2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
64 * Virtual memory object module.
67 #include <sys/cdefs.h>
68 __FBSDID("$FreeBSD$");
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/blockcount.h>
75 #include <sys/cpuset.h>
76 #include <sys/limits.h>
79 #include <sys/mount.h>
80 #include <sys/kernel.h>
81 #include <sys/pctrie.h>
82 #include <sys/sysctl.h>
83 #include <sys/mutex.h>
84 #include <sys/proc.h> /* for curproc, pageproc */
85 #include <sys/refcount.h>
86 #include <sys/socket.h>
87 #include <sys/resourcevar.h>
88 #include <sys/refcount.h>
89 #include <sys/rwlock.h>
91 #include <sys/vnode.h>
92 #include <sys/vmmeter.h>
96 #include <vm/vm_param.h>
98 #include <vm/vm_map.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_page.h>
101 #include <vm/vm_pageout.h>
102 #include <vm/vm_pager.h>
103 #include <vm/vm_phys.h>
104 #include <vm/vm_pagequeue.h>
105 #include <vm/swap_pager.h>
106 #include <vm/vm_kern.h>
107 #include <vm/vm_extern.h>
108 #include <vm/vm_radix.h>
109 #include <vm/vm_reserv.h>
112 static int old_msync;
113 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
114 "Use old (insecure) msync behavior");
116 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
117 int pagerflags, int flags, boolean_t *allclean,
119 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
120 boolean_t *allclean);
121 static void vm_object_backing_remove(vm_object_t object);
124 * Virtual memory objects maintain the actual data
125 * associated with allocated virtual memory. A given
126 * page of memory exists within exactly one object.
128 * An object is only deallocated when all "references"
129 * are given up. Only one "reference" to a given
130 * region of an object should be writeable.
132 * Associated with each object is a list of all resident
133 * memory pages belonging to that object; this list is
134 * maintained by the "vm_page" module, and locked by the object's
137 * Each object also records a "pager" routine which is
138 * used to retrieve (and store) pages to the proper backing
139 * storage. In addition, objects may be backed by other
140 * objects from which they were virtual-copied.
142 * The only items within the object structure which are
143 * modified after time of creation are:
144 * reference count locked by object's lock
145 * pager routine locked by object's lock
149 struct object_q vm_object_list;
150 struct mtx vm_object_list_mtx; /* lock for object list and count */
152 struct vm_object kernel_object_store;
154 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
157 static COUNTER_U64_DEFINE_EARLY(object_collapses);
158 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
160 "VM object collapses");
162 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
163 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
165 "VM object bypasses");
167 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
168 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
169 &object_collapse_waits,
170 "Number of sleeps for collapse");
172 static uma_zone_t obj_zone;
174 static int vm_object_zinit(void *mem, int size, int flags);
177 static void vm_object_zdtor(void *mem, int size, void *arg);
180 vm_object_zdtor(void *mem, int size, void *arg)
184 object = (vm_object_t)mem;
185 KASSERT(object->ref_count == 0,
186 ("object %p ref_count = %d", object, object->ref_count));
187 KASSERT(TAILQ_EMPTY(&object->memq),
188 ("object %p has resident pages in its memq", object));
189 KASSERT(vm_radix_is_empty(&object->rtree),
190 ("object %p has resident pages in its trie", object));
191 #if VM_NRESERVLEVEL > 0
192 KASSERT(LIST_EMPTY(&object->rvq),
193 ("object %p has reservations",
196 KASSERT(!vm_object_busied(object),
197 ("object %p busy = %d", object, blockcount_read(&object->busy)));
198 KASSERT(object->resident_page_count == 0,
199 ("object %p resident_page_count = %d",
200 object, object->resident_page_count));
201 KASSERT(object->shadow_count == 0,
202 ("object %p shadow_count = %d",
203 object, object->shadow_count));
204 KASSERT(object->type == OBJT_DEAD,
205 ("object %p has non-dead type %d",
206 object, object->type));
211 vm_object_zinit(void *mem, int size, int flags)
215 object = (vm_object_t)mem;
216 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
218 /* These are true for any object that has been freed */
219 object->type = OBJT_DEAD;
220 vm_radix_init(&object->rtree);
221 refcount_init(&object->ref_count, 0);
222 blockcount_init(&object->paging_in_progress);
223 blockcount_init(&object->busy);
224 object->resident_page_count = 0;
225 object->shadow_count = 0;
226 object->flags = OBJ_DEAD;
228 mtx_lock(&vm_object_list_mtx);
229 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
230 mtx_unlock(&vm_object_list_mtx);
235 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
236 vm_object_t object, void *handle)
239 TAILQ_INIT(&object->memq);
240 LIST_INIT(&object->shadow_head);
243 if (type == OBJT_SWAP)
244 pctrie_init(&object->un_pager.swp.swp_blks);
247 * Ensure that swap_pager_swapoff() iteration over object_list
248 * sees up to date type and pctrie head if it observed
251 atomic_thread_fence_rel();
253 object->pg_color = 0;
254 object->flags = flags;
256 object->domain.dr_policy = NULL;
257 object->generation = 1;
258 object->cleangeneration = 1;
259 refcount_init(&object->ref_count, 1);
260 object->memattr = VM_MEMATTR_DEFAULT;
263 object->handle = handle;
264 object->backing_object = NULL;
265 object->backing_object_offset = (vm_ooffset_t) 0;
266 #if VM_NRESERVLEVEL > 0
267 LIST_INIT(&object->rvq);
269 umtx_shm_object_init(object);
275 * Initialize the VM objects module.
280 TAILQ_INIT(&vm_object_list);
281 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
283 rw_init(&kernel_object->lock, "kernel vm object");
284 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
285 VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
286 #if VM_NRESERVLEVEL > 0
287 kernel_object->flags |= OBJ_COLORED;
288 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
290 kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
293 * The lock portion of struct vm_object must be type stable due
294 * to vm_pageout_fallback_object_lock locking a vm object
295 * without holding any references to it.
297 * paging_in_progress is valid always. Lockless references to
298 * the objects may acquire pip and then check OBJ_DEAD.
300 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
306 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
312 vm_object_clear_flag(vm_object_t object, u_short bits)
315 VM_OBJECT_ASSERT_WLOCKED(object);
316 object->flags &= ~bits;
320 * Sets the default memory attribute for the specified object. Pages
321 * that are allocated to this object are by default assigned this memory
324 * Presently, this function must be called before any pages are allocated
325 * to the object. In the future, this requirement may be relaxed for
326 * "default" and "swap" objects.
329 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
332 VM_OBJECT_ASSERT_WLOCKED(object);
333 switch (object->type) {
341 if (!TAILQ_EMPTY(&object->memq))
342 return (KERN_FAILURE);
345 return (KERN_INVALID_ARGUMENT);
347 panic("vm_object_set_memattr: object %p is of undefined type",
350 object->memattr = memattr;
351 return (KERN_SUCCESS);
355 vm_object_pip_add(vm_object_t object, short i)
359 blockcount_acquire(&object->paging_in_progress, i);
363 vm_object_pip_wakeup(vm_object_t object)
366 vm_object_pip_wakeupn(object, 1);
370 vm_object_pip_wakeupn(vm_object_t object, short i)
374 blockcount_release(&object->paging_in_progress, i);
378 * Atomically drop the object lock and wait for pip to drain. This protects
379 * from sleep/wakeup races due to identity changes. The lock is not re-acquired
383 vm_object_pip_sleep(vm_object_t object, const char *waitid)
386 (void)blockcount_sleep(&object->paging_in_progress, &object->lock,
387 waitid, PVM | PDROP);
391 vm_object_pip_wait(vm_object_t object, const char *waitid)
394 VM_OBJECT_ASSERT_WLOCKED(object);
396 blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
401 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
404 VM_OBJECT_ASSERT_UNLOCKED(object);
406 blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
410 * vm_object_allocate:
412 * Returns a new object with the given size.
415 vm_object_allocate(objtype_t type, vm_pindex_t size)
422 panic("vm_object_allocate: can't create OBJT_DEAD");
429 flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
432 flags = OBJ_FICTITIOUS;
435 flags = OBJ_UNMANAGED;
441 panic("vm_object_allocate: type %d is undefined", type);
443 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
444 _vm_object_allocate(type, size, flags, object, NULL);
450 * vm_object_allocate_anon:
452 * Returns a new default object of the given size and marked as
453 * anonymous memory for special split/collapse handling. Color
454 * to be initialized by the caller.
