2 * Copyright (c) 1991, 1993, 2013
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94
35 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36 * All rights reserved.
38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
40 * Permission to use, copy, modify and distribute this software and
41 * its documentation is hereby granted, provided that both the copyright
42 * notice and this permission notice appear in all copies of the
43 * software, derivative works or modified versions, and any portions
44 * thereof, and that both notices appear in supporting documentation.
46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
50 * Carnegie Mellon requests users of this software to return to
52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
53 * School of Computer Science
54 * Carnegie Mellon University
55 * Pittsburgh PA 15213-3890
57 * any improvements or extensions that they make and grant Carnegie the
58 * rights to redistribute these changes.
60 * $FreeBSD: src/sys/vm/vm_object.c,v 1.171.2.8 2003/05/26 19:17:56 alc Exp $
64 * Virtual memory object module.
67 #include <sys/param.h>
68 #include <sys/systm.h>
69 #include <sys/proc.h> /* for curproc, pageproc */
70 #include <sys/thread.h>
71 #include <sys/vnode.h>
72 #include <sys/vmmeter.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/refcount.h>
80 #include <vm/vm_param.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_pager.h>
87 #include <vm/swap_pager.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_extern.h>
90 #include <vm/vm_zone.h>
92 #include <vm/vm_page2.h>
94 #include <machine/specialreg.h>
96 #define EASY_SCAN_FACTOR 8
98 static void vm_object_qcollapse(vm_object_t object,
99 vm_object_t backing_object);
100 static void vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
102 static void vm_object_lock_init(vm_object_t);
106 * Virtual memory objects maintain the actual data
107 * associated with allocated virtual memory. A given
108 * page of memory exists within exactly one object.
110 * An object is only deallocated when all "references"
111 * are given up. Only one "reference" to a given
112 * region of an object should be writeable.
114 * Associated with each object is a list of all resident
115 * memory pages belonging to that object; this list is
116 * maintained by the "vm_page" module, and locked by the object's
119 * Each object also records a "pager" routine which is
120 * used to retrieve (and store) pages to the proper backing
121 * storage. In addition, objects may be backed by other
122 * objects from which they were virtual-copied.
124 * The only items within the object structure which are
125 * modified after time of creation are:
126 * reference count locked by object's lock
127 * pager routine locked by object's lock
131 struct object_q vm_object_list; /* locked by vmobj_token */
132 struct vm_object kernel_object;
134 static long vm_object_count; /* locked by vmobj_token */
136 static long object_collapses;
137 static long object_bypasses;
138 static int next_index;
139 static vm_zone_t obj_zone;
140 static struct vm_zone obj_zone_store;
141 #define VM_OBJECTS_INIT 256
142 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
145 * Misc low level routines
148 vm_object_lock_init(vm_object_t obj)
150 #if defined(DEBUG_LOCKS)
153 obj->debug_hold_bitmap = 0;
154 obj->debug_hold_ovfl = 0;
155 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) {
156 obj->debug_hold_thrs[i] = NULL;
157 obj->debug_hold_file[i] = NULL;
158 obj->debug_hold_line[i] = 0;
164 vm_object_lock_swap(void)
170 vm_object_lock(vm_object_t obj)
172 lwkt_gettoken(&obj->token);
176 * Returns TRUE on sucesss
179 vm_object_lock_try(vm_object_t obj)
181 return(lwkt_trytoken(&obj->token));
185 vm_object_lock_shared(vm_object_t obj)
187 lwkt_gettoken_shared(&obj->token);
191 vm_object_unlock(vm_object_t obj)
193 lwkt_reltoken(&obj->token);
197 vm_object_upgrade(vm_object_t obj)
199 lwkt_reltoken(&obj->token);
200 lwkt_gettoken(&obj->token);
204 vm_object_downgrade(vm_object_t obj)
206 lwkt_reltoken(&obj->token);
207 lwkt_gettoken_shared(&obj->token);
211 vm_object_assert_held(vm_object_t obj)
213 ASSERT_LWKT_TOKEN_HELD(&obj->token);
218 vm_object_hold(vm_object_t obj)
220 debugvm_object_hold(vm_object_t obj, char *file, int line)
223 KKASSERT(obj != NULL);
226 * Object must be held (object allocation is stable due to callers
227 * context, typically already holding the token on a parent object)
228 * prior to potentially blocking on the lock, otherwise the object
229 * can get ripped away from us.
231 refcount_acquire(&obj->hold_count);
234 #if defined(DEBUG_LOCKS)
239 mask = ~obj->debug_hold_bitmap;
241 if (mask == 0xFFFFFFFFU) {
242 if (obj->debug_hold_ovfl == 0)
243 obj->debug_hold_ovfl = 1;
247 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask,
249 obj->debug_hold_bitmap |= (1 << i);
250 obj->debug_hold_thrs[i] = curthread;
251 obj->debug_hold_file[i] = file;
252 obj->debug_hold_line[i] = line;
261 vm_object_hold_try(vm_object_t obj)
263 debugvm_object_hold_try(vm_object_t obj, char *file, int line)
266 KKASSERT(obj != NULL);
269 * Object must be held (object allocation is stable due to callers
270 * context, typically already holding the token on a parent object)
271 * prior to potentially blocking on the lock, otherwise the object
272 * can get ripped away from us.
274 refcount_acquire(&obj->hold_count);
275 if (vm_object_lock_try(obj) == 0) {
276 if (refcount_release(&obj->hold_count)) {
277 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD))
278 zfree(obj_zone, obj);
283 #if defined(DEBUG_LOCKS)
288 mask = ~obj->debug_hold_bitmap;
290 if (mask == 0xFFFFFFFFU) {
291 if (obj->debug_hold_ovfl == 0)
292 obj->debug_hold_ovfl = 1;
296 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask,
298 obj->debug_hold_bitmap |= (1 << i);
299 obj->debug_hold_thrs[i] = curthread;
300 obj->debug_hold_file[i] = file;
301 obj->debug_hold_line[i] = line;
311 vm_object_hold_shared(vm_object_t obj)
313 debugvm_object_hold_shared(vm_object_t obj, char *file, int line)
316 KKASSERT(obj != NULL);
319 * Object must be held (object allocation is stable due to callers
320 * context, typically already holding the token on a parent object)
321 * prior to potentially blocking on the lock, otherwise the object
322 * can get ripped away from us.
324 refcount_acquire(&obj->hold_count);
325 vm_object_lock_shared(obj);
327 #if defined(DEBUG_LOCKS)
332 mask = ~obj->debug_hold_bitmap;
334 if (mask == 0xFFFFFFFFU) {
335 if (obj->debug_hold_ovfl == 0)
336 obj->debug_hold_ovfl = 1;
340 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask,
342 obj->debug_hold_bitmap |= (1 << i);
343 obj->debug_hold_thrs[i] = curthread;
344 obj->debug_hold_file[i] = file;
345 obj->debug_hold_line[i] = line;
355 * Obtain either a shared or exclusive lock on VM object
356 * based on whether this is a terminal vnode object or not.
360 vm_object_hold_maybe_shared(vm_object_t obj)
362 debugvm_object_hold_maybe_shared(vm_object_t obj, char *file, int line)
365 if (vm_shared_fault &&
366 obj->type == OBJT_VNODE &&
367 obj->backing_object == NULL) {
368 vm_object_hold_shared(obj);
379 * Drop the token and hold_count on the object.
381 * WARNING! Token might be shared.
384 vm_object_drop(vm_object_t obj)
389 #if defined(DEBUG_LOCKS)
393 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) {
394 if ((obj->debug_hold_bitmap & (1 << i)) &&
395 (obj->debug_hold_thrs[i] == curthread)) {
396 obj->debug_hold_bitmap &= ~(1 << i);
397 obj->debug_hold_thrs[i] = NULL;
398 obj->debug_hold_file[i] = NULL;
399 obj->debug_hold_line[i] = 0;
405 if (found == 0 && obj->debug_hold_ovfl == 0)
406 panic("vm_object: attempt to drop hold on non-self-held obj");
410 * No new holders should be possible once we drop hold_count 1->0 as
411 * there is no longer any way to reference the object.
413 KKASSERT(obj->hold_count > 0);
414 if (refcount_release(&obj->hold_count)) {
415 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD)) {
416 vm_object_unlock(obj);
417 zfree(obj_zone, obj);
419 vm_object_unlock(obj);
422 vm_object_unlock(obj);
427 * Initialize a freshly allocated object, returning a held object.
429 * Used only by vm_object_allocate() and zinitna().
