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 vm_object kernel_object
;
133 static long vm_object_count
;
135 static long object_collapses
;
136 static long object_bypasses
;
137 static int next_index
;
138 static vm_zone_t obj_zone
;
139 static struct vm_zone obj_zone_store
;
140 #define VM_OBJECTS_INIT 256
141 static struct vm_object vm_objects_init
[VM_OBJECTS_INIT
];
143 struct object_q vm_object_lists
[VMOBJ_HSIZE
];
144 struct lwkt_token vmobj_tokens
[VMOBJ_HSIZE
];
146 #if defined(DEBUG_LOCKS)
148 #define vm_object_vndeallocate(obj, vpp) \
149 debugvm_object_vndeallocate(obj, vpp, __FILE__, __LINE__)
152 * Debug helper to track hold/drop/ref/deallocate calls.
155 debugvm_object_add(vm_object_t obj
, char *file
, int line
, int addrem
)
159 i
= atomic_fetchadd_int(&obj
->debug_index
, 1);
160 i
= i
& (VMOBJ_DEBUG_ARRAY_SIZE
- 1);
161 ksnprintf(obj
->debug_hold_thrs
[i
],
162 sizeof(obj
->debug_hold_thrs
[i
]),
164 (addrem
== -1 ? '-' : (addrem
== 1 ? '+' : '=')),
165 (curthread
->td_proc
? curthread
->td_proc
->p_pid
: -1),
168 obj
->debug_hold_file
[i
] = file
;
169 obj
->debug_hold_line
[i
] = line
;
171 /* Uncomment for debugging obj refs/derefs in reproducable cases */
172 if (strcmp(curthread
->td_comm
, "sshd") == 0) {
173 kprintf("%d %p refs=%d ar=%d file: %s/%d\n",
174 (curthread
->td_proc
? curthread
->td_proc
->p_pid
: -1),
175 obj
, obj
->ref_count
, addrem
, file
, line
);
183 * Misc low level routines
186 vm_object_lock_init(vm_object_t obj
)
188 #if defined(DEBUG_LOCKS)
191 obj
->debug_index
= 0;
192 for (i
= 0; i
< VMOBJ_DEBUG_ARRAY_SIZE
; i
++) {
193 obj
->debug_hold_thrs
[i
][0] = 0;
194 obj
->debug_hold_file
[i
] = NULL
;
195 obj
->debug_hold_line
[i
] = 0;
201 vm_object_lock_swap(void)
207 vm_object_lock(vm_object_t obj
)
209 lwkt_gettoken(&obj
->token
);
213 * Returns TRUE on sucesss
216 vm_object_lock_try(vm_object_t obj
)
218 return(lwkt_trytoken(&obj
->token
));
222 vm_object_lock_shared(vm_object_t obj
)
224 lwkt_gettoken_shared(&obj
->token
);
228 vm_object_unlock(vm_object_t obj
)
230 lwkt_reltoken(&obj
->token
);
234 vm_object_upgrade(vm_object_t obj
)
236 lwkt_reltoken(&obj
->token
);
237 lwkt_gettoken(&obj
->token
);
241 vm_object_downgrade(vm_object_t obj
)
243 lwkt_reltoken(&obj
->token
);
244 lwkt_gettoken_shared(&obj
->token
);
248 vm_object_assert_held(vm_object_t obj
)
250 ASSERT_LWKT_TOKEN_HELD(&obj
->token
);
256 globaldata_t gd
= mycpu
;
259 pg_color
= (int)(intptr_t)gd
->gd_curthread
>> 10;
260 pg_color
+= ++gd
->gd_quick_color
;
266 VMOBJDEBUG(vm_object_hold
)(vm_object_t obj VMOBJDBARGS
)
268 KKASSERT(obj
!= NULL
);
271 * Object must be held (object allocation is stable due to callers
272 * context, typically already holding the token on a parent object)
273 * prior to potentially blocking on the lock, otherwise the object
274 * can get ripped away from us.
276 refcount_acquire(&obj
->hold_count
);
279 #if defined(DEBUG_LOCKS)
280 debugvm_object_add(obj
, file
, line
, 1);
285 VMOBJDEBUG(vm_object_hold_try
)(vm_object_t obj VMOBJDBARGS
)
287 KKASSERT(obj
!= NULL
);
290 * Object must be held (object allocation is stable due to callers
291 * context, typically already holding the token on a parent object)
292 * prior to potentially blocking on the lock, otherwise the object
293 * can get ripped away from us.
295 refcount_acquire(&obj
->hold_count
);
296 if (vm_object_lock_try(obj
) == 0) {
297 if (refcount_release(&obj
->hold_count
)) {
298 if (obj
->ref_count
== 0 && (obj
->flags
& OBJ_DEAD
))
299 zfree(obj_zone
, obj
);
304 #if defined(DEBUG_LOCKS)
305 debugvm_object_add(obj
, file
, line
, 1);
311 VMOBJDEBUG(vm_object_hold_shared
)(vm_object_t obj VMOBJDBARGS
)
313 KKASSERT(obj
!= NULL
);
316 * Object must be held (object allocation is stable due to callers
317 * context, typically already holding the token on a parent object)
318 * prior to potentially blocking on the lock, otherwise the object
319 * can get ripped away from us.
321 refcount_acquire(&obj
->hold_count
);
322 vm_object_lock_shared(obj
);
324 #if defined(DEBUG_LOCKS)
325 debugvm_object_add(obj
, file
, line
, 1);
330 * Drop the token and hold_count on the object.
332 * WARNING! Token might be shared.
335 VMOBJDEBUG(vm_object_drop
)(vm_object_t obj VMOBJDBARGS
)
341 * No new holders should be possible once we drop hold_count 1->0 as
342 * there is no longer any way to reference the object.
344 KKASSERT(obj
->hold_count
> 0);
345 if (refcount_release(&obj
->hold_count
)) {
346 #if defined(DEBUG_LOCKS)
347 debugvm_object_add(obj
, file
, line
, -1);
350 if (obj
->ref_count
== 0 && (obj
->flags
& OBJ_DEAD
)) {
351 vm_object_unlock(obj
);
352 zfree(obj_zone
, obj
);
354 vm_object_unlock(obj
);
357 #if defined(DEBUG_LOCKS)
358 debugvm_object_add(obj
, file
, line
, -1);
360 vm_object_unlock(obj
);
365 * Initialize a freshly allocated object, returning a held object.
367 * Used only by vm_object_allocate() and zinitna().
372 _vm_object_allocate(objtype_t type
, vm_pindex_t size
, vm_object_t object
)
377 RB_INIT(&object
->rb_memq
);
378 LIST_INIT(&object
->shadow_head
);
379 lwkt_token_init(&object
->token
, "vmobj");
383 object
->ref_count
= 1;
384 object
->memattr
= VM_MEMATTR_DEFAULT
;
385 object
->hold_count
= 0;
387 if ((object
->type
== OBJT_DEFAULT
) || (object
->type
== OBJT_SWAP
))
388 vm_object_set_flag(object
, OBJ_ONEMAPPING
);
389 object
->paging_in_progress
= 0;
390 object
->resident_page_count
= 0;
391 object
->agg_pv_list_count
= 0;
392 object
->shadow_count
= 0;
393 /* cpu localization twist */
394 object
->pg_color
= vm_quickcolor();
395 if ( size
> (PQ_L2_SIZE
/ 3 + PQ_PRIME1
))
396 incr
= PQ_L2_SIZE
/ 3 + PQ_PRIME1
;
399 next_index
= (next_index
+ incr
) & PQ_L2_MASK
;
400 object
->handle
= NULL
;
401 object
->backing_object
= NULL
;
402 object
->backing_object_offset
= (vm_ooffset_t
)0;
404 object
->generation
++;
405 object
->swblock_count
= 0;
406 RB_INIT(&object
->swblock_root
);
407 vm_object_lock_init(object
);
408 pmap_object_init(object
);
410 vm_object_hold(object
);
412 n
= VMOBJ_HASH(object
);
413 atomic_add_long(&vm_object_count
, 1);
414 lwkt_gettoken(&vmobj_tokens
[n
]);
415 TAILQ_INSERT_TAIL(&vm_object_lists
[n
], object
, object_list
);
416 lwkt_reltoken(&vmobj_tokens
[n
]);
420 * Initialize the VM objects module.
422 * Called from the low level boot code only.
429 for (i
= 0; i
< VMOBJ_HSIZE
; ++i
) {
430 TAILQ_INIT(&vm_object_lists
[i
]);
431 lwkt_token_init(&vmobj_tokens
[i
], "vmobjlst");
434 _vm_object_allocate(OBJT_DEFAULT
, OFF_TO_IDX(KvaEnd
),
436 vm_object_drop(&kernel_object
);
438 obj_zone
= &obj_zone_store
;
439 zbootinit(obj_zone
, "VM OBJECT", sizeof (struct vm_object
),
440 vm_objects_init
, VM_OBJECTS_INIT
);
444 vm_object_init2(void)
446 zinitna(obj_zone
, NULL
, NULL
, 0, 0, ZONE_PANICFAIL
, 1);
450 * Allocate and return a new object of the specified type and size.
455 vm_object_allocate(objtype_t type
, vm_pindex_t size
)
459 result
= (vm_object_t
) zalloc(obj_zone
);
461 _vm_object_allocate(type
, size
, result
);
462 vm_object_drop(result
);
468 * This version returns a held object, allowing further atomic initialization
472 vm_object_allocate_hold(objtype_t type
, vm_pindex_t size
)
476 result
= (vm_object_t
) zalloc(obj_zone
);
478 _vm_object_allocate(type
, size
, result
);
484 * Add an additional reference to a vm_object. The object must already be
485 * held. The original non-lock version is no longer supported. The object
486 * must NOT be chain locked by anyone at the time the reference is added.
488 * Referencing a chain-locked object can blow up the fairly sensitive
489 * ref_count and shadow_count tests in the deallocator. Most callers
490 * will call vm_object_chain_wait() prior to calling
491 * vm_object_reference_locked() to avoid the case.
493 * The object must be held, but may be held shared if desired (hence why
494 * we use an atomic op).
497 VMOBJDEBUG(vm_object_reference_locked
)(vm_object_t object VMOBJDBARGS
)
499 KKASSERT(object
!= NULL
);
500 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
501 KKASSERT((object
->chainlk
& (CHAINLK_EXCL
| CHAINLK_MASK
)) == 0);
502 atomic_add_int(&object
->ref_count
, 1);
503 if (object
->type
== OBJT_VNODE
) {
504 vref(object
->handle
);
505 /* XXX what if the vnode is being destroyed? */
507 #if defined(DEBUG_LOCKS)
508 debugvm_object_add(object
, file
, line
, 1);
513 * This version is only allowed for vnode objects.
