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>
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/blockcount.h>
73 #include <sys/cpuset.h>
74 #include <sys/limits.h>
77 #include <sys/mount.h>
78 #include <sys/kernel.h>
79 #include <sys/pctrie.h>
80 #include <sys/sysctl.h>
81 #include <sys/mutex.h>
82 #include <sys/proc.h> /* for curproc, pageproc */
83 #include <sys/refcount.h>
84 #include <sys/socket.h>
85 #include <sys/resourcevar.h>
86 #include <sys/refcount.h>
87 #include <sys/rwlock.h>
89 #include <sys/vnode.h>
90 #include <sys/vmmeter.h>
94 #include <vm/vm_param.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_object.h>
98 #include <vm/vm_page.h>
99 #include <vm/vm_pageout.h>
100 #include <vm/vm_pager.h>
101 #include <vm/vm_phys.h>
102 #include <vm/vm_pagequeue.h>
103 #include <vm/swap_pager.h>
104 #include <vm/vm_kern.h>
105 #include <vm/vm_extern.h>
106 #include <vm/vm_radix.h>
107 #include <vm/vm_reserv.h>
110 static int old_msync;
111 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
112 "Use old (insecure) msync behavior");
114 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
115 int pagerflags, int flags, boolean_t *allclean,
117 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
118 boolean_t *allclean);
119 static void vm_object_backing_remove(vm_object_t object);
122 * Virtual memory objects maintain the actual data
123 * associated with allocated virtual memory. A given
124 * page of memory exists within exactly one object.
126 * An object is only deallocated when all "references"
127 * are given up. Only one "reference" to a given
128 * region of an object should be writeable.
130 * Associated with each object is a list of all resident
131 * memory pages belonging to that object; this list is
132 * maintained by the "vm_page" module, and locked by the object's
135 * Each object also records a "pager" routine which is
136 * used to retrieve (and store) pages to the proper backing
137 * storage. In addition, objects may be backed by other
138 * objects from which they were virtual-copied.
140 * The only items within the object structure which are
141 * modified after time of creation are:
142 * reference count locked by object's lock
143 * pager routine locked by object's lock
147 struct object_q vm_object_list;
148 struct mtx vm_object_list_mtx; /* lock for object list and count */
150 struct vm_object kernel_object_store;
152 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
155 static COUNTER_U64_DEFINE_EARLY(object_collapses);
156 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
158 "VM object collapses");
160 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
161 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
163 "VM object bypasses");
165 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
166 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
167 &object_collapse_waits,
168 "Number of sleeps for collapse");
170 static uma_zone_t obj_zone;
172 static int vm_object_zinit(void *mem, int size, int flags);
175 static void vm_object_zdtor(void *mem, int size, void *arg);
178 vm_object_zdtor(void *mem, int size, void *arg)
182 object = (vm_object_t)mem;
183 KASSERT(object->ref_count == 0,
184 ("object %p ref_count = %d", object, object->ref_count));
185 KASSERT(TAILQ_EMPTY(&object->memq),
186 ("object %p has resident pages in its memq", object));
187 KASSERT(vm_radix_is_empty(&object->rtree),
188 ("object %p has resident pages in its trie", object));
189 #if VM_NRESERVLEVEL > 0
190 KASSERT(LIST_EMPTY(&object->rvq),
191 ("object %p has reservations",
194 KASSERT(!vm_object_busied(object),
195 ("object %p busy = %d", object, blockcount_read(&object->busy)));
196 KASSERT(object->resident_page_count == 0,
197 ("object %p resident_page_count = %d",
198 object, object->resident_page_count));
199 KASSERT(atomic_load_int(&object->shadow_count) == 0,
200 ("object %p shadow_count = %d",
201 object, atomic_load_int(&object->shadow_count)));
202 KASSERT(object->type == OBJT_DEAD,
203 ("object %p has non-dead type %d",
204 object, object->type));
205 KASSERT(object->charge == 0 && object->cred == NULL,
206 ("object %p has non-zero charge %ju (%p)",
207 object, (uintmax_t)object->charge, object->cred));
212 vm_object_zinit(void *mem, int size, int flags)
216 object = (vm_object_t)mem;
217 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
219 /* These are true for any object that has been freed */
220 object->type = OBJT_DEAD;
221 vm_radix_init(&object->rtree);
222 refcount_init(&object->ref_count, 0);
223 blockcount_init(&object->paging_in_progress);
224 blockcount_init(&object->busy);
225 object->resident_page_count = 0;
226 atomic_store_int(&object->shadow_count, 0);
227 object->flags = OBJ_DEAD;
229 mtx_lock(&vm_object_list_mtx);
230 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
231 mtx_unlock(&vm_object_list_mtx);
236 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
237 vm_object_t object, void *handle)
240 TAILQ_INIT(&object->memq);
241 LIST_INIT(&object->shadow_head);
244 object->flags = flags;
245 if ((flags & OBJ_SWAP) != 0)
246 pctrie_init(&object->un_pager.swp.swp_blks);
249 * Ensure that swap_pager_swapoff() iteration over object_list
250 * sees up to date type and pctrie head if it observed
253 atomic_thread_fence_rel();
255 object->pg_color = 0;
257 object->domain.dr_policy = NULL;
258 object->generation = 1;
259 object->cleangeneration = 1;
260 refcount_init(&object->ref_count, 1);
261 object->memattr = VM_MEMATTR_DEFAULT;
264 object->handle = handle;
265 object->backing_object = NULL;
266 object->backing_object_offset = (vm_ooffset_t) 0;
267 #if VM_NRESERVLEVEL > 0
268 LIST_INIT(&object->rvq);
270 umtx_shm_object_init(object);
276 * Initialize the VM objects module.
