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(atomic_load_int(&object->shadow_count) == 0,
202 ("object %p shadow_count = %d",
203 object, atomic_load_int(&object->shadow_count)));
204 KASSERT(object->type == OBJT_DEAD,
205 ("object %p has non-dead type %d",
206 object, object->type));
207 KASSERT(object->charge == 0 && object->cred == NULL,
208 ("object %p has non-zero charge %ju (%p)",
209 object, (uintmax_t)object->charge, object->cred));
214 vm_object_zinit(void *mem, int size, int flags)
218 object = (vm_object_t)mem;
219 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
221 /* These are true for any object that has been freed */
222 object->type = OBJT_DEAD;
223 vm_radix_init(&object->rtree);
224 refcount_init(&object->ref_count, 0);
225 blockcount_init(&object->paging_in_progress);
226 blockcount_init(&object->busy);
227 object->resident_page_count = 0;
228 atomic_store_int(&object->shadow_count, 0);
229 object->flags = OBJ_DEAD;
231 mtx_lock(&vm_object_list_mtx);
232 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
233 mtx_unlock(&vm_object_list_mtx);
238 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
239 vm_object_t object, void *handle)
242 TAILQ_INIT(&object->memq);
243 LIST_INIT(&object->shadow_head);
246 object->flags = flags;
247 if ((flags & OBJ_SWAP) != 0)
248 pctrie_init(&object->un_pager.swp.swp_blks);
251 * Ensure that swap_pager_swapoff() iteration over object_list
252 * sees up to date type and pctrie head if it observed
255 atomic_thread_fence_rel();
257 object->pg_color = 0;
259 object->domain.dr_policy = NULL;
260 object->generation = 1;
261 object->cleangeneration = 1;
262 refcount_init(&object->ref_count, 1);
263 object->memattr = VM_MEMATTR_DEFAULT;
266 object->handle = handle;
267 object->backing_object = NULL;
268 object->backing_object_offset = (vm_ooffset_t) 0;
269 #if VM_NRESERVLEVEL > 0
270 LIST_INIT(&object->rvq);
272 umtx_shm_object_init(object);
278 * Initialize the VM objects module.
283 TAILQ_INIT(&vm_object_list);
284 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
286 rw_init(&kernel_object->lock, "kernel vm object");
287 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
288 VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
289 #if VM_NRESERVLEVEL > 0
290 kernel_object->flags |= OBJ_COLORED;
291 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
293 kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
296 * The lock portion of struct vm_object must be type stable due
297 * to vm_pageout_fallback_object_lock locking a vm object
298 * without holding any references to it.
300 * paging_in_progress is valid always. Lockless references to
301 * the objects may acquire pip and then check OBJ_DEAD.
303 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
309 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
315 vm_object_clear_flag(vm_object_t object, u_short bits)
318 VM_OBJECT_ASSERT_WLOCKED(object);
319 object->flags &= ~bits;
323 * Sets the default memory attribute for the specified object. Pages
324 * that are allocated to this object are by default assigned this memory
327 * Presently, this function must be called before any pages are allocated
328 * to the object. In the future, this requirement may be relaxed for
329 * "default" and "swap" objects.
332 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
335 VM_OBJECT_ASSERT_WLOCKED(object);
337 if (object->type == OBJT_DEAD)
338 return (KERN_INVALID_ARGUMENT);
339 if (!TAILQ_EMPTY(&object->memq))
340 return (KERN_FAILURE);
342 object->memattr = memattr;
343 return (KERN_SUCCESS);
347 vm_object_pip_add(vm_object_t object, short i)
351 blockcount_acquire(&object->paging_in_progress, i);
355 vm_object_pip_wakeup(vm_object_t object)
358 vm_object_pip_wakeupn(object, 1);
362 vm_object_pip_wakeupn(vm_object_t object, short i)
366 blockcount_release(&object->paging_in_progress, i);
370 * Atomically drop the object lock and wait for pip to drain. This protects
371 * from sleep/wakeup races due to identity changes. The lock is not re-acquired
375 vm_object_pip_sleep(vm_object_t object, const char *waitid)
378 (void)blockcount_sleep(&object->paging_in_progress, &object->lock,
379 waitid, PVM | PDROP);
383 vm_object_pip_wait(vm_object_t object, const char *waitid)
386 VM_OBJECT_ASSERT_WLOCKED(object);
388 blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
393 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
396 VM_OBJECT_ASSERT_UNLOCKED(object);
398 blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
402 * vm_object_allocate:
404 * Returns a new object with the given size.
407 vm_object_allocate(objtype_t type, vm_pindex_t size)
414 panic("vm_object_allocate: can't create OBJT_DEAD");
419 flags = OBJ_COLORED | OBJ_SWAP;
423 flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
426 flags = OBJ_FICTITIOUS;
429 flags = OBJ_UNMANAGED;
435 panic("vm_object_allocate: type %d is undefined or dynamic",
438 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
439 _vm_object_allocate(type, size, flags, object, NULL);
445 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags)
449 MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */);
450 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
451 _vm_object_allocate(dyntype, size, flags, object, NULL);
457 * vm_object_allocate_anon:
459 * Returns a new default object of the given size and marked as
460 * anonymous memory for special split/collapse handling. Color
461 * to be initialized by the caller.
464 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
465 struct ucred *cred, vm_size_t charge)
467 vm_object_t handle, object;
469 if (backing_object == NULL)
471 else if ((backing_object->flags & OBJ_ANON) != 0)
472 handle = backing_object->handle;
474 handle = backing_object;
475 object = uma_zalloc(obj_zone, M_WAITOK);
476 _vm_object_allocate(OBJT_DEFAULT, size, OBJ_ANON | OBJ_ONEMAPPING,
479 object->charge = cred != NULL ? charge : 0;
484 vm_object_reference_vnode(vm_object_t object)
489 * vnode objects need the lock for the first reference
490 * to serialize with vnode_object_deallocate().
492 if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
493 VM_OBJECT_RLOCK(object);
494 old = refcount_acquire(&object->ref_count);
495 if (object->type == OBJT_VNODE && old == 0)
496 vref(object->handle);
497 VM_OBJECT_RUNLOCK(object);
502 * vm_object_reference:
504 * Acquires a reference to the given object.
507 vm_object_reference(vm_object_t object)
513 if (object->type == OBJT_VNODE)
514 vm_object_reference_vnode(object);
516 refcount_acquire(&object->ref_count);
517 KASSERT((object->flags & OBJ_DEAD) == 0,
518 ("vm_object_reference: Referenced dead object."));
522 * vm_object_reference_locked:
524 * Gets another reference to the given object.
526 * The object must be locked.
529 vm_object_reference_locked(vm_object_t object)
533 VM_OBJECT_ASSERT_LOCKED(object);
534 old = refcount_acquire(&object->ref_count);
535 if (object->type == OBJT_VNODE && old == 0)
536 vref(object->handle);
537 KASSERT((object->flags & OBJ_DEAD) == 0,
538 ("vm_object_reference: Referenced dead object."));
542 * Handle deallocating an object of type OBJT_VNODE.
545 vm_object_deallocate_vnode(vm_object_t object)
547 struct vnode *vp = (struct vnode *) object->handle;
550 KASSERT(object->type == OBJT_VNODE,
551 ("vm_object_deallocate_vnode: not a vnode object"));
552 KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
554 /* Object lock to protect handle lookup. */
555 last = refcount_release(&object->ref_count);
556 VM_OBJECT_RUNLOCK(object);
561 if (!umtx_shm_vnobj_persistent)
562 umtx_shm_object_terminated(object);
564 /* vrele may need the vnode lock. */
569 * We dropped a reference on an object and discovered that it had a
570 * single remaining shadow. This is a sibling of the reference we
571 * dropped. Attempt to collapse the sibling and backing object.
574 vm_object_deallocate_anon(vm_object_t backing_object)
578 /* Fetch the final shadow. */
579 object = LIST_FIRST(&backing_object->shadow_head);
580 KASSERT(object != NULL &&
581 atomic_load_int(&backing_object->shadow_count) == 1,
582 ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
583 backing_object->ref_count,
584 atomic_load_int(&backing_object->shadow_count)));
585 KASSERT((object->flags & OBJ_ANON) != 0,
586 ("invalid shadow object %p", object));
588 if (!VM_OBJECT_TRYWLOCK(object)) {
590 * Prevent object from disappearing since we do not have a
593 vm_object_pip_add(object, 1);
594 VM_OBJECT_WUNLOCK(backing_object);
595 VM_OBJECT_WLOCK(object);
596 vm_object_pip_wakeup(object);
598 VM_OBJECT_WUNLOCK(backing_object);
601 * Check for a collapse/terminate race with the last reference holder.