516 VMOBJDEBUG(vm_object_reference_quick
)(vm_object_t object VMOBJDBARGS
)
518 KKASSERT(object
->type
== OBJT_VNODE
);
519 atomic_add_int(&object
->ref_count
, 1);
520 vref(object
->handle
);
521 #if defined(DEBUG_LOCKS)
522 debugvm_object_add(object
, file
, line
, 1);
527 * Object OBJ_CHAINLOCK lock handling.
529 * The caller can chain-lock backing objects recursively and then
530 * use vm_object_chain_release_all() to undo the whole chain.
532 * Chain locks are used to prevent collapses and are only applicable
533 * to OBJT_DEFAULT and OBJT_SWAP objects. Chain locking operations
534 * on other object types are ignored. This is also important because
535 * it allows e.g. the vnode underlying a memory mapping to take concurrent
538 * The object must usually be held on entry, though intermediate
539 * objects need not be held on release. The object must be held exclusively,
540 * NOT shared. Note that the prefault path checks the shared state and
541 * avoids using the chain functions.
544 vm_object_chain_wait(vm_object_t object
, int shared
)
546 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
548 uint32_t chainlk
= object
->chainlk
;
552 if (chainlk
& (CHAINLK_EXCL
| CHAINLK_EXCLREQ
)) {
553 tsleep_interlock(object
, 0);
554 if (atomic_cmpset_int(&object
->chainlk
,
556 chainlk
| CHAINLK_WAIT
)) {
557 tsleep(object
, PINTERLOCKED
,
566 if (chainlk
& (CHAINLK_MASK
| CHAINLK_EXCL
)) {
567 tsleep_interlock(object
, 0);
568 if (atomic_cmpset_int(&object
->chainlk
,
570 chainlk
| CHAINLK_WAIT
))
572 tsleep(object
, PINTERLOCKED
,
577 if (atomic_cmpset_int(&object
->chainlk
,
579 chainlk
& ~CHAINLK_WAIT
))
581 if (chainlk
& CHAINLK_WAIT
)
593 vm_object_chain_acquire(vm_object_t object
, int shared
)
595 if (object
->type
!= OBJT_DEFAULT
&& object
->type
!= OBJT_SWAP
)
597 if (vm_shared_fault
== 0)
601 uint32_t chainlk
= object
->chainlk
;
605 if (chainlk
& (CHAINLK_EXCL
| CHAINLK_EXCLREQ
)) {
606 tsleep_interlock(object
, 0);
607 if (atomic_cmpset_int(&object
->chainlk
,
609 chainlk
| CHAINLK_WAIT
)) {
610 tsleep(object
, PINTERLOCKED
,
614 } else if (atomic_cmpset_int(&object
->chainlk
,
615 chainlk
, chainlk
+ 1)) {
620 if (chainlk
& (CHAINLK_MASK
| CHAINLK_EXCL
)) {
621 tsleep_interlock(object
, 0);
622 if (atomic_cmpset_int(&object
->chainlk
,
627 tsleep(object
, PINTERLOCKED
,
632 if (atomic_cmpset_int(&object
->chainlk
,
634 (chainlk
| CHAINLK_EXCL
) &
637 if (chainlk
& CHAINLK_WAIT
)
649 vm_object_chain_release(vm_object_t object
)
651 /*ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));*/
652 if (object
->type
!= OBJT_DEFAULT
&& object
->type
!= OBJT_SWAP
)
654 KKASSERT(object
->chainlk
& (CHAINLK_MASK
| CHAINLK_EXCL
));
656 uint32_t chainlk
= object
->chainlk
;
659 if (chainlk
& CHAINLK_MASK
) {
660 if ((chainlk
& CHAINLK_MASK
) == 1 &&
661 atomic_cmpset_int(&object
->chainlk
,
663 (chainlk
- 1) & ~CHAINLK_WAIT
)) {
664 if (chainlk
& CHAINLK_WAIT
)
668 if ((chainlk
& CHAINLK_MASK
) > 1 &&
669 atomic_cmpset_int(&object
->chainlk
,
670 chainlk
, chainlk
- 1)) {
675 KKASSERT(chainlk
& CHAINLK_EXCL
);
676 if (atomic_cmpset_int(&object
->chainlk
,
678 chainlk
& ~(CHAINLK_EXCL
|
680 if (chainlk
& CHAINLK_WAIT
)
689 * Release the chain from first_object through and including stopobj.
690 * The caller is typically holding the first and last object locked
691 * (shared or exclusive) to prevent destruction races.
693 * We release stopobj first as an optimization as this object is most
694 * likely to be shared across multiple processes.
697 vm_object_chain_release_all(vm_object_t first_object
, vm_object_t stopobj
)
699 vm_object_t backing_object
;
702 vm_object_chain_release(stopobj
);
703 object
= first_object
;
705 while (object
!= stopobj
) {
707 backing_object
= object
->backing_object
;
708 vm_object_chain_release(object
);
709 object
= backing_object
;
714 * Dereference an object and its underlying vnode. The object may be
715 * held shared. On return the object will remain held.
717 * This function may return a vnode in *vpp which the caller must release
718 * after the caller drops its own lock. If vpp is NULL, we assume that
719 * the caller was holding an exclusive lock on the object and we vrele()
723 VMOBJDEBUG(vm_object_vndeallocate
)(vm_object_t object
, struct vnode
**vpp
726 struct vnode
*vp
= (struct vnode
*) object
->handle
;
728 KASSERT(object
->type
== OBJT_VNODE
,
729 ("vm_object_vndeallocate: not a vnode object"));
730 KASSERT(vp
!= NULL
, ("vm_object_vndeallocate: missing vp"));
731 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
733 if (object
->ref_count
== 0) {
734 vprint("vm_object_vndeallocate", vp
);
735 panic("vm_object_vndeallocate: bad object reference count");
739 int count
= object
->ref_count
;
742 vm_object_upgrade(object
);
743 if (atomic_cmpset_int(&object
->ref_count
, count
, 0)) {
744 vclrflags(vp
, VTEXT
);
748 if (atomic_cmpset_int(&object
->ref_count
,
755 #if defined(DEBUG_LOCKS)
756 debugvm_object_add(object
, file
, line
, -1);
760 * vrele or return the vp to vrele. We can only safely vrele(vp)
761 * if the object was locked exclusively. But there are two races
764 * We had to upgrade the object above to safely clear VTEXT
765 * but the alternative path where the shared lock is retained
766 * can STILL race to 0 in other paths and cause our own vrele()
767 * to terminate the vnode. We can't allow that if the VM object
768 * is still locked shared.
777 * Release a reference to the specified object, gained either through a
778 * vm_object_allocate or a vm_object_reference call. When all references
779 * are gone, storage associated with this object may be relinquished.
781 * The caller does not have to hold the object locked but must have control
782 * over the reference in question in order to guarantee that the object
783 * does not get ripped out from under us.
785 * XXX Currently all deallocations require an exclusive lock.
788 VMOBJDEBUG(vm_object_deallocate
)(vm_object_t object VMOBJDBARGS
)
797 count
= object
->ref_count
;
801 * If decrementing the count enters into special handling
802 * territory (0, 1, or 2) we have to do it the hard way.
803 * Fortunate though, objects with only a few refs like this
804 * are not likely to be heavily contended anyway.
806 * For vnode objects we only care about 1->0 transitions.
808 if (count
<= 3 || (object
->type
== OBJT_VNODE
&& count
<= 1)) {
809 #if defined(DEBUG_LOCKS)
810 debugvm_object_add(object
, file
, line
, 0);
812 vm_object_hold(object
);
813 vm_object_deallocate_locked(object
);
814 vm_object_drop(object
);
819 * Try to decrement ref_count without acquiring a hold on
820 * the object. This is particularly important for the exec*()
821 * and exit*() code paths because the program binary may
822 * have a great deal of sharing and an exclusive lock will
823 * crowbar performance in those circumstances.
825 if (object
->type
== OBJT_VNODE
) {
826 vp
= (struct vnode
*)object
->handle
;
827 if (atomic_cmpset_int(&object
->ref_count
,
829 #if defined(DEBUG_LOCKS)
830 debugvm_object_add(object
, file
, line
, -1);
838 if (atomic_cmpset_int(&object
->ref_count
,
840 #if defined(DEBUG_LOCKS)
841 debugvm_object_add(object
, file
, line
, -1);
852 VMOBJDEBUG(vm_object_deallocate_locked
)(vm_object_t object VMOBJDBARGS
)
854 struct vm_object_dealloc_list
*dlist
= NULL
;
855 struct vm_object_dealloc_list
*dtmp
;
860 * We may chain deallocate object, but additional objects may
861 * collect on the dlist which also have to be deallocated. We
862 * must avoid a recursion, vm_object chains can get deep.
866 while (object
!= NULL
) {
868 * vnode case, caller either locked the object exclusively
869 * or this is a recursion with must_drop != 0 and the vnode
870 * object will be locked shared.
872 * If locked shared we have to drop the object before we can
873 * call vrele() or risk a shared/exclusive livelock.
875 if (object
->type
== OBJT_VNODE
) {
876 ASSERT_LWKT_TOKEN_HELD(&object
->token
);
878 struct vnode
*tmp_vp
;
880 vm_object_vndeallocate(object
, &tmp_vp
);
881 vm_object_drop(object
);
886 vm_object_vndeallocate(object
, NULL
);
890 ASSERT_LWKT_TOKEN_HELD_EXCL(&object
->token
);
893 * Normal case (object is locked exclusively)
895 if (object
->ref_count
== 0) {
896 panic("vm_object_deallocate: object deallocated "
897 "too many times: %d", object
->type
);
899 if (object
->ref_count
> 2) {
900 atomic_add_int(&object
->ref_count
, -1);
901 #if defined(DEBUG_LOCKS)
902 debugvm_object_add(object
, file
, line
, -1);
908 * Here on ref_count of one or two, which are special cases for
911 * Nominal ref_count > 1 case if the second ref is not from
914 * (ONEMAPPING only applies to DEFAULT AND SWAP objects)
916 if (object
->ref_count
== 2 && object
->shadow_count
== 0) {
917 if (object
->type
== OBJT_DEFAULT
||
918 object
->type
== OBJT_SWAP
) {
919 vm_object_set_flag(object
, OBJ_ONEMAPPING
);
921 atomic_add_int(&object
->ref_count
, -1);
922 #if defined(DEBUG_LOCKS)
923 debugvm_object_add(object
, file
, line
, -1);
929 * If the second ref is from a shadow we chain along it
930 * upwards if object's handle is exhausted.