457 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
458 struct ucred *cred, vm_size_t charge)
460 vm_object_t handle, object;
462 if (backing_object == NULL)
464 else if ((backing_object->flags & OBJ_ANON) != 0)
465 handle = backing_object->handle;
467 handle = backing_object;
468 object = uma_zalloc(obj_zone, M_WAITOK);
469 _vm_object_allocate(OBJT_DEFAULT, size, OBJ_ANON | OBJ_ONEMAPPING,
472 object->charge = cred != NULL ? charge : 0;
477 vm_object_reference_vnode(vm_object_t object)
482 * vnode objects need the lock for the first reference
483 * to serialize with vnode_object_deallocate().
485 if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
486 VM_OBJECT_RLOCK(object);
487 old = refcount_acquire(&object->ref_count);
488 if (object->type == OBJT_VNODE && old == 0)
489 vref(object->handle);
490 VM_OBJECT_RUNLOCK(object);
495 * vm_object_reference:
497 * Acquires a reference to the given object.
500 vm_object_reference(vm_object_t object)
506 if (object->type == OBJT_VNODE)
507 vm_object_reference_vnode(object);
509 refcount_acquire(&object->ref_count);
510 KASSERT((object->flags & OBJ_DEAD) == 0,
511 ("vm_object_reference: Referenced dead object."));
515 * vm_object_reference_locked:
517 * Gets another reference to the given object.
519 * The object must be locked.
522 vm_object_reference_locked(vm_object_t object)
526 VM_OBJECT_ASSERT_LOCKED(object);
527 old = refcount_acquire(&object->ref_count);
528 if (object->type == OBJT_VNODE && old == 0)
529 vref(object->handle);
530 KASSERT((object->flags & OBJ_DEAD) == 0,
531 ("vm_object_reference: Referenced dead object."));
535 * Handle deallocating an object of type OBJT_VNODE.
538 vm_object_deallocate_vnode(vm_object_t object)
540 struct vnode *vp = (struct vnode *) object->handle;
543 KASSERT(object->type == OBJT_VNODE,
544 ("vm_object_deallocate_vnode: not a vnode object"));
545 KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
547 /* Object lock to protect handle lookup. */
548 last = refcount_release(&object->ref_count);
549 VM_OBJECT_RUNLOCK(object);
554 if (!umtx_shm_vnobj_persistent)
555 umtx_shm_object_terminated(object);
557 /* vrele may need the vnode lock. */
562 * We dropped a reference on an object and discovered that it had a
563 * single remaining shadow. This is a sibling of the reference we
564 * dropped. Attempt to collapse the sibling and backing object.
567 vm_object_deallocate_anon(vm_object_t backing_object)
571 /* Fetch the final shadow. */
572 object = LIST_FIRST(&backing_object->shadow_head);
573 KASSERT(object != NULL && backing_object->shadow_count == 1,
574 ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
575 backing_object->ref_count, backing_object->shadow_count));
576 KASSERT((object->flags & (OBJ_TMPFS_NODE | OBJ_ANON)) == OBJ_ANON,
577 ("invalid shadow object %p", object));
579 if (!VM_OBJECT_TRYWLOCK(object)) {
581 * Prevent object from disappearing since we do not have a
584 vm_object_pip_add(object, 1);
585 VM_OBJECT_WUNLOCK(backing_object);
586 VM_OBJECT_WLOCK(object);
587 vm_object_pip_wakeup(object);
589 VM_OBJECT_WUNLOCK(backing_object);
592 * Check for a collapse/terminate race with the last reference holder.
594 if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
595 !refcount_acquire_if_not_zero(&object->ref_count)) {
596 VM_OBJECT_WUNLOCK(object);
599 backing_object = object->backing_object;
600 if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
601 vm_object_collapse(object);
602 VM_OBJECT_WUNLOCK(object);
608 * vm_object_deallocate:
610 * Release a reference to the specified object,
611 * gained either through a vm_object_allocate
612 * or a vm_object_reference call. When all references
613 * are gone, storage associated with this object
614 * may be relinquished.
616 * No object may be locked.
619 vm_object_deallocate(vm_object_t object)
624 while (object != NULL) {
626 * If the reference count goes to 0 we start calling
627 * vm_object_terminate() on the object chain. A ref count
628 * of 1 may be a special case depending on the shadow count
629 * being 0 or 1. These cases require a write lock on the
632 if ((object->flags & OBJ_ANON) == 0)
633 released = refcount_release_if_gt(&object->ref_count, 1);
635 released = refcount_release_if_gt(&object->ref_count, 2);
639 if (object->type == OBJT_VNODE) {
640 VM_OBJECT_RLOCK(object);
641 if (object->type == OBJT_VNODE) {
642 vm_object_deallocate_vnode(object);
645 VM_OBJECT_RUNLOCK(object);
648 VM_OBJECT_WLOCK(object);
649 KASSERT(object->ref_count > 0,
650 ("vm_object_deallocate: object deallocated too many times: %d",
654 * If this is not the final reference to an anonymous
655 * object we may need to collapse the shadow chain.
657 if (!refcount_release(&object->ref_count)) {
658 if (object->ref_count > 1 ||
659 object->shadow_count == 0) {
660 if ((object->flags & OBJ_ANON) != 0 &&
661 object->ref_count == 1)
662 vm_object_set_flag(object,
664 VM_OBJECT_WUNLOCK(object);
668 /* Handle collapsing last ref on anonymous objects. */
669 object = vm_object_deallocate_anon(object);
674 * Handle the final reference to an object. We restart
675 * the loop with the backing object to avoid recursion.
677 umtx_shm_object_terminated(object);
678 temp = object->backing_object;
680 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
681 ("shadowed tmpfs v_object 2 %p", object));
682 vm_object_backing_remove(object);
685 KASSERT((object->flags & OBJ_DEAD) == 0,
686 ("vm_object_deallocate: Terminating dead object."));
687 vm_object_set_flag(object, OBJ_DEAD);
688 vm_object_terminate(object);
694 * vm_object_destroy removes the object from the global object list
695 * and frees the space for the object.
698 vm_object_destroy(vm_object_t object)
702 * Release the allocation charge.
704 if (object->cred != NULL) {
705 swap_release_by_cred(object->charge, object->cred);
707 crfree(object->cred);
712 * Free the space for the object.
714 uma_zfree(obj_zone, object);
718 vm_object_backing_remove_locked(vm_object_t object)
720 vm_object_t backing_object;
722 backing_object = object->backing_object;
723 VM_OBJECT_ASSERT_WLOCKED(object);
724 VM_OBJECT_ASSERT_WLOCKED(backing_object);
726 KASSERT((object->flags & OBJ_COLLAPSING) == 0,
727 ("vm_object_backing_remove: Removing collapsing object."));
729 if ((object->flags & OBJ_SHADOWLIST) != 0) {
730 LIST_REMOVE(object, shadow_list);
731 backing_object->shadow_count--;
732 object->flags &= ~OBJ_SHADOWLIST;
734 object->backing_object = NULL;
738 vm_object_backing_remove(vm_object_t object)
740 vm_object_t backing_object;
742 VM_OBJECT_ASSERT_WLOCKED(object);
744 if ((object->flags & OBJ_SHADOWLIST) != 0) {
745 backing_object = object->backing_object;
746 VM_OBJECT_WLOCK(backing_object);
747 vm_object_backing_remove_locked(object);
748 VM_OBJECT_WUNLOCK(backing_object);
750 object->backing_object = NULL;
754 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
757 VM_OBJECT_ASSERT_WLOCKED(object);
759 if ((backing_object->flags & OBJ_ANON) != 0) {
760 VM_OBJECT_ASSERT_WLOCKED(backing_object);
761 LIST_INSERT_HEAD(&backing_object->shadow_head, object,
763 backing_object->shadow_count++;
764 object->flags |= OBJ_SHADOWLIST;
766 object->backing_object = backing_object;
770 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
773 VM_OBJECT_ASSERT_WLOCKED(object);
775 if ((backing_object->flags & OBJ_ANON) != 0) {
776 VM_OBJECT_WLOCK(backing_object);
777 vm_object_backing_insert_locked(object, backing_object);
778 VM_OBJECT_WUNLOCK(backing_object);
780 object->backing_object = backing_object;
784 * Insert an object into a backing_object's shadow list with an additional
785 * reference to the backing_object added.
