]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/vm/vm_object.c
Remove xpt_lock mutex.
[FreeBSD/FreeBSD.git] / sys / vm / vm_object.c
1 /*-
2  * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3  *
4  * Copyright (c) 1991, 1993
5  *      The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * The Mach Operating System project at Carnegie-Mellon University.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *      from: @(#)vm_object.c   8.5 (Berkeley) 3/22/94
35  *
36  *
37  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38  * All rights reserved.
39  *
40  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41  *
42  * Permission to use, copy, modify and distribute this software and
43  * its documentation is hereby granted, provided that both the copyright
44  * notice and this permission notice appear in all copies of the
45  * software, derivative works or modified versions, and any portions
46  * thereof, and that both notices appear in supporting documentation.
47  *
48  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51  *
52  * Carnegie Mellon requests users of this software to return to
53  *
54  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
55  *  School of Computer Science
56  *  Carnegie Mellon University
57  *  Pittsburgh PA 15213-3890
58  *
59  * any improvements or extensions that they make and grant Carnegie the
60  * rights to redistribute these changes.
61  */
62
63 /*
64  *      Virtual memory object module.
65  */
66
67 #include <sys/cdefs.h>
68 __FBSDID("$FreeBSD$");
69
70 #include "opt_vm.h"
71
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/cpuset.h>
75 #include <sys/lock.h>
76 #include <sys/mman.h>
77 #include <sys/mount.h>
78 #include <sys/kernel.h>
79 #include <sys/pctrie.h>
80 #include <sys/sysctl.h>
81 #include <sys/mutex.h>
82 #include <sys/proc.h>           /* for curproc, pageproc */
83 #include <sys/refcount.h>
84 #include <sys/socket.h>
85 #include <sys/resourcevar.h>
86 #include <sys/refcount.h>
87 #include <sys/rwlock.h>
88 #include <sys/user.h>
89 #include <sys/vnode.h>
90 #include <sys/vmmeter.h>
91 #include <sys/sx.h>
92
93 #include <vm/vm.h>
94 #include <vm/vm_param.h>
95 #include <vm/pmap.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_object.h>
98 #include <vm/vm_page.h>
99 #include <vm/vm_pageout.h>
100 #include <vm/vm_pager.h>
101 #include <vm/vm_phys.h>
102 #include <vm/vm_pagequeue.h>
103 #include <vm/swap_pager.h>
104 #include <vm/vm_kern.h>
105 #include <vm/vm_extern.h>
106 #include <vm/vm_radix.h>
107 #include <vm/vm_reserv.h>
108 #include <vm/uma.h>
109
110 static int old_msync;
111 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
112     "Use old (insecure) msync behavior");
113
114 static int      vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
115                     int pagerflags, int flags, boolean_t *allclean,
116                     boolean_t *eio);
117 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
118                     boolean_t *allclean);
119 static void     vm_object_qcollapse(vm_object_t object);
120 static void     vm_object_vndeallocate(vm_object_t object);
121 static void     vm_object_backing_remove(vm_object_t object);
122
123 /*
124  *      Virtual memory objects maintain the actual data
125  *      associated with allocated virtual memory.  A given
126  *      page of memory exists within exactly one object.
127  *
128  *      An object is only deallocated when all "references"
129  *      are given up.  Only one "reference" to a given
130  *      region of an object should be writeable.
131  *
132  *      Associated with each object is a list of all resident
133  *      memory pages belonging to that object; this list is
134  *      maintained by the "vm_page" module, and locked by the object's
135  *      lock.
136  *
137  *      Each object also records a "pager" routine which is
138  *      used to retrieve (and store) pages to the proper backing
139  *      storage.  In addition, objects may be backed by other
140  *      objects from which they were virtual-copied.
141  *
142  *      The only items within the object structure which are
143  *      modified after time of creation are:
144  *              reference count         locked by object's lock
145  *              pager routine           locked by object's lock
146  *
147  */
148
149 struct object_q vm_object_list;
150 struct mtx vm_object_list_mtx;  /* lock for object list and count */
151
152 struct vm_object kernel_object_store;
153
154 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
155     "VM object stats");
156
157 static counter_u64_t object_collapses = EARLY_COUNTER;
158 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
159     &object_collapses,
160     "VM object collapses");
161
162 static counter_u64_t object_bypasses = EARLY_COUNTER;
163 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
164     &object_bypasses,
165     "VM object bypasses");
166
167 static void
168 counter_startup(void)
169 {
170
171         object_collapses = counter_u64_alloc(M_WAITOK);
172         object_bypasses = counter_u64_alloc(M_WAITOK);
173 }
174 SYSINIT(object_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL);
175
176 static uma_zone_t obj_zone;
177
178 static int vm_object_zinit(void *mem, int size, int flags);
179
180 #ifdef INVARIANTS
181 static void vm_object_zdtor(void *mem, int size, void *arg);
182
183 static void
184 vm_object_zdtor(void *mem, int size, void *arg)
185 {
186         vm_object_t object;
187
188         object = (vm_object_t)mem;
189         KASSERT(object->ref_count == 0,
190             ("object %p ref_count = %d", object, object->ref_count));
191         KASSERT(TAILQ_EMPTY(&object->memq),
192             ("object %p has resident pages in its memq", object));
193         KASSERT(vm_radix_is_empty(&object->rtree),
194             ("object %p has resident pages in its trie", object));
195 #if VM_NRESERVLEVEL > 0
196         KASSERT(LIST_EMPTY(&object->rvq),
197             ("object %p has reservations",
198             object));
199 #endif
200         KASSERT(REFCOUNT_COUNT(object->paging_in_progress) == 0,
201             ("object %p paging_in_progress = %d",
202             object, REFCOUNT_COUNT(object->paging_in_progress)));
203         KASSERT(object->busy == 0,
204             ("object %p busy = %d",
205             object, object->busy));
206         KASSERT(object->resident_page_count == 0,
207             ("object %p resident_page_count = %d",
208             object, object->resident_page_count));
209         KASSERT(object->shadow_count == 0,
210             ("object %p shadow_count = %d",
211             object, object->shadow_count));
212         KASSERT(object->type == OBJT_DEAD,
213             ("object %p has non-dead type %d",
214             object, object->type));
215 }
216 #endif
217
218 static int
219 vm_object_zinit(void *mem, int size, int flags)
220 {
221         vm_object_t object;
222
223         object = (vm_object_t)mem;
224         rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
225
226         /* These are true for any object that has been freed */
227         object->type = OBJT_DEAD;
228         vm_radix_init(&object->rtree);
229         refcount_init(&object->ref_count, 0);
230         refcount_init(&object->paging_in_progress, 0);
231         refcount_init(&object->busy, 0);
232         object->resident_page_count = 0;
233         object->shadow_count = 0;
234         object->flags = OBJ_DEAD;
235
236         mtx_lock(&vm_object_list_mtx);
237         TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
238         mtx_unlock(&vm_object_list_mtx);
239         return (0);
240 }
241
242 static void
243 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
244     vm_object_t object)
245 {
246
247         TAILQ_INIT(&object->memq);
248         LIST_INIT(&object->shadow_head);
249
250         object->type = type;
251         if (type == OBJT_SWAP)
252                 pctrie_init(&object->un_pager.swp.swp_blks);
253
254         /*
255          * Ensure that swap_pager_swapoff() iteration over object_list
256          * sees up to date type and pctrie head if it observed
257          * non-dead object.
258          */
259         atomic_thread_fence_rel();
260
261         object->pg_color = 0;
262         object->flags = flags;
263         object->size = size;
264         object->domain.dr_policy = NULL;
265         object->generation = 1;
266         object->cleangeneration = 1;
267         refcount_init(&object->ref_count, 1);
268         object->memattr = VM_MEMATTR_DEFAULT;
269         object->cred = NULL;
270         object->charge = 0;
271         object->handle = NULL;
272         object->backing_object = NULL;
273         object->backing_object_offset = (vm_ooffset_t) 0;
274 #if VM_NRESERVLEVEL > 0
275         LIST_INIT(&object->rvq);
276 #endif
277         umtx_shm_object_init(object);
278 }
279
280 /*
281  *      vm_object_init:
282  *
283  *      Initialize the VM objects module.
284  */
285 void
286 vm_object_init(void)
287 {
288         TAILQ_INIT(&vm_object_list);
289         mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
290         
291         rw_init(&kernel_object->lock, "kernel vm object");
292         _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
293             VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object);
294 #if VM_NRESERVLEVEL > 0
295         kernel_object->flags |= OBJ_COLORED;
296         kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
297 #endif
298
299         /*
300          * The lock portion of struct vm_object must be type stable due
301          * to vm_pageout_fallback_object_lock locking a vm object
302          * without holding any references to it.
303          */
304         obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
305 #ifdef INVARIANTS
306             vm_object_zdtor,
307 #else
308             NULL,
309 #endif
310             vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
311
312         vm_radix_zinit();
313 }
314
315 void
316 vm_object_clear_flag(vm_object_t object, u_short bits)
317 {
318
319         VM_OBJECT_ASSERT_WLOCKED(object);
320         object->flags &= ~bits;
321 }
322
323 /*
324  *      Sets the default memory attribute for the specified object.  Pages
325  *      that are allocated to this object are by default assigned this memory
326  *      attribute.
327  *
328  *      Presently, this function must be called before any pages are allocated
329  *      to the object.  In the future, this requirement may be relaxed for
330  *      "default" and "swap" objects.
331  */
332 int
333 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
334 {
335
336         VM_OBJECT_ASSERT_WLOCKED(object);
337         switch (object->type) {
338         case OBJT_DEFAULT:
339         case OBJT_DEVICE:
340         case OBJT_MGTDEVICE:
341         case OBJT_PHYS:
342         case OBJT_SG:
343         case OBJT_SWAP:
344         case OBJT_VNODE:
345                 if (!TAILQ_EMPTY(&object->memq))
346                         return (KERN_FAILURE);
347                 break;
348         case OBJT_DEAD:
349                 return (KERN_INVALID_ARGUMENT);
350         default:
351                 panic("vm_object_set_memattr: object %p is of undefined type",
352                     object);
353         }
354         object->memattr = memattr;
355         return (KERN_SUCCESS);
356 }
357
358 void
359 vm_object_pip_add(vm_object_t object, short i)
360 {
361
362         refcount_acquiren(&object->paging_in_progress, i);
363 }
364
365 void
366 vm_object_pip_wakeup(vm_object_t object)
367 {
368
369         refcount_release(&object->paging_in_progress);
370 }
371
372 void
373 vm_object_pip_wakeupn(vm_object_t object, short i)
374 {
375
376         refcount_releasen(&object->paging_in_progress, i);
377 }
378
379 void
380 vm_object_pip_wait(vm_object_t object, char *waitid)
381 {
382
383         VM_OBJECT_ASSERT_WLOCKED(object);
384
385         while (REFCOUNT_COUNT(object->paging_in_progress) > 0) {
386                 VM_OBJECT_WUNLOCK(object);
387                 refcount_wait(&object->paging_in_progress, waitid, PVM);
388                 VM_OBJECT_WLOCK(object);
389         }
390 }
391
392 void
393 vm_object_pip_wait_unlocked(vm_object_t object, char *waitid)
394 {
395
396         VM_OBJECT_ASSERT_UNLOCKED(object);
397
398         while (REFCOUNT_COUNT(object->paging_in_progress) > 0)
399                 refcount_wait(&object->paging_in_progress, waitid, PVM);
400 }
401
402 /*
403  *      vm_object_allocate:
404  *
405  *      Returns a new object with the given size.