434 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
438 RB_INIT(&object->rb_memq);
439 LIST_INIT(&object->shadow_head);
440 lwkt_token_init(&object->token, "vmobj");
444 object->ref_count = 1;
445 object->memattr = VM_MEMATTR_DEFAULT;
446 object->hold_count = 0;
448 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
449 vm_object_set_flag(object, OBJ_ONEMAPPING);
450 object->paging_in_progress = 0;
451 object->resident_page_count = 0;
452 object->agg_pv_list_count = 0;
453 object->shadow_count = 0;
454 /* cpu localization twist */
455 object->pg_color = (int)(intptr_t)curthread;
456 if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
457 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
460 next_index = (next_index + incr) & PQ_L2_MASK;
461 object->handle = NULL;
462 object->backing_object = NULL;
463 object->backing_object_offset = (vm_ooffset_t)0;
465 object->generation++;
466 object->swblock_count = 0;
467 RB_INIT(&object->swblock_root);
468 vm_object_lock_init(object);
469 pmap_object_init(object);
471 vm_object_hold(object);
472 lwkt_gettoken(&vmobj_token);
473 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
475 lwkt_reltoken(&vmobj_token);
479 * Initialize the VM objects module.
481 * Called from the low level boot code only.
486 TAILQ_INIT(&vm_object_list);
488 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd),
490 vm_object_drop(&kernel_object);
492 obj_zone = &obj_zone_store;
493 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
494 vm_objects_init, VM_OBJECTS_INIT);
498 vm_object_init2(void)
500 zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1);
504 * Allocate and return a new object of the specified type and size.
509 vm_object_allocate(objtype_t type, vm_pindex_t size)
513 result = (vm_object_t) zalloc(obj_zone);
515 _vm_object_allocate(type, size, result);
516 vm_object_drop(result);
522 * This version returns a held object, allowing further atomic initialization
526 vm_object_allocate_hold(objtype_t type, vm_pindex_t size)
530 result = (vm_object_t) zalloc(obj_zone);
532 _vm_object_allocate(type, size, result);
538 * Add an additional reference to a vm_object. The object must already be
539 * held. The original non-lock version is no longer supported. The object
540 * must NOT be chain locked by anyone at the time the reference is added.
542 * Referencing a chain-locked object can blow up the fairly sensitive
543 * ref_count and shadow_count tests in the deallocator. Most callers
544 * will call vm_object_chain_wait() prior to calling
545 * vm_object_reference_locked() to avoid the case.
547 * The object must be held, but may be held shared if desired (hence why
548 * we use an atomic op).
551 vm_object_reference_locked(vm_object_t object)
553 KKASSERT(object != NULL);
554 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
555 KKASSERT((object->chainlk & (CHAINLK_EXCL | CHAINLK_MASK)) == 0);
556 atomic_add_int(&object->ref_count, 1);
557 if (object->type == OBJT_VNODE) {
558 vref(object->handle);
559 /* XXX what if the vnode is being destroyed? */
564 * This version is only allowed for vnode objects.
567 vm_object_reference_quick(vm_object_t object)
569 KKASSERT(object->type == OBJT_VNODE);
570 atomic_add_int(&object->ref_count, 1);
571 vref(object->handle);
575 * Object OBJ_CHAINLOCK lock handling.
577 * The caller can chain-lock backing objects recursively and then
578 * use vm_object_chain_release_all() to undo the whole chain.
580 * Chain locks are used to prevent collapses and are only applicable
581 * to OBJT_DEFAULT and OBJT_SWAP objects. Chain locking operations
582 * on other object types are ignored. This is also important because
583 * it allows e.g. the vnode underlying a memory mapping to take concurrent
586 * The object must usually be held on entry, though intermediate
587 * objects need not be held on release. The object must be held exclusively,
588 * NOT shared. Note that the prefault path checks the shared state and
589 * avoids using the chain functions.
592 vm_object_chain_wait(vm_object_t object, int shared)
594 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
596 uint32_t chainlk = object->chainlk;
600 if (chainlk & (CHAINLK_EXCL | CHAINLK_EXCLREQ)) {
601 tsleep_interlock(object, 0);
602 if (atomic_cmpset_int(&object->chainlk,
604 chainlk | CHAINLK_WAIT)) {
605 tsleep(object, PINTERLOCKED,
614 if (chainlk & (CHAINLK_MASK | CHAINLK_EXCL)) {
615 tsleep_interlock(object, 0);
616 if (atomic_cmpset_int(&object->chainlk,
618 chainlk | CHAINLK_WAIT))
620 tsleep(object, PINTERLOCKED,
625 if (atomic_cmpset_int(&object->chainlk,
627 chainlk & ~CHAINLK_WAIT))
629 if (chainlk & CHAINLK_WAIT)
641 vm_object_chain_acquire(vm_object_t object, int shared)
643 if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP)
645 if (vm_shared_fault == 0)
649 uint32_t chainlk = object->chainlk;
653 if (chainlk & (CHAINLK_EXCL | CHAINLK_EXCLREQ)) {
654 tsleep_interlock(object, 0);
655 if (atomic_cmpset_int(&object->chainlk,
657 chainlk | CHAINLK_WAIT)) {
658 tsleep(object, PINTERLOCKED,
662 } else if (atomic_cmpset_int(&object->chainlk,
663 chainlk, chainlk + 1)) {
668 if (chainlk & (CHAINLK_MASK | CHAINLK_EXCL)) {
669 tsleep_interlock(object, 0);
670 if (atomic_cmpset_int(&object->chainlk,
675 tsleep(object, PINTERLOCKED,
680 if (atomic_cmpset_int(&object->chainlk,
682 (chainlk | CHAINLK_EXCL) &
685 if (chainlk & CHAINLK_WAIT)
697 vm_object_chain_release(vm_object_t object)
699 /*ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));*/
700 if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP)
702 KKASSERT(object->chainlk & (CHAINLK_MASK | CHAINLK_EXCL));
704 uint32_t chainlk = object->chainlk;
707 if (chainlk & CHAINLK_MASK) {
708 if ((chainlk & CHAINLK_MASK) == 1 &&
709 atomic_cmpset_int(&object->chainlk,
711 (chainlk - 1) & ~CHAINLK_WAIT)) {
712 if (chainlk & CHAINLK_WAIT)
716 if ((chainlk & CHAINLK_MASK) > 1 &&
717 atomic_cmpset_int(&object->chainlk,
718 chainlk, chainlk - 1)) {
723 KKASSERT(chainlk & CHAINLK_EXCL);
724 if (atomic_cmpset_int(&object->chainlk,
726 chainlk & ~(CHAINLK_EXCL |
728 if (chainlk & CHAINLK_WAIT)
737 * Release the chain from first_object through and including stopobj.
738 * The caller is typically holding the first and last object locked
739 * (shared or exclusive) to prevent destruction races.
741 * We release stopobj first as an optimization as this object is most
742 * likely to be shared across multiple processes.
745 vm_object_chain_release_all(vm_object_t first_object, vm_object_t stopobj)
747 vm_object_t backing_object;
750 vm_object_chain_release(stopobj);
751 object = first_object;
753 while (object != stopobj) {
756 /* shouldn't need this since chain is held */
757 if (object != first_object)
758 vm_object_hold(object);
760 backing_object = object->backing_object;
761 vm_object_chain_release(object);
763 if (object != first_object)
764 vm_object_drop(object);
766 object = backing_object;
771 * Dereference an object and its underlying vnode.
773 * The object must be held exclusively and will remain held on return.
774 * (We don't need an atomic op due to the exclusivity).
777 vm_object_vndeallocate(vm_object_t object)
779 struct vnode *vp = (struct vnode *) object->handle;
781 KASSERT(object->type == OBJT_VNODE,
782 ("vm_object_vndeallocate: not a vnode object"));
783 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
784 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
786 if (object->ref_count == 0) {
787 vprint("vm_object_vndeallocate", vp);
788 panic("vm_object_vndeallocate: bad object reference count");
791 atomic_add_int(&object->ref_count, -1);
792 if (object->ref_count == 0)
793 vclrflags(vp, VTEXT);
798 * Release a reference to the specified object, gained either through a
799 * vm_object_allocate or a vm_object_reference call. When all references
800 * are gone, storage associated with this object may be relinquished.
802 * The caller does not have to hold the object locked but must have control
803 * over the reference in question in order to guarantee that the object
804 * does not get ripped out from under us.
806 * XXX Currently all deallocations require an exclusive lock.
809 vm_object_deallocate(vm_object_t object)
817 count = object->ref_count;
821 * If decrementing the count enters into special handling
822 * territory (0, 1, or 2) we have to do it the hard way.
823 * Fortunate though, objects with only a few refs like this
824 * are not likely to be heavily contended anyway.
827 vm_object_hold(object);
828 vm_object_deallocate_locked(object);
829 vm_object_drop(object);
834 * Try to decrement ref_count without acquiring a hold on
835 * the object. This is particularly important for the exec*()
836 * and exit*() code paths because the program binary may
837 * have a great deal of sharing and an exclusive lock will
838 * crowbar performance in those circumstances.
