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 object_collapses;
134 static long object_bypasses;
136 struct vm_object_hash vm_object_hash[VMOBJ_HSIZE];
138 MALLOC_DEFINE(M_VM_OBJECT, "vm_object", "vm_object structures");
140 #if defined(DEBUG_LOCKS)
142 #define vm_object_vndeallocate(obj, vpp) \
143 debugvm_object_vndeallocate(obj, vpp, __FILE__, __LINE__)
146 * Debug helper to track hold/drop/ref/deallocate calls.
149 debugvm_object_add(vm_object_t obj, char *file, int line, int addrem)
153 i = atomic_fetchadd_int(&obj->debug_index, 1);
154 i = i & (VMOBJ_DEBUG_ARRAY_SIZE - 1);
155 ksnprintf(obj->debug_hold_thrs[i],
156 sizeof(obj->debug_hold_thrs[i]),
158 (addrem == -1 ? '-' : (addrem == 1 ? '+' : '=')),
159 (curthread->td_proc ? curthread->td_proc->p_pid : -1),
162 obj->debug_hold_file[i] = file;
163 obj->debug_hold_line[i] = line;
165 /* Uncomment for debugging obj refs/derefs in reproducable cases */
166 if (strcmp(curthread->td_comm, "sshd") == 0) {
167 kprintf("%d %p refs=%d ar=%d file: %s/%d\n",
168 (curthread->td_proc ? curthread->td_proc->p_pid : -1),
169 obj, obj->ref_count, addrem, file, line);
177 * Misc low level routines
180 vm_object_lock_init(vm_object_t obj)
182 #if defined(DEBUG_LOCKS)
185 obj->debug_index = 0;
186 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) {
187 obj->debug_hold_thrs[i][0] = 0;
188 obj->debug_hold_file[i] = NULL;
189 obj->debug_hold_line[i] = 0;
195 vm_object_lock_swap(void)
201 vm_object_lock(vm_object_t obj)
203 lwkt_gettoken(&obj->token);
207 * Returns TRUE on sucesss
210 vm_object_lock_try(vm_object_t obj)
212 return(lwkt_trytoken(&obj->token));
216 vm_object_lock_shared(vm_object_t obj)
218 lwkt_gettoken_shared(&obj->token);
222 vm_object_unlock(vm_object_t obj)
224 lwkt_reltoken(&obj->token);
228 vm_object_upgrade(vm_object_t obj)
230 lwkt_reltoken(&obj->token);
231 lwkt_gettoken(&obj->token);
235 vm_object_downgrade(vm_object_t obj)
237 lwkt_reltoken(&obj->token);
238 lwkt_gettoken_shared(&obj->token);
242 vm_object_assert_held(vm_object_t obj)
244 ASSERT_LWKT_TOKEN_HELD(&obj->token);
250 globaldata_t gd = mycpu;
253 pg_color = (int)(intptr_t)gd->gd_curthread >> 10;
254 pg_color += gd->gd_quick_color;
255 gd->gd_quick_color += PQ_PRIME2;
261 VMOBJDEBUG(vm_object_hold)(vm_object_t obj VMOBJDBARGS)
263 KKASSERT(obj != NULL);
266 * Object must be held (object allocation is stable due to callers
267 * context, typically already holding the token on a parent object)
268 * prior to potentially blocking on the lock, otherwise the object
269 * can get ripped away from us.
271 refcount_acquire(&obj->hold_count);
274 #if defined(DEBUG_LOCKS)
275 debugvm_object_add(obj, file, line, 1);
280 VMOBJDEBUG(vm_object_hold_try)(vm_object_t obj VMOBJDBARGS)
282 KKASSERT(obj != NULL);
285 * Object must be held (object allocation is stable due to callers
286 * context, typically already holding the token on a parent object)
287 * prior to potentially blocking on the lock, otherwise the object
288 * can get ripped away from us.
290 refcount_acquire(&obj->hold_count);
291 if (vm_object_lock_try(obj) == 0) {
292 if (refcount_release(&obj->hold_count)) {
293 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD))
294 kfree(obj, M_VM_OBJECT);
299 #if defined(DEBUG_LOCKS)
300 debugvm_object_add(obj, file, line, 1);
306 VMOBJDEBUG(vm_object_hold_shared)(vm_object_t obj VMOBJDBARGS)
308 KKASSERT(obj != NULL);
311 * Object must be held (object allocation is stable due to callers
312 * context, typically already holding the token on a parent object)
313 * prior to potentially blocking on the lock, otherwise the object
314 * can get ripped away from us.
316 refcount_acquire(&obj->hold_count);
317 vm_object_lock_shared(obj);
319 #if defined(DEBUG_LOCKS)
320 debugvm_object_add(obj, file, line, 1);
325 * Drop the token and hold_count on the object.
327 * WARNING! Token might be shared.
330 VMOBJDEBUG(vm_object_drop)(vm_object_t obj VMOBJDBARGS)
336 * No new holders should be possible once we drop hold_count 1->0 as
337 * there is no longer any way to reference the object.
339 KKASSERT(obj->hold_count > 0);
340 if (refcount_release(&obj->hold_count)) {
341 #if defined(DEBUG_LOCKS)
342 debugvm_object_add(obj, file, line, -1);
345 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD)) {
346 vm_object_unlock(obj);
347 kfree(obj, M_VM_OBJECT);
349 vm_object_unlock(obj);
352 #if defined(DEBUG_LOCKS)
353 debugvm_object_add(obj, file, line, -1);
355 vm_object_unlock(obj);
360 * Initialize a freshly allocated object, returning a held object.
362 * Used only by vm_object_allocate(), zinitna() and vm_object_init().
367 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
369 struct vm_object_hash *hash;
371 RB_INIT(&object->rb_memq);
372 LIST_INIT(&object->shadow_head);
373 lwkt_token_init(&object->token, "vmobj");
377 object->ref_count = 1;
378 object->memattr = VM_MEMATTR_DEFAULT;
379 object->hold_count = 0;
381 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
382 vm_object_set_flag(object, OBJ_ONEMAPPING);
383 object->paging_in_progress = 0;
384 object->resident_page_count = 0;
385 object->agg_pv_list_count = 0;
386 object->shadow_count = 0;
387 /* cpu localization twist */
388 object->pg_color = vm_quickcolor();
389 object->handle = NULL;
390 object->backing_object = NULL;
391 object->backing_object_offset = (vm_ooffset_t)0;
393 object->generation++;
394 object->swblock_count = 0;
395 RB_INIT(&object->swblock_root);
396 vm_object_lock_init(object);
397 pmap_object_init(object);
399 vm_object_hold(object);
401 hash = VMOBJ_HASH(object);
402 lwkt_gettoken(&hash->token);
403 TAILQ_INSERT_TAIL(&hash->list, object, object_list);
404 lwkt_reltoken(&hash->token);
408 * Initialize a VM object.
411 vm_object_init(vm_object_t object, vm_pindex_t size)
413 _vm_object_allocate(OBJT_DEFAULT, size, object);
414 vm_object_drop(object);
418 * Initialize the VM objects module.
420 * Called from the low level boot code only. Note that this occurs before
421 * kmalloc is initialized so we cannot allocate any VM objects.
424 vm_object_init1(void)
428 for (i = 0; i < VMOBJ_HSIZE; ++i) {
429 TAILQ_INIT(&vm_object_hash[i].list);
430 lwkt_token_init(&vm_object_hash[i].token, "vmobjlst");
433 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd),
435 vm_object_drop(&kernel_object);
439 vm_object_init2(void)
441 kmalloc_set_unlimited(M_VM_OBJECT);
445 * Allocate and return a new object of the specified type and size.
450 vm_object_allocate(objtype_t type, vm_pindex_t size)
454 obj = kmalloc(sizeof(*obj), M_VM_OBJECT, M_INTWAIT|M_ZERO);
455 _vm_object_allocate(type, size, obj);
462 * This version returns a held object, allowing further atomic initialization
466 vm_object_allocate_hold(objtype_t type, vm_pindex_t size)
470 obj = kmalloc(sizeof(*obj), M_VM_OBJECT, M_INTWAIT|M_ZERO);
471 _vm_object_allocate(type, size, obj);
477 * Add an additional reference to a vm_object. The object must already be
478 * held. The original non-lock version is no longer supported. The object
479 * must NOT be chain locked by anyone at the time the reference is added.
481 * Referencing a chain-locked object can blow up the fairly sensitive
482 * ref_count and shadow_count tests in the deallocator. Most callers
483 * will call vm_object_chain_wait() prior to calling
484 * vm_object_reference_locked() to avoid the case.
486 * The object must be held, but may be held shared if desired (hence why
487 * we use an atomic op).
490 VMOBJDEBUG(vm_object_reference_locked)(vm_object_t object VMOBJDBARGS)
492 KKASSERT(object != NULL);
493 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
494 KKASSERT((object->chainlk & (CHAINLK_EXCL | CHAINLK_MASK)) == 0);
495 atomic_add_int(&object->ref_count, 1);
496 if (object->type == OBJT_VNODE) {
497 vref(object->handle);
498 /* XXX what if the vnode is being destroyed? */
500 #if defined(DEBUG_LOCKS)
501 debugvm_object_add(object, file, line, 1);
506 * This version is only allowed for vnode objects.