406  */
407 vm_object_t
408 vm_object_allocate(objtype_t type, vm_pindex_t size)
409 {
410         vm_object_t object;
411         u_short flags;
412
413         switch (type) {
414         case OBJT_DEAD:
415                 panic("vm_object_allocate: can't create OBJT_DEAD");
416         case OBJT_DEFAULT:
417         case OBJT_SWAP:
418                 flags = OBJ_COLORED;
419                 break;
420         case OBJT_DEVICE:
421         case OBJT_SG:
422                 flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
423                 break;
424         case OBJT_MGTDEVICE:
425                 flags = OBJ_FICTITIOUS;
426                 break;
427         case OBJT_PHYS:
428                 flags = OBJ_UNMANAGED;
429                 break;
430         case OBJT_VNODE:
431                 flags = 0;
432                 break;
433         default:
434                 panic("vm_object_allocate: type %d is undefined", type);
435         }
436         object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
437         _vm_object_allocate(type, size, flags, object);
438
439         return (object);
440 }
441
442 /*
443  *      vm_object_allocate_anon:
444  *
445  *      Returns a new default object of the given size and marked as
446  *      anonymous memory for special split/collapse handling.  Color
447  *      to be initialized by the caller.
448  */
449 vm_object_t
450 vm_object_allocate_anon(vm_pindex_t size)
451 {
452         vm_object_t object;
453
454         object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
455         _vm_object_allocate(OBJT_DEFAULT, size, OBJ_ANON | OBJ_ONEMAPPING,
456             object);
457
458         return (object);
459 }
460
461
462 /*
463  *      vm_object_reference:
464  *
465  *      Gets another reference to the given object.  Note: OBJ_DEAD
466  *      objects can be referenced during final cleaning.
467  */
468 void
469 vm_object_reference(vm_object_t object)
470 {
471         if (object == NULL)
472                 return;
473         VM_OBJECT_RLOCK(object);
474         vm_object_reference_locked(object);
475         VM_OBJECT_RUNLOCK(object);
476 }
477
478 /*
479  *      vm_object_reference_locked:
480  *
481  *      Gets another reference to the given object.
482  *
483  *      The object must be locked.
484  */
485 void
486 vm_object_reference_locked(vm_object_t object)
487 {
488         struct vnode *vp;
489
490         VM_OBJECT_ASSERT_LOCKED(object);
491         refcount_acquire(&object->ref_count);
492         if (object->type == OBJT_VNODE) {
493                 vp = object->handle;
494                 vref(vp);
495         }
496 }
497
498 /*
499  * Handle deallocating an object of type OBJT_VNODE.
500  */
501 static void
502 vm_object_vndeallocate(vm_object_t object)
503 {
504         struct vnode *vp = (struct vnode *) object->handle;
505
506         KASSERT(object->type == OBJT_VNODE,
507             ("vm_object_vndeallocate: not a vnode object"));
508         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
509
510         if (refcount_release(&object->ref_count) &&
511             !umtx_shm_vnobj_persistent)
512                 umtx_shm_object_terminated(object);
513
514         VM_OBJECT_RUNLOCK(object);
515         /* vrele may need the vnode lock. */
516         vrele(vp);
517 }
518
519 /*
520  *      vm_object_deallocate:
521  *
522  *      Release a reference to the specified object,
523  *      gained either through a vm_object_allocate
524  *      or a vm_object_reference call.  When all references
525  *      are gone, storage associated with this object
526  *      may be relinquished.
527  *
528  *      No object may be locked.
529  */
530 void
531 vm_object_deallocate(vm_object_t object)
532 {
533         vm_object_t temp;
534         bool released;
535
536         while (object != NULL) {
537                 VM_OBJECT_RLOCK(object);
538                 if (object->type == OBJT_VNODE) {
539                         vm_object_vndeallocate(object);
540                         return;
541                 }
542
543                 /*
544                  * If the reference count goes to 0 we start calling
545                  * vm_object_terminate() on the object chain.  A ref count
546                  * of 1 may be a special case depending on the shadow count
547                  * being 0 or 1.  These cases require a write lock on the
548                  * object.
549                  */
550                 if ((object->flags & OBJ_ANON) == 0)
551                         released = refcount_release_if_gt(&object->ref_count, 1);
552                 else
553                         released = refcount_release_if_gt(&object->ref_count, 2);
554                 VM_OBJECT_RUNLOCK(object);
555                 if (released)
556                         return;
557
558                 VM_OBJECT_WLOCK(object);
559                 KASSERT(object->ref_count != 0,
560                         ("vm_object_deallocate: object deallocated too many times: %d", object->type));
561
562                 refcount_release(&object->ref_count);
563                 if (object->ref_count > 1) {
564                         VM_OBJECT_WUNLOCK(object);
565                         return;
566                 } else if (object->ref_count == 1) {
567                         if (object->shadow_count == 0 &&
568                             object->handle == NULL &&
569                             (object->flags & OBJ_ANON) != 0) {
570                                 vm_object_set_flag(object, OBJ_ONEMAPPING);
571                         } else if ((object->shadow_count == 1) &&
572                             (object->handle == NULL) &&
573                             (object->flags & OBJ_ANON) != 0) {
574                                 vm_object_t robject;
575
576                                 robject = LIST_FIRST(&object->shadow_head);
577                                 KASSERT(robject != NULL,
578                                     ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
579                                          object->ref_count,
580                                          object->shadow_count));
581                                 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0,
582                                     ("shadowed tmpfs v_object %p", object));
583                                 if (!VM_OBJECT_TRYWLOCK(robject)) {
584                                         /*
585                                          * Avoid a potential deadlock.
586                                          */
587                                         refcount_acquire(&object->ref_count);
588                                         VM_OBJECT_WUNLOCK(object);
589                                         /*
590                                          * More likely than not the thread
591                                          * holding robject's lock has lower
592                                          * priority than the current thread.
593                                          * Let the lower priority thread run.
594                                          */
595                                         pause("vmo_de", 1);
596                                         continue;
597                                 }
598                                 /*
599                                  * Collapse object into its shadow unless its
600                                  * shadow is dead.  In that case, object will
601                                  * be deallocated by the thread that is
602                                  * deallocating its shadow.
603                                  */
604                                 if ((robject->flags &
605                                     (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON &&
606                                     robject->handle == NULL) {
607
608                                         refcount_acquire(&robject->ref_count);
609 retry:
610                                         if (REFCOUNT_COUNT(robject->paging_in_progress) > 0) {
611                                                 VM_OBJECT_WUNLOCK(object);
612                                                 vm_object_pip_wait(robject,
613                                                     "objde1");
614                                                 temp = robject->backing_object;
615                                                 if (object == temp) {
616                                                         VM_OBJECT_WLOCK(object);
617                                                         goto retry;
618                                                 }
619                                         } else if (REFCOUNT_COUNT(object->paging_in_progress) > 0) {
620                                                 VM_OBJECT_WUNLOCK(robject);
621                                                 VM_OBJECT_WUNLOCK(object);
622                                                 refcount_wait(
623                                                     &object->paging_in_progress,
624                                                     "objde2", PVM);
625                                                 VM_OBJECT_WLOCK(robject);
626                                                 temp = robject->backing_object;
627                                                 if (object == temp) {
628                                                         VM_OBJECT_WLOCK(object);
629                                                         goto retry;
630                                                 }
631                                         } else
632                                                 VM_OBJECT_WUNLOCK(object);
633
634                                         if (robject->ref_count == 1) {
635                                                 robject->ref_count--;
636                                                 object = robject;
637                                                 goto doterm;
638                                         }
639                                         object = robject;
640                                         vm_object_collapse(object);
641                                         VM_OBJECT_WUNLOCK(object);
642                                         continue;
643                                 }
644                                 VM_OBJECT_WUNLOCK(robject);
645                         }
646                         VM_OBJECT_WUNLOCK(object);
647                         return;
648                 }
649 doterm:
650                 umtx_shm_object_terminated(object);
651                 temp = object->backing_object;
652                 if (temp != NULL) {
653                         KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
654                             ("shadowed tmpfs v_object 2 %p", object));
655                         vm_object_backing_remove(object);
656                 }
657                 /*
658                  * Don't double-terminate, we could be in a termination
659                  * recursion due to the terminate having to sync data
660                  * to disk.
661                  */
662                 if ((object->flags & OBJ_DEAD) == 0) {
663                         vm_object_set_flag(object, OBJ_DEAD);
664                         vm_object_terminate(object);
665                 } else
666                         VM_OBJECT_WUNLOCK(object);
667                 object = temp;
668         }
669 }
670
671 /*
672  *      vm_object_destroy removes the object from the global object list
673  *      and frees the space for the object.
674  */
675 void
676 vm_object_destroy(vm_object_t object)
677 {
678
679         /*
680          * Release the allocation charge.
681          */
682         if (object->cred != NULL) {
683                 swap_release_by_cred(object->charge, object->cred);
684                 object->charge = 0;
685                 crfree(object->cred);
686                 object->cred = NULL;
687         }
688
689         /*
690          * Free the space for the object.
691          */
692         uma_zfree(obj_zone, object);
693 }
694
695 static void
696 vm_object_backing_remove_locked(vm_object_t object)
697 {
698         vm_object_t backing_object;
699
700         backing_object = object->backing_object;
701         VM_OBJECT_ASSERT_WLOCKED(object);
702         VM_OBJECT_ASSERT_WLOCKED(backing_object);
703
704         if ((object->flags & OBJ_SHADOWLIST) != 0) {
705                 LIST_REMOVE(object, shadow_list);
706                 backing_object->shadow_count--;
707                 object->flags &= ~OBJ_SHADOWLIST;
708         }
709         object->backing_object = NULL;
710 }
711
712 static void
713 vm_object_backing_remove(vm_object_t object)
714 {
715         vm_object_t backing_object;
716
717         VM_OBJECT_ASSERT_WLOCKED(object);
718
719         if ((object->flags & OBJ_SHADOWLIST) != 0) {
720                 backing_object = object->backing_object;
721                 VM_OBJECT_WLOCK(backing_object);
722                 vm_object_backing_remove_locked(object);
723                 VM_OBJECT_WUNLOCK(backing_object);
724         } else
725                 object->backing_object = NULL;
726 }
727
728 static void
729 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
730 {
731
732         VM_OBJECT_ASSERT_WLOCKED(object);
733
734         if ((backing_object->flags & OBJ_ANON) != 0) {
735                 VM_OBJECT_ASSERT_WLOCKED(backing_object);
736                 LIST_INSERT_HEAD(&backing_object->shadow_head, object,
737                     shadow_list);
738                 backing_object->shadow_count++;
739                 object->flags |= OBJ_SHADOWLIST;
740         }
741         object->backing_object = backing_object;
742 }
743
744 static void
745 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
746 {
747
748         VM_OBJECT_ASSERT_WLOCKED(object);
749
750         if ((backing_object->flags & OBJ_ANON) != 0) {
751                 VM_OBJECT_WLOCK(backing_object);
752                 vm_object_backing_insert_locked(object, backing_object);
753                 VM_OBJECT_WUNLOCK(backing_object);
754         } else
755                 object->backing_object = backing_object;
756 }
757
758
759 /*
760  *      vm_object_terminate_pages removes any remaining pageable pages
761  *      from the object and resets the object to an empty state.