840 if (object->type == OBJT_VNODE) {
841 vp = (struct vnode *)object->handle;
842 if (atomic_cmpset_int(&object->ref_count,
849 if (atomic_cmpset_int(&object->ref_count,
860 vm_object_deallocate_locked(vm_object_t object)
862 struct vm_object_dealloc_list *dlist = NULL;
863 struct vm_object_dealloc_list *dtmp;
868 * We may chain deallocate object, but additional objects may
869 * collect on the dlist which also have to be deallocated. We
870 * must avoid a recursion, vm_object chains can get deep.
873 while (object != NULL) {
874 ASSERT_LWKT_TOKEN_HELD_EXCL(&object->token);
877 * Don't rip a ref_count out from under an object undergoing
878 * collapse, it will confuse the collapse code.
880 vm_object_chain_wait(object);
882 if (object->type == OBJT_VNODE) {
883 vm_object_vndeallocate(object);
887 if (object->ref_count == 0) {
888 panic("vm_object_deallocate: object deallocated "
889 "too many times: %d", object->type);
891 if (object->ref_count > 2) {
892 atomic_add_int(&object->ref_count, -1);
897 * Here on ref_count of one or two, which are special cases for
900 * Nominal ref_count > 1 case if the second ref is not from
903 * (ONEMAPPING only applies to DEFAULT AND SWAP objects)
905 if (object->ref_count == 2 && object->shadow_count == 0) {
906 if (object->type == OBJT_DEFAULT ||
907 object->type == OBJT_SWAP) {
908 vm_object_set_flag(object, OBJ_ONEMAPPING);
910 atomic_add_int(&object->ref_count, -1);
915 * If the second ref is from a shadow we chain along it
916 * upwards if object's handle is exhausted.
918 * We have to decrement object->ref_count before potentially
919 * collapsing the first shadow object or the collapse code
920 * will not be able to handle the degenerate case to remove
921 * object. However, if we do it too early the object can
922 * get ripped out from under us.
924 if (object->ref_count == 2 && object->shadow_count == 1 &&
925 object->handle == NULL && (object->type == OBJT_DEFAULT ||
926 object->type == OBJT_SWAP)) {
927 temp = LIST_FIRST(&object->shadow_head);
928 KKASSERT(temp != NULL);
929 vm_object_hold(temp);
932 * Wait for any paging to complete so the collapse
933 * doesn't (or isn't likely to) qcollapse. pip
934 * waiting must occur before we acquire the
938 temp->paging_in_progress ||
939 object->paging_in_progress
941 vm_object_pip_wait(temp, "objde1");
942 vm_object_pip_wait(object, "objde2");
946 * If the parent is locked we have to give up, as
947 * otherwise we would be acquiring locks in the
948 * wrong order and potentially deadlock.
950 if (temp->chainlk & (CHAINLK_EXCL | CHAINLK_MASK)) {
951 vm_object_drop(temp);
954 vm_object_chain_acquire(temp, 0);
957 * Recheck/retry after the hold and the paging
958 * wait, both of which can block us.
960 if (object->ref_count != 2 ||
961 object->shadow_count != 1 ||
963 LIST_FIRST(&object->shadow_head) != temp ||
964 (object->type != OBJT_DEFAULT &&
965 object->type != OBJT_SWAP)) {
966 vm_object_chain_release(temp);
967 vm_object_drop(temp);
972 * We can safely drop object's ref_count now.
974 KKASSERT(object->ref_count == 2);
975 atomic_add_int(&object->ref_count, -1);
978 * If our single parent is not collapseable just
979 * decrement ref_count (2->1) and stop.
981 if (temp->handle || (temp->type != OBJT_DEFAULT &&
982 temp->type != OBJT_SWAP)) {
983 vm_object_chain_release(temp);
984 vm_object_drop(temp);
989 * At this point we have already dropped object's
990 * ref_count so it is possible for a race to
991 * deallocate obj out from under us. Any collapse
992 * will re-check the situation. We must not block
993 * until we are able to collapse.
995 * Bump temp's ref_count to avoid an unwanted
996 * degenerate recursion (can't call
997 * vm_object_reference_locked() because it asserts
998 * that CHAINLOCK is not set).
1000 atomic_add_int(&temp->ref_count, 1);
1001 KKASSERT(temp->ref_count > 1);
1004 * Collapse temp, then deallocate the extra ref
1007 vm_object_collapse(temp, &dlist);
1008 vm_object_chain_release(temp);
1010 vm_object_lock_swap();
1011 vm_object_drop(object);
1019 * Drop the ref and handle termination on the 1->0 transition.
1020 * We may have blocked above so we have to recheck.
1023 KKASSERT(object->ref_count != 0);
1024 if (object->ref_count >= 2) {
1025 atomic_add_int(&object->ref_count, -1);
1028 KKASSERT(object->ref_count == 1);
1031 * 1->0 transition. Chain through the backing_object.
1032 * Maintain the ref until we've located the backing object,
1035 while ((temp = object->backing_object) != NULL) {
1036 vm_object_hold(temp);
1037 if (temp == object->backing_object)
1039 vm_object_drop(temp);
1043 * 1->0 transition verified, retry if ref_count is no longer
1044 * 1. Otherwise disconnect the backing_object (temp) and
1047 if (object->ref_count != 1) {
1048 vm_object_drop(temp);
1053 * It shouldn't be possible for the object to be chain locked
1054 * if we're removing the last ref on it.
1056 KKASSERT((object->chainlk & (CHAINLK_EXCL|CHAINLK_MASK)) == 0);
1059 if (object->flags & OBJ_ONSHADOW) {
1060 LIST_REMOVE(object, shadow_list);
1061 temp->shadow_count--;
1063 vm_object_clear_flag(object, OBJ_ONSHADOW);
1065 object->backing_object = NULL;
1068 atomic_add_int(&object->ref_count, -1);
1069 if ((object->flags & OBJ_DEAD) == 0)
1070 vm_object_terminate(object);
1071 if (must_drop && temp)
1072 vm_object_lock_swap();
1074 vm_object_drop(object);
1078 if (must_drop && object)
1079 vm_object_drop(object);
1082 * Additional tail recursion on dlist. Avoid a recursion. Objects
1083 * on the dlist have a hold count but are not locked.
1085 if ((dtmp = dlist) != NULL) {
1087 object = dtmp->object;
1088 kfree(dtmp, M_TEMP);
1090 vm_object_lock(object); /* already held, add lock */
1091 must_drop = 1; /* and we're responsible for it */
1097 * Destroy the specified object, freeing up related resources.
1099 * The object must have zero references.
1101 * The object must held. The caller is responsible for dropping the object
1102 * after terminate returns. Terminate does NOT drop the object.
1104 static int vm_object_terminate_callback(vm_page_t p, void *data);
1107 vm_object_terminate(vm_object_t object)
1110 * Make sure no one uses us. Once we set OBJ_DEAD we should be
1111 * able to safely block.
1113 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1114 KKASSERT((object->flags & OBJ_DEAD) == 0);
1115 vm_object_set_flag(object, OBJ_DEAD);
1118 * Wait for the pageout daemon to be done with the object
1120 vm_object_pip_wait(object, "objtrm1");
1122 KASSERT(!object->paging_in_progress,
1123 ("vm_object_terminate: pageout in progress"));
1126 * Clean and free the pages, as appropriate. All references to the
1127 * object are gone, so we don't need to lock it.
1129 if (object->type == OBJT_VNODE) {
1133 * Clean pages and flush buffers.
1135 * NOTE! TMPFS buffer flushes do not typically flush the
1136 * actual page to swap as this would be highly
1137 * inefficient, and normal filesystems usually wrap
1138 * page flushes with buffer cache buffers.
1140 * To deal with this we have to call vinvalbuf() both
1141 * before and after the vm_object_page_clean().
1143 vp = (struct vnode *) object->handle;
1144 vinvalbuf(vp, V_SAVE, 0, 0);
1145 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
1146 vinvalbuf(vp, V_SAVE, 0, 0);
1150 * Wait for any I/O to complete, after which there had better not
1151 * be any references left on the object.
1153 vm_object_pip_wait(object, "objtrm2");
1155 if (object->ref_count != 0) {
1156 panic("vm_object_terminate: object with references, "
1157 "ref_count=%d", object->ref_count);
1161 * Cleanup any shared pmaps associated with this object.
1163 pmap_object_free(object);
1166 * Now free any remaining pages. For internal objects, this also
1167 * removes them from paging queues. Don't free wired pages, just
1168 * remove them from the object.
1170 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1171 vm_object_terminate_callback, NULL);
1174 * Let the pager know object is dead.
1176 vm_pager_deallocate(object);
1179 * Wait for the object hold count to hit 1, clean out pages as
1180 * we go. vmobj_token interlocks any race conditions that might
1181 * pick the object up from the vm_object_list after we have cleared
1185 if (RB_ROOT(&object->rb_memq) == NULL)
1187 kprintf("vm_object_terminate: Warning, object %p "
1188 "still has %d pages\n",
1189 object, object->resident_page_count);
1190 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1191 vm_object_terminate_callback, NULL);
1195 * There had better not be any pages left
1197 KKASSERT(object->resident_page_count == 0);
1200 * Remove the object from the global object list.