281 TAILQ_INIT(&vm_object_list);
282 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
284 rw_init(&kernel_object->lock, "kernel vm object");
285 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
286 VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
287 #if VM_NRESERVLEVEL > 0
288 kernel_object->flags |= OBJ_COLORED;
289 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
291 kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
294 * The lock portion of struct vm_object must be type stable due
295 * to vm_pageout_fallback_object_lock locking a vm object
296 * without holding any references to it.
298 * paging_in_progress is valid always. Lockless references to
299 * the objects may acquire pip and then check OBJ_DEAD.
301 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
307 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
313 vm_object_clear_flag(vm_object_t object, u_short bits)
316 VM_OBJECT_ASSERT_WLOCKED(object);
317 object->flags &= ~bits;
321 * Sets the default memory attribute for the specified object. Pages
322 * that are allocated to this object are by default assigned this memory
325 * Presently, this function must be called before any pages are allocated
326 * to the object. In the future, this requirement may be relaxed for
327 * "default" and "swap" objects.
330 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
333 VM_OBJECT_ASSERT_WLOCKED(object);
335 if (object->type == OBJT_DEAD)
336 return (KERN_INVALID_ARGUMENT);
337 if (!TAILQ_EMPTY(&object->memq))
338 return (KERN_FAILURE);
340 object->memattr = memattr;
341 return (KERN_SUCCESS);
345 vm_object_pip_add(vm_object_t object, short i)
349 blockcount_acquire(&object->paging_in_progress, i);
353 vm_object_pip_wakeup(vm_object_t object)
356 vm_object_pip_wakeupn(object, 1);
360 vm_object_pip_wakeupn(vm_object_t object, short i)
364 blockcount_release(&object->paging_in_progress, i);
368 * Atomically drop the object lock and wait for pip to drain. This protects
369 * from sleep/wakeup races due to identity changes. The lock is not re-acquired
373 vm_object_pip_sleep(vm_object_t object, const char *waitid)
376 (void)blockcount_sleep(&object->paging_in_progress, &object->lock,
377 waitid, PVM | PDROP);
381 vm_object_pip_wait(vm_object_t object, const char *waitid)
384 VM_OBJECT_ASSERT_WLOCKED(object);
386 blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
391 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
394 VM_OBJECT_ASSERT_UNLOCKED(object);
396 blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
400 * vm_object_allocate:
402 * Returns a new object with the given size.
405 vm_object_allocate(objtype_t type, vm_pindex_t size)
412 panic("vm_object_allocate: can't create OBJT_DEAD");
417 flags = OBJ_COLORED | OBJ_SWAP;
421 flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
424 flags = OBJ_FICTITIOUS;
427 flags = OBJ_UNMANAGED;
433 panic("vm_object_allocate: type %d is undefined or dynamic",
436 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
437 _vm_object_allocate(type, size, flags, object, NULL);
443 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags)
447 MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */);
448 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
449 _vm_object_allocate(dyntype, size, flags, object, NULL);
455 * vm_object_allocate_anon:
457 * Returns a new default object of the given size and marked as
458 * anonymous memory for special split/collapse handling. Color
459 * to be initialized by the caller.
462 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
463 struct ucred *cred, vm_size_t charge)
465 vm_object_t handle, object;
467 if (backing_object == NULL)
469 else if ((backing_object->flags & OBJ_ANON) != 0)
470 handle = backing_object->handle;
472 handle = backing_object;
473 object = uma_zalloc(obj_zone, M_WAITOK);
474 _vm_object_allocate(OBJT_DEFAULT, size, OBJ_ANON | OBJ_ONEMAPPING,
477 object->charge = cred != NULL ? charge : 0;
482 vm_object_reference_vnode(vm_object_t object)
487 * vnode objects need the lock for the first reference
488 * to serialize with vnode_object_deallocate().
