]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/vm/vm_object.c
Remove white spaces.
[FreeBSD/FreeBSD.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  * 4. Neither the name of the University nor the names of its contributors
17  *    may be used to endorse or promote products derived from this software
18  *    without specific prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  *
32  *      from: @(#)vm_object.c   8.5 (Berkeley) 3/22/94
33  *
34  *
35  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36  * All rights reserved.
37  *
38  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39  *
40  * Permission to use, copy, modify and distribute this software and
41  * its documentation is hereby granted, provided that both the copyright
42  * notice and this permission notice appear in all copies of the
43  * software, derivative works or modified versions, and any portions
44  * thereof, and that both notices appear in supporting documentation.
45  *
46  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49  *
50  * Carnegie Mellon requests users of this software to return to
51  *
52  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
53  *  School of Computer Science
54  *  Carnegie Mellon University
55  *  Pittsburgh PA 15213-3890
56  *
57  * any improvements or extensions that they make and grant Carnegie the
58  * rights to redistribute these changes.
59  */
60
61 /*
62  *      Virtual memory object module.
63  */
64
65 #include <sys/cdefs.h>
66 __FBSDID("$FreeBSD$");
67
68 #include "opt_vm.h"
69
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/lock.h>
73 #include <sys/mman.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/mutex.h>
78 #include <sys/proc.h>           /* for curproc, pageproc */
79 #include <sys/socket.h>
80 #include <sys/resourcevar.h>
81 #include <sys/vnode.h>
82 #include <sys/vmmeter.h>
83 #include <sys/sx.h>
84
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_pager.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_reserv.h>
97 #include <vm/uma.h>
98
99 static int old_msync;
100 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
101     "Use old (insecure) msync behavior");
102
103 static int      vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
104                     int pagerflags, int flags, boolean_t *clearobjflags,
105                     boolean_t *eio);
106 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
107                     boolean_t *clearobjflags);
108 static void     vm_object_qcollapse(vm_object_t object);
109 static void     vm_object_vndeallocate(vm_object_t object);
110
111 /*
112  *      Virtual memory objects maintain the actual data
113  *      associated with allocated virtual memory.  A given
114  *      page of memory exists within exactly one object.
115  *
116  *      An object is only deallocated when all "references"
117  *      are given up.  Only one "reference" to a given
118  *      region of an object should be writeable.
119  *
120  *      Associated with each object is a list of all resident
121  *      memory pages belonging to that object; this list is
122  *      maintained by the "vm_page" module, and locked by the object's
123  *      lock.
124  *
125  *      Each object also records a "pager" routine which is
126  *      used to retrieve (and store) pages to the proper backing
127  *      storage.  In addition, objects may be backed by other
128  *      objects from which they were virtual-copied.
129  *
130  *      The only items within the object structure which are
131  *      modified after time of creation are:
132  *              reference count         locked by object's lock
133  *              pager routine           locked by object's lock
134  *
135  */
136
137 struct object_q vm_object_list;
138 struct mtx vm_object_list_mtx;  /* lock for object list and count */
139
140 struct vm_object kernel_object_store;
141 struct vm_object kmem_object_store;
142
143 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
144     "VM object stats");
145
146 static long object_collapses;
147 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
148     &object_collapses, 0, "VM object collapses");
149
150 static long object_bypasses;
151 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
152     &object_bypasses, 0, "VM object bypasses");
153
154 static uma_zone_t obj_zone;
155
156 static int vm_object_zinit(void *mem, int size, int flags);
157
158 #ifdef INVARIANTS
159 static void vm_object_zdtor(void *mem, int size, void *arg);
160
161 static void
162 vm_object_zdtor(void *mem, int size, void *arg)
163 {
164         vm_object_t object;
165
166         object = (vm_object_t)mem;
167         KASSERT(TAILQ_EMPTY(&object->memq),
168             ("object %p has resident pages",
169             object));
170 #if VM_NRESERVLEVEL > 0
171         KASSERT(LIST_EMPTY(&object->rvq),
172             ("object %p has reservations",
173             object));
174 #endif
175         KASSERT(object->cache == NULL,
176             ("object %p has cached pages",
177             object));
178         KASSERT(object->paging_in_progress == 0,
179             ("object %p paging_in_progress = %d",
180             object, object->paging_in_progress));
181         KASSERT(object->resident_page_count == 0,
182             ("object %p resident_page_count = %d",
183             object, object->resident_page_count));
184         KASSERT(object->shadow_count == 0,
185             ("object %p shadow_count = %d",
186             object, object->shadow_count));
187 }
188 #endif
189
190 static int
191 vm_object_zinit(void *mem, int size, int flags)
192 {
193         vm_object_t object;
194
195         object = (vm_object_t)mem;
196         bzero(&object->mtx, sizeof(object->mtx));
197         VM_OBJECT_LOCK_INIT(object, "standard object");
198
199         /* These are true for any object that has been freed */
200         object->paging_in_progress = 0;
201         object->resident_page_count = 0;
202         object->shadow_count = 0;
203         return (0);
204 }
205
206 void
207 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
208 {
209
210         TAILQ_INIT(&object->memq);
211         LIST_INIT(&object->shadow_head);
212
213         object->root = NULL;
214         object->type = type;
215         switch (type) {
216         case OBJT_DEAD:
217                 panic("_vm_object_allocate: can't create OBJT_DEAD");
218         case OBJT_DEFAULT:
219         case OBJT_SWAP:
220                 object->flags = OBJ_ONEMAPPING;
221                 break;
222         case OBJT_DEVICE:
223         case OBJT_SG:
224                 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
225                 break;
226         case OBJT_MGTDEVICE:
227                 object->flags = OBJ_FICTITIOUS;
228                 break;
229         case OBJT_PHYS:
230                 object->flags = OBJ_UNMANAGED;
231                 break;
232         case OBJT_VNODE:
233                 object->flags = 0;
234                 break;
235         default:
236                 panic("_vm_object_allocate: type %d is undefined", type);
237         }
238         object->size = size;
239         object->generation = 1;
240         object->ref_count = 1;
241         object->memattr = VM_MEMATTR_DEFAULT;
242         object->cred = NULL;
243         object->charge = 0;
244         object->pg_color = 0;
245         object->handle = NULL;
246         object->backing_object = NULL;
247         object->backing_object_offset = (vm_ooffset_t) 0;
248 #if VM_NRESERVLEVEL > 0
249         LIST_INIT(&object->rvq);
250 #endif
251         object->cache = NULL;
252
253         mtx_lock(&vm_object_list_mtx);
254         TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
255         mtx_unlock(&vm_object_list_mtx);
256 }
257
258 /*
259  *      vm_object_init:
260  *
261  *      Initialize the VM objects module.
262  */
263 void
264 vm_object_init(void)
265 {
266         TAILQ_INIT(&vm_object_list);
267         mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
268         
269         VM_OBJECT_LOCK_INIT(kernel_object, "kernel object");
270         _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
271             kernel_object);
272 #if VM_NRESERVLEVEL > 0
273         kernel_object->flags |= OBJ_COLORED;
274         kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
275 #endif
276
277         VM_OBJECT_LOCK_INIT(kmem_object, "kmem object");
278         _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
279             kmem_object);
280 #if VM_NRESERVLEVEL > 0
281         kmem_object->flags |= OBJ_COLORED;
282         kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
283 #endif
284
285         /*
286          * The lock portion of struct vm_object must be type stable due
287          * to vm_pageout_fallback_object_lock locking a vm object
288          * without holding any references to it.
289          */
290         obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
291 #ifdef INVARIANTS
292             vm_object_zdtor,
293 #else
294             NULL,
295 #endif
296             vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
297 }
298
299 void
300 vm_object_clear_flag(vm_object_t object, u_short bits)
301 {
302
303         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
304         object->flags &= ~bits;
305 }
306
307 /*
308  *      Sets the default memory attribute for the specified object.  Pages
309  *      that are allocated to this object are by default assigned this memory
310  *      attribute.
311  *
312  *      Presently, this function must be called before any pages are allocated
313  *      to the object.  In the future, this requirement may be relaxed for
314  *      "default" and "swap" objects.
315  */
316 int
317 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
318 {
319
320         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
321         switch (object->type) {
322         case OBJT_DEFAULT:
323         case OBJT_DEVICE:
324         case OBJT_MGTDEVICE:
325         case OBJT_PHYS:
326         case OBJT_SG:
327         case OBJT_SWAP:
328         case OBJT_VNODE:
329                 if (!TAILQ_EMPTY(&object->memq))
330                         return (KERN_FAILURE);
331                 break;
332         case OBJT_DEAD:
333                 return (KERN_INVALID_ARGUMENT);
334         default:
335                 panic("vm_object_set_memattr: object %p is of undefined type",
336                     object);
337         }
338         object->memattr = memattr;
339         return (KERN_SUCCESS);
340 }
341
342 void
343 vm_object_pip_add(vm_object_t object, short i)
344 {
345
346         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
347         object->paging_in_progress += i;
348 }
349
350 void
351 vm_object_pip_subtract(vm_object_t object, short i)
352 {
353
354         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
355         object->paging_in_progress -= i;
356 }
357
358 void
359 vm_object_pip_wakeup(vm_object_t object)
360 {
361
362         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
363         object->paging_in_progress--;
364         if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
365                 vm_object_clear_flag(object, OBJ_PIPWNT);
366                 wakeup(object);
367         }
368 }
369
370 void
371 vm_object_pip_wakeupn(vm_object_t object, short i)
372 {
373
374         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
375         if (i)
376                 object->paging_in_progress -= i;
377         if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
378                 vm_object_clear_flag(object, OBJ_PIPWNT);
379                 wakeup(object);
380         }
381 }
382
383 void
384 vm_object_pip_wait(vm_object_t object, char *waitid)
385 {
386
387         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
388         while (object->paging_in_progress) {
389                 object->flags |= OBJ_PIPWNT;
390                 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
391         }
392 }
393
394 /*
395  *      vm_object_allocate:
396  *
397  *      Returns a new object with the given size.