603 if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
604 !refcount_acquire_if_not_zero(&object->ref_count)) {
605 VM_OBJECT_WUNLOCK(object);
608 backing_object = object->backing_object;
609 if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
610 vm_object_collapse(object);
611 VM_OBJECT_WUNLOCK(object);
617 * vm_object_deallocate:
619 * Release a reference to the specified object,
620 * gained either through a vm_object_allocate
621 * or a vm_object_reference call. When all references
622 * are gone, storage associated with this object
623 * may be relinquished.
625 * No object may be locked.
628 vm_object_deallocate(vm_object_t object)
633 while (object != NULL) {
635 * If the reference count goes to 0 we start calling
636 * vm_object_terminate() on the object chain. A ref count
637 * of 1 may be a special case depending on the shadow count
638 * being 0 or 1. These cases require a write lock on the
641 if ((object->flags & OBJ_ANON) == 0)
642 released = refcount_release_if_gt(&object->ref_count, 1);
644 released = refcount_release_if_gt(&object->ref_count, 2);
648 if (object->type == OBJT_VNODE) {
649 VM_OBJECT_RLOCK(object);
650 if (object->type == OBJT_VNODE) {
651 vm_object_deallocate_vnode(object);
654 VM_OBJECT_RUNLOCK(object);
657 VM_OBJECT_WLOCK(object);
658 KASSERT(object->ref_count > 0,
659 ("vm_object_deallocate: object deallocated too many times: %d",
663 * If this is not the final reference to an anonymous
664 * object we may need to collapse the shadow chain.
666 if (!refcount_release(&object->ref_count)) {
667 if (object->ref_count > 1 ||
668 atomic_load_int(&object->shadow_count) == 0) {
669 if ((object->flags & OBJ_ANON) != 0 &&
670 object->ref_count == 1)
671 vm_object_set_flag(object,
673 VM_OBJECT_WUNLOCK(object);
677 /* Handle collapsing last ref on anonymous objects. */
678 object = vm_object_deallocate_anon(object);
683 * Handle the final reference to an object. We restart
684 * the loop with the backing object to avoid recursion.
686 umtx_shm_object_terminated(object);
687 temp = object->backing_object;
689 KASSERT(object->type == OBJT_DEFAULT ||
690 object->type == OBJT_SWAP,
691 ("shadowed tmpfs v_object 2 %p", object));
692 vm_object_backing_remove(object);
695 KASSERT((object->flags & OBJ_DEAD) == 0,
696 ("vm_object_deallocate: Terminating dead object."));
697 vm_object_set_flag(object, OBJ_DEAD);
698 vm_object_terminate(object);
704 * vm_object_destroy removes the object from the global object list
705 * and frees the space for the object.
708 vm_object_destroy(vm_object_t object)
712 * Release the allocation charge.
714 if (object->cred != NULL) {
715 swap_release_by_cred(object->charge, object->cred);
717 crfree(object->cred);
722 * Free the space for the object.
724 uma_zfree(obj_zone, object);
728 vm_object_sub_shadow(vm_object_t object)
730 KASSERT(object->shadow_count >= 1,
731 ("object %p sub_shadow count zero", object));
732 atomic_subtract_int(&object->shadow_count, 1);
736 vm_object_backing_remove_locked(vm_object_t object)
738 vm_object_t backing_object;
740 backing_object = object->backing_object;
741 VM_OBJECT_ASSERT_WLOCKED(object);
742 VM_OBJECT_ASSERT_WLOCKED(backing_object);
744 KASSERT((object->flags & OBJ_COLLAPSING) == 0,
745 ("vm_object_backing_remove: Removing collapsing object."));
747 vm_object_sub_shadow(backing_object);
748 if ((object->flags & OBJ_SHADOWLIST) != 0) {
749 LIST_REMOVE(object, shadow_list);
750 vm_object_clear_flag(object, OBJ_SHADOWLIST);
752 object->backing_object = NULL;
756 vm_object_backing_remove(vm_object_t object)
758 vm_object_t backing_object;
760 VM_OBJECT_ASSERT_WLOCKED(object);
762 backing_object = object->backing_object;
763 if ((object->flags & OBJ_SHADOWLIST) != 0) {
764 VM_OBJECT_WLOCK(backing_object);
765 vm_object_backing_remove_locked(object);
766 VM_OBJECT_WUNLOCK(backing_object);
768 object->backing_object = NULL;
769 vm_object_sub_shadow(backing_object);
774 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
777 VM_OBJECT_ASSERT_WLOCKED(object);
779 atomic_add_int(&backing_object->shadow_count, 1);
780 if ((backing_object->flags & OBJ_ANON) != 0) {
781 VM_OBJECT_ASSERT_WLOCKED(backing_object);
782 LIST_INSERT_HEAD(&backing_object->shadow_head, object,
784 vm_object_set_flag(object, OBJ_SHADOWLIST);
786 object->backing_object = backing_object;
790 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
793 VM_OBJECT_ASSERT_WLOCKED(object);
795 if ((backing_object->flags & OBJ_ANON) != 0) {
796 VM_OBJECT_WLOCK(backing_object);
797 vm_object_backing_insert_locked(object, backing_object);
798 VM_OBJECT_WUNLOCK(backing_object);
800 object->backing_object = backing_object;
801 atomic_add_int(&backing_object->shadow_count, 1);
806 * Insert an object into a backing_object's shadow list with an additional
807 * reference to the backing_object added.
810 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
813 VM_OBJECT_ASSERT_WLOCKED(object);
815 if ((backing_object->flags & OBJ_ANON) != 0) {
816 VM_OBJECT_WLOCK(backing_object);
817 KASSERT((backing_object->flags & OBJ_DEAD) == 0,
818 ("shadowing dead anonymous object"));
819 vm_object_reference_locked(backing_object);
820 vm_object_backing_insert_locked(object, backing_object);
821 vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
822 VM_OBJECT_WUNLOCK(backing_object);
824 vm_object_reference(backing_object);
825 atomic_add_int(&backing_object->shadow_count, 1);
826 object->backing_object = backing_object;
831 * Transfer a backing reference from backing_object to object.
834 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
836 vm_object_t new_backing_object;
839 * Note that the reference to backing_object->backing_object
840 * moves from within backing_object to within object.
842 vm_object_backing_remove_locked(object);
843 new_backing_object = backing_object->backing_object;
844 if (new_backing_object == NULL)
846 if ((new_backing_object->flags & OBJ_ANON) != 0) {
847 VM_OBJECT_WLOCK(new_backing_object);
848 vm_object_backing_remove_locked(backing_object);
849 vm_object_backing_insert_locked(object, new_backing_object);
850 VM_OBJECT_WUNLOCK(new_backing_object);
853 * shadow_count for new_backing_object is left
854 * unchanged, its reference provided by backing_object
855 * is replaced by object.
857 object->backing_object = new_backing_object;
858 backing_object->backing_object = NULL;
863 * Wait for a concurrent collapse to settle.
866 vm_object_collapse_wait(vm_object_t object)
869 VM_OBJECT_ASSERT_WLOCKED(object);
871 while ((object->flags & OBJ_COLLAPSING) != 0) {
872 vm_object_pip_wait(object, "vmcolwait");
873 counter_u64_add(object_collapse_waits, 1);
878 * Waits for a backing object to clear a pending collapse and returns
879 * it locked if it is an ANON object.
882 vm_object_backing_collapse_wait(vm_object_t object)
884 vm_object_t backing_object;
886 VM_OBJECT_ASSERT_WLOCKED(object);
889 backing_object = object->backing_object;
890 if (backing_object == NULL ||
891 (backing_object->flags & OBJ_ANON) == 0)
893 VM_OBJECT_WLOCK(backing_object);
894 if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
896 VM_OBJECT_WUNLOCK(object);
897 vm_object_pip_sleep(backing_object, "vmbckwait");
898 counter_u64_add(object_collapse_waits, 1);
899 VM_OBJECT_WLOCK(object);
901 return (backing_object);
905 * vm_object_terminate_pages removes any remaining pageable pages
906 * from the object and resets the object to an empty state.
