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);
249 object->un_pager.swp.writemappings = 0;
253 * Ensure that swap_pager_swapoff() iteration over object_list
254 * sees up to date type and pctrie head if it observed
257 atomic_thread_fence_rel();
259 object->pg_color = 0;
261 object->domain.dr_policy = NULL;
262 object->generation = 1;
263 object->cleangeneration = 1;
264 refcount_init(&object->ref_count, 1);
265 object->memattr = VM_MEMATTR_DEFAULT;
268 object->handle = handle;
269 object->backing_object = NULL;
270 object->backing_object_offset = (vm_ooffset_t) 0;
271 #if VM_NRESERVLEVEL > 0
272 LIST_INIT(&object->rvq);
274 umtx_shm_object_init(object);
280 * Initialize the VM objects module.
285 TAILQ_INIT(&vm_object_list);
286 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
288 rw_init(&kernel_object->lock, "kernel vm object");
289 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
290 VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
291 #if VM_NRESERVLEVEL > 0
292 kernel_object->flags |= OBJ_COLORED;
293 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
295 kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
298 * The lock portion of struct vm_object must be type stable due
299 * to vm_pageout_fallback_object_lock locking a vm object
300 * without holding any references to it.
302 * paging_in_progress is valid always. Lockless references to
303 * the objects may acquire pip and then check OBJ_DEAD.
305 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
311 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
317 vm_object_clear_flag(vm_object_t object, u_short bits)
320 VM_OBJECT_ASSERT_WLOCKED(object);
321 object->flags &= ~bits;
325 * Sets the default memory attribute for the specified object. Pages
326 * that are allocated to this object are by default assigned this memory
329 * Presently, this function must be called before any pages are allocated
330 * to the object. In the future, this requirement may be relaxed for
331 * "default" and "swap" objects.
334 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
337 VM_OBJECT_ASSERT_WLOCKED(object);
339 if (object->type == OBJT_DEAD)
340 return (KERN_INVALID_ARGUMENT);
341 if (!TAILQ_EMPTY(&object->memq))
342 return (KERN_FAILURE);
344 object->memattr = memattr;
345 return (KERN_SUCCESS);
349 vm_object_pip_add(vm_object_t object, short i)
353 blockcount_acquire(&object->paging_in_progress, i);
357 vm_object_pip_wakeup(vm_object_t object)
360 vm_object_pip_wakeupn(object, 1);
364 vm_object_pip_wakeupn(vm_object_t object, short i)
368 blockcount_release(&object->paging_in_progress, i);
372 * Atomically drop the object lock and wait for pip to drain. This protects
373 * from sleep/wakeup races due to identity changes. The lock is not re-acquired
377 vm_object_pip_sleep(vm_object_t object, const char *waitid)
380 (void)blockcount_sleep(&object->paging_in_progress, &object->lock,
381 waitid, PVM | PDROP);
385 vm_object_pip_wait(vm_object_t object, const char *waitid)
388 VM_OBJECT_ASSERT_WLOCKED(object);
390 blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
395 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
398 VM_OBJECT_ASSERT_UNLOCKED(object);
400 blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
404 * vm_object_allocate:
406 * Returns a new object with the given size.
409 vm_object_allocate(objtype_t type, vm_pindex_t size)
416 panic("vm_object_allocate: can't create OBJT_DEAD");
418 flags = OBJ_COLORED | OBJ_SWAP;
422 flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
425 flags = OBJ_FICTITIOUS;
428 flags = OBJ_UNMANAGED;
434 panic("vm_object_allocate: type %d is undefined or dynamic",
437 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
438 _vm_object_allocate(type, size, flags, object, NULL);
444 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags)
448 MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */);
449 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
450 _vm_object_allocate(dyntype, size, flags, object, NULL);
456 * vm_object_allocate_anon:
458 * Returns a new default object of the given size and marked as
459 * anonymous memory for special split/collapse handling. Color
460 * to be initialized by the caller.
463 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
464 struct ucred *cred, vm_size_t charge)
466 vm_object_t handle, object;
468 if (backing_object == NULL)
470 else if ((backing_object->flags & OBJ_ANON) != 0)
471 handle = backing_object->handle;
473 handle = backing_object;
474 object = uma_zalloc(obj_zone, M_WAITOK);
475 _vm_object_allocate(OBJT_SWAP, size,
476 OBJ_ANON | OBJ_ONEMAPPING | OBJ_SWAP, object, handle);
478 object->charge = cred != NULL ? charge : 0;
483 vm_object_reference_vnode(vm_object_t object)
488 * vnode objects need the lock for the first reference
489 * to serialize with vnode_object_deallocate().
491 if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
492 VM_OBJECT_RLOCK(object);
493 old = refcount_acquire(&object->ref_count);
494 if (object->type == OBJT_VNODE && old == 0)
495 vref(object->handle);
496 VM_OBJECT_RUNLOCK(object);
501 * vm_object_reference:
503 * Acquires a reference to the given object.
506 vm_object_reference(vm_object_t object)
512 if (object->type == OBJT_VNODE)
513 vm_object_reference_vnode(object);
515 refcount_acquire(&object->ref_count);
516 KASSERT((object->flags & OBJ_DEAD) == 0,
517 ("vm_object_reference: Referenced dead object."));
521 * vm_object_reference_locked:
523 * Gets another reference to the given object.
525 * The object must be locked.
528 vm_object_reference_locked(vm_object_t object)
532 VM_OBJECT_ASSERT_LOCKED(object);
533 old = refcount_acquire(&object->ref_count);
534 if (object->type == OBJT_VNODE && old == 0)
535 vref(object->handle);
536 KASSERT((object->flags & OBJ_DEAD) == 0,
537 ("vm_object_reference: Referenced dead object."));
541 * Handle deallocating an object of type OBJT_VNODE.
544 vm_object_deallocate_vnode(vm_object_t object)
546 struct vnode *vp = (struct vnode *) object->handle;
549 KASSERT(object->type == OBJT_VNODE,
550 ("vm_object_deallocate_vnode: not a vnode object"));
551 KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
553 /* Object lock to protect handle lookup. */
554 last = refcount_release(&object->ref_count);
555 VM_OBJECT_RUNLOCK(object);
560 if (!umtx_shm_vnobj_persistent)
561 umtx_shm_object_terminated(object);
563 /* vrele may need the vnode lock. */
568 * We dropped a reference on an object and discovered that it had a
569 * single remaining shadow. This is a sibling of the reference we
570 * dropped. Attempt to collapse the sibling and backing object.
573 vm_object_deallocate_anon(vm_object_t backing_object)
577 /* Fetch the final shadow. */
578 object = LIST_FIRST(&backing_object->shadow_head);
579 KASSERT(object != NULL &&
580 atomic_load_int(&backing_object->shadow_count) == 1,
581 ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
582 backing_object->ref_count,
583 atomic_load_int(&backing_object->shadow_count)));
584 KASSERT((object->flags & OBJ_ANON) != 0,
585 ("invalid shadow object %p", object));
587 if (!VM_OBJECT_TRYWLOCK(object)) {
589 * Prevent object from disappearing since we do not have a
592 vm_object_pip_add(object, 1);
593 VM_OBJECT_WUNLOCK(backing_object);
594 VM_OBJECT_WLOCK(object);
595 vm_object_pip_wakeup(object);
597 VM_OBJECT_WUNLOCK(backing_object);
600 * Check for a collapse/terminate race with the last reference holder.
602 if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
603 !refcount_acquire_if_not_zero(&object->ref_count)) {
604 VM_OBJECT_WUNLOCK(object);
607 backing_object = object->backing_object;
608 if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
609 vm_object_collapse(object);
610 VM_OBJECT_WUNLOCK(object);
616 * vm_object_deallocate:
618 * Release a reference to the specified object,
619 * gained either through a vm_object_allocate
620 * or a vm_object_reference call. When all references
621 * are gone, storage associated with this object
622 * may be relinquished.