398  */
399 vm_object_t
400 vm_object_allocate(objtype_t type, vm_pindex_t size)
401 {
402         vm_object_t object;
403
404         object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
405         _vm_object_allocate(type, size, object);
406         return (object);
407 }
408
409
410 /*
411  *      vm_object_reference:
412  *
413  *      Gets another reference to the given object.  Note: OBJ_DEAD
414  *      objects can be referenced during final cleaning.
415  */
416 void
417 vm_object_reference(vm_object_t object)
418 {
419         if (object == NULL)
420                 return;
421         VM_OBJECT_LOCK(object);
422         vm_object_reference_locked(object);
423         VM_OBJECT_UNLOCK(object);
424 }
425
426 /*
427  *      vm_object_reference_locked:
428  *
429  *      Gets another reference to the given object.
430  *
431  *      The object must be locked.
432  */
433 void
434 vm_object_reference_locked(vm_object_t object)
435 {
436         struct vnode *vp;
437
438         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
439         object->ref_count++;
440         if (object->type == OBJT_VNODE) {
441                 vp = object->handle;
442                 vref(vp);
443         }
444 }
445
446 /*
447  * Handle deallocating an object of type OBJT_VNODE.
448  */
449 static void
450 vm_object_vndeallocate(vm_object_t object)
451 {
452         struct vnode *vp = (struct vnode *) object->handle;
453
454         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
455         KASSERT(object->type == OBJT_VNODE,
456             ("vm_object_vndeallocate: not a vnode object"));
457         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
458 #ifdef INVARIANTS
459         if (object->ref_count == 0) {
460                 vprint("vm_object_vndeallocate", vp);
461                 panic("vm_object_vndeallocate: bad object reference count");
462         }
463 #endif
464
465         if (object->ref_count > 1) {
466                 object->ref_count--;
467                 VM_OBJECT_UNLOCK(object);
468                 /* vrele may need the vnode lock. */
469                 vrele(vp);
470         } else {
471                 vhold(vp);
472                 VM_OBJECT_UNLOCK(object);
473                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
474                 vdrop(vp);
475                 VM_OBJECT_LOCK(object);
476                 object->ref_count--;
477                 if (object->type == OBJT_DEAD) {
478                         VM_OBJECT_UNLOCK(object);
479                         VOP_UNLOCK(vp, 0);
480                 } else {
481                         if (object->ref_count == 0)
482                                 VOP_UNSET_TEXT(vp);
483                         VM_OBJECT_UNLOCK(object);
484                         vput(vp);
485                 }
486         }
487 }
488
489 /*
490  *      vm_object_deallocate:
491  *
492  *      Release a reference to the specified object,
493  *      gained either through a vm_object_allocate
494  *      or a vm_object_reference call.  When all references
495  *      are gone, storage associated with this object
496  *      may be relinquished.
497  *
498  *      No object may be locked.
499  */
500 void
501 vm_object_deallocate(vm_object_t object)
502 {
503         vm_object_t temp;
504
505         while (object != NULL) {
506                 VM_OBJECT_LOCK(object);
507                 if (object->type == OBJT_VNODE) {
508                         vm_object_vndeallocate(object);
509                         return;
510                 }
511
512                 KASSERT(object->ref_count != 0,
513                         ("vm_object_deallocate: object deallocated too many times: %d", object->type));
514
515                 /*
516                  * If the reference count goes to 0 we start calling
517                  * vm_object_terminate() on the object chain.
518                  * A ref count of 1 may be a special case depending on the
519                  * shadow count being 0 or 1.
520                  */
521                 object->ref_count--;
522                 if (object->ref_count > 1) {
523                         VM_OBJECT_UNLOCK(object);
524                         return;
525                 } else if (object->ref_count == 1) {
526                         if (object->shadow_count == 0 &&
527                             object->handle == NULL &&
528                             (object->type == OBJT_DEFAULT ||
529                              object->type == OBJT_SWAP)) {
530                                 vm_object_set_flag(object, OBJ_ONEMAPPING);
531                         } else if ((object->shadow_count == 1) &&
532                             (object->handle == NULL) &&
533                             (object->type == OBJT_DEFAULT ||
534                              object->type == OBJT_SWAP)) {
535                                 vm_object_t robject;
536
537                                 robject = LIST_FIRST(&object->shadow_head);
538                                 KASSERT(robject != NULL,
539                                     ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
540                                          object->ref_count,
541                                          object->shadow_count));
542                                 if (!VM_OBJECT_TRYLOCK(robject)) {
543                                         /*
544                                          * Avoid a potential deadlock.
545                                          */
546                                         object->ref_count++;
547                                         VM_OBJECT_UNLOCK(object);
548                                         /*
549                                          * More likely than not the thread
550                                          * holding robject's lock has lower
551                                          * priority than the current thread.
552                                          * Let the lower priority thread run.
553                                          */
554                                         pause("vmo_de", 1);
555                                         continue;
556                                 }
557                                 /*
558                                  * Collapse object into its shadow unless its
559                                  * shadow is dead.  In that case, object will
560                                  * be deallocated by the thread that is
561                                  * deallocating its shadow.
562                                  */
563                                 if ((robject->flags & OBJ_DEAD) == 0 &&
564                                     (robject->handle == NULL) &&
565                                     (robject->type == OBJT_DEFAULT ||
566                                      robject->type == OBJT_SWAP)) {
567
568                                         robject->ref_count++;
569 retry:
570                                         if (robject->paging_in_progress) {
571                                                 VM_OBJECT_UNLOCK(object);
572                                                 vm_object_pip_wait(robject,
573                                                     "objde1");
574                                                 temp = robject->backing_object;
575                                                 if (object == temp) {
576                                                         VM_OBJECT_LOCK(object);
577                                                         goto retry;
578                                                 }
579                                         } else if (object->paging_in_progress) {
580                                                 VM_OBJECT_UNLOCK(robject);
581                                                 object->flags |= OBJ_PIPWNT;
582                                                 VM_OBJECT_SLEEP(object, object,
583                                                     PDROP | PVM, "objde2", 0);
584                                                 VM_OBJECT_LOCK(robject);
585                                                 temp = robject->backing_object;
586                                                 if (object == temp) {
587                                                         VM_OBJECT_LOCK(object);
588                                                         goto retry;
589                                                 }
590                                         } else
591                                                 VM_OBJECT_UNLOCK(object);
592
593                                         if (robject->ref_count == 1) {
594                                                 robject->ref_count--;
595                                                 object = robject;
596                                                 goto doterm;
597                                         }
598                                         object = robject;
599                                         vm_object_collapse(object);
600                                         VM_OBJECT_UNLOCK(object);
601                                         continue;
602                                 }
603                                 VM_OBJECT_UNLOCK(robject);
604                         }
605                         VM_OBJECT_UNLOCK(object);
606                         return;
607                 }
608 doterm:
609                 temp = object->backing_object;
610                 if (temp != NULL) {
611                         VM_OBJECT_LOCK(temp);
612                         LIST_REMOVE(object, shadow_list);
613                         temp->shadow_count--;
614                         VM_OBJECT_UNLOCK(temp);
615                         object->backing_object = NULL;
616                 }
617                 /*
618                  * Don't double-terminate, we could be in a termination
619                  * recursion due to the terminate having to sync data
620                  * to disk.
621                  */
622                 if ((object->flags & OBJ_DEAD) == 0)
623                         vm_object_terminate(object);
624                 else
625                         VM_OBJECT_UNLOCK(object);
626                 object = temp;
627         }
628 }
629
630 /*
631  *      vm_object_destroy removes the object from the global object list
632  *      and frees the space for the object.
633  */
634 void
635 vm_object_destroy(vm_object_t object)
636 {
637
638         /*
639          * Remove the object from the global object list.
640          */
641         mtx_lock(&vm_object_list_mtx);
642         TAILQ_REMOVE(&vm_object_list, object, object_list);
643         mtx_unlock(&vm_object_list_mtx);
644
645         /*
646          * Release the allocation charge.
647          */
648         if (object->cred != NULL) {
649                 KASSERT(object->type == OBJT_DEFAULT ||
650                     object->type == OBJT_SWAP,
651                     ("vm_object_terminate: non-swap obj %p has cred",
652                      object));
653                 swap_release_by_cred(object->charge, object->cred);
654                 object->charge = 0;
655                 crfree(object->cred);
656                 object->cred = NULL;
657         }
658
659         /*
660          * Free the space for the object.
661          */
662         uma_zfree(obj_zone, object);
663 }
664
665 /*
666  *      vm_object_terminate actually destroys the specified object, freeing
667  *      up all previously used resources.
668  *
669  *      The object must be locked.
670  *      This routine may block.
671  */
672 void
673 vm_object_terminate(vm_object_t object)
674 {
675         vm_page_t p, p_next;
676
677         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
678
679         /*
680          * Make sure no one uses us.
