390                  * the parent bypass the object if the parent happens to shadow
1391                  * all the resident pages in the entire backing object.
1392                  *
1393                  * This is ignoring pager-backed pages such as swap pages.
1394                  * vm_object_backing_scan fails the shadowing test in this
1395                  * case.
1396                  */
1397
1398                 if (backing_object->ref_count == 1) {
1399                         /*
1400                          * If there is exactly one reference to the backing
1401                          * object, we can collapse it into the parent.  
1402                          */
1403                         vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1404
1405                         /*
1406                          * Move the pager from backing_object to object.
1407                          */
1408
1409                         if (backing_object->type == OBJT_SWAP) {
1410                                 vm_object_pip_add(backing_object, 1);
1411
1412                                 /*
1413                                  * scrap the paging_offset junk and do a 
1414                                  * discrete copy.  This also removes major 
1415                                  * assumptions about how the swap-pager 
1416                                  * works from where it doesn't belong.  The
1417                                  * new swapper is able to optimize the
1418                                  * destroy-source case.
1419                                  */
1420
1421                                 vm_object_pip_add(object, 1);
1422                                 swap_pager_copy(
1423                                     backing_object,
1424                                     object,
1425                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1426                                 vm_object_pip_wakeup(object);
1427
1428                                 vm_object_pip_wakeup(backing_object);
1429                         }
1430                         /*
1431                          * Object now shadows whatever backing_object did.
1432                          * Note that the reference to 
1433                          * backing_object->backing_object moves from within 
1434                          * backing_object to within object.
1435                          */
1436
1437                         LIST_REMOVE(object, shadow_list);
1438                         object->backing_object->shadow_count--;
1439                         object->backing_object->generation++;
1440                         if (backing_object->backing_object) {
1441                                 LIST_REMOVE(backing_object, shadow_list);
1442                                 backing_object->backing_object->shadow_count--;
1443                                 backing_object->backing_object->generation++;
1444                         }
1445                         object->backing_object = backing_object->backing_object;
1446                         if (object->backing_object) {
1447                                 LIST_INSERT_HEAD(
1448                                     &object->backing_object->shadow_head,
1449                                     object, 
1450                                     shadow_list
1451                                 );
1452                                 object->backing_object->shadow_count++;
1453                                 object->backing_object->generation++;
1454                         }
1455
1456                         object->backing_object_offset +=
1457                             backing_object->backing_object_offset;
1458
1459                         /*
1460                          * Discard backing_object.
1461                          *
1462                          * Since the backing object has no pages, no pager left,
1463                          * and no object references within it, all that is
1464                          * necessary is to dispose of it.
1465                          */
1466
1467                         KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1468                         KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object));
1469                         crit_enter();
1470                         TAILQ_REMOVE(
1471                             &vm_object_list, 
1472                             backing_object,
1473                             object_list
1474                         );
1475                         vm_object_count--;
1476                         crit_exit();
1477
1478                         zfree(obj_zone, backing_object);
1479
1480                         object_collapses++;
1481                 } else {
1482                         vm_object_t new_backing_object;
1483
1484                         /*
1485                          * If we do not entirely shadow the backing object,
1486                          * there is nothing we can do so we give up.
1487                          */
1488
1489                         if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1490                                 break;
1491                         }
1492
1493                         /*
1494                          * Make the parent shadow the next object in the
1495                          * chain.  Deallocating backing_object will not remove
1496                          * it, since its reference count is at least 2.
1497                          */
1498
1499                         LIST_REMOVE(object, shadow_list);
1500                         backing_object->shadow_count--;
1501                         backing_object->generation++;
1502
1503                         new_backing_object = backing_object->backing_object;
1504                         if ((object->backing_object = new_backing_object) != NULL) {
1505                                 vm_object_reference(new_backing_object);
1506                                 LIST_INSERT_HEAD(
1507                                     &new_backing_object->shadow_head,
1508                                     object,
1509                                     shadow_list
1510                                 );
1511                                 new_backing_object->shadow_count++;
1512                                 new_backing_object->generation++;
1513                                 object->backing_object_offset +=
1514                                         backing_object->backing_object_offset;
1515                         }
1516
1517                         /*
1518                          * Drop the reference count on backing_object. Since
1519                          * its ref_count was at least 2, it will not vanish;
1520                          * so we don't need to call vm_object_deallocate, but
1521                          * we do anyway.