624 * No object may be locked.
627 vm_object_deallocate(vm_object_t object)
632 while (object != NULL) {
634 * If the reference count goes to 0 we start calling
635 * vm_object_terminate() on the object chain. A ref count
636 * of 1 may be a special case depending on the shadow count
637 * being 0 or 1. These cases require a write lock on the
640 if ((object->flags & OBJ_ANON) == 0)
641 released = refcount_release_if_gt(&object->ref_count, 1);
643 released = refcount_release_if_gt(&object->ref_count, 2);
647 if (object->type == OBJT_VNODE) {
648 VM_OBJECT_RLOCK(object);
649 if (object->type == OBJT_VNODE) {
650 vm_object_deallocate_vnode(object);
653 VM_OBJECT_RUNLOCK(object);
656 VM_OBJECT_WLOCK(object);
657 KASSERT(object->ref_count > 0,
658 ("vm_object_deallocate: object deallocated too many times: %d",
662 * If this is not the final reference to an anonymous
663 * object we may need to collapse the shadow chain.
665 if (!refcount_release(&object->ref_count)) {
666 if (object->ref_count > 1 ||
667 atomic_load_int(&object->shadow_count) == 0) {
668 if ((object->flags & OBJ_ANON) != 0 &&
669 object->ref_count == 1)
670 vm_object_set_flag(object,
672 VM_OBJECT_WUNLOCK(object);
676 /* Handle collapsing last ref on anonymous objects. */
677 object = vm_object_deallocate_anon(object);
682 * Handle the final reference to an object. We restart
683 * the loop with the backing object to avoid recursion.
685 umtx_shm_object_terminated(object);
686 temp = object->backing_object;
688 KASSERT(object->type == OBJT_SWAP,
689 ("shadowed tmpfs v_object 2 %p", object));
690 vm_object_backing_remove(object);
693 KASSERT((object->flags & OBJ_DEAD) == 0,
694 ("vm_object_deallocate: Terminating dead object."));
695 vm_object_set_flag(object, OBJ_DEAD);
696 vm_object_terminate(object);
702 vm_object_destroy(vm_object_t object)
704 uma_zfree(obj_zone, object);
708 vm_object_sub_shadow(vm_object_t object)
710 KASSERT(object->shadow_count >= 1,
711 ("object %p sub_shadow count zero", object));
712 atomic_subtract_int(&object->shadow_count, 1);
716 vm_object_backing_remove_locked(vm_object_t object)
718 vm_object_t backing_object;
720 backing_object = object->backing_object;
721 VM_OBJECT_ASSERT_WLOCKED(object);
722 VM_OBJECT_ASSERT_WLOCKED(backing_object);
724 KASSERT((object->flags & OBJ_COLLAPSING) == 0,
725 ("vm_object_backing_remove: Removing collapsing object."));
727 vm_object_sub_shadow(backing_object);
728 if ((object->flags & OBJ_SHADOWLIST) != 0) {
729 LIST_REMOVE(object, shadow_list);
730 vm_object_clear_flag(object, OBJ_SHADOWLIST);
732 object->backing_object = NULL;
736 vm_object_backing_remove(vm_object_t object)
738 vm_object_t backing_object;
740 VM_OBJECT_ASSERT_WLOCKED(object);
742 backing_object = object->backing_object;
743 if ((object->flags & OBJ_SHADOWLIST) != 0) {
744 VM_OBJECT_WLOCK(backing_object);
745 vm_object_backing_remove_locked(object);
746 VM_OBJECT_WUNLOCK(backing_object);
748 object->backing_object = NULL;
749 vm_object_sub_shadow(backing_object);
754 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
757 VM_OBJECT_ASSERT_WLOCKED(object);
759 atomic_add_int(&backing_object->shadow_count, 1);
760 if ((backing_object->flags & OBJ_ANON) != 0) {
761 VM_OBJECT_ASSERT_WLOCKED(backing_object);
762 LIST_INSERT_HEAD(&backing_object->shadow_head, object,
764 vm_object_set_flag(object, 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;
781 atomic_add_int(&backing_object->shadow_count, 1);
786 * Insert an object into a backing_object's shadow list with an additional
787 * reference to the backing_object added.
790 vm_object_backing_insert_ref(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 KASSERT((backing_object->flags & OBJ_DEAD) == 0,
798 ("shadowing dead anonymous object"));
799 vm_object_reference_locked(backing_object);
800 vm_object_backing_insert_locked(object, backing_object);
801 vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
802 VM_OBJECT_WUNLOCK(backing_object);
804 vm_object_reference(backing_object);
805 atomic_add_int(&backing_object->shadow_count, 1);
806 object->backing_object = backing_object;
811 * Transfer a backing reference from backing_object to object.
814 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
816 vm_object_t new_backing_object;
819 * Note that the reference to backing_object->backing_object
820 * moves from within backing_object to within object.
822 vm_object_backing_remove_locked(object);
823 new_backing_object = backing_object->backing_object;
824 if (new_backing_object == NULL)
826 if ((new_backing_object->flags & OBJ_ANON) != 0) {
827 VM_OBJECT_WLOCK(new_backing_object);
828 vm_object_backing_remove_locked(backing_object);
829 vm_object_backing_insert_locked(object, new_backing_object);
830 VM_OBJECT_WUNLOCK(new_backing_object);
833 * shadow_count for new_backing_object is left
834 * unchanged, its reference provided by backing_object
835 * is replaced by object.
837 object->backing_object = new_backing_object;
838 backing_object->backing_object = NULL;
843 * Wait for a concurrent collapse to settle.
846 vm_object_collapse_wait(vm_object_t object)
849 VM_OBJECT_ASSERT_WLOCKED(object);
851 while ((object->flags & OBJ_COLLAPSING) != 0) {
852 vm_object_pip_wait(object, "vmcolwait");
853 counter_u64_add(object_collapse_waits, 1);
858 * Waits for a backing object to clear a pending collapse and returns
859 * it locked if it is an ANON object.
862 vm_object_backing_collapse_wait(vm_object_t object)
864 vm_object_t backing_object;
866 VM_OBJECT_ASSERT_WLOCKED(object);
869 backing_object = object->backing_object;
870 if (backing_object == NULL ||
871 (backing_object->flags & OBJ_ANON) == 0)
873 VM_OBJECT_WLOCK(backing_object);
874 if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
876 VM_OBJECT_WUNLOCK(object);
877 vm_object_pip_sleep(backing_object, "vmbckwait");
878 counter_u64_add(object_collapse_waits, 1);
879 VM_OBJECT_WLOCK(object);
881 return (backing_object);
885 * vm_object_terminate_pages removes any remaining pageable pages
886 * from the object and resets the object to an empty state.
889 vm_object_terminate_pages(vm_object_t object)
893 VM_OBJECT_ASSERT_WLOCKED(object);
896 * Free any remaining pageable pages. This also removes them from the
897 * paging queues. However, don't free wired pages, just remove them
898 * from the object. Rather than incrementally removing each page from
899 * the object, the page and object are reset to any empty state.
901 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
902 vm_page_assert_unbusied(p);
903 KASSERT(p->object == object &&
904 (p->ref_count & VPRC_OBJREF) != 0,
905 ("vm_object_terminate_pages: page %p is inconsistent", p));
908 if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
915 * If the object contained any pages, then reset it to an empty state.
916 * None of the object's fields, including "resident_page_count", were
917 * modified by the preceding loop.
919 if (object->resident_page_count != 0) {
920 vm_radix_reclaim_allnodes(&object->rtree);
921 TAILQ_INIT(&object->memq);
922 object->resident_page_count = 0;
923 if (object->type == OBJT_VNODE)
924 vdrop(object->handle);
929 * vm_object_terminate actually destroys the specified object, freeing
930 * up all previously used resources.
932 * The object must be locked.
933 * This routine may block.