681          */
682         vm_object_set_flag(object, OBJ_DEAD);
683
684         /*
685          * wait for the pageout daemon to be done with the object
686          */
687         vm_object_pip_wait(object, "objtrm");
688
689         KASSERT(!object->paging_in_progress,
690                 ("vm_object_terminate: pageout in progress"));
691
692         /*
693          * Clean and free the pages, as appropriate. All references to the
694          * object are gone, so we don't need to lock it.
695          */
696         if (object->type == OBJT_VNODE) {
697                 struct vnode *vp = (struct vnode *)object->handle;
698
699                 /*
700                  * Clean pages and flush buffers.
701                  */
702                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
703                 VM_OBJECT_UNLOCK(object);
704
705                 vinvalbuf(vp, V_SAVE, 0, 0);
706
707                 VM_OBJECT_LOCK(object);
708         }
709
710         KASSERT(object->ref_count == 0, 
711                 ("vm_object_terminate: object with references, ref_count=%d",
712                 object->ref_count));
713
714         /*
715          * Free any remaining pageable pages.  This also removes them from the
716          * paging queues.  However, don't free wired pages, just remove them
717          * from the object.  Rather than incrementally removing each page from
718          * the object, the page and object are reset to any empty state. 
719          */
720         TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
721                 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
722                     ("vm_object_terminate: freeing busy page %p", p));
723                 vm_page_lock(p);
724                 /*
725                  * Optimize the page's removal from the object by resetting
726                  * its "object" field.  Specifically, if the page is not
727                  * wired, then the effect of this assignment is that
728                  * vm_page_free()'s call to vm_page_remove() will return
729                  * immediately without modifying the page or the object.
730                  */ 
731                 p->object = NULL;
732                 if (p->wire_count == 0) {
733                         vm_page_free(p);
734                         PCPU_INC(cnt.v_pfree);
735                 }
736                 vm_page_unlock(p);
737         }
738         /*
739          * If the object contained any pages, then reset it to an empty state.
740          * None of the object's fields, including "resident_page_count", were
741          * modified by the preceding loop.
742          */
743         if (object->resident_page_count != 0) {
744                 object->root = NULL;
745                 TAILQ_INIT(&object->memq);
746                 object->resident_page_count = 0;
747                 if (object->type == OBJT_VNODE)
748                         vdrop(object->handle);
749         }
750
751 #if VM_NRESERVLEVEL > 0
752         if (__predict_false(!LIST_EMPTY(&object->rvq)))
753                 vm_reserv_break_all(object);
754 #endif
755         if (__predict_false(object->cache != NULL))
756                 vm_page_cache_free(object, 0, 0);
757
758         /*
759          * Let the pager know object is dead.
760          */
761         vm_pager_deallocate(object);
762         VM_OBJECT_UNLOCK(object);
763
764         vm_object_destroy(object);
765 }
766
767 /*
768  * Make the page read-only so that we can clear the object flags.  However, if
769  * this is a nosync mmap then the object is likely to stay dirty so do not
770  * mess with the page and do not clear the object flags.  Returns TRUE if the
771  * page should be flushed, and FALSE otherwise.
772  */
773 static boolean_t
774 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
775 {
776
777         /*
778          * If we have been asked to skip nosync pages and this is a
779          * nosync page, skip it.  Note that the object flags were not
780          * cleared in this case so we do not have to set them.
781          */
782         if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
783                 *clearobjflags = FALSE;
784                 return (FALSE);
785         } else {
786                 pmap_remove_write(p);
787                 return (p->dirty != 0);
788         }
789 }
790
791 /*
792  *      vm_object_page_clean
793  *
794  *      Clean all dirty pages in the specified range of object.  Leaves page 
795  *      on whatever queue it is currently on.   If NOSYNC is set then do not
796  *      write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
797  *      leaving the object dirty.
798  *
799  *      When stuffing pages asynchronously, allow clustering.  XXX we need a
800  *      synchronous clustering mode implementation.
801  *
802  *      Odd semantics: if start == end, we clean everything.
803  *
804  *      The object must be locked.
805  *
806  *      Returns FALSE if some page from the range was not written, as
807  *      reported by the pager, and TRUE otherwise.
808  */
809 boolean_t
810 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
811     int flags)
812 {
813         vm_page_t np, p;
814         vm_pindex_t pi, tend, tstart;
815         int curgeneration, n, pagerflags;
816         boolean_t clearobjflags, eio, res;
817
818         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
819         KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
820         if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
821             object->resident_page_count == 0)
822                 return (TRUE);
823
824         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
825             VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
826         pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
827
828         tstart = OFF_TO_IDX(start);
829         tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
830         clearobjflags = tstart == 0 && tend >= object->size;
831         res = TRUE;
832
833 rescan:
834         curgeneration = object->generation;
835
836         for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
837                 pi = p->pindex;
838                 if (pi >= tend)
839                         break;
840                 np = TAILQ_NEXT(p, listq);
841                 if (p->valid == 0)
842                         continue;
843                 if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
844                         if (object->generation != curgeneration) {
845                                 if ((flags & OBJPC_SYNC) != 0)
846                                         goto rescan;
847                                 else
848                                         clearobjflags = FALSE;
849                         }
850                         np = vm_page_find_least(object, pi);
851                         continue;
852                 }
853                 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
854                         continue;
855
856                 n = vm_object_page_collect_flush(object, p, pagerflags,
857                     flags, &clearobjflags, &eio);
858                 if (eio) {
859                         res = FALSE;
860                         clearobjflags = FALSE;
861                 }
862                 if (object->generation != curgeneration) {
863                         if ((flags & OBJPC_SYNC) != 0)
864                                 goto rescan;
865                         else
866                                 clearobjflags = FALSE;
867                 }
868
869                 /*
870                  * If the VOP_PUTPAGES() did a truncated write, so
871                  * that even the first page of the run is not fully
872                  * written, vm_pageout_flush() returns 0 as the run
873                  * length.  Since the condition that caused truncated
874                  * write may be permanent, e.g. exhausted free space,
875                  * accepting n == 0 would cause an infinite loop.
876                  *
877                  * Forwarding the iterator leaves the unwritten page
878                  * behind, but there is not much we can do there if
879                  * filesystem refuses to write it.
880                  */
881                 if (n == 0) {
882                         n = 1;
883                         clearobjflags = FALSE;
884                 }
885                 np = vm_page_find_least(object, pi + n);
886         }
887 #if 0
888         VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
889 #endif
890
891         if (clearobjflags)
892                 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
893         return (res);
894 }
895
896 static int
897 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
898     int flags, boolean_t *clearobjflags, boolean_t *eio)
899 {
900         vm_page_t ma[vm_pageout_page_count], p_first, tp;
901         int count, i, mreq, runlen;
902
903         vm_page_lock_assert(p, MA_NOTOWNED);
904         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
905
906         count = 1;
907         mreq = 0;
908
909         for (tp = p; count < vm_pageout_page_count; count++) {
910                 tp = vm_page_next(tp);
911                 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
912                         break;
913                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
914                         break;
915         }
916
917         for (p_first = p; count < vm_pageout_page_count; count++) {
918                 tp = vm_page_prev(p_first);
919                 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
920                         break;
921                 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
922                         break;
923                 p_first = tp;
924                 mreq++;
925         }
926
927         for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
928                 ma[i] = tp;
929
930         vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
931         return (runlen);
932 }
933
934 /*
935  * Note that there is absolutely no sense in writing out
936  * anonymous objects, so we track down the vnode object
937  * to write out.
938  * We invalidate (remove) all pages from the address space
939  * for semantic correctness.
940  *
941  * If the backing object is a device object with unmanaged pages, then any
942  * mappings to the specified range of pages must be removed before this
943  * function is called.
944  *
945  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
946  * may start out with a NULL object.
947  */
948 boolean_t
949 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
950     boolean_t syncio, boolean_t invalidate)
951 {
952         vm_object_t backing_object;
953         struct vnode *vp;
954         struct mount *mp;
955         int error, flags, fsync_after;
956         boolean_t res;
957
958         if (object == NULL)
959                 return (TRUE);
960         res = TRUE;
961         error = 0;
962         VM_OBJECT_LOCK(object);
963         while ((backing_object = object->backing_object) != NULL) {
964                 VM_OBJECT_LOCK(backing_object);
965                 offset += object->backing_object_offset;
966                 VM_OBJECT_UNLOCK(object);
967                 object = backing_object;
968                 if (object->size < OFF_TO_IDX(offset + size))
969                         size = IDX_TO_OFF(object->size) - offset;
970         }
971         /*
972          * Flush pages if writing is allowed, invalidate them
973          * if invalidation requested.  Pages undergoing I/O
974          * will be ignored by vm_object_page_remove().
975          *
976          * We cannot lock the vnode and then wait for paging
977          * to complete without deadlocking against vm_fault.
978          * Instead we simply call vm_object_page_remove() and
979          * allow it to block internally on a page-by-page
980          * basis when it encounters pages undergoing async
981          * I/O.
982          */
983         if (object->type == OBJT_VNODE &&
984             (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
985                 vp = object->handle;
986                 VM_OBJECT_UNLOCK(object);
987                 (void) vn_start_write(vp, &mp, V_WAIT);
988                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
989                 if (syncio && !invalidate && offset == 0 &&
990                     OFF_TO_IDX(size) == object->size) {
991                         /*
992                          * If syncing the whole mapping of the file,
993                          * it is faster to schedule all the writes in
994                          * async mode, also allowing the clustering,
995                          * and then wait for i/o to complete.