1202 lwkt_gettoken(&vmobj_token);
1203 TAILQ_REMOVE(&vm_object_list, object, object_list);
1205 lwkt_reltoken(&vmobj_token);
1206 vm_object_dead_wakeup(object);
1208 if (object->ref_count != 0) {
1209 panic("vm_object_terminate2: object with references, "
1210 "ref_count=%d", object->ref_count);
1214 * NOTE: The object hold_count is at least 1, so we cannot zfree()
1215 * the object here. See vm_object_drop().
1220 * The caller must hold the object.
1223 vm_object_terminate_callback(vm_page_t p, void *data __unused)
1228 vm_page_busy_wait(p, TRUE, "vmpgtrm");
1229 if (object != p->object) {
1230 kprintf("vm_object_terminate: Warning: Encountered "
1231 "busied page %p on queue %d\n", p, p->queue);
1233 } else if (p->wire_count == 0) {
1235 * NOTE: p->dirty and PG_NEED_COMMIT are ignored.
1238 mycpu->gd_cnt.v_pfree++;
1240 if (p->queue != PQ_NONE)
1241 kprintf("vm_object_terminate: Warning: Encountered "
1242 "wired page %p on queue %d\n", p, p->queue);
1251 * The object is dead but still has an object<->pager association. Sleep
1252 * and return. The caller typically retests the association in a loop.
1254 * The caller must hold the object.
1257 vm_object_dead_sleep(vm_object_t object, const char *wmesg)
1259 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1260 if (object->handle) {
1261 vm_object_set_flag(object, OBJ_DEADWNT);
1262 tsleep(object, 0, wmesg, 0);
1263 /* object may be invalid after this point */
1268 * Wakeup anyone waiting for the object<->pager disassociation on
1271 * The caller must hold the object.
1274 vm_object_dead_wakeup(vm_object_t object)
1276 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1277 if (object->flags & OBJ_DEADWNT) {
1278 vm_object_clear_flag(object, OBJ_DEADWNT);
1284 * Clean all dirty pages in the specified range of object. Leaves page
1285 * on whatever queue it is currently on. If NOSYNC is set then do not
1286 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
1287 * leaving the object dirty.
1289 * When stuffing pages asynchronously, allow clustering. XXX we need a
1290 * synchronous clustering mode implementation.
1292 * Odd semantics: if start == end, we clean everything.
1294 * The object must be locked? XXX
1296 static int vm_object_page_clean_pass1(struct vm_page *p, void *data);
1297 static int vm_object_page_clean_pass2(struct vm_page *p, void *data);
1300 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1303 struct rb_vm_page_scan_info info;
1309 vm_object_hold(object);
1310 if (object->type != OBJT_VNODE ||
1311 (object->flags & OBJ_MIGHTBEDIRTY) == 0) {
1312 vm_object_drop(object);
1316 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ?
1317 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1318 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
1320 vp = object->handle;
1323 * Interlock other major object operations. This allows us to
1324 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY.
1326 vm_object_set_flag(object, OBJ_CLEANING);
1329 * Handle 'entire object' case
1331 info.start_pindex = start;
1333 info.end_pindex = object->size - 1;
1335 info.end_pindex = end - 1;
1337 wholescan = (start == 0 && info.end_pindex == object->size - 1);
1339 info.pagerflags = pagerflags;
1340 info.object = object;
1343 * If cleaning the entire object do a pass to mark the pages read-only.
1344 * If everything worked out ok, clear OBJ_WRITEABLE and
1349 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1350 vm_object_page_clean_pass1, &info);
1351 if (info.error == 0) {
1352 vm_object_clear_flag(object,
1353 OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1354 if (object->type == OBJT_VNODE &&
1355 (vp = (struct vnode *)object->handle) != NULL) {
1357 (vp->v_mount->mnt_kern_flag & MNTK_THR_SYNC)) {
1360 vclrflags(vp, VOBJDIRTY);
1367 * Do a pass to clean all the dirty pages we find.
1371 generation = object->generation;
1372 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1373 vm_object_page_clean_pass2, &info);
1374 } while (info.error || generation != object->generation);
1376 vm_object_clear_flag(object, OBJ_CLEANING);
1377 vm_object_drop(object);
1381 * The caller must hold the object.
1385 vm_object_page_clean_pass1(struct vm_page *p, void *data)
1387 struct rb_vm_page_scan_info *info = data;
1389 vm_page_flag_set(p, PG_CLEANCHK);
1390 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1392 } else if (vm_page_busy_try(p, FALSE) == 0) {
1393 vm_page_protect(p, VM_PROT_READ); /* must not block */
1403 * The caller must hold the object
1407 vm_object_page_clean_pass2(struct vm_page *p, void *data)
1409 struct rb_vm_page_scan_info *info = data;
1413 * Do not mess with pages that were inserted after we started
1414 * the cleaning pass.
1416 if ((p->flags & PG_CLEANCHK) == 0)
1419 generation = info->object->generation;
1420 vm_page_busy_wait(p, TRUE, "vpcwai");
1421 if (p->object != info->object ||
1422 info->object->generation != generation) {
1429 * Before wasting time traversing the pmaps, check for trivial
1430 * cases where the page cannot be dirty.
1432 if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) {
1433 KKASSERT((p->dirty & p->valid) == 0 &&
1434 (p->flags & PG_NEED_COMMIT) == 0);
1440 * Check whether the page is dirty or not. The page has been set
1441 * to be read-only so the check will not race a user dirtying the
1444 vm_page_test_dirty(p);
1445 if ((p->dirty & p->valid) == 0 && (p->flags & PG_NEED_COMMIT) == 0) {
1446 vm_page_flag_clear(p, PG_CLEANCHK);
1452 * If we have been asked to skip nosync pages and this is a
1453 * nosync page, skip it. Note that the object flags were
1454 * not cleared in this case (because pass1 will have returned an
1455 * error), so we do not have to set them.
1457 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1458 vm_page_flag_clear(p, PG_CLEANCHK);
1464 * Flush as many pages as we can. PG_CLEANCHK will be cleared on
1465 * the pages that get successfully flushed. Set info->error if
1466 * we raced an object modification.
1468 vm_object_page_collect_flush(info->object, p, info->pagerflags);
1476 * Collect the specified page and nearby pages and flush them out.
1477 * The number of pages flushed is returned. The passed page is busied
1478 * by the caller and we are responsible for its disposition.
1480 * The caller must hold the object.
1483 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags)
1491 vm_page_t ma[BLIST_MAX_ALLOC];
1493 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1496 page_base = pi % BLIST_MAX_ALLOC;
1504 tp = vm_page_lookup_busy_try(object, pi - page_base + ib,
1510 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1511 (tp->flags & PG_CLEANCHK) == 0) {
1515 if ((tp->queue - tp->pc) == PQ_CACHE) {
1516 vm_page_flag_clear(tp, PG_CLEANCHK);
1520 vm_page_test_dirty(tp);
1521 if ((tp->dirty & tp->valid) == 0 &&
1522 (tp->flags & PG_NEED_COMMIT) == 0) {
1523 vm_page_flag_clear(tp, PG_CLEANCHK);
1532 while (is < BLIST_MAX_ALLOC &&
1533 pi - page_base + is < object->size) {
1536 tp = vm_page_lookup_busy_try(object, pi - page_base + is,
1542 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1543 (tp->flags & PG_CLEANCHK) == 0) {
1547 if ((tp->queue - tp->pc) == PQ_CACHE) {
1548 vm_page_flag_clear(tp, PG_CLEANCHK);
1552 vm_page_test_dirty(tp);
1553 if ((tp->dirty & tp->valid) == 0 &&
1554 (tp->flags & PG_NEED_COMMIT) == 0) {
1555 vm_page_flag_clear(tp, PG_CLEANCHK);
1564 * All pages in the ma[] array are busied now
1566 for (i = ib; i < is; ++i) {
1567 vm_page_flag_clear(ma[i], PG_CLEANCHK);
1568 vm_page_hold(ma[i]); /* XXX need this any more? */
1570 vm_pageout_flush(&ma[ib], is - ib, pagerflags);
1571 for (i = ib; i < is; ++i) /* XXX need this any more? */
1572 vm_page_unhold(ma[i]);
1576 * Same as vm_object_pmap_copy, except range checking really
1577 * works, and is meant for small sections of an object.
1579 * This code protects resident pages by making them read-only
1580 * and is typically called on a fork or split when a page
1581 * is converted to copy-on-write.
1583 * NOTE: If the page is already at VM_PROT_NONE, calling
1584 * vm_page_protect will have no effect.