509 VMOBJDEBUG(vm_object_reference_quick)(vm_object_t object VMOBJDBARGS)
511 KKASSERT(object->type == OBJT_VNODE);
512 atomic_add_int(&object->ref_count, 1);
513 vref(object->handle);
514 #if defined(DEBUG_LOCKS)
515 debugvm_object_add(object, file, line, 1);
520 * Object OBJ_CHAINLOCK lock handling.
522 * The caller can chain-lock backing objects recursively and then
523 * use vm_object_chain_release_all() to undo the whole chain.
525 * Chain locks are used to prevent collapses and are only applicable
526 * to OBJT_DEFAULT and OBJT_SWAP objects. Chain locking operations
527 * on other object types are ignored. This is also important because
528 * it allows e.g. the vnode underlying a memory mapping to take concurrent
531 * The object must usually be held on entry, though intermediate
532 * objects need not be held on release. The object must be held exclusively,
533 * NOT shared. Note that the prefault path checks the shared state and
534 * avoids using the chain functions.
537 vm_object_chain_wait(vm_object_t object, int shared)
539 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
541 uint32_t chainlk = object->chainlk;
545 if (chainlk & (CHAINLK_EXCL | CHAINLK_EXCLREQ)) {
546 tsleep_interlock(object, 0);
547 if (atomic_cmpset_int(&object->chainlk,
549 chainlk | CHAINLK_WAIT)) {
550 tsleep(object, PINTERLOCKED,
559 if (chainlk & (CHAINLK_MASK | CHAINLK_EXCL)) {
560 tsleep_interlock(object, 0);
561 if (atomic_cmpset_int(&object->chainlk,
563 chainlk | CHAINLK_WAIT))
565 tsleep(object, PINTERLOCKED,
570 if (atomic_cmpset_int(&object->chainlk,
572 chainlk & ~CHAINLK_WAIT))
574 if (chainlk & CHAINLK_WAIT)
586 vm_object_chain_acquire(vm_object_t object, int shared)
588 if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP)
590 if (vm_shared_fault == 0)
594 uint32_t chainlk = object->chainlk;
598 if (chainlk & (CHAINLK_EXCL | CHAINLK_EXCLREQ)) {
599 tsleep_interlock(object, 0);
600 if (atomic_cmpset_int(&object->chainlk,
602 chainlk | CHAINLK_WAIT)) {
603 tsleep(object, PINTERLOCKED,
607 } else if (atomic_cmpset_int(&object->chainlk,
608 chainlk, chainlk + 1)) {
613 if (chainlk & (CHAINLK_MASK | CHAINLK_EXCL)) {
614 tsleep_interlock(object, 0);
615 if (atomic_cmpset_int(&object->chainlk,
620 tsleep(object, PINTERLOCKED,
625 if (atomic_cmpset_int(&object->chainlk,
627 (chainlk | CHAINLK_EXCL) &
630 if (chainlk & CHAINLK_WAIT)
642 vm_object_chain_release(vm_object_t object)
644 /*ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));*/
645 if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP)
647 KKASSERT(object->chainlk & (CHAINLK_MASK | CHAINLK_EXCL));
649 uint32_t chainlk = object->chainlk;
652 if (chainlk & CHAINLK_MASK) {
653 if ((chainlk & CHAINLK_MASK) == 1 &&
654 atomic_cmpset_int(&object->chainlk,
656 (chainlk - 1) & ~CHAINLK_WAIT)) {
657 if (chainlk & CHAINLK_WAIT)
661 if ((chainlk & CHAINLK_MASK) > 1 &&
662 atomic_cmpset_int(&object->chainlk,
663 chainlk, chainlk - 1)) {
668 KKASSERT(chainlk & CHAINLK_EXCL);
669 if (atomic_cmpset_int(&object->chainlk,
671 chainlk & ~(CHAINLK_EXCL |
673 if (chainlk & CHAINLK_WAIT)
682 * Release the chain from first_object through and including stopobj.
683 * The caller is typically holding the first and last object locked
684 * (shared or exclusive) to prevent destruction races.
686 * We release stopobj first as an optimization as this object is most
687 * likely to be shared across multiple processes.
690 vm_object_chain_release_all(vm_object_t first_object, vm_object_t stopobj)
692 vm_object_t backing_object;
695 vm_object_chain_release(stopobj);
696 object = first_object;
698 while (object != stopobj) {
700 backing_object = object->backing_object;
701 vm_object_chain_release(object);
702 object = backing_object;
707 * Dereference an object and its underlying vnode. The object may be
708 * held shared. On return the object will remain held.
710 * This function may return a vnode in *vpp which the caller must release
711 * after the caller drops its own lock. If vpp is NULL, we assume that
712 * the caller was holding an exclusive lock on the object and we vrele()
716 VMOBJDEBUG(vm_object_vndeallocate)(vm_object_t object, struct vnode **vpp
719 struct vnode *vp = (struct vnode *) object->handle;
721 KASSERT(object->type == OBJT_VNODE,
722 ("vm_object_vndeallocate: not a vnode object"));
723 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
724 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
726 if (object->ref_count == 0) {
727 vprint("vm_object_vndeallocate", vp);
728 panic("vm_object_vndeallocate: bad object reference count");
732 int count = object->ref_count;
735 vm_object_upgrade(object);
736 if (atomic_cmpset_int(&object->ref_count, count, 0)) {
737 vclrflags(vp, VTEXT);
741 if (atomic_cmpset_int(&object->ref_count,
748 #if defined(DEBUG_LOCKS)
749 debugvm_object_add(object, file, line, -1);
753 * vrele or return the vp to vrele. We can only safely vrele(vp)
754 * if the object was locked exclusively. But there are two races
757 * We had to upgrade the object above to safely clear VTEXT
758 * but the alternative path where the shared lock is retained
759 * can STILL race to 0 in other paths and cause our own vrele()
760 * to terminate the vnode. We can't allow that if the VM object
761 * is still locked shared.
770 * Release a reference to the specified object, gained either through a
771 * vm_object_allocate or a vm_object_reference call. When all references
772 * are gone, storage associated with this object may be relinquished.
774 * The caller does not have to hold the object locked but must have control
775 * over the reference in question in order to guarantee that the object
776 * does not get ripped out from under us.
778 * XXX Currently all deallocations require an exclusive lock.
781 VMOBJDEBUG(vm_object_deallocate)(vm_object_t object VMOBJDBARGS)
790 count = object->ref_count;
794 * If decrementing the count enters into special handling
795 * territory (0, 1, or 2) we have to do it the hard way.
796 * Fortunate though, objects with only a few refs like this
797 * are not likely to be heavily contended anyway.
799 * For vnode objects we only care about 1->0 transitions.
801 if (count <= 3 || (object->type == OBJT_VNODE && count <= 1)) {
802 #if defined(DEBUG_LOCKS)
803 debugvm_object_add(object, file, line, 0);
805 vm_object_hold(object);
806 vm_object_deallocate_locked(object);
807 vm_object_drop(object);
812 * Try to decrement ref_count without acquiring a hold on
813 * the object. This is particularly important for the exec*()
814 * and exit*() code paths because the program binary may
815 * have a great deal of sharing and an exclusive lock will
816 * crowbar performance in those circumstances.
818 if (object->type == OBJT_VNODE) {
819 vp = (struct vnode *)object->handle;
820 if (atomic_cmpset_int(&object->ref_count,
822 #if defined(DEBUG_LOCKS)
823 debugvm_object_add(object, file, line, -1);
831 if (atomic_cmpset_int(&object->ref_count,
833 #if defined(DEBUG_LOCKS)
834 debugvm_object_add(object, file, line, -1);
845 VMOBJDEBUG(vm_object_deallocate_locked)(vm_object_t object VMOBJDBARGS)
847 struct vm_object_dealloc_list *dlist = NULL;
848 struct vm_object_dealloc_list *dtmp;
853 * We may chain deallocate object, but additional objects may
854 * collect on the dlist which also have to be deallocated. We
855 * must avoid a recursion, vm_object chains can get deep.
859 while (object != NULL) {
861 * vnode case, caller either locked the object exclusively
862 * or this is a recursion with must_drop != 0 and the vnode
863 * object will be locked shared.
865 * If locked shared we have to drop the object before we can
866 * call vrele() or risk a shared/exclusive livelock.
868 if (object->type == OBJT_VNODE) {
869 ASSERT_LWKT_TOKEN_HELD(&object->token);
871 struct vnode *tmp_vp;
873 vm_object_vndeallocate(object, &tmp_vp);
874 vm_object_drop(object);
879 vm_object_vndeallocate(object, NULL);
883 ASSERT_LWKT_TOKEN_HELD_EXCL(&object->token);
886 * Normal case (object is locked exclusively)
888 if (object->ref_count == 0) {
889 panic("vm_object_deallocate: object deallocated "
890 "too many times: %d", object->type);
892 if (object->ref_count > 2) {
893 atomic_add_int(&object->ref_count, -1);
894 #if defined(DEBUG_LOCKS)
895 debugvm_object_add(object, file, line, -1);
901 * Here on ref_count of one or two, which are special cases for
904 * Nominal ref_count > 1 case if the second ref is not from
907 * (ONEMAPPING only applies to DEFAULT AND SWAP objects)
909 if (object->ref_count == 2 && object->shadow_count == 0) {
910 if (object->type == OBJT_DEFAULT ||
911 object->type == OBJT_SWAP) {
912 vm_object_set_flag(object, OBJ_ONEMAPPING);
914 atomic_add_int(&object->ref_count, -1);
915 #if defined(DEBUG_LOCKS)
916 debugvm_object_add(object, file, line, -1);
922 * If the second ref is from a shadow we chain along it
923 * upwards if object's handle is exhausted.