909 vm_object_terminate_pages(vm_object_t object)
913 VM_OBJECT_ASSERT_WLOCKED(object);
916 * Free any remaining pageable pages. This also removes them from the
917 * paging queues. However, don't free wired pages, just remove them
918 * from the object. Rather than incrementally removing each page from
919 * the object, the page and object are reset to any empty state.
921 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
922 vm_page_assert_unbusied(p);
923 KASSERT(p->object == object &&
924 (p->ref_count & VPRC_OBJREF) != 0,
925 ("vm_object_terminate_pages: page %p is inconsistent", p));
928 if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
935 * If the object contained any pages, then reset it to an empty state.
936 * None of the object's fields, including "resident_page_count", were
937 * modified by the preceding loop.
939 if (object->resident_page_count != 0) {
940 vm_radix_reclaim_allnodes(&object->rtree);
941 TAILQ_INIT(&object->memq);
942 object->resident_page_count = 0;
943 if (object->type == OBJT_VNODE)
944 vdrop(object->handle);
949 * vm_object_terminate actually destroys the specified object, freeing
950 * up all previously used resources.
952 * The object must be locked.
953 * This routine may block.
956 vm_object_terminate(vm_object_t object)
959 VM_OBJECT_ASSERT_WLOCKED(object);
960 KASSERT((object->flags & OBJ_DEAD) != 0,
961 ("terminating non-dead obj %p", object));
962 KASSERT((object->flags & OBJ_COLLAPSING) == 0,
963 ("terminating collapsing obj %p", object));
964 KASSERT(object->backing_object == NULL,
965 ("terminating shadow obj %p", object));
968 * Wait for the pageout daemon and other current users to be
969 * done with the object. Note that new paging_in_progress
970 * users can come after this wait, but they must check
971 * OBJ_DEAD flag set (without unlocking the object), and avoid
972 * the object being terminated.
974 vm_object_pip_wait(object, "objtrm");
976 KASSERT(object->ref_count == 0,
977 ("vm_object_terminate: object with references, ref_count=%d",
980 if ((object->flags & OBJ_PG_DTOR) == 0)
981 vm_object_terminate_pages(object);
983 #if VM_NRESERVLEVEL > 0
984 if (__predict_false(!LIST_EMPTY(&object->rvq)))
985 vm_reserv_break_all(object);
988 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
989 (object->flags & OBJ_SWAP) != 0,
990 ("%s: non-swap obj %p has cred", __func__, object));
993 * Let the pager know object is dead.
995 vm_pager_deallocate(object);
996 VM_OBJECT_WUNLOCK(object);
998 vm_object_destroy(object);
1002 * Make the page read-only so that we can clear the object flags. However, if
1003 * this is a nosync mmap then the object is likely to stay dirty so do not
1004 * mess with the page and do not clear the object flags. Returns TRUE if the
1005 * page should be flushed, and FALSE otherwise.
1008 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
1011 vm_page_assert_busied(p);
1014 * If we have been asked to skip nosync pages and this is a
1015 * nosync page, skip it. Note that the object flags were not
1016 * cleared in this case so we do not have to set them.
1018 if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
1022 pmap_remove_write(p);
1023 return (p->dirty != 0);
1028 * vm_object_page_clean
1030 * Clean all dirty pages in the specified range of object. Leaves page
1031 * on whatever queue it is currently on. If NOSYNC is set then do not
1032 * write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1033 * leaving the object dirty.
1035 * For swap objects backing tmpfs regular files, do not flush anything,
1036 * but remove write protection on the mapped pages to update mtime through
1039 * When stuffing pages asynchronously, allow clustering. XXX we need a
1040 * synchronous clustering mode implementation.
1042 * Odd semantics: if start == end, we clean everything.
1044 * The object must be locked.
1046 * Returns FALSE if some page from the range was not written, as
1047 * reported by the pager, and TRUE otherwise.
1050 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1054 vm_pindex_t pi, tend, tstart;
1055 int curgeneration, n, pagerflags;
1056 boolean_t eio, res, allclean;
1058 VM_OBJECT_ASSERT_WLOCKED(object);
1060 if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1063 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1064 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1065 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1067 tstart = OFF_TO_IDX(start);
1068 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1069 allclean = tstart == 0 && tend >= object->size;
1073 curgeneration = object->generation;
1075 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1079 np = TAILQ_NEXT(p, listq);
1080 if (vm_page_none_valid(p))
1082 if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1083 if (object->generation != curgeneration &&
1084 (flags & OBJPC_SYNC) != 0)
1086 np = vm_page_find_least(object, pi);
1089 if (!vm_object_page_remove_write(p, flags, &allclean)) {
1093 if (object->type == OBJT_VNODE) {
1094 n = vm_object_page_collect_flush(object, p, pagerflags,
1095 flags, &allclean, &eio);
1100 if (object->generation != curgeneration &&
1101 (flags & OBJPC_SYNC) != 0)
1105 * If the VOP_PUTPAGES() did a truncated write, so
1106 * that even the first page of the run is not fully
1107 * written, vm_pageout_flush() returns 0 as the run
1108 * length. Since the condition that caused truncated
1109 * write may be permanent, e.g. exhausted free space,
1110 * accepting n == 0 would cause an infinite loop.
1112 * Forwarding the iterator leaves the unwritten page
1113 * behind, but there is not much we can do there if
1114 * filesystem refuses to write it.
1124 np = vm_page_find_least(object, pi + n);
1127 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1131 * Leave updating cleangeneration for tmpfs objects to tmpfs
1132 * scan. It needs to update mtime, which happens for other
1133 * filesystems during page writeouts.
1135 if (allclean && object->type == OBJT_VNODE)
1136 object->cleangeneration = curgeneration;
1141 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1142 int flags, boolean_t *allclean, boolean_t *eio)
1144 vm_page_t ma[vm_pageout_page_count], p_first, tp;
1145 int count, i, mreq, runlen;
1147 vm_page_lock_assert(p, MA_NOTOWNED);
1148 vm_page_assert_xbusied(p);
1149 VM_OBJECT_ASSERT_WLOCKED(object);
1154 for (tp = p; count < vm_pageout_page_count; count++) {
1155 tp = vm_page_next(tp);
1156 if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1158 if (!vm_object_page_remove_write(tp, flags, allclean)) {
1159 vm_page_xunbusy(tp);
1164 for (p_first = p; count < vm_pageout_page_count; count++) {
1165 tp = vm_page_prev(p_first);
1166 if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1168 if (!vm_object_page_remove_write(tp, flags, allclean)) {
1169 vm_page_xunbusy(tp);
1176 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1179 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1184 * Note that there is absolutely no sense in writing out
1185 * anonymous objects, so we track down the vnode object
1187 * We invalidate (remove) all pages from the address space
1188 * for semantic correctness.
1190 * If the backing object is a device object with unmanaged pages, then any
1191 * mappings to the specified range of pages must be removed before this
1192 * function is called.
1194 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1195 * may start out with a NULL object.
1198 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1199 boolean_t syncio, boolean_t invalidate)
1201 vm_object_t backing_object;
1204 int error, flags, fsync_after;
1211 VM_OBJECT_WLOCK(object);
1212 while ((backing_object = object->backing_object) != NULL) {
1213 VM_OBJECT_WLOCK(backing_object);
1214 offset += object->backing_object_offset;
1215 VM_OBJECT_WUNLOCK(object);
1216 object = backing_object;
1217 if (object->size < OFF_TO_IDX(offset + size))
1218 size = IDX_TO_OFF(object->size) - offset;
1221 * Flush pages if writing is allowed, invalidate them
1222 * if invalidation requested. Pages undergoing I/O
1223 * will be ignored by vm_object_page_remove().
1225 * We cannot lock the vnode and then wait for paging
1226 * to complete without deadlocking against vm_fault.