932 * We have to decrement object->ref_count before potentially
933 * collapsing the first shadow object or the collapse code
934 * will not be able to handle the degenerate case to remove
935 * object. However, if we do it too early the object can
936 * get ripped out from under us.
938 if (object
->ref_count
== 2 && object
->shadow_count
== 1 &&
939 object
->handle
== NULL
&& (object
->type
== OBJT_DEFAULT
||
940 object
->type
== OBJT_SWAP
)) {
941 temp
= LIST_FIRST(&object
->shadow_head
);
942 KKASSERT(temp
!= NULL
);
943 vm_object_hold(temp
);
946 * Wait for any paging to complete so the collapse
947 * doesn't (or isn't likely to) qcollapse. pip
948 * waiting must occur before we acquire the
952 temp
->paging_in_progress
||
953 object
->paging_in_progress
955 vm_object_pip_wait(temp
, "objde1");
956 vm_object_pip_wait(object
, "objde2");
960 * If the parent is locked we have to give up, as
961 * otherwise we would be acquiring locks in the
962 * wrong order and potentially deadlock.
964 if (temp
->chainlk
& (CHAINLK_EXCL
| CHAINLK_MASK
)) {
965 vm_object_drop(temp
);
968 vm_object_chain_acquire(temp
, 0);
971 * Recheck/retry after the hold and the paging
972 * wait, both of which can block us.
974 if (object
->ref_count
!= 2 ||
975 object
->shadow_count
!= 1 ||
977 LIST_FIRST(&object
->shadow_head
) != temp
||
978 (object
->type
!= OBJT_DEFAULT
&&
979 object
->type
!= OBJT_SWAP
)) {
980 vm_object_chain_release(temp
);
981 vm_object_drop(temp
);
986 * We can safely drop object's ref_count now.
988 KKASSERT(object
->ref_count
== 2);
989 atomic_add_int(&object
->ref_count
, -1);
990 #if defined(DEBUG_LOCKS)
991 debugvm_object_add(object
, file
, line
, -1);
995 * If our single parent is not collapseable just
996 * decrement ref_count (2->1) and stop.
998 if (temp
->handle
|| (temp
->type
!= OBJT_DEFAULT
&&
999 temp
->type
!= OBJT_SWAP
)) {
1000 vm_object_chain_release(temp
);
1001 vm_object_drop(temp
);
1006 * At this point we have already dropped object's
1007 * ref_count so it is possible for a race to
1008 * deallocate obj out from under us. Any collapse
1009 * will re-check the situation. We must not block
1010 * until we are able to collapse.
1012 * Bump temp's ref_count to avoid an unwanted
1013 * degenerate recursion (can't call
1014 * vm_object_reference_locked() because it asserts
1015 * that CHAINLOCK is not set).
1017 atomic_add_int(&temp
->ref_count
, 1);
1018 KKASSERT(temp
->ref_count
> 1);
1021 * Collapse temp, then deallocate the extra ref
1024 vm_object_collapse(temp
, &dlist
);
1025 vm_object_chain_release(temp
);
1027 vm_object_lock_swap();
1028 vm_object_drop(object
);
1036 * Drop the ref and handle termination on the 1->0 transition.
1037 * We may have blocked above so we have to recheck.
1040 KKASSERT(object
->ref_count
!= 0);
1041 if (object
->ref_count
>= 2) {
1042 atomic_add_int(&object
->ref_count
, -1);
1043 #if defined(DEBUG_LOCKS)
1044 debugvm_object_add(object
, file
, line
, -1);
1048 KKASSERT(object
->ref_count
== 1);
1051 * 1->0 transition. Chain through the backing_object.
1052 * Maintain the ref until we've located the backing object,
1055 while ((temp
= object
->backing_object
) != NULL
) {
1056 if (temp
->type
== OBJT_VNODE
)
1057 vm_object_hold_shared(temp
);
1059 vm_object_hold(temp
);
1060 if (temp
== object
->backing_object
)
1062 vm_object_drop(temp
);
1066 * 1->0 transition verified, retry if ref_count is no longer
1067 * 1. Otherwise disconnect the backing_object (temp) and
1070 if (object
->ref_count
!= 1) {
1071 vm_object_drop(temp
);
1076 * It shouldn't be possible for the object to be chain locked
1077 * if we're removing the last ref on it.
1079 * Removing object from temp's shadow list requires dropping
1080 * temp, which we will do on loop.
1082 * NOTE! vnodes do not use the shadow list, but still have
1083 * the backing_object reference.
1085 KKASSERT((object
->chainlk
& (CHAINLK_EXCL
|CHAINLK_MASK
)) == 0);
1088 if (object
->flags
& OBJ_ONSHADOW
) {
1089 LIST_REMOVE(object
, shadow_list
);
1090 temp
->shadow_count
--;
1092 vm_object_clear_flag(object
, OBJ_ONSHADOW
);
1094 object
->backing_object
= NULL
;
1097 atomic_add_int(&object
->ref_count
, -1);
1098 if ((object
->flags
& OBJ_DEAD
) == 0)
1099 vm_object_terminate(object
);
1100 if (must_drop
&& temp
)
1101 vm_object_lock_swap();
1103 vm_object_drop(object
);
1108 if (must_drop
&& object
)
1109 vm_object_drop(object
);
1112 * Additional tail recursion on dlist. Avoid a recursion. Objects
1113 * on the dlist have a hold count but are not locked.
1115 if ((dtmp
= dlist
) != NULL
) {
1117 object
= dtmp
->object
;
1118 kfree(dtmp
, M_TEMP
);
1120 vm_object_lock(object
); /* already held, add lock */
1121 must_drop
= 1; /* and we're responsible for it */
1127 * Destroy the specified object, freeing up related resources.
1129 * The object must have zero references.
1131 * The object must held. The caller is responsible for dropping the object
1132 * after terminate returns. Terminate does NOT drop the object.
1134 static int vm_object_terminate_callback(vm_page_t p
, void *data
);
1137 vm_object_terminate(vm_object_t object
)
1139 struct rb_vm_page_scan_info info
;
1143 * Make sure no one uses us. Once we set OBJ_DEAD we should be
1144 * able to safely block.
1146 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1147 KKASSERT((object
->flags
& OBJ_DEAD
) == 0);
1148 vm_object_set_flag(object
, OBJ_DEAD
);
1151 * Wait for the pageout daemon to be done with the object
1153 vm_object_pip_wait(object
, "objtrm1");
1155 KASSERT(!object
->paging_in_progress
,
1156 ("vm_object_terminate: pageout in progress"));
1159 * Clean and free the pages, as appropriate. All references to the
1160 * object are gone, so we don't need to lock it.
1162 if (object
->type
== OBJT_VNODE
) {
1166 * Clean pages and flush buffers.
1168 * NOTE! TMPFS buffer flushes do not typically flush the
1169 * actual page to swap as this would be highly
1170 * inefficient, and normal filesystems usually wrap
1171 * page flushes with buffer cache buffers.
1173 * To deal with this we have to call vinvalbuf() both
1174 * before and after the vm_object_page_clean().
1176 vp
= (struct vnode
*) object
->handle
;
1177 vinvalbuf(vp
, V_SAVE
, 0, 0);
1178 vm_object_page_clean(object
, 0, 0, OBJPC_SYNC
);
1179 vinvalbuf(vp
, V_SAVE
, 0, 0);
1183 * Wait for any I/O to complete, after which there had better not
1184 * be any references left on the object.
1186 vm_object_pip_wait(object
, "objtrm2");
1188 if (object
->ref_count
!= 0) {
1189 panic("vm_object_terminate: object with references, "
1190 "ref_count=%d", object
->ref_count
);
1194 * Cleanup any shared pmaps associated with this object.
1196 pmap_object_free(object
);
1199 * Now free any remaining pages. For internal objects, this also
1200 * removes them from paging queues. Don't free wired pages, just
1201 * remove them from the object.
1204 info
.object
= object
;
1205 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, NULL
,
1206 vm_object_terminate_callback
, &info
);
1209 * Let the pager know object is dead.
1211 vm_pager_deallocate(object
);
1214 * Wait for the object hold count to hit 1, clean out pages as
1215 * we go. vmobj_token interlocks any race conditions that might
1216 * pick the object up from the vm_object_list after we have cleared
1220 if (RB_ROOT(&object
->rb_memq
) == NULL
)
1222 kprintf("vm_object_terminate: Warning, object %p "
1223 "still has %d pages\n",
1224 object
, object
->resident_page_count
);
1225 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, NULL
,
1226 vm_object_terminate_callback
, &info
);
1230 * There had better not be any pages left
1232 KKASSERT(object
->resident_page_count
== 0);
1235 * Remove the object from the global object list.
1237 n
= VMOBJ_HASH(object
);
1238 lwkt_gettoken(&vmobj_tokens
[n
]);
1239 TAILQ_REMOVE(&vm_object_lists
[n
], object
, object_list
);
1240 lwkt_reltoken(&vmobj_tokens
[n
]);
1241 atomic_add_long(&vm_object_count
, -1);
1243 if (object
->ref_count
!= 0) {
1244 panic("vm_object_terminate2: object with references, "
1245 "ref_count=%d", object
->ref_count
);
1249 * NOTE: The object hold_count is at least 1, so we cannot zfree()
1250 * the object here. See vm_object_drop().
1255 * The caller must hold the object.
1258 vm_object_terminate_callback(vm_page_t p
, void *data
)
1260 struct rb_vm_page_scan_info
*info
= data
;
1263 if ((++info
->count
& 63) == 0)
1266 if (object
!= info
->object
) {
1267 kprintf("vm_object_terminate_callback: obj/pg race %p/%p\n",
1271 vm_page_busy_wait(p
, TRUE
, "vmpgtrm");
1272 if (object
!= p
->object
) {
1273 kprintf("vm_object_terminate: Warning: Encountered "
1274 "busied page %p on queue %d\n", p
, p
->queue
);
1276 } else if (p
->wire_count
== 0) {
1278 * NOTE: p->dirty and PG_NEED_COMMIT are ignored.
1281 mycpu
->gd_cnt
.v_pfree
++;
1283 if (p
->queue
!= PQ_NONE
)
1284 kprintf("vm_object_terminate: Warning: Encountered "
1285 "wired page %p on queue %d\n", p
, p
->queue
);
1293 * Clean all dirty pages in the specified range of object. Leaves page
1294 * on whatever queue it is currently on. If NOSYNC is set then do not
1295 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
1296 * leaving the object dirty.