996                          */
997                         flags = 0;
998                         fsync_after = TRUE;
999                 } else {
1000                         flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1001                         flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1002                         fsync_after = FALSE;
1003                 }
1004                 VM_OBJECT_LOCK(object);
1005                 res = vm_object_page_clean(object, offset, offset + size,
1006                     flags);
1007                 VM_OBJECT_UNLOCK(object);
1008                 if (fsync_after)
1009                         error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1010                 VOP_UNLOCK(vp, 0);
1011                 vn_finished_write(mp);
1012                 if (error != 0)
1013                         res = FALSE;
1014                 VM_OBJECT_LOCK(object);
1015         }
1016         if ((object->type == OBJT_VNODE ||
1017              object->type == OBJT_DEVICE) && invalidate) {
1018                 if (object->type == OBJT_DEVICE)
1019                         /*
1020                          * The option OBJPR_NOTMAPPED must be passed here
1021                          * because vm_object_page_remove() cannot remove
1022                          * unmanaged mappings.
1023                          */
1024                         flags = OBJPR_NOTMAPPED;
1025                 else if (old_msync)
1026                         flags = 0;
1027                 else
1028                         flags = OBJPR_CLEANONLY;
1029                 vm_object_page_remove(object, OFF_TO_IDX(offset),
1030                     OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1031         }
1032         VM_OBJECT_UNLOCK(object);
1033         return (res);
1034 }
1035
1036 /*
1037  *      vm_object_madvise:
1038  *
1039  *      Implements the madvise function at the object/page level.
1040  *
1041  *      MADV_WILLNEED   (any object)
1042  *
1043  *          Activate the specified pages if they are resident.
1044  *
1045  *      MADV_DONTNEED   (any object)
1046  *
1047  *          Deactivate the specified pages if they are resident.
1048  *
1049  *      MADV_FREE       (OBJT_DEFAULT/OBJT_SWAP objects,
1050  *                       OBJ_ONEMAPPING only)
1051  *
1052  *          Deactivate and clean the specified pages if they are
1053  *          resident.  This permits the process to reuse the pages
1054  *          without faulting or the kernel to reclaim the pages
1055  *          without I/O.
1056  */
1057 void
1058 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1059     int advise)
1060 {
1061         vm_pindex_t tpindex;
1062         vm_object_t backing_object, tobject;
1063         vm_page_t m;
1064
1065         if (object == NULL)
1066                 return;
1067         VM_OBJECT_LOCK(object);
1068         /*
1069          * Locate and adjust resident pages
1070          */
1071         for (; pindex < end; pindex += 1) {
1072 relookup:
1073                 tobject = object;
1074                 tpindex = pindex;
1075 shadowlookup:
1076                 /*
1077                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1078                  * and those pages must be OBJ_ONEMAPPING.
1079                  */
1080                 if (advise == MADV_FREE) {
1081                         if ((tobject->type != OBJT_DEFAULT &&
1082                              tobject->type != OBJT_SWAP) ||
1083                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
1084                                 goto unlock_tobject;
1085                         }
1086                 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1087                         goto unlock_tobject;
1088                 m = vm_page_lookup(tobject, tpindex);
1089                 if (m == NULL && advise == MADV_WILLNEED) {
1090                         /*
1091                          * If the page is cached, reactivate it.
1092                          */
1093                         m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1094                             VM_ALLOC_NOBUSY);
1095                 }
1096                 if (m == NULL) {
1097                         /*
1098                          * There may be swap even if there is no backing page
1099                          */
1100                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1101                                 swap_pager_freespace(tobject, tpindex, 1);
1102                         /*
1103                          * next object
1104                          */
1105                         backing_object = tobject->backing_object;
1106                         if (backing_object == NULL)
1107                                 goto unlock_tobject;
1108                         VM_OBJECT_LOCK(backing_object);
1109                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1110                         if (tobject != object)
1111                                 VM_OBJECT_UNLOCK(tobject);
1112                         tobject = backing_object;
1113                         goto shadowlookup;
1114                 } else if (m->valid != VM_PAGE_BITS_ALL)
1115                         goto unlock_tobject;
1116                 /*
1117                  * If the page is not in a normal state, skip it.
1118                  */
1119                 vm_page_lock(m);
1120                 if (m->hold_count != 0 || m->wire_count != 0) {
1121                         vm_page_unlock(m);
1122                         goto unlock_tobject;
1123                 }
1124                 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1125                     ("vm_object_madvise: page %p is fictitious", m));
1126                 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1127                     ("vm_object_madvise: page %p is not managed", m));
1128                 if ((m->oflags & VPO_BUSY) || m->busy) {
1129                         if (advise == MADV_WILLNEED) {
1130                                 /*
1131                                  * Reference the page before unlocking and
1132                                  * sleeping so that the page daemon is less
1133                                  * likely to reclaim it. 
1134                                  */
1135                                 vm_page_aflag_set(m, PGA_REFERENCED);
1136                         }
1137                         vm_page_unlock(m);
1138                         if (object != tobject)
1139                                 VM_OBJECT_UNLOCK(object);
1140                         m->oflags |= VPO_WANTED;
1141                         VM_OBJECT_SLEEP(tobject, m, PDROP | PVM, "madvpo", 0);
1142                         VM_OBJECT_LOCK(object);
1143                         goto relookup;
1144                 }
1145                 if (advise == MADV_WILLNEED) {
1146                         vm_page_activate(m);
1147                 } else if (advise == MADV_DONTNEED) {
1148                         vm_page_dontneed(m);
1149                 } else if (advise == MADV_FREE) {
1150                         /*
1151                          * Mark the page clean.  This will allow the page
1152                          * to be freed up by the system.  However, such pages
1153                          * are often reused quickly by malloc()/free()
1154                          * so we do not do anything that would cause
1155                          * a page fault if we can help it.
1156                          *
1157                          * Specifically, we do not try to actually free
1158                          * the page now nor do we try to put it in the
1159                          * cache (which would cause a page fault on reuse).
1160                          *
1161                          * But we do make the page is freeable as we
1162                          * can without actually taking the step of unmapping
1163                          * it.
1164                          */
1165                         pmap_clear_modify(m);
1166                         m->dirty = 0;
1167                         m->act_count = 0;
1168                         vm_page_dontneed(m);
1169                 }
1170                 vm_page_unlock(m);
1171                 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1172                         swap_pager_freespace(tobject, tpindex, 1);
1173 unlock_tobject:
1174                 if (tobject != object)
1175                         VM_OBJECT_UNLOCK(tobject);
1176         }       
1177         VM_OBJECT_UNLOCK(object);
1178 }
1179
1180 /*
1181  *      vm_object_shadow:
1182  *
1183  *      Create a new object which is backed by the
1184  *      specified existing object range.  The source
1185  *      object reference is deallocated.
1186  *
1187  *      The new object and offset into that object
1188  *      are returned in the source parameters.
1189  */
1190 void
1191 vm_object_shadow(
1192         vm_object_t *object,    /* IN/OUT */
1193         vm_ooffset_t *offset,   /* IN/OUT */
1194         vm_size_t length)
1195 {
1196         vm_object_t source;
1197         vm_object_t result;
1198
1199         source = *object;
1200
1201         /*
1202          * Don't create the new object if the old object isn't shared.
1203          */
1204         if (source != NULL) {
1205                 VM_OBJECT_LOCK(source);
1206                 if (source->ref_count == 1 &&
1207                     source->handle == NULL &&
1208                     (source->type == OBJT_DEFAULT ||
1209                      source->type == OBJT_SWAP)) {
1210                         VM_OBJECT_UNLOCK(source);
1211                         return;
1212                 }
1213                 VM_OBJECT_UNLOCK(source);
1214         }
1215
1216         /*
1217          * Allocate a new object with the given length.
1218          */
1219         result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1220
1221         /*
1222          * The new object shadows the source object, adding a reference to it.
1223          * Our caller changes his reference to point to the new object,
1224          * removing a reference to the source object.  Net result: no change
1225          * of reference count.
1226          *
1227          * Try to optimize the result object's page color when shadowing
1228          * in order to maintain page coloring consistency in the combined 
1229          * shadowed object.
1230          */
1231         result->backing_object = source;
1232         /*
1233          * Store the offset into the source object, and fix up the offset into
1234          * the new object.
1235          */
1236         result->backing_object_offset = *offset;
1237         if (source != NULL) {
1238                 VM_OBJECT_LOCK(source);
1239                 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1240                 source->shadow_count++;
1241 #if VM_NRESERVLEVEL > 0
1242                 result->flags |= source->flags & OBJ_COLORED;
1243                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1244                     ((1 << (VM_NFREEORDER - 1)) - 1);
1245 #endif
1246                 VM_OBJECT_UNLOCK(source);
1247         }
1248
1249
1250         /*
1251          * Return the new things
1252          */
1253         *offset = 0;
1254         *object = result;
1255 }
1256
1257 /*
1258  *      vm_object_split:
1259  *
1260  * Split the pages in a map entry into a new object.  This affords
1261  * easier removal of unused pages, and keeps object inheritance from
1262  * being a negative impact on memory usage.