936 vm_object_terminate(vm_object_t object)
939 VM_OBJECT_ASSERT_WLOCKED(object);
940 KASSERT((object->flags & OBJ_DEAD) != 0,
941 ("terminating non-dead obj %p", object));
942 KASSERT((object->flags & OBJ_COLLAPSING) == 0,
943 ("terminating collapsing obj %p", object));
944 KASSERT(object->backing_object == NULL,
945 ("terminating shadow obj %p", object));
948 * Wait for the pageout daemon and other current users to be
949 * done with the object. Note that new paging_in_progress
950 * users can come after this wait, but they must check
951 * OBJ_DEAD flag set (without unlocking the object), and avoid
952 * the object being terminated.
954 vm_object_pip_wait(object, "objtrm");
956 KASSERT(object->ref_count == 0,
957 ("vm_object_terminate: object with references, ref_count=%d",
960 if ((object->flags & OBJ_PG_DTOR) == 0)
961 vm_object_terminate_pages(object);
963 #if VM_NRESERVLEVEL > 0
964 if (__predict_false(!LIST_EMPTY(&object->rvq)))
965 vm_reserv_break_all(object);
968 KASSERT(object->cred == NULL || (object->flags & OBJ_SWAP) != 0,
969 ("%s: non-swap obj %p has cred", __func__, object));
972 * Let the pager know object is dead.
974 vm_pager_deallocate(object);
975 VM_OBJECT_WUNLOCK(object);
977 vm_object_destroy(object);
981 * Make the page read-only so that we can clear the object flags. However, if
982 * this is a nosync mmap then the object is likely to stay dirty so do not
983 * mess with the page and do not clear the object flags. Returns TRUE if the
984 * page should be flushed, and FALSE otherwise.
987 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
990 vm_page_assert_busied(p);
993 * If we have been asked to skip nosync pages and this is a
994 * nosync page, skip it. Note that the object flags were not
995 * cleared in this case so we do not have to set them.
997 if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
1001 pmap_remove_write(p);
1002 return (p->dirty != 0);
1007 * vm_object_page_clean
1009 * Clean all dirty pages in the specified range of object. Leaves page
1010 * on whatever queue it is currently on. If NOSYNC is set then do not
1011 * write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1012 * leaving the object dirty.
1014 * For swap objects backing tmpfs regular files, do not flush anything,
1015 * but remove write protection on the mapped pages to update mtime through
1018 * When stuffing pages asynchronously, allow clustering. XXX we need a
1019 * synchronous clustering mode implementation.
1021 * Odd semantics: if start == end, we clean everything.
1023 * The object must be locked.
1025 * Returns FALSE if some page from the range was not written, as
1026 * reported by the pager, and TRUE otherwise.
1029 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1033 vm_pindex_t pi, tend, tstart;
1034 int curgeneration, n, pagerflags;
1035 boolean_t eio, res, allclean;
1037 VM_OBJECT_ASSERT_WLOCKED(object);
1039 if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1042 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1043 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1044 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1046 tstart = OFF_TO_IDX(start);
1047 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1048 allclean = tstart == 0 && tend >= object->size;
1052 curgeneration = object->generation;
1054 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1058 np = TAILQ_NEXT(p, listq);
1059 if (vm_page_none_valid(p))
1061 if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1062 if (object->generation != curgeneration &&
1063 (flags & OBJPC_SYNC) != 0)
1065 np = vm_page_find_least(object, pi);
1068 if (!vm_object_page_remove_write(p, flags, &allclean)) {
1072 if (object->type == OBJT_VNODE) {
1073 n = vm_object_page_collect_flush(object, p, pagerflags,
1074 flags, &allclean, &eio);
1079 if (object->generation != curgeneration &&
1080 (flags & OBJPC_SYNC) != 0)
1084 * If the VOP_PUTPAGES() did a truncated write, so
1085 * that even the first page of the run is not fully
1086 * written, vm_pageout_flush() returns 0 as the run
1087 * length. Since the condition that caused truncated
1088 * write may be permanent, e.g. exhausted free space,
1089 * accepting n == 0 would cause an infinite loop.
1091 * Forwarding the iterator leaves the unwritten page
1092 * behind, but there is not much we can do there if
1093 * filesystem refuses to write it.
1103 np = vm_page_find_least(object, pi + n);
1106 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1110 * Leave updating cleangeneration for tmpfs objects to tmpfs
1111 * scan. It needs to update mtime, which happens for other
1112 * filesystems during page writeouts.
1114 if (allclean && object->type == OBJT_VNODE)
1115 object->cleangeneration = curgeneration;
1120 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1121 int flags, boolean_t *allclean, boolean_t *eio)
1123 vm_page_t ma[vm_pageout_page_count], p_first, tp;
1124 int count, i, mreq, runlen;
1126 vm_page_lock_assert(p, MA_NOTOWNED);
1127 vm_page_assert_xbusied(p);
1128 VM_OBJECT_ASSERT_WLOCKED(object);
1133 for (tp = p; count < vm_pageout_page_count; count++) {
1134 tp = vm_page_next(tp);
1135 if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1137 if (!vm_object_page_remove_write(tp, flags, allclean)) {
1138 vm_page_xunbusy(tp);
1143 for (p_first = p; count < vm_pageout_page_count; count++) {
1144 tp = vm_page_prev(p_first);
1145 if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1147 if (!vm_object_page_remove_write(tp, flags, allclean)) {
1148 vm_page_xunbusy(tp);
1155 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1158 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1163 * Note that there is absolutely no sense in writing out
1164 * anonymous objects, so we track down the vnode object
1166 * We invalidate (remove) all pages from the address space
1167 * for semantic correctness.
1169 * If the backing object is a device object with unmanaged pages, then any
1170 * mappings to the specified range of pages must be removed before this
1171 * function is called.
1173 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1174 * may start out with a NULL object.
1177 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1178 boolean_t syncio, boolean_t invalidate)
1180 vm_object_t backing_object;
1183 int error, flags, fsync_after;
1190 VM_OBJECT_WLOCK(object);
1191 while ((backing_object = object->backing_object) != NULL) {
1192 VM_OBJECT_WLOCK(backing_object);
1193 offset += object->backing_object_offset;
1194 VM_OBJECT_WUNLOCK(object);
1195 object = backing_object;
1196 if (object->size < OFF_TO_IDX(offset + size))
1197 size = IDX_TO_OFF(object->size) - offset;
1200 * Flush pages if writing is allowed, invalidate them
1201 * if invalidation requested. Pages undergoing I/O
1202 * will be ignored by vm_object_page_remove().
1204 * We cannot lock the vnode and then wait for paging
1205 * to complete without deadlocking against vm_fault.
1206 * Instead we simply call vm_object_page_remove() and
1207 * allow it to block internally on a page-by-page
1208 * basis when it encounters pages undergoing async
1211 if (object->type == OBJT_VNODE &&
1212 vm_object_mightbedirty(object) != 0 &&
1213 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1214 VM_OBJECT_WUNLOCK(object);
1215 (void) vn_start_write(vp, &mp, V_WAIT);
1216 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1217 if (syncio && !invalidate && offset == 0 &&
1218 atop(size) == object->size) {
1220 * If syncing the whole mapping of the file,
1221 * it is faster to schedule all the writes in
1222 * async mode, also allowing the clustering,
1223 * and then wait for i/o to complete.
1228 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1229 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1230 fsync_after = FALSE;
1232 VM_OBJECT_WLOCK(object);
1233 res = vm_object_page_clean(object, offset, offset + size,
1235 VM_OBJECT_WUNLOCK(object);
1237 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1239 vn_finished_write(mp);
1242 VM_OBJECT_WLOCK(object);
1244 if ((object->type == OBJT_VNODE ||
1245 object->type == OBJT_DEVICE) && invalidate) {
1246 if (object->type == OBJT_DEVICE)
1248 * The option OBJPR_NOTMAPPED must be passed here
1249 * because vm_object_page_remove() cannot remove
1250 * unmanaged mappings.