1298 * When stuffing pages asynchronously, allow clustering. XXX we need a
1299 * synchronous clustering mode implementation.
1301 * Odd semantics: if start == end, we clean everything.
1303 * The object must be locked? XXX
1305 static int vm_object_page_clean_pass1(struct vm_page
*p
, void *data
);
1306 static int vm_object_page_clean_pass2(struct vm_page
*p
, void *data
);
1309 vm_object_page_clean(vm_object_t object
, vm_pindex_t start
, vm_pindex_t end
,
1312 struct rb_vm_page_scan_info info
;
1318 vm_object_hold(object
);
1319 if (object
->type
!= OBJT_VNODE
||
1320 (object
->flags
& OBJ_MIGHTBEDIRTY
) == 0) {
1321 vm_object_drop(object
);
1325 pagerflags
= (flags
& (OBJPC_SYNC
| OBJPC_INVAL
)) ?
1326 VM_PAGER_PUT_SYNC
: VM_PAGER_CLUSTER_OK
;
1327 pagerflags
|= (flags
& OBJPC_INVAL
) ? VM_PAGER_PUT_INVAL
: 0;
1329 vp
= object
->handle
;
1332 * Interlock other major object operations. This allows us to
1333 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY.
1335 vm_object_set_flag(object
, OBJ_CLEANING
);
1338 * Handle 'entire object' case
1340 info
.start_pindex
= start
;
1342 info
.end_pindex
= object
->size
- 1;
1344 info
.end_pindex
= end
- 1;
1346 wholescan
= (start
== 0 && info
.end_pindex
== object
->size
- 1);
1348 info
.pagerflags
= pagerflags
;
1349 info
.object
= object
;
1353 * If cleaning the entire object do a pass to mark the pages read-only.
1354 * If everything worked out ok, clear OBJ_WRITEABLE and
1359 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, rb_vm_page_scancmp
,
1360 vm_object_page_clean_pass1
, &info
);
1361 if (info
.error
== 0) {
1362 vm_object_clear_flag(object
,
1363 OBJ_WRITEABLE
|OBJ_MIGHTBEDIRTY
);
1364 if (object
->type
== OBJT_VNODE
&&
1365 (vp
= (struct vnode
*)object
->handle
) != NULL
) {
1367 * Use new-style interface to clear VISDIRTY
1368 * because the vnode is not necessarily removed
1369 * from the syncer list(s) as often as it was
1370 * under the old interface, which can leave
1371 * the vnode on the syncer list after reclaim.
1379 * Do a pass to clean all the dirty pages we find.
1383 generation
= object
->generation
;
1384 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, rb_vm_page_scancmp
,
1385 vm_object_page_clean_pass2
, &info
);
1386 } while (info
.error
|| generation
!= object
->generation
);
1388 vm_object_clear_flag(object
, OBJ_CLEANING
);
1389 vm_object_drop(object
);
1393 * The caller must hold the object.
1397 vm_object_page_clean_pass1(struct vm_page
*p
, void *data
)
1399 struct rb_vm_page_scan_info
*info
= data
;
1401 if ((++info
->count
& 63) == 0)
1403 if (p
->object
!= info
->object
||
1404 p
->pindex
< info
->start_pindex
||
1405 p
->pindex
> info
->end_pindex
) {
1406 kprintf("vm_object_page_clean_pass1: obj/pg race %p/%p\n",
1410 vm_page_flag_set(p
, PG_CLEANCHK
);
1411 if ((info
->limit
& OBJPC_NOSYNC
) && (p
->flags
& PG_NOSYNC
)) {
1413 } else if (vm_page_busy_try(p
, FALSE
) == 0) {
1414 if (p
->object
== info
->object
)
1415 vm_page_protect(p
, VM_PROT_READ
);
1424 * The caller must hold the object
1428 vm_object_page_clean_pass2(struct vm_page
*p
, void *data
)
1430 struct rb_vm_page_scan_info
*info
= data
;
1433 if (p
->object
!= info
->object
||
1434 p
->pindex
< info
->start_pindex
||
1435 p
->pindex
> info
->end_pindex
) {
1436 kprintf("vm_object_page_clean_pass2: obj/pg race %p/%p\n",
1442 * Do not mess with pages that were inserted after we started
1443 * the cleaning pass.
1445 if ((p
->flags
& PG_CLEANCHK
) == 0)
1448 generation
= info
->object
->generation
;
1449 vm_page_busy_wait(p
, TRUE
, "vpcwai");
1451 if (p
->object
!= info
->object
||
1452 p
->pindex
< info
->start_pindex
||
1453 p
->pindex
> info
->end_pindex
||
1454 info
->object
->generation
!= generation
) {
1461 * Before wasting time traversing the pmaps, check for trivial
1462 * cases where the page cannot be dirty.
1464 if (p
->valid
== 0 || (p
->queue
- p
->pc
) == PQ_CACHE
) {
1465 KKASSERT((p
->dirty
& p
->valid
) == 0 &&
1466 (p
->flags
& PG_NEED_COMMIT
) == 0);
1472 * Check whether the page is dirty or not. The page has been set
1473 * to be read-only so the check will not race a user dirtying the
1476 vm_page_test_dirty(p
);
1477 if ((p
->dirty
& p
->valid
) == 0 && (p
->flags
& PG_NEED_COMMIT
) == 0) {
1478 vm_page_flag_clear(p
, PG_CLEANCHK
);
1484 * If we have been asked to skip nosync pages and this is a
1485 * nosync page, skip it. Note that the object flags were
1486 * not cleared in this case (because pass1 will have returned an
1487 * error), so we do not have to set them.
1489 if ((info
->limit
& OBJPC_NOSYNC
) && (p
->flags
& PG_NOSYNC
)) {
1490 vm_page_flag_clear(p
, PG_CLEANCHK
);
1496 * Flush as many pages as we can. PG_CLEANCHK will be cleared on
1497 * the pages that get successfully flushed. Set info->error if
1498 * we raced an object modification.
1500 vm_object_page_collect_flush(info
->object
, p
, info
->pagerflags
);
1501 /* vm_wait_nominal(); this can deadlock the system in syncer/pageout */
1503 if ((++info
->count
& 63) == 0)
1510 * Collect the specified page and nearby pages and flush them out.
1511 * The number of pages flushed is returned. The passed page is busied
1512 * by the caller and we are responsible for its disposition.
1514 * The caller must hold the object.
1517 vm_object_page_collect_flush(vm_object_t object
, vm_page_t p
, int pagerflags
)
1525 vm_page_t ma
[BLIST_MAX_ALLOC
];
1527 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1530 page_base
= pi
% BLIST_MAX_ALLOC
;
1538 tp
= vm_page_lookup_busy_try(object
, pi
- page_base
+ ib
,
1544 if ((pagerflags
& VM_PAGER_IGNORE_CLEANCHK
) == 0 &&
1545 (tp
->flags
& PG_CLEANCHK
) == 0) {
1549 if ((tp
->queue
- tp
->pc
) == PQ_CACHE
) {
1550 vm_page_flag_clear(tp
, PG_CLEANCHK
);
1554 vm_page_test_dirty(tp
);
1555 if ((tp
->dirty
& tp
->valid
) == 0 &&
1556 (tp
->flags
& PG_NEED_COMMIT
) == 0) {
1557 vm_page_flag_clear(tp
, PG_CLEANCHK
);
1566 while (is
< BLIST_MAX_ALLOC
&&
1567 pi
- page_base
+ is
< object
->size
) {
1570 tp
= vm_page_lookup_busy_try(object
, pi
- page_base
+ is
,
1576 if ((pagerflags
& VM_PAGER_IGNORE_CLEANCHK
) == 0 &&
1577 (tp
->flags
& PG_CLEANCHK
) == 0) {
1581 if ((tp
->queue
- tp
->pc
) == PQ_CACHE
) {
1582 vm_page_flag_clear(tp
, PG_CLEANCHK
);
1586 vm_page_test_dirty(tp
);
1587 if ((tp
->dirty
& tp
->valid
) == 0 &&
1588 (tp
->flags
& PG_NEED_COMMIT
) == 0) {
1589 vm_page_flag_clear(tp
, PG_CLEANCHK
);
1598 * All pages in the ma[] array are busied now
1600 for (i
= ib
; i
< is
; ++i
) {
1601 vm_page_flag_clear(ma
[i
], PG_CLEANCHK
);
1602 vm_page_hold(ma
[i
]); /* XXX need this any more? */
1604 vm_pageout_flush(&ma
[ib
], is
- ib
, pagerflags
);
1605 for (i
= ib
; i
< is
; ++i
) /* XXX need this any more? */
1606 vm_page_unhold(ma
[i
]);
1610 * Same as vm_object_pmap_copy, except range checking really
1611 * works, and is meant for small sections of an object.
1613 * This code protects resident pages by making them read-only
1614 * and is typically called on a fork or split when a page
1615 * is converted to copy-on-write.
1617 * NOTE: If the page is already at VM_PROT_NONE, calling
1618 * vm_page_protect will have no effect.
1621 vm_object_pmap_copy_1(vm_object_t object
, vm_pindex_t start
, vm_pindex_t end
)
1626 if (object
== NULL
|| (object
->flags
& OBJ_WRITEABLE
) == 0)
1629 vm_object_hold(object
);
1630 for (idx
= start
; idx
< end
; idx
++) {
1631 p
= vm_page_lookup(object
, idx
);
1634 vm_page_protect(p
, VM_PROT_READ
);
1636 vm_object_drop(object
);
1640 * Removes all physical pages in the specified object range from all
1643 * The object must *not* be locked.
1646 static int vm_object_pmap_remove_callback(vm_page_t p
, void *data
);
1649 vm_object_pmap_remove(vm_object_t object
, vm_pindex_t start
, vm_pindex_t end
)
1651 struct rb_vm_page_scan_info info
;
1655 info
.start_pindex
= start
;
1656 info
.end_pindex
= end
- 1;
1658 info
.object
= object
;
1660 vm_object_hold(object
);
1661 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, rb_vm_page_scancmp
,
1662 vm_object_pmap_remove_callback
, &info
);
1663 if (start
== 0 && end
== object
->size
)
1664 vm_object_clear_flag(object
, OBJ_WRITEABLE
);
1665 vm_object_drop(object
);
1669 * The caller must hold the object
1672 vm_object_pmap_remove_callback(vm_page_t p
, void *data
)
1674 struct rb_vm_page_scan_info
*info
= data
;
1676 if ((++info
->count
& 63) == 0)
1679 if (info
->object
!= p
->object
||
1680 p
->pindex
< info
->start_pindex
||
1681 p
->pindex
> info
->end_pindex
) {
1682 kprintf("vm_object_pmap_remove_callback: obj/pg race %p/%p\n",
1687 vm_page_protect(p
, VM_PROT_NONE
);
1693 * Implements the madvise function at the object/page level.