1587 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1592 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
1595 vm_object_hold(object);
1596 for (idx = start; idx < end; idx++) {
1597 p = vm_page_lookup(object, idx);
1600 vm_page_protect(p, VM_PROT_READ);
1602 vm_object_drop(object);
1606 * Removes all physical pages in the specified object range from all
1609 * The object must *not* be locked.
1612 static int vm_object_pmap_remove_callback(vm_page_t p, void *data);
1615 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1617 struct rb_vm_page_scan_info info;
1621 info.start_pindex = start;
1622 info.end_pindex = end - 1;
1624 vm_object_hold(object);
1625 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1626 vm_object_pmap_remove_callback, &info);
1627 if (start == 0 && end == object->size)
1628 vm_object_clear_flag(object, OBJ_WRITEABLE);
1629 vm_object_drop(object);
1633 * The caller must hold the object
1636 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused)
1638 vm_page_protect(p, VM_PROT_NONE);
1643 * Implements the madvise function at the object/page level.
1645 * MADV_WILLNEED (any object)
1647 * Activate the specified pages if they are resident.
1649 * MADV_DONTNEED (any object)
1651 * Deactivate the specified pages if they are resident.
1653 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only)
1655 * Deactivate and clean the specified pages if they are
1656 * resident. This permits the process to reuse the pages
1657 * without faulting or the kernel to reclaim the pages
1663 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1665 vm_pindex_t end, tpindex;
1666 vm_object_t tobject;
1674 end = pindex + count;
1676 vm_object_hold(object);
1680 * Locate and adjust resident pages
1682 for (; pindex < end; pindex += 1) {
1684 if (tobject != object)
1685 vm_object_drop(tobject);
1690 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1691 * and those pages must be OBJ_ONEMAPPING.
1693 if (advise == MADV_FREE) {
1694 if ((tobject->type != OBJT_DEFAULT &&
1695 tobject->type != OBJT_SWAP) ||
1696 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1701 m = vm_page_lookup_busy_try(tobject, tpindex, TRUE, &error);
1704 vm_page_sleep_busy(m, TRUE, "madvpo");
1709 * There may be swap even if there is no backing page
1711 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1712 swap_pager_freespace(tobject, tpindex, 1);
1717 while ((xobj = tobject->backing_object) != NULL) {
1718 KKASSERT(xobj != object);
1719 vm_object_hold(xobj);
1720 if (xobj == tobject->backing_object)
1722 vm_object_drop(xobj);
1726 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1727 if (tobject != object) {
1728 vm_object_lock_swap();
1729 vm_object_drop(tobject);
1736 * If the page is not in a normal active state, we skip it.
1737 * If the page is not managed there are no page queues to
1738 * mess with. Things can break if we mess with pages in
1739 * any of the below states.
1741 if (m->wire_count ||
1742 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
1743 m->valid != VM_PAGE_BITS_ALL
1750 * Theoretically once a page is known not to be busy, an
1751 * interrupt cannot come along and rip it out from under us.
1754 if (advise == MADV_WILLNEED) {
1755 vm_page_activate(m);
1756 } else if (advise == MADV_DONTNEED) {
1757 vm_page_dontneed(m);
1758 } else if (advise == MADV_FREE) {
1760 * Mark the page clean. This will allow the page
1761 * to be freed up by the system. However, such pages
1762 * are often reused quickly by malloc()/free()
1763 * so we do not do anything that would cause
1764 * a page fault if we can help it.
1766 * Specifically, we do not try to actually free
1767 * the page now nor do we try to put it in the
1768 * cache (which would cause a page fault on reuse).
1770 * But we do make the page is freeable as we
1771 * can without actually taking the step of unmapping
1774 pmap_clear_modify(m);
1777 vm_page_dontneed(m);
1778 if (tobject->type == OBJT_SWAP)
1779 swap_pager_freespace(tobject, tpindex, 1);
1783 if (tobject != object)
1784 vm_object_drop(tobject);
1785 vm_object_drop(object);
1789 * Create a new object which is backed by the specified existing object
1790 * range. Replace the pointer and offset that was pointing at the existing
1791 * object with the pointer/offset for the new object.
1793 * No other requirements.
1796 vm_object_shadow(vm_object_t *objectp, vm_ooffset_t *offset, vm_size_t length,
1806 * Don't create the new object if the old object isn't shared.
1807 * We have to chain wait before adding the reference to avoid
1808 * racing a collapse or deallocation.
1810 * Add the additional ref to source here to avoid racing a later
1811 * collapse or deallocation. Clear the ONEMAPPING flag whether
1812 * addref is TRUE or not in this case because the original object
1817 if (source->type != OBJT_VNODE) {
1819 vm_object_hold(source);
1820 vm_object_chain_wait(source, 0);
1821 if (source->ref_count == 1 &&
1822 source->handle == NULL &&
1823 (source->type == OBJT_DEFAULT ||
1824 source->type == OBJT_SWAP)) {
1826 vm_object_reference_locked(source);
1827 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1829 vm_object_drop(source);
1832 vm_object_reference_locked(source);
1833 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1835 vm_object_reference_quick(source);
1836 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1841 * Allocate a new object with the given length. The new object
1842 * is returned referenced but we may have to add another one.
1843 * If we are adding a second reference we must clear OBJ_ONEMAPPING.
1844 * (typically because the caller is about to clone a vm_map_entry).
1846 * The source object currently has an extra reference to prevent
1847 * collapses into it while we mess with its shadow list, which
1848 * we will remove later in this routine.
1850 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1851 panic("vm_object_shadow: no object for shadowing");
1852 vm_object_hold(result);
1854 vm_object_reference_locked(result);
1855 vm_object_clear_flag(result, OBJ_ONEMAPPING);
1859 * The new object shadows the source object. Chain wait before
1860 * adjusting shadow_count or the shadow list to avoid races.
1862 * Try to optimize the result object's page color when shadowing
1863 * in order to maintain page coloring consistency in the combined
1866 * SHADOWING IS NOT APPLICABLE TO OBJT_VNODE OBJECTS
1868 KKASSERT(result->backing_object == NULL);
1869 result->backing_object = source;
1871 if (useshadowlist) {
1872 vm_object_chain_wait(source, 0);
1873 LIST_INSERT_HEAD(&source->shadow_head,
1874 result, shadow_list);
1875 source->shadow_count++;
1876 source->generation++;
1877 vm_object_set_flag(result, OBJ_ONSHADOW);
1879 /* cpu localization twist */
1880 result->pg_color = (int)(intptr_t)curthread;
1884 * Adjust the return storage. Drop the ref on source before
1887 result->backing_object_offset = *offset;
1888 vm_object_drop(result);
1891 if (useshadowlist) {
1892 vm_object_deallocate_locked(source);
1893 vm_object_drop(source);
1895 vm_object_deallocate(source);
1900 * Return the new things
1905 #define OBSC_TEST_ALL_SHADOWED 0x0001
1906 #define OBSC_COLLAPSE_NOWAIT 0x0002
1907 #define OBSC_COLLAPSE_WAIT 0x0004
1909 static int vm_object_backing_scan_callback(vm_page_t p, void *data);
1912 * The caller must hold the object.
1915 vm_object_backing_scan(vm_object_t object, vm_object_t backing_object, int op)
1917 struct rb_vm_page_scan_info info;
1919 vm_object_assert_held(object);
1920 vm_object_assert_held(backing_object);
1922 KKASSERT(backing_object == object->backing_object);
1923 info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1926 * Initial conditions
1928 if (op & OBSC_TEST_ALL_SHADOWED) {
1930 * We do not want to have to test for the existence of
1931 * swap pages in the backing object. XXX but with the
1932 * new swapper this would be pretty easy to do.
1934 * XXX what about anonymous MAP_SHARED memory that hasn't
1935 * been ZFOD faulted yet? If we do not test for this, the
1936 * shadow test may succeed! XXX
1938 if (backing_object->type != OBJT_DEFAULT)
1941 if (op & OBSC_COLLAPSE_WAIT) {
1942 KKASSERT((backing_object->flags & OBJ_DEAD) == 0);
1943 vm_object_set_flag(backing_object, OBJ_DEAD);
1944 lwkt_gettoken(&vmobj_token);
1945 TAILQ_REMOVE(&vm_object_list, backing_object, object_list);
1947 lwkt_reltoken(&vmobj_token);
1948 vm_object_dead_wakeup(backing_object);
1952 * Our scan. We have to retry if a negative error code is returned,
1953 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that
1954 * the scan had to be stopped because the parent does not completely
1957 info.object = object;
1958 info.backing_object = backing_object;
1962 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL,
1963 vm_object_backing_scan_callback,
1965 } while (info.error < 0);
1971 * The caller must hold the object.