1252 flags = OBJPR_NOTMAPPED;
1256 flags = OBJPR_CLEANONLY;
1257 vm_object_page_remove(object, OFF_TO_IDX(offset),
1258 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1260 VM_OBJECT_WUNLOCK(object);
1265 * Determine whether the given advice can be applied to the object. Advice is
1266 * not applied to unmanaged pages since they never belong to page queues, and
1267 * since MADV_FREE is destructive, it can apply only to anonymous pages that
1268 * have been mapped at most once.
1271 vm_object_advice_applies(vm_object_t object, int advice)
1274 if ((object->flags & OBJ_UNMANAGED) != 0)
1276 if (advice != MADV_FREE)
1278 return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1279 (OBJ_ONEMAPPING | OBJ_ANON));
1283 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1287 if (advice == MADV_FREE)
1288 vm_pager_freespace(object, pindex, size);
1292 * vm_object_madvise:
1294 * Implements the madvise function at the object/page level.
1296 * MADV_WILLNEED (any object)
1298 * Activate the specified pages if they are resident.
1300 * MADV_DONTNEED (any object)
1302 * Deactivate the specified pages if they are resident.
1304 * MADV_FREE (OBJT_SWAP objects, OBJ_ONEMAPPING only)
1306 * Deactivate and clean the specified pages if they are
1307 * resident. This permits the process to reuse the pages
1308 * without faulting or the kernel to reclaim the pages
1312 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1315 vm_pindex_t tpindex;
1316 vm_object_t backing_object, tobject;
1323 VM_OBJECT_WLOCK(object);
1324 if (!vm_object_advice_applies(object, advice)) {
1325 VM_OBJECT_WUNLOCK(object);
1328 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1332 * If the next page isn't resident in the top-level object, we
1333 * need to search the shadow chain. When applying MADV_FREE, we
1334 * take care to release any swap space used to store
1335 * non-resident pages.
1337 if (m == NULL || pindex < m->pindex) {
1339 * Optimize a common case: if the top-level object has
1340 * no backing object, we can skip over the non-resident
1341 * range in constant time.
1343 if (object->backing_object == NULL) {
1344 tpindex = (m != NULL && m->pindex < end) ?
1346 vm_object_madvise_freespace(object, advice,
1347 pindex, tpindex - pindex);
1348 if ((pindex = tpindex) == end)
1355 vm_object_madvise_freespace(tobject, advice,
1358 * Prepare to search the next object in the
1361 backing_object = tobject->backing_object;
1362 if (backing_object == NULL)
1364 VM_OBJECT_WLOCK(backing_object);
1366 OFF_TO_IDX(tobject->backing_object_offset);
1367 if (tobject != object)
1368 VM_OBJECT_WUNLOCK(tobject);
1369 tobject = backing_object;
1370 if (!vm_object_advice_applies(tobject, advice))
1372 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
1377 m = TAILQ_NEXT(m, listq);
1381 * If the page is not in a normal state, skip it. The page
1382 * can not be invalidated while the object lock is held.
1384 if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1386 KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1387 ("vm_object_madvise: page %p is fictitious", tm));
1388 KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1389 ("vm_object_madvise: page %p is not managed", tm));
1390 if (vm_page_tryxbusy(tm) == 0) {
1391 if (object != tobject)
1392 VM_OBJECT_WUNLOCK(object);
1393 if (advice == MADV_WILLNEED) {
1395 * Reference the page before unlocking and
1396 * sleeping so that the page daemon is less
1397 * likely to reclaim it.
1399 vm_page_aflag_set(tm, PGA_REFERENCED);
1401 if (!vm_page_busy_sleep(tm, "madvpo", 0))
1402 VM_OBJECT_WUNLOCK(tobject);
1405 vm_page_advise(tm, advice);
1406 vm_page_xunbusy(tm);
1407 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1409 if (tobject != object)
1410 VM_OBJECT_WUNLOCK(tobject);
1412 VM_OBJECT_WUNLOCK(object);
1418 * Create a new object which is backed by the
1419 * specified existing object range. The source
1420 * object reference is deallocated.
1422 * The new object and offset into that object
1423 * are returned in the source parameters.
1426 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1427 struct ucred *cred, bool shared)
1435 * Don't create the new object if the old object isn't shared.
1437 * If we hold the only reference we can guarantee that it won't
1438 * increase while we have the map locked. Otherwise the race is
1439 * harmless and we will end up with an extra shadow object that
1440 * will be collapsed later.
1442 if (source != NULL && source->ref_count == 1 &&
1443 (source->flags & OBJ_ANON) != 0)
1447 * Allocate a new object with the given length.
1449 result = vm_object_allocate_anon(atop(length), source, cred, length);
1452 * Store the offset into the source object, and fix up the offset into
1455 result->backing_object_offset = *offset;
1457 if (shared || source != NULL) {
1458 VM_OBJECT_WLOCK(result);
1461 * The new object shadows the source object, adding a
1462 * reference to it. Our caller changes his reference
1463 * to point to the new object, removing a reference to
1464 * the source object. Net result: no change of
1465 * reference count, unless the caller needs to add one
1466 * more reference due to forking a shared map entry.
1469 vm_object_reference_locked(result);
1470 vm_object_clear_flag(result, OBJ_ONEMAPPING);
1474 * Try to optimize the result object's page color when
1475 * shadowing in order to maintain page coloring
1476 * consistency in the combined shadowed object.
1478 if (source != NULL) {
1479 vm_object_backing_insert(result, source);
1480 result->domain = source->domain;
1481 #if VM_NRESERVLEVEL > 0
1482 vm_object_set_flag(result,
1483 (source->flags & OBJ_COLORED));
1484 result->pg_color = (source->pg_color +
1485 OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1489 VM_OBJECT_WUNLOCK(result);
1493 * Return the new things
1502 * Split the pages in a map entry into a new object. This affords
1503 * easier removal of unused pages, and keeps object inheritance from
1504 * being a negative impact on memory usage.
1507 vm_object_split(vm_map_entry_t entry)
1509 vm_page_t m, m_next;
1510 vm_object_t orig_object, new_object, backing_object;
1511 vm_pindex_t idx, offidxstart;
1514 orig_object = entry->object.vm_object;
1515 KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1516 ("vm_object_split: Splitting object with multiple mappings."));
1517 if ((orig_object->flags & OBJ_ANON) == 0)
1519 if (orig_object->ref_count <= 1)
1521 VM_OBJECT_WUNLOCK(orig_object);
1523 offidxstart = OFF_TO_IDX(entry->offset);
1524 size = atop(entry->end - entry->start);
1526 new_object = vm_object_allocate_anon(size, orig_object,
1527 orig_object->cred, ptoa(size));
1530 * We must wait for the orig_object to complete any in-progress
1531 * collapse so that the swap blocks are stable below. The
1532 * additional reference on backing_object by new object will
1533 * prevent further collapse operations until split completes.
1535 VM_OBJECT_WLOCK(orig_object);
1536 vm_object_collapse_wait(orig_object);
1539 * At this point, the new object is still private, so the order in
1540 * which the original and new objects are locked does not matter.
1542 VM_OBJECT_WLOCK(new_object);
1543 new_object->domain = orig_object->domain;
1544 backing_object = orig_object->backing_object;
1545 if (backing_object != NULL) {
1546 vm_object_backing_insert_ref(new_object, backing_object);
1547 new_object->backing_object_offset =
1548 orig_object->backing_object_offset + entry->offset;
1550 if (orig_object->cred != NULL) {
1551 crhold(orig_object->cred);
1552 KASSERT(orig_object->charge >= ptoa(size),
1553 ("orig_object->charge < 0"));
1554 orig_object->charge -= ptoa(size);
1558 * Mark the split operation so that swap_pager_getpages() knows
1559 * that the object is in transition.