1695 * MADV_WILLNEED (any object)
1697 * Activate the specified pages if they are resident.
1699 * MADV_DONTNEED (any object)
1701 * Deactivate the specified pages if they are resident.
1703 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only)
1705 * Deactivate and clean the specified pages if they are
1706 * resident. This permits the process to reuse the pages
1707 * without faulting or the kernel to reclaim the pages
1713 vm_object_madvise(vm_object_t object
, vm_pindex_t pindex
, int count
, int advise
)
1715 vm_pindex_t end
, tpindex
;
1716 vm_object_t tobject
;
1724 end
= pindex
+ count
;
1726 vm_object_hold(object
);
1730 * Locate and adjust resident pages
1732 for (; pindex
< end
; pindex
+= 1) {
1734 if (tobject
!= object
)
1735 vm_object_drop(tobject
);
1740 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1741 * and those pages must be OBJ_ONEMAPPING.
1743 if (advise
== MADV_FREE
) {
1744 if ((tobject
->type
!= OBJT_DEFAULT
&&
1745 tobject
->type
!= OBJT_SWAP
) ||
1746 (tobject
->flags
& OBJ_ONEMAPPING
) == 0) {
1751 m
= vm_page_lookup_busy_try(tobject
, tpindex
, TRUE
, &error
);
1754 vm_page_sleep_busy(m
, TRUE
, "madvpo");
1759 * There may be swap even if there is no backing page
1761 if (advise
== MADV_FREE
&& tobject
->type
== OBJT_SWAP
)
1762 swap_pager_freespace(tobject
, tpindex
, 1);
1767 while ((xobj
= tobject
->backing_object
) != NULL
) {
1768 KKASSERT(xobj
!= object
);
1769 vm_object_hold(xobj
);
1770 if (xobj
== tobject
->backing_object
)
1772 vm_object_drop(xobj
);
1776 tpindex
+= OFF_TO_IDX(tobject
->backing_object_offset
);
1777 if (tobject
!= object
) {
1778 vm_object_lock_swap();
1779 vm_object_drop(tobject
);
1786 * If the page is not in a normal active state, we skip it.
1787 * If the page is not managed there are no page queues to
1788 * mess with. Things can break if we mess with pages in
1789 * any of the below states.
1791 if (m
->wire_count
||
1792 (m
->flags
& (PG_UNMANAGED
| PG_NEED_COMMIT
)) ||
1793 m
->valid
!= VM_PAGE_BITS_ALL
1800 * Theoretically once a page is known not to be busy, an
1801 * interrupt cannot come along and rip it out from under us.
1804 if (advise
== MADV_WILLNEED
) {
1805 vm_page_activate(m
);
1806 } else if (advise
== MADV_DONTNEED
) {
1807 vm_page_dontneed(m
);
1808 } else if (advise
== MADV_FREE
) {
1810 * Mark the page clean. This will allow the page
1811 * to be freed up by the system. However, such pages
1812 * are often reused quickly by malloc()/free()
1813 * so we do not do anything that would cause
1814 * a page fault if we can help it.
1816 * Specifically, we do not try to actually free
1817 * the page now nor do we try to put it in the
1818 * cache (which would cause a page fault on reuse).
1820 * But we do make the page is freeable as we
1821 * can without actually taking the step of unmapping
1824 pmap_clear_modify(m
);
1827 vm_page_dontneed(m
);
1828 if (tobject
->type
== OBJT_SWAP
)
1829 swap_pager_freespace(tobject
, tpindex
, 1);
1833 if (tobject
!= object
)
1834 vm_object_drop(tobject
);
1835 vm_object_drop(object
);
1839 * Create a new object which is backed by the specified existing object
1840 * range. Replace the pointer and offset that was pointing at the existing
1841 * object with the pointer/offset for the new object.
1843 * If addref is non-zero the returned object is given an additional reference.
1844 * This mechanic exists to avoid the situation where refs might be 1 and
1845 * race against a collapse when the caller intends to bump it. So the
1846 * caller cannot add the ref after the fact. Used when the caller is
1847 * duplicating a vm_map_entry.
1849 * No other requirements.
1852 vm_object_shadow(vm_object_t
*objectp
, vm_ooffset_t
*offset
, vm_size_t length
,
1862 * Don't create the new object if the old object isn't shared.
1863 * We have to chain wait before adding the reference to avoid
1864 * racing a collapse or deallocation.
1866 * Clear OBJ_ONEMAPPING flag when shadowing.
1868 * The caller owns a ref on source via *objectp which we are going
1869 * to replace. This ref is inherited by the backing_object assignment.
1870 * from nobject and does not need to be incremented here.
1872 * However, we add a temporary extra reference to the original source
1873 * prior to holding nobject in case we block, to avoid races where
1874 * someone else might believe that the source can be collapsed.
1878 if (source
->type
!= OBJT_VNODE
) {
1880 vm_object_hold(source
);
1881 vm_object_chain_wait(source
, 0);
1882 if (source
->ref_count
== 1 &&
1883 source
->handle
== NULL
&&
1884 (source
->type
== OBJT_DEFAULT
||
1885 source
->type
== OBJT_SWAP
)) {
1887 vm_object_reference_locked(source
);
1888 vm_object_clear_flag(source
,
1891 vm_object_drop(source
);
1894 vm_object_reference_locked(source
);
1895 vm_object_clear_flag(source
, OBJ_ONEMAPPING
);
1897 vm_object_reference_quick(source
);
1898 vm_object_clear_flag(source
, OBJ_ONEMAPPING
);
1903 * Allocate a new object with the given length. The new object
1904 * is returned referenced but we may have to add another one.
1905 * If we are adding a second reference we must clear OBJ_ONEMAPPING.
1906 * (typically because the caller is about to clone a vm_map_entry).
1908 * The source object currently has an extra reference to prevent
1909 * collapses into it while we mess with its shadow list, which
1910 * we will remove later in this routine.
1912 * The target object may require a second reference if asked for one
1915 result
= vm_object_allocate(OBJT_DEFAULT
, length
);
1917 panic("vm_object_shadow: no object for shadowing");
1918 vm_object_hold(result
);
1920 vm_object_reference_locked(result
);
1921 vm_object_clear_flag(result
, OBJ_ONEMAPPING
);
1925 * The new object shadows the source object. Chain wait before
1926 * adjusting shadow_count or the shadow list to avoid races.
1928 * Try to optimize the result object's page color when shadowing
1929 * in order to maintain page coloring consistency in the combined
1932 * The backing_object reference to source requires adding a ref to
1933 * source. We simply inherit the ref from the original *objectp
1934 * (which we are replacing) so no additional refs need to be added.
1935 * (we must still clean up the extra ref we had to prevent collapse
1938 * SHADOWING IS NOT APPLICABLE TO OBJT_VNODE OBJECTS
1940 KKASSERT(result
->backing_object
== NULL
);
1941 result
->backing_object
= source
;
1943 if (useshadowlist
) {
1944 vm_object_chain_wait(source
, 0);
1945 LIST_INSERT_HEAD(&source
->shadow_head
,
1946 result
, shadow_list
);
1947 source
->shadow_count
++;
1948 source
->generation
++;
1949 vm_object_set_flag(result
, OBJ_ONSHADOW
);
1951 /* cpu localization twist */
1952 result
->pg_color
= vm_quickcolor();
1956 * Adjust the return storage. Drop the ref on source before
1959 result
->backing_object_offset
= *offset
;
1960 vm_object_drop(result
);
1963 if (useshadowlist
) {
1964 vm_object_deallocate_locked(source
);
1965 vm_object_drop(source
);
1967 vm_object_deallocate(source
);
1972 * Return the new things
1977 #define OBSC_TEST_ALL_SHADOWED 0x0001
1978 #define OBSC_COLLAPSE_NOWAIT 0x0002
1979 #define OBSC_COLLAPSE_WAIT 0x0004
1981 static int vm_object_backing_scan_callback(vm_page_t p
, void *data
);
1984 * The caller must hold the object.
1987 vm_object_backing_scan(vm_object_t object
, vm_object_t backing_object
, int op
)
1989 struct rb_vm_page_scan_info info
;
1992 vm_object_assert_held(object
);
1993 vm_object_assert_held(backing_object
);
1995 KKASSERT(backing_object
== object
->backing_object
);
1996 info
.backing_offset_index
= OFF_TO_IDX(object
->backing_object_offset
);
1999 * Initial conditions
2001 if (op
& OBSC_TEST_ALL_SHADOWED
) {
2003 * We do not want to have to test for the existence of
2004 * swap pages in the backing object. XXX but with the
2005 * new swapper this would be pretty easy to do.
2007 * XXX what about anonymous MAP_SHARED memory that hasn't
2008 * been ZFOD faulted yet? If we do not test for this, the
2009 * shadow test may succeed! XXX
2011 if (backing_object
->type
!= OBJT_DEFAULT
)
2014 if (op
& OBSC_COLLAPSE_WAIT
) {
2015 KKASSERT((backing_object
->flags
& OBJ_DEAD
) == 0);
2016 vm_object_set_flag(backing_object
, OBJ_DEAD
);
2018 n
= VMOBJ_HASH(backing_object
);
2019 lwkt_gettoken(&vmobj_tokens
[n
]);
2020 TAILQ_REMOVE(&vm_object_lists
[n
], backing_object
, object_list
);
2021 lwkt_reltoken(&vmobj_tokens
[n
]);
2022 atomic_add_long(&vm_object_count
, -1);
2026 * Our scan. We have to retry if a negative error code is returned,
2027 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that
2028 * the scan had to be stopped because the parent does not completely
2031 info
.object
= object
;
2032 info
.backing_object
= backing_object
;
2036 vm_page_rb_tree_RB_SCAN(&backing_object
->rb_memq
, NULL
,
2037 vm_object_backing_scan_callback
,
2039 } while (info
.error
< 0);
2045 * The caller must hold the object.