1522                          */
1523                         vm_object_deallocate(backing_object);
1524                         object_bypasses++;
1525                 }
1526
1527                 /*
1528                  * Try again with this object's new backing object.
1529                  */
1530         }
1531 }
1532
1533 /*
1534  *      vm_object_page_remove: [internal]
1535  *
1536  *      Removes all physical pages in the specified
1537  *      object range from the object's list of pages.
1538  */
1539 void
1540 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1541     boolean_t clean_only)
1542 {
1543         vm_page_t p, next;
1544         unsigned int size;
1545         int all;
1546
1547         if (object == NULL || object->resident_page_count == 0)
1548                 return;
1549
1550         all = ((end == 0) && (start == 0));
1551
1552         /*
1553          * Since physically-backed objects do not use managed pages, we can't
1554          * remove pages from the object (we must instead remove the page
1555          * references, and then destroy the object).
1556          */
1557         KASSERT(object->type != OBJT_PHYS, 
1558                 ("attempt to remove pages from a physical object"));
1559
1560         /*
1561          * Indicating that the object is undergoing paging.
1562          *
1563          * spl protection is required to avoid a race between the memq scan,
1564          * an interrupt unbusy/free, and the busy check.
1565          */
1566         vm_object_pip_add(object, 1);
1567         crit_enter();
1568 again:
1569         size = end - start;
1570         if (all || size > object->resident_page_count / 4) {
1571                 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
1572                         next = TAILQ_NEXT(p, listq);
1573                         if (all || ((start <= p->pindex) && (p->pindex < end))) {
1574                                 if (p->wire_count != 0) {
1575                                         vm_page_protect(p, VM_PROT_NONE);
1576                                         if (!clean_only)
1577                                                 p->valid = 0;
1578                                         continue;
1579                                 }
1580
1581                                 /*
1582                                  * The busy flags are only cleared at
1583                                  * interrupt -- minimize the spl transitions
1584                                  */
1585
1586                                 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1587                                         goto again;
1588
1589                                 if (clean_only && p->valid) {
1590                                         vm_page_test_dirty(p);
1591                                         if (p->valid & p->dirty)
1592                                                 continue;
1593                                 }
1594
1595                                 vm_page_busy(p);
1596                                 vm_page_protect(p, VM_PROT_NONE);
1597                                 vm_page_free(p);
1598                         }
1599                 }
1600         } else {
1601                 while (size > 0) {
1602                         if ((p = vm_page_lookup(object, start)) != 0) {
1603                                 if (p->wire_count != 0) {
1604                                         vm_page_protect(p, VM_PROT_NONE);
1605                                         if (!clean_only)
1606                                                 p->valid = 0;
1607                                         start += 1;
1608                                         size -= 1;
1609                                         continue;
1610                                 }
1611
1612                                 /*
1613                                  * The busy flags are only cleared at
1614                                  * interrupt -- minimize the spl transitions
1615                                  */
1616                                 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1617                                         goto again;
1618
1619                                 if (clean_only && p->valid) {
1620                                         vm_page_test_dirty(p);
1621                                         if (p->valid & p->dirty) {
1622                                                 start += 1;
1623                                                 size -= 1;
1624                                                 continue;
1625                                         }
1626                                 }
1627
1628                                 vm_page_busy(p);
1629                                 vm_page_protect(p, VM_PROT_NONE);
1630                                 vm_page_free(p);
1631                         }
1632                         start += 1;
1633                         size -= 1;
1634                 }
1635         }
1636         crit_exit();
1637         vm_object_pip_wakeup(object);
1638 }
1639
1640 /*
1641  *      Routine:        vm_object_coalesce
1642  *      Function:       Coalesces two objects backing up adjoining
1643  *                      regions of memory into a single object.