1974 vm_object_backing_scan_callback(vm_page_t p, void *data)
1976 struct rb_vm_page_scan_info *info = data;
1977 vm_object_t backing_object;
1980 vm_pindex_t new_pindex;
1981 vm_pindex_t backing_offset_index;
1985 new_pindex = pindex - info->backing_offset_index;
1987 object = info->object;
1988 backing_object = info->backing_object;
1989 backing_offset_index = info->backing_offset_index;
1991 if (op & OBSC_TEST_ALL_SHADOWED) {
1995 * Ignore pages outside the parent object's range
1996 * and outside the parent object's mapping of the
1999 * note that we do not busy the backing object's
2002 if (pindex < backing_offset_index ||
2003 new_pindex >= object->size
2009 * See if the parent has the page or if the parent's
2010 * object pager has the page. If the parent has the
2011 * page but the page is not valid, the parent's
2012 * object pager must have the page.
2014 * If this fails, the parent does not completely shadow
2015 * the object and we might as well give up now.
2017 pp = vm_page_lookup(object, new_pindex);
2018 if ((pp == NULL || pp->valid == 0) &&
2019 !vm_pager_has_page(object, new_pindex)
2021 info->error = 0; /* problemo */
2022 return(-1); /* stop the scan */
2027 * Check for busy page. Note that we may have lost (p) when we
2028 * possibly blocked above.
2030 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
2033 if (vm_page_busy_try(p, TRUE)) {
2034 if (op & OBSC_COLLAPSE_NOWAIT) {
2038 * If we slept, anything could have
2039 * happened. Ask that the scan be restarted.
2041 * Since the object is marked dead, the
2042 * backing offset should not have changed.
2044 vm_page_sleep_busy(p, TRUE, "vmocol");
2051 * If (p) is no longer valid restart the scan.
2053 if (p->object != backing_object || p->pindex != pindex) {
2054 kprintf("vm_object_backing_scan: Warning: page "
2055 "%p ripped out from under us\n", p);
2061 if (op & OBSC_COLLAPSE_NOWAIT) {
2062 if (p->valid == 0 ||
2064 (p->flags & PG_NEED_COMMIT)) {
2069 /* XXX what if p->valid == 0 , hold_count, etc? */
2073 p->object == backing_object,
2074 ("vm_object_qcollapse(): object mismatch")
2078 * Destroy any associated swap
2080 if (backing_object->type == OBJT_SWAP)
2081 swap_pager_freespace(backing_object, p->pindex, 1);
2084 p->pindex < backing_offset_index ||
2085 new_pindex >= object->size
2088 * Page is out of the parent object's range, we
2089 * can simply destroy it.
2091 vm_page_protect(p, VM_PROT_NONE);
2096 pp = vm_page_lookup(object, new_pindex);
2097 if (pp != NULL || vm_pager_has_page(object, new_pindex)) {
2099 * page already exists in parent OR swap exists
2100 * for this location in the parent. Destroy
2101 * the original page from the backing object.
2103 * Leave the parent's page alone
2105 vm_page_protect(p, VM_PROT_NONE);
2111 * Page does not exist in parent, rename the
2112 * page from the backing object to the main object.
2114 * If the page was mapped to a process, it can remain
2115 * mapped through the rename.
2117 if ((p->queue - p->pc) == PQ_CACHE)
2118 vm_page_deactivate(p);
2120 vm_page_rename(p, object, new_pindex);
2122 /* page automatically made dirty by rename */
2128 * This version of collapse allows the operation to occur earlier and
2129 * when paging_in_progress is true for an object... This is not a complete
2130 * operation, but should plug 99.9% of the rest of the leaks.
2132 * The caller must hold the object and backing_object and both must be
2135 * (only called from vm_object_collapse)
2138 vm_object_qcollapse(vm_object_t object, vm_object_t backing_object)
2140 if (backing_object->ref_count == 1) {
2141 atomic_add_int(&backing_object->ref_count, 2);
2142 vm_object_backing_scan(object, backing_object,
2143 OBSC_COLLAPSE_NOWAIT);
2144 atomic_add_int(&backing_object->ref_count, -2);
2149 * Collapse an object with the object backing it. Pages in the backing
2150 * object are moved into the parent, and the backing object is deallocated.
2151 * Any conflict is resolved in favor of the parent's existing pages.
2153 * object must be held and chain-locked on call.
2155 * The caller must have an extra ref on object to prevent a race from
2156 * destroying it during the collapse.
2159 vm_object_collapse(vm_object_t object, struct vm_object_dealloc_list **dlistp)
2161 struct vm_object_dealloc_list *dlist = NULL;
2162 vm_object_t backing_object;
2165 * Only one thread is attempting a collapse at any given moment.
2166 * There are few restrictions for (object) that callers of this
2167 * function check so reentrancy is likely.
2169 KKASSERT(object != NULL);
2170 vm_object_assert_held(object);
2171 KKASSERT(object->chainlk & (CHAINLK_MASK | CHAINLK_EXCL));
2178 * We can only collapse a DEFAULT/SWAP object with a
2179 * DEFAULT/SWAP object.
2181 if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP) {
2182 backing_object = NULL;
2186 backing_object = object->backing_object;
2187 if (backing_object == NULL)
2189 if (backing_object->type != OBJT_DEFAULT &&
2190 backing_object->type != OBJT_SWAP) {
2191 backing_object = NULL;
2196 * Hold the backing_object and check for races
2198 vm_object_hold(backing_object);
2199 if (backing_object != object->backing_object ||
2200 (backing_object->type != OBJT_DEFAULT &&
2201 backing_object->type != OBJT_SWAP)) {
2202 vm_object_drop(backing_object);
2207 * Chain-lock the backing object too because if we
2208 * successfully merge its pages into the top object we
2209 * will collapse backing_object->backing_object as the
2210 * new backing_object. Re-check that it is still our
2213 vm_object_chain_acquire(backing_object, 0);
2214 if (backing_object != object->backing_object) {
2215 vm_object_chain_release(backing_object);
2216 vm_object_drop(backing_object);
2221 * we check the backing object first, because it is most likely
2224 if (backing_object->handle != NULL ||
2225 (backing_object->type != OBJT_DEFAULT &&
2226 backing_object->type != OBJT_SWAP) ||
2227 (backing_object->flags & OBJ_DEAD) ||
2228 object->handle != NULL ||
2229 (object->type != OBJT_DEFAULT &&
2230 object->type != OBJT_SWAP) ||
2231 (object->flags & OBJ_DEAD)) {
2236 * If paging is in progress we can't do a normal collapse.
2239 object->paging_in_progress != 0 ||
2240 backing_object->paging_in_progress != 0
2242 vm_object_qcollapse(object, backing_object);
2247 * We know that we can either collapse the backing object (if
2248 * the parent is the only reference to it) or (perhaps) have
2249 * the parent bypass the object if the parent happens to shadow
2250 * all the resident pages in the entire backing object.
2252 * This is ignoring pager-backed pages such as swap pages.
2253 * vm_object_backing_scan fails the shadowing test in this
2256 if (backing_object->ref_count == 1) {
2258 * If there is exactly one reference to the backing
2259 * object, we can collapse it into the parent.
2261 KKASSERT(object->backing_object == backing_object);
2262 vm_object_backing_scan(object, backing_object,
2263 OBSC_COLLAPSE_WAIT);
2266 * Move the pager from backing_object to object.
2268 if (backing_object->type == OBJT_SWAP) {
2269 vm_object_pip_add(backing_object, 1);
2272 * scrap the paging_offset junk and do a
2273 * discrete copy. This also removes major
2274 * assumptions about how the swap-pager
2275 * works from where it doesn't belong. The
2276 * new swapper is able to optimize the
2277 * destroy-source case.
2279 vm_object_pip_add(object, 1);
2280 swap_pager_copy(backing_object, object,
2281 OFF_TO_IDX(object->backing_object_offset),
2283 vm_object_pip_wakeup(object);
2284 vm_object_pip_wakeup(backing_object);
2288 * Object now shadows whatever backing_object did.
2289 * Remove object from backing_object's shadow_list.
2291 KKASSERT(object->backing_object == backing_object);
2292 if (object->flags & OBJ_ONSHADOW) {
2293 LIST_REMOVE(object, shadow_list);
2294 backing_object->shadow_count--;
2295 backing_object->generation++;
2296 vm_object_clear_flag(object, OBJ_ONSHADOW);
2300 * backing_object->backing_object moves from within
2301 * backing_object to within object.
2303 * OBJT_VNODE bbobj's should have empty shadow lists.