788 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
791 VM_OBJECT_ASSERT_WLOCKED(object);
793 if ((backing_object->flags & OBJ_ANON) != 0) {
794 VM_OBJECT_WLOCK(backing_object);
795 KASSERT((backing_object->flags & OBJ_DEAD) == 0,
796 ("shadowing dead anonymous object"));
797 vm_object_reference_locked(backing_object);
798 vm_object_backing_insert_locked(object, backing_object);
799 vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
800 VM_OBJECT_WUNLOCK(backing_object);
802 vm_object_reference(backing_object);
803 object->backing_object = backing_object;
808 * Transfer a backing reference from backing_object to object.
811 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
813 vm_object_t new_backing_object;
816 * Note that the reference to backing_object->backing_object
817 * moves from within backing_object to within object.
819 vm_object_backing_remove_locked(object);
820 new_backing_object = backing_object->backing_object;
821 if (new_backing_object == NULL)
823 if ((new_backing_object->flags & OBJ_ANON) != 0) {
824 VM_OBJECT_WLOCK(new_backing_object);
825 vm_object_backing_remove_locked(backing_object);
826 vm_object_backing_insert_locked(object, new_backing_object);
827 VM_OBJECT_WUNLOCK(new_backing_object);
829 object->backing_object = new_backing_object;
830 backing_object->backing_object = NULL;
835 * Wait for a concurrent collapse to settle.
838 vm_object_collapse_wait(vm_object_t object)
841 VM_OBJECT_ASSERT_WLOCKED(object);
843 while ((object->flags & OBJ_COLLAPSING) != 0) {
844 vm_object_pip_wait(object, "vmcolwait");
845 counter_u64_add(object_collapse_waits, 1);
850 * Waits for a backing object to clear a pending collapse and returns
851 * it locked if it is an ANON object.
854 vm_object_backing_collapse_wait(vm_object_t object)
856 vm_object_t backing_object;
858 VM_OBJECT_ASSERT_WLOCKED(object);
861 backing_object = object->backing_object;
862 if (backing_object == NULL ||
863 (backing_object->flags & OBJ_ANON) == 0)
865 VM_OBJECT_WLOCK(backing_object);
866 if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
868 VM_OBJECT_WUNLOCK(object);
869 vm_object_pip_sleep(backing_object, "vmbckwait");
870 counter_u64_add(object_collapse_waits, 1);
871 VM_OBJECT_WLOCK(object);
873 return (backing_object);
877 * vm_object_terminate_pages removes any remaining pageable pages
878 * from the object and resets the object to an empty state.
881 vm_object_terminate_pages(vm_object_t object)
885 VM_OBJECT_ASSERT_WLOCKED(object);
888 * Free any remaining pageable pages. This also removes them from the
889 * paging queues. However, don't free wired pages, just remove them
890 * from the object. Rather than incrementally removing each page from
891 * the object, the page and object are reset to any empty state.
893 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
894 vm_page_assert_unbusied(p);
895 KASSERT(p->object == object &&
896 (p->ref_count & VPRC_OBJREF) != 0,
897 ("vm_object_terminate_pages: page %p is inconsistent", p));
900 if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
907 * If the object contained any pages, then reset it to an empty state.
908 * None of the object's fields, including "resident_page_count", were
909 * modified by the preceding loop.
911 if (object->resident_page_count != 0) {
912 vm_radix_reclaim_allnodes(&object->rtree);
913 TAILQ_INIT(&object->memq);
914 object->resident_page_count = 0;
915 if (object->type == OBJT_VNODE)
916 vdrop(object->handle);
921 * vm_object_terminate actually destroys the specified object, freeing
922 * up all previously used resources.
924 * The object must be locked.
925 * This routine may block.
928 vm_object_terminate(vm_object_t object)
931 VM_OBJECT_ASSERT_WLOCKED(object);
932 KASSERT((object->flags & OBJ_DEAD) != 0,
933 ("terminating non-dead obj %p", object));
934 KASSERT((object->flags & OBJ_COLLAPSING) == 0,
935 ("terminating collapsing obj %p", object));
936 KASSERT(object->backing_object == NULL,
937 ("terminating shadow obj %p", object));
940 * Wait for the pageout daemon and other current users to be
941 * done with the object. Note that new paging_in_progress
942 * users can come after this wait, but they must check
943 * OBJ_DEAD flag set (without unlocking the object), and avoid
944 * the object being terminated.
946 vm_object_pip_wait(object, "objtrm");
948 KASSERT(object->ref_count == 0,
949 ("vm_object_terminate: object with references, ref_count=%d",
952 if ((object->flags & OBJ_PG_DTOR) == 0)
953 vm_object_terminate_pages(object);
955 #if VM_NRESERVLEVEL > 0
956 if (__predict_false(!LIST_EMPTY(&object->rvq)))
957 vm_reserv_break_all(object);
960 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
961 object->type == OBJT_SWAP,
962 ("%s: non-swap obj %p has cred", __func__, object));
965 * Let the pager know object is dead.
967 vm_pager_deallocate(object);
968 VM_OBJECT_WUNLOCK(object);
970 vm_object_destroy(object);
974 * Make the page read-only so that we can clear the object flags. However, if
975 * this is a nosync mmap then the object is likely to stay dirty so do not
976 * mess with the page and do not clear the object flags. Returns TRUE if the
977 * page should be flushed, and FALSE otherwise.
980 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
983 vm_page_assert_busied(p);
986 * If we have been asked to skip nosync pages and this is a
987 * nosync page, skip it. Note that the object flags were not
988 * cleared in this case so we do not have to set them.
990 if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
994 pmap_remove_write(p);
995 return (p->dirty != 0);
1000 * vm_object_page_clean
1002 * Clean all dirty pages in the specified range of object. Leaves page
1003 * on whatever queue it is currently on. If NOSYNC is set then do not
1004 * write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1005 * leaving the object dirty.
1007 * For swap objects backing tmpfs regular files, do not flush anything,
1008 * but remove write protection on the mapped pages to update mtime through
1011 * When stuffing pages asynchronously, allow clustering. XXX we need a
1012 * synchronous clustering mode implementation.
1014 * Odd semantics: if start == end, we clean everything.
1016 * The object must be locked.
1018 * Returns FALSE if some page from the range was not written, as
1019 * reported by the pager, and TRUE otherwise.
1022 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1026 vm_pindex_t pi, tend, tstart;
1027 int curgeneration, n, pagerflags;
1028 boolean_t eio, res, allclean;
1030 VM_OBJECT_ASSERT_WLOCKED(object);
1032 if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1035 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1036 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1037 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1039 tstart = OFF_TO_IDX(start);
1040 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1041 allclean = tstart == 0 && tend >= object->size;
1045 curgeneration = object->generation;
1047 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1051 np = TAILQ_NEXT(p, listq);
1052 if (vm_page_none_valid(p))
1054 if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1055 if (object->generation != curgeneration &&
1056 (flags & OBJPC_SYNC) != 0)
1058 np = vm_page_find_least(object, pi);
1061 if (!vm_object_page_remove_write(p, flags, &allclean)) {
1065 if (object->type == OBJT_VNODE) {
1066 n = vm_object_page_collect_flush(object, p, pagerflags,
1067 flags, &allclean, &eio);
1072 if (object->generation != curgeneration &&
1073 (flags & OBJPC_SYNC) != 0)
1077 * If the VOP_PUTPAGES() did a truncated write, so
1078 * that even the first page of the run is not fully
1079 * written, vm_pageout_flush() returns 0 as the run
1080 * length. Since the condition that caused truncated
1081 * write may be permanent, e.g. exhausted free space,
1082 * accepting n == 0 would cause an infinite loop.
1084 * Forwarding the iterator leaves the unwritten page
1085 * behind, but there is not much we can do there if
1086 * filesystem refuses to write it.
1096 np = vm_page_find_least(object, pi + n);
1099 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1103 * Leave updating cleangeneration for tmpfs objects to tmpfs
1104 * scan. It needs to update mtime, which happens for other
1105 * filesystems during page writeouts.
1107 if (allclean && object->type == OBJT_VNODE)
1108 object->cleangeneration = curgeneration;
1113 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1114 int flags, boolean_t *allclean, boolean_t *eio)
1116 vm_page_t ma[vm_pageout_page_count], p_first, tp;
1117 int count, i, mreq, runlen;
1119 vm_page_lock_assert(p, MA_NOTOWNED);
1120 vm_page_assert_xbusied(p);
1121 VM_OBJECT_ASSERT_WLOCKED(object);
1126 for (tp = p; count < vm_pageout_page_count; count++) {
1127 tp = vm_page_next(tp);
1128 if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1130 if (!vm_object_page_remove_write(tp, flags, allclean)) {
1131 vm_page_xunbusy(tp);
1136 for (p_first = p; count < vm_pageout_page_count; count++) {
1137 tp = vm_page_prev(p_first);
1138 if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1140 if (!vm_object_page_remove_write(tp, flags, allclean)) {
1141 vm_page_xunbusy(tp);
1148 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1151 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1156 * Note that there is absolutely no sense in writing out
1157 * anonymous objects, so we track down the vnode object
1159 * We invalidate (remove) all pages from the address space
1160 * for semantic correctness.