925 * We have to decrement object->ref_count before potentially
926 * collapsing the first shadow object or the collapse code
927 * will not be able to handle the degenerate case to remove
928 * object. However, if we do it too early the object can
929 * get ripped out from under us.
931 if (object->ref_count == 2 && object->shadow_count == 1 &&
932 object->handle == NULL && (object->type == OBJT_DEFAULT ||
933 object->type == OBJT_SWAP)) {
934 temp = LIST_FIRST(&object->shadow_head);
935 KKASSERT(temp != NULL);
936 vm_object_hold(temp);
939 * Wait for any paging to complete so the collapse
940 * doesn't (or isn't likely to) qcollapse. pip
941 * waiting must occur before we acquire the
945 temp->paging_in_progress ||
946 object->paging_in_progress
948 vm_object_pip_wait(temp, "objde1");
949 vm_object_pip_wait(object, "objde2");
953 * If the parent is locked we have to give up, as
954 * otherwise we would be acquiring locks in the
955 * wrong order and potentially deadlock.
957 if (temp->chainlk & (CHAINLK_EXCL | CHAINLK_MASK)) {
958 vm_object_drop(temp);
961 vm_object_chain_acquire(temp, 0);
964 * Recheck/retry after the hold and the paging
965 * wait, both of which can block us.
967 if (object->ref_count != 2 ||
968 object->shadow_count != 1 ||
970 LIST_FIRST(&object->shadow_head) != temp ||
971 (object->type != OBJT_DEFAULT &&
972 object->type != OBJT_SWAP)) {
973 vm_object_chain_release(temp);
974 vm_object_drop(temp);
979 * We can safely drop object's ref_count now.
981 KKASSERT(object->ref_count == 2);
982 atomic_add_int(&object->ref_count, -1);
983 #if defined(DEBUG_LOCKS)
984 debugvm_object_add(object, file, line, -1);
988 * If our single parent is not collapseable just
989 * decrement ref_count (2->1) and stop.
991 if (temp->handle || (temp->type != OBJT_DEFAULT &&
992 temp->type != OBJT_SWAP)) {
993 vm_object_chain_release(temp);
994 vm_object_drop(temp);
999 * At this point we have already dropped object's
1000 * ref_count so it is possible for a race to
1001 * deallocate obj out from under us. Any collapse
1002 * will re-check the situation. We must not block
1003 * until we are able to collapse.
1005 * Bump temp's ref_count to avoid an unwanted
1006 * degenerate recursion (can't call
1007 * vm_object_reference_locked() because it asserts
1008 * that CHAINLOCK is not set).
1010 atomic_add_int(&temp->ref_count, 1);
1011 KKASSERT(temp->ref_count > 1);
1014 * Collapse temp, then deallocate the extra ref
1017 vm_object_collapse(temp, &dlist);
1018 vm_object_chain_release(temp);
1020 vm_object_lock_swap();
1021 vm_object_drop(object);
1029 * Drop the ref and handle termination on the 1->0 transition.
1030 * We may have blocked above so we have to recheck.
1033 KKASSERT(object->ref_count != 0);
1034 if (object->ref_count >= 2) {
1035 atomic_add_int(&object->ref_count, -1);
1036 #if defined(DEBUG_LOCKS)
1037 debugvm_object_add(object, file, line, -1);
1041 KKASSERT(object->ref_count == 1);
1044 * 1->0 transition. Chain through the backing_object.
1045 * Maintain the ref until we've located the backing object,
1048 while ((temp = object->backing_object) != NULL) {
1049 if (temp->type == OBJT_VNODE)
1050 vm_object_hold_shared(temp);
1052 vm_object_hold(temp);
1053 if (temp == object->backing_object)
1055 vm_object_drop(temp);
1059 * 1->0 transition verified, retry if ref_count is no longer
1060 * 1. Otherwise disconnect the backing_object (temp) and
1063 if (object->ref_count != 1) {
1064 vm_object_drop(temp);
1069 * It shouldn't be possible for the object to be chain locked
1070 * if we're removing the last ref on it.
1072 * Removing object from temp's shadow list requires dropping
1073 * temp, which we will do on loop.
1075 * NOTE! vnodes do not use the shadow list, but still have
1076 * the backing_object reference.
1078 KKASSERT((object->chainlk & (CHAINLK_EXCL|CHAINLK_MASK)) == 0);
1081 if (object->flags & OBJ_ONSHADOW) {
1082 LIST_REMOVE(object, shadow_list);
1083 temp->shadow_count--;
1085 vm_object_clear_flag(object, OBJ_ONSHADOW);
1087 object->backing_object = NULL;
1090 atomic_add_int(&object->ref_count, -1);
1091 if ((object->flags & OBJ_DEAD) == 0)
1092 vm_object_terminate(object);
1093 if (must_drop && temp)
1094 vm_object_lock_swap();
1096 vm_object_drop(object);
1101 if (must_drop && object)
1102 vm_object_drop(object);
1105 * Additional tail recursion on dlist. Avoid a recursion. Objects
1106 * on the dlist have a hold count but are not locked.
1108 if ((dtmp = dlist) != NULL) {
1110 object = dtmp->object;
1111 kfree(dtmp, M_TEMP);
1113 vm_object_lock(object); /* already held, add lock */
1114 must_drop = 1; /* and we're responsible for it */
1120 * Destroy the specified object, freeing up related resources.
1122 * The object must have zero references.
1124 * The object must held. The caller is responsible for dropping the object
1125 * after terminate returns. Terminate does NOT drop the object.
1127 static int vm_object_terminate_callback(vm_page_t p, void *data);
1130 vm_object_terminate(vm_object_t object)
1132 struct rb_vm_page_scan_info info;
1133 struct vm_object_hash *hash;
1136 * Make sure no one uses us. Once we set OBJ_DEAD we should be
1137 * able to safely block.
1139 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1140 KKASSERT((object->flags & OBJ_DEAD) == 0);
1141 vm_object_set_flag(object, OBJ_DEAD);
1144 * Wait for the pageout daemon to be done with the object
1146 vm_object_pip_wait(object, "objtrm1");
1148 KASSERT(!object->paging_in_progress,
1149 ("vm_object_terminate: pageout in progress"));
1152 * Clean and free the pages, as appropriate. All references to the
1153 * object are gone, so we don't need to lock it.
1155 if (object->type == OBJT_VNODE) {
1159 * Clean pages and flush buffers.
1161 * NOTE! TMPFS buffer flushes do not typically flush the
1162 * actual page to swap as this would be highly
1163 * inefficient, and normal filesystems usually wrap
1164 * page flushes with buffer cache buffers.
1166 * To deal with this we have to call vinvalbuf() both
1167 * before and after the vm_object_page_clean().
1169 vp = (struct vnode *) object->handle;
1170 vinvalbuf(vp, V_SAVE, 0, 0);
1171 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
1172 vinvalbuf(vp, V_SAVE, 0, 0);
1176 * Wait for any I/O to complete, after which there had better not
1177 * be any references left on the object.
1179 vm_object_pip_wait(object, "objtrm2");
1181 if (object->ref_count != 0) {
1182 panic("vm_object_terminate: object with references, "
1183 "ref_count=%d", object->ref_count);
1187 * Cleanup any shared pmaps associated with this object.
1189 pmap_object_free(object);
1192 * Now free any remaining pages. For internal objects, this also
1193 * removes them from paging queues. Don't free wired pages, just
1194 * remove them from the object.
1197 info.object = object;
1198 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1199 vm_object_terminate_callback, &info);
1202 * Let the pager know object is dead.
1204 vm_pager_deallocate(object);
1207 * Wait for the object hold count to hit 1, clean out pages as
1208 * we go. vmobj_token interlocks any race conditions that might
1209 * pick the object up from the vm_object_list after we have cleared
1213 if (RB_ROOT(&object->rb_memq) == NULL)
1215 kprintf("vm_object_terminate: Warning, object %p "
1216 "still has %ld pages\n",
1217 object, object->resident_page_count);
1218 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1219 vm_object_terminate_callback, &info);
1223 * There had better not be any pages left
1225 KKASSERT(object->resident_page_count == 0);
1228 * Remove the object from the global object list.
1230 hash = VMOBJ_HASH(object);
1231 lwkt_gettoken(&hash->token);
1232 TAILQ_REMOVE(&hash->list, object, object_list);
1233 lwkt_reltoken(&hash->token);
1235 if (object->ref_count != 0) {
1236 panic("vm_object_terminate2: object with references, "
1237 "ref_count=%d", object->ref_count);
1241 * NOTE: The object hold_count is at least 1, so we cannot kfree()
1242 * the object here. See vm_object_drop().