490 if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
491 VM_OBJECT_RLOCK(object);
492 old = refcount_acquire(&object->ref_count);
493 if (object->type == OBJT_VNODE && old == 0)
494 vref(object->handle);
495 VM_OBJECT_RUNLOCK(object);
500 * vm_object_reference:
502 * Acquires a reference to the given object.
505 vm_object_reference(vm_object_t object)
511 if (object->type == OBJT_VNODE)
512 vm_object_reference_vnode(object);
514 refcount_acquire(&object->ref_count);
515 KASSERT((object->flags & OBJ_DEAD) == 0,
516 ("vm_object_reference: Referenced dead object."));
520 * vm_object_reference_locked:
522 * Gets another reference to the given object.
524 * The object must be locked.
527 vm_object_reference_locked(vm_object_t object)
531 VM_OBJECT_ASSERT_LOCKED(object);
532 old = refcount_acquire(&object->ref_count);
533 if (object->type == OBJT_VNODE && old == 0)
534 vref(object->handle);
535 KASSERT((object->flags & OBJ_DEAD) == 0,
536 ("vm_object_reference: Referenced dead object."));
540 * Handle deallocating an object of type OBJT_VNODE.
543 vm_object_deallocate_vnode(vm_object_t object)
545 struct vnode *vp = (struct vnode *) object->handle;
548 KASSERT(object->type == OBJT_VNODE,
549 ("vm_object_deallocate_vnode: not a vnode object"));
550 KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
552 /* Object lock to protect handle lookup. */
553 last = refcount_release(&object->ref_count);
554 VM_OBJECT_RUNLOCK(object);
559 if (!umtx_shm_vnobj_persistent)
560 umtx_shm_object_terminated(object);
562 /* vrele may need the vnode lock. */
567 * We dropped a reference on an object and discovered that it had a
568 * single remaining shadow. This is a sibling of the reference we
569 * dropped. Attempt to collapse the sibling and backing object.
572 vm_object_deallocate_anon(vm_object_t backing_object)
576 /* Fetch the final shadow. */
577 object = LIST_FIRST(&backing_object->shadow_head);
578 KASSERT(object != NULL &&
579 atomic_load_int(&backing_object->shadow_count) == 1,
580 ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
581 backing_object->ref_count,
582 atomic_load_int(&backing_object->shadow_count)));
583 KASSERT((object->flags & OBJ_ANON) != 0,
584 ("invalid shadow object %p", object));
586 if (!VM_OBJECT_TRYWLOCK(object)) {
588 * Prevent object from disappearing since we do not have a
591 vm_object_pip_add(object, 1);
592 VM_OBJECT_WUNLOCK(backing_object);
593 VM_OBJECT_WLOCK(object);
594 vm_object_pip_wakeup(object);
596 VM_OBJECT_WUNLOCK(backing_object);
599 * Check for a collapse/terminate race with the last reference holder.
601 if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
602 !refcount_acquire_if_not_zero(&object->ref_count)) {
603 VM_OBJECT_WUNLOCK(object);
606 backing_object = object->backing_object;
607 if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
608 vm_object_collapse(object);
609 VM_OBJECT_WUNLOCK(object);
615 * vm_object_deallocate:
617 * Release a reference to the specified object,
618 * gained either through a vm_object_allocate
619 * or a vm_object_reference call. When all references
620 * are gone, storage associated with this object
621 * may be relinquished.
623 * No object may be locked.
626 vm_object_deallocate(vm_object_t object)
631 while (object != NULL) {
633 * If the reference count goes to 0 we start calling
634 * vm_object_terminate() on the object chain. A ref count
635 * of 1 may be a special case depending on the shadow count
636 * being 0 or 1. These cases require a write lock on the
639 if ((object->flags & OBJ_ANON) == 0)
640 released = refcount_release_if_gt(&object->ref_count, 1);
642 released = refcount_release_if_gt(&object->ref_count, 2);
646 if (object->type == OBJT_VNODE) {
647 VM_OBJECT_RLOCK(object);
648 if (object->type == OBJT_VNODE) {
649 vm_object_deallocate_vnode(object);
652 VM_OBJECT_RUNLOCK(object);
655 VM_OBJECT_WLOCK(object);
656 KASSERT(object->ref_count > 0,
657 ("vm_object_deallocate: object deallocated too many times: %d",
661 * If this is not the final reference to an anonymous
662 * object we may need to collapse the shadow chain.