762  */
763 static void
764 vm_object_terminate_pages(vm_object_t object)
765 {
766         vm_page_t p, p_next;
767
768         VM_OBJECT_ASSERT_WLOCKED(object);
769
770         /*
771          * Free any remaining pageable pages.  This also removes them from the
772          * paging queues.  However, don't free wired pages, just remove them
773          * from the object.  Rather than incrementally removing each page from
774          * the object, the page and object are reset to any empty state. 
775          */
776         TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
777                 vm_page_assert_unbusied(p);
778                 KASSERT(p->object == object &&
779                     (p->ref_count & VPRC_OBJREF) != 0,
780                     ("vm_object_terminate_pages: page %p is inconsistent", p));
781
782                 p->object = NULL;
783                 if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
784                         VM_CNT_INC(v_pfree);
785                         vm_page_free(p);
786                 }
787         }
788
789         /*
790          * If the object contained any pages, then reset it to an empty state.
791          * None of the object's fields, including "resident_page_count", were
792          * modified by the preceding loop.
793          */
794         if (object->resident_page_count != 0) {
795                 vm_radix_reclaim_allnodes(&object->rtree);
796                 TAILQ_INIT(&object->memq);
797                 object->resident_page_count = 0;
798                 if (object->type == OBJT_VNODE)
799                         vdrop(object->handle);
800         }
801 }
802
803 /*
804  *      vm_object_terminate actually destroys the specified object, freeing
805  *      up all previously used resources.
806  *
807  *      The object must be locked.
808  *      This routine may block.
809  */
810 void
811 vm_object_terminate(vm_object_t object)
812 {
813         VM_OBJECT_ASSERT_WLOCKED(object);
814         KASSERT((object->flags & OBJ_DEAD) != 0,
815             ("terminating non-dead obj %p", object));
816
817         /*
818          * wait for the pageout daemon to be done with the object
819          */
820         vm_object_pip_wait(object, "objtrm");
821
822         KASSERT(!REFCOUNT_COUNT(object->paging_in_progress),
823                 ("vm_object_terminate: pageout in progress"));
824
825         KASSERT(object->ref_count == 0, 
826                 ("vm_object_terminate: object with references, ref_count=%d",
827                 object->ref_count));
828
829         if ((object->flags & OBJ_PG_DTOR) == 0)
830                 vm_object_terminate_pages(object);
831
832 #if VM_NRESERVLEVEL > 0
833         if (__predict_false(!LIST_EMPTY(&object->rvq)))
834                 vm_reserv_break_all(object);
835 #endif
836
837         KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
838             object->type == OBJT_SWAP,
839             ("%s: non-swap obj %p has cred", __func__, object));
840
841         /*
842          * Let the pager know object is dead.
843          */
844         vm_pager_deallocate(object);
845         VM_OBJECT_WUNLOCK(object);
846
847         vm_object_destroy(object);
848 }
849
850 /*
851  * Make the page read-only so that we can clear the object flags.  However, if
852  * this is a nosync mmap then the object is likely to stay dirty so do not
853  * mess with the page and do not clear the object flags.  Returns TRUE if the
854  * page should be flushed, and FALSE otherwise.
855  */
856 static boolean_t
857 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
858 {
859
860         vm_page_assert_busied(p);
861
862         /*
863          * If we have been asked to skip nosync pages and this is a
864          * nosync page, skip it.  Note that the object flags were not
865          * cleared in this case so we do not have to set them.
866          */
867         if ((flags & OBJPC_NOSYNC) != 0 && (p->aflags & PGA_NOSYNC) != 0) {
868                 *allclean = FALSE;
869                 return (FALSE);
870         } else {
871                 pmap_remove_write(p);
872                 return (p->dirty != 0);
873         }
874 }
875
876 /*
877  *      vm_object_page_clean
878  *
879  *      Clean all dirty pages in the specified range of object.  Leaves page 
880  *      on whatever queue it is currently on.   If NOSYNC is set then do not
881  *      write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
882  *      leaving the object dirty.
883  *
884  *      When stuffing pages asynchronously, allow clustering.  XXX we need a
885  *      synchronous clustering mode implementation.
886  *
887  *      Odd semantics: if start == end, we clean everything.
888  *
889  *      The object must be locked.
890  *
891  *      Returns FALSE if some page from the range was not written, as
892  *      reported by the pager, and TRUE otherwise.
893  */
894 boolean_t
895 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
896     int flags)
897 {
898         vm_page_t np, p;
899         vm_pindex_t pi, tend, tstart;
900         int curgeneration, n, pagerflags;
901         boolean_t eio, res, allclean;
902
903         VM_OBJECT_ASSERT_WLOCKED(object);
904
905         if (object->type != OBJT_VNODE || !vm_object_mightbedirty(object) ||
906             object->resident_page_count == 0)
907                 return (TRUE);
908
909         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
910             VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
911         pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
912
913         tstart = OFF_TO_IDX(start);
914         tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
915         allclean = tstart == 0 && tend >= object->size;
916         res = TRUE;
917
918 rescan:
919         curgeneration = object->generation;
920
921         for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
922                 pi = p->pindex;
923                 if (pi >= tend)
924                         break;
925                 np = TAILQ_NEXT(p, listq);
926                 if (vm_page_none_valid(p))
927                         continue;
928                 if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
929                         if (object->generation != curgeneration &&
930                             (flags & OBJPC_SYNC) != 0)
931                                 goto rescan;
932                         np = vm_page_find_least(object, pi);
933                         continue;
934                 }
935                 if (!vm_object_page_remove_write(p, flags, &allclean)) {
936                         vm_page_xunbusy(p);
937                         continue;
938                 }
939
940                 n = vm_object_page_collect_flush(object, p, pagerflags,
941                     flags, &allclean, &eio);
942                 if (eio) {
943                         res = FALSE;
944                         allclean = FALSE;
945                 }
946                 if (object->generation != curgeneration &&
947                     (flags & OBJPC_SYNC) != 0)
948                         goto rescan;
949
950                 /*
951                  * If the VOP_PUTPAGES() did a truncated write, so
952                  * that even the first page of the run is not fully
953                  * written, vm_pageout_flush() returns 0 as the run
954                  * length.  Since the condition that caused truncated
955                  * write may be permanent, e.g. exhausted free space,
956                  * accepting n == 0 would cause an infinite loop.
957                  *
958                  * Forwarding the iterator leaves the unwritten page
959                  * behind, but there is not much we can do there if
960                  * filesystem refuses to write it.
961                  */
962                 if (n == 0) {
963                         n = 1;
964                         allclean = FALSE;
965                 }
966                 np = vm_page_find_least(object, pi + n);
967         }
968 #if 0
969         VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
970 #endif
971
972         if (allclean)
973                 object->cleangeneration = curgeneration;
974         return (res);
975 }
976
977 static int
978 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
979     int flags, boolean_t *allclean, boolean_t *eio)
980 {
981         vm_page_t ma[vm_pageout_page_count], p_first, tp;
982         int count, i, mreq, runlen;
983
984         vm_page_lock_assert(p, MA_NOTOWNED);
985         vm_page_assert_xbusied(p);
986         VM_OBJECT_ASSERT_WLOCKED(object);
987
988         count = 1;
989         mreq = 0;
990
991         for (tp = p; count < vm_pageout_page_count; count++) {
992                 tp = vm_page_next(tp);
993                 if (tp == NULL || vm_page_tryxbusy(tp) == 0)
994                         break;
995                 if (!vm_object_page_remove_write(tp, flags, allclean)) {
996                         vm_page_xunbusy(tp);
997                         break;
998                 }
999         }
1000
1001         for (p_first = p; count < vm_pageout_page_count; count++) {
1002                 tp = vm_page_prev(p_first);
1003                 if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1004                         break;
1005                 if (!vm_object_page_remove_write(tp, flags, allclean)) {
1006                         vm_page_xunbusy(tp);
1007                         break;
1008                 }
1009                 p_first = tp;
1010                 mreq++;
1011         }
1012
1013         for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1014                 ma[i] = tp;
1015
1016         vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1017         return (runlen);
1018 }
1019
1020 /*
1021  * Note that there is absolutely no sense in writing out
1022  * anonymous objects, so we track down the vnode object
1023  * to write out.
1024  * We invalidate (remove) all pages from the address space
1025  * for semantic correctness.
1026  *
1027  * If the backing object is a device object with unmanaged pages, then any
1028  * mappings to the specified range of pages must be removed before this
1029  * function is called.
1030  *
1031  * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1032  * may start out with a NULL object.
1033  */
1034 boolean_t
1035 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1036     boolean_t syncio, boolean_t invalidate)
1037 {
1038         vm_object_t backing_object;
1039         struct vnode *vp;
1040         struct mount *mp;
1041         int error, flags, fsync_after;
1042         boolean_t res;
1043
1044         if (object == NULL)
1045                 return (TRUE);
1046         res = TRUE;
1047         error = 0;
1048         VM_OBJECT_WLOCK(object);
1049         while ((backing_object = object->backing_object) != NULL) {
1050                 VM_OBJECT_WLOCK(backing_object);
1051                 offset += object->backing_object_offset;
1052                 VM_OBJECT_WUNLOCK(object);
1053                 object = backing_object;
1054                 if (object->size < OFF_TO_IDX(offset + size))
1055                         size = IDX_TO_OFF(object->size) - offset;
1056         }
1057         /*
1058          * Flush pages if writing is allowed, invalidate them
1059          * if invalidation requested.  Pages undergoing I/O
1060          * will be ignored by vm_object_page_remove().
1061          *
1062          * We cannot lock the vnode and then wait for paging
1063          * to complete without deadlocking against vm_fault.
1064          * Instead we simply call vm_object_page_remove() and
1065          * allow it to block internally on a page-by-page
1066          * basis when it encounters pages undergoing async
1067          * I/O.
1068          */
1069         if (object->type == OBJT_VNODE &&
1070             vm_object_mightbedirty(object) != 0 &&
1071             ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1072                 VM_OBJECT_WUNLOCK(object);
1073                 (void) vn_start_write(vp, &mp, V_WAIT);
1074                 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1075                 if (syncio && !invalidate && offset == 0 &&
1076                     atop(size) == object->size) {
1077                         /*
1078                          * If syncing the whole mapping of the file,
1079                          * it is faster to schedule all the writes in
1080                          * async mode, also allowing the clustering,
1081                          * and then wait for i/o to complete.