1561 vm_object_set_flag(orig_object, OBJ_SPLIT);
1566 m = vm_page_find_least(orig_object, offidxstart);
1567 KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1568 ("%s: object %p was repopulated", __func__, orig_object));
1569 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1571 m_next = TAILQ_NEXT(m, listq);
1574 * We must wait for pending I/O to complete before we can
1577 * We do not have to VM_PROT_NONE the page as mappings should
1578 * not be changed by this operation.
1580 if (vm_page_tryxbusy(m) == 0) {
1581 VM_OBJECT_WUNLOCK(new_object);
1582 if (vm_page_busy_sleep(m, "spltwt", 0))
1583 VM_OBJECT_WLOCK(orig_object);
1584 VM_OBJECT_WLOCK(new_object);
1589 * The page was left invalid. Likely placed there by
1590 * an incomplete fault. Just remove and ignore.
1592 if (vm_page_none_valid(m)) {
1593 if (vm_page_remove(m))
1598 /* vm_page_rename() will dirty the page. */
1599 if (vm_page_rename(m, new_object, idx)) {
1601 VM_OBJECT_WUNLOCK(new_object);
1602 VM_OBJECT_WUNLOCK(orig_object);
1604 VM_OBJECT_WLOCK(orig_object);
1605 VM_OBJECT_WLOCK(new_object);
1609 #if VM_NRESERVLEVEL > 0
1611 * If some of the reservation's allocated pages remain with
1612 * the original object, then transferring the reservation to
1613 * the new object is neither particularly beneficial nor
1614 * particularly harmful as compared to leaving the reservation
1615 * with the original object. If, however, all of the
1616 * reservation's allocated pages are transferred to the new
1617 * object, then transferring the reservation is typically
1618 * beneficial. Determining which of these two cases applies
1619 * would be more costly than unconditionally renaming the
1622 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1627 * swap_pager_copy() can sleep, in which case the orig_object's
1628 * and new_object's locks are released and reacquired.
1630 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1632 TAILQ_FOREACH(m, &new_object->memq, listq)
1635 vm_object_clear_flag(orig_object, OBJ_SPLIT);
1636 VM_OBJECT_WUNLOCK(orig_object);
1637 VM_OBJECT_WUNLOCK(new_object);
1638 entry->object.vm_object = new_object;
1639 entry->offset = 0LL;
1640 vm_object_deallocate(orig_object);
1641 VM_OBJECT_WLOCK(new_object);
1645 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1647 vm_object_t backing_object;
1649 VM_OBJECT_ASSERT_WLOCKED(object);
1650 backing_object = object->backing_object;
1651 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1653 KASSERT(p == NULL || p->object == object || p->object == backing_object,
1654 ("invalid ownership %p %p %p", p, object, backing_object));
1655 /* The page is only NULL when rename fails. */
1657 VM_OBJECT_WUNLOCK(object);
1658 VM_OBJECT_WUNLOCK(backing_object);
1660 VM_OBJECT_WLOCK(object);
1661 } else if (p->object == object) {
1662 VM_OBJECT_WUNLOCK(backing_object);
1663 if (vm_page_busy_sleep(p, "vmocol", 0))
1664 VM_OBJECT_WLOCK(object);
1666 VM_OBJECT_WUNLOCK(object);
1667 if (!vm_page_busy_sleep(p, "vmocol", 0))
1668 VM_OBJECT_WUNLOCK(backing_object);
1669 VM_OBJECT_WLOCK(object);
1671 VM_OBJECT_WLOCK(backing_object);
1672 return (TAILQ_FIRST(&backing_object->memq));
1676 vm_object_scan_all_shadowed(vm_object_t object)
1678 vm_object_t backing_object;
1680 vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1682 VM_OBJECT_ASSERT_WLOCKED(object);
1683 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1685 backing_object = object->backing_object;
1687 if ((backing_object->flags & OBJ_ANON) == 0)
1690 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1691 p = vm_page_find_least(backing_object, pi);
1692 ps = swap_pager_find_least(backing_object, pi);
1695 * Only check pages inside the parent object's range and
1696 * inside the parent object's mapping of the backing object.
1699 if (p != NULL && p->pindex < pi)
1700 p = TAILQ_NEXT(p, listq);
1702 ps = swap_pager_find_least(backing_object, pi);
1703 if (p == NULL && ps >= backing_object->size)
1708 pi = MIN(p->pindex, ps);
1710 new_pindex = pi - backing_offset_index;
1711 if (new_pindex >= object->size)
1716 * If the backing object page is busy a
1717 * grandparent or older page may still be
1718 * undergoing CoW. It is not safe to collapse
1719 * the backing object until it is quiesced.
1721 if (vm_page_tryxbusy(p) == 0)
1725 * We raced with the fault handler that left
1726 * newly allocated invalid page on the object
1727 * queue and retried.
1729 if (!vm_page_all_valid(p))
1734 * See if the parent has the page or if the parent's object
1735 * pager has the page. If the parent has the page but the page
1736 * is not valid, the parent's object pager must have the page.
1738 * If this fails, the parent does not completely shadow the
1739 * object and we might as well give up now.
1741 pp = vm_page_lookup(object, new_pindex);
1744 * The valid check here is stable due to object lock
1745 * being required to clear valid and initiate paging.
1746 * Busy of p disallows fault handler to validate pp.
1748 if ((pp == NULL || vm_page_none_valid(pp)) &&
1749 !vm_pager_has_page(object, new_pindex, NULL, NULL))
1763 vm_object_collapse_scan(vm_object_t object)
1765 vm_object_t backing_object;
1766 vm_page_t next, p, pp;
1767 vm_pindex_t backing_offset_index, new_pindex;
1769 VM_OBJECT_ASSERT_WLOCKED(object);
1770 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1772 backing_object = object->backing_object;
1773 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1778 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1779 next = TAILQ_NEXT(p, listq);
1780 new_pindex = p->pindex - backing_offset_index;
1783 * Check for busy page
1785 if (vm_page_tryxbusy(p) == 0) {
1786 next = vm_object_collapse_scan_wait(object, p);
1790 KASSERT(object->backing_object == backing_object,
1791 ("vm_object_collapse_scan: backing object mismatch %p != %p",
1792 object->backing_object, backing_object));
1793 KASSERT(p->object == backing_object,
1794 ("vm_object_collapse_scan: object mismatch %p != %p",
1795 p->object, backing_object));
1797 if (p->pindex < backing_offset_index ||
1798 new_pindex >= object->size) {
1799 vm_pager_freespace(backing_object, p->pindex, 1);
1801 KASSERT(!pmap_page_is_mapped(p),
1802 ("freeing mapped page %p", p));
1803 if (vm_page_remove(p))
1808 if (!vm_page_all_valid(p)) {
1809 KASSERT(!pmap_page_is_mapped(p),
1810 ("freeing mapped page %p", p));
1811 if (vm_page_remove(p))
1816 pp = vm_page_lookup(object, new_pindex);
1817 if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1820 * The page in the parent is busy and possibly not
1821 * (yet) valid. Until its state is finalized by the
1822 * busy bit owner, we can't tell whether it shadows the
1825 next = vm_object_collapse_scan_wait(object, pp);
1829 if (pp != NULL && vm_page_none_valid(pp)) {
1831 * The page was invalid in the parent. Likely placed
1832 * there by an incomplete fault. Just remove and
1833 * ignore. p can replace it.
1835 if (vm_page_remove(pp))
1840 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1843 * The page already exists in the parent OR swap exists
1844 * for this location in the parent. Leave the parent's
1845 * page alone. Destroy the original page from the
1848 vm_pager_freespace(backing_object, p->pindex, 1);
1849 KASSERT(!pmap_page_is_mapped(p),
1850 ("freeing mapped page %p", p));
1851 if (vm_page_remove(p))
1854 vm_page_xunbusy(pp);
1859 * Page does not exist in parent, rename the page from the
1860 * backing object to the main object.
1862 * If the page was mapped to a process, it can remain mapped
1863 * through the rename. vm_page_rename() will dirty the page.