1263  */
1264 void
1265 vm_object_split(vm_map_entry_t entry)
1266 {
1267         vm_page_t m, m_next;
1268         vm_object_t orig_object, new_object, source;
1269         vm_pindex_t idx, offidxstart;
1270         vm_size_t size;
1271
1272         orig_object = entry->object.vm_object;
1273         if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1274                 return;
1275         if (orig_object->ref_count <= 1)
1276                 return;
1277         VM_OBJECT_UNLOCK(orig_object);
1278
1279         offidxstart = OFF_TO_IDX(entry->offset);
1280         size = atop(entry->end - entry->start);
1281
1282         /*
1283          * If swap_pager_copy() is later called, it will convert new_object
1284          * into a swap object.
1285          */
1286         new_object = vm_object_allocate(OBJT_DEFAULT, size);
1287
1288         /*
1289          * At this point, the new object is still private, so the order in
1290          * which the original and new objects are locked does not matter.
1291          */
1292         VM_OBJECT_LOCK(new_object);
1293         VM_OBJECT_LOCK(orig_object);
1294         source = orig_object->backing_object;
1295         if (source != NULL) {
1296                 VM_OBJECT_LOCK(source);
1297                 if ((source->flags & OBJ_DEAD) != 0) {
1298                         VM_OBJECT_UNLOCK(source);
1299                         VM_OBJECT_UNLOCK(orig_object);
1300                         VM_OBJECT_UNLOCK(new_object);
1301                         vm_object_deallocate(new_object);
1302                         VM_OBJECT_LOCK(orig_object);
1303                         return;
1304                 }
1305                 LIST_INSERT_HEAD(&source->shadow_head,
1306                                   new_object, shadow_list);
1307                 source->shadow_count++;
1308                 vm_object_reference_locked(source);     /* for new_object */
1309                 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1310                 VM_OBJECT_UNLOCK(source);
1311                 new_object->backing_object_offset = 
1312                         orig_object->backing_object_offset + entry->offset;
1313                 new_object->backing_object = source;
1314         }
1315         if (orig_object->cred != NULL) {
1316                 new_object->cred = orig_object->cred;
1317                 crhold(orig_object->cred);
1318                 new_object->charge = ptoa(size);
1319                 KASSERT(orig_object->charge >= ptoa(size),
1320                     ("orig_object->charge < 0"));
1321                 orig_object->charge -= ptoa(size);
1322         }
1323 retry:
1324         m = vm_page_find_least(orig_object, offidxstart);
1325         for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1326             m = m_next) {
1327                 m_next = TAILQ_NEXT(m, listq);
1328
1329                 /*
1330                  * We must wait for pending I/O to complete before we can
1331                  * rename the page.
1332                  *
1333                  * We do not have to VM_PROT_NONE the page as mappings should
1334                  * not be changed by this operation.
1335                  */
1336                 if ((m->oflags & VPO_BUSY) || m->busy) {
1337                         VM_OBJECT_UNLOCK(new_object);
1338                         m->oflags |= VPO_WANTED;
1339                         VM_OBJECT_SLEEP(orig_object, m, PVM, "spltwt", 0);
1340                         VM_OBJECT_LOCK(new_object);
1341                         goto retry;
1342                 }
1343 #if VM_NRESERVLEVEL > 0
1344                 /*
1345                  * If some of the reservation's allocated pages remain with
1346                  * the original object, then transferring the reservation to
1347                  * the new object is neither particularly beneficial nor
1348                  * particularly harmful as compared to leaving the reservation
1349                  * with the original object.  If, however, all of the
1350                  * reservation's allocated pages are transferred to the new
1351                  * object, then transferring the reservation is typically
1352                  * beneficial.  Determining which of these two cases applies
1353                  * would be more costly than unconditionally renaming the
1354                  * reservation.
1355                  */
1356                 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1357 #endif
1358                 vm_page_lock(m);
1359                 vm_page_rename(m, new_object, idx);
1360                 vm_page_unlock(m);
1361                 /* page automatically made dirty by rename and cache handled */
1362                 vm_page_busy(m);
1363         }
1364         if (orig_object->type == OBJT_SWAP) {
1365                 /*
1366                  * swap_pager_copy() can sleep, in which case the orig_object's
1367                  * and new_object's locks are released and reacquired. 
1368                  */
1369                 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1370
1371                 /*
1372                  * Transfer any cached pages from orig_object to new_object.
1373                  * If swap_pager_copy() found swapped out pages within the
1374                  * specified range of orig_object, then it changed
1375                  * new_object's type to OBJT_SWAP when it transferred those
1376                  * pages to new_object.  Otherwise, new_object's type
1377                  * should still be OBJT_DEFAULT and orig_object should not
1378                  * contain any cached pages within the specified range.
1379                  */
1380                 if (__predict_false(orig_object->cache != NULL))
1381                         vm_page_cache_transfer(orig_object, offidxstart,
1382                             new_object);
1383         }
1384         VM_OBJECT_UNLOCK(orig_object);
1385         TAILQ_FOREACH(m, &new_object->memq, listq)
1386                 vm_page_wakeup(m);
1387         VM_OBJECT_UNLOCK(new_object);
1388         entry->object.vm_object = new_object;
1389         entry->offset = 0LL;
1390         vm_object_deallocate(orig_object);
1391         VM_OBJECT_LOCK(new_object);
1392 }
1393
1394 #define OBSC_TEST_ALL_SHADOWED  0x0001
1395 #define OBSC_COLLAPSE_NOWAIT    0x0002
1396 #define OBSC_COLLAPSE_WAIT      0x0004
1397
1398 static int
1399 vm_object_backing_scan(vm_object_t object, int op)
1400 {
1401         int r = 1;
1402         vm_page_t p;
1403         vm_object_t backing_object;
1404         vm_pindex_t backing_offset_index;
1405
1406         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1407         VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1408
1409         backing_object = object->backing_object;
1410         backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1411
1412         /*
1413          * Initial conditions
1414          */
1415         if (op & OBSC_TEST_ALL_SHADOWED) {
1416                 /*
1417                  * We do not want to have to test for the existence of cache
1418                  * or swap pages in the backing object.  XXX but with the
1419                  * new swapper this would be pretty easy to do.
1420                  *
1421                  * XXX what about anonymous MAP_SHARED memory that hasn't
1422                  * been ZFOD faulted yet?  If we do not test for this, the
1423                  * shadow test may succeed! XXX
1424                  */
1425                 if (backing_object->type != OBJT_DEFAULT) {
1426                         return (0);
1427                 }
1428         }
1429         if (op & OBSC_COLLAPSE_WAIT) {
1430                 vm_object_set_flag(backing_object, OBJ_DEAD);
1431         }
1432
1433         /*
1434          * Our scan
1435          */
1436         p = TAILQ_FIRST(&backing_object->memq);
1437         while (p) {
1438                 vm_page_t next = TAILQ_NEXT(p, listq);
1439                 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1440
1441                 if (op & OBSC_TEST_ALL_SHADOWED) {
1442                         vm_page_t pp;
1443
1444                         /*
1445                          * Ignore pages outside the parent object's range
1446                          * and outside the parent object's mapping of the 
1447                          * backing object.
1448                          *
1449                          * note that we do not busy the backing object's
1450                          * page.
1451                          */
1452                         if (
1453                             p->pindex < backing_offset_index ||
1454                             new_pindex >= object->size
1455                         ) {
1456                                 p = next;
1457                                 continue;
1458                         }
1459
1460                         /*
1461                          * See if the parent has the page or if the parent's
1462                          * object pager has the page.  If the parent has the
1463                          * page but the page is not valid, the parent's
1464                          * object pager must have the page.
1465                          *
1466                          * If this fails, the parent does not completely shadow
1467                          * the object and we might as well give up now.
1468                          */
1469
1470                         pp = vm_page_lookup(object, new_pindex);
1471                         if (
1472                             (pp == NULL || pp->valid == 0) &&
1473                             !vm_pager_has_page(object, new_pindex, NULL, NULL)
1474                         ) {
1475                                 r = 0;
1476                                 break;
1477                         }
1478                 }
1479
1480                 /*
1481                  * Check for busy page
1482                  */
1483                 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1484                         vm_page_t pp;
1485
1486                         if (op & OBSC_COLLAPSE_NOWAIT) {
1487                                 if ((p->oflags & VPO_BUSY) ||
1488                                     !p->valid || 
1489                                     p->busy) {
1490                                         p = next;
1491                                         continue;
1492                                 }
1493                         } else if (op & OBSC_COLLAPSE_WAIT) {
1494                                 if ((p->oflags & VPO_BUSY) || p->busy) {
1495                                         VM_OBJECT_UNLOCK(object);
1496                                         p->oflags |= VPO_WANTED;
1497                                         VM_OBJECT_SLEEP(backing_object, p,
1498                                             PDROP | PVM, "vmocol", 0);
1499                                         VM_OBJECT_LOCK(object);
1500                                         VM_OBJECT_LOCK(backing_object);
1501                                         /*
1502                                          * If we slept, anything could have
1503                                          * happened.  Since the object is
1504                                          * marked dead, the backing offset
1505                                          * should not have changed so we
1506                                          * just restart our scan.
1507                                          */
1508                                         p = TAILQ_FIRST(&backing_object->memq);
1509                                         continue;
1510                                 }
1511                         }
1512
1513                         KASSERT(
1514                             p->object == backing_object,
1515                             ("vm_object_backing_scan: object mismatch")
1516                         );
1517
1518                         /*
1519                          * Destroy any associated swap
1520                          */
1521                         if (backing_object->type == OBJT_SWAP) {
1522                                 swap_pager_freespace(
1523                                     backing_object, 
1524                                     p->pindex,
1525                                     1
1526                                 );
1527                         }
1528
1529                         if (
1530                             p->pindex < backing_offset_index ||
1531                             new_pindex >= object->size
1532                         ) {
1533                                 /*
1534                                  * Page is out of the parent object's range, we 
1535                                  * can simply destroy it. 