2305 while ((bbobj = backing_object->backing_object) != NULL) {
2306 if (bbobj->type == OBJT_VNODE)
2307 vm_object_hold_shared(bbobj);
2309 vm_object_hold(bbobj);
2310 if (bbobj == backing_object->backing_object)
2312 vm_object_drop(bbobj);
2315 if (backing_object->flags & OBJ_ONSHADOW) {
2316 /* not locked exclusively if vnode */
2317 KKASSERT(bbobj->type != OBJT_VNODE);
2318 LIST_REMOVE(backing_object,
2320 bbobj->shadow_count--;
2321 bbobj->generation++;
2322 vm_object_clear_flag(backing_object,
2325 backing_object->backing_object = NULL;
2327 object->backing_object = bbobj;
2329 if (bbobj->type != OBJT_VNODE) {
2330 LIST_INSERT_HEAD(&bbobj->shadow_head,
2331 object, shadow_list);
2332 bbobj->shadow_count++;
2333 bbobj->generation++;
2334 vm_object_set_flag(object,
2339 object->backing_object_offset +=
2340 backing_object->backing_object_offset;
2342 vm_object_drop(bbobj);
2345 * Discard the old backing_object. Nothing should be
2346 * able to ref it, other than a vm_map_split(),
2347 * and vm_map_split() will stall on our chain lock.
2348 * And we control the parent so it shouldn't be
2349 * possible for it to go away either.
2351 * Since the backing object has no pages, no pager
2352 * left, and no object references within it, all
2353 * that is necessary is to dispose of it.
2355 KASSERT(backing_object->ref_count == 1,
2356 ("backing_object %p was somehow "
2357 "re-referenced during collapse!",
2359 KASSERT(RB_EMPTY(&backing_object->rb_memq),
2360 ("backing_object %p somehow has left "
2361 "over pages during collapse!",
2365 * The object can be destroyed.
2367 * XXX just fall through and dodealloc instead
2368 * of forcing destruction?
2370 atomic_add_int(&backing_object->ref_count, -1);
2371 if ((backing_object->flags & OBJ_DEAD) == 0)
2372 vm_object_terminate(backing_object);
2377 * If we do not entirely shadow the backing object,
2378 * there is nothing we can do so we give up.
2380 if (vm_object_backing_scan(object, backing_object,
2381 OBSC_TEST_ALL_SHADOWED) == 0) {
2386 * bbobj is backing_object->backing_object. Since
2387 * object completely shadows backing_object we can
2388 * bypass it and become backed by bbobj instead.
2390 * The shadow list for vnode backing objects is not
2391 * used and a shared hold is allowed.
2393 while ((bbobj = backing_object->backing_object) != NULL) {
2394 if (bbobj->type == OBJT_VNODE)
2395 vm_object_hold_shared(bbobj);
2397 vm_object_hold(bbobj);
2398 if (bbobj == backing_object->backing_object)
2400 vm_object_drop(bbobj);
2404 * Make object shadow bbobj instead of backing_object.
2405 * Remove object from backing_object's shadow list.
2407 * Deallocating backing_object will not remove
2408 * it, since its reference count is at least 2.
2410 KKASSERT(object->backing_object == backing_object);
2411 if (object->flags & OBJ_ONSHADOW) {
2412 LIST_REMOVE(object, shadow_list);
2413 backing_object->shadow_count--;
2414 backing_object->generation++;
2415 vm_object_clear_flag(object, OBJ_ONSHADOW);
2419 * Add a ref to bbobj, bbobj now shadows object.
2421 * NOTE: backing_object->backing_object still points
2422 * to bbobj. That relationship remains intact
2423 * because backing_object has > 1 ref, so
2424 * someone else is pointing to it (hence why
2425 * we can't collapse it into object and can
2426 * only handle the all-shadowed bypass case).
2429 if (bbobj->type != OBJT_VNODE) {
2430 vm_object_chain_wait(bbobj, 0);
2431 vm_object_reference_locked(bbobj);
2432 LIST_INSERT_HEAD(&bbobj->shadow_head,
2433 object, shadow_list);
2434 bbobj->shadow_count++;
2435 bbobj->generation++;
2436 vm_object_set_flag(object,
2439 vm_object_reference_quick(bbobj);
2441 object->backing_object_offset +=
2442 backing_object->backing_object_offset;
2443 object->backing_object = bbobj;
2444 vm_object_drop(bbobj);
2446 object->backing_object = NULL;
2450 * Drop the reference count on backing_object. To
2451 * handle ref_count races properly we can't assume
2452 * that the ref_count is still at least 2 so we
2453 * have to actually call vm_object_deallocate()
2454 * (after clearing the chainlock).
2461 * Ok, we want to loop on the new object->bbobj association,
2462 * possibly collapsing it further. However if dodealloc is
2463 * non-zero we have to deallocate the backing_object which
2464 * itself can potentially undergo a collapse, creating a
2465 * recursion depth issue with the LWKT token subsystem.
2467 * In the case where we must deallocate the backing_object
2468 * it is possible now that the backing_object has a single
2469 * shadow count on some other object (not represented here
2470 * as yet), since it no longer shadows us. Thus when we
2471 * call vm_object_deallocate() it may attempt to collapse
2472 * itself into its remaining parent.
2475 struct vm_object_dealloc_list *dtmp;
2477 vm_object_chain_release(backing_object);
2478 vm_object_unlock(backing_object);
2479 /* backing_object remains held */
2482 * Auto-deallocation list for caller convenience.
2487 dtmp = kmalloc(sizeof(*dtmp), M_TEMP, M_WAITOK);
2488 dtmp->object = backing_object;
2489 dtmp->next = *dlistp;
2492 vm_object_chain_release(backing_object);
2493 vm_object_drop(backing_object);
2495 /* backing_object = NULL; not needed */
2500 * Clean up any left over backing_object
2502 if (backing_object) {
2503 vm_object_chain_release(backing_object);
2504 vm_object_drop(backing_object);
2508 * Clean up any auto-deallocation list. This is a convenience
2509 * for top-level callers so they don't have to pass &dlist.
2510 * Do not clean up any caller-passed dlistp, the caller will
2514 vm_object_deallocate_list(&dlist);
2519 * vm_object_collapse() may collect additional objects in need of
2520 * deallocation. This routine deallocates these objects. The
2521 * deallocation itself can trigger additional collapses (which the
2522 * deallocate function takes care of). This procedure is used to
2523 * reduce procedural recursion since these vm_object shadow chains
2524 * can become quite long.
2527 vm_object_deallocate_list(struct vm_object_dealloc_list **dlistp)
2529 struct vm_object_dealloc_list *dlist;
2531 while ((dlist = *dlistp) != NULL) {
2532 *dlistp = dlist->next;
2533 vm_object_lock(dlist->object);
2534 vm_object_deallocate_locked(dlist->object);
2535 vm_object_drop(dlist->object);
2536 kfree(dlist, M_TEMP);
2541 * Removes all physical pages in the specified object range from the
2542 * object's list of pages.
2546 static int vm_object_page_remove_callback(vm_page_t p, void *data);
2549 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2550 boolean_t clean_only)
2552 struct rb_vm_page_scan_info info;
2556 * Degenerate cases and assertions
2558 vm_object_hold(object);
2559 if (object == NULL ||
2560 (object->resident_page_count == 0 && object->swblock_count == 0)) {
2561 vm_object_drop(object);
2564 KASSERT(object->type != OBJT_PHYS,
2565 ("attempt to remove pages from a physical object"));
2568 * Indicate that paging is occuring on the object
2570 vm_object_pip_add(object, 1);
2573 * Figure out the actual removal range and whether we are removing
2574 * the entire contents of the object or not. If removing the entire
2575 * contents, be sure to get all pages, even those that might be
2576 * beyond the end of the object.
2578 info.start_pindex = start;
2580 info.end_pindex = (vm_pindex_t)-1;
2582 info.end_pindex = end - 1;
2583 info.limit = clean_only;
2584 all = (start == 0 && info.end_pindex >= object->size - 1);
2587 * Loop until we are sure we have gotten them all.
2591 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2592 vm_object_page_remove_callback, &info);
2593 } while (info.error);
2596 * Remove any related swap if throwing away pages, or for
2597 * non-swap objects (the swap is a clean copy in that case).
2599 if (object->type != OBJT_SWAP || clean_only == FALSE) {
2601 swap_pager_freespace_all(object);
2603 swap_pager_freespace(object, info.start_pindex,
2604 info.end_pindex - info.start_pindex + 1);
2610 vm_object_pip_wakeup(object);
2611 vm_object_drop(object);
2615 * The caller must hold the object
2618 vm_object_page_remove_callback(vm_page_t p, void *data)
2620 struct rb_vm_page_scan_info *info = data;
2622 if (vm_page_busy_try(p, TRUE)) {
2623 vm_page_sleep_busy(p, TRUE, "vmopar");
2629 * Wired pages cannot be destroyed, but they can be invalidated
2630 * and we do so if clean_only (limit) is not set.
2632 * WARNING! The page may be wired due to being part of a buffer
2633 * cache buffer, and the buffer might be marked B_CACHE.
2634 * This is fine as part of a truncation but VFSs must be
2635 * sure to fix the buffer up when re-extending the file.
2637 * NOTE! PG_NEED_COMMIT is ignored.
2639 if (p->wire_count != 0) {
2640 vm_page_protect(p, VM_PROT_NONE);
2641 if (info->limit == 0)
2648 * limit is our clean_only flag. If set and the page is dirty or
2649 * requires a commit, do not free it. If set and the page is being
2650 * held by someone, do not free it.