1082                          */
1083                         flags = 0;
1084                         fsync_after = TRUE;
1085                 } else {
1086                         flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1087                         flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1088                         fsync_after = FALSE;
1089                 }
1090                 VM_OBJECT_WLOCK(object);
1091                 res = vm_object_page_clean(object, offset, offset + size,
1092                     flags);
1093                 VM_OBJECT_WUNLOCK(object);
1094                 if (fsync_after)
1095                         error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1096                 VOP_UNLOCK(vp, 0);
1097                 vn_finished_write(mp);
1098                 if (error != 0)
1099                         res = FALSE;
1100                 VM_OBJECT_WLOCK(object);
1101         }
1102         if ((object->type == OBJT_VNODE ||
1103              object->type == OBJT_DEVICE) && invalidate) {
1104                 if (object->type == OBJT_DEVICE)
1105                         /*
1106                          * The option OBJPR_NOTMAPPED must be passed here
1107                          * because vm_object_page_remove() cannot remove
1108                          * unmanaged mappings.
1109                          */
1110                         flags = OBJPR_NOTMAPPED;
1111                 else if (old_msync)
1112                         flags = 0;
1113                 else
1114                         flags = OBJPR_CLEANONLY;
1115                 vm_object_page_remove(object, OFF_TO_IDX(offset),
1116                     OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1117         }
1118         VM_OBJECT_WUNLOCK(object);
1119         return (res);
1120 }
1121
1122 /*
1123  * Determine whether the given advice can be applied to the object.  Advice is
1124  * not applied to unmanaged pages since they never belong to page queues, and
1125  * since MADV_FREE is destructive, it can apply only to anonymous pages that
1126  * have been mapped at most once.
1127  */
1128 static bool
1129 vm_object_advice_applies(vm_object_t object, int advice)
1130 {
1131
1132         if ((object->flags & OBJ_UNMANAGED) != 0)
1133                 return (false);
1134         if (advice != MADV_FREE)
1135                 return (true);
1136         return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1137             (OBJ_ONEMAPPING | OBJ_ANON));
1138 }
1139
1140 static void
1141 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1142     vm_size_t size)
1143 {
1144
1145         if (advice == MADV_FREE && object->type == OBJT_SWAP)
1146                 swap_pager_freespace(object, pindex, size);
1147 }
1148
1149 /*
1150  *      vm_object_madvise:
1151  *
1152  *      Implements the madvise function at the object/page level.
1153  *
1154  *      MADV_WILLNEED   (any object)
1155  *
1156  *          Activate the specified pages if they are resident.
1157  *
1158  *      MADV_DONTNEED   (any object)
1159  *
1160  *          Deactivate the specified pages if they are resident.
1161  *
1162  *      MADV_FREE       (OBJT_DEFAULT/OBJT_SWAP objects,
1163  *                       OBJ_ONEMAPPING only)
1164  *
1165  *          Deactivate and clean the specified pages if they are
1166  *          resident.  This permits the process to reuse the pages
1167  *          without faulting or the kernel to reclaim the pages
1168  *          without I/O.
1169  */
1170 void
1171 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1172     int advice)
1173 {
1174         vm_pindex_t tpindex;
1175         vm_object_t backing_object, tobject;
1176         vm_page_t m, tm;
1177
1178         if (object == NULL)
1179                 return;
1180
1181 relookup:
1182         VM_OBJECT_WLOCK(object);
1183         if (!vm_object_advice_applies(object, advice)) {
1184                 VM_OBJECT_WUNLOCK(object);
1185                 return;
1186         }
1187         for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1188                 tobject = object;
1189
1190                 /*
1191                  * If the next page isn't resident in the top-level object, we
1192                  * need to search the shadow chain.  When applying MADV_FREE, we
1193                  * take care to release any swap space used to store
1194                  * non-resident pages.
1195                  */
1196                 if (m == NULL || pindex < m->pindex) {
1197                         /*
1198                          * Optimize a common case: if the top-level object has
1199                          * no backing object, we can skip over the non-resident
1200                          * range in constant time.
1201                          */
1202                         if (object->backing_object == NULL) {
1203                                 tpindex = (m != NULL && m->pindex < end) ?
1204                                     m->pindex : end;
1205                                 vm_object_madvise_freespace(object, advice,
1206                                     pindex, tpindex - pindex);
1207                                 if ((pindex = tpindex) == end)
1208                                         break;
1209                                 goto next_page;
1210                         }
1211
1212                         tpindex = pindex;
1213                         do {
1214                                 vm_object_madvise_freespace(tobject, advice,
1215                                     tpindex, 1);
1216                                 /*
1217                                  * Prepare to search the next object in the
1218                                  * chain.
1219                                  */
1220                                 backing_object = tobject->backing_object;
1221                                 if (backing_object == NULL)
1222                                         goto next_pindex;
1223                                 VM_OBJECT_WLOCK(backing_object);
1224                                 tpindex +=
1225                                     OFF_TO_IDX(tobject->backing_object_offset);
1226                                 if (tobject != object)
1227                                         VM_OBJECT_WUNLOCK(tobject);
1228                                 tobject = backing_object;
1229                                 if (!vm_object_advice_applies(tobject, advice))
1230                                         goto next_pindex;
1231                         } while ((tm = vm_page_lookup(tobject, tpindex)) ==
1232                             NULL);
1233                 } else {
1234 next_page:
1235                         tm = m;
1236                         m = TAILQ_NEXT(m, listq);
1237                 }
1238
1239                 /*
1240                  * If the page is not in a normal state, skip it.  The page
1241                  * can not be invalidated while the object lock is held.
1242                  */
1243                 if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1244                         goto next_pindex;
1245                 KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1246                     ("vm_object_madvise: page %p is fictitious", tm));
1247                 KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1248                     ("vm_object_madvise: page %p is not managed", tm));
1249                 if (vm_page_tryxbusy(tm) == 0) {
1250                         if (object != tobject)
1251                                 VM_OBJECT_WUNLOCK(object);
1252                         if (advice == MADV_WILLNEED) {
1253                                 /*
1254                                  * Reference the page before unlocking and
1255                                  * sleeping so that the page daemon is less
1256                                  * likely to reclaim it.
1257                                  */
1258                                 vm_page_aflag_set(tm, PGA_REFERENCED);
1259                         }
1260                         vm_page_busy_sleep(tm, "madvpo", false);
1261                         goto relookup;
1262                 }
1263                 vm_page_lock(tm);
1264                 vm_page_advise(tm, advice);
1265                 vm_page_unlock(tm);
1266                 vm_page_xunbusy(tm);
1267                 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1268 next_pindex:
1269                 if (tobject != object)
1270                         VM_OBJECT_WUNLOCK(tobject);
1271         }
1272         VM_OBJECT_WUNLOCK(object);
1273 }
1274
1275 /*
1276  *      vm_object_shadow:
1277  *
1278  *      Create a new object which is backed by the
1279  *      specified existing object range.  The source
1280  *      object reference is deallocated.
1281  *
1282  *      The new object and offset into that object
1283  *      are returned in the source parameters.
1284  */
1285 void
1286 vm_object_shadow(
1287         vm_object_t *object,    /* IN/OUT */
1288         vm_ooffset_t *offset,   /* IN/OUT */
1289         vm_size_t length)
1290 {
1291         vm_object_t source;
1292         vm_object_t result;
1293
1294         source = *object;
1295
1296         /*
1297          * Don't create the new object if the old object isn't shared.
1298          *
1299          * If we hold the only reference we can guarantee that it won't
1300          * increase while we have the map locked.  Otherwise the race is
1301          * harmless and we will end up with an extra shadow object that
1302          * will be collapsed later.
1303          */
1304         if (source != NULL && source->ref_count == 1 &&
1305             source->handle == NULL && (source->flags & OBJ_ANON) != 0)
1306                 return;
1307
1308         /*
1309          * Allocate a new object with the given length.
1310          */
1311         result = vm_object_allocate_anon(atop(length));
1312
1313         /*
1314          * Store the offset into the source object, and fix up the offset into
1315          * the new object.
1316          */
1317         result->backing_object_offset = *offset;
1318
1319         /*
1320          * The new object shadows the source object, adding a reference to it.
1321          * Our caller changes his reference to point to the new object,
1322          * removing a reference to the source object.  Net result: no change
1323          * of reference count.
1324          *
1325          * Try to optimize the result object's page color when shadowing
1326          * in order to maintain page coloring consistency in the combined 
1327          * shadowed object.
1328          */
1329         if (source != NULL) {
1330                 VM_OBJECT_WLOCK(result);
1331                 vm_object_backing_insert(result, source);
1332                 result->domain = source->domain;
1333 #if VM_NRESERVLEVEL > 0
1334                 result->flags |= source->flags & OBJ_COLORED;
1335                 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1336                     ((1 << (VM_NFREEORDER - 1)) - 1);
1337 #endif
1338                 VM_OBJECT_WUNLOCK(result);
1339         }
1340
1341         /*
1342          * Return the new things
1343          */
1344         *offset = 0;
1345         *object = result;
1346 }
1347
1348 /*
1349  *      vm_object_split:
1350  *
1351  * Split the pages in a map entry into a new object.  This affords
1352  * easier removal of unused pages, and keeps object inheritance from
1353  * being a negative impact on memory usage.
1354  */
1355 void
1356 vm_object_split(vm_map_entry_t entry)
1357 {
1358         vm_page_t m, m_next;
1359         vm_object_t orig_object, new_object, source;
1360         vm_pindex_t idx, offidxstart;
1361         vm_size_t size;
1362
1363         orig_object = entry->object.vm_object;
1364         if ((orig_object->flags & OBJ_ANON) == 0)
1365                 return;
1366         if (orig_object->ref_count <= 1)
1367                 return;
1368         VM_OBJECT_WUNLOCK(orig_object);
1369
1370         offidxstart = OFF_TO_IDX(entry->offset);
1371         size = atop(entry->end - entry->start);
1372
1373         /*
1374          * If swap_pager_copy() is later called, it will convert new_object
1375          * into a swap object.
1376          */
1377         new_object = vm_object_allocate_anon(size);
1378
1379         /*
1380          * At this point, the new object is still private, so the order in
1381          * which the original and new objects are locked does not matter.
1382          */
1383         VM_OBJECT_WLOCK(new_object);
1384         VM_OBJECT_WLOCK(orig_object);
1385         new_object->domain = orig_object->domain;
1386         source = orig_object->backing_object;
1387         if (source != NULL) {
1388                 if ((source->flags & (OBJ_ANON | OBJ_DEAD)) != 0) {
1389                         VM_OBJECT_WLOCK(source);
1390                         if ((source->flags & OBJ_DEAD) != 0) {
1391                                 VM_OBJECT_WUNLOCK(source);
1392                                 VM_OBJECT_WUNLOCK(orig_object);
1393                                 VM_OBJECT_WUNLOCK(new_object);
1394                                 vm_object_deallocate(new_object);
1395                                 VM_OBJECT_WLOCK(orig_object);
1396                                 return;
1397                         }
1398                         vm_object_backing_insert_locked(new_object, source);
1399                         vm_object_reference_locked(source);     /* for new_object */
1400                         vm_object_clear_flag(source, OBJ_ONEMAPPING);
1401                         VM_OBJECT_WUNLOCK(source);
1402                 } else {
1403                         vm_object_backing_insert(new_object, source);
1404                         vm_object_reference(source);
1405                 }
1406                 new_object->backing_object_offset = 
1407                         orig_object->backing_object_offset + entry->offset;
1408         }
1409         if (orig_object->cred != NULL) {
1410                 new_object->cred = orig_object->cred;
1411                 crhold(orig_object->cred);
1412                 new_object->charge = ptoa(size);
1413                 KASSERT(orig_object->charge >= ptoa(size),
1414                     ("orig_object->charge < 0"));
1415                 orig_object->charge -= ptoa(size);
1416         }
1417 retry:
1418         m = vm_page_find_least(orig_object, offidxstart);
1419         for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1420             m = m_next) {
1421                 m_next = TAILQ_NEXT(m, listq);
1422
1423                 /*
1424                  * We must wait for pending I/O to complete before we can
1425                  * rename the page.