1644  *
1645  *      returns TRUE if objects were combined.
1646  *
1647  *      NOTE:   Only works at the moment if the second object is NULL -
1648  *              if it's not, which object do we lock first?
1649  *
1650  *      Parameters:
1651  *              prev_object     First object to coalesce
1652  *              prev_offset     Offset into prev_object
1653  *              next_object     Second object into coalesce
1654  *              next_offset     Offset into next_object
1655  *
1656  *              prev_size       Size of reference to prev_object
1657  *              next_size       Size of reference to next_object
1658  *
1659  *      Conditions:
1660  *      The object must *not* be locked.
1661  */
1662 boolean_t
1663 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
1664     vm_size_t prev_size, vm_size_t next_size)
1665 {
1666         vm_pindex_t next_pindex;
1667
1668         if (prev_object == NULL) {
1669                 return (TRUE);
1670         }
1671
1672         if (prev_object->type != OBJT_DEFAULT &&
1673             prev_object->type != OBJT_SWAP) {
1674                 return (FALSE);
1675         }
1676
1677         /*
1678          * Try to collapse the object first
1679          */
1680         vm_object_collapse(prev_object);
1681
1682         /*
1683          * Can't coalesce if: . more than one reference . paged out . shadows
1684          * another object . has a copy elsewhere (any of which mean that the
1685          * pages not mapped to prev_entry may be in use anyway)
1686          */
1687
1688         if (prev_object->backing_object != NULL) {
1689                 return (FALSE);
1690         }
1691
1692         prev_size >>= PAGE_SHIFT;
1693         next_size >>= PAGE_SHIFT;
1694         next_pindex = prev_pindex + prev_size;
1695
1696         if ((prev_object->ref_count > 1) &&
1697             (prev_object->size != next_pindex)) {
1698                 return (FALSE);
1699         }
1700
1701         /*
1702          * Remove any pages that may still be in the object from a previous
1703          * deallocation.
1704          */
1705         if (next_pindex < prev_object->size) {
1706                 vm_object_page_remove(prev_object,
1707                                       next_pindex,
1708                                       next_pindex + next_size, FALSE);
1709                 if (prev_object->type == OBJT_SWAP)
1710                         swap_pager_freespace(prev_object,
1711                                              next_pindex, next_size);
1712         }
1713
1714         /*
1715          * Extend the object if necessary.
1716          */
1717         if (next_pindex + next_size > prev_object->size)
1718                 prev_object->size = next_pindex + next_size;
1719
1720         return (TRUE);
1721 }
1722
1723 void
1724 vm_object_set_writeable_dirty(vm_object_t object)
1725 {
1726         struct vnode *vp;
1727
1728         vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1729         if (object->type == OBJT_VNODE &&
1730             (vp = (struct vnode *)object->handle) != NULL) {
1731                 if ((vp->v_flag & VOBJDIRTY) == 0) {
1732                         vsetflags(vp, VOBJDIRTY);
1733                 }
1734         }
1735 }
1736
1737
1738
1739 #include "opt_ddb.h"
1740 #ifdef DDB
1741 #include <sys/kernel.h>
1742
1743 #include <sys/cons.h>
1744
1745 #include <ddb/ddb.h>
1746
1747 static int      _vm_object_in_map (vm_map_t map, vm_object_t object,
1748                                        vm_map_entry_t entry);
1749 static int      vm_object_in_map (vm_object_t object);
1750
1751 static int
1752 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
1753 {
1754         vm_map_t tmpm;
1755         vm_map_entry_t tmpe;
1756         vm_object_t obj;
1757         int entcount;
1758
1759         if (map == 0)
1760                 return 0;
1761
1762         if (entry == 0) {
1763                 tmpe = map->header.next;
1764                 entcount = map->nentries;
1765                 while (entcount-- && (tmpe != &map->header)) {
1766                         if( _vm_object_in_map(map, object, tmpe)) {
1767                                 return 1;
1768                         }