2652 if (info->limit && p->valid) {
2653 vm_page_test_dirty(p);
2654 if ((p->valid & p->dirty) || (p->flags & PG_NEED_COMMIT)) {
2659 if (p->hold_count) {
2669 vm_page_protect(p, VM_PROT_NONE);
2675 * Coalesces two objects backing up adjoining regions of memory into a
2678 * returns TRUE if objects were combined.
2680 * NOTE: Only works at the moment if the second object is NULL -
2681 * if it's not, which object do we lock first?
2684 * prev_object First object to coalesce
2685 * prev_offset Offset into prev_object
2686 * next_object Second object into coalesce
2687 * next_offset Offset into next_object
2689 * prev_size Size of reference to prev_object
2690 * next_size Size of reference to next_object
2692 * The caller does not need to hold (prev_object) but must have a stable
2693 * pointer to it (typically by holding the vm_map locked).
2696 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
2697 vm_size_t prev_size, vm_size_t next_size)
2699 vm_pindex_t next_pindex;
2701 if (prev_object == NULL)
2704 vm_object_hold(prev_object);
2706 if (prev_object->type != OBJT_DEFAULT &&
2707 prev_object->type != OBJT_SWAP) {
2708 vm_object_drop(prev_object);
2713 * Try to collapse the object first
2715 vm_object_chain_acquire(prev_object, 0);
2716 vm_object_collapse(prev_object, NULL);
2719 * Can't coalesce if: . more than one reference . paged out . shadows
2720 * another object . has a copy elsewhere (any of which mean that the
2721 * pages not mapped to prev_entry may be in use anyway)
2724 if (prev_object->backing_object != NULL) {
2725 vm_object_chain_release(prev_object);
2726 vm_object_drop(prev_object);
2730 prev_size >>= PAGE_SHIFT;
2731 next_size >>= PAGE_SHIFT;
2732 next_pindex = prev_pindex + prev_size;
2734 if ((prev_object->ref_count > 1) &&
2735 (prev_object->size != next_pindex)) {
2736 vm_object_chain_release(prev_object);
2737 vm_object_drop(prev_object);
2742 * Remove any pages that may still be in the object from a previous
2745 if (next_pindex < prev_object->size) {
2746 vm_object_page_remove(prev_object,
2748 next_pindex + next_size, FALSE);
2749 if (prev_object->type == OBJT_SWAP)
2750 swap_pager_freespace(prev_object,
2751 next_pindex, next_size);
2755 * Extend the object if necessary.
2757 if (next_pindex + next_size > prev_object->size)
2758 prev_object->size = next_pindex + next_size;
2760 vm_object_chain_release(prev_object);
2761 vm_object_drop(prev_object);
2766 * Make the object writable and flag is being possibly dirty.
2768 * The object might not be held (or might be held but held shared),
2769 * the related vnode is probably not held either. Object and vnode are
2770 * stable by virtue of the vm_page busied by the caller preventing
2773 * If the related mount is flagged MNTK_THR_SYNC we need to call
2774 * vsetobjdirty(). Filesystems using this option usually shortcut
2775 * synchronization by only scanning the syncer list.
2778 vm_object_set_writeable_dirty(vm_object_t object)
2782 /*vm_object_assert_held(object);*/
2784 * Avoid contention in vm fault path by checking the state before
2785 * issuing an atomic op on it.
2787 if ((object->flags & (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) !=
2788 (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) {
2789 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
2791 if (object->type == OBJT_VNODE &&
2792 (vp = (struct vnode *)object->handle) != NULL) {
2793 if ((vp->v_flag & VOBJDIRTY) == 0) {
2795 (vp->v_mount->mnt_kern_flag & MNTK_THR_SYNC)) {
2798 vsetflags(vp, VOBJDIRTY);
2804 #include "opt_ddb.h"
2806 #include <sys/kernel.h>
2808 #include <sys/cons.h>
2810 #include <ddb/ddb.h>
2812 static int _vm_object_in_map (vm_map_t map, vm_object_t object,
2813 vm_map_entry_t entry);
2814 static int vm_object_in_map (vm_object_t object);
2817 * The caller must hold the object.
2820 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2823 vm_map_entry_t tmpe;
2824 vm_object_t obj, nobj;
2830 tmpe = map->header.next;
2831 entcount = map->nentries;
2832 while (entcount-- && (tmpe != &map->header)) {
2833 if( _vm_object_in_map(map, object, tmpe)) {
2840 switch(entry->maptype) {
2841 case VM_MAPTYPE_SUBMAP:
2842 tmpm = entry->object.sub_map;
2843 tmpe = tmpm->header.next;
2844 entcount = tmpm->nentries;
2845 while (entcount-- && tmpe != &tmpm->header) {
2846 if( _vm_object_in_map(tmpm, object, tmpe)) {
2852 case VM_MAPTYPE_NORMAL:
2853 case VM_MAPTYPE_VPAGETABLE:
2854 obj = entry->object.vm_object;
2856 if (obj == object) {
2857 if (obj != entry->object.vm_object)
2858 vm_object_drop(obj);
2861 while ((nobj = obj->backing_object) != NULL) {
2862 vm_object_hold(nobj);
2863 if (nobj == obj->backing_object)
2865 vm_object_drop(nobj);
2867 if (obj != entry->object.vm_object) {
2869 vm_object_lock_swap();
2870 vm_object_drop(obj);
2881 static int vm_object_in_map_callback(struct proc *p, void *data);
2883 struct vm_object_in_map_info {
2892 vm_object_in_map(vm_object_t object)
2894 struct vm_object_in_map_info info;
2897 info.object = object;
2899 allproc_scan(vm_object_in_map_callback, &info);
2902 if( _vm_object_in_map(&kernel_map, object, 0))
2904 if( _vm_object_in_map(&pager_map, object, 0))
2906 if( _vm_object_in_map(&buffer_map, object, 0))
2915 vm_object_in_map_callback(struct proc *p, void *data)
2917 struct vm_object_in_map_info *info = data;
2920 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) {
2928 DB_SHOW_COMMAND(vmochk, vm_object_check)
2933 * make sure that internal objs are in a map somewhere
2934 * and none have zero ref counts.
2936 for (object = TAILQ_FIRST(&vm_object_list);
2938 object = TAILQ_NEXT(object, object_list)) {
2939 if (object->type == OBJT_MARKER)
2941 if (object->handle == NULL &&
2942 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2943 if (object->ref_count == 0) {
2944 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2945 (long)object->size);
2947 if (!vm_object_in_map(object)) {
2949 "vmochk: internal obj is not in a map: "
2950 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2951 object->ref_count, (u_long)object->size,
2952 (u_long)object->size,
2953 (void *)object->backing_object);
2962 DB_SHOW_COMMAND(object, vm_object_print_static)
2964 /* XXX convert args. */
2965 vm_object_t object = (vm_object_t)addr;
2966 boolean_t full = have_addr;
2970 /* XXX count is an (unused) arg. Avoid shadowing it. */
2971 #define count was_count
2979 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
2980 object, (int)object->type, (u_long)object->size,
2981 object->resident_page_count, object->ref_count, object->flags);
2983 * XXX no %qd in kernel. Truncate object->backing_object_offset.
2985 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
2986 object->shadow_count,
2987 object->backing_object ? object->backing_object->ref_count : 0,
2988 object->backing_object, (long)object->backing_object_offset);
2995 RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) {
2997 db_iprintf("memory:=");
2998 else if (count == 6) {
3006 db_printf("(off=0x%lx,page=0x%lx)",
3007 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
3018 * XXX need this non-static entry for calling from vm_map_print.
3023 vm_object_print(/* db_expr_t */ long addr,
3024 boolean_t have_addr,
3025 /* db_expr_t */ long count,
3028 vm_object_print_static(addr, have_addr, count, modif);
3034 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
3039 for (object = TAILQ_FIRST(&vm_object_list);
3041 object = TAILQ_NEXT(object, object_list)) {
3042 vm_pindex_t idx, fidx;
3044 vm_paddr_t pa = -1, padiff;
3048 if (object->type == OBJT_MARKER)
3050 db_printf("new object: %p\n", (void *)object);
3060 osize = object->size;
3063 for (idx = 0; idx < osize; idx++) {
3064 m = vm_page_lookup(object, idx);
3067 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
3068 (long)fidx, rcount, (long)pa);
3083 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
3088 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
3089 padiff >>= PAGE_SHIFT;
3090 padiff &= PQ_L2_MASK;
3092 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
3096 db_printf(" index(%ld)run(%d)pa(0x%lx)",
3097 (long)fidx, rcount, (long)pa);
3098 db_printf("pd(%ld)\n", (long)padiff);
3108 pa = VM_PAGE_TO_PHYS(m);
3112 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
3113 (long)fidx, rcount, (long)pa);