1865 if (vm_page_rename(p, object, new_pindex)) {
1867 next = vm_object_collapse_scan_wait(object, NULL);
1871 /* Use the old pindex to free the right page. */
1872 vm_pager_freespace(backing_object, new_pindex +
1873 backing_offset_index, 1);
1875 #if VM_NRESERVLEVEL > 0
1877 * Rename the reservation.
1879 vm_reserv_rename(p, object, backing_object,
1880 backing_offset_index);
1888 * vm_object_collapse:
1890 * Collapse an object with the object backing it.
1891 * Pages in the backing object are moved into the
1892 * parent, and the backing object is deallocated.
1895 vm_object_collapse(vm_object_t object)
1897 vm_object_t backing_object, new_backing_object;
1899 VM_OBJECT_ASSERT_WLOCKED(object);
1902 KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1903 ("collapsing invalid object"));
1906 * Wait for the backing_object to finish any pending
1907 * collapse so that the caller sees the shortest possible
1910 backing_object = vm_object_backing_collapse_wait(object);
1911 if (backing_object == NULL)
1914 KASSERT(object->ref_count > 0 &&
1915 object->ref_count > atomic_load_int(&object->shadow_count),
1916 ("collapse with invalid ref %d or shadow %d count.",
1917 object->ref_count, atomic_load_int(&object->shadow_count)));
1918 KASSERT((backing_object->flags &
1919 (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1920 ("vm_object_collapse: Backing object already collapsing."));
1921 KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1922 ("vm_object_collapse: object is already collapsing."));
1925 * We know that we can either collapse the backing object if
1926 * the parent is the only reference to it, or (perhaps) have
1927 * the parent bypass the object if the parent happens to shadow
1928 * all the resident pages in the entire backing object.
1930 if (backing_object->ref_count == 1) {
1931 KASSERT(atomic_load_int(&backing_object->shadow_count)
1933 ("vm_object_collapse: shadow_count: %d",
1934 atomic_load_int(&backing_object->shadow_count)));
1935 vm_object_pip_add(object, 1);
1936 vm_object_set_flag(object, OBJ_COLLAPSING);
1937 vm_object_pip_add(backing_object, 1);
1938 vm_object_set_flag(backing_object, OBJ_DEAD);
1941 * If there is exactly one reference to the backing
1942 * object, we can collapse it into the parent.
1944 vm_object_collapse_scan(object);
1946 #if VM_NRESERVLEVEL > 0
1948 * Break any reservations from backing_object.
1950 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1951 vm_reserv_break_all(backing_object);
1955 * Move the pager from backing_object to object.
1957 * swap_pager_copy() can sleep, in which case the
1958 * backing_object's and object's locks are released and
1961 swap_pager_copy(backing_object, object,
1962 OFF_TO_IDX(object->backing_object_offset), TRUE);
1965 * Object now shadows whatever backing_object did.
1967 vm_object_clear_flag(object, OBJ_COLLAPSING);
1968 vm_object_backing_transfer(object, backing_object);
1969 object->backing_object_offset +=
1970 backing_object->backing_object_offset;
1971 VM_OBJECT_WUNLOCK(object);
1972 vm_object_pip_wakeup(object);
1975 * Discard backing_object.
1977 * Since the backing object has no pages, no pager left,
1978 * and no object references within it, all that is
1979 * necessary is to dispose of it.
1981 KASSERT(backing_object->ref_count == 1, (
1982 "backing_object %p was somehow re-referenced during collapse!",
1984 vm_object_pip_wakeup(backing_object);
1985 (void)refcount_release(&backing_object->ref_count);
1986 vm_object_terminate(backing_object);
1987 counter_u64_add(object_collapses, 1);
1988 VM_OBJECT_WLOCK(object);
1991 * If we do not entirely shadow the backing object,
1992 * there is nothing we can do so we give up.
1994 * The object lock and backing_object lock must not
1995 * be dropped during this sequence.
1997 if (!vm_object_scan_all_shadowed(object)) {
1998 VM_OBJECT_WUNLOCK(backing_object);
2003 * Make the parent shadow the next object in the
2004 * chain. Deallocating backing_object will not remove
2005 * it, since its reference count is at least 2.
2007 vm_object_backing_remove_locked(object);
2008 new_backing_object = backing_object->backing_object;
2009 if (new_backing_object != NULL) {
2010 vm_object_backing_insert_ref(object,
2011 new_backing_object);
2012 object->backing_object_offset +=
2013 backing_object->backing_object_offset;
2017 * Drop the reference count on backing_object. Since
2018 * its ref_count was at least 2, it will not vanish.
2020 (void)refcount_release(&backing_object->ref_count);
2021 KASSERT(backing_object->ref_count >= 1, (
2022 "backing_object %p was somehow dereferenced during collapse!",
2024 VM_OBJECT_WUNLOCK(backing_object);
2025 counter_u64_add(object_bypasses, 1);
2029 * Try again with this object's new backing object.
2035 * vm_object_page_remove:
2037 * For the given object, either frees or invalidates each of the
2038 * specified pages. In general, a page is freed. However, if a page is
2039 * wired for any reason other than the existence of a managed, wired
2040 * mapping, then it may be invalidated but not removed from the object.
2041 * Pages are specified by the given range ["start", "end") and the option
2042 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
2043 * extends from "start" to the end of the object. If the option
2044 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2045 * specified range are affected. If the option OBJPR_NOTMAPPED is
2046 * specified, then the pages within the specified range must have no
2047 * mappings. Otherwise, if this option is not specified, any mappings to
2048 * the specified pages are removed before the pages are freed or
2051 * In general, this operation should only be performed on objects that
2052 * contain managed pages. There are, however, two exceptions. First, it
2053 * is performed on the kernel and kmem objects by vm_map_entry_delete().
2054 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2055 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
2056 * not be specified and the option OBJPR_NOTMAPPED must be specified.
2058 * The object must be locked.
2061 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2066 VM_OBJECT_ASSERT_WLOCKED(object);
2067 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2068 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2069 ("vm_object_page_remove: illegal options for object %p", object));
2070 if (object->resident_page_count == 0)
2072 vm_object_pip_add(object, 1);
2074 p = vm_page_find_least(object, start);
2077 * Here, the variable "p" is either (1) the page with the least pindex
2078 * greater than or equal to the parameter "start" or (2) NULL.
2080 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2081 next = TAILQ_NEXT(p, listq);
2084 * Skip invalid pages if asked to do so. Try to avoid acquiring
2085 * the busy lock, as some consumers rely on this to avoid
2088 * A thread may concurrently transition the page from invalid to
2089 * valid using only the busy lock, so the result of this check
2090 * is immediately stale. It is up to consumers to handle this,
2091 * for instance by ensuring that all invalid->valid transitions
2092 * happen with a mutex held, as may be possible for a
2095 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2099 * If the page is wired for any reason besides the existence
2100 * of managed, wired mappings, then it cannot be freed. For
2101 * example, fictitious pages, which represent device memory,
2102 * are inherently wired and cannot be freed. They can,
2103 * however, be invalidated if the option OBJPR_CLEANONLY is
2106 if (vm_page_tryxbusy(p) == 0) {
2107 if (vm_page_busy_sleep(p, "vmopar", 0))
2108 VM_OBJECT_WLOCK(object);
2111 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2115 if (vm_page_wired(p)) {
2117 if ((options & OBJPR_NOTMAPPED) == 0 &&
2118 object->ref_count != 0)
2120 if ((options & OBJPR_CLEANONLY) == 0) {
2127 KASSERT((p->flags & PG_FICTITIOUS) == 0,
2128 ("vm_object_page_remove: page %p is fictitious", p));
2129 if ((options & OBJPR_CLEANONLY) != 0 &&
2130 !vm_page_none_valid(p)) {
2131 if ((options & OBJPR_NOTMAPPED) == 0 &&
2132 object->ref_count != 0 &&
2133 !vm_page_try_remove_write(p))
2135 if (p->dirty != 0) {
2140 if ((options & OBJPR_NOTMAPPED) == 0 &&
2141 object->ref_count != 0 && !vm_page_try_remove_all(p))
2145 vm_object_pip_wakeup(object);
2147 vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2152 * vm_object_page_noreuse:
2154 * For the given object, attempt to move the specified pages to
2155 * the head of the inactive queue. This bypasses regular LRU
2156 * operation and allows the pages to be reused quickly under memory
2157 * pressure. If a page is wired for any reason, then it will not
2158 * be queued. Pages are specified by the range ["start", "end").