2048 vm_object_backing_scan_callback(vm_page_t p
, void *data
)
2050 struct rb_vm_page_scan_info
*info
= data
;
2051 vm_object_t backing_object
;
2054 vm_pindex_t new_pindex
;
2055 vm_pindex_t backing_offset_index
;
2059 new_pindex
= pindex
- info
->backing_offset_index
;
2061 object
= info
->object
;
2062 backing_object
= info
->backing_object
;
2063 backing_offset_index
= info
->backing_offset_index
;
2065 if (op
& OBSC_TEST_ALL_SHADOWED
) {
2069 * Ignore pages outside the parent object's range
2070 * and outside the parent object's mapping of the
2073 * note that we do not busy the backing object's
2076 if (pindex
< backing_offset_index
||
2077 new_pindex
>= object
->size
2083 * See if the parent has the page or if the parent's
2084 * object pager has the page. If the parent has the
2085 * page but the page is not valid, the parent's
2086 * object pager must have the page.
2088 * If this fails, the parent does not completely shadow
2089 * the object and we might as well give up now.
2091 pp
= vm_page_lookup(object
, new_pindex
);
2092 if ((pp
== NULL
|| pp
->valid
== 0) &&
2093 !vm_pager_has_page(object
, new_pindex
)
2095 info
->error
= 0; /* problemo */
2096 return(-1); /* stop the scan */
2101 * Check for busy page. Note that we may have lost (p) when we
2102 * possibly blocked above.
2104 if (op
& (OBSC_COLLAPSE_WAIT
| OBSC_COLLAPSE_NOWAIT
)) {
2107 if (vm_page_busy_try(p
, TRUE
)) {
2108 if (op
& OBSC_COLLAPSE_NOWAIT
) {
2112 * If we slept, anything could have
2113 * happened. Ask that the scan be restarted.
2115 * Since the object is marked dead, the
2116 * backing offset should not have changed.
2118 vm_page_sleep_busy(p
, TRUE
, "vmocol");
2125 * If (p) is no longer valid restart the scan.
2127 if (p
->object
!= backing_object
|| p
->pindex
!= pindex
) {
2128 kprintf("vm_object_backing_scan: Warning: page "
2129 "%p ripped out from under us\n", p
);
2135 if (op
& OBSC_COLLAPSE_NOWAIT
) {
2136 if (p
->valid
== 0 ||
2138 (p
->flags
& PG_NEED_COMMIT
)) {
2143 /* XXX what if p->valid == 0 , hold_count, etc? */
2147 p
->object
== backing_object
,
2148 ("vm_object_qcollapse(): object mismatch")
2152 * Destroy any associated swap
2154 if (backing_object
->type
== OBJT_SWAP
)
2155 swap_pager_freespace(backing_object
, p
->pindex
, 1);
2158 p
->pindex
< backing_offset_index
||
2159 new_pindex
>= object
->size
2162 * Page is out of the parent object's range, we
2163 * can simply destroy it.
2165 vm_page_protect(p
, VM_PROT_NONE
);
2170 pp
= vm_page_lookup(object
, new_pindex
);
2171 if (pp
!= NULL
|| vm_pager_has_page(object
, new_pindex
)) {
2173 * page already exists in parent OR swap exists
2174 * for this location in the parent. Destroy
2175 * the original page from the backing object.
2177 * Leave the parent's page alone
2179 vm_page_protect(p
, VM_PROT_NONE
);
2185 * Page does not exist in parent, rename the
2186 * page from the backing object to the main object.
2188 * If the page was mapped to a process, it can remain
2189 * mapped through the rename.
2191 if ((p
->queue
- p
->pc
) == PQ_CACHE
)
2192 vm_page_deactivate(p
);
2194 vm_page_rename(p
, object
, new_pindex
);
2196 /* page automatically made dirty by rename */
2202 * This version of collapse allows the operation to occur earlier and
2203 * when paging_in_progress is true for an object... This is not a complete
2204 * operation, but should plug 99.9% of the rest of the leaks.
2206 * The caller must hold the object and backing_object and both must be
2209 * (only called from vm_object_collapse)
2212 vm_object_qcollapse(vm_object_t object
, vm_object_t backing_object
)
2214 if (backing_object
->ref_count
== 1) {
2215 atomic_add_int(&backing_object
->ref_count
, 2);
2216 #if defined(DEBUG_LOCKS)
2217 debugvm_object_add(backing_object
, "qcollapse", 1, 2);
2219 vm_object_backing_scan(object
, backing_object
,
2220 OBSC_COLLAPSE_NOWAIT
);
2221 atomic_add_int(&backing_object
->ref_count
, -2);
2222 #if defined(DEBUG_LOCKS)
2223 debugvm_object_add(backing_object
, "qcollapse", 2, -2);
2229 * Collapse an object with the object backing it. Pages in the backing
2230 * object are moved into the parent, and the backing object is deallocated.
2231 * Any conflict is resolved in favor of the parent's existing pages.
2233 * object must be held and chain-locked on call.
2235 * The caller must have an extra ref on object to prevent a race from
2236 * destroying it during the collapse.
2239 vm_object_collapse(vm_object_t object
, struct vm_object_dealloc_list
**dlistp
)
2241 struct vm_object_dealloc_list
*dlist
= NULL
;
2242 vm_object_t backing_object
;
2245 * Only one thread is attempting a collapse at any given moment.
2246 * There are few restrictions for (object) that callers of this
2247 * function check so reentrancy is likely.
2249 KKASSERT(object
!= NULL
);
2250 vm_object_assert_held(object
);
2251 KKASSERT(object
->chainlk
& (CHAINLK_MASK
| CHAINLK_EXCL
));
2258 * We can only collapse a DEFAULT/SWAP object with a
2259 * DEFAULT/SWAP object.
2261 if (object
->type
!= OBJT_DEFAULT
&& object
->type
!= OBJT_SWAP
) {
2262 backing_object
= NULL
;
2266 backing_object
= object
->backing_object
;
2267 if (backing_object
== NULL
)
2269 if (backing_object
->type
!= OBJT_DEFAULT
&&
2270 backing_object
->type
!= OBJT_SWAP
) {
2271 backing_object
= NULL
;
2276 * Hold the backing_object and check for races
2278 vm_object_hold(backing_object
);
2279 if (backing_object
!= object
->backing_object
||
2280 (backing_object
->type
!= OBJT_DEFAULT
&&
2281 backing_object
->type
!= OBJT_SWAP
)) {
2282 vm_object_drop(backing_object
);
2287 * Chain-lock the backing object too because if we
2288 * successfully merge its pages into the top object we
2289 * will collapse backing_object->backing_object as the
2290 * new backing_object. Re-check that it is still our
2293 vm_object_chain_acquire(backing_object
, 0);
2294 if (backing_object
!= object
->backing_object
) {
2295 vm_object_chain_release(backing_object
);
2296 vm_object_drop(backing_object
);
2301 * we check the backing object first, because it is most likely
2304 if (backing_object
->handle
!= NULL
||
2305 (backing_object
->type
!= OBJT_DEFAULT
&&
2306 backing_object
->type
!= OBJT_SWAP
) ||
2307 (backing_object
->flags
& OBJ_DEAD
) ||
2308 object
->handle
!= NULL
||
2309 (object
->type
!= OBJT_DEFAULT
&&
2310 object
->type
!= OBJT_SWAP
) ||
2311 (object
->flags
& OBJ_DEAD
)) {
2316 * If paging is in progress we can't do a normal collapse.
2319 object
->paging_in_progress
!= 0 ||
2320 backing_object
->paging_in_progress
!= 0
2322 vm_object_qcollapse(object
, backing_object
);
2327 * We know that we can either collapse the backing object (if
2328 * the parent is the only reference to it) or (perhaps) have
2329 * the parent bypass the object if the parent happens to shadow
2330 * all the resident pages in the entire backing object.
2332 * This is ignoring pager-backed pages such as swap pages.
2333 * vm_object_backing_scan fails the shadowing test in this
2336 if (backing_object
->ref_count
== 1) {
2338 * If there is exactly one reference to the backing
2339 * object, we can collapse it into the parent.
2341 KKASSERT(object
->backing_object
== backing_object
);
2342 vm_object_backing_scan(object
, backing_object
,
2343 OBSC_COLLAPSE_WAIT
);
2346 * Move the pager from backing_object to object.
2348 if (backing_object
->type
== OBJT_SWAP
) {
2349 vm_object_pip_add(backing_object
, 1);
2352 * scrap the paging_offset junk and do a
2353 * discrete copy. This also removes major
2354 * assumptions about how the swap-pager
2355 * works from where it doesn't belong. The
2356 * new swapper is able to optimize the
2357 * destroy-source case.
2359 vm_object_pip_add(object
, 1);
2360 swap_pager_copy(backing_object
, object
,
2361 OFF_TO_IDX(object
->backing_object_offset
),
2363 vm_object_pip_wakeup(object
);
2364 vm_object_pip_wakeup(backing_object
);
2368 * Object now shadows whatever backing_object did.
2369 * Remove object from backing_object's shadow_list.
2371 * Removing object from backing_objects shadow list
2372 * requires releasing object, which we will do below.
2374 KKASSERT(object
->backing_object
== backing_object
);
2375 if (object
->flags
& OBJ_ONSHADOW
) {
2376 LIST_REMOVE(object
, shadow_list
);
2377 backing_object
->shadow_count
--;
2378 backing_object
->generation
++;
2379 vm_object_clear_flag(object
, OBJ_ONSHADOW
);
2383 * backing_object->backing_object moves from within
2384 * backing_object to within object.
2386 * OBJT_VNODE bbobj's should have empty shadow lists.
2388 while ((bbobj
= backing_object
->backing_object
) != NULL
) {
2389 if (bbobj
->type
== OBJT_VNODE
)
2390 vm_object_hold_shared(bbobj
);
2392 vm_object_hold(bbobj
);
2393 if (bbobj
== backing_object
->backing_object
)
2395 vm_object_drop(bbobj
);
2399 * We are removing backing_object from bbobj's
2400 * shadow list and adding object to bbobj's shadow
2401 * list, so the ref_count on bbobj is unchanged.
2404 if (backing_object
->flags
& OBJ_ONSHADOW
) {
2405 /* not locked exclusively if vnode */
2406 KKASSERT(bbobj
->type
!= OBJT_VNODE
);
2407 LIST_REMOVE(backing_object
,
2409 bbobj
->shadow_count
--;
2410 bbobj
->generation
++;
2411 vm_object_clear_flag(backing_object
,
2414 backing_object
->backing_object
= NULL
;
2416 object
->backing_object
= bbobj
;
2418 if (bbobj
->type
!= OBJT_VNODE
) {
2419 LIST_INSERT_HEAD(&bbobj
->shadow_head
,
2420 object
, shadow_list
);
2421 bbobj
->shadow_count
++;
2422 bbobj
->generation
++;
2423 vm_object_set_flag(object
,
2428 object
->backing_object_offset
+=
2429 backing_object
->backing_object_offset
;
2431 vm_object_drop(bbobj
);
2434 * Discard the old backing_object. Nothing should be
2435 * able to ref it, other than a vm_map_split(),
2436 * and vm_map_split() will stall on our chain lock.