1769                         tmpe = tmpe->next;
1770                 }
1771         } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
1772                 tmpm = entry->object.sub_map;
1773                 tmpe = tmpm->header.next;
1774                 entcount = tmpm->nentries;
1775                 while (entcount-- && tmpe != &tmpm->header) {
1776                         if( _vm_object_in_map(tmpm, object, tmpe)) {
1777                                 return 1;
1778                         }
1779                         tmpe = tmpe->next;
1780                 }
1781         } else if ((obj = entry->object.vm_object) != NULL) {
1782                 for(; obj; obj=obj->backing_object)
1783                         if( obj == object) {
1784                                 return 1;
1785                         }
1786         }
1787         return 0;
1788 }
1789
1790 static int vm_object_in_map_callback(struct proc *p, void *data);
1791
1792 struct vm_object_in_map_info {
1793         vm_object_t object;
1794         int rv;
1795 };
1796
1797 static int
1798 vm_object_in_map(vm_object_t object)
1799 {
1800         struct vm_object_in_map_info info;
1801
1802         info.rv = 0;
1803         info.object = object;
1804
1805         allproc_scan(vm_object_in_map_callback, &info);
1806         if (info.rv)
1807                 return 1;
1808         if( _vm_object_in_map( kernel_map, object, 0))
1809                 return 1;
1810         if( _vm_object_in_map( pager_map, object, 0))
1811                 return 1;
1812         if( _vm_object_in_map( buffer_map, object, 0))
1813                 return 1;
1814         return 0;
1815 }
1816
1817 static int
1818 vm_object_in_map_callback(struct proc *p, void *data)
1819 {
1820         struct vm_object_in_map_info *info = data;
1821
1822         if (p->p_vmspace) {
1823                 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) {
1824                         info->rv = 1;
1825                         return -1;
1826                 }
1827         }
1828         return (0);
1829 }
1830
1831 DB_SHOW_COMMAND(vmochk, vm_object_check)
1832 {
1833         vm_object_t object;
1834
1835         /*
1836          * make sure that internal objs are in a map somewhere
1837          * and none have zero ref counts.
1838          */
1839         for (object = TAILQ_FIRST(&vm_object_list);
1840                         object != NULL;
1841                         object = TAILQ_NEXT(object, object_list)) {
1842                 if (object->handle == NULL &&
1843                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1844                         if (object->ref_count == 0) {
1845                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
1846                                         (long)object->size);
1847                         }
1848                         if (!vm_object_in_map(object)) {
1849                                 db_printf(
1850                         "vmochk: internal obj is not in a map: "
1851                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1852                                     object->ref_count, (u_long)object->size, 
1853                                     (u_long)object->size,
1854                                     (void *)object->backing_object);
1855                         }
1856                 }
1857         }
1858 }
1859
1860 /*
1861  *      vm_object_print:        [ debug ]
1862  */
1863 DB_SHOW_COMMAND(object, vm_object_print_static)
1864 {
1865         /* XXX convert args. */
1866         vm_object_t object = (vm_object_t)addr;
1867         boolean_t full = have_addr;
1868
1869         vm_page_t p;
1870
1871         /* XXX count is an (unused) arg.  Avoid shadowing it. */
1872 #define count   was_count
1873
1874         int count;
1875
1876         if (object == NULL)
1877                 return;
1878
1879         db_iprintf(
1880             "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1881             object, (int)object->type, (u_long)object->size,
1882             object->resident_page_count, object->ref_count, object->flags);
1883         /*
1884          * XXX no %qd in kernel.  Truncate object->backing_object_offset.