1247 * The caller must hold the object.
1250 vm_object_terminate_callback(vm_page_t p, void *data)
1252 struct rb_vm_page_scan_info *info = data;
1255 if ((++info->count & 63) == 0)
1258 if (object != info->object) {
1259 kprintf("vm_object_terminate_callback: obj/pg race %p/%p\n",
1263 vm_page_busy_wait(p, TRUE, "vmpgtrm");
1264 if (object != p->object) {
1265 kprintf("vm_object_terminate: Warning: Encountered "
1266 "busied page %p on queue %d\n", p, p->queue);
1268 } else if (p->wire_count == 0) {
1270 * NOTE: p->dirty and PG_NEED_COMMIT are ignored.
1273 mycpu->gd_cnt.v_pfree++;
1275 if (p->queue != PQ_NONE)
1276 kprintf("vm_object_terminate: Warning: Encountered "
1277 "wired page %p on queue %d\n", p, p->queue);
1285 * Clean all dirty pages in the specified range of object. Leaves page
1286 * on whatever queue it is currently on. If NOSYNC is set then do not
1287 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
1288 * leaving the object dirty.
1290 * When stuffing pages asynchronously, allow clustering. XXX we need a
1291 * synchronous clustering mode implementation.
1293 * Odd semantics: if start == end, we clean everything.
1295 * The object must be locked? XXX
1297 static int vm_object_page_clean_pass1(struct vm_page *p, void *data);
1298 static int vm_object_page_clean_pass2(struct vm_page *p, void *data);
1301 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1304 struct rb_vm_page_scan_info info;
1310 vm_object_hold(object);
1311 if (object->type != OBJT_VNODE ||
1312 (object->flags & OBJ_MIGHTBEDIRTY) == 0) {
1313 vm_object_drop(object);
1317 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ?
1318 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1319 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
1321 vp = object->handle;
1324 * Interlock other major object operations. This allows us to
1325 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY.
1327 vm_object_set_flag(object, OBJ_CLEANING);
1330 * Handle 'entire object' case
1332 info.start_pindex = start;
1334 info.end_pindex = object->size - 1;
1336 info.end_pindex = end - 1;
1338 wholescan = (start == 0 && info.end_pindex == object->size - 1);
1340 info.pagerflags = pagerflags;
1341 info.object = object;
1344 * If cleaning the entire object do a pass to mark the pages read-only.
1345 * If everything worked out ok, clear OBJ_WRITEABLE and
1351 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1352 vm_object_page_clean_pass1, &info);
1353 if (info.error == 0) {
1354 vm_object_clear_flag(object,
1355 OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1356 if (object->type == OBJT_VNODE &&
1357 (vp = (struct vnode *)object->handle) != NULL) {
1359 * Use new-style interface to clear VISDIRTY
1360 * because the vnode is not necessarily removed
1361 * from the syncer list(s) as often as it was
1362 * under the old interface, which can leave
1363 * the vnode on the syncer list after reclaim.
1371 * Do a pass to clean all the dirty pages we find.
1376 generation = object->generation;
1377 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1378 vm_object_page_clean_pass2, &info);
1379 } while (info.error || generation != object->generation);
1381 vm_object_clear_flag(object, OBJ_CLEANING);
1382 vm_object_drop(object);
1386 * The caller must hold the object.
1390 vm_object_page_clean_pass1(struct vm_page *p, void *data)
1392 struct rb_vm_page_scan_info *info = data;
1394 if ((++info->count & 63) == 0)
1396 if (p->object != info->object ||
1397 p->pindex < info->start_pindex ||
1398 p->pindex > info->end_pindex) {
1399 kprintf("vm_object_page_clean_pass1: obj/pg race %p/%p\n",
1403 vm_page_flag_set(p, PG_CLEANCHK);
1404 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1406 } else if (vm_page_busy_try(p, FALSE) == 0) {
1407 if (p->object == info->object)
1408 vm_page_protect(p, VM_PROT_READ);
1417 * The caller must hold the object
1421 vm_object_page_clean_pass2(struct vm_page *p, void *data)
1423 struct rb_vm_page_scan_info *info = data;
1426 if (p->object != info->object ||
1427 p->pindex < info->start_pindex ||
1428 p->pindex > info->end_pindex) {
1429 kprintf("vm_object_page_clean_pass2: obj/pg race %p/%p\n",
1435 * Do not mess with pages that were inserted after we started
1436 * the cleaning pass.
1438 if ((p->flags & PG_CLEANCHK) == 0)
1441 generation = info->object->generation;
1442 vm_page_busy_wait(p, TRUE, "vpcwai");
1444 if (p->object != info->object ||
1445 p->pindex < info->start_pindex ||
1446 p->pindex > info->end_pindex ||
1447 info->object->generation != generation) {
1454 * Before wasting time traversing the pmaps, check for trivial
1455 * cases where the page cannot be dirty.
1457 if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) {
1458 KKASSERT((p->dirty & p->valid) == 0 &&
1459 (p->flags & PG_NEED_COMMIT) == 0);
1465 * Check whether the page is dirty or not. The page has been set
1466 * to be read-only so the check will not race a user dirtying the
1469 vm_page_test_dirty(p);
1470 if ((p->dirty & p->valid) == 0 && (p->flags & PG_NEED_COMMIT) == 0) {
1471 vm_page_flag_clear(p, PG_CLEANCHK);
1477 * If we have been asked to skip nosync pages and this is a
1478 * nosync page, skip it. Note that the object flags were
1479 * not cleared in this case (because pass1 will have returned an
1480 * error), so we do not have to set them.
1482 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1483 vm_page_flag_clear(p, PG_CLEANCHK);
1489 * Flush as many pages as we can. PG_CLEANCHK will be cleared on
1490 * the pages that get successfully flushed. Set info->error if
1491 * we raced an object modification.
1493 vm_object_page_collect_flush(info->object, p, info->pagerflags);
1494 /* vm_wait_nominal(); this can deadlock the system in syncer/pageout */
1496 if ((++info->count & 63) == 0)
1503 * Collect the specified page and nearby pages and flush them out.
1504 * The number of pages flushed is returned. The passed page is busied
1505 * by the caller and we are responsible for its disposition.
1507 * The caller must hold the object.
1510 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags)
1518 vm_page_t ma[BLIST_MAX_ALLOC];
1520 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1523 page_base = pi % BLIST_MAX_ALLOC;
1531 tp = vm_page_lookup_busy_try(object, pi - page_base + ib,
1537 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1538 (tp->flags & PG_CLEANCHK) == 0) {
1542 if ((tp->queue - tp->pc) == PQ_CACHE) {
1543 vm_page_flag_clear(tp, PG_CLEANCHK);
1547 vm_page_test_dirty(tp);
1548 if ((tp->dirty & tp->valid) == 0 &&
1549 (tp->flags & PG_NEED_COMMIT) == 0) {
1550 vm_page_flag_clear(tp, PG_CLEANCHK);
1559 while (is < BLIST_MAX_ALLOC &&
1560 pi - page_base + is < object->size) {
1563 tp = vm_page_lookup_busy_try(object, pi - page_base + is,
1569 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1570 (tp->flags & PG_CLEANCHK) == 0) {
1574 if ((tp->queue - tp->pc) == PQ_CACHE) {
1575 vm_page_flag_clear(tp, PG_CLEANCHK);
1579 vm_page_test_dirty(tp);
1580 if ((tp->dirty & tp->valid) == 0 &&
1581 (tp->flags & PG_NEED_COMMIT) == 0) {
1582 vm_page_flag_clear(tp, PG_CLEANCHK);
1591 * All pages in the ma[] array are busied now
1593 for (i = ib; i < is; ++i) {
1594 vm_page_flag_clear(ma[i], PG_CLEANCHK);
1595 vm_page_hold(ma[i]); /* XXX need this any more? */
1597 vm_pageout_flush(&ma[ib], is - ib, pagerflags);
1598 for (i = ib; i < is; ++i) /* XXX need this any more? */
1599 vm_page_unhold(ma[i]);
1603 * Same as vm_object_pmap_copy, except range checking really
1604 * works, and is meant for small sections of an object.
1606 * This code protects resident pages by making them read-only
1607 * and is typically called on a fork or split when a page
1608 * is converted to copy-on-write.
1610 * NOTE: If the page is already at VM_PROT_NONE, calling
1611 * vm_page_protect will have no effect.
1614 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1619 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
1622 vm_object_hold(object);
1623 for (idx = start; idx < end; idx++) {
1624 p = vm_page_lookup(object, idx);
1627 vm_page_protect(p, VM_PROT_READ);
1629 vm_object_drop(object);
1633 * Removes all physical pages in the specified object range from all
1636 * The object must *not* be locked.