2437 * And we control the parent so it shouldn't be
2438 * possible for it to go away either.
2440 * Since the backing object has no pages, no pager
2441 * left, and no object references within it, all
2442 * that is necessary is to dispose of it.
2444 KASSERT(backing_object
->ref_count
== 1,
2445 ("backing_object %p was somehow "
2446 "re-referenced during collapse!",
2448 KASSERT(RB_EMPTY(&backing_object
->rb_memq
),
2449 ("backing_object %p somehow has left "
2450 "over pages during collapse!",
2454 * The object can be destroyed.
2456 * XXX just fall through and dodealloc instead
2457 * of forcing destruction?
2459 atomic_add_int(&backing_object
->ref_count
, -1);
2460 #if defined(DEBUG_LOCKS)
2461 debugvm_object_add(backing_object
, "collapse", 1, -1);
2463 if ((backing_object
->flags
& OBJ_DEAD
) == 0)
2464 vm_object_terminate(backing_object
);
2469 * If we do not entirely shadow the backing object,
2470 * there is nothing we can do so we give up.
2472 if (vm_object_backing_scan(object
, backing_object
,
2473 OBSC_TEST_ALL_SHADOWED
) == 0) {
2478 * bbobj is backing_object->backing_object. Since
2479 * object completely shadows backing_object we can
2480 * bypass it and become backed by bbobj instead.
2482 * The shadow list for vnode backing objects is not
2483 * used and a shared hold is allowed.
2485 while ((bbobj
= backing_object
->backing_object
) != NULL
) {
2486 if (bbobj
->type
== OBJT_VNODE
)
2487 vm_object_hold_shared(bbobj
);
2489 vm_object_hold(bbobj
);
2490 if (bbobj
== backing_object
->backing_object
)
2492 vm_object_drop(bbobj
);
2496 * Make object shadow bbobj instead of backing_object.
2497 * Remove object from backing_object's shadow list.
2499 * Deallocating backing_object will not remove
2500 * it, since its reference count is at least 2.
2502 * Removing object from backing_object's shadow
2503 * list requires releasing a ref, which we do
2504 * below by setting dodealloc to 1.
2506 KKASSERT(object
->backing_object
== backing_object
);
2507 if (object
->flags
& OBJ_ONSHADOW
) {
2508 LIST_REMOVE(object
, shadow_list
);
2509 backing_object
->shadow_count
--;
2510 backing_object
->generation
++;
2511 vm_object_clear_flag(object
, OBJ_ONSHADOW
);
2515 * Add a ref to bbobj, bbobj now shadows object.
2517 * NOTE: backing_object->backing_object still points
2518 * to bbobj. That relationship remains intact
2519 * because backing_object has > 1 ref, so
2520 * someone else is pointing to it (hence why
2521 * we can't collapse it into object and can
2522 * only handle the all-shadowed bypass case).
2525 if (bbobj
->type
!= OBJT_VNODE
) {
2526 vm_object_chain_wait(bbobj
, 0);
2527 vm_object_reference_locked(bbobj
);
2528 LIST_INSERT_HEAD(&bbobj
->shadow_head
,
2529 object
, shadow_list
);
2530 bbobj
->shadow_count
++;
2531 bbobj
->generation
++;
2532 vm_object_set_flag(object
,
2535 vm_object_reference_quick(bbobj
);
2537 object
->backing_object_offset
+=
2538 backing_object
->backing_object_offset
;
2539 object
->backing_object
= bbobj
;
2540 vm_object_drop(bbobj
);
2542 object
->backing_object
= NULL
;
2546 * Drop the reference count on backing_object. To
2547 * handle ref_count races properly we can't assume
2548 * that the ref_count is still at least 2 so we
2549 * have to actually call vm_object_deallocate()
2550 * (after clearing the chainlock).
2557 * Ok, we want to loop on the new object->bbobj association,
2558 * possibly collapsing it further. However if dodealloc is
2559 * non-zero we have to deallocate the backing_object which
2560 * itself can potentially undergo a collapse, creating a
2561 * recursion depth issue with the LWKT token subsystem.
2563 * In the case where we must deallocate the backing_object
2564 * it is possible now that the backing_object has a single
2565 * shadow count on some other object (not represented here
2566 * as yet), since it no longer shadows us. Thus when we
2567 * call vm_object_deallocate() it may attempt to collapse
2568 * itself into its remaining parent.
2571 struct vm_object_dealloc_list
*dtmp
;
2573 vm_object_chain_release(backing_object
);
2574 vm_object_unlock(backing_object
);
2575 /* backing_object remains held */
2578 * Auto-deallocation list for caller convenience.
2583 dtmp
= kmalloc(sizeof(*dtmp
), M_TEMP
, M_WAITOK
);
2584 dtmp
->object
= backing_object
;
2585 dtmp
->next
= *dlistp
;
2588 vm_object_chain_release(backing_object
);
2589 vm_object_drop(backing_object
);
2591 /* backing_object = NULL; not needed */
2596 * Clean up any left over backing_object
2598 if (backing_object
) {
2599 vm_object_chain_release(backing_object
);
2600 vm_object_drop(backing_object
);
2604 * Clean up any auto-deallocation list. This is a convenience
2605 * for top-level callers so they don't have to pass &dlist.
2606 * Do not clean up any caller-passed dlistp, the caller will
2610 vm_object_deallocate_list(&dlist
);
2615 * vm_object_collapse() may collect additional objects in need of
2616 * deallocation. This routine deallocates these objects. The
2617 * deallocation itself can trigger additional collapses (which the
2618 * deallocate function takes care of). This procedure is used to
2619 * reduce procedural recursion since these vm_object shadow chains
2620 * can become quite long.
2623 vm_object_deallocate_list(struct vm_object_dealloc_list
**dlistp
)
2625 struct vm_object_dealloc_list
*dlist
;
2627 while ((dlist
= *dlistp
) != NULL
) {
2628 *dlistp
= dlist
->next
;
2629 vm_object_lock(dlist
->object
);
2630 vm_object_deallocate_locked(dlist
->object
);
2631 vm_object_drop(dlist
->object
);
2632 kfree(dlist
, M_TEMP
);
2637 * Removes all physical pages in the specified object range from the
2638 * object's list of pages.
2642 static int vm_object_page_remove_callback(vm_page_t p
, void *data
);
2645 vm_object_page_remove(vm_object_t object
, vm_pindex_t start
, vm_pindex_t end
,
2646 boolean_t clean_only
)
2648 struct rb_vm_page_scan_info info
;
2652 * Degenerate cases and assertions
2654 vm_object_hold(object
);
2655 if (object
== NULL
||
2656 (object
->resident_page_count
== 0 && object
->swblock_count
== 0)) {
2657 vm_object_drop(object
);
2660 KASSERT(object
->type
!= OBJT_PHYS
,
2661 ("attempt to remove pages from a physical object"));
2664 * Indicate that paging is occuring on the object
2666 vm_object_pip_add(object
, 1);
2669 * Figure out the actual removal range and whether we are removing
2670 * the entire contents of the object or not. If removing the entire
2671 * contents, be sure to get all pages, even those that might be
2672 * beyond the end of the object.
2674 info
.object
= object
;
2675 info
.start_pindex
= start
;
2677 info
.end_pindex
= (vm_pindex_t
)-1;
2679 info
.end_pindex
= end
- 1;
2680 info
.limit
= clean_only
;
2681 all
= (start
== 0 && info
.end_pindex
>= object
->size
- 1);
2684 * Loop until we are sure we have gotten them all.
2688 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, rb_vm_page_scancmp
,
2689 vm_object_page_remove_callback
, &info
);
2690 } while (info
.error
);
2693 * Remove any related swap if throwing away pages, or for
2694 * non-swap objects (the swap is a clean copy in that case).
2696 if (object
->type
!= OBJT_SWAP
|| clean_only
== FALSE
) {
2698 swap_pager_freespace_all(object
);
2700 swap_pager_freespace(object
, info
.start_pindex
,
2701 info
.end_pindex
- info
.start_pindex
+ 1);
2707 vm_object_pip_wakeup(object
);
2708 vm_object_drop(object
);
2712 * The caller must hold the object
2715 vm_object_page_remove_callback(vm_page_t p
, void *data
)
2717 struct rb_vm_page_scan_info
*info
= data
;
2719 if ((++info
->count
& 63) == 0)
2722 if (info
->object
!= p
->object
||
2723 p
->pindex
< info
->start_pindex
||
2724 p
->pindex
> info
->end_pindex
) {
2725 kprintf("vm_object_page_remove_callbackA: obj/pg race %p/%p\n",
2729 if (vm_page_busy_try(p
, TRUE
)) {
2730 vm_page_sleep_busy(p
, TRUE
, "vmopar");
2734 if (info
->object
!= p
->object
) {
2735 /* this should never happen */
2736 kprintf("vm_object_page_remove_callbackB: obj/pg race %p/%p\n",
2743 * Wired pages cannot be destroyed, but they can be invalidated
2744 * and we do so if clean_only (limit) is not set.
2746 * WARNING! The page may be wired due to being part of a buffer
2747 * cache buffer, and the buffer might be marked B_CACHE.
2748 * This is fine as part of a truncation but VFSs must be
2749 * sure to fix the buffer up when re-extending the file.
2751 * NOTE! PG_NEED_COMMIT is ignored.
2753 if (p
->wire_count
!= 0) {
2754 vm_page_protect(p
, VM_PROT_NONE
);
2755 if (info
->limit
== 0)
2762 * limit is our clean_only flag. If set and the page is dirty or
2763 * requires a commit, do not free it. If set and the page is being
2764 * held by someone, do not free it.
2766 if (info
->limit
&& p
->valid
) {
2767 vm_page_test_dirty(p
);
2768 if ((p
->valid
& p
->dirty
) || (p
->flags
& PG_NEED_COMMIT
)) {
2777 vm_page_protect(p
, VM_PROT_NONE
);
2784 * Coalesces two objects backing up adjoining regions of memory into a
2787 * returns TRUE if objects were combined.
2789 * NOTE: Only works at the moment if the second object is NULL -
2790 * if it's not, which object do we lock first?