1426                  *
1427                  * We do not have to VM_PROT_NONE the page as mappings should
1428                  * not be changed by this operation.
1429                  */
1430                 if (vm_page_tryxbusy(m) == 0) {
1431                         VM_OBJECT_WUNLOCK(new_object);
1432                         vm_page_sleep_if_busy(m, "spltwt");
1433                         VM_OBJECT_WLOCK(new_object);
1434                         goto retry;
1435                 }
1436
1437                 /* vm_page_rename() will dirty the page. */
1438                 if (vm_page_rename(m, new_object, idx)) {
1439                         vm_page_xunbusy(m);
1440                         VM_OBJECT_WUNLOCK(new_object);
1441                         VM_OBJECT_WUNLOCK(orig_object);
1442                         vm_radix_wait();
1443                         VM_OBJECT_WLOCK(orig_object);
1444                         VM_OBJECT_WLOCK(new_object);
1445                         goto retry;
1446                 }
1447
1448 #if VM_NRESERVLEVEL > 0
1449                 /*
1450                  * If some of the reservation's allocated pages remain with
1451                  * the original object, then transferring the reservation to
1452                  * the new object is neither particularly beneficial nor
1453                  * particularly harmful as compared to leaving the reservation
1454                  * with the original object.  If, however, all of the
1455                  * reservation's allocated pages are transferred to the new
1456                  * object, then transferring the reservation is typically
1457                  * beneficial.  Determining which of these two cases applies
1458                  * would be more costly than unconditionally renaming the
1459                  * reservation.
1460                  */
1461                 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1462 #endif
1463                 if (orig_object->type != OBJT_SWAP)
1464                         vm_page_xunbusy(m);
1465         }
1466         if (orig_object->type == OBJT_SWAP) {
1467                 /*
1468                  * swap_pager_copy() can sleep, in which case the orig_object's
1469                  * and new_object's locks are released and reacquired. 
1470                  */
1471                 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1472                 TAILQ_FOREACH(m, &new_object->memq, listq)
1473                         vm_page_xunbusy(m);
1474         }
1475         VM_OBJECT_WUNLOCK(orig_object);
1476         VM_OBJECT_WUNLOCK(new_object);
1477         entry->object.vm_object = new_object;
1478         entry->offset = 0LL;
1479         vm_object_deallocate(orig_object);
1480         VM_OBJECT_WLOCK(new_object);
1481 }
1482
1483 #define OBSC_COLLAPSE_NOWAIT    0x0002
1484 #define OBSC_COLLAPSE_WAIT      0x0004
1485
1486 static vm_page_t
1487 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next,
1488     int op)
1489 {
1490         vm_object_t backing_object;
1491
1492         VM_OBJECT_ASSERT_WLOCKED(object);
1493         backing_object = object->backing_object;
1494         VM_OBJECT_ASSERT_WLOCKED(backing_object);
1495
1496         KASSERT(p == NULL || p->object == object || p->object == backing_object,
1497             ("invalid ownership %p %p %p", p, object, backing_object));
1498         if ((op & OBSC_COLLAPSE_NOWAIT) != 0)
1499                 return (next);
1500         /* The page is only NULL when rename fails. */
1501         if (p == NULL) {
1502                 vm_radix_wait();
1503         } else {
1504                 if (p->object == object)
1505                         VM_OBJECT_WUNLOCK(backing_object);
1506                 else
1507                         VM_OBJECT_WUNLOCK(object);
1508                 vm_page_busy_sleep(p, "vmocol", false);
1509         }
1510         VM_OBJECT_WLOCK(object);
1511         VM_OBJECT_WLOCK(backing_object);
1512         return (TAILQ_FIRST(&backing_object->memq));
1513 }
1514
1515 static bool
1516 vm_object_scan_all_shadowed(vm_object_t object)
1517 {
1518         vm_object_t backing_object;
1519         vm_page_t p, pp;
1520         vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1521
1522         VM_OBJECT_ASSERT_WLOCKED(object);
1523         VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1524
1525         backing_object = object->backing_object;
1526
1527         if ((backing_object->flags & OBJ_ANON) == 0)
1528                 return (false);
1529
1530         pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1531         p = vm_page_find_least(backing_object, pi);
1532         ps = swap_pager_find_least(backing_object, pi);
1533
1534         /*
1535          * Only check pages inside the parent object's range and
1536          * inside the parent object's mapping of the backing object.
1537          */
1538         for (;; pi++) {
1539                 if (p != NULL && p->pindex < pi)
1540                         p = TAILQ_NEXT(p, listq);
1541                 if (ps < pi)
1542                         ps = swap_pager_find_least(backing_object, pi);
1543                 if (p == NULL && ps >= backing_object->size)
1544                         break;
1545                 else if (p == NULL)
1546                         pi = ps;
1547                 else
1548                         pi = MIN(p->pindex, ps);
1549
1550                 new_pindex = pi - backing_offset_index;
1551                 if (new_pindex >= object->size)
1552                         break;
1553
1554                 /*
1555                  * See if the parent has the page or if the parent's object
1556                  * pager has the page.  If the parent has the page but the page
1557                  * is not valid, the parent's object pager must have the page.
1558                  *
1559                  * If this fails, the parent does not completely shadow the
1560                  * object and we might as well give up now.
1561                  */
1562                 pp = vm_page_lookup(object, new_pindex);
1563                 /*
1564                  * The valid check here is stable due to object lock being
1565                  * required to clear valid and initiate paging.
1566                  */
1567                 if ((pp == NULL || vm_page_none_valid(pp)) &&
1568                     !vm_pager_has_page(object, new_pindex, NULL, NULL))
1569                         return (false);
1570         }
1571         return (true);
1572 }
1573
1574 static bool
1575 vm_object_collapse_scan(vm_object_t object, int op)
1576 {
1577         vm_object_t backing_object;
1578         vm_page_t next, p, pp;
1579         vm_pindex_t backing_offset_index, new_pindex;
1580
1581         VM_OBJECT_ASSERT_WLOCKED(object);
1582         VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1583
1584         backing_object = object->backing_object;
1585         backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1586
1587         /*
1588          * Initial conditions
1589          */
1590         if ((op & OBSC_COLLAPSE_WAIT) != 0)
1591                 vm_object_set_flag(backing_object, OBJ_DEAD);
1592
1593         /*
1594          * Our scan
1595          */
1596         for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1597                 next = TAILQ_NEXT(p, listq);
1598                 new_pindex = p->pindex - backing_offset_index;
1599
1600                 /*
1601                  * Check for busy page
1602                  */
1603                 if (vm_page_tryxbusy(p) == 0) {
1604                         next = vm_object_collapse_scan_wait(object, p, next, op);
1605                         continue;
1606                 }
1607
1608                 KASSERT(p->object == backing_object,
1609                     ("vm_object_collapse_scan: object mismatch"));
1610
1611                 if (p->pindex < backing_offset_index ||
1612                     new_pindex >= object->size) {
1613                         if (backing_object->type == OBJT_SWAP)
1614                                 swap_pager_freespace(backing_object, p->pindex,
1615                                     1);
1616
1617                         KASSERT(!pmap_page_is_mapped(p),
1618                             ("freeing mapped page %p", p));
1619                         if (vm_page_remove(p))
1620                                 vm_page_free(p);
1621                         else
1622                                 vm_page_xunbusy(p);
1623                         continue;
1624                 }
1625
1626                 pp = vm_page_lookup(object, new_pindex);
1627                 if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1628                         vm_page_xunbusy(p);
1629                         /*
1630                          * The page in the parent is busy and possibly not
1631                          * (yet) valid.  Until its state is finalized by the
1632                          * busy bit owner, we can't tell whether it shadows the
1633                          * original page.  Therefore, we must either skip it
1634                          * and the original (backing_object) page or wait for
1635                          * its state to be finalized.
1636                          *
1637                          * This is due to a race with vm_fault() where we must
1638                          * unbusy the original (backing_obj) page before we can
1639                          * (re)lock the parent.  Hence we can get here.
1640                          */
1641                         next = vm_object_collapse_scan_wait(object, pp, next,
1642                             op);
1643                         continue;
1644                 }
1645
1646                 KASSERT(pp == NULL || !vm_page_none_valid(pp),
1647                     ("unbusy invalid page %p", pp));
1648
1649                 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1650                         NULL)) {
1651                         /*
1652                          * The page already exists in the parent OR swap exists
1653                          * for this location in the parent.  Leave the parent's
1654                          * page alone.  Destroy the original page from the
1655                          * backing object.
1656                          */
1657                         if (backing_object->type == OBJT_SWAP)
1658                                 swap_pager_freespace(backing_object, p->pindex,
1659                                     1);
1660                         KASSERT(!pmap_page_is_mapped(p),
1661                             ("freeing mapped page %p", p));
1662                         if (vm_page_remove(p))
1663                                 vm_page_free(p);
1664                         else
1665                                 vm_page_xunbusy(p);
1666                         if (pp != NULL)
1667                                 vm_page_xunbusy(pp);
1668                         continue;
1669                 }
1670
1671                 /*
1672                  * Page does not exist in parent, rename the page from the
1673                  * backing object to the main object.
1674                  *
1675                  * If the page was mapped to a process, it can remain mapped
1676                  * through the rename.  vm_page_rename() will dirty the page.
1677                  */
1678                 if (vm_page_rename(p, object, new_pindex)) {
1679                         vm_page_xunbusy(p);
1680                         if (pp != NULL)
1681                                 vm_page_xunbusy(pp);
1682                         next = vm_object_collapse_scan_wait(object, NULL, next,
1683                             op);
1684                         continue;
1685                 }
1686
1687                 /* Use the old pindex to free the right page. */
1688                 if (backing_object->type == OBJT_SWAP)
1689                         swap_pager_freespace(backing_object,
1690                             new_pindex + backing_offset_index, 1);
1691
1692 #if VM_NRESERVLEVEL > 0
1693                 /*
1694                  * Rename the reservation.
1695                  */
1696                 vm_reserv_rename(p, object, backing_object,
1697                     backing_offset_index);
1698 #endif
1699                 vm_page_xunbusy(p);
1700         }
1701         return (true);
1702 }
1703
1704
1705 /*
1706  * this version of collapse allows the operation to occur earlier and
1707  * when paging_in_progress is true for an object...  This is not a complete
1708  * operation, but should plug 99.9% of the rest of the leaks.
1709  */
1710 static void
1711 vm_object_qcollapse(vm_object_t object)
1712 {
1713         vm_object_t backing_object = object->backing_object;
1714
1715         VM_OBJECT_ASSERT_WLOCKED(object);
1716         VM_OBJECT_ASSERT_WLOCKED(backing_object);
1717
1718         if (backing_object->ref_count != 1)
1719                 return;
1720
1721         vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT);
1722 }
1723
1724 /*
1725  *      vm_object_collapse:
1726  *
1727  *      Collapse an object with the object backing it.