1639 static int vm_object_pmap_remove_callback(vm_page_t p, void *data);
1642 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1644 struct rb_vm_page_scan_info info;
1648 info.start_pindex = start;
1649 info.end_pindex = end - 1;
1651 info.object = object;
1653 vm_object_hold(object);
1654 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1655 vm_object_pmap_remove_callback, &info);
1656 if (start == 0 && end == object->size)
1657 vm_object_clear_flag(object, OBJ_WRITEABLE);
1658 vm_object_drop(object);
1662 * The caller must hold the object
1665 vm_object_pmap_remove_callback(vm_page_t p, void *data)
1667 struct rb_vm_page_scan_info *info = data;
1669 if ((++info->count & 63) == 0)
1672 if (info->object != p->object ||
1673 p->pindex < info->start_pindex ||
1674 p->pindex > info->end_pindex) {
1675 kprintf("vm_object_pmap_remove_callback: obj/pg race %p/%p\n",
1680 vm_page_protect(p, VM_PROT_NONE);
1686 * Implements the madvise function at the object/page level.
1688 * MADV_WILLNEED (any object)
1690 * Activate the specified pages if they are resident.
1692 * MADV_DONTNEED (any object)
1694 * Deactivate the specified pages if they are resident.
1696 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only)
1698 * Deactivate and clean the specified pages if they are
1699 * resident. This permits the process to reuse the pages
1700 * without faulting or the kernel to reclaim the pages
1706 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1708 vm_pindex_t end, tpindex;
1709 vm_object_t tobject;
1717 end = pindex + count;
1719 vm_object_hold(object);
1723 * Locate and adjust resident pages
1725 for (; pindex < end; pindex += 1) {
1727 if (tobject != object)
1728 vm_object_drop(tobject);
1733 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1734 * and those pages must be OBJ_ONEMAPPING.
1736 if (advise == MADV_FREE) {
1737 if ((tobject->type != OBJT_DEFAULT &&
1738 tobject->type != OBJT_SWAP) ||
1739 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1744 m = vm_page_lookup_busy_try(tobject, tpindex, TRUE, &error);
1747 vm_page_sleep_busy(m, TRUE, "madvpo");
1752 * There may be swap even if there is no backing page
1754 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1755 swap_pager_freespace(tobject, tpindex, 1);
1760 while ((xobj = tobject->backing_object) != NULL) {
1761 KKASSERT(xobj != object);
1762 vm_object_hold(xobj);
1763 if (xobj == tobject->backing_object)
1765 vm_object_drop(xobj);
1769 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1770 if (tobject != object) {
1771 vm_object_lock_swap();
1772 vm_object_drop(tobject);
1779 * If the page is not in a normal active state, we skip it.
1780 * If the page is not managed there are no page queues to
1781 * mess with. Things can break if we mess with pages in
1782 * any of the below states.
1784 if (m->wire_count ||
1785 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
1786 m->valid != VM_PAGE_BITS_ALL
1793 * Theoretically once a page is known not to be busy, an
1794 * interrupt cannot come along and rip it out from under us.
1797 if (advise == MADV_WILLNEED) {
1798 vm_page_activate(m);
1799 } else if (advise == MADV_DONTNEED) {
1800 vm_page_dontneed(m);
1801 } else if (advise == MADV_FREE) {
1803 * Mark the page clean. This will allow the page
1804 * to be freed up by the system. However, such pages
1805 * are often reused quickly by malloc()/free()
1806 * so we do not do anything that would cause
1807 * a page fault if we can help it.
1809 * Specifically, we do not try to actually free
1810 * the page now nor do we try to put it in the
1811 * cache (which would cause a page fault on reuse).
1813 * But we do make the page is freeable as we
1814 * can without actually taking the step of unmapping
1817 pmap_clear_modify(m);
1820 vm_page_dontneed(m);
1821 if (tobject->type == OBJT_SWAP)
1822 swap_pager_freespace(tobject, tpindex, 1);
1826 if (tobject != object)
1827 vm_object_drop(tobject);
1828 vm_object_drop(object);
1832 * Create a new object which is backed by the specified existing object
1833 * range. Replace the pointer and offset that was pointing at the existing
1834 * object with the pointer/offset for the new object.
1836 * If addref is non-zero the returned object is given an additional reference.
1837 * This mechanic exists to avoid the situation where refs might be 1 and
1838 * race against a collapse when the caller intends to bump it. So the
1839 * caller cannot add the ref after the fact. Used when the caller is
1840 * duplicating a vm_map_entry.
1842 * No other requirements.
1845 vm_object_shadow(vm_object_t *objectp, vm_ooffset_t *offset, vm_size_t length,
1855 * Don't create the new object if the old object isn't shared.
1856 * We have to chain wait before adding the reference to avoid
1857 * racing a collapse or deallocation.
1859 * Clear OBJ_ONEMAPPING flag when shadowing.
1861 * The caller owns a ref on source via *objectp which we are going
1862 * to replace. This ref is inherited by the backing_object assignment.
1863 * from nobject and does not need to be incremented here.
1865 * However, we add a temporary extra reference to the original source
1866 * prior to holding nobject in case we block, to avoid races where
1867 * someone else might believe that the source can be collapsed.
1871 if (source->type != OBJT_VNODE) {
1873 vm_object_hold(source);
1874 vm_object_chain_wait(source, 0);
1875 if (source->ref_count == 1 &&
1876 source->handle == NULL &&
1877 (source->type == OBJT_DEFAULT ||
1878 source->type == OBJT_SWAP)) {
1880 vm_object_reference_locked(source);
1881 vm_object_clear_flag(source,
1884 vm_object_drop(source);
1887 vm_object_reference_locked(source);
1888 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1890 vm_object_reference_quick(source);
1891 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1896 * Allocate a new object with the given length. The new object
1897 * is returned referenced but we may have to add another one.
1898 * If we are adding a second reference we must clear OBJ_ONEMAPPING.
1899 * (typically because the caller is about to clone a vm_map_entry).
1901 * The source object currently has an extra reference to prevent
1902 * collapses into it while we mess with its shadow list, which
1903 * we will remove later in this routine.
1905 * The target object may require a second reference if asked for one
1908 result = vm_object_allocate(OBJT_DEFAULT, length);
1910 panic("vm_object_shadow: no object for shadowing");
1911 vm_object_hold(result);
1913 vm_object_reference_locked(result);
1914 vm_object_clear_flag(result, OBJ_ONEMAPPING);
1918 * The new object shadows the source object. Chain wait before
1919 * adjusting shadow_count or the shadow list to avoid races.
1921 * Try to optimize the result object's page color when shadowing
1922 * in order to maintain page coloring consistency in the combined
1925 * The backing_object reference to source requires adding a ref to
1926 * source. We simply inherit the ref from the original *objectp
1927 * (which we are replacing) so no additional refs need to be added.
1928 * (we must still clean up the extra ref we had to prevent collapse
1931 * SHADOWING IS NOT APPLICABLE TO OBJT_VNODE OBJECTS
1933 KKASSERT(result->backing_object == NULL);
1934 result->backing_object = source;
1936 if (useshadowlist) {
1937 vm_object_chain_wait(source, 0);
1938 LIST_INSERT_HEAD(&source->shadow_head,
1939 result, shadow_list);
1940 source->shadow_count++;
1941 source->generation++;
1942 vm_object_set_flag(result, OBJ_ONSHADOW);
1944 /* cpu localization twist */
1945 result->pg_color = vm_quickcolor();
1949 * Adjust the return storage. Drop the ref on source before
1952 result->backing_object_offset = *offset;
1953 vm_object_drop(result);
1956 if (useshadowlist) {
1957 vm_object_deallocate_locked(source);
1958 vm_object_drop(source);
1960 vm_object_deallocate(source);
1965 * Return the new things
1970 #define OBSC_TEST_ALL_SHADOWED 0x0001
1971 #define OBSC_COLLAPSE_NOWAIT 0x0002
1972 #define OBSC_COLLAPSE_WAIT 0x0004
1974 static int vm_object_backing_scan_callback(vm_page_t p, void *data);
1977 * The caller must hold the object.
1980 vm_object_backing_scan(vm_object_t object, vm_object_t backing_object, int op)
1982 struct rb_vm_page_scan_info info;
1983 struct vm_object_hash *hash;
1985 vm_object_assert_held(object);
1986 vm_object_assert_held(backing_object);
1988 KKASSERT(backing_object == object->backing_object);
1989 info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1992 * Initial conditions
1994 if (op & OBSC_TEST_ALL_SHADOWED) {
1996 * We do not want to have to test for the existence of
1997 * swap pages in the backing object. XXX but with the
1998 * new swapper this would be pretty easy to do.
2000 * XXX what about anonymous MAP_SHARED memory that hasn't
2001 * been ZFOD faulted yet? If we do not test for this, the
2002 * shadow test may succeed! XXX
2004 if (backing_object->type != OBJT_DEFAULT)
2007 if (op & OBSC_COLLAPSE_WAIT) {
2008 KKASSERT((backing_object->flags & OBJ_DEAD) == 0);
2009 vm_object_set_flag(backing_object, OBJ_DEAD);
2011 hash = VMOBJ_HASH(backing_object);
2012 lwkt_gettoken(&hash->token);
2013 TAILQ_REMOVE(&hash->list, backing_object, object_list);
2014 lwkt_reltoken(&hash->token);
2018 * Our scan. We have to retry if a negative error code is returned,
2019 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that
2020 * the scan had to be stopped because the parent does not completely
2023 info.object = object;
2024 info.backing_object = backing_object;
2029 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL,
2030 vm_object_backing_scan_callback,
2032 } while (info.error < 0);
2038 * The caller must hold the object.