1162 * If the backing object is a device object with unmanaged pages, then any
1163 * mappings to the specified range of pages must be removed before this
1164 * function is called.
1166 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1167 * may start out with a NULL object.
1170 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1171 boolean_t syncio, boolean_t invalidate)
1173 vm_object_t backing_object;
1176 int error, flags, fsync_after;
1183 VM_OBJECT_WLOCK(object);
1184 while ((backing_object = object->backing_object) != NULL) {
1185 VM_OBJECT_WLOCK(backing_object);
1186 offset += object->backing_object_offset;
1187 VM_OBJECT_WUNLOCK(object);
1188 object = backing_object;
1189 if (object->size < OFF_TO_IDX(offset + size))
1190 size = IDX_TO_OFF(object->size) - offset;
1193 * Flush pages if writing is allowed, invalidate them
1194 * if invalidation requested. Pages undergoing I/O
1195 * will be ignored by vm_object_page_remove().
1197 * We cannot lock the vnode and then wait for paging
1198 * to complete without deadlocking against vm_fault.
1199 * Instead we simply call vm_object_page_remove() and
1200 * allow it to block internally on a page-by-page
1201 * basis when it encounters pages undergoing async
1204 if (object->type == OBJT_VNODE &&
1205 vm_object_mightbedirty(object) != 0 &&
1206 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1207 VM_OBJECT_WUNLOCK(object);
1208 (void) vn_start_write(vp, &mp, V_WAIT);
1209 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1210 if (syncio && !invalidate && offset == 0 &&
1211 atop(size) == object->size) {
1213 * If syncing the whole mapping of the file,
1214 * it is faster to schedule all the writes in
1215 * async mode, also allowing the clustering,
1216 * and then wait for i/o to complete.
1221 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1222 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1223 fsync_after = FALSE;
1225 VM_OBJECT_WLOCK(object);
1226 res = vm_object_page_clean(object, offset, offset + size,
1228 VM_OBJECT_WUNLOCK(object);
1230 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1232 vn_finished_write(mp);
1235 VM_OBJECT_WLOCK(object);
1237 if ((object->type == OBJT_VNODE ||
1238 object->type == OBJT_DEVICE) && invalidate) {
1239 if (object->type == OBJT_DEVICE)
1241 * The option OBJPR_NOTMAPPED must be passed here
1242 * because vm_object_page_remove() cannot remove
1243 * unmanaged mappings.
1245 flags = OBJPR_NOTMAPPED;
1249 flags = OBJPR_CLEANONLY;
1250 vm_object_page_remove(object, OFF_TO_IDX(offset),
1251 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1253 VM_OBJECT_WUNLOCK(object);
1258 * Determine whether the given advice can be applied to the object. Advice is
1259 * not applied to unmanaged pages since they never belong to page queues, and
1260 * since MADV_FREE is destructive, it can apply only to anonymous pages that
1261 * have been mapped at most once.
1264 vm_object_advice_applies(vm_object_t object, int advice)
1267 if ((object->flags & OBJ_UNMANAGED) != 0)
1269 if (advice != MADV_FREE)
1271 return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1272 (OBJ_ONEMAPPING | OBJ_ANON));
1276 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1280 if (advice == MADV_FREE)
1281 vm_pager_freespace(object, pindex, size);
1285 * vm_object_madvise:
1287 * Implements the madvise function at the object/page level.
1289 * MADV_WILLNEED (any object)
1291 * Activate the specified pages if they are resident.
1293 * MADV_DONTNEED (any object)
1295 * Deactivate the specified pages if they are resident.
1297 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1298 * OBJ_ONEMAPPING only)
1300 * Deactivate and clean the specified pages if they are
1301 * resident. This permits the process to reuse the pages
1302 * without faulting or the kernel to reclaim the pages
1306 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1309 vm_pindex_t tpindex;
1310 vm_object_t backing_object, tobject;
1317 VM_OBJECT_WLOCK(object);
1318 if (!vm_object_advice_applies(object, advice)) {
1319 VM_OBJECT_WUNLOCK(object);
1322 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1326 * If the next page isn't resident in the top-level object, we
1327 * need to search the shadow chain. When applying MADV_FREE, we
1328 * take care to release any swap space used to store
1329 * non-resident pages.
1331 if (m == NULL || pindex < m->pindex) {
1333 * Optimize a common case: if the top-level object has
1334 * no backing object, we can skip over the non-resident
1335 * range in constant time.
1337 if (object->backing_object == NULL) {
1338 tpindex = (m != NULL && m->pindex < end) ?
1340 vm_object_madvise_freespace(object, advice,
1341 pindex, tpindex - pindex);
1342 if ((pindex = tpindex) == end)
1349 vm_object_madvise_freespace(tobject, advice,
1352 * Prepare to search the next object in the
1355 backing_object = tobject->backing_object;
1356 if (backing_object == NULL)
1358 VM_OBJECT_WLOCK(backing_object);
1360 OFF_TO_IDX(tobject->backing_object_offset);
1361 if (tobject != object)
1362 VM_OBJECT_WUNLOCK(tobject);
1363 tobject = backing_object;
1364 if (!vm_object_advice_applies(tobject, advice))
1366 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
1371 m = TAILQ_NEXT(m, listq);
1375 * If the page is not in a normal state, skip it. The page
1376 * can not be invalidated while the object lock is held.
1378 if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1380 KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1381 ("vm_object_madvise: page %p is fictitious", tm));
1382 KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1383 ("vm_object_madvise: page %p is not managed", tm));
1384 if (vm_page_tryxbusy(tm) == 0) {
1385 if (object != tobject)
1386 VM_OBJECT_WUNLOCK(object);
1387 if (advice == MADV_WILLNEED) {
1389 * Reference the page before unlocking and
1390 * sleeping so that the page daemon is less
1391 * likely to reclaim it.
1393 vm_page_aflag_set(tm, PGA_REFERENCED);
1395 vm_page_busy_sleep(tm, "madvpo", false);
1398 vm_page_advise(tm, advice);
1399 vm_page_xunbusy(tm);
1400 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1402 if (tobject != object)
1403 VM_OBJECT_WUNLOCK(tobject);
1405 VM_OBJECT_WUNLOCK(object);
1411 * Create a new object which is backed by the
1412 * specified existing object range. The source
1413 * object reference is deallocated.
1415 * The new object and offset into that object
1416 * are returned in the source parameters.
1419 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1420 struct ucred *cred, bool shared)
1428 * Don't create the new object if the old object isn't shared.
1430 * If we hold the only reference we can guarantee that it won't
1431 * increase while we have the map locked. Otherwise the race is
1432 * harmless and we will end up with an extra shadow object that
1433 * will be collapsed later.
1435 if (source != NULL && source->ref_count == 1 &&
1436 (source->flags & OBJ_ANON) != 0)
1440 * Allocate a new object with the given length.
1442 result = vm_object_allocate_anon(atop(length), source, cred, length);
1445 * Store the offset into the source object, and fix up the offset into
1448 result->backing_object_offset = *offset;
1450 if (shared || source != NULL) {
1451 VM_OBJECT_WLOCK(result);
1454 * The new object shadows the source object, adding a
1455 * reference to it. Our caller changes his reference
1456 * to point to the new object, removing a reference to
1457 * the source object. Net result: no change of
1458 * reference count, unless the caller needs to add one
1459 * more reference due to forking a shared map entry.
1462 vm_object_reference_locked(result);
1463 vm_object_clear_flag(result, OBJ_ONEMAPPING);
1467 * Try to optimize the result object's page color when
1468 * shadowing in order to maintain page coloring
1469 * consistency in the combined shadowed object.
1471 if (source != NULL) {
1472 vm_object_backing_insert(result, source);
1473 result->domain = source->domain;
1474 #if VM_NRESERVLEVEL > 0
1475 result->flags |= source->flags & OBJ_COLORED;
1476 result->pg_color = (source->pg_color +
1477 OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1481 VM_OBJECT_WUNLOCK(result);
1485 * Return the new things
1494 * Split the pages in a map entry into a new object. This affords
1495 * easier removal of unused pages, and keeps object inheritance from
1496 * being a negative impact on memory usage.