1536                                  */
1537                                 vm_page_lock(p);
1538                                 KASSERT(!pmap_page_is_mapped(p),
1539                                     ("freeing mapped page %p", p));
1540                                 if (p->wire_count == 0)
1541                                         vm_page_free(p);
1542                                 else
1543                                         vm_page_remove(p);
1544                                 vm_page_unlock(p);
1545                                 p = next;
1546                                 continue;
1547                         }
1548
1549                         pp = vm_page_lookup(object, new_pindex);
1550                         if (
1551                             (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1552                             (pp != NULL && pp->valid == 0)
1553                         ) {
1554                                 /*
1555                                  * The page in the parent is not (yet) valid.
1556                                  * We don't know anything about the state of
1557                                  * the original page.  It might be mapped,
1558                                  * so we must avoid the next if here.
1559                                  *
1560                                  * This is due to a race in vm_fault() where
1561                                  * we must unbusy the original (backing_obj)
1562                                  * page before we can (re)lock the parent.
1563                                  * Hence we can get here.
1564                                  */
1565                                 p = next;
1566                                 continue;
1567                         }
1568                         if (
1569                             pp != NULL ||
1570                             vm_pager_has_page(object, new_pindex, NULL, NULL)
1571                         ) {
1572                                 /*
1573                                  * page already exists in parent OR swap exists
1574                                  * for this location in the parent.  Destroy 
1575                                  * the original page from the backing object.
1576                                  *
1577                                  * Leave the parent's page alone
1578                                  */
1579                                 vm_page_lock(p);
1580                                 KASSERT(!pmap_page_is_mapped(p),
1581                                     ("freeing mapped page %p", p));
1582                                 if (p->wire_count == 0)
1583                                         vm_page_free(p);
1584                                 else
1585                                         vm_page_remove(p);
1586                                 vm_page_unlock(p);
1587                                 p = next;
1588                                 continue;
1589                         }
1590
1591 #if VM_NRESERVLEVEL > 0
1592                         /*
1593                          * Rename the reservation.
1594                          */
1595                         vm_reserv_rename(p, object, backing_object,
1596                             backing_offset_index);
1597 #endif
1598
1599                         /*
1600                          * Page does not exist in parent, rename the
1601                          * page from the backing object to the main object. 
1602                          *
1603                          * If the page was mapped to a process, it can remain 
1604                          * mapped through the rename.
1605                          */
1606                         vm_page_lock(p);
1607                         vm_page_rename(p, object, new_pindex);
1608                         vm_page_unlock(p);
1609                         /* page automatically made dirty by rename */
1610                 }
1611                 p = next;
1612         }
1613         return (r);
1614 }
1615
1616
1617 /*
1618  * this version of collapse allows the operation to occur earlier and
1619  * when paging_in_progress is true for an object...  This is not a complete
1620  * operation, but should plug 99.9% of the rest of the leaks.
1621  */
1622 static void
1623 vm_object_qcollapse(vm_object_t object)
1624 {
1625         vm_object_t backing_object = object->backing_object;
1626
1627         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1628         VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1629
1630         if (backing_object->ref_count != 1)
1631                 return;
1632
1633         vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1634 }
1635
1636 /*
1637  *      vm_object_collapse:
1638  *
1639  *      Collapse an object with the object backing it.
1640  *      Pages in the backing object are moved into the
1641  *      parent, and the backing object is deallocated.
1642  */
1643 void
1644 vm_object_collapse(vm_object_t object)
1645 {
1646         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1647         
1648         while (TRUE) {
1649                 vm_object_t backing_object;
1650
1651                 /*
1652                  * Verify that the conditions are right for collapse:
1653                  *
1654                  * The object exists and the backing object exists.
1655                  */
1656                 if ((backing_object = object->backing_object) == NULL)
1657                         break;
1658
1659                 /*
1660                  * we check the backing object first, because it is most likely
1661                  * not collapsable.
1662                  */
1663                 VM_OBJECT_LOCK(backing_object);
1664                 if (backing_object->handle != NULL ||
1665                     (backing_object->type != OBJT_DEFAULT &&
1666                      backing_object->type != OBJT_SWAP) ||
1667                     (backing_object->flags & OBJ_DEAD) ||
1668                     object->handle != NULL ||
1669                     (object->type != OBJT_DEFAULT &&
1670                      object->type != OBJT_SWAP) ||
1671                     (object->flags & OBJ_DEAD)) {
1672                         VM_OBJECT_UNLOCK(backing_object);
1673                         break;
1674                 }
1675
1676                 if (
1677                     object->paging_in_progress != 0 ||
1678                     backing_object->paging_in_progress != 0
1679                 ) {
1680                         vm_object_qcollapse(object);
1681                         VM_OBJECT_UNLOCK(backing_object);
1682                         break;
1683                 }
1684                 /*
1685                  * We know that we can either collapse the backing object (if
1686                  * the parent is the only reference to it) or (perhaps) have
1687                  * the parent bypass the object if the parent happens to shadow
1688                  * all the resident pages in the entire backing object.
1689                  *
1690                  * This is ignoring pager-backed pages such as swap pages.
1691                  * vm_object_backing_scan fails the shadowing test in this
1692                  * case.
1693                  */
1694                 if (backing_object->ref_count == 1) {
1695                         /*
1696                          * If there is exactly one reference to the backing
1697                          * object, we can collapse it into the parent.  
1698                          */
1699                         vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1700
1701 #if VM_NRESERVLEVEL > 0
1702                         /*
1703                          * Break any reservations from backing_object.
1704                          */
1705                         if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1706                                 vm_reserv_break_all(backing_object);
1707 #endif
1708
1709                         /*
1710                          * Move the pager from backing_object to object.
1711                          */
1712                         if (backing_object->type == OBJT_SWAP) {
1713                                 /*
1714                                  * swap_pager_copy() can sleep, in which case
1715                                  * the backing_object's and object's locks are
1716                                  * released and reacquired.
1717                                  * Since swap_pager_copy() is being asked to
1718                                  * destroy the source, it will change the
1719                                  * backing_object's type to OBJT_DEFAULT.
1720                                  */
1721                                 swap_pager_copy(
1722                                     backing_object,
1723                                     object,
1724                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1725
1726                                 /*
1727                                  * Free any cached pages from backing_object.
1728                                  */
1729                                 if (__predict_false(backing_object->cache != NULL))
1730                                         vm_page_cache_free(backing_object, 0, 0);
1731                         }
1732                         /*
1733                          * Object now shadows whatever backing_object did.
1734                          * Note that the reference to 
1735                          * backing_object->backing_object moves from within 
1736                          * backing_object to within object.
1737                          */
1738                         LIST_REMOVE(object, shadow_list);
1739                         backing_object->shadow_count--;
1740                         if (backing_object->backing_object) {
1741                                 VM_OBJECT_LOCK(backing_object->backing_object);
1742                                 LIST_REMOVE(backing_object, shadow_list);
1743                                 LIST_INSERT_HEAD(
1744                                     &backing_object->backing_object->shadow_head,
1745                                     object, shadow_list);
1746                                 /*
1747                                  * The shadow_count has not changed.
1748                                  */
1749                                 VM_OBJECT_UNLOCK(backing_object->backing_object);
1750                         }
1751                         object->backing_object = backing_object->backing_object;
1752                         object->backing_object_offset +=
1753                             backing_object->backing_object_offset;
1754
1755                         /*
1756                          * Discard backing_object.
1757                          *
1758                          * Since the backing object has no pages, no pager left,
1759                          * and no object references within it, all that is
1760                          * necessary is to dispose of it.
1761                          */
1762                         KASSERT(backing_object->ref_count == 1, (
1763 "backing_object %p was somehow re-referenced during collapse!",
1764                             backing_object));
1765                         VM_OBJECT_UNLOCK(backing_object);
1766                         vm_object_destroy(backing_object);
1767
1768                         object_collapses++;
1769                 } else {
1770                         vm_object_t new_backing_object;
1771
1772                         /*
1773                          * If we do not entirely shadow the backing object,
1774                          * there is nothing we can do so we give up.
1775                          */
1776                         if (object->resident_page_count != object->size &&
1777                             vm_object_backing_scan(object,
1778                             OBSC_TEST_ALL_SHADOWED) == 0) {
1779                                 VM_OBJECT_UNLOCK(backing_object);
1780                                 break;
1781                         }
1782
1783                         /*
1784                          * Make the parent shadow the next object in the
1785                          * chain.  Deallocating backing_object will not remove
1786                          * it, since its reference count is at least 2.
1787                          */
1788                         LIST_REMOVE(object, shadow_list);
1789                         backing_object->shadow_count--;
1790
1791                         new_backing_object = backing_object->backing_object;
1792                         if ((object->backing_object = new_backing_object) != NULL) {
1793                                 VM_OBJECT_LOCK(new_backing_object);
1794                                 LIST_INSERT_HEAD(
1795                                     &new_backing_object->shadow_head,
1796                                     object,
1797                                     shadow_list
1798                                 );
1799                                 new_backing_object->shadow_count++;
1800                                 vm_object_reference_locked(new_backing_object);
1801                                 VM_OBJECT_UNLOCK(new_backing_object);
1802                                 object->backing_object_offset +=
1803                                         backing_object->backing_object_offset;
1804                         }
1805
1806                         /*
1807                          * Drop the reference count on backing_object. Since
1808                          * its ref_count was at least 2, it will not vanish.