664 if (!refcount_release(&object->ref_count)) {
665 if (object->ref_count > 1 ||
666 atomic_load_int(&object->shadow_count) == 0) {
667 if ((object->flags & OBJ_ANON) != 0 &&
668 object->ref_count == 1)
669 vm_object_set_flag(object,
671 VM_OBJECT_WUNLOCK(object);
675 /* Handle collapsing last ref on anonymous objects. */
676 object = vm_object_deallocate_anon(object);
681 * Handle the final reference to an object. We restart
682 * the loop with the backing object to avoid recursion.
684 umtx_shm_object_terminated(object);
685 temp = object->backing_object;
687 KASSERT(object->type == OBJT_DEFAULT ||
688 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 removes the object from the global object list
703 * and frees the space for the object.
706 vm_object_destroy(vm_object_t object)
710 * Release the allocation charge.
712 if (object->cred != NULL) {
713 swap_release_by_cred(object->charge, object->cred);
715 crfree(object->cred);
720 * Free the space for the object.
722 uma_zfree(obj_zone, object);
726 vm_object_sub_shadow(vm_object_t object)
728 KASSERT(object->shadow_count >= 1,
729 ("object %p sub_shadow count zero", object));
730 atomic_subtract_int(&object->shadow_count, 1);
734 vm_object_backing_remove_locked(vm_object_t object)
736 vm_object_t backing_object;
738 backing_object = object->backing_object;
739 VM_OBJECT_ASSERT_WLOCKED(object);
740 VM_OBJECT_ASSERT_WLOCKED(backing_object);
742 KASSERT((object->flags & OBJ_COLLAPSING) == 0,
743 ("vm_object_backing_remove: Removing collapsing object."));
745 vm_object_sub_shadow(backing_object);
746 if ((object->flags & OBJ_SHADOWLIST) != 0) {
747 LIST_REMOVE(object, shadow_list);
748 vm_object_clear_flag(object, OBJ_SHADOWLIST);
750 object->backing_object = NULL;
754 vm_object_backing_remove(vm_object_t object)
756 vm_object_t backing_object;
758 VM_OBJECT_ASSERT_WLOCKED(object);
760 backing_object = object->backing_object;
761 if ((object->flags & OBJ_SHADOWLIST) != 0) {
762 VM_OBJECT_WLOCK(backing_object);
763 vm_object_backing_remove_locked(object);
764 VM_OBJECT_WUNLOCK(backing_object);
766 object->backing_object = NULL;
767 vm_object_sub_shadow(backing_object);
772 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
775 VM_OBJECT_ASSERT_WLOCKED(object);
777 atomic_add_int(&backing_object->shadow_count, 1);
778 if ((backing_object->flags & OBJ_ANON) != 0) {
779 VM_OBJECT_ASSERT_WLOCKED(backing_object);
780 LIST_INSERT_HEAD(&backing_object->shadow_head, object,
782 vm_object_set_flag(object, OBJ_SHADOWLIST);
784 object->backing_object = backing_object;
788 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
791 VM_OBJECT_ASSERT_WLOCKED(object);
793 if ((backing_object->flags & OBJ_ANON) != 0) {
794 VM_OBJECT_WLOCK(backing_object);
795 vm_object_backing_insert_locked(object, backing_object);
796 VM_OBJECT_WUNLOCK(backing_object);
798 object->backing_object = backing_object;
799 atomic_add_int(&backing_object->shadow_count, 1);
804 * Insert an object into a backing_object's shadow list with an additional
805 * reference to the backing_object added.
808 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
811 VM_OBJECT_ASSERT_WLOCKED(object);
813 if ((backing_object->flags & OBJ_ANON) != 0) {
814 VM_OBJECT_WLOCK(backing_object);
815 KASSERT((backing_object->flags & OBJ_DEAD) == 0,
816 ("shadowing dead anonymous object"));
817 vm_object_reference_locked(backing_object);
818 vm_object_backing_insert_locked(object, backing_object);
819 vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
820 VM_OBJECT_WUNLOCK(backing_object);
822 vm_object_reference(backing_object);
823 atomic_add_int(&backing_object->shadow_count, 1);
824 object->backing_object = backing_object;
829 * Transfer a backing reference from backing_object to object.
832 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
834 vm_object_t new_backing_object;
837 * Note that the reference to backing_object->backing_object
838 * moves from within backing_object to within object.
840 vm_object_backing_remove_locked(object);
841 new_backing_object = backing_object->backing_object;
842 if (new_backing_object == NULL)
844 if ((new_backing_object->flags & OBJ_ANON) != 0) {
845 VM_OBJECT_WLOCK(new_backing_object);
846 vm_object_backing_remove_locked(backing_object);
847 vm_object_backing_insert_locked(object, new_backing_object);
848 VM_OBJECT_WUNLOCK(new_backing_object);
851 * shadow_count for new_backing_object is left
852 * unchanged, its reference provided by backing_object
853 * is replaced by object.