1499 vm_object_split(vm_map_entry_t entry)
1501 vm_page_t m, m_busy, m_next;
1502 vm_object_t orig_object, new_object, backing_object;
1503 vm_pindex_t idx, offidxstart;
1506 orig_object = entry->object.vm_object;
1507 KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1508 ("vm_object_split: Splitting object with multiple mappings."));
1509 if ((orig_object->flags & OBJ_ANON) == 0)
1511 if (orig_object->ref_count <= 1)
1513 VM_OBJECT_WUNLOCK(orig_object);
1515 offidxstart = OFF_TO_IDX(entry->offset);
1516 size = atop(entry->end - entry->start);
1519 * If swap_pager_copy() is later called, it will convert new_object
1520 * into a swap object.
1522 new_object = vm_object_allocate_anon(size, orig_object,
1523 orig_object->cred, ptoa(size));
1526 * We must wait for the orig_object to complete any in-progress
1527 * collapse so that the swap blocks are stable below. The
1528 * additional reference on backing_object by new object will
1529 * prevent further collapse operations until split completes.
1531 VM_OBJECT_WLOCK(orig_object);
1532 vm_object_collapse_wait(orig_object);
1535 * At this point, the new object is still private, so the order in
1536 * which the original and new objects are locked does not matter.
1538 VM_OBJECT_WLOCK(new_object);
1539 new_object->domain = orig_object->domain;
1540 backing_object = orig_object->backing_object;
1541 if (backing_object != NULL) {
1542 vm_object_backing_insert_ref(new_object, backing_object);
1543 new_object->backing_object_offset =
1544 orig_object->backing_object_offset + entry->offset;
1546 if (orig_object->cred != NULL) {
1547 crhold(orig_object->cred);
1548 KASSERT(orig_object->charge >= ptoa(size),
1549 ("orig_object->charge < 0"));
1550 orig_object->charge -= ptoa(size);
1554 * Mark the split operation so that swap_pager_getpages() knows
1555 * that the object is in transition.
1557 vm_object_set_flag(orig_object, OBJ_SPLIT);
1563 m = vm_page_find_least(orig_object, offidxstart);
1564 KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1565 ("%s: object %p was repopulated", __func__, orig_object));
1566 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1568 m_next = TAILQ_NEXT(m, listq);
1571 * We must wait for pending I/O to complete before we can
1574 * We do not have to VM_PROT_NONE the page as mappings should
1575 * not be changed by this operation.
1577 if (vm_page_tryxbusy(m) == 0) {
1578 VM_OBJECT_WUNLOCK(new_object);
1579 vm_page_sleep_if_busy(m, "spltwt");
1580 VM_OBJECT_WLOCK(new_object);
1585 * The page was left invalid. Likely placed there by
1586 * an incomplete fault. Just remove and ignore.
1588 if (vm_page_none_valid(m)) {
1589 if (vm_page_remove(m))
1594 /* vm_page_rename() will dirty the page. */
1595 if (vm_page_rename(m, new_object, idx)) {
1597 VM_OBJECT_WUNLOCK(new_object);
1598 VM_OBJECT_WUNLOCK(orig_object);
1600 VM_OBJECT_WLOCK(orig_object);
1601 VM_OBJECT_WLOCK(new_object);
1605 #if VM_NRESERVLEVEL > 0
1607 * If some of the reservation's allocated pages remain with
1608 * the original object, then transferring the reservation to
1609 * the new object is neither particularly beneficial nor
1610 * particularly harmful as compared to leaving the reservation
1611 * with the original object. If, however, all of the
1612 * reservation's allocated pages are transferred to the new
1613 * object, then transferring the reservation is typically
1614 * beneficial. Determining which of these two cases applies
1615 * would be more costly than unconditionally renaming the
1618 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1622 * orig_object's type may change while sleeping, so keep track
1623 * of the beginning of the busied range.
1625 if (orig_object->type != OBJT_SWAP)
1627 else if (m_busy == NULL)
1630 if (orig_object->type == OBJT_SWAP) {
1632 * swap_pager_copy() can sleep, in which case the orig_object's
1633 * and new_object's locks are released and reacquired.
1635 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1637 TAILQ_FOREACH_FROM(m_busy, &new_object->memq, listq)
1638 vm_page_xunbusy(m_busy);
1640 vm_object_clear_flag(orig_object, OBJ_SPLIT);
1641 VM_OBJECT_WUNLOCK(orig_object);
1642 VM_OBJECT_WUNLOCK(new_object);
1643 entry->object.vm_object = new_object;
1644 entry->offset = 0LL;
1645 vm_object_deallocate(orig_object);
1646 VM_OBJECT_WLOCK(new_object);
1650 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1652 vm_object_t backing_object;
1654 VM_OBJECT_ASSERT_WLOCKED(object);
1655 backing_object = object->backing_object;
1656 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1658 KASSERT(p == NULL || p->object == object || p->object == backing_object,
1659 ("invalid ownership %p %p %p", p, object, backing_object));
1660 /* The page is only NULL when rename fails. */
1662 VM_OBJECT_WUNLOCK(object);
1663 VM_OBJECT_WUNLOCK(backing_object);
1666 if (p->object == object)
1667 VM_OBJECT_WUNLOCK(backing_object);
1669 VM_OBJECT_WUNLOCK(object);
1670 vm_page_busy_sleep(p, "vmocol", false);
1672 VM_OBJECT_WLOCK(object);
1673 VM_OBJECT_WLOCK(backing_object);
1674 return (TAILQ_FIRST(&backing_object->memq));
1678 vm_object_scan_all_shadowed(vm_object_t object)
1680 vm_object_t backing_object;
1682 vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1684 VM_OBJECT_ASSERT_WLOCKED(object);
1685 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1687 backing_object = object->backing_object;
1689 if ((backing_object->flags & OBJ_ANON) == 0)
1692 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1693 p = vm_page_find_least(backing_object, pi);
1694 ps = swap_pager_find_least(backing_object, pi);
1697 * Only check pages inside the parent object's range and
1698 * inside the parent object's mapping of the backing object.
1701 if (p != NULL && p->pindex < pi)
1702 p = TAILQ_NEXT(p, listq);
1704 ps = swap_pager_find_least(backing_object, pi);
1705 if (p == NULL && ps >= backing_object->size)
1710 pi = MIN(p->pindex, ps);
1712 new_pindex = pi - backing_offset_index;
1713 if (new_pindex >= object->size)
1718 * If the backing object page is busy a
1719 * grandparent or older page may still be
1720 * undergoing CoW. It is not safe to collapse
1721 * the backing object until it is quiesced.
1723 if (vm_page_tryxbusy(p) == 0)
1727 * We raced with the fault handler that left
1728 * newly allocated invalid page on the object
1729 * queue and retried.
1731 if (!vm_page_all_valid(p))
1736 * See if the parent has the page or if the parent's object
1737 * pager has the page. If the parent has the page but the page
1738 * is not valid, the parent's object pager must have the page.
1740 * If this fails, the parent does not completely shadow the
1741 * object and we might as well give up now.
1743 pp = vm_page_lookup(object, new_pindex);
1746 * The valid check here is stable due to object lock
1747 * being required to clear valid and initiate paging.
1748 * Busy of p disallows fault handler to validate pp.