1227 * Instead we simply call vm_object_page_remove() and
1228 * allow it to block internally on a page-by-page
1229 * basis when it encounters pages undergoing async
1232 if (object->type == OBJT_VNODE &&
1233 vm_object_mightbedirty(object) != 0 &&
1234 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1235 VM_OBJECT_WUNLOCK(object);
1236 (void)vn_start_write(vp, &mp, V_WAIT);
1237 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1238 if (syncio && !invalidate && offset == 0 &&
1239 atop(size) == object->size) {
1241 * If syncing the whole mapping of the file,
1242 * it is faster to schedule all the writes in
1243 * async mode, also allowing the clustering,
1244 * and then wait for i/o to complete.
1249 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1250 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1251 fsync_after = FALSE;
1253 VM_OBJECT_WLOCK(object);
1254 res = vm_object_page_clean(object, offset, offset + size,
1256 VM_OBJECT_WUNLOCK(object);
1259 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1260 if (error != ERELOOKUP)
1264 * Allow SU/bufdaemon to handle more
1265 * dependencies in the meantime.
1268 vn_finished_write(mp);
1270 (void)vn_start_write(vp, &mp, V_WAIT);
1271 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1275 vn_finished_write(mp);
1278 VM_OBJECT_WLOCK(object);
1280 if ((object->type == OBJT_VNODE ||
1281 object->type == OBJT_DEVICE) && invalidate) {
1282 if (object->type == OBJT_DEVICE)
1284 * The option OBJPR_NOTMAPPED must be passed here
1285 * because vm_object_page_remove() cannot remove
1286 * unmanaged mappings.
1288 flags = OBJPR_NOTMAPPED;
1292 flags = OBJPR_CLEANONLY;
1293 vm_object_page_remove(object, OFF_TO_IDX(offset),
1294 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1296 VM_OBJECT_WUNLOCK(object);
1301 * Determine whether the given advice can be applied to the object. Advice is
1302 * not applied to unmanaged pages since they never belong to page queues, and
1303 * since MADV_FREE is destructive, it can apply only to anonymous pages that
1304 * have been mapped at most once.
1307 vm_object_advice_applies(vm_object_t object, int advice)
1310 if ((object->flags & OBJ_UNMANAGED) != 0)
1312 if (advice != MADV_FREE)
1314 return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1315 (OBJ_ONEMAPPING | OBJ_ANON));
1319 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1323 if (advice == MADV_FREE)
1324 vm_pager_freespace(object, pindex, size);
1328 * vm_object_madvise:
1330 * Implements the madvise function at the object/page level.
1332 * MADV_WILLNEED (any object)
1334 * Activate the specified pages if they are resident.
1336 * MADV_DONTNEED (any object)
1338 * Deactivate the specified pages if they are resident.
1340 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1341 * OBJ_ONEMAPPING only)
1343 * Deactivate and clean the specified pages if they are
1344 * resident. This permits the process to reuse the pages
1345 * without faulting or the kernel to reclaim the pages
1349 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1352 vm_pindex_t tpindex;
1353 vm_object_t backing_object, tobject;
1360 VM_OBJECT_WLOCK(object);
1361 if (!vm_object_advice_applies(object, advice)) {
1362 VM_OBJECT_WUNLOCK(object);
1365 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1369 * If the next page isn't resident in the top-level object, we
1370 * need to search the shadow chain. When applying MADV_FREE, we
1371 * take care to release any swap space used to store
1372 * non-resident pages.
1374 if (m == NULL || pindex < m->pindex) {
1376 * Optimize a common case: if the top-level object has
1377 * no backing object, we can skip over the non-resident
1378 * range in constant time.
1380 if (object->backing_object == NULL) {
1381 tpindex = (m != NULL && m->pindex < end) ?
1383 vm_object_madvise_freespace(object, advice,
1384 pindex, tpindex - pindex);
1385 if ((pindex = tpindex) == end)
1392 vm_object_madvise_freespace(tobject, advice,
1395 * Prepare to search the next object in the
1398 backing_object = tobject->backing_object;
1399 if (backing_object == NULL)
1401 VM_OBJECT_WLOCK(backing_object);
1403 OFF_TO_IDX(tobject->backing_object_offset);
1404 if (tobject != object)
1405 VM_OBJECT_WUNLOCK(tobject);
1406 tobject = backing_object;
1407 if (!vm_object_advice_applies(tobject, advice))
1409 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
1414 m = TAILQ_NEXT(m, listq);
1418 * If the page is not in a normal state, skip it. The page
1419 * can not be invalidated while the object lock is held.
1421 if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1423 KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1424 ("vm_object_madvise: page %p is fictitious", tm));
1425 KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1426 ("vm_object_madvise: page %p is not managed", tm));
1427 if (vm_page_tryxbusy(tm) == 0) {
1428 if (object != tobject)
1429 VM_OBJECT_WUNLOCK(object);
1430 if (advice == MADV_WILLNEED) {
1432 * Reference the page before unlocking and
1433 * sleeping so that the page daemon is less
1434 * likely to reclaim it.
1436 vm_page_aflag_set(tm, PGA_REFERENCED);
1438 if (!vm_page_busy_sleep(tm, "madvpo", 0))
1439 VM_OBJECT_WUNLOCK(tobject);
1442 vm_page_advise(tm, advice);
1443 vm_page_xunbusy(tm);
1444 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1446 if (tobject != object)
1447 VM_OBJECT_WUNLOCK(tobject);
1449 VM_OBJECT_WUNLOCK(object);
1455 * Create a new object which is backed by the
1456 * specified existing object range. The source
1457 * object reference is deallocated.
1459 * The new object and offset into that object
1460 * are returned in the source parameters.
1463 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1464 struct ucred *cred, bool shared)
1472 * Don't create the new object if the old object isn't shared.
1474 * If we hold the only reference we can guarantee that it won't
1475 * increase while we have the map locked. Otherwise the race is
1476 * harmless and we will end up with an extra shadow object that
1477 * will be collapsed later.
1479 if (source != NULL && source->ref_count == 1 &&
1480 (source->flags & OBJ_ANON) != 0)
1484 * Allocate a new object with the given length.
1486 result = vm_object_allocate_anon(atop(length), source, cred, length);
1489 * Store the offset into the source object, and fix up the offset into
1492 result->backing_object_offset = *offset;
1494 if (shared || source != NULL) {
1495 VM_OBJECT_WLOCK(result);
1498 * The new object shadows the source object, adding a
1499 * reference to it. Our caller changes his reference
1500 * to point to the new object, removing a reference to
1501 * the source object. Net result: no change of
1502 * reference count, unless the caller needs to add one
1503 * more reference due to forking a shared map entry.
1506 vm_object_reference_locked(result);
1507 vm_object_clear_flag(result, OBJ_ONEMAPPING);
1511 * Try to optimize the result object's page color when
1512 * shadowing in order to maintain page coloring
1513 * consistency in the combined shadowed object.
1515 if (source != NULL) {
1516 vm_object_backing_insert(result, source);
1517 result->domain = source->domain;
1518 #if VM_NRESERVLEVEL > 0
1519 vm_object_set_flag(result,
1520 (source->flags & OBJ_COLORED));
1521 result->pg_color = (source->pg_color +
1522 OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1526 VM_OBJECT_WUNLOCK(result);
1530 * Return the new things
1539 * Split the pages in a map entry into a new object. This affords
1540 * easier removal of unused pages, and keeps object inheritance from
1541 * being a negative impact on memory usage.
1544 vm_object_split(vm_map_entry_t entry)
1546 vm_page_t m, m_busy, m_next;
1547 vm_object_t orig_object, new_object, backing_object;
1548 vm_pindex_t idx, offidxstart;
1551 orig_object = entry->object.vm_object;
1552 KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1553 ("vm_object_split: Splitting object with multiple mappings."));
1554 if ((orig_object->flags & OBJ_ANON) == 0)
1556 if (orig_object->ref_count <= 1)
1558 VM_OBJECT_WUNLOCK(orig_object);
1560 offidxstart = OFF_TO_IDX(entry->offset);
1561 size = atop(entry->end - entry->start);
1564 * If swap_pager_copy() is later called, it will convert new_object
1565 * into a swap object.
1567 new_object = vm_object_allocate_anon(size, orig_object,
1568 orig_object->cred, ptoa(size));
1571 * We must wait for the orig_object to complete any in-progress
1572 * collapse so that the swap blocks are stable below. The
1573 * additional reference on backing_object by new object will
1574 * prevent further collapse operations until split completes.