1809                          */
1810                         backing_object->ref_count--;
1811                         VM_OBJECT_UNLOCK(backing_object);
1812                         object_bypasses++;
1813                 }
1814
1815                 /*
1816                  * Try again with this object's new backing object.
1817                  */
1818         }
1819 }
1820
1821 /*
1822  *      vm_object_page_remove:
1823  *
1824  *      For the given object, either frees or invalidates each of the
1825  *      specified pages.  In general, a page is freed.  However, if a page is
1826  *      wired for any reason other than the existence of a managed, wired
1827  *      mapping, then it may be invalidated but not removed from the object.
1828  *      Pages are specified by the given range ["start", "end") and the option
1829  *      OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
1830  *      extends from "start" to the end of the object.  If the option
1831  *      OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1832  *      specified range are affected.  If the option OBJPR_NOTMAPPED is
1833  *      specified, then the pages within the specified range must have no
1834  *      mappings.  Otherwise, if this option is not specified, any mappings to
1835  *      the specified pages are removed before the pages are freed or
1836  *      invalidated.
1837  *
1838  *      In general, this operation should only be performed on objects that
1839  *      contain managed pages.  There are, however, two exceptions.  First, it
1840  *      is performed on the kernel and kmem objects by vm_map_entry_delete().
1841  *      Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1842  *      backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
1843  *      not be specified and the option OBJPR_NOTMAPPED must be specified.
1844  *
1845  *      The object must be locked.
1846  */
1847 void
1848 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1849     int options)
1850 {
1851         vm_page_t p, next;
1852         int wirings;
1853
1854         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1855         KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1856             (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1857             ("vm_object_page_remove: illegal options for object %p", object));
1858         if (object->resident_page_count == 0)
1859                 goto skipmemq;
1860         vm_object_pip_add(object, 1);
1861 again:
1862         p = vm_page_find_least(object, start);
1863
1864         /*
1865          * Here, the variable "p" is either (1) the page with the least pindex
1866          * greater than or equal to the parameter "start" or (2) NULL. 
1867          */
1868         for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1869                 next = TAILQ_NEXT(p, listq);
1870
1871                 /*
1872                  * If the page is wired for any reason besides the existence
1873                  * of managed, wired mappings, then it cannot be freed.  For
1874                  * example, fictitious pages, which represent device memory,
1875                  * are inherently wired and cannot be freed.  They can,
1876                  * however, be invalidated if the option OBJPR_CLEANONLY is
1877                  * not specified.
1878                  */
1879                 vm_page_lock(p);
1880                 if ((wirings = p->wire_count) != 0 &&
1881                     (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1882                         if ((options & OBJPR_NOTMAPPED) == 0) {
1883                                 pmap_remove_all(p);
1884                                 /* Account for removal of wired mappings. */
1885                                 if (wirings != 0)
1886                                         p->wire_count -= wirings;
1887                         }
1888                         if ((options & OBJPR_CLEANONLY) == 0) {
1889                                 p->valid = 0;
1890                                 vm_page_undirty(p);
1891                         }
1892                         vm_page_unlock(p);
1893                         continue;
1894                 }
1895                 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1896                         goto again;
1897                 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1898                     ("vm_object_page_remove: page %p is fictitious", p));
1899                 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1900                         if ((options & OBJPR_NOTMAPPED) == 0)
1901                                 pmap_remove_write(p);
1902                         if (p->dirty) {
1903                                 vm_page_unlock(p);
1904                                 continue;
1905                         }
1906                 }
1907                 if ((options & OBJPR_NOTMAPPED) == 0) {
1908                         pmap_remove_all(p);
1909                         /* Account for removal of wired mappings. */
1910                         if (wirings != 0) {
1911                                 KASSERT(p->wire_count == wirings,
1912                                     ("inconsistent wire count %d %d %p",
1913                                     p->wire_count, wirings, p));
1914                                 p->wire_count = 0;
1915                                 atomic_subtract_int(&cnt.v_wire_count, 1);
1916                         }
1917                 }
1918                 vm_page_free(p);
1919                 vm_page_unlock(p);
1920         }
1921         vm_object_pip_wakeup(object);
1922 skipmemq:
1923         if (__predict_false(object->cache != NULL))
1924                 vm_page_cache_free(object, start, end);
1925 }
1926
1927 /*
1928  *      vm_object_page_cache:
1929  *
1930  *      For the given object, attempt to move the specified clean
1931  *      pages to the cache queue.  If a page is wired for any reason,
1932  *      then it will not be changed.  Pages are specified by the given
1933  *      range ["start", "end").  As a special case, if "end" is zero,
1934  *      then the range extends from "start" to the end of the object.
1935  *      Any mappings to the specified pages are removed before the
1936  *      pages are moved to the cache queue.
1937  *
1938  *      This operation should only be performed on objects that
1939  *      contain non-fictitious, managed pages.
1940  *
1941  *      The object must be locked.
1942  */
1943 void
1944 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1945 {
1946         struct mtx *mtx, *new_mtx;
1947         vm_page_t p, next;
1948
1949         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1950         KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1951             ("vm_object_page_cache: illegal object %p", object));
1952         if (object->resident_page_count == 0)
1953                 return;
1954         p = vm_page_find_least(object, start);
1955
1956         /*
1957          * Here, the variable "p" is either (1) the page with the least pindex
1958          * greater than or equal to the parameter "start" or (2) NULL. 
1959          */
1960         mtx = NULL;
1961         for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1962                 next = TAILQ_NEXT(p, listq);
1963
1964                 /*
1965                  * Avoid releasing and reacquiring the same page lock.
1966                  */
1967                 new_mtx = vm_page_lockptr(p);
1968                 if (mtx != new_mtx) {
1969                         if (mtx != NULL)
1970                                 mtx_unlock(mtx);
1971                         mtx = new_mtx;
1972                         mtx_lock(mtx);
1973                 }
1974                 vm_page_try_to_cache(p);
1975         }
1976         if (mtx != NULL)
1977                 mtx_unlock(mtx);
1978 }
1979
1980 /*
1981  *      Populate the specified range of the object with valid pages.  Returns
1982  *      TRUE if the range is successfully populated and FALSE otherwise.
1983  *
1984  *      Note: This function should be optimized to pass a larger array of
1985  *      pages to vm_pager_get_pages() before it is applied to a non-
1986  *      OBJT_DEVICE object.
1987  *
1988  *      The object must be locked.
1989  */
1990 boolean_t
1991 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1992 {
1993         vm_page_t m, ma[1];
1994         vm_pindex_t pindex;
1995         int rv;
1996
1997         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1998         for (pindex = start; pindex < end; pindex++) {
1999                 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
2000                     VM_ALLOC_RETRY);
2001                 if (m->valid != VM_PAGE_BITS_ALL) {
2002                         ma[0] = m;
2003                         rv = vm_pager_get_pages(object, ma, 1, 0);
2004                         m = vm_page_lookup(object, pindex);
2005                         if (m == NULL)
2006                                 break;
2007                         if (rv != VM_PAGER_OK) {
2008                                 vm_page_lock(m);
2009                                 vm_page_free(m);
2010                                 vm_page_unlock(m);
2011                                 break;
2012                         }
2013                 }
2014                 /*
2015                  * Keep "m" busy because a subsequent iteration may unlock
2016                  * the object.
2017                  */
2018         }
2019         if (pindex > start) {
2020                 m = vm_page_lookup(object, start);
2021                 while (m != NULL && m->pindex < pindex) {
2022                         vm_page_wakeup(m);
2023                         m = TAILQ_NEXT(m, listq);
2024                 }
2025         }
2026         return (pindex == end);
2027 }
2028
2029 /*
2030  *      Routine:        vm_object_coalesce
2031  *      Function:       Coalesces two objects backing up adjoining
2032  *                      regions of memory into a single object.
2033  *
2034  *      returns TRUE if objects were combined.
2035  *
2036  *      NOTE:   Only works at the moment if the second object is NULL -
2037  *              if it's not, which object do we lock first?
2038  *
2039  *      Parameters:
2040  *              prev_object     First object to coalesce
2041  *              prev_offset     Offset into prev_object
2042  *              prev_size       Size of reference to prev_object
2043  *              next_size       Size of reference to the second object
2044  *              reserved        Indicator that extension region has
2045  *                              swap accounted for
2046  *
2047  *      Conditions:
2048  *      The object must *not* be locked.
2049  */
2050 boolean_t
2051 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2052     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2053 {
2054         vm_pindex_t next_pindex;
2055
2056         if (prev_object == NULL)
2057                 return (TRUE);
2058         VM_OBJECT_LOCK(prev_object);
2059         if (prev_object->type != OBJT_DEFAULT &&
2060             prev_object->type != OBJT_SWAP) {
2061                 VM_OBJECT_UNLOCK(prev_object);
2062                 return (FALSE);
2063         }
2064
2065         /*
2066          * Try to collapse the object first
2067          */
2068         vm_object_collapse(prev_object);
2069
2070         /*
2071          * Can't coalesce if: . more than one reference . paged out . shadows
2072          * another object . has a copy elsewhere (any of which mean that the
2073          * pages not mapped to prev_entry may be in use anyway)
2074          */
2075         if (prev_object->backing_object != NULL) {
2076                 VM_OBJECT_UNLOCK(prev_object);
2077                 return (FALSE);
2078         }
2079
2080         prev_size >>= PAGE_SHIFT;
2081         next_size >>= PAGE_SHIFT;
2082         next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2083
2084         if ((prev_object->ref_count > 1) &&
2085             (prev_object->size != next_pindex)) {
2086                 VM_OBJECT_UNLOCK(prev_object);
2087                 return (FALSE);
2088         }
2089
2090         /*
2091          * Account for the charge.