1728  *      Pages in the backing object are moved into the
1729  *      parent, and the backing object is deallocated.
1730  */
1731 void
1732 vm_object_collapse(vm_object_t object)
1733 {
1734         vm_object_t backing_object, new_backing_object;
1735
1736         VM_OBJECT_ASSERT_WLOCKED(object);
1737
1738         while (TRUE) {
1739                 /*
1740                  * Verify that the conditions are right for collapse:
1741                  *
1742                  * The object exists and the backing object exists.
1743                  */
1744                 if ((backing_object = object->backing_object) == NULL)
1745                         break;
1746
1747                 /*
1748                  * we check the backing object first, because it is most likely
1749                  * not collapsable.
1750                  */
1751                 if ((backing_object->flags & OBJ_ANON) == 0)
1752                         break;
1753                 VM_OBJECT_WLOCK(backing_object);
1754                 if (backing_object->handle != NULL ||
1755                     (backing_object->flags & OBJ_DEAD) != 0 ||
1756                     object->handle != NULL ||
1757                     (object->flags & OBJ_DEAD) != 0) {
1758                         VM_OBJECT_WUNLOCK(backing_object);
1759                         break;
1760                 }
1761
1762                 if (REFCOUNT_COUNT(object->paging_in_progress) > 0 ||
1763                     REFCOUNT_COUNT(backing_object->paging_in_progress) > 0) {
1764                         vm_object_qcollapse(object);
1765                         VM_OBJECT_WUNLOCK(backing_object);
1766                         break;
1767                 }
1768
1769                 /*
1770                  * We know that we can either collapse the backing object (if
1771                  * the parent is the only reference to it) or (perhaps) have
1772                  * the parent bypass the object if the parent happens to shadow
1773                  * all the resident pages in the entire backing object.
1774                  *
1775                  * This is ignoring pager-backed pages such as swap pages.
1776                  * vm_object_collapse_scan fails the shadowing test in this
1777                  * case.
1778                  */
1779                 if (backing_object->ref_count == 1) {
1780                         vm_object_pip_add(object, 1);
1781                         vm_object_pip_add(backing_object, 1);
1782
1783                         /*
1784                          * If there is exactly one reference to the backing
1785                          * object, we can collapse it into the parent.
1786                          */
1787                         vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT);
1788
1789 #if VM_NRESERVLEVEL > 0
1790                         /*
1791                          * Break any reservations from backing_object.
1792                          */
1793                         if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1794                                 vm_reserv_break_all(backing_object);
1795 #endif
1796
1797                         /*
1798                          * Move the pager from backing_object to object.
1799                          */
1800                         if (backing_object->type == OBJT_SWAP) {
1801                                 /*
1802                                  * swap_pager_copy() can sleep, in which case
1803                                  * the backing_object's and object's locks are
1804                                  * released and reacquired.
1805                                  * Since swap_pager_copy() is being asked to
1806                                  * destroy the source, it will change the
1807                                  * backing_object's type to OBJT_DEFAULT.
1808                                  */
1809                                 swap_pager_copy(
1810                                     backing_object,
1811                                     object,
1812                                     OFF_TO_IDX(object->backing_object_offset), TRUE);
1813                         }
1814                         /*
1815                          * Object now shadows whatever backing_object did.
1816                          * Note that the reference to 
1817                          * backing_object->backing_object moves from within 
1818                          * backing_object to within object.
1819                          */
1820                         vm_object_backing_remove_locked(object);
1821                         new_backing_object = backing_object->backing_object;
1822                         if (new_backing_object != NULL) {
1823                                 VM_OBJECT_WLOCK(new_backing_object);
1824                                 vm_object_backing_remove_locked(backing_object);
1825                                 vm_object_backing_insert_locked(object,
1826                                     new_backing_object);
1827                                 VM_OBJECT_WUNLOCK(new_backing_object);
1828                         }
1829                         object->backing_object_offset +=
1830                             backing_object->backing_object_offset;
1831
1832                         /*
1833                          * Discard backing_object.
1834                          *
1835                          * Since the backing object has no pages, no pager left,
1836                          * and no object references within it, all that is
1837                          * necessary is to dispose of it.
1838                          */
1839                         KASSERT(backing_object->ref_count == 1, (
1840 "backing_object %p was somehow re-referenced during collapse!",
1841                             backing_object));
1842                         vm_object_pip_wakeup(backing_object);
1843                         backing_object->type = OBJT_DEAD;
1844                         backing_object->ref_count = 0;
1845                         VM_OBJECT_WUNLOCK(backing_object);
1846                         vm_object_destroy(backing_object);
1847
1848                         vm_object_pip_wakeup(object);
1849                         counter_u64_add(object_collapses, 1);
1850                 } else {
1851                         /*
1852                          * If we do not entirely shadow the backing object,
1853                          * there is nothing we can do so we give up.
1854                          */
1855                         if (object->resident_page_count != object->size &&
1856                             !vm_object_scan_all_shadowed(object)) {
1857                                 VM_OBJECT_WUNLOCK(backing_object);
1858                                 break;
1859                         }
1860
1861                         /*
1862                          * Make the parent shadow the next object in the
1863                          * chain.  Deallocating backing_object will not remove
1864                          * it, since its reference count is at least 2.
1865                          */
1866                         vm_object_backing_remove_locked(object);
1867
1868                         new_backing_object = backing_object->backing_object;
1869                         if (new_backing_object != NULL) {
1870                                 vm_object_backing_insert(object,
1871                                     new_backing_object);
1872                                 vm_object_reference(new_backing_object);
1873                                 object->backing_object_offset +=
1874                                         backing_object->backing_object_offset;
1875                         }
1876
1877                         /*
1878                          * Drop the reference count on backing_object. Since
1879                          * its ref_count was at least 2, it will not vanish.
1880                          */
1881                         refcount_release(&backing_object->ref_count);
1882                         VM_OBJECT_WUNLOCK(backing_object);
1883                         counter_u64_add(object_bypasses, 1);
1884                 }
1885
1886                 /*
1887                  * Try again with this object's new backing object.
1888                  */
1889         }
1890 }
1891
1892 /*
1893  *      vm_object_page_remove:
1894  *
1895  *      For the given object, either frees or invalidates each of the
1896  *      specified pages.  In general, a page is freed.  However, if a page is
1897  *      wired for any reason other than the existence of a managed, wired
1898  *      mapping, then it may be invalidated but not removed from the object.
1899  *      Pages are specified by the given range ["start", "end") and the option
1900  *      OBJPR_CLEANONLY.  As a special case, if "end" is zero, then the range
1901  *      extends from "start" to the end of the object.  If the option
1902  *      OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1903  *      specified range are affected.  If the option OBJPR_NOTMAPPED is
1904  *      specified, then the pages within the specified range must have no
1905  *      mappings.  Otherwise, if this option is not specified, any mappings to
1906  *      the specified pages are removed before the pages are freed or
1907  *      invalidated.
1908  *
1909  *      In general, this operation should only be performed on objects that
1910  *      contain managed pages.  There are, however, two exceptions.  First, it
1911  *      is performed on the kernel and kmem objects by vm_map_entry_delete().
1912  *      Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1913  *      backed pages.  In both of these cases, the option OBJPR_CLEANONLY must
1914  *      not be specified and the option OBJPR_NOTMAPPED must be specified.
1915  *
1916  *      The object must be locked.
1917  */
1918 void
1919 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1920     int options)
1921 {
1922         vm_page_t p, next;
1923
1924         VM_OBJECT_ASSERT_WLOCKED(object);
1925         KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1926             (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1927             ("vm_object_page_remove: illegal options for object %p", object));
1928         if (object->resident_page_count == 0)
1929                 return;
1930         vm_object_pip_add(object, 1);
1931 again:
1932         p = vm_page_find_least(object, start);
1933
1934         /*
1935          * Here, the variable "p" is either (1) the page with the least pindex
1936          * greater than or equal to the parameter "start" or (2) NULL. 
1937          */
1938         for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1939                 next = TAILQ_NEXT(p, listq);
1940
1941                 /*
1942                  * If the page is wired for any reason besides the existence
1943                  * of managed, wired mappings, then it cannot be freed.  For
1944                  * example, fictitious pages, which represent device memory,
1945                  * are inherently wired and cannot be freed.  They can,
1946                  * however, be invalidated if the option OBJPR_CLEANONLY is
1947                  * not specified.
1948                  */
1949                 if (vm_page_tryxbusy(p) == 0) {
1950                         vm_page_sleep_if_busy(p, "vmopar");
1951                         goto again;
1952                 }
1953                 if (vm_page_wired(p)) {
1954 wired:
1955                         if ((options & OBJPR_NOTMAPPED) == 0 &&
1956                             object->ref_count != 0)
1957                                 pmap_remove_all(p);
1958                         if ((options & OBJPR_CLEANONLY) == 0) {
1959                                 vm_page_invalid(p);
1960                                 vm_page_undirty(p);
1961                         }
1962                         vm_page_xunbusy(p);
1963                         continue;
1964                 }
1965                 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1966                     ("vm_object_page_remove: page %p is fictitious", p));
1967                 if ((options & OBJPR_CLEANONLY) != 0 &&
1968                     !vm_page_none_valid(p)) {
1969                         if ((options & OBJPR_NOTMAPPED) == 0 &&
1970                             object->ref_count != 0 &&
1971                             !vm_page_try_remove_write(p))
1972                                 goto wired;
1973                         if (p->dirty != 0) {
1974                                 vm_page_xunbusy(p);
1975                                 continue;
1976                         }
1977                 }
1978                 if ((options & OBJPR_NOTMAPPED) == 0 &&
1979                     object->ref_count != 0 && !vm_page_try_remove_all(p))
1980                         goto wired;
1981                 vm_page_free(p);
1982         }
1983         vm_object_pip_wakeup(object);
1984 }
1985
1986 /*
1987  *      vm_object_page_noreuse:
1988  *
1989  *      For the given object, attempt to move the specified pages to
1990  *      the head of the inactive queue.  This bypasses regular LRU
1991  *      operation and allows the pages to be reused quickly under memory
1992  *      pressure.  If a page is wired for any reason, then it will not
1993  *      be queued.  Pages are specified by the range ["start", "end").
1994  *      As a special case, if "end" is zero, then the range extends from
1995  *      "start" to the end of the object.
1996  *
1997  *      This operation should only be performed on objects that
1998  *      contain non-fictitious, managed pages.
1999  *
2000  *      The object must be locked.
2001  */
2002 void
2003 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2004 {
2005         struct mtx *mtx;
2006         vm_page_t p, next;
2007
2008         VM_OBJECT_ASSERT_LOCKED(object);
2009         KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2010             ("vm_object_page_noreuse: illegal object %p", object));
2011         if (object->resident_page_count == 0)
2012                 return;
2013         p = vm_page_find_least(object, start);
2014
2015         /*
2016          * Here, the variable "p" is either (1) the page with the least pindex
2017          * greater than or equal to the parameter "start" or (2) NULL. 