1576 VM_OBJECT_WLOCK(orig_object);
1577 vm_object_collapse_wait(orig_object);
1580 * At this point, the new object is still private, so the order in
1581 * which the original and new objects are locked does not matter.
1583 VM_OBJECT_WLOCK(new_object);
1584 new_object->domain = orig_object->domain;
1585 backing_object = orig_object->backing_object;
1586 if (backing_object != NULL) {
1587 vm_object_backing_insert_ref(new_object, backing_object);
1588 new_object->backing_object_offset =
1589 orig_object->backing_object_offset + entry->offset;
1591 if (orig_object->cred != NULL) {
1592 crhold(orig_object->cred);
1593 KASSERT(orig_object->charge >= ptoa(size),
1594 ("orig_object->charge < 0"));
1595 orig_object->charge -= ptoa(size);
1599 * Mark the split operation so that swap_pager_getpages() knows
1600 * that the object is in transition.
1602 vm_object_set_flag(orig_object, OBJ_SPLIT);
1608 m = vm_page_find_least(orig_object, offidxstart);
1609 KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1610 ("%s: object %p was repopulated", __func__, orig_object));
1611 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1613 m_next = TAILQ_NEXT(m, listq);
1616 * We must wait for pending I/O to complete before we can
1619 * We do not have to VM_PROT_NONE the page as mappings should
1620 * not be changed by this operation.
1622 if (vm_page_tryxbusy(m) == 0) {
1623 VM_OBJECT_WUNLOCK(new_object);
1624 if (vm_page_busy_sleep(m, "spltwt", 0))
1625 VM_OBJECT_WLOCK(orig_object);
1626 VM_OBJECT_WLOCK(new_object);
1631 * The page was left invalid. Likely placed there by
1632 * an incomplete fault. Just remove and ignore.
1634 if (vm_page_none_valid(m)) {
1635 if (vm_page_remove(m))
1640 /* vm_page_rename() will dirty the page. */
1641 if (vm_page_rename(m, new_object, idx)) {
1643 VM_OBJECT_WUNLOCK(new_object);
1644 VM_OBJECT_WUNLOCK(orig_object);
1646 VM_OBJECT_WLOCK(orig_object);
1647 VM_OBJECT_WLOCK(new_object);
1651 #if VM_NRESERVLEVEL > 0
1653 * If some of the reservation's allocated pages remain with
1654 * the original object, then transferring the reservation to
1655 * the new object is neither particularly beneficial nor
1656 * particularly harmful as compared to leaving the reservation
1657 * with the original object. If, however, all of the
1658 * reservation's allocated pages are transferred to the new
1659 * object, then transferring the reservation is typically
1660 * beneficial. Determining which of these two cases applies
1661 * would be more costly than unconditionally renaming the
1664 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1668 * orig_object's type may change while sleeping, so keep track
1669 * of the beginning of the busied range.
1671 if (orig_object->type != OBJT_SWAP)
1673 else if (m_busy == NULL)
1676 if ((orig_object->flags & OBJ_SWAP) != 0) {
1678 * swap_pager_copy() can sleep, in which case the orig_object's
1679 * and new_object's locks are released and reacquired.
1681 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1683 TAILQ_FOREACH_FROM(m_busy, &new_object->memq, listq)
1684 vm_page_xunbusy(m_busy);
1686 vm_object_clear_flag(orig_object, OBJ_SPLIT);
1687 VM_OBJECT_WUNLOCK(orig_object);
1688 VM_OBJECT_WUNLOCK(new_object);
1689 entry->object.vm_object = new_object;
1690 entry->offset = 0LL;
1691 vm_object_deallocate(orig_object);
1692 VM_OBJECT_WLOCK(new_object);
1696 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1698 vm_object_t backing_object;
1700 VM_OBJECT_ASSERT_WLOCKED(object);
1701 backing_object = object->backing_object;
1702 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1704 KASSERT(p == NULL || p->object == object || p->object == backing_object,
1705 ("invalid ownership %p %p %p", p, object, backing_object));
1706 /* The page is only NULL when rename fails. */
1708 VM_OBJECT_WUNLOCK(object);
1709 VM_OBJECT_WUNLOCK(backing_object);
1711 VM_OBJECT_WLOCK(object);
1712 } else if (p->object == object) {
1713 VM_OBJECT_WUNLOCK(backing_object);
1714 if (vm_page_busy_sleep(p, "vmocol", 0))
1715 VM_OBJECT_WLOCK(object);
1717 VM_OBJECT_WUNLOCK(object);
1718 if (!vm_page_busy_sleep(p, "vmocol", 0))
1719 VM_OBJECT_WUNLOCK(backing_object);
1720 VM_OBJECT_WLOCK(object);
1722 VM_OBJECT_WLOCK(backing_object);
1723 return (TAILQ_FIRST(&backing_object->memq));
1727 vm_object_scan_all_shadowed(vm_object_t object)
1729 vm_object_t backing_object;
1731 vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1733 VM_OBJECT_ASSERT_WLOCKED(object);
1734 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1736 backing_object = object->backing_object;
1738 if ((backing_object->flags & OBJ_ANON) == 0)
1741 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1742 p = vm_page_find_least(backing_object, pi);
1743 ps = swap_pager_find_least(backing_object, pi);
1746 * Only check pages inside the parent object's range and
1747 * inside the parent object's mapping of the backing object.
1750 if (p != NULL && p->pindex < pi)
1751 p = TAILQ_NEXT(p, listq);
1753 ps = swap_pager_find_least(backing_object, pi);
1754 if (p == NULL && ps >= backing_object->size)
1759 pi = MIN(p->pindex, ps);
1761 new_pindex = pi - backing_offset_index;
1762 if (new_pindex >= object->size)
1767 * If the backing object page is busy a
1768 * grandparent or older page may still be
1769 * undergoing CoW. It is not safe to collapse
1770 * the backing object until it is quiesced.
1772 if (vm_page_tryxbusy(p) == 0)
1776 * We raced with the fault handler that left
1777 * newly allocated invalid page on the object
1778 * queue and retried.
1780 if (!vm_page_all_valid(p))
1785 * See if the parent has the page or if the parent's object
1786 * pager has the page. If the parent has the page but the page
1787 * is not valid, the parent's object pager must have the page.
1789 * If this fails, the parent does not completely shadow the
1790 * object and we might as well give up now.
1792 pp = vm_page_lookup(object, new_pindex);
1795 * The valid check here is stable due to object lock
1796 * being required to clear valid and initiate paging.
1797 * Busy of p disallows fault handler to validate pp.
1799 if ((pp == NULL || vm_page_none_valid(pp)) &&
1800 !vm_pager_has_page(object, new_pindex, NULL, NULL))
1814 vm_object_collapse_scan(vm_object_t object)
1816 vm_object_t backing_object;
1817 vm_page_t next, p, pp;
1818 vm_pindex_t backing_offset_index, new_pindex;
1820 VM_OBJECT_ASSERT_WLOCKED(object);
1821 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1823 backing_object = object->backing_object;
1824 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1829 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1830 next = TAILQ_NEXT(p, listq);
1831 new_pindex = p->pindex - backing_offset_index;
1834 * Check for busy page
1836 if (vm_page_tryxbusy(p) == 0) {
1837 next = vm_object_collapse_scan_wait(object, p);
1841 KASSERT(object->backing_object == backing_object,
1842 ("vm_object_collapse_scan: backing object mismatch %p != %p",
1843 object->backing_object, backing_object));
1844 KASSERT(p->object == backing_object,
1845 ("vm_object_collapse_scan: object mismatch %p != %p",
1846 p->object, backing_object));
1848 if (p->pindex < backing_offset_index ||
1849 new_pindex >= object->size) {
1850 vm_pager_freespace(backing_object, p->pindex, 1);
1852 KASSERT(!pmap_page_is_mapped(p),
1853 ("freeing mapped page %p", p));
1854 if (vm_page_remove(p))
1859 if (!vm_page_all_valid(p)) {
1860 KASSERT(!pmap_page_is_mapped(p),
1861 ("freeing mapped page %p", p));
1862 if (vm_page_remove(p))
1867 pp = vm_page_lookup(object, new_pindex);
1868 if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1871 * The page in the parent is busy and possibly not
1872 * (yet) valid. Until its state is finalized by the
1873 * busy bit owner, we can't tell whether it shadows the
1876 next = vm_object_collapse_scan_wait(object, pp);
1880 if (pp != NULL && vm_page_none_valid(pp)) {
1882 * The page was invalid in the parent. Likely placed
1883 * there by an incomplete fault. Just remove and
1884 * ignore. p can replace it.