2793 * prev_object First object to coalesce
2794 * prev_offset Offset into prev_object
2795 * next_object Second object into coalesce
2796 * next_offset Offset into next_object
2798 * prev_size Size of reference to prev_object
2799 * next_size Size of reference to next_object
2801 * The caller does not need to hold (prev_object) but must have a stable
2802 * pointer to it (typically by holding the vm_map locked).
2805 vm_object_coalesce(vm_object_t prev_object
, vm_pindex_t prev_pindex
,
2806 vm_size_t prev_size
, vm_size_t next_size
)
2808 vm_pindex_t next_pindex
;
2810 if (prev_object
== NULL
)
2813 vm_object_hold(prev_object
);
2815 if (prev_object
->type
!= OBJT_DEFAULT
&&
2816 prev_object
->type
!= OBJT_SWAP
) {
2817 vm_object_drop(prev_object
);
2822 * Try to collapse the object first
2824 vm_object_chain_acquire(prev_object
, 0);
2825 vm_object_collapse(prev_object
, NULL
);
2828 * Can't coalesce if: . more than one reference . paged out . shadows
2829 * another object . has a copy elsewhere (any of which mean that the
2830 * pages not mapped to prev_entry may be in use anyway)
2833 if (prev_object
->backing_object
!= NULL
) {
2834 vm_object_chain_release(prev_object
);
2835 vm_object_drop(prev_object
);
2839 prev_size
>>= PAGE_SHIFT
;
2840 next_size
>>= PAGE_SHIFT
;
2841 next_pindex
= prev_pindex
+ prev_size
;
2843 if ((prev_object
->ref_count
> 1) &&
2844 (prev_object
->size
!= next_pindex
)) {
2845 vm_object_chain_release(prev_object
);
2846 vm_object_drop(prev_object
);
2851 * Remove any pages that may still be in the object from a previous
2854 if (next_pindex
< prev_object
->size
) {
2855 vm_object_page_remove(prev_object
,
2857 next_pindex
+ next_size
, FALSE
);
2858 if (prev_object
->type
== OBJT_SWAP
)
2859 swap_pager_freespace(prev_object
,
2860 next_pindex
, next_size
);
2864 * Extend the object if necessary.
2866 if (next_pindex
+ next_size
> prev_object
->size
)
2867 prev_object
->size
= next_pindex
+ next_size
;
2869 vm_object_chain_release(prev_object
);
2870 vm_object_drop(prev_object
);
2875 * Make the object writable and flag is being possibly dirty.
2877 * The object might not be held (or might be held but held shared),
2878 * the related vnode is probably not held either. Object and vnode are
2879 * stable by virtue of the vm_page busied by the caller preventing
2882 * If the related mount is flagged MNTK_THR_SYNC we need to call
2883 * vsetobjdirty(). Filesystems using this option usually shortcut
2884 * synchronization by only scanning the syncer list.
2887 vm_object_set_writeable_dirty(vm_object_t object
)
2891 /*vm_object_assert_held(object);*/
2893 * Avoid contention in vm fault path by checking the state before
2894 * issuing an atomic op on it.
2896 if ((object
->flags
& (OBJ_WRITEABLE
|OBJ_MIGHTBEDIRTY
)) !=
2897 (OBJ_WRITEABLE
|OBJ_MIGHTBEDIRTY
)) {
2898 vm_object_set_flag(object
, OBJ_WRITEABLE
|OBJ_MIGHTBEDIRTY
);
2900 if (object
->type
== OBJT_VNODE
&&
2901 (vp
= (struct vnode
*)object
->handle
) != NULL
) {
2902 if ((vp
->v_flag
& VOBJDIRTY
) == 0) {
2904 (vp
->v_mount
->mnt_kern_flag
& MNTK_THR_SYNC
)) {
2906 * New style THR_SYNC places vnodes on the
2907 * syncer list more deterministically.
2912 * Old style scan would not necessarily place
2913 * a vnode on the syncer list when possibly
2914 * modified via mmap.
2916 vsetflags(vp
, VOBJDIRTY
);
2922 #include "opt_ddb.h"
2924 #include <sys/kernel.h>
2926 #include <sys/cons.h>
2928 #include <ddb/ddb.h>
2930 static int _vm_object_in_map (vm_map_t map
, vm_object_t object
,
2931 vm_map_entry_t entry
);
2932 static int vm_object_in_map (vm_object_t object
);
2935 * The caller must hold the object.
2938 _vm_object_in_map(vm_map_t map
, vm_object_t object
, vm_map_entry_t entry
)
2941 vm_map_entry_t tmpe
;
2942 vm_object_t obj
, nobj
;
2948 tmpe
= map
->header
.next
;
2949 entcount
= map
->nentries
;
2950 while (entcount
-- && (tmpe
!= &map
->header
)) {
2951 if( _vm_object_in_map(map
, object
, tmpe
)) {
2958 switch(entry
->maptype
) {
2959 case VM_MAPTYPE_SUBMAP
:
2960 tmpm
= entry
->object
.sub_map
;
2961 tmpe
= tmpm
->header
.next
;
2962 entcount
= tmpm
->nentries
;
2963 while (entcount
-- && tmpe
!= &tmpm
->header
) {
2964 if( _vm_object_in_map(tmpm
, object
, tmpe
)) {
2970 case VM_MAPTYPE_NORMAL
:
2971 case VM_MAPTYPE_VPAGETABLE
:
2972 obj
= entry
->object
.vm_object
;
2974 if (obj
== object
) {
2975 if (obj
!= entry
->object
.vm_object
)
2976 vm_object_drop(obj
);
2979 while ((nobj
= obj
->backing_object
) != NULL
) {
2980 vm_object_hold(nobj
);
2981 if (nobj
== obj
->backing_object
)
2983 vm_object_drop(nobj
);
2985 if (obj
!= entry
->object
.vm_object
) {
2987 vm_object_lock_swap();
2988 vm_object_drop(obj
);
2999 static int vm_object_in_map_callback(struct proc
*p
, void *data
);
3001 struct vm_object_in_map_info
{
3010 vm_object_in_map(vm_object_t object
)
3012 struct vm_object_in_map_info info
;
3015 info
.object
= object
;
3017 allproc_scan(vm_object_in_map_callback
, &info
);
3020 if( _vm_object_in_map(&kernel_map
, object
, 0))
3022 if( _vm_object_in_map(&pager_map
, object
, 0))
3024 if( _vm_object_in_map(&buffer_map
, object
, 0))
3033 vm_object_in_map_callback(struct proc
*p
, void *data
)
3035 struct vm_object_in_map_info
*info
= data
;
3038 if (_vm_object_in_map(&p
->p_vmspace
->vm_map
, info
->object
, 0)) {
3046 DB_SHOW_COMMAND(vmochk
, vm_object_check
)
3052 * make sure that internal objs are in a map somewhere
3053 * and none have zero ref counts.
3055 for (n
= 0; n
< VMOBJ_HSIZE
; ++n
) {
3056 for (object
= TAILQ_FIRST(&vm_object_lists
[n
]);
3058 object
= TAILQ_NEXT(object
, object_list
)) {
3059 if (object
->type
== OBJT_MARKER
)
3061 if (object
->handle
!= NULL
||
3062 (object
->type
!= OBJT_DEFAULT
&&
3063 object
->type
!= OBJT_SWAP
)) {
3066 if (object
->ref_count
== 0) {
3067 db_printf("vmochk: internal obj has "
3068 "zero ref count: %ld\n",
3069 (long)object
->size
);
3071 if (vm_object_in_map(object
))
3073 db_printf("vmochk: internal obj is not in a map: "
3074 "ref: %d, size: %lu: 0x%lx, "
3075 "backing_object: %p\n",
3076 object
->ref_count
, (u_long
)object
->size
,
3077 (u_long
)object
->size
,
3078 (void *)object
->backing_object
);
3086 DB_SHOW_COMMAND(object
, vm_object_print_static
)
3088 /* XXX convert args. */
3089 vm_object_t object
= (vm_object_t
)addr
;
3090 boolean_t full
= have_addr
;
3094 /* XXX count is an (unused) arg. Avoid shadowing it. */
3095 #define count was_count
3103 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
3104 object
, (int)object
->type
, (u_long
)object
->size
,
3105 object
->resident_page_count
, object
->ref_count
, object
->flags
);
3107 * XXX no %qd in kernel. Truncate object->backing_object_offset.
3109 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
3110 object
->shadow_count
,
3111 object
->backing_object
? object
->backing_object
->ref_count
: 0,
3112 object
->backing_object
, (long)object
->backing_object_offset
);
3119 RB_FOREACH(p
, vm_page_rb_tree
, &object
->rb_memq
) {
3121 db_iprintf("memory:=");
3122 else if (count
== 6) {
3130 db_printf("(off=0x%lx,page=0x%lx)",
3131 (u_long
) p
->pindex
, (u_long
) VM_PAGE_TO_PHYS(p
));
3142 * XXX need this non-static entry for calling from vm_map_print.
3147 vm_object_print(/* db_expr_t */ long addr
,
3148 boolean_t have_addr
,
3149 /* db_expr_t */ long count
,
3152 vm_object_print_static(addr
, have_addr
, count
, modif
);
3158 DB_SHOW_COMMAND(vmopag
, vm_object_print_pages
)
3165 for (n
= 0; n
< VMOBJ_HSIZE
; ++n
) {
3166 for (object
= TAILQ_FIRST(&vm_object_lists
[n
]);
3168 object
= TAILQ_NEXT(object
, object_list
)) {
3169 vm_pindex_t idx
, fidx
;
3171 vm_paddr_t pa
= -1, padiff
;
3175 if (object
->type
== OBJT_MARKER
)
3177 db_printf("new object: %p\n", (void *)object
);
3187 osize
= object
->size
;
3190 for (idx
= 0; idx
< osize
; idx
++) {
3191 m
= vm_page_lookup(object
, idx
);
3194 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
3195 (long)fidx
, rcount
, (long)pa
);
3209 (VM_PAGE_TO_PHYS(m
) == pa
+ rcount
* PAGE_SIZE
)) {
3214 padiff
= pa
+ rcount
* PAGE_SIZE
- VM_PAGE_TO_PHYS(m
);
3215 padiff
>>= PAGE_SHIFT
;
3216 padiff
&= PQ_L2_MASK
;
3218 pa
= VM_PAGE_TO_PHYS(m
) - rcount
* PAGE_SIZE
;
3222 db_printf(" index(%ld)run(%d)pa(0x%lx)",
3223 (long)fidx
, rcount
, (long)pa
);
3224 db_printf("pd(%ld)\n", (long)padiff
);
3234 pa
= VM_PAGE_TO_PHYS(m
);
3238 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
3239 (long)fidx
, rcount
, (long)pa
);