2018          */
2019         mtx = NULL;
2020         for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2021                 next = TAILQ_NEXT(p, listq);
2022                 vm_page_change_lock(p, &mtx);
2023                 vm_page_deactivate_noreuse(p);
2024         }
2025         if (mtx != NULL)
2026                 mtx_unlock(mtx);
2027 }
2028
2029 /*
2030  *      Populate the specified range of the object with valid pages.  Returns
2031  *      TRUE if the range is successfully populated and FALSE otherwise.
2032  *
2033  *      Note: This function should be optimized to pass a larger array of
2034  *      pages to vm_pager_get_pages() before it is applied to a non-
2035  *      OBJT_DEVICE object.
2036  *
2037  *      The object must be locked.
2038  */
2039 boolean_t
2040 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2041 {
2042         vm_page_t m;
2043         vm_pindex_t pindex;
2044         int rv;
2045
2046         VM_OBJECT_ASSERT_WLOCKED(object);
2047         for (pindex = start; pindex < end; pindex++) {
2048                 rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2049                 if (rv != VM_PAGER_OK)
2050                         break;
2051
2052                 /*
2053                  * Keep "m" busy because a subsequent iteration may unlock
2054                  * the object.
2055                  */
2056         }
2057         if (pindex > start) {
2058                 m = vm_page_lookup(object, start);
2059                 while (m != NULL && m->pindex < pindex) {
2060                         vm_page_xunbusy(m);
2061                         m = TAILQ_NEXT(m, listq);
2062                 }
2063         }
2064         return (pindex == end);
2065 }
2066
2067 /*
2068  *      Routine:        vm_object_coalesce
2069  *      Function:       Coalesces two objects backing up adjoining
2070  *                      regions of memory into a single object.
2071  *
2072  *      returns TRUE if objects were combined.
2073  *
2074  *      NOTE:   Only works at the moment if the second object is NULL -
2075  *              if it's not, which object do we lock first?
2076  *
2077  *      Parameters:
2078  *              prev_object     First object to coalesce
2079  *              prev_offset     Offset into prev_object
2080  *              prev_size       Size of reference to prev_object
2081  *              next_size       Size of reference to the second object
2082  *              reserved        Indicator that extension region has
2083  *                              swap accounted for
2084  *
2085  *      Conditions:
2086  *      The object must *not* be locked.
2087  */
2088 boolean_t
2089 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2090     vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2091 {
2092         vm_pindex_t next_pindex;
2093
2094         if (prev_object == NULL)
2095                 return (TRUE);
2096         if ((prev_object->flags & OBJ_ANON) == 0)
2097                 return (FALSE);
2098
2099         VM_OBJECT_WLOCK(prev_object);
2100         /*
2101          * Try to collapse the object first
2102          */
2103         vm_object_collapse(prev_object);
2104
2105         /*
2106          * Can't coalesce if: . more than one reference . paged out . shadows
2107          * another object . has a copy elsewhere (any of which mean that the
2108          * pages not mapped to prev_entry may be in use anyway)
2109          */
2110         if (prev_object->backing_object != NULL) {
2111                 VM_OBJECT_WUNLOCK(prev_object);
2112                 return (FALSE);
2113         }
2114
2115         prev_size >>= PAGE_SHIFT;
2116         next_size >>= PAGE_SHIFT;
2117         next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2118
2119         if (prev_object->ref_count > 1 &&
2120             prev_object->size != next_pindex &&
2121             (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2122                 VM_OBJECT_WUNLOCK(prev_object);
2123                 return (FALSE);
2124         }
2125
2126         /*
2127          * Account for the charge.
2128          */
2129         if (prev_object->cred != NULL) {
2130
2131                 /*
2132                  * If prev_object was charged, then this mapping,
2133                  * although not charged now, may become writable
2134                  * later. Non-NULL cred in the object would prevent
2135                  * swap reservation during enabling of the write
2136                  * access, so reserve swap now. Failed reservation
2137                  * cause allocation of the separate object for the map
2138                  * entry, and swap reservation for this entry is
2139                  * managed in appropriate time.
2140                  */
2141                 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2142                     prev_object->cred)) {
2143                         VM_OBJECT_WUNLOCK(prev_object);
2144                         return (FALSE);
2145                 }
2146                 prev_object->charge += ptoa(next_size);
2147         }
2148
2149         /*
2150          * Remove any pages that may still be in the object from a previous
2151          * deallocation.
2152          */
2153         if (next_pindex < prev_object->size) {
2154                 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2155                     next_size, 0);
2156                 if (prev_object->type == OBJT_SWAP)
2157                         swap_pager_freespace(prev_object,
2158                                              next_pindex, next_size);
2159 #if 0
2160                 if (prev_object->cred != NULL) {
2161                         KASSERT(prev_object->charge >=
2162                             ptoa(prev_object->size - next_pindex),
2163                             ("object %p overcharged 1 %jx %jx", prev_object,
2164                                 (uintmax_t)next_pindex, (uintmax_t)next_size));
2165                         prev_object->charge -= ptoa(prev_object->size -
2166                             next_pindex);
2167                 }
2168 #endif
2169         }
2170
2171         /*
2172          * Extend the object if necessary.
2173          */
2174         if (next_pindex + next_size > prev_object->size)
2175                 prev_object->size = next_pindex + next_size;
2176
2177         VM_OBJECT_WUNLOCK(prev_object);
2178         return (TRUE);
2179 }
2180
2181 void
2182 vm_object_set_writeable_dirty(vm_object_t object)
2183 {
2184
2185         VM_OBJECT_ASSERT_LOCKED(object);
2186
2187         /* Only set for vnodes & tmpfs */
2188         if (object->type != OBJT_VNODE &&
2189             (object->flags & OBJ_TMPFS_NODE) == 0)
2190                 return;
2191         atomic_add_int(&object->generation, 1);
2192 }
2193
2194 /*
2195  *      vm_object_unwire:
2196  *
2197  *      For each page offset within the specified range of the given object,
2198  *      find the highest-level page in the shadow chain and unwire it.  A page
2199  *      must exist at every page offset, and the highest-level page must be
2200  *      wired.
2201  */
2202 void
2203 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2204     uint8_t queue)
2205 {
2206         vm_object_t tobject, t1object;
2207         vm_page_t m, tm;
2208         vm_pindex_t end_pindex, pindex, tpindex;
2209         int depth, locked_depth;
2210
2211         KASSERT((offset & PAGE_MASK) == 0,
2212             ("vm_object_unwire: offset is not page aligned"));
2213         KASSERT((length & PAGE_MASK) == 0,
2214             ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2215         /* The wired count of a fictitious page never changes. */
2216         if ((object->flags & OBJ_FICTITIOUS) != 0)
2217                 return;
2218         pindex = OFF_TO_IDX(offset);
2219         end_pindex = pindex + atop(length);
2220 again:
2221         locked_depth = 1;
2222         VM_OBJECT_RLOCK(object);
2223         m = vm_page_find_least(object, pindex);
2224         while (pindex < end_pindex) {
2225                 if (m == NULL || pindex < m->pindex) {
2226                         /*
2227                          * The first object in the shadow chain doesn't
2228                          * contain a page at the current index.  Therefore,
2229                          * the page must exist in a backing object.
2230                          */
2231                         tobject = object;
2232                         tpindex = pindex;
2233                         depth = 0;
2234                         do {
2235                                 tpindex +=
2236                                     OFF_TO_IDX(tobject->backing_object_offset);
2237                                 tobject = tobject->backing_object;
2238                                 KASSERT(tobject != NULL,
2239                                     ("vm_object_unwire: missing page"));
2240                                 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2241                                         goto next_page;
2242                                 depth++;
2243                                 if (depth == locked_depth) {
2244                                         locked_depth++;
2245                                         VM_OBJECT_RLOCK(tobject);
2246                                 }
2247                         } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2248                             NULL);
2249                 } else {
2250                         tm = m;
2251                         m = TAILQ_NEXT(m, listq);
2252                 }
2253                 if (vm_page_trysbusy(tm) == 0) {
2254                         for (tobject = object; locked_depth >= 1;
2255                             locked_depth--) {
2256                                 t1object = tobject->backing_object;
2257                                 if (tm->object != tobject)
2258                                         VM_OBJECT_RUNLOCK(tobject);
2259                                 tobject = t1object;
2260                         }
2261                         vm_page_busy_sleep(tm, "unwbo", true);
2262                         goto again;
2263                 }
2264                 vm_page_unwire(tm, queue);
2265                 vm_page_sunbusy(tm);
2266 next_page:
2267                 pindex++;
2268         }
2269         /* Release the accumulated object locks. */
2270         for (tobject = object; locked_depth >= 1; locked_depth--) {
2271                 t1object = tobject->backing_object;
2272                 VM_OBJECT_RUNLOCK(tobject);
2273                 tobject = t1object;
2274         }
2275 }
2276
2277 /*
2278  * Return the vnode for the given object, or NULL if none exists.
2279  * For tmpfs objects, the function may return NULL if there is
2280  * no vnode allocated at the time of the call.
2281  */
2282 struct vnode *
2283 vm_object_vnode(vm_object_t object)
2284 {
2285         struct vnode *vp;
2286
2287         VM_OBJECT_ASSERT_LOCKED(object);
2288         if (object->type == OBJT_VNODE) {
2289                 vp = object->handle;
2290                 KASSERT(vp != NULL, ("%s: OBJT_VNODE has no vnode", __func__));
2291         } else if (object->type == OBJT_SWAP &&
2292             (object->flags & OBJ_TMPFS) != 0) {
2293                 vp = object->un_pager.swp.swp_tmpfs;
2294                 KASSERT(vp != NULL, ("%s: OBJT_TMPFS has no vnode", __func__));
2295         } else {
2296                 vp = NULL;
2297         }
2298         return (vp);
2299 }
2300
2301
2302 /*
2303  * Busy the vm object.  This prevents new pages belonging to the object from
2304  * becoming busy.  Existing pages persist as busy.  Callers are responsible
2305  * for checking page state before proceeding.
2306  */
2307 void
2308 vm_object_busy(vm_object_t obj)
2309 {
2310
2311         VM_OBJECT_ASSERT_LOCKED(obj);
2312
2313         refcount_acquire(&obj->busy);
2314         /* The fence is required to order loads of page busy. */
2315         atomic_thread_fence_acq_rel();
2316 }
2317
2318 void
2319 vm_object_unbusy(vm_object_t obj)
2320 {
2321
2322
2323         refcount_release(&obj->busy);
2324 }
2325
2326 void
2327 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2328 {
2329
2330         VM_OBJECT_ASSERT_UNLOCKED(obj);
2331
2332         if (obj->busy)
2333                 refcount_sleep(&obj->busy, wmesg, PVM);
2334 }
2335
2336 /*
2337  * Return the kvme type of the given object.
2338  * If vpp is not NULL, set it to the object's vm_object_vnode() or NULL.