2041 vm_object_backing_scan_callback(vm_page_t p, void *data)
2043 struct rb_vm_page_scan_info *info = data;
2044 vm_object_t backing_object;
2047 vm_pindex_t new_pindex;
2048 vm_pindex_t backing_offset_index;
2052 new_pindex = pindex - info->backing_offset_index;
2054 object = info->object;
2055 backing_object = info->backing_object;
2056 backing_offset_index = info->backing_offset_index;
2058 if (op & OBSC_TEST_ALL_SHADOWED) {
2062 * Ignore pages outside the parent object's range
2063 * and outside the parent object's mapping of the
2066 * note that we do not busy the backing object's
2069 if (pindex < backing_offset_index ||
2070 new_pindex >= object->size
2076 * See if the parent has the page or if the parent's
2077 * object pager has the page. If the parent has the
2078 * page but the page is not valid, the parent's
2079 * object pager must have the page.
2081 * If this fails, the parent does not completely shadow
2082 * the object and we might as well give up now.
2084 pp = vm_page_lookup(object, new_pindex);
2085 if ((pp == NULL || pp->valid == 0) &&
2086 !vm_pager_has_page(object, new_pindex)
2088 info->error = 0; /* problemo */
2089 return(-1); /* stop the scan */
2094 * Check for busy page. Note that we may have lost (p) when we
2095 * possibly blocked above.
2097 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
2100 if (vm_page_busy_try(p, TRUE)) {
2101 if (op & OBSC_COLLAPSE_NOWAIT) {
2105 * If we slept, anything could have
2106 * happened. Ask that the scan be restarted.
2108 * Since the object is marked dead, the
2109 * backing offset should not have changed.
2111 vm_page_sleep_busy(p, TRUE, "vmocol");
2118 * If (p) is no longer valid restart the scan.
2120 if (p->object != backing_object || p->pindex != pindex) {
2121 kprintf("vm_object_backing_scan: Warning: page "
2122 "%p ripped out from under us\n", p);
2128 if (op & OBSC_COLLAPSE_NOWAIT) {
2129 if (p->valid == 0 ||
2131 (p->flags & PG_NEED_COMMIT)) {
2136 /* XXX what if p->valid == 0 , hold_count, etc? */
2140 p->object == backing_object,
2141 ("vm_object_qcollapse(): object mismatch")
2145 * Destroy any associated swap
2147 if (backing_object->type == OBJT_SWAP)
2148 swap_pager_freespace(backing_object, p->pindex, 1);
2151 p->pindex < backing_offset_index ||
2152 new_pindex >= object->size
2155 * Page is out of the parent object's range, we
2156 * can simply destroy it.
2158 vm_page_protect(p, VM_PROT_NONE);
2163 pp = vm_page_lookup(object, new_pindex);
2164 if (pp != NULL || vm_pager_has_page(object, new_pindex)) {
2166 * page already exists in parent OR swap exists
2167 * for this location in the parent. Destroy
2168 * the original page from the backing object.
2170 * Leave the parent's page alone
2172 vm_page_protect(p, VM_PROT_NONE);
2178 * Page does not exist in parent, rename the
2179 * page from the backing object to the main object.
2181 * If the page was mapped to a process, it can remain
2182 * mapped through the rename.
2184 if ((p->queue - p->pc) == PQ_CACHE)
2185 vm_page_deactivate(p);
2187 vm_page_rename(p, object, new_pindex);
2189 /* page automatically made dirty by rename */
2195 * This version of collapse allows the operation to occur earlier and
2196 * when paging_in_progress is true for an object... This is not a complete
2197 * operation, but should plug 99.9% of the rest of the leaks.
2199 * The caller must hold the object and backing_object and both must be
2202 * (only called from vm_object_collapse)
2205 vm_object_qcollapse(vm_object_t object, vm_object_t backing_object)
2207 if (backing_object->ref_count == 1) {
2208 atomic_add_int(&backing_object->ref_count, 2);
2209 #if defined(DEBUG_LOCKS)
2210 debugvm_object_add(backing_object, "qcollapse", 1, 2);
2212 vm_object_backing_scan(object, backing_object,
2213 OBSC_COLLAPSE_NOWAIT);
2214 atomic_add_int(&backing_object->ref_count, -2);
2215 #if defined(DEBUG_LOCKS)
2216 debugvm_object_add(backing_object, "qcollapse", 2, -2);
2222 * Collapse an object with the object backing it. Pages in the backing
2223 * object are moved into the parent, and the backing object is deallocated.
2224 * Any conflict is resolved in favor of the parent's existing pages.
2226 * object must be held and chain-locked on call.
2228 * The caller must have an extra ref on object to prevent a race from
2229 * destroying it during the collapse.
2232 vm_object_collapse(vm_object_t object, struct vm_object_dealloc_list **dlistp)
2234 struct vm_object_dealloc_list *dlist = NULL;
2235 vm_object_t backing_object;
2238 * Only one thread is attempting a collapse at any given moment.
2239 * There are few restrictions for (object) that callers of this
2240 * function check so reentrancy is likely.
2242 KKASSERT(object != NULL);
2243 vm_object_assert_held(object);
2244 KKASSERT(object->chainlk & (CHAINLK_MASK | CHAINLK_EXCL));
2251 * We can only collapse a DEFAULT/SWAP object with a
2252 * DEFAULT/SWAP object.
2254 if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP) {
2255 backing_object = NULL;
2259 backing_object = object->backing_object;
2260 if (backing_object == NULL)
2262 if (backing_object->type != OBJT_DEFAULT &&
2263 backing_object->type != OBJT_SWAP) {
2264 backing_object = NULL;
2269 * Hold the backing_object and check for races
2271 vm_object_hold(backing_object);
2272 if (backing_object != object->backing_object ||
2273 (backing_object->type != OBJT_DEFAULT &&
2274 backing_object->type != OBJT_SWAP)) {
2275 vm_object_drop(backing_object);
2280 * Chain-lock the backing object too because if we
2281 * successfully merge its pages into the top object we
2282 * will collapse backing_object->backing_object as the
2283 * new backing_object. Re-check that it is still our
2286 vm_object_chain_acquire(backing_object, 0);
2287 if (backing_object != object->backing_object) {
2288 vm_object_chain_release(backing_object);
2289 vm_object_drop(backing_object);
2294 * we check the backing object first, because it is most likely
2297 if (backing_object->handle != NULL ||
2298 (backing_object->type != OBJT_DEFAULT &&
2299 backing_object->type != OBJT_SWAP) ||
2300 (backing_object->flags & OBJ_DEAD) ||
2301 object->handle != NULL ||
2302 (object->type != OBJT_DEFAULT &&
2303 object->type != OBJT_SWAP) ||
2304 (object->flags & OBJ_DEAD)) {
2309 * If paging is in progress we can't do a normal collapse.
2312 object->paging_in_progress != 0 ||
2313 backing_object->paging_in_progress != 0
2315 vm_object_qcollapse(object, backing_object);
2320 * We know that we can either collapse the backing object (if
2321 * the parent is the only reference to it) or (perhaps) have
2322 * the parent bypass the object if the parent happens to shadow
2323 * all the resident pages in the entire backing object.
2325 * This is ignoring pager-backed pages such as swap pages.
2326 * vm_object_backing_scan fails the shadowing test in this
2329 if (backing_object->ref_count == 1) {
2331 * If there is exactly one reference to the backing
2332 * object, we can collapse it into the parent.
2334 KKASSERT(object->backing_object == backing_object);
2335 vm_object_backing_scan(object, backing_object,
2336 OBSC_COLLAPSE_WAIT);
2339 * Move the pager from backing_object to object.
2341 if (backing_object->type == OBJT_SWAP) {
2342 vm_object_pip_add(backing_object, 1);
2345 * scrap the paging_offset junk and do a
2346 * discrete copy. This also removes major
2347 * assumptions about how the swap-pager
2348 * works from where it doesn't belong. The
2349 * new swapper is able to optimize the
2350 * destroy-source case.
2352 vm_object_pip_add(object, 1);
2353 swap_pager_copy(backing_object, object,
2354 OFF_TO_IDX(object->backing_object_offset),
2356 vm_object_pip_wakeup(object);
2357 vm_object_pip_wakeup(backing_object);
2361 * Object now shadows whatever backing_object did.
2362 * Remove object from backing_object's shadow_list.
2364 * Removing object from backing_objects shadow list
2365 * requires releasing object, which we will do below.
2367 KKASSERT(object->backing_object == backing_object);
2368 if (object->flags & OBJ_ONSHADOW) {
2369 LIST_REMOVE(object, shadow_list);
2370 backing_object->shadow_count--;
2371 backing_object->generation++;
2372 vm_object_clear_flag(object, OBJ_ONSHADOW);
2376 * backing_object->backing_object moves from within
2377 * backing_object to within object.
2379 * OBJT_VNODE bbobj's should have empty shadow lists.