855 object->backing_object = new_backing_object;
856 backing_object->backing_object = NULL;
861 * Wait for a concurrent collapse to settle.
864 vm_object_collapse_wait(vm_object_t object)
867 VM_OBJECT_ASSERT_WLOCKED(object);
869 while ((object->flags & OBJ_COLLAPSING) != 0) {
870 vm_object_pip_wait(object, "vmcolwait");
871 counter_u64_add(object_collapse_waits, 1);
876 * Waits for a backing object to clear a pending collapse and returns
877 * it locked if it is an ANON object.
880 vm_object_backing_collapse_wait(vm_object_t object)
882 vm_object_t backing_object;
884 VM_OBJECT_ASSERT_WLOCKED(object);
887 backing_object = object->backing_object;
888 if (backing_object == NULL ||
889 (backing_object->flags & OBJ_ANON) == 0)
891 VM_OBJECT_WLOCK(backing_object);
892 if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
894 VM_OBJECT_WUNLOCK(object);
895 vm_object_pip_sleep(backing_object, "vmbckwait");
896 counter_u64_add(object_collapse_waits, 1);
897 VM_OBJECT_WLOCK(object);
899 return (backing_object);
903 * vm_object_terminate_pages removes any remaining pageable pages
904 * from the object and resets the object to an empty state.
907 vm_object_terminate_pages(vm_object_t object)
911 VM_OBJECT_ASSERT_WLOCKED(object);
914 * Free any remaining pageable pages. This also removes them from the
915 * paging queues. However, don't free wired pages, just remove them
916 * from the object. Rather than incrementally removing each page from
917 * the object, the page and object are reset to any empty state.
919 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
920 vm_page_assert_unbusied(p);
921 KASSERT(p->object == object &&
922 (p->ref_count & VPRC_OBJREF) != 0,
923 ("vm_object_terminate_pages: page %p is inconsistent", p));
926 if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
933 * If the object contained any pages, then reset it to an empty state.
934 * None of the object's fields, including "resident_page_count", were
935 * modified by the preceding loop.
937 if (object->resident_page_count != 0) {
938 vm_radix_reclaim_allnodes(&object->rtree);
939 TAILQ_INIT(&object->memq);
940 object->resident_page_count = 0;
941 if (object->type == OBJT_VNODE)
942 vdrop(object->handle);
947 * vm_object_terminate actually destroys the specified object, freeing
948 * up all previously used resources.
950 * The object must be locked.
951 * This routine may block.
954 vm_object_terminate(vm_object_t object)
957 VM_OBJECT_ASSERT_WLOCKED(object);
958 KASSERT((object->flags & OBJ_DEAD) != 0,
959 ("terminating non-dead obj %p", object));
960 KASSERT((object->flags & OBJ_COLLAPSING) == 0,
961 ("terminating collapsing obj %p", object));
962 KASSERT(object->backing_object == NULL,
963 ("terminating shadow obj %p", object));
966 * Wait for the pageout daemon and other current users to be
967 * done with the object. Note that new paging_in_progress
968 * users can come after this wait, but they must check
969 * OBJ_DEAD flag set (without unlocking the object), and avoid
970 * the object being terminated.
972 vm_object_pip_wait(object, "objtrm");
974 KASSERT(object->ref_count == 0,
975 ("vm_object_terminate: object with references, ref_count=%d",
978 if ((object->flags & OBJ_PG_DTOR) == 0)
979 vm_object_terminate_pages(object);
981 #if VM_NRESERVLEVEL > 0
982 if (__predict_false(!LIST_EMPTY(&object->rvq)))
983 vm_reserv_break_all(object);
986 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
987 (object->flags & OBJ_SWAP) != 0,
988 ("%s: non-swap obj %p has cred", __func__, object));
991 * Let the pager know object is dead.
993 vm_pager_deallocate(object);
994 VM_OBJECT_WUNLOCK(object);
996 vm_object_destroy(object);
1000 * Make the page read-only so that we can clear the object flags. However, if
1001 * this is a nosync mmap then the object is likely to stay dirty so do not
1002 * mess with the page and do not clear the object flags. Returns TRUE if the
1003 * page should be flushed, and FALSE otherwise.
1006 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
1009 vm_page_assert_busied(p);
1012 * If we have been asked to skip nosync pages and this is a
1013 * nosync page, skip it. Note that the object flags were not
1014 * cleared in this case so we do not have to set them.
1016 if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
1020 pmap_remove_write(p);
1021 return (p->dirty != 0);
1026 * vm_object_page_clean
1028 * Clean all dirty pages in the specified range of object. Leaves page
1029 * on whatever queue it is currently on. If NOSYNC is set then do not
1030 * write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1031 * leaving the object dirty.