1885          */
1886         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1887             object->shadow_count, 
1888             object->backing_object ? object->backing_object->ref_count : 0,
1889             object->backing_object, (long)object->backing_object_offset);
1890
1891         if (!full)
1892                 return;
1893
1894         db_indent += 2;
1895         count = 0;
1896         for (p = TAILQ_FIRST(&object->memq); p != NULL; p = TAILQ_NEXT(p, listq)) {
1897                 if (count == 0)
1898                         db_iprintf("memory:=");
1899                 else if (count == 6) {
1900                         db_printf("\n");
1901                         db_iprintf(" ...");
1902                         count = 0;
1903                 } else
1904                         db_printf(",");
1905                 count++;
1906
1907                 db_printf("(off=0x%lx,page=0x%lx)",
1908                     (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1909         }
1910         if (count != 0)
1911                 db_printf("\n");
1912         db_indent -= 2;
1913 }
1914
1915 /* XXX. */
1916 #undef count
1917
1918 /* XXX need this non-static entry for calling from vm_map_print. */
1919 void
1920 vm_object_print(/* db_expr_t */ long addr,
1921                 boolean_t have_addr,
1922                 /* db_expr_t */ long count,
1923                 char *modif)
1924 {
1925         vm_object_print_static(addr, have_addr, count, modif);
1926 }
1927
1928 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1929 {
1930         vm_object_t object;
1931         int nl = 0;
1932         int c;
1933         for (object = TAILQ_FIRST(&vm_object_list);
1934                         object != NULL;
1935                         object = TAILQ_NEXT(object, object_list)) {
1936                 vm_pindex_t idx, fidx;
1937                 vm_pindex_t osize;
1938                 vm_paddr_t pa = -1, padiff;
1939                 int rcount;
1940                 vm_page_t m;
1941
1942                 db_printf("new object: %p\n", (void *)object);
1943                 if ( nl > 18) {
1944                         c = cngetc();
1945                         if (c != ' ')
1946                                 return;
1947                         nl = 0;
1948                 }
1949                 nl++;
1950                 rcount = 0;
1951                 fidx = 0;
1952                 osize = object->size;
1953                 if (osize > 128)
1954                         osize = 128;
1955                 for (idx = 0; idx < osize; idx++) {
1956                         m = vm_page_lookup(object, idx);
1957                         if (m == NULL) {
1958                                 if (rcount) {
1959                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1960                                                 (long)fidx, rcount, (long)pa);
1961                                         if ( nl > 18) {
1962                                                 c = cngetc();
1963                                                 if (c != ' ')
1964                                                         return;
1965                                                 nl = 0;
1966                                         }
1967                                         nl++;
1968                                         rcount = 0;
1969                                 }
1970                                 continue;
1971                         }
1972
1973                                 
1974                         if (rcount &&
1975                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1976                                 ++rcount;
1977                                 continue;
1978                         }
1979                         if (rcount) {
1980                                 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1981                                 padiff >>= PAGE_SHIFT;
1982                                 padiff &= PQ_L2_MASK;
1983                                 if (padiff == 0) {
1984                                         pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1985                                         ++rcount;
1986                                         continue;
1987                                 }
1988                                 db_printf(" index(%ld)run(%d)pa(0x%lx)",
1989                                         (long)fidx, rcount, (long)pa);
1990                                 db_printf("pd(%ld)\n", (long)padiff);
1991                                 if ( nl > 18) {
1992                                         c = cngetc();
1993                                         if (c != ' ')
1994                                                 return;
1995                                         nl = 0;
1996                                 }
1997                                 nl++;
1998                         }
1999                         fidx = idx;
2000                         pa = VM_PAGE_TO_PHYS(m);
2001                         rcount = 1;
2002                 }
2003                 if (rcount) {
2004                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2005                                 (long)fidx, rcount, (long)pa);
2006                         if ( nl > 18) {
2007                                 c = cngetc();
2008                                 if (c != ' ')
2009                                         return;
2010                                 nl = 0;
2011                         }
2012                         nl++;
2013                 }
2014         }
2015 }
2016 #endif /* DDB */