2381 while ((bbobj = backing_object->backing_object) != NULL) {
2382 if (bbobj->type == OBJT_VNODE)
2383 vm_object_hold_shared(bbobj);
2385 vm_object_hold(bbobj);
2386 if (bbobj == backing_object->backing_object)
2388 vm_object_drop(bbobj);
2392 * We are removing backing_object from bbobj's
2393 * shadow list and adding object to bbobj's shadow
2394 * list, so the ref_count on bbobj is unchanged.
2397 if (backing_object->flags & OBJ_ONSHADOW) {
2398 /* not locked exclusively if vnode */
2399 KKASSERT(bbobj->type != OBJT_VNODE);
2400 LIST_REMOVE(backing_object,
2402 bbobj->shadow_count--;
2403 bbobj->generation++;
2404 vm_object_clear_flag(backing_object,
2407 backing_object->backing_object = NULL;
2409 object->backing_object = bbobj;
2411 if (bbobj->type != OBJT_VNODE) {
2412 LIST_INSERT_HEAD(&bbobj->shadow_head,
2413 object, shadow_list);
2414 bbobj->shadow_count++;
2415 bbobj->generation++;
2416 vm_object_set_flag(object,
2421 object->backing_object_offset +=
2422 backing_object->backing_object_offset;
2424 vm_object_drop(bbobj);
2427 * Discard the old backing_object. Nothing should be
2428 * able to ref it, other than a vm_map_split(),
2429 * and vm_map_split() will stall on our chain lock.
2430 * And we control the parent so it shouldn't be
2431 * possible for it to go away either.
2433 * Since the backing object has no pages, no pager
2434 * left, and no object references within it, all
2435 * that is necessary is to dispose of it.
2437 KASSERT(backing_object->ref_count == 1,
2438 ("backing_object %p was somehow "
2439 "re-referenced during collapse!",
2441 KASSERT(RB_EMPTY(&backing_object->rb_memq),
2442 ("backing_object %p somehow has left "
2443 "over pages during collapse!",
2447 * The object can be destroyed.
2449 * XXX just fall through and dodealloc instead
2450 * of forcing destruction?
2452 atomic_add_int(&backing_object->ref_count, -1);
2453 #if defined(DEBUG_LOCKS)
2454 debugvm_object_add(backing_object, "collapse", 1, -1);
2456 if ((backing_object->flags & OBJ_DEAD) == 0)
2457 vm_object_terminate(backing_object);
2462 * If we do not entirely shadow the backing object,
2463 * there is nothing we can do so we give up.
2465 if (vm_object_backing_scan(object, backing_object,
2466 OBSC_TEST_ALL_SHADOWED) == 0) {
2471 * bbobj is backing_object->backing_object. Since
2472 * object completely shadows backing_object we can
2473 * bypass it and become backed by bbobj instead.
2475 * The shadow list for vnode backing objects is not
2476 * used and a shared hold is allowed.
2478 while ((bbobj = backing_object->backing_object) != NULL) {
2479 if (bbobj->type == OBJT_VNODE)
2480 vm_object_hold_shared(bbobj);
2482 vm_object_hold(bbobj);
2483 if (bbobj == backing_object->backing_object)
2485 vm_object_drop(bbobj);
2489 * Make object shadow bbobj instead of backing_object.
2490 * Remove object from backing_object's shadow list.
2492 * Deallocating backing_object will not remove
2493 * it, since its reference count is at least 2.
2495 * Removing object from backing_object's shadow
2496 * list requires releasing a ref, which we do
2497 * below by setting dodealloc to 1.
2499 KKASSERT(object->backing_object == backing_object);
2500 if (object->flags & OBJ_ONSHADOW) {
2501 LIST_REMOVE(object, shadow_list);
2502 backing_object->shadow_count--;
2503 backing_object->generation++;
2504 vm_object_clear_flag(object, OBJ_ONSHADOW);
2508 * Add a ref to bbobj, bbobj now shadows object.
2510 * NOTE: backing_object->backing_object still points
2511 * to bbobj. That relationship remains intact
2512 * because backing_object has > 1 ref, so
2513 * someone else is pointing to it (hence why
2514 * we can't collapse it into object and can
2515 * only handle the all-shadowed bypass case).
2518 if (bbobj->type != OBJT_VNODE) {
2519 vm_object_chain_wait(bbobj, 0);
2520 vm_object_reference_locked(bbobj);
2521 LIST_INSERT_HEAD(&bbobj->shadow_head,
2522 object, shadow_list);
2523 bbobj->shadow_count++;
2524 bbobj->generation++;
2525 vm_object_set_flag(object,
2528 vm_object_reference_quick(bbobj);
2530 object->backing_object_offset +=
2531 backing_object->backing_object_offset;
2532 object->backing_object = bbobj;
2533 vm_object_drop(bbobj);
2535 object->backing_object = NULL;
2539 * Drop the reference count on backing_object. To
2540 * handle ref_count races properly we can't assume
2541 * that the ref_count is still at least 2 so we
2542 * have to actually call vm_object_deallocate()
2543 * (after clearing the chainlock).
2550 * Ok, we want to loop on the new object->bbobj association,
2551 * possibly collapsing it further. However if dodealloc is
2552 * non-zero we have to deallocate the backing_object which
2553 * itself can potentially undergo a collapse, creating a
2554 * recursion depth issue with the LWKT token subsystem.
2556 * In the case where we must deallocate the backing_object
2557 * it is possible now that the backing_object has a single
2558 * shadow count on some other object (not represented here
2559 * as yet), since it no longer shadows us. Thus when we
2560 * call vm_object_deallocate() it may attempt to collapse
2561 * itself into its remaining parent.
2564 struct vm_object_dealloc_list *dtmp;
2566 vm_object_chain_release(backing_object);
2567 vm_object_unlock(backing_object);
2568 /* backing_object remains held */
2571 * Auto-deallocation list for caller convenience.
2576 dtmp = kmalloc(sizeof(*dtmp), M_TEMP, M_WAITOK);
2577 dtmp->object = backing_object;
2578 dtmp->next = *dlistp;
2581 vm_object_chain_release(backing_object);
2582 vm_object_drop(backing_object);
2584 /* backing_object = NULL; not needed */
2589 * Clean up any left over backing_object
2591 if (backing_object) {
2592 vm_object_chain_release(backing_object);
2593 vm_object_drop(backing_object);
2597 * Clean up any auto-deallocation list. This is a convenience
2598 * for top-level callers so they don't have to pass &dlist.
2599 * Do not clean up any caller-passed dlistp, the caller will
2603 vm_object_deallocate_list(&dlist);
2608 * vm_object_collapse() may collect additional objects in need of
2609 * deallocation. This routine deallocates these objects. The
2610 * deallocation itself can trigger additional collapses (which the
2611 * deallocate function takes care of). This procedure is used to
2612 * reduce procedural recursion since these vm_object shadow chains
2613 * can become quite long.
2616 vm_object_deallocate_list(struct vm_object_dealloc_list **dlistp)
2618 struct vm_object_dealloc_list *dlist;
2620 while ((dlist = *dlistp) != NULL) {
2621 *dlistp = dlist->next;
2622 vm_object_lock(dlist->object);
2623 vm_object_deallocate_locked(dlist->object);
2624 vm_object_drop(dlist->object);
2625 kfree(dlist, M_TEMP);
2630 * Removes all physical pages in the specified object range from the
2631 * object's list of pages.
2635 static int vm_object_page_remove_callback(vm_page_t p, void *data);
2638 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2639 boolean_t clean_only)
2641 struct rb_vm_page_scan_info info;
2645 * Degenerate cases and assertions
2647 vm_object_hold(object);
2648 if (object == NULL ||
2649 (object->resident_page_count == 0 && object->swblock_count == 0)) {
2650 vm_object_drop(object);
2653 KASSERT(object->type != OBJT_PHYS,
2654 ("attempt to remove pages from a physical object"));
2657 * Indicate that paging is occuring on the object
2659 vm_object_pip_add(object, 1);
2662 * Figure out the actual removal range and whether we are removing
2663 * the entire contents of the object or not. If removing the entire
2664 * contents, be sure to get all pages, even those that might be
2665 * beyond the end of the object.
2667 info.object = object;
2668 info.start_pindex = start;
2670 info.end_pindex = (vm_pindex_t)-1;
2672 info.end_pindex = end - 1;
2673 info.limit = clean_only;
2675 all = (start == 0 && info.end_pindex >= object->size - 1);
2678 * Loop until we are sure we have gotten them all.
2682 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2683 vm_object_page_remove_callback, &info);
2684 } while (info.error);
2687 * Remove any related swap if throwing away pages, or for
2688 * non-swap objects (the swap is a clean copy in that case).
2690 if (object->type != OBJT_SWAP || clean_only == FALSE) {
2692 swap_pager_freespace_all(object);
2694 swap_pager_freespace(object, info.start_pindex,
2695 info.end_pindex - info.start_pindex + 1);
2701 vm_object_pip_wakeup(object);
2702 vm_object_drop(object);
2706 * The caller must hold the object.
2708 * NOTE: User yields are allowed when removing more than one page, but not
2709 * allowed if only removing one page (the path for single page removals
2710 * might hold a spinlock).