1033 * For swap objects backing tmpfs regular files, do not flush anything,
1034 * but remove write protection on the mapped pages to update mtime through
1037 * When stuffing pages asynchronously, allow clustering. XXX we need a
1038 * synchronous clustering mode implementation.
1040 * Odd semantics: if start == end, we clean everything.
1042 * The object must be locked.
1044 * Returns FALSE if some page from the range was not written, as
1045 * reported by the pager, and TRUE otherwise.
1048 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1052 vm_pindex_t pi, tend, tstart;
1053 int curgeneration, n, pagerflags;
1054 boolean_t eio, res, allclean;
1056 VM_OBJECT_ASSERT_WLOCKED(object);
1058 if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1061 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1062 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1063 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1065 tstart = OFF_TO_IDX(start);
1066 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1067 allclean = tstart == 0 && tend >= object->size;
1071 curgeneration = object->generation;
1073 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1077 np = TAILQ_NEXT(p, listq);
1078 if (vm_page_none_valid(p))
1080 if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1081 if (object->generation != curgeneration &&
1082 (flags & OBJPC_SYNC) != 0)
1084 np = vm_page_find_least(object, pi);
1087 if (!vm_object_page_remove_write(p, flags, &allclean)) {
1091 if (object->type == OBJT_VNODE) {
1092 n = vm_object_page_collect_flush(object, p, pagerflags,
1093 flags, &allclean, &eio);
1098 if (object->generation != curgeneration &&
1099 (flags & OBJPC_SYNC) != 0)
1103 * If the VOP_PUTPAGES() did a truncated write, so
1104 * that even the first page of the run is not fully
1105 * written, vm_pageout_flush() returns 0 as the run
1106 * length. Since the condition that caused truncated
1107 * write may be permanent, e.g. exhausted free space,
1108 * accepting n == 0 would cause an infinite loop.
1110 * Forwarding the iterator leaves the unwritten page
1111 * behind, but there is not much we can do there if
1112 * filesystem refuses to write it.
1122 np = vm_page_find_least(object, pi + n);
1125 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1129 * Leave updating cleangeneration for tmpfs objects to tmpfs
1130 * scan. It needs to update mtime, which happens for other
1131 * filesystems during page writeouts.
1133 if (allclean && object->type == OBJT_VNODE)
1134 object->cleangeneration = curgeneration;
1139 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1140 int flags, boolean_t *allclean, boolean_t *eio)
1142 vm_page_t ma[vm_pageout_page_count], p_first, tp;
1143 int count, i, mreq, runlen;
1145 vm_page_lock_assert(p, MA_NOTOWNED);
1146 vm_page_assert_xbusied(p);
1147 VM_OBJECT_ASSERT_WLOCKED(object);
1152 for (tp = p; count < vm_pageout_page_count; count++) {
1153 tp = vm_page_next(tp);
1154 if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1156 if (!vm_object_page_remove_write(tp, flags, allclean)) {
1157 vm_page_xunbusy(tp);
1162 for (p_first = p; count < vm_pageout_page_count; count++) {
1163 tp = vm_page_prev(p_first);
1164 if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1166 if (!vm_object_page_remove_write(tp, flags, allclean)) {
1167 vm_page_xunbusy(tp);
1174 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1177 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1182 * Note that there is absolutely no sense in writing out
1183 * anonymous objects, so we track down the vnode object
1185 * We invalidate (remove) all pages from the address space
1186 * for semantic correctness.
1188 * If the backing object is a device object with unmanaged pages, then any
1189 * mappings to the specified range of pages must be removed before this
1190 * function is called.
1192 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1193 * may start out with a NULL object.
1196 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1197 boolean_t syncio, boolean_t invalidate)
1199 vm_object_t backing_object;
1202 int error, flags, fsync_after;
1209 VM_OBJECT_WLOCK(object);
1210 while ((backing_object = object->backing_object) != NULL) {
1211 VM_OBJECT_WLOCK(backing_object);
1212 offset += object->backing_object_offset;
1213 VM_OBJECT_WUNLOCK(object);
1214 object = backing_object;
1215 if (object->size < OFF_TO_IDX(offset + size))
1216 size = IDX_TO_OFF(object->size) - offset;
1219 * Flush pages if writing is allowed, invalidate them
1220 * if invalidation requested. Pages undergoing I/O
1221 * will be ignored by vm_object_page_remove().
1223 * We cannot lock the vnode and then wait for paging
1224 * to complete without deadlocking against vm_fault.