1886 if (vm_page_remove(pp))
1891 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1894 * The page already exists in the parent OR swap exists
1895 * for this location in the parent. Leave the parent's
1896 * page alone. Destroy the original page from the
1899 vm_pager_freespace(backing_object, p->pindex, 1);
1900 KASSERT(!pmap_page_is_mapped(p),
1901 ("freeing mapped page %p", p));
1902 if (vm_page_remove(p))
1905 vm_page_xunbusy(pp);
1910 * Page does not exist in parent, rename the page from the
1911 * backing object to the main object.
1913 * If the page was mapped to a process, it can remain mapped
1914 * through the rename. vm_page_rename() will dirty the page.
1916 if (vm_page_rename(p, object, new_pindex)) {
1918 next = vm_object_collapse_scan_wait(object, NULL);
1922 /* Use the old pindex to free the right page. */
1923 vm_pager_freespace(backing_object, new_pindex +
1924 backing_offset_index, 1);
1926 #if VM_NRESERVLEVEL > 0
1928 * Rename the reservation.
1930 vm_reserv_rename(p, object, backing_object,
1931 backing_offset_index);
1939 * vm_object_collapse:
1941 * Collapse an object with the object backing it.
1942 * Pages in the backing object are moved into the
1943 * parent, and the backing object is deallocated.
1946 vm_object_collapse(vm_object_t object)
1948 vm_object_t backing_object, new_backing_object;
1950 VM_OBJECT_ASSERT_WLOCKED(object);
1953 KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1954 ("collapsing invalid object"));
1957 * Wait for the backing_object to finish any pending
1958 * collapse so that the caller sees the shortest possible
1961 backing_object = vm_object_backing_collapse_wait(object);
1962 if (backing_object == NULL)
1965 KASSERT(object->ref_count > 0 &&
1966 object->ref_count > atomic_load_int(&object->shadow_count),
1967 ("collapse with invalid ref %d or shadow %d count.",
1968 object->ref_count, atomic_load_int(&object->shadow_count)));
1969 KASSERT((backing_object->flags &
1970 (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1971 ("vm_object_collapse: Backing object already collapsing."));
1972 KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1973 ("vm_object_collapse: object is already collapsing."));
1976 * We know that we can either collapse the backing object if
1977 * the parent is the only reference to it, or (perhaps) have
1978 * the parent bypass the object if the parent happens to shadow
1979 * all the resident pages in the entire backing object.
1981 if (backing_object->ref_count == 1) {
1982 KASSERT(atomic_load_int(&backing_object->shadow_count)
1984 ("vm_object_collapse: shadow_count: %d",
1985 atomic_load_int(&backing_object->shadow_count)));
1986 vm_object_pip_add(object, 1);
1987 vm_object_set_flag(object, OBJ_COLLAPSING);
1988 vm_object_pip_add(backing_object, 1);
1989 vm_object_set_flag(backing_object, OBJ_DEAD);
1992 * If there is exactly one reference to the backing
1993 * object, we can collapse it into the parent.
1995 vm_object_collapse_scan(object);
1997 #if VM_NRESERVLEVEL > 0
1999 * Break any reservations from backing_object.
2001 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
2002 vm_reserv_break_all(backing_object);
2006 * Move the pager from backing_object to object.
2008 if ((backing_object->flags & OBJ_SWAP) != 0) {
2010 * swap_pager_copy() can sleep, in which case
2011 * the backing_object's and object's locks are
2012 * released and reacquired.
2013 * Since swap_pager_copy() is being asked to
2014 * destroy backing_object, it will change the
2015 * type to OBJT_DEFAULT.
2020 OFF_TO_IDX(object->backing_object_offset), TRUE);
2024 * Object now shadows whatever backing_object did.
2026 vm_object_clear_flag(object, OBJ_COLLAPSING);
2027 vm_object_backing_transfer(object, backing_object);
2028 object->backing_object_offset +=
2029 backing_object->backing_object_offset;
2030 VM_OBJECT_WUNLOCK(object);
2031 vm_object_pip_wakeup(object);
2034 * Discard backing_object.
2036 * Since the backing object has no pages, no pager left,
2037 * and no object references within it, all that is
2038 * necessary is to dispose of it.
2040 KASSERT(backing_object->ref_count == 1, (
2041 "backing_object %p was somehow re-referenced during collapse!",
2043 vm_object_pip_wakeup(backing_object);
2044 (void)refcount_release(&backing_object->ref_count);
2045 vm_object_terminate(backing_object);
2046 counter_u64_add(object_collapses, 1);
2047 VM_OBJECT_WLOCK(object);
2050 * If we do not entirely shadow the backing object,
2051 * there is nothing we can do so we give up.
2053 * The object lock and backing_object lock must not
2054 * be dropped during this sequence.
2056 if (!vm_object_scan_all_shadowed(object)) {
2057 VM_OBJECT_WUNLOCK(backing_object);
2062 * Make the parent shadow the next object in the
2063 * chain. Deallocating backing_object will not remove
2064 * it, since its reference count is at least 2.
2066 vm_object_backing_remove_locked(object);
2067 new_backing_object = backing_object->backing_object;
2068 if (new_backing_object != NULL) {
2069 vm_object_backing_insert_ref(object,
2070 new_backing_object);
2071 object->backing_object_offset +=
2072 backing_object->backing_object_offset;
2076 * Drop the reference count on backing_object. Since
2077 * its ref_count was at least 2, it will not vanish.
2079 (void)refcount_release(&backing_object->ref_count);
2080 KASSERT(backing_object->ref_count >= 1, (
2081 "backing_object %p was somehow dereferenced during collapse!",
2083 VM_OBJECT_WUNLOCK(backing_object);
2084 counter_u64_add(object_bypasses, 1);
2088 * Try again with this object's new backing object.
2094 * vm_object_page_remove:
2096 * For the given object, either frees or invalidates each of the
2097 * specified pages. In general, a page is freed. However, if a page is
2098 * wired for any reason other than the existence of a managed, wired
2099 * mapping, then it may be invalidated but not removed from the object.
2100 * Pages are specified by the given range ["start", "end") and the option
2101 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
2102 * extends from "start" to the end of the object. If the option
2103 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2104 * specified range are affected. If the option OBJPR_NOTMAPPED is
2105 * specified, then the pages within the specified range must have no
2106 * mappings. Otherwise, if this option is not specified, any mappings to
2107 * the specified pages are removed before the pages are freed or
2110 * In general, this operation should only be performed on objects that
2111 * contain managed pages. There are, however, two exceptions. First, it
2112 * is performed on the kernel and kmem objects by vm_map_entry_delete().
2113 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2114 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
2115 * not be specified and the option OBJPR_NOTMAPPED must be specified.
2117 * The object must be locked.
2120 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2125 VM_OBJECT_ASSERT_WLOCKED(object);
2126 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2127 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2128 ("vm_object_page_remove: illegal options for object %p", object));
2129 if (object->resident_page_count == 0)
2131 vm_object_pip_add(object, 1);
2133 p = vm_page_find_least(object, start);
2136 * Here, the variable "p" is either (1) the page with the least pindex
2137 * greater than or equal to the parameter "start" or (2) NULL.