2092          */
2093         if (prev_object->cred != NULL) {
2094
2095                 /*
2096                  * If prev_object was charged, then this mapping,
2097                  * althought not charged now, may become writable
2098                  * later. Non-NULL cred in the object would prevent
2099                  * swap reservation during enabling of the write
2100                  * access, so reserve swap now. Failed reservation
2101                  * cause allocation of the separate object for the map
2102                  * entry, and swap reservation for this entry is
2103                  * managed in appropriate time.
2104                  */
2105                 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2106                     prev_object->cred)) {
2107                         return (FALSE);
2108                 }
2109                 prev_object->charge += ptoa(next_size);
2110         }
2111
2112         /*
2113          * Remove any pages that may still be in the object from a previous
2114          * deallocation.
2115          */
2116         if (next_pindex < prev_object->size) {
2117                 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2118                     next_size, 0);
2119                 if (prev_object->type == OBJT_SWAP)
2120                         swap_pager_freespace(prev_object,
2121                                              next_pindex, next_size);
2122 #if 0
2123                 if (prev_object->cred != NULL) {
2124                         KASSERT(prev_object->charge >=
2125                             ptoa(prev_object->size - next_pindex),
2126                             ("object %p overcharged 1 %jx %jx", prev_object,
2127                                 (uintmax_t)next_pindex, (uintmax_t)next_size));
2128                         prev_object->charge -= ptoa(prev_object->size -
2129                             next_pindex);
2130                 }
2131 #endif
2132         }
2133
2134         /*
2135          * Extend the object if necessary.
2136          */
2137         if (next_pindex + next_size > prev_object->size)
2138                 prev_object->size = next_pindex + next_size;
2139
2140         VM_OBJECT_UNLOCK(prev_object);
2141         return (TRUE);
2142 }
2143
2144 void
2145 vm_object_set_writeable_dirty(vm_object_t object)
2146 {
2147
2148         VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2149         if (object->type != OBJT_VNODE)
2150                 return;
2151         object->generation++;
2152         if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2153                 return;
2154         vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2155 }
2156
2157 #include "opt_ddb.h"
2158 #ifdef DDB
2159 #include <sys/kernel.h>
2160
2161 #include <sys/cons.h>
2162
2163 #include <ddb/ddb.h>
2164
2165 static int
2166 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2167 {
2168         vm_map_t tmpm;
2169         vm_map_entry_t tmpe;
2170         vm_object_t obj;
2171         int entcount;
2172
2173         if (map == 0)
2174                 return 0;
2175
2176         if (entry == 0) {
2177                 tmpe = map->header.next;
2178                 entcount = map->nentries;
2179                 while (entcount-- && (tmpe != &map->header)) {
2180                         if (_vm_object_in_map(map, object, tmpe)) {
2181                                 return 1;
2182                         }
2183                         tmpe = tmpe->next;
2184                 }
2185         } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2186                 tmpm = entry->object.sub_map;
2187                 tmpe = tmpm->header.next;
2188                 entcount = tmpm->nentries;
2189                 while (entcount-- && tmpe != &tmpm->header) {
2190                         if (_vm_object_in_map(tmpm, object, tmpe)) {
2191                                 return 1;
2192                         }
2193                         tmpe = tmpe->next;
2194                 }
2195         } else if ((obj = entry->object.vm_object) != NULL) {
2196                 for (; obj; obj = obj->backing_object)
2197                         if (obj == object) {
2198                                 return 1;
2199                         }
2200         }
2201         return 0;
2202 }
2203
2204 static int
2205 vm_object_in_map(vm_object_t object)
2206 {
2207         struct proc *p;
2208
2209         /* sx_slock(&allproc_lock); */
2210         FOREACH_PROC_IN_SYSTEM(p) {
2211                 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2212                         continue;
2213                 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2214                         /* sx_sunlock(&allproc_lock); */
2215                         return 1;
2216                 }
2217         }
2218         /* sx_sunlock(&allproc_lock); */
2219         if (_vm_object_in_map(kernel_map, object, 0))
2220                 return 1;
2221         if (_vm_object_in_map(kmem_map, object, 0))
2222                 return 1;
2223         if (_vm_object_in_map(pager_map, object, 0))
2224                 return 1;
2225         if (_vm_object_in_map(buffer_map, object, 0))
2226                 return 1;
2227         return 0;
2228 }
2229
2230 DB_SHOW_COMMAND(vmochk, vm_object_check)
2231 {
2232         vm_object_t object;
2233
2234         /*
2235          * make sure that internal objs are in a map somewhere
2236          * and none have zero ref counts.
2237          */
2238         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2239                 if (object->handle == NULL &&
2240                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2241                         if (object->ref_count == 0) {
2242                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2243                                         (long)object->size);
2244                         }
2245                         if (!vm_object_in_map(object)) {
2246                                 db_printf(
2247                         "vmochk: internal obj is not in a map: "
2248                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2249                                     object->ref_count, (u_long)object->size, 
2250                                     (u_long)object->size,
2251                                     (void *)object->backing_object);
2252                         }
2253                 }
2254         }
2255 }
2256
2257 /*
2258  *      vm_object_print:        [ debug ]
2259  */
2260 DB_SHOW_COMMAND(object, vm_object_print_static)
2261 {
2262         /* XXX convert args. */
2263         vm_object_t object = (vm_object_t)addr;
2264         boolean_t full = have_addr;
2265
2266         vm_page_t p;
2267
2268         /* XXX count is an (unused) arg.  Avoid shadowing it. */
2269 #define count   was_count
2270
2271         int count;
2272
2273         if (object == NULL)
2274                 return;
2275
2276         db_iprintf(
2277             "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2278             object, (int)object->type, (uintmax_t)object->size,
2279             object->resident_page_count, object->ref_count, object->flags,
2280             object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2281         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2282             object->shadow_count, 
2283             object->backing_object ? object->backing_object->ref_count : 0,
2284             object->backing_object, (uintmax_t)object->backing_object_offset);
2285
2286         if (!full)
2287                 return;
2288
2289         db_indent += 2;
2290         count = 0;
2291         TAILQ_FOREACH(p, &object->memq, listq) {
2292                 if (count == 0)
2293                         db_iprintf("memory:=");
2294                 else if (count == 6) {
2295                         db_printf("\n");
2296                         db_iprintf(" ...");
2297                         count = 0;
2298                 } else
2299                         db_printf(",");
2300                 count++;
2301
2302                 db_printf("(off=0x%jx,page=0x%jx)",
2303                     (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2304         }
2305         if (count != 0)
2306                 db_printf("\n");
2307         db_indent -= 2;
2308 }
2309
2310 /* XXX. */
2311 #undef count
2312
2313 /* XXX need this non-static entry for calling from vm_map_print. */
2314 void
2315 vm_object_print(
2316         /* db_expr_t */ long addr,
2317         boolean_t have_addr,
2318         /* db_expr_t */ long count,
2319         char *modif)
2320 {
2321         vm_object_print_static(addr, have_addr, count, modif);
2322 }
2323
2324 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2325 {
2326         vm_object_t object;
2327         vm_pindex_t fidx;
2328         vm_paddr_t pa;
2329         vm_page_t m, prev_m;
2330         int rcount, nl, c;
2331
2332         nl = 0;
2333         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2334                 db_printf("new object: %p\n", (void *)object);
2335                 if (nl > 18) {
2336                         c = cngetc();
2337                         if (c != ' ')
2338                                 return;
2339                         nl = 0;
2340                 }
2341                 nl++;
2342                 rcount = 0;
2343                 fidx = 0;
2344                 pa = -1;
2345                 TAILQ_FOREACH(m, &object->memq, listq) {
2346                         if (m->pindex > 128)
2347                                 break;
2348                         if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2349                             prev_m->pindex + 1 != m->pindex) {
2350                                 if (rcount) {
2351                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2352                                                 (long)fidx, rcount, (long)pa);
2353                                         if (nl > 18) {
2354                                                 c = cngetc();
2355                                                 if (c != ' ')
2356                                                         return;
2357                                                 nl = 0;
2358                                         }
2359                                         nl++;
2360                                         rcount = 0;
2361                                 }
2362                         }                               
2363                         if (rcount &&
2364                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2365                                 ++rcount;
2366                                 continue;
2367                         }
2368                         if (rcount) {
2369                                 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2370                                         (long)fidx, rcount, (long)pa);
2371                                 if (nl > 18) {
2372                                         c = cngetc();
2373                                         if (c != ' ')
2374                                                 return;
2375                                         nl = 0;
2376                                 }
2377                                 nl++;
2378                         }
2379                         fidx = m->pindex;
2380                         pa = VM_PAGE_TO_PHYS(m);
2381                         rcount = 1;
2382                 }
2383                 if (rcount) {
2384                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2385                                 (long)fidx, rcount, (long)pa);
2386                         if (nl > 18) {
2387                                 c = cngetc();
2388                                 if (c != ' ')
2389                                         return;
2390                                 nl = 0;
2391                         }
2392                         nl++;
2393                 }
2394         }
2395 }
2396 #endif /* DDB */