1750 if ((pp == NULL || vm_page_none_valid(pp)) &&
1751 !vm_pager_has_page(object, new_pindex, NULL, NULL))
1765 vm_object_collapse_scan(vm_object_t object)
1767 vm_object_t backing_object;
1768 vm_page_t next, p, pp;
1769 vm_pindex_t backing_offset_index, new_pindex;
1771 VM_OBJECT_ASSERT_WLOCKED(object);
1772 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1774 backing_object = object->backing_object;
1775 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1780 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1781 next = TAILQ_NEXT(p, listq);
1782 new_pindex = p->pindex - backing_offset_index;
1785 * Check for busy page
1787 if (vm_page_tryxbusy(p) == 0) {
1788 next = vm_object_collapse_scan_wait(object, p);
1792 KASSERT(object->backing_object == backing_object,
1793 ("vm_object_collapse_scan: backing object mismatch %p != %p",
1794 object->backing_object, backing_object));
1795 KASSERT(p->object == backing_object,
1796 ("vm_object_collapse_scan: object mismatch %p != %p",
1797 p->object, backing_object));
1799 if (p->pindex < backing_offset_index ||
1800 new_pindex >= object->size) {
1801 vm_pager_freespace(backing_object, p->pindex, 1);
1803 KASSERT(!pmap_page_is_mapped(p),
1804 ("freeing mapped page %p", p));
1805 if (vm_page_remove(p))
1810 if (!vm_page_all_valid(p)) {
1811 KASSERT(!pmap_page_is_mapped(p),
1812 ("freeing mapped page %p", p));
1813 if (vm_page_remove(p))
1818 pp = vm_page_lookup(object, new_pindex);
1819 if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1822 * The page in the parent is busy and possibly not
1823 * (yet) valid. Until its state is finalized by the
1824 * busy bit owner, we can't tell whether it shadows the
1827 next = vm_object_collapse_scan_wait(object, pp);
1831 if (pp != NULL && vm_page_none_valid(pp)) {
1833 * The page was invalid in the parent. Likely placed
1834 * there by an incomplete fault. Just remove and
1835 * ignore. p can replace it.
1837 if (vm_page_remove(pp))
1842 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1845 * The page already exists in the parent OR swap exists
1846 * for this location in the parent. Leave the parent's
1847 * page alone. Destroy the original page from the
1850 vm_pager_freespace(backing_object, p->pindex, 1);
1851 KASSERT(!pmap_page_is_mapped(p),
1852 ("freeing mapped page %p", p));
1853 if (vm_page_remove(p))
1856 vm_page_xunbusy(pp);
1861 * Page does not exist in parent, rename the page from the
1862 * backing object to the main object.
1864 * If the page was mapped to a process, it can remain mapped
1865 * through the rename. vm_page_rename() will dirty the page.
1867 if (vm_page_rename(p, object, new_pindex)) {
1869 next = vm_object_collapse_scan_wait(object, NULL);
1873 /* Use the old pindex to free the right page. */
1874 vm_pager_freespace(backing_object, new_pindex +
1875 backing_offset_index, 1);
1877 #if VM_NRESERVLEVEL > 0
1879 * Rename the reservation.
1881 vm_reserv_rename(p, object, backing_object,
1882 backing_offset_index);
1890 * vm_object_collapse:
1892 * Collapse an object with the object backing it.
1893 * Pages in the backing object are moved into the
1894 * parent, and the backing object is deallocated.
1897 vm_object_collapse(vm_object_t object)
1899 vm_object_t backing_object, new_backing_object;
1901 VM_OBJECT_ASSERT_WLOCKED(object);
1904 KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1905 ("collapsing invalid object"));
1908 * Wait for the backing_object to finish any pending
1909 * collapse so that the caller sees the shortest possible
1912 backing_object = vm_object_backing_collapse_wait(object);
1913 if (backing_object == NULL)
1916 KASSERT(object->ref_count > 0 &&
1917 object->ref_count > object->shadow_count,
1918 ("collapse with invalid ref %d or shadow %d count.",
1919 object->ref_count, object->shadow_count));
1920 KASSERT((backing_object->flags &
1921 (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1922 ("vm_object_collapse: Backing object already collapsing."));
1923 KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1924 ("vm_object_collapse: object is already collapsing."));
1927 * We know that we can either collapse the backing object if
1928 * the parent is the only reference to it, or (perhaps) have
1929 * the parent bypass the object if the parent happens to shadow
1930 * all the resident pages in the entire backing object.
1932 if (backing_object->ref_count == 1) {
1933 KASSERT(backing_object->shadow_count == 1,
1934 ("vm_object_collapse: shadow_count: %d",
1935 backing_object->shadow_count));
1936 vm_object_pip_add(object, 1);
1937 vm_object_set_flag(object, OBJ_COLLAPSING);
1938 vm_object_pip_add(backing_object, 1);
1939 vm_object_set_flag(backing_object, OBJ_DEAD);
1942 * If there is exactly one reference to the backing
1943 * object, we can collapse it into the parent.
1945 vm_object_collapse_scan(object);
1947 #if VM_NRESERVLEVEL > 0
1949 * Break any reservations from backing_object.
1951 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1952 vm_reserv_break_all(backing_object);
1956 * Move the pager from backing_object to object.
1958 if (backing_object->type == OBJT_SWAP) {
1960 * swap_pager_copy() can sleep, in which case
1961 * the backing_object's and object's locks are
1962 * released and reacquired.
1963 * Since swap_pager_copy() is being asked to
1964 * destroy backing_object, it will change the
1965 * type to OBJT_DEFAULT.
1970 OFF_TO_IDX(object->backing_object_offset), TRUE);
1974 * Object now shadows whatever backing_object did.
1976 vm_object_clear_flag(object, OBJ_COLLAPSING);
1977 vm_object_backing_transfer(object, backing_object);
1978 object->backing_object_offset +=
1979 backing_object->backing_object_offset;
1980 VM_OBJECT_WUNLOCK(object);
1981 vm_object_pip_wakeup(object);
1984 * Discard backing_object.
1986 * Since the backing object has no pages, no pager left,
1987 * and no object references within it, all that is
1988 * necessary is to dispose of it.
1990 KASSERT(backing_object->ref_count == 1, (
1991 "backing_object %p was somehow re-referenced during collapse!",
1993 vm_object_pip_wakeup(backing_object);
1994 (void)refcount_release(&backing_object->ref_count);
1995 vm_object_terminate(backing_object);
1996 counter_u64_add(object_collapses, 1);
1997 VM_OBJECT_WLOCK(object);
2000 * If we do not entirely shadow the backing object,
2001 * there is nothing we can do so we give up.
2003 * The object lock and backing_object lock must not
2004 * be dropped during this sequence.
2006 if (!vm_object_scan_all_shadowed(object)) {
2007 VM_OBJECT_WUNLOCK(backing_object);
2012 * Make the parent shadow the next object in the
2013 * chain. Deallocating backing_object will not remove
2014 * it, since its reference count is at least 2.
2016 vm_object_backing_remove_locked(object);
2017 new_backing_object = backing_object->backing_object;
2018 if (new_backing_object != NULL) {
2019 vm_object_backing_insert_ref(object,
2020 new_backing_object);
2021 object->backing_object_offset +=
2022 backing_object->backing_object_offset;
2026 * Drop the reference count on backing_object. Since
2027 * its ref_count was at least 2, it will not vanish.
2029 (void)refcount_release(&backing_object->ref_count);
2030 KASSERT(backing_object->ref_count >= 1, (
2031 "backing_object %p was somehow dereferenced during collapse!",
2033 VM_OBJECT_WUNLOCK(backing_object);
2034 counter_u64_add(object_bypasses, 1);
2038 * Try again with this object's new backing object.
2044 * vm_object_page_remove:
2046 * For the given object, either frees or invalidates each of the
2047 * specified pages. In general, a page is freed. However, if a page is
2048 * wired for any reason other than the existence of a managed, wired
2049 * mapping, then it may be invalidated but not removed from the object.
2050 * Pages are specified by the given range ["start", "end") and the option
2051 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
2052 * extends from "start" to the end of the object. If the option
2053 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2054 * specified range are affected. If the option OBJPR_NOTMAPPED is
2055 * specified, then the pages within the specified range must have no
2056 * mappings. Otherwise, if this option is not specified, any mappings to
2057 * the specified pages are removed before the pages are freed or
2060 * In general, this operation should only be performed on objects that
2061 * contain managed pages. There are, however, two exceptions. First, it
2062 * is performed on the kernel and kmem objects by vm_map_entry_delete().
2063 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2064 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
2065 * not be specified and the option OBJPR_NOTMAPPED must be specified.
2067 * The object must be locked.
2070 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2075 VM_OBJECT_ASSERT_WLOCKED(object);
2076 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2077 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2078 ("vm_object_page_remove: illegal options for object %p", object));
2079 if (object->resident_page_count == 0)
2081 vm_object_pip_add(object, 1);
2083 p = vm_page_find_least(object, start);
2086 * Here, the variable "p" is either (1) the page with the least pindex
2087 * greater than or equal to the parameter "start" or (2) NULL.