2139 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2140 next = TAILQ_NEXT(p, listq);
2143 * Skip invalid pages if asked to do so. Try to avoid acquiring
2144 * the busy lock, as some consumers rely on this to avoid
2147 * A thread may concurrently transition the page from invalid to
2148 * valid using only the busy lock, so the result of this check
2149 * is immediately stale. It is up to consumers to handle this,
2150 * for instance by ensuring that all invalid->valid transitions
2151 * happen with a mutex held, as may be possible for a
2154 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2158 * If the page is wired for any reason besides the existence
2159 * of managed, wired mappings, then it cannot be freed. For
2160 * example, fictitious pages, which represent device memory,
2161 * are inherently wired and cannot be freed. They can,
2162 * however, be invalidated if the option OBJPR_CLEANONLY is
2165 if (vm_page_tryxbusy(p) == 0) {
2166 if (vm_page_busy_sleep(p, "vmopar", 0))
2167 VM_OBJECT_WLOCK(object);
2170 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2174 if (vm_page_wired(p)) {
2176 if ((options & OBJPR_NOTMAPPED) == 0 &&
2177 object->ref_count != 0)
2179 if ((options & OBJPR_CLEANONLY) == 0) {
2186 KASSERT((p->flags & PG_FICTITIOUS) == 0,
2187 ("vm_object_page_remove: page %p is fictitious", p));
2188 if ((options & OBJPR_CLEANONLY) != 0 &&
2189 !vm_page_none_valid(p)) {
2190 if ((options & OBJPR_NOTMAPPED) == 0 &&
2191 object->ref_count != 0 &&
2192 !vm_page_try_remove_write(p))
2194 if (p->dirty != 0) {
2199 if ((options & OBJPR_NOTMAPPED) == 0 &&
2200 object->ref_count != 0 && !vm_page_try_remove_all(p))
2204 vm_object_pip_wakeup(object);
2206 vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2211 * vm_object_page_noreuse:
2213 * For the given object, attempt to move the specified pages to
2214 * the head of the inactive queue. This bypasses regular LRU
2215 * operation and allows the pages to be reused quickly under memory
2216 * pressure. If a page is wired for any reason, then it will not
2217 * be queued. Pages are specified by the range ["start", "end").
2218 * As a special case, if "end" is zero, then the range extends from
2219 * "start" to the end of the object.
2221 * This operation should only be performed on objects that
2222 * contain non-fictitious, managed pages.
2224 * The object must be locked.
2227 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2231 VM_OBJECT_ASSERT_LOCKED(object);
2232 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2233 ("vm_object_page_noreuse: illegal object %p", object));
2234 if (object->resident_page_count == 0)
2236 p = vm_page_find_least(object, start);
2239 * Here, the variable "p" is either (1) the page with the least pindex
2240 * greater than or equal to the parameter "start" or (2) NULL.
2242 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2243 next = TAILQ_NEXT(p, listq);
2244 vm_page_deactivate_noreuse(p);
2249 * Populate the specified range of the object with valid pages. Returns
2250 * TRUE if the range is successfully populated and FALSE otherwise.
2252 * Note: This function should be optimized to pass a larger array of
2253 * pages to vm_pager_get_pages() before it is applied to a non-
2254 * OBJT_DEVICE object.
2256 * The object must be locked.
2259 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2265 VM_OBJECT_ASSERT_WLOCKED(object);
2266 for (pindex = start; pindex < end; pindex++) {
2267 rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2268 if (rv != VM_PAGER_OK)
2272 * Keep "m" busy because a subsequent iteration may unlock
2276 if (pindex > start) {
2277 m = vm_page_lookup(object, start);
2278 while (m != NULL && m->pindex < pindex) {
2280 m = TAILQ_NEXT(m, listq);
2283 return (pindex == end);
2287 * Routine: vm_object_coalesce
2288 * Function: Coalesces two objects backing up adjoining
2289 * regions of memory into a single object.
2291 * returns TRUE if objects were combined.
2293 * NOTE: Only works at the moment if the second object is NULL -
2294 * if it's not, which object do we lock first?
2297 * prev_object First object to coalesce
2298 * prev_offset Offset into prev_object
2299 * prev_size Size of reference to prev_object
2300 * next_size Size of reference to the second object
2301 * reserved Indicator that extension region has
2302 * swap accounted for
2305 * The object must *not* be locked.
2308 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2309 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2311 vm_pindex_t next_pindex;
2313 if (prev_object == NULL)
2315 if ((prev_object->flags & OBJ_ANON) == 0)
2318 VM_OBJECT_WLOCK(prev_object);
2320 * Try to collapse the object first.
2322 vm_object_collapse(prev_object);
2325 * Can't coalesce if: . more than one reference . paged out . shadows
2326 * another object . has a copy elsewhere (any of which mean that the
2327 * pages not mapped to prev_entry may be in use anyway)
2329 if (prev_object->backing_object != NULL) {
2330 VM_OBJECT_WUNLOCK(prev_object);
2334 prev_size >>= PAGE_SHIFT;
2335 next_size >>= PAGE_SHIFT;
2336 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2338 if (prev_object->ref_count > 1 &&
2339 prev_object->size != next_pindex &&
2340 (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2341 VM_OBJECT_WUNLOCK(prev_object);
2346 * Account for the charge.
2348 if (prev_object->cred != NULL) {
2350 * If prev_object was charged, then this mapping,
2351 * although not charged now, may become writable
2352 * later. Non-NULL cred in the object would prevent
2353 * swap reservation during enabling of the write
2354 * access, so reserve swap now. Failed reservation
2355 * cause allocation of the separate object for the map
2356 * entry, and swap reservation for this entry is
2357 * managed in appropriate time.
2359 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2360 prev_object->cred)) {
2361 VM_OBJECT_WUNLOCK(prev_object);
2364 prev_object->charge += ptoa(next_size);
2368 * Remove any pages that may still be in the object from a previous
2371 if (next_pindex < prev_object->size) {
2372 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2375 if (prev_object->cred != NULL) {
2376 KASSERT(prev_object->charge >=
2377 ptoa(prev_object->size - next_pindex),
2378 ("object %p overcharged 1 %jx %jx", prev_object,
2379 (uintmax_t)next_pindex, (uintmax_t)next_size));
2380 prev_object->charge -= ptoa(prev_object->size -
2387 * Extend the object if necessary.
2389 if (next_pindex + next_size > prev_object->size)
2390 prev_object->size = next_pindex + next_size;
2392 VM_OBJECT_WUNLOCK(prev_object);
2397 vm_object_set_writeable_dirty_(vm_object_t object)
2399 atomic_add_int(&object->generation, 1);
2403 vm_object_mightbedirty_(vm_object_t object)
2405 return (object->generation != object->cleangeneration);
2411 * For each page offset within the specified range of the given object,
2412 * find the highest-level page in the shadow chain and unwire it. A page
2413 * must exist at every page offset, and the highest-level page must be
2417 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2420 vm_object_t tobject, t1object;
2422 vm_pindex_t end_pindex, pindex, tpindex;
2423 int depth, locked_depth;
2425 KASSERT((offset & PAGE_MASK) == 0,
2426 ("vm_object_unwire: offset is not page aligned"));
2427 KASSERT((length & PAGE_MASK) == 0,
2428 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2429 /* The wired count of a fictitious page never changes. */
2430 if ((object->flags & OBJ_FICTITIOUS) != 0)
2432 pindex = OFF_TO_IDX(offset);
2433 end_pindex = pindex + atop(length);
2436 VM_OBJECT_RLOCK(object);
2437 m = vm_page_find_least(object, pindex);
2438 while (pindex < end_pindex) {
2439 if (m == NULL || pindex < m->pindex) {
2441 * The first object in the shadow chain doesn't
2442 * contain a page at the current index. Therefore,
2443 * the page must exist in a backing object.
2450 OFF_TO_IDX(tobject->backing_object_offset);
2451 tobject = tobject->backing_object;
2452 KASSERT(tobject != NULL,
2453 ("vm_object_unwire: missing page"));
2454 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2457 if (depth == locked_depth) {
2459 VM_OBJECT_RLOCK(tobject);
2461 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2465 m = TAILQ_NEXT(m, listq);
2467 if (vm_page_trysbusy(tm) == 0) {
2468 for (tobject = object; locked_depth >= 1;
2470 t1object = tobject->backing_object;
2471 if (tm->object != tobject)
2472 VM_OBJECT_RUNLOCK(tobject);
2475 tobject = tm->object;
2476 if (!vm_page_busy_sleep(tm, "unwbo",
2477 VM_ALLOC_IGN_SBUSY))
2478 VM_OBJECT_RUNLOCK(tobject);
2481 vm_page_unwire(tm, queue);
2482 vm_page_sunbusy(tm);
2486 /* Release the accumulated object locks. */
2487 for (tobject = object; locked_depth >= 1; locked_depth--) {
2488 t1object = tobject->backing_object;
2489 VM_OBJECT_RUNLOCK(tobject);
2495 * Return the vnode for the given object, or NULL if none exists.
2496 * For tmpfs objects, the function may return NULL if there is
2497 * no vnode allocated at the time of the call.