2713 vm_object_page_remove_callback(vm_page_t p, void *data)
2715 struct rb_vm_page_scan_info *info = data;
2717 if ((++info->count & 63) == 0)
2720 if (info->object != p->object ||
2721 p->pindex < info->start_pindex ||
2722 p->pindex > info->end_pindex) {
2723 kprintf("vm_object_page_remove_callbackA: obj/pg race %p/%p\n",
2727 if (vm_page_busy_try(p, TRUE)) {
2728 vm_page_sleep_busy(p, TRUE, "vmopar");
2732 if (info->object != p->object) {
2733 /* this should never happen */
2734 kprintf("vm_object_page_remove_callbackB: obj/pg race %p/%p\n",
2741 * Wired pages cannot be destroyed, but they can be invalidated
2742 * and we do so if clean_only (limit) is not set.
2744 * WARNING! The page may be wired due to being part of a buffer
2745 * cache buffer, and the buffer might be marked B_CACHE.
2746 * This is fine as part of a truncation but VFSs must be
2747 * sure to fix the buffer up when re-extending the file.
2749 * NOTE! PG_NEED_COMMIT is ignored.
2751 if (p->wire_count != 0) {
2752 vm_page_protect(p, VM_PROT_NONE);
2753 if (info->limit == 0)
2760 * limit is our clean_only flag. If set and the page is dirty or
2761 * requires a commit, do not free it. If set and the page is being
2762 * held by someone, do not free it.
2764 if (info->limit && p->valid) {
2765 vm_page_test_dirty(p);
2766 if ((p->valid & p->dirty) || (p->flags & PG_NEED_COMMIT)) {
2775 vm_page_protect(p, VM_PROT_NONE);
2782 * Coalesces two objects backing up adjoining regions of memory into a
2785 * returns TRUE if objects were combined.
2787 * NOTE: Only works at the moment if the second object is NULL -
2788 * if it's not, which object do we lock first?
2791 * prev_object First object to coalesce
2792 * prev_offset Offset into prev_object
2793 * next_object Second object into coalesce
2794 * next_offset Offset into next_object
2796 * prev_size Size of reference to prev_object
2797 * next_size Size of reference to next_object
2799 * The caller does not need to hold (prev_object) but must have a stable
2800 * pointer to it (typically by holding the vm_map locked).
2803 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
2804 vm_size_t prev_size, vm_size_t next_size)
2806 vm_pindex_t next_pindex;
2808 if (prev_object == NULL)
2811 vm_object_hold(prev_object);
2813 if (prev_object->type != OBJT_DEFAULT &&
2814 prev_object->type != OBJT_SWAP) {
2815 vm_object_drop(prev_object);
2820 * Try to collapse the object first
2822 vm_object_chain_acquire(prev_object, 0);
2823 vm_object_collapse(prev_object, NULL);
2826 * Can't coalesce if: . more than one reference . paged out . shadows
2827 * another object . has a copy elsewhere (any of which mean that the
2828 * pages not mapped to prev_entry may be in use anyway)
2831 if (prev_object->backing_object != NULL) {
2832 vm_object_chain_release(prev_object);
2833 vm_object_drop(prev_object);
2837 prev_size >>= PAGE_SHIFT;
2838 next_size >>= PAGE_SHIFT;
2839 next_pindex = prev_pindex + prev_size;
2841 if ((prev_object->ref_count > 1) &&
2842 (prev_object->size != next_pindex)) {
2843 vm_object_chain_release(prev_object);
2844 vm_object_drop(prev_object);
2849 * Remove any pages that may still be in the object from a previous
2852 if (next_pindex < prev_object->size) {
2853 vm_object_page_remove(prev_object,
2855 next_pindex + next_size, FALSE);
2856 if (prev_object->type == OBJT_SWAP)
2857 swap_pager_freespace(prev_object,
2858 next_pindex, next_size);
2862 * Extend the object if necessary.
2864 if (next_pindex + next_size > prev_object->size)
2865 prev_object->size = next_pindex + next_size;
2867 vm_object_chain_release(prev_object);
2868 vm_object_drop(prev_object);
2873 * Make the object writable and flag is being possibly dirty.
2875 * The object might not be held (or might be held but held shared),
2876 * the related vnode is probably not held either. Object and vnode are
2877 * stable by virtue of the vm_page busied by the caller preventing
2880 * If the related mount is flagged MNTK_THR_SYNC we need to call
2881 * vsetobjdirty(). Filesystems using this option usually shortcut
2882 * synchronization by only scanning the syncer list.
2885 vm_object_set_writeable_dirty(vm_object_t object)
2889 /*vm_object_assert_held(object);*/
2891 * Avoid contention in vm fault path by checking the state before
2892 * issuing an atomic op on it.
2894 if ((object->flags & (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) !=
2895 (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) {
2896 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
2898 if (object->type == OBJT_VNODE &&
2899 (vp = (struct vnode *)object->handle) != NULL) {
2900 if ((vp->v_flag & VOBJDIRTY) == 0) {
2902 (vp->v_mount->mnt_kern_flag & MNTK_THR_SYNC)) {
2904 * New style THR_SYNC places vnodes on the
2905 * syncer list more deterministically.
2910 * Old style scan would not necessarily place
2911 * a vnode on the syncer list when possibly
2912 * modified via mmap.
2914 vsetflags(vp, VOBJDIRTY);
2920 #include "opt_ddb.h"
2922 #include <sys/kernel.h>
2924 #include <sys/cons.h>
2926 #include <ddb/ddb.h>
2928 static int _vm_object_in_map (vm_map_t map, vm_object_t object,
2929 vm_map_entry_t entry);
2930 static int vm_object_in_map (vm_object_t object);
2933 * The caller must hold the object.
2936 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2939 vm_map_entry_t tmpe;
2940 vm_object_t obj, nobj;
2946 tmpe = map->header.next;
2947 entcount = map->nentries;
2948 while (entcount-- && (tmpe != &map->header)) {
2949 if( _vm_object_in_map(map, object, tmpe)) {
2956 switch(entry->maptype) {
2957 case VM_MAPTYPE_SUBMAP:
2958 tmpm = entry->object.sub_map;
2959 tmpe = tmpm->header.next;
2960 entcount = tmpm->nentries;
2961 while (entcount-- && tmpe != &tmpm->header) {
2962 if( _vm_object_in_map(tmpm, object, tmpe)) {
2968 case VM_MAPTYPE_NORMAL:
2969 case VM_MAPTYPE_VPAGETABLE:
2970 obj = entry->object.vm_object;
2972 if (obj == object) {
2973 if (obj != entry->object.vm_object)
2974 vm_object_drop(obj);
2977 while ((nobj = obj->backing_object) != NULL) {
2978 vm_object_hold(nobj);
2979 if (nobj == obj->backing_object)
2981 vm_object_drop(nobj);
2983 if (obj != entry->object.vm_object) {
2985 vm_object_lock_swap();
2986 vm_object_drop(obj);
2997 static int vm_object_in_map_callback(struct proc *p, void *data);
2999 struct vm_object_in_map_info {
3008 vm_object_in_map(vm_object_t object)
3010 struct vm_object_in_map_info info;
3013 info.object = object;
3015 allproc_scan(vm_object_in_map_callback, &info);
3018 if( _vm_object_in_map(&kernel_map, object, 0))
3020 if( _vm_object_in_map(&pager_map, object, 0))
3022 if( _vm_object_in_map(&buffer_map, object, 0))
3031 vm_object_in_map_callback(struct proc *p, void *data)
3033 struct vm_object_in_map_info *info = data;
3036 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) {
3044 DB_SHOW_COMMAND(vmochk, vm_object_check)
3046 struct vm_object_hash *hash;
3051 * make sure that internal objs are in a map somewhere
3052 * and none have zero ref counts.
3054 for (n = 0; n < VMOBJ_HSIZE; ++n) {
3055 hash = &vm_object_hash[n];
3056 for (object = TAILQ_FIRST(&hash->list);
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=%ld, 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)
3160 struct vm_object_hash *hash;
3166 for (n = 0; n < VMOBJ_HSIZE; ++n) {
3167 hash = &vm_object_hash[n];
3168 for (object = TAILQ_FIRST(&hash->list);
3170 object = TAILQ_NEXT(object, object_list)) {
3171 vm_pindex_t idx, fidx;
3173 vm_paddr_t pa = -1, padiff;
3177 if (object->type == OBJT_MARKER)
3179 db_printf("new object: %p\n", (void *)object);
3189 osize = object->size;
3192 for (idx = 0; idx < osize; idx++) {
3193 m = vm_page_lookup(object, idx);
3196 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
3197 (long)fidx, rcount, (long)pa);
3211 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
3216 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
3217 padiff >>= PAGE_SHIFT;
3218 padiff &= PQ_L2_MASK;
3220 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
3224 db_printf(" index(%ld)run(%d)pa(0x%lx)",
3225 (long)fidx, rcount, (long)pa);
3226 db_printf("pd(%ld)\n", (long)padiff);
3236 pa = VM_PAGE_TO_PHYS(m);
3240 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
3241 (long)fidx, rcount, (long)pa);