2159 * As a special case, if "end" is zero, then the range extends from
2160 * "start" to the end of the object.
2162 * This operation should only be performed on objects that
2163 * contain non-fictitious, managed pages.
2165 * The object must be locked.
2168 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2172 VM_OBJECT_ASSERT_LOCKED(object);
2173 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2174 ("vm_object_page_noreuse: illegal object %p", object));
2175 if (object->resident_page_count == 0)
2177 p = vm_page_find_least(object, start);
2180 * Here, the variable "p" is either (1) the page with the least pindex
2181 * greater than or equal to the parameter "start" or (2) NULL.
2183 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2184 next = TAILQ_NEXT(p, listq);
2185 vm_page_deactivate_noreuse(p);
2190 * Populate the specified range of the object with valid pages. Returns
2191 * TRUE if the range is successfully populated and FALSE otherwise.
2193 * Note: This function should be optimized to pass a larger array of
2194 * pages to vm_pager_get_pages() before it is applied to a non-
2195 * OBJT_DEVICE object.
2197 * The object must be locked.
2200 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2206 VM_OBJECT_ASSERT_WLOCKED(object);
2207 for (pindex = start; pindex < end; pindex++) {
2208 rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2209 if (rv != VM_PAGER_OK)
2213 * Keep "m" busy because a subsequent iteration may unlock
2217 if (pindex > start) {
2218 m = vm_page_lookup(object, start);
2219 while (m != NULL && m->pindex < pindex) {
2221 m = TAILQ_NEXT(m, listq);
2224 return (pindex == end);
2228 * Routine: vm_object_coalesce
2229 * Function: Coalesces two objects backing up adjoining
2230 * regions of memory into a single object.
2232 * returns TRUE if objects were combined.
2234 * NOTE: Only works at the moment if the second object is NULL -
2235 * if it's not, which object do we lock first?
2238 * prev_object First object to coalesce
2239 * prev_offset Offset into prev_object
2240 * prev_size Size of reference to prev_object
2241 * next_size Size of reference to the second object
2242 * reserved Indicator that extension region has
2243 * swap accounted for
2246 * The object must *not* be locked.
2249 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2250 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2252 vm_pindex_t next_pindex;
2254 if (prev_object == NULL)
2256 if ((prev_object->flags & OBJ_ANON) == 0)
2259 VM_OBJECT_WLOCK(prev_object);
2261 * Try to collapse the object first.
2263 vm_object_collapse(prev_object);
2266 * Can't coalesce if: . more than one reference . paged out . shadows
2267 * another object . has a copy elsewhere (any of which mean that the
2268 * pages not mapped to prev_entry may be in use anyway)
2270 if (prev_object->backing_object != NULL) {
2271 VM_OBJECT_WUNLOCK(prev_object);
2275 prev_size >>= PAGE_SHIFT;
2276 next_size >>= PAGE_SHIFT;
2277 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2279 if (prev_object->ref_count > 1 &&
2280 prev_object->size != next_pindex &&
2281 (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2282 VM_OBJECT_WUNLOCK(prev_object);
2287 * Account for the charge.
2289 if (prev_object->cred != NULL) {
2291 * If prev_object was charged, then this mapping,
2292 * although not charged now, may become writable
2293 * later. Non-NULL cred in the object would prevent
2294 * swap reservation during enabling of the write
2295 * access, so reserve swap now. Failed reservation
2296 * cause allocation of the separate object for the map
2297 * entry, and swap reservation for this entry is
2298 * managed in appropriate time.
2300 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2301 prev_object->cred)) {
2302 VM_OBJECT_WUNLOCK(prev_object);
2305 prev_object->charge += ptoa(next_size);
2309 * Remove any pages that may still be in the object from a previous
2312 if (next_pindex < prev_object->size) {
2313 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2316 if (prev_object->cred != NULL) {
2317 KASSERT(prev_object->charge >=
2318 ptoa(prev_object->size - next_pindex),
2319 ("object %p overcharged 1 %jx %jx", prev_object,
2320 (uintmax_t)next_pindex, (uintmax_t)next_size));
2321 prev_object->charge -= ptoa(prev_object->size -
2328 * Extend the object if necessary.
2330 if (next_pindex + next_size > prev_object->size)
2331 prev_object->size = next_pindex + next_size;
2333 VM_OBJECT_WUNLOCK(prev_object);
2338 vm_object_set_writeable_dirty_(vm_object_t object)
2340 atomic_add_int(&object->generation, 1);
2344 vm_object_mightbedirty_(vm_object_t object)
2346 return (object->generation != object->cleangeneration);
2352 * For each page offset within the specified range of the given object,
2353 * find the highest-level page in the shadow chain and unwire it. A page
2354 * must exist at every page offset, and the highest-level page must be
2358 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2361 vm_object_t tobject, t1object;
2363 vm_pindex_t end_pindex, pindex, tpindex;
2364 int depth, locked_depth;
2366 KASSERT((offset & PAGE_MASK) == 0,
2367 ("vm_object_unwire: offset is not page aligned"));
2368 KASSERT((length & PAGE_MASK) == 0,
2369 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2370 /* The wired count of a fictitious page never changes. */
2371 if ((object->flags & OBJ_FICTITIOUS) != 0)
2373 pindex = OFF_TO_IDX(offset);
2374 end_pindex = pindex + atop(length);
2377 VM_OBJECT_RLOCK(object);
2378 m = vm_page_find_least(object, pindex);
2379 while (pindex < end_pindex) {
2380 if (m == NULL || pindex < m->pindex) {
2382 * The first object in the shadow chain doesn't
2383 * contain a page at the current index. Therefore,
2384 * the page must exist in a backing object.
2391 OFF_TO_IDX(tobject->backing_object_offset);
2392 tobject = tobject->backing_object;
2393 KASSERT(tobject != NULL,
2394 ("vm_object_unwire: missing page"));
2395 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2398 if (depth == locked_depth) {
2400 VM_OBJECT_RLOCK(tobject);
2402 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2406 m = TAILQ_NEXT(m, listq);
2408 if (vm_page_trysbusy(tm) == 0) {
2409 for (tobject = object; locked_depth >= 1;
2411 t1object = tobject->backing_object;
2412 if (tm->object != tobject)
2413 VM_OBJECT_RUNLOCK(tobject);
2416 tobject = tm->object;
2417 if (!vm_page_busy_sleep(tm, "unwbo",
2418 VM_ALLOC_IGN_SBUSY))
2419 VM_OBJECT_RUNLOCK(tobject);
2422 vm_page_unwire(tm, queue);
2423 vm_page_sunbusy(tm);
2427 /* Release the accumulated object locks. */
2428 for (tobject = object; locked_depth >= 1; locked_depth--) {
2429 t1object = tobject->backing_object;
2430 VM_OBJECT_RUNLOCK(tobject);
2436 * Return the vnode for the given object, or NULL if none exists.
2437 * For tmpfs objects, the function may return NULL if there is
2438 * no vnode allocated at the time of the call.
2441 vm_object_vnode(vm_object_t object)
2445 VM_OBJECT_ASSERT_LOCKED(object);
2446 vm_pager_getvp(object, &vp, NULL);
2451 * Busy the vm object. This prevents new pages belonging to the object from
2452 * becoming busy. Existing pages persist as busy. Callers are responsible
2453 * for checking page state before proceeding.