1225 * Instead we simply call vm_object_page_remove() and
1226 * allow it to block internally on a page-by-page
1227 * basis when it encounters pages undergoing async
1230 if (object->type == OBJT_VNODE &&
1231 vm_object_mightbedirty(object) != 0 &&
1232 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1233 VM_OBJECT_WUNLOCK(object);
1234 (void)vn_start_write(vp, &mp, V_WAIT);
1235 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1236 if (syncio && !invalidate && offset == 0 &&
1237 atop(size) == object->size) {
1239 * If syncing the whole mapping of the file,
1240 * it is faster to schedule all the writes in
1241 * async mode, also allowing the clustering,
1242 * and then wait for i/o to complete.
1247 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1248 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1249 fsync_after = FALSE;
1251 VM_OBJECT_WLOCK(object);
1252 res = vm_object_page_clean(object, offset, offset + size,
1254 VM_OBJECT_WUNLOCK(object);
1257 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1258 if (error != ERELOOKUP)
1262 * Allow SU/bufdaemon to handle more
1263 * dependencies in the meantime.
1266 vn_finished_write(mp);
1268 (void)vn_start_write(vp, &mp, V_WAIT);
1269 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1273 vn_finished_write(mp);
1276 VM_OBJECT_WLOCK(object);
1278 if ((object->type == OBJT_VNODE ||
1279 object->type == OBJT_DEVICE) && invalidate) {
1280 if (object->type == OBJT_DEVICE)
1282 * The option OBJPR_NOTMAPPED must be passed here
1283 * because vm_object_page_remove() cannot remove
1284 * unmanaged mappings.
1286 flags = OBJPR_NOTMAPPED;
1290 flags = OBJPR_CLEANONLY;
1291 vm_object_page_remove(object, OFF_TO_IDX(offset),
1292 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1294 VM_OBJECT_WUNLOCK(object);
1299 * Determine whether the given advice can be applied to the object. Advice is
1300 * not applied to unmanaged pages since they never belong to page queues, and
1301 * since MADV_FREE is destructive, it can apply only to anonymous pages that
1302 * have been mapped at most once.
1305 vm_object_advice_applies(vm_object_t object, int advice)
1308 if ((object->flags & OBJ_UNMANAGED) != 0)
1310 if (advice != MADV_FREE)
1312 return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1313 (OBJ_ONEMAPPING | OBJ_ANON));
1317 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1321 if (advice == MADV_FREE)
1322 vm_pager_freespace(object, pindex, size);
1326 * vm_object_madvise:
1328 * Implements the madvise function at the object/page level.
1330 * MADV_WILLNEED (any object)
1332 * Activate the specified pages if they are resident.
1334 * MADV_DONTNEED (any object)
1336 * Deactivate the specified pages if they are resident.
1338 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1339 * OBJ_ONEMAPPING only)
1341 * Deactivate and clean the specified pages if they are
1342 * resident. This permits the process to reuse the pages
1343 * without faulting or the kernel to reclaim the pages
1347 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1350 vm_pindex_t tpindex;
1351 vm_object_t backing_object, tobject;
1358 VM_OBJECT_WLOCK(object);
1359 if (!vm_object_advice_applies(object, advice)) {
1360 VM_OBJECT_WUNLOCK(object);
1363 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1367 * If the next page isn't resident in the top-level object, we
1368 * need to search the shadow chain. When applying MADV_FREE, we
1369 * take care to release any swap space used to store
1370 * non-resident pages.
1372 if (m == NULL || pindex < m->pindex) {
1374 * Optimize a common case: if the top-level object has
1375 * no backing object, we can skip over the non-resident
1376 * range in constant time.
1378 if (object->backing_object == NULL) {
1379 tpindex = (m != NULL && m->pindex < end) ?
1381 vm_object_madvise_freespace(object, advice,
1382 pindex, tpindex - pindex);
1383 if ((pindex = tpindex) == end)
1390 vm_object_madvise_freespace(tobject, advice,
1393 * Prepare to search the next object in the
1396 backing_object = tobject->backing_object;
1397 if (backing_object == NULL)
1399 VM_OBJECT_WLOCK(backing_object);
1401 OFF_TO_IDX(tobject->backing_object_offset);
1402 if (tobject != object)
1403 VM_OBJECT_WUNLOCK(tobject);
1404 tobject = backing_object;
1405 if (!vm_object_advice_applies(tobject, advice))
1407 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
1412 m = TAILQ_NEXT(m, listq);
1416 * If the page is not in a normal state, skip it. The page
1417 * can not be invalidated while the object lock is held.
1419 if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1421 KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1422 ("vm_object_madvise: page %p is fictitious", tm));
1423 KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1424 ("vm_object_madvise: page %p is not managed", tm));
1425 if (vm_page_tryxbusy(tm) == 0) {
1426 if (object != tobject)
1427 VM_OBJECT_WUNLOCK(object);
1428 if (advice == MADV_WILLNEED) {
1430 * Reference the page before unlocking and
1431 * sleeping so that the page daemon is less
1432 * likely to reclaim it.