2089 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2090 next = TAILQ_NEXT(p, listq);
2093 * If the page is wired for any reason besides the existence
2094 * of managed, wired mappings, then it cannot be freed. For
2095 * example, fictitious pages, which represent device memory,
2096 * are inherently wired and cannot be freed. They can,
2097 * however, be invalidated if the option OBJPR_CLEANONLY is
2100 if (vm_page_tryxbusy(p) == 0) {
2101 vm_page_sleep_if_busy(p, "vmopar");
2104 if (vm_page_wired(p)) {
2106 if ((options & OBJPR_NOTMAPPED) == 0 &&
2107 object->ref_count != 0)
2109 if ((options & OBJPR_CLEANONLY) == 0) {
2116 KASSERT((p->flags & PG_FICTITIOUS) == 0,
2117 ("vm_object_page_remove: page %p is fictitious", p));
2118 if ((options & OBJPR_CLEANONLY) != 0 &&
2119 !vm_page_none_valid(p)) {
2120 if ((options & OBJPR_NOTMAPPED) == 0 &&
2121 object->ref_count != 0 &&
2122 !vm_page_try_remove_write(p))
2124 if (p->dirty != 0) {
2129 if ((options & OBJPR_NOTMAPPED) == 0 &&
2130 object->ref_count != 0 && !vm_page_try_remove_all(p))
2134 vm_object_pip_wakeup(object);
2136 vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2141 * vm_object_page_noreuse:
2143 * For the given object, attempt to move the specified pages to
2144 * the head of the inactive queue. This bypasses regular LRU
2145 * operation and allows the pages to be reused quickly under memory
2146 * pressure. If a page is wired for any reason, then it will not
2147 * be queued. Pages are specified by the range ["start", "end").
2148 * As a special case, if "end" is zero, then the range extends from
2149 * "start" to the end of the object.
2151 * This operation should only be performed on objects that
2152 * contain non-fictitious, managed pages.
2154 * The object must be locked.
2157 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2161 VM_OBJECT_ASSERT_LOCKED(object);
2162 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2163 ("vm_object_page_noreuse: illegal object %p", object));
2164 if (object->resident_page_count == 0)
2166 p = vm_page_find_least(object, start);
2169 * Here, the variable "p" is either (1) the page with the least pindex
2170 * greater than or equal to the parameter "start" or (2) NULL.
2172 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2173 next = TAILQ_NEXT(p, listq);
2174 vm_page_deactivate_noreuse(p);
2179 * Populate the specified range of the object with valid pages. Returns
2180 * TRUE if the range is successfully populated and FALSE otherwise.
2182 * Note: This function should be optimized to pass a larger array of
2183 * pages to vm_pager_get_pages() before it is applied to a non-
2184 * OBJT_DEVICE object.
2186 * The object must be locked.
2189 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2195 VM_OBJECT_ASSERT_WLOCKED(object);
2196 for (pindex = start; pindex < end; pindex++) {
2197 rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2198 if (rv != VM_PAGER_OK)
2202 * Keep "m" busy because a subsequent iteration may unlock
2206 if (pindex > start) {
2207 m = vm_page_lookup(object, start);
2208 while (m != NULL && m->pindex < pindex) {
2210 m = TAILQ_NEXT(m, listq);
2213 return (pindex == end);
2217 * Routine: vm_object_coalesce
2218 * Function: Coalesces two objects backing up adjoining
2219 * regions of memory into a single object.
2221 * returns TRUE if objects were combined.
2223 * NOTE: Only works at the moment if the second object is NULL -
2224 * if it's not, which object do we lock first?
2227 * prev_object First object to coalesce
2228 * prev_offset Offset into prev_object
2229 * prev_size Size of reference to prev_object
2230 * next_size Size of reference to the second object
2231 * reserved Indicator that extension region has
2232 * swap accounted for
2235 * The object must *not* be locked.
2238 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2239 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2241 vm_pindex_t next_pindex;
2243 if (prev_object == NULL)
2245 if ((prev_object->flags & OBJ_ANON) == 0)
2248 VM_OBJECT_WLOCK(prev_object);
2250 * Try to collapse the object first.
2252 vm_object_collapse(prev_object);
2255 * Can't coalesce if: . more than one reference . paged out . shadows
2256 * another object . has a copy elsewhere (any of which mean that the
2257 * pages not mapped to prev_entry may be in use anyway)
2259 if (prev_object->backing_object != NULL) {
2260 VM_OBJECT_WUNLOCK(prev_object);
2264 prev_size >>= PAGE_SHIFT;
2265 next_size >>= PAGE_SHIFT;
2266 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2268 if (prev_object->ref_count > 1 &&
2269 prev_object->size != next_pindex &&
2270 (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2271 VM_OBJECT_WUNLOCK(prev_object);
2276 * Account for the charge.
2278 if (prev_object->cred != NULL) {
2280 * If prev_object was charged, then this mapping,
2281 * although not charged now, may become writable
2282 * later. Non-NULL cred in the object would prevent
2283 * swap reservation during enabling of the write
2284 * access, so reserve swap now. Failed reservation
2285 * cause allocation of the separate object for the map
2286 * entry, and swap reservation for this entry is
2287 * managed in appropriate time.
2289 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2290 prev_object->cred)) {
2291 VM_OBJECT_WUNLOCK(prev_object);
2294 prev_object->charge += ptoa(next_size);
2298 * Remove any pages that may still be in the object from a previous
2301 if (next_pindex < prev_object->size) {
2302 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2305 if (prev_object->cred != NULL) {
2306 KASSERT(prev_object->charge >=
2307 ptoa(prev_object->size - next_pindex),
2308 ("object %p overcharged 1 %jx %jx", prev_object,
2309 (uintmax_t)next_pindex, (uintmax_t)next_size));
2310 prev_object->charge -= ptoa(prev_object->size -
2317 * Extend the object if necessary.
2319 if (next_pindex + next_size > prev_object->size)
2320 prev_object->size = next_pindex + next_size;
2322 VM_OBJECT_WUNLOCK(prev_object);
2327 vm_object_set_writeable_dirty_(vm_object_t object)
2329 atomic_add_int(&object->generation, 1);
2333 vm_object_mightbedirty_(vm_object_t object)
2335 return (object->generation != object->cleangeneration);
2341 * For each page offset within the specified range of the given object,
2342 * find the highest-level page in the shadow chain and unwire it. A page
2343 * must exist at every page offset, and the highest-level page must be
2347 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2350 vm_object_t tobject, t1object;
2352 vm_pindex_t end_pindex, pindex, tpindex;
2353 int depth, locked_depth;
2355 KASSERT((offset & PAGE_MASK) == 0,
2356 ("vm_object_unwire: offset is not page aligned"));
2357 KASSERT((length & PAGE_MASK) == 0,
2358 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2359 /* The wired count of a fictitious page never changes. */
2360 if ((object->flags & OBJ_FICTITIOUS) != 0)
2362 pindex = OFF_TO_IDX(offset);
2363 end_pindex = pindex + atop(length);
2366 VM_OBJECT_RLOCK(object);
2367 m = vm_page_find_least(object, pindex);
2368 while (pindex < end_pindex) {
2369 if (m == NULL || pindex < m->pindex) {
2371 * The first object in the shadow chain doesn't
2372 * contain a page at the current index. Therefore,
2373 * the page must exist in a backing object.
2380 OFF_TO_IDX(tobject->backing_object_offset);
2381 tobject = tobject->backing_object;
2382 KASSERT(tobject != NULL,
2383 ("vm_object_unwire: missing page"));
2384 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2387 if (depth == locked_depth) {
2389 VM_OBJECT_RLOCK(tobject);
2391 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2395 m = TAILQ_NEXT(m, listq);
2397 if (vm_page_trysbusy(tm) == 0) {
2398 for (tobject = object; locked_depth >= 1;
2400 t1object = tobject->backing_object;
2401 if (tm->object != tobject)
2402 VM_OBJECT_RUNLOCK(tobject);
2405 vm_page_busy_sleep(tm, "unwbo", true);
2408 vm_page_unwire(tm, queue);
2409 vm_page_sunbusy(tm);
2413 /* Release the accumulated object locks. */
2414 for (tobject = object; locked_depth >= 1; locked_depth--) {
2415 t1object = tobject->backing_object;
2416 VM_OBJECT_RUNLOCK(tobject);
2422 * Return the vnode for the given object, or NULL if none exists.
2423 * For tmpfs objects, the function may return NULL if there is
2424 * no vnode allocated at the time of the call.