2500 vm_object_vnode(vm_object_t object)
2504 VM_OBJECT_ASSERT_LOCKED(object);
2505 vm_pager_getvp(object, &vp, NULL);
2510 * Busy the vm object. This prevents new pages belonging to the object from
2511 * becoming busy. Existing pages persist as busy. Callers are responsible
2512 * for checking page state before proceeding.
2515 vm_object_busy(vm_object_t obj)
2518 VM_OBJECT_ASSERT_LOCKED(obj);
2520 blockcount_acquire(&obj->busy, 1);
2521 /* The fence is required to order loads of page busy. */
2522 atomic_thread_fence_acq_rel();
2526 vm_object_unbusy(vm_object_t obj)
2529 blockcount_release(&obj->busy, 1);
2533 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2536 VM_OBJECT_ASSERT_UNLOCKED(obj);
2538 (void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2542 * This function aims to determine if the object is mapped,
2543 * specifically, if it is referenced by a vm_map_entry. Because
2544 * objects occasionally acquire transient references that do not
2545 * represent a mapping, the method used here is inexact. However, it
2546 * has very low overhead and is good enough for the advisory
2547 * vm.vmtotal sysctl.
2550 vm_object_is_active(vm_object_t obj)
2553 return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2557 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2559 struct kinfo_vmobject *kvo;
2560 char *fullpath, *freepath;
2568 if (req->oldptr == NULL) {
2570 * If an old buffer has not been provided, generate an
2571 * estimate of the space needed for a subsequent call.
2573 mtx_lock(&vm_object_list_mtx);
2575 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2576 if (obj->type == OBJT_DEAD)
2580 mtx_unlock(&vm_object_list_mtx);
2581 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2585 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO);
2589 * VM objects are type stable and are never removed from the
2590 * list once added. This allows us to safely read obj->object_list
2591 * after reacquiring the VM object lock.
2593 mtx_lock(&vm_object_list_mtx);
2594 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2595 if (obj->type == OBJT_DEAD ||
2596 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2598 VM_OBJECT_RLOCK(obj);
2599 if (obj->type == OBJT_DEAD ||
2600 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2601 VM_OBJECT_RUNLOCK(obj);
2604 mtx_unlock(&vm_object_list_mtx);
2605 kvo->kvo_size = ptoa(obj->size);
2606 kvo->kvo_resident = obj->resident_page_count;
2607 kvo->kvo_ref_count = obj->ref_count;
2608 kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2609 kvo->kvo_memattr = obj->memattr;
2610 kvo->kvo_active = 0;
2611 kvo->kvo_inactive = 0;
2613 TAILQ_FOREACH(m, &obj->memq, listq) {
2615 * A page may belong to the object but be
2616 * dequeued and set to PQ_NONE while the
2617 * object lock is not held. This makes the
2618 * reads of m->queue below racy, and we do not
2619 * count pages set to PQ_NONE. However, this
2620 * sysctl is only meant to give an
2621 * approximation of the system anyway.
2623 if (m->a.queue == PQ_ACTIVE)
2625 else if (m->a.queue == PQ_INACTIVE)
2626 kvo->kvo_inactive++;
2630 kvo->kvo_vn_fileid = 0;
2631 kvo->kvo_vn_fsid = 0;
2632 kvo->kvo_vn_fsid_freebsd11 = 0;
2636 kvo->kvo_type = vm_object_kvme_type(obj, swap_only ? NULL : &vp);
2639 } else if ((obj->flags & OBJ_ANON) != 0) {
2640 MPASS(kvo->kvo_type == KVME_TYPE_DEFAULT ||
2641 kvo->kvo_type == KVME_TYPE_SWAP);
2642 kvo->kvo_me = (uintptr_t)obj;
2643 /* tmpfs objs are reported as vnodes */
2644 kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2645 sp = swap_pager_swapped_pages(obj);
2646 kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2648 VM_OBJECT_RUNLOCK(obj);
2650 vn_fullpath(vp, &fullpath, &freepath);
2651 vn_lock(vp, LK_SHARED | LK_RETRY);
2652 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2653 kvo->kvo_vn_fileid = va.va_fileid;
2654 kvo->kvo_vn_fsid = va.va_fsid;
2655 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2661 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2662 if (freepath != NULL)
2663 free(freepath, M_TEMP);
2665 /* Pack record size down */
2666 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2667 + strlen(kvo->kvo_path) + 1;
2668 kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2670 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2672 mtx_lock(&vm_object_list_mtx);
2676 mtx_unlock(&vm_object_list_mtx);
2682 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2684 return (vm_object_list_handler(req, false));
2687 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2688 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2689 "List of VM objects");
2692 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2694 return (vm_object_list_handler(req, true));
2698 * This sysctl returns list of the anonymous or swap objects. Intent
2699 * is to provide stripped optimized list useful to analyze swap use.
2700 * Since technically non-swap (default) objects participate in the
2701 * shadow chains, and are converted to swap type as needed by swap
2702 * pager, we must report them.
2704 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2705 CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2706 sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2707 "List of swap VM objects");
2709 #include "opt_ddb.h"
2711 #include <sys/kernel.h>
2713 #include <sys/cons.h>
2715 #include <ddb/ddb.h>
2718 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2721 vm_map_entry_t tmpe;
2728 VM_MAP_ENTRY_FOREACH(tmpe, map) {
2729 if (_vm_object_in_map(map, object, tmpe)) {
2733 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2734 tmpm = entry->object.sub_map;
2735 VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2736 if (_vm_object_in_map(tmpm, object, tmpe)) {
2740 } else if ((obj = entry->object.vm_object) != NULL) {
2741 for (; obj; obj = obj->backing_object)
2742 if (obj == object) {
2750 vm_object_in_map(vm_object_t object)
2754 /* sx_slock(&allproc_lock); */
2755 FOREACH_PROC_IN_SYSTEM(p) {
2756 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2758 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2759 /* sx_sunlock(&allproc_lock); */
2763 /* sx_sunlock(&allproc_lock); */
2764 if (_vm_object_in_map(kernel_map, object, 0))
2769 DB_SHOW_COMMAND(vmochk, vm_object_check)
2774 * make sure that internal objs are in a map somewhere
2775 * and none have zero ref counts.
2777 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2778 if ((object->flags & OBJ_ANON) != 0) {
2779 if (object->ref_count == 0) {
2780 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2781 (long)object->size);
2783 if (!vm_object_in_map(object)) {
2785 "vmochk: internal obj is not in a map: "
2786 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2787 object->ref_count, (u_long)object->size,
2788 (u_long)object->size,
2789 (void *)object->backing_object);
2798 * vm_object_print: [ debug ]
2800 DB_SHOW_COMMAND(object, vm_object_print_static)
2802 /* XXX convert args. */
2803 vm_object_t object = (vm_object_t)addr;
2804 boolean_t full = have_addr;
2808 /* XXX count is an (unused) arg. Avoid shadowing it. */
2809 #define count was_count
2817 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2818 object, (int)object->type, (uintmax_t)object->size,
2819 object->resident_page_count, object->ref_count, object->flags,
2820 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2821 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2822 atomic_load_int(&object->shadow_count),
2823 object->backing_object ? object->backing_object->ref_count : 0,
2824 object->backing_object, (uintmax_t)object->backing_object_offset);
2831 TAILQ_FOREACH(p, &object->memq, listq) {
2833 db_iprintf("memory:=");
2834 else if (count == 6) {
2842 db_printf("(off=0x%jx,page=0x%jx)",
2843 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2856 /* XXX need this non-static entry for calling from vm_map_print. */
2859 /* db_expr_t */ long addr,
2860 boolean_t have_addr,
2861 /* db_expr_t */ long count,
2864 vm_object_print_static(addr, have_addr, count, modif);
2867 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2872 vm_page_t m, prev_m;
2875 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2876 db_printf("new object: %p\n", (void *)object);
2883 TAILQ_FOREACH(m, &object->memq, listq) {
2884 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2885 prev_m->pindex + 1 != m->pindex) {
2887 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2888 (long)fidx, rcount, (long)pa);
2895 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2900 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2901 (long)fidx, rcount, (long)pa);
2906 pa = VM_PAGE_TO_PHYS(m);
2910 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2911 (long)fidx, rcount, (long)pa);