2339  */
2340 int
2341 vm_object_kvme_type(vm_object_t object, struct vnode **vpp)
2342 {
2343
2344         VM_OBJECT_ASSERT_LOCKED(object);
2345         if (vpp != NULL)
2346                 *vpp = vm_object_vnode(object);
2347         switch (object->type) {
2348         case OBJT_DEFAULT:
2349                 return (KVME_TYPE_DEFAULT);
2350         case OBJT_VNODE:
2351                 return (KVME_TYPE_VNODE);
2352         case OBJT_SWAP:
2353                 if ((object->flags & OBJ_TMPFS_NODE) != 0)
2354                         return (KVME_TYPE_VNODE);
2355                 return (KVME_TYPE_SWAP);
2356         case OBJT_DEVICE:
2357                 return (KVME_TYPE_DEVICE);
2358         case OBJT_PHYS:
2359                 return (KVME_TYPE_PHYS);
2360         case OBJT_DEAD:
2361                 return (KVME_TYPE_DEAD);
2362         case OBJT_SG:
2363                 return (KVME_TYPE_SG);
2364         case OBJT_MGTDEVICE:
2365                 return (KVME_TYPE_MGTDEVICE);
2366         default:
2367                 return (KVME_TYPE_UNKNOWN);
2368         }
2369 }
2370
2371 static int
2372 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2373 {
2374         struct kinfo_vmobject *kvo;
2375         char *fullpath, *freepath;
2376         struct vnode *vp;
2377         struct vattr va;
2378         vm_object_t obj;
2379         vm_page_t m;
2380         int count, error;
2381
2382         if (req->oldptr == NULL) {
2383                 /*
2384                  * If an old buffer has not been provided, generate an
2385                  * estimate of the space needed for a subsequent call.
2386                  */
2387                 mtx_lock(&vm_object_list_mtx);
2388                 count = 0;
2389                 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2390                         if (obj->type == OBJT_DEAD)
2391                                 continue;
2392                         count++;
2393                 }
2394                 mtx_unlock(&vm_object_list_mtx);
2395                 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2396                     count * 11 / 10));
2397         }
2398
2399         kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK);
2400         error = 0;
2401
2402         /*
2403          * VM objects are type stable and are never removed from the
2404          * list once added.  This allows us to safely read obj->object_list
2405          * after reacquiring the VM object lock.
2406          */
2407         mtx_lock(&vm_object_list_mtx);
2408         TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2409                 if (obj->type == OBJT_DEAD)
2410                         continue;
2411                 VM_OBJECT_RLOCK(obj);
2412                 if (obj->type == OBJT_DEAD) {
2413                         VM_OBJECT_RUNLOCK(obj);
2414                         continue;
2415                 }
2416                 mtx_unlock(&vm_object_list_mtx);
2417                 kvo->kvo_size = ptoa(obj->size);
2418                 kvo->kvo_resident = obj->resident_page_count;
2419                 kvo->kvo_ref_count = obj->ref_count;
2420                 kvo->kvo_shadow_count = obj->shadow_count;
2421                 kvo->kvo_memattr = obj->memattr;
2422                 kvo->kvo_active = 0;
2423                 kvo->kvo_inactive = 0;
2424                 TAILQ_FOREACH(m, &obj->memq, listq) {
2425                         /*
2426                          * A page may belong to the object but be
2427                          * dequeued and set to PQ_NONE while the
2428                          * object lock is not held.  This makes the
2429                          * reads of m->queue below racy, and we do not
2430                          * count pages set to PQ_NONE.  However, this
2431                          * sysctl is only meant to give an
2432                          * approximation of the system anyway.
2433                          */
2434                         if (m->queue == PQ_ACTIVE)
2435                                 kvo->kvo_active++;
2436                         else if (m->queue == PQ_INACTIVE)
2437                                 kvo->kvo_inactive++;
2438                 }
2439
2440                 kvo->kvo_vn_fileid = 0;
2441                 kvo->kvo_vn_fsid = 0;
2442                 kvo->kvo_vn_fsid_freebsd11 = 0;
2443                 freepath = NULL;
2444                 fullpath = "";
2445                 kvo->kvo_type = vm_object_kvme_type(obj, &vp);
2446                 if (vp != NULL)
2447                         vref(vp);
2448                 VM_OBJECT_RUNLOCK(obj);
2449                 if (vp != NULL) {
2450                         vn_fullpath(curthread, vp, &fullpath, &freepath);
2451                         vn_lock(vp, LK_SHARED | LK_RETRY);
2452                         if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2453                                 kvo->kvo_vn_fileid = va.va_fileid;
2454                                 kvo->kvo_vn_fsid = va.va_fsid;
2455                                 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2456                                                                 /* truncate */
2457                         }
2458                         vput(vp);
2459                 }
2460
2461                 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2462                 if (freepath != NULL)
2463                         free(freepath, M_TEMP);
2464
2465                 /* Pack record size down */
2466                 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2467                     + strlen(kvo->kvo_path) + 1;
2468                 kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2469                     sizeof(uint64_t));
2470                 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2471                 mtx_lock(&vm_object_list_mtx);
2472                 if (error)
2473                         break;
2474         }
2475         mtx_unlock(&vm_object_list_mtx);
2476         free(kvo, M_TEMP);
2477         return (error);
2478 }
2479 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2480     CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2481     "List of VM objects");
2482
2483 #include "opt_ddb.h"
2484 #ifdef DDB
2485 #include <sys/kernel.h>
2486
2487 #include <sys/cons.h>
2488
2489 #include <ddb/ddb.h>
2490
2491 static int
2492 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2493 {
2494         vm_map_t tmpm;
2495         vm_map_entry_t tmpe;
2496         vm_object_t obj;
2497
2498         if (map == 0)
2499                 return 0;
2500
2501         if (entry == 0) {
2502                 VM_MAP_ENTRY_FOREACH(tmpe, map) {
2503                         if (_vm_object_in_map(map, object, tmpe)) {
2504                                 return 1;
2505                         }
2506                 }
2507         } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2508                 tmpm = entry->object.sub_map;
2509                 VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2510                         if (_vm_object_in_map(tmpm, object, tmpe)) {
2511                                 return 1;
2512                         }
2513                 }
2514         } else if ((obj = entry->object.vm_object) != NULL) {
2515                 for (; obj; obj = obj->backing_object)
2516                         if (obj == object) {
2517                                 return 1;
2518                         }
2519         }
2520         return 0;
2521 }
2522
2523 static int
2524 vm_object_in_map(vm_object_t object)
2525 {
2526         struct proc *p;
2527
2528         /* sx_slock(&allproc_lock); */
2529         FOREACH_PROC_IN_SYSTEM(p) {
2530                 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2531                         continue;
2532                 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2533                         /* sx_sunlock(&allproc_lock); */
2534                         return 1;
2535                 }
2536         }
2537         /* sx_sunlock(&allproc_lock); */
2538         if (_vm_object_in_map(kernel_map, object, 0))
2539                 return 1;
2540         return 0;
2541 }
2542
2543 DB_SHOW_COMMAND(vmochk, vm_object_check)
2544 {
2545         vm_object_t object;
2546
2547         /*
2548          * make sure that internal objs are in a map somewhere
2549          * and none have zero ref counts.
2550          */
2551         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2552                 if (object->handle == NULL &&
2553                     (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2554                         if (object->ref_count == 0) {
2555                                 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2556                                         (long)object->size);
2557                         }
2558                         if (!vm_object_in_map(object)) {
2559                                 db_printf(
2560                         "vmochk: internal obj is not in a map: "
2561                         "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2562                                     object->ref_count, (u_long)object->size, 
2563                                     (u_long)object->size,
2564                                     (void *)object->backing_object);
2565                         }
2566                 }
2567         }
2568 }
2569
2570 /*
2571  *      vm_object_print:        [ debug ]
2572  */
2573 DB_SHOW_COMMAND(object, vm_object_print_static)
2574 {
2575         /* XXX convert args. */
2576         vm_object_t object = (vm_object_t)addr;
2577         boolean_t full = have_addr;
2578
2579         vm_page_t p;
2580
2581         /* XXX count is an (unused) arg.  Avoid shadowing it. */
2582 #define count   was_count
2583
2584         int count;
2585
2586         if (object == NULL)
2587                 return;
2588
2589         db_iprintf(
2590             "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2591             object, (int)object->type, (uintmax_t)object->size,
2592             object->resident_page_count, object->ref_count, object->flags,
2593             object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2594         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2595             object->shadow_count, 
2596             object->backing_object ? object->backing_object->ref_count : 0,
2597             object->backing_object, (uintmax_t)object->backing_object_offset);
2598
2599         if (!full)
2600                 return;
2601
2602         db_indent += 2;
2603         count = 0;
2604         TAILQ_FOREACH(p, &object->memq, listq) {
2605                 if (count == 0)
2606                         db_iprintf("memory:=");
2607                 else if (count == 6) {
2608                         db_printf("\n");
2609                         db_iprintf(" ...");
2610                         count = 0;
2611                 } else
2612                         db_printf(",");
2613                 count++;
2614
2615                 db_printf("(off=0x%jx,page=0x%jx)",
2616                     (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2617         }
2618         if (count != 0)
2619                 db_printf("\n");
2620         db_indent -= 2;
2621 }
2622
2623 /* XXX. */
2624 #undef count
2625
2626 /* XXX need this non-static entry for calling from vm_map_print. */
2627 void
2628 vm_object_print(
2629         /* db_expr_t */ long addr,
2630         boolean_t have_addr,
2631         /* db_expr_t */ long count,
2632         char *modif)
2633 {
2634         vm_object_print_static(addr, have_addr, count, modif);
2635 }
2636
2637 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2638 {
2639         vm_object_t object;
2640         vm_pindex_t fidx;
2641         vm_paddr_t pa;
2642         vm_page_t m, prev_m;
2643         int rcount, nl, c;
2644
2645         nl = 0;
2646         TAILQ_FOREACH(object, &vm_object_list, object_list) {
2647                 db_printf("new object: %p\n", (void *)object);
2648                 if (nl > 18) {
2649                         c = cngetc();
2650                         if (c != ' ')
2651                                 return;
2652                         nl = 0;
2653                 }
2654                 nl++;
2655                 rcount = 0;
2656                 fidx = 0;
2657                 pa = -1;
2658                 TAILQ_FOREACH(m, &object->memq, listq) {
2659                         if (m->pindex > 128)
2660                                 break;
2661                         if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2662                             prev_m->pindex + 1 != m->pindex) {
2663                                 if (rcount) {
2664                                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2665                                                 (long)fidx, rcount, (long)pa);
2666                                         if (nl > 18) {
2667                                                 c = cngetc();
2668                                                 if (c != ' ')
2669                                                         return;
2670                                                 nl = 0;
2671                                         }
2672                                         nl++;
2673                                         rcount = 0;
2674                                 }
2675                         }                               
2676                         if (rcount &&
2677                                 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2678                                 ++rcount;
2679                                 continue;
2680                         }
2681                         if (rcount) {
2682                                 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2683                                         (long)fidx, rcount, (long)pa);
2684                                 if (nl > 18) {
2685                                         c = cngetc();
2686                                         if (c != ' ')
2687                                                 return;
2688                                         nl = 0;
2689                                 }
2690                                 nl++;
2691                         }
2692                         fidx = m->pindex;
2693                         pa = VM_PAGE_TO_PHYS(m);
2694                         rcount = 1;
2695                 }
2696                 if (rcount) {
2697                         db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2698                                 (long)fidx, rcount, (long)pa);
2699                         if (nl > 18) {
2700                                 c = cngetc();
2701                                 if (c != ' ')
2702                                         return;
2703                                 nl = 0;
2704                         }
2705                         nl++;
2706                 }
2707         }
2708 }
2709 #endif /* DDB */