2427 vm_object_vnode(vm_object_t object)
2431 VM_OBJECT_ASSERT_LOCKED(object);
2432 if (object->type == OBJT_VNODE) {
2433 vp = object->handle;
2434 KASSERT(vp != NULL, ("%s: OBJT_VNODE has no vnode", __func__));
2435 } else if (object->type == OBJT_SWAP &&
2436 (object->flags & OBJ_TMPFS) != 0) {
2437 vp = object->un_pager.swp.swp_tmpfs;
2438 KASSERT(vp != NULL, ("%s: OBJT_TMPFS has no vnode", __func__));
2446 * Busy the vm object. This prevents new pages belonging to the object from
2447 * becoming busy. Existing pages persist as busy. Callers are responsible
2448 * for checking page state before proceeding.
2451 vm_object_busy(vm_object_t obj)
2454 VM_OBJECT_ASSERT_LOCKED(obj);
2456 blockcount_acquire(&obj->busy, 1);
2457 /* The fence is required to order loads of page busy. */
2458 atomic_thread_fence_acq_rel();
2462 vm_object_unbusy(vm_object_t obj)
2465 blockcount_release(&obj->busy, 1);
2469 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2472 VM_OBJECT_ASSERT_UNLOCKED(obj);
2474 (void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2478 * Return the kvme type of the given object.
2479 * If vpp is not NULL, set it to the object's vm_object_vnode() or NULL.
2482 vm_object_kvme_type(vm_object_t object, struct vnode **vpp)
2485 VM_OBJECT_ASSERT_LOCKED(object);
2487 *vpp = vm_object_vnode(object);
2488 switch (object->type) {
2490 return (KVME_TYPE_DEFAULT);
2492 return (KVME_TYPE_VNODE);
2494 if ((object->flags & OBJ_TMPFS_NODE) != 0)
2495 return (KVME_TYPE_VNODE);
2496 return (KVME_TYPE_SWAP);
2498 return (KVME_TYPE_DEVICE);
2500 return (KVME_TYPE_PHYS);
2502 return (KVME_TYPE_DEAD);
2504 return (KVME_TYPE_SG);
2505 case OBJT_MGTDEVICE:
2506 return (KVME_TYPE_MGTDEVICE);
2508 return (KVME_TYPE_UNKNOWN);
2513 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2515 struct kinfo_vmobject *kvo;
2516 char *fullpath, *freepath;
2524 if (req->oldptr == NULL) {
2526 * If an old buffer has not been provided, generate an
2527 * estimate of the space needed for a subsequent call.
2529 mtx_lock(&vm_object_list_mtx);
2531 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2532 if (obj->type == OBJT_DEAD)
2536 mtx_unlock(&vm_object_list_mtx);
2537 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2541 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK);
2545 * VM objects are type stable and are never removed from the
2546 * list once added. This allows us to safely read obj->object_list
2547 * after reacquiring the VM object lock.
2549 mtx_lock(&vm_object_list_mtx);
2550 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2551 if (obj->type == OBJT_DEAD)
2553 VM_OBJECT_RLOCK(obj);
2554 if (obj->type == OBJT_DEAD) {
2555 VM_OBJECT_RUNLOCK(obj);
2558 mtx_unlock(&vm_object_list_mtx);
2559 kvo->kvo_size = ptoa(obj->size);
2560 kvo->kvo_resident = obj->resident_page_count;
2561 kvo->kvo_ref_count = obj->ref_count;
2562 kvo->kvo_shadow_count = obj->shadow_count;
2563 kvo->kvo_memattr = obj->memattr;
2564 kvo->kvo_active = 0;
2565 kvo->kvo_inactive = 0;
2566 TAILQ_FOREACH(m, &obj->memq, listq) {
2568 * A page may belong to the object but be
2569 * dequeued and set to PQ_NONE while the
2570 * object lock is not held. This makes the
2571 * reads of m->queue below racy, and we do not
2572 * count pages set to PQ_NONE. However, this
2573 * sysctl is only meant to give an
2574 * approximation of the system anyway.
2576 if (m->a.queue == PQ_ACTIVE)
2578 else if (m->a.queue == PQ_INACTIVE)
2579 kvo->kvo_inactive++;
2582 kvo->kvo_vn_fileid = 0;
2583 kvo->kvo_vn_fsid = 0;
2584 kvo->kvo_vn_fsid_freebsd11 = 0;
2587 kvo->kvo_type = vm_object_kvme_type(obj, &vp);
2590 } else if ((obj->flags & OBJ_ANON) != 0) {
2591 MPASS(kvo->kvo_type == KVME_TYPE_DEFAULT ||
2592 kvo->kvo_type == KVME_TYPE_SWAP);
2593 kvo->kvo_me = (uintptr_t)obj;
2594 /* tmpfs objs are reported as vnodes */
2595 kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2596 sp = swap_pager_swapped_pages(obj);
2597 kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2599 VM_OBJECT_RUNLOCK(obj);
2601 vn_fullpath(vp, &fullpath, &freepath);
2602 vn_lock(vp, LK_SHARED | LK_RETRY);
2603 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2604 kvo->kvo_vn_fileid = va.va_fileid;
2605 kvo->kvo_vn_fsid = va.va_fsid;
2606 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2612 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2613 if (freepath != NULL)
2614 free(freepath, M_TEMP);
2616 /* Pack record size down */
2617 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2618 + strlen(kvo->kvo_path) + 1;
2619 kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2621 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2622 mtx_lock(&vm_object_list_mtx);
2626 mtx_unlock(&vm_object_list_mtx);
2630 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2631 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2632 "List of VM objects");
2634 #include "opt_ddb.h"
2636 #include <sys/kernel.h>
2638 #include <sys/cons.h>
2640 #include <ddb/ddb.h>
2643 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2646 vm_map_entry_t tmpe;
2653 VM_MAP_ENTRY_FOREACH(tmpe, map) {
2654 if (_vm_object_in_map(map, object, tmpe)) {
2658 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2659 tmpm = entry->object.sub_map;
2660 VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2661 if (_vm_object_in_map(tmpm, object, tmpe)) {
2665 } else if ((obj = entry->object.vm_object) != NULL) {
2666 for (; obj; obj = obj->backing_object)
2667 if (obj == object) {
2675 vm_object_in_map(vm_object_t object)
2679 /* sx_slock(&allproc_lock); */
2680 FOREACH_PROC_IN_SYSTEM(p) {
2681 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2683 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2684 /* sx_sunlock(&allproc_lock); */
2688 /* sx_sunlock(&allproc_lock); */
2689 if (_vm_object_in_map(kernel_map, object, 0))
2694 DB_SHOW_COMMAND(vmochk, vm_object_check)
2699 * make sure that internal objs are in a map somewhere
2700 * and none have zero ref counts.
2702 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2703 if ((object->flags & OBJ_ANON) != 0) {
2704 if (object->ref_count == 0) {
2705 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2706 (long)object->size);
2708 if (!vm_object_in_map(object)) {
2710 "vmochk: internal obj is not in a map: "
2711 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2712 object->ref_count, (u_long)object->size,
2713 (u_long)object->size,
2714 (void *)object->backing_object);
2723 * vm_object_print: [ debug ]
2725 DB_SHOW_COMMAND(object, vm_object_print_static)
2727 /* XXX convert args. */
2728 vm_object_t object = (vm_object_t)addr;
2729 boolean_t full = have_addr;
2733 /* XXX count is an (unused) arg. Avoid shadowing it. */
2734 #define count was_count
2742 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2743 object, (int)object->type, (uintmax_t)object->size,
2744 object->resident_page_count, object->ref_count, object->flags,
2745 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2746 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2747 object->shadow_count,
2748 object->backing_object ? object->backing_object->ref_count : 0,
2749 object->backing_object, (uintmax_t)object->backing_object_offset);
2756 TAILQ_FOREACH(p, &object->memq, listq) {
2758 db_iprintf("memory:=");
2759 else if (count == 6) {
2767 db_printf("(off=0x%jx,page=0x%jx)",
2768 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2781 /* XXX need this non-static entry for calling from vm_map_print. */
2784 /* db_expr_t */ long addr,
2785 boolean_t have_addr,
2786 /* db_expr_t */ long count,
2789 vm_object_print_static(addr, have_addr, count, modif);
2792 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2797 vm_page_t m, prev_m;
2801 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2802 db_printf("new object: %p\n", (void *)object);
2813 TAILQ_FOREACH(m, &object->memq, listq) {
2814 if (m->pindex > 128)
2816 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2817 prev_m->pindex + 1 != m->pindex) {
2819 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2820 (long)fidx, rcount, (long)pa);
2832 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2837 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2838 (long)fidx, rcount, (long)pa);
2848 pa = VM_PAGE_TO_PHYS(m);
2852 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2853 (long)fidx, rcount, (long)pa);