2456 vm_object_busy(vm_object_t obj)
2459 VM_OBJECT_ASSERT_LOCKED(obj);
2461 blockcount_acquire(&obj->busy, 1);
2462 /* The fence is required to order loads of page busy. */
2463 atomic_thread_fence_acq_rel();
2467 vm_object_unbusy(vm_object_t obj)
2470 blockcount_release(&obj->busy, 1);
2474 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2477 VM_OBJECT_ASSERT_UNLOCKED(obj);
2479 (void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2483 * This function aims to determine if the object is mapped,
2484 * specifically, if it is referenced by a vm_map_entry. Because
2485 * objects occasionally acquire transient references that do not
2486 * represent a mapping, the method used here is inexact. However, it
2487 * has very low overhead and is good enough for the advisory
2488 * vm.vmtotal sysctl.
2491 vm_object_is_active(vm_object_t obj)
2494 return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2498 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2500 struct kinfo_vmobject *kvo;
2501 char *fullpath, *freepath;
2509 if (req->oldptr == NULL) {
2511 * If an old buffer has not been provided, generate an
2512 * estimate of the space needed for a subsequent call.
2514 mtx_lock(&vm_object_list_mtx);
2516 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2517 if (obj->type == OBJT_DEAD)
2521 mtx_unlock(&vm_object_list_mtx);
2522 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2526 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK);
2530 * VM objects are type stable and are never removed from the
2531 * list once added. This allows us to safely read obj->object_list
2532 * after reacquiring the VM object lock.
2534 mtx_lock(&vm_object_list_mtx);
2535 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2536 if (obj->type == OBJT_DEAD ||
2537 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2539 VM_OBJECT_RLOCK(obj);
2540 if (obj->type == OBJT_DEAD ||
2541 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2542 VM_OBJECT_RUNLOCK(obj);
2545 mtx_unlock(&vm_object_list_mtx);
2546 kvo->kvo_size = ptoa(obj->size);
2547 kvo->kvo_resident = obj->resident_page_count;
2548 kvo->kvo_ref_count = obj->ref_count;
2549 kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2550 kvo->kvo_memattr = obj->memattr;
2551 kvo->kvo_active = 0;
2552 kvo->kvo_inactive = 0;
2554 TAILQ_FOREACH(m, &obj->memq, listq) {
2556 * A page may belong to the object but be
2557 * dequeued and set to PQ_NONE while the
2558 * object lock is not held. This makes the
2559 * reads of m->queue below racy, and we do not
2560 * count pages set to PQ_NONE. However, this
2561 * sysctl is only meant to give an
2562 * approximation of the system anyway.
2564 if (m->a.queue == PQ_ACTIVE)
2566 else if (m->a.queue == PQ_INACTIVE)
2567 kvo->kvo_inactive++;
2571 kvo->kvo_vn_fileid = 0;
2572 kvo->kvo_vn_fsid = 0;
2573 kvo->kvo_vn_fsid_freebsd11 = 0;
2577 kvo->kvo_type = vm_object_kvme_type(obj, swap_only ? NULL : &vp);
2580 } else if ((obj->flags & OBJ_ANON) != 0) {
2581 MPASS(kvo->kvo_type == KVME_TYPE_DEFAULT ||
2582 kvo->kvo_type == KVME_TYPE_SWAP);
2583 kvo->kvo_me = (uintptr_t)obj;
2584 /* tmpfs objs are reported as vnodes */
2585 kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2586 sp = swap_pager_swapped_pages(obj);
2587 kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2589 VM_OBJECT_RUNLOCK(obj);
2591 vn_fullpath(vp, &fullpath, &freepath);
2592 vn_lock(vp, LK_SHARED | LK_RETRY);
2593 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2594 kvo->kvo_vn_fileid = va.va_fileid;
2595 kvo->kvo_vn_fsid = va.va_fsid;
2596 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2602 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2603 if (freepath != NULL)
2604 free(freepath, M_TEMP);
2606 /* Pack record size down */
2607 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2608 + strlen(kvo->kvo_path) + 1;
2609 kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2611 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2613 mtx_lock(&vm_object_list_mtx);
2617 mtx_unlock(&vm_object_list_mtx);
2623 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2625 return (vm_object_list_handler(req, false));
2628 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2629 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2630 "List of VM objects");
2633 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2635 return (vm_object_list_handler(req, true));
2639 * This sysctl returns list of the anonymous or swap objects. Intent
2640 * is to provide stripped optimized list useful to analyze swap use.
2641 * Since technically non-swap (default) objects participate in the
2642 * shadow chains, and are converted to swap type as needed by swap
2643 * pager, we must report them.
2645 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2646 CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2647 sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2648 "List of swap VM objects");
2650 #include "opt_ddb.h"
2652 #include <sys/kernel.h>
2654 #include <sys/cons.h>
2656 #include <ddb/ddb.h>
2659 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2662 vm_map_entry_t tmpe;
2669 VM_MAP_ENTRY_FOREACH(tmpe, map) {
2670 if (_vm_object_in_map(map, object, tmpe)) {
2674 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2675 tmpm = entry->object.sub_map;
2676 VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2677 if (_vm_object_in_map(tmpm, object, tmpe)) {
2681 } else if ((obj = entry->object.vm_object) != NULL) {
2682 for (; obj; obj = obj->backing_object)
2683 if (obj == object) {
2691 vm_object_in_map(vm_object_t object)
2695 /* sx_slock(&allproc_lock); */
2696 FOREACH_PROC_IN_SYSTEM(p) {
2697 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2699 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2700 /* sx_sunlock(&allproc_lock); */
2704 /* sx_sunlock(&allproc_lock); */
2705 if (_vm_object_in_map(kernel_map, object, 0))
2710 DB_SHOW_COMMAND(vmochk, vm_object_check)
2715 * make sure that internal objs are in a map somewhere
2716 * and none have zero ref counts.
2718 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2719 if ((object->flags & OBJ_ANON) != 0) {
2720 if (object->ref_count == 0) {
2721 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2722 (long)object->size);
2724 if (!vm_object_in_map(object)) {
2726 "vmochk: internal obj is not in a map: "
2727 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2728 object->ref_count, (u_long)object->size,
2729 (u_long)object->size,
2730 (void *)object->backing_object);
2739 * vm_object_print: [ debug ]
2741 DB_SHOW_COMMAND(object, vm_object_print_static)
2743 /* XXX convert args. */
2744 vm_object_t object = (vm_object_t)addr;
2745 boolean_t full = have_addr;
2749 /* XXX count is an (unused) arg. Avoid shadowing it. */
2750 #define count was_count
2758 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2759 object, (int)object->type, (uintmax_t)object->size,
2760 object->resident_page_count, object->ref_count, object->flags,
2761 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2762 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2763 atomic_load_int(&object->shadow_count),
2764 object->backing_object ? object->backing_object->ref_count : 0,
2765 object->backing_object, (uintmax_t)object->backing_object_offset);
2772 TAILQ_FOREACH(p, &object->memq, listq) {
2774 db_iprintf("memory:=");
2775 else if (count == 6) {
2783 db_printf("(off=0x%jx,page=0x%jx)",
2784 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2797 /* XXX need this non-static entry for calling from vm_map_print. */
2800 /* db_expr_t */ long addr,
2801 boolean_t have_addr,
2802 /* db_expr_t */ long count,
2805 vm_object_print_static(addr, have_addr, count, modif);
2808 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2813 vm_page_t m, prev_m;
2816 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2817 db_printf("new object: %p\n", (void *)object);
2824 TAILQ_FOREACH(m, &object->memq, listq) {
2825 if (m->pindex > 128)
2827 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2828 prev_m->pindex + 1 != m->pindex) {
2830 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2831 (long)fidx, rcount, (long)pa);
2838 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2843 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2844 (long)fidx, rcount, (long)pa);
2849 pa = VM_PAGE_TO_PHYS(m);
2853 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2854 (long)fidx, rcount, (long)pa);