1434 vm_page_aflag_set(tm, PGA_REFERENCED);
1436 if (!vm_page_busy_sleep(tm, "madvpo", 0))
1437 VM_OBJECT_WUNLOCK(tobject);
1440 vm_page_advise(tm, advice);
1441 vm_page_xunbusy(tm);
1442 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1444 if (tobject != object)
1445 VM_OBJECT_WUNLOCK(tobject);
1447 VM_OBJECT_WUNLOCK(object);
1453 * Create a new object which is backed by the
1454 * specified existing object range. The source
1455 * object reference is deallocated.
1457 * The new object and offset into that object
1458 * are returned in the source parameters.
1461 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1462 struct ucred *cred, bool shared)
1470 * Don't create the new object if the old object isn't shared.
1472 * If we hold the only reference we can guarantee that it won't
1473 * increase while we have the map locked. Otherwise the race is
1474 * harmless and we will end up with an extra shadow object that
1475 * will be collapsed later.
1477 if (source != NULL && source->ref_count == 1 &&
1478 (source->flags & OBJ_ANON) != 0)
1482 * Allocate a new object with the given length.
1484 result = vm_object_allocate_anon(atop(length), source, cred, length);
1487 * Store the offset into the source object, and fix up the offset into
1490 result->backing_object_offset = *offset;
1492 if (shared || source != NULL) {
1493 VM_OBJECT_WLOCK(result);
1496 * The new object shadows the source object, adding a
1497 * reference to it. Our caller changes his reference
1498 * to point to the new object, removing a reference to
1499 * the source object. Net result: no change of
1500 * reference count, unless the caller needs to add one
1501 * more reference due to forking a shared map entry.
1504 vm_object_reference_locked(result);
1505 vm_object_clear_flag(result, OBJ_ONEMAPPING);
1509 * Try to optimize the result object's page color when
1510 * shadowing in order to maintain page coloring
1511 * consistency in the combined shadowed object.
1513 if (source != NULL) {
1514 vm_object_backing_insert(result, source);
1515 result->domain = source->domain;
1516 #if VM_NRESERVLEVEL > 0
1517 vm_object_set_flag(result,
1518 (source->flags & OBJ_COLORED));
1519 result->pg_color = (source->pg_color +
1520 OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1524 VM_OBJECT_WUNLOCK(result);
1528 * Return the new things
1537 * Split the pages in a map entry into a new object. This affords
1538 * easier removal of unused pages, and keeps object inheritance from
1539 * being a negative impact on memory usage.
1542 vm_object_split(vm_map_entry_t entry)
1544 vm_page_t m, m_busy, m_next;
1545 vm_object_t orig_object, new_object, backing_object;
1546 vm_pindex_t idx, offidxstart;
1549 orig_object = entry->object.vm_object;
1550 KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1551 ("vm_object_split: Splitting object with multiple mappings."));
1552 if ((orig_object->flags & OBJ_ANON) == 0)
1554 if (orig_object->ref_count <= 1)
1556 VM_OBJECT_WUNLOCK(orig_object);
1558 offidxstart = OFF_TO_IDX(entry->offset);
1559 size = atop(entry->end - entry->start);
1562 * If swap_pager_copy() is later called, it will convert new_object
1563 * into a swap object.
1565 new_object = vm_object_allocate_anon(size, orig_object,
1566 orig_object->cred, ptoa(size));
1569 * We must wait for the orig_object to complete any in-progress
1570 * collapse so that the swap blocks are stable below. The
1571 * additional reference on backing_object by new object will
1572 * prevent further collapse operations until split completes.
1574 VM_OBJECT_WLOCK(orig_object);
1575 vm_object_collapse_wait(orig_object);
1578 * At this point, the new object is still private, so the order in
1579 * which the original and new objects are locked does not matter.
1581 VM_OBJECT_WLOCK(new_object);
1582 new_object->domain = orig_object->domain;
1583 backing_object = orig_object->backing_object;
1584 if (backing_object != NULL) {
1585 vm_object_backing_insert_ref(new_object, backing_object);
1586 new_object->backing_object_offset =
1587 orig_object->backing_object_offset + entry->offset;
1589 if (orig_object->cred != NULL) {
1590 crhold(orig_object->cred);
1591 KASSERT(orig_object->charge >= ptoa(size),
1592 ("orig_object->charge < 0"));
1593 orig_object->charge -= ptoa(size);
1597 * Mark the split operation so that swap_pager_getpages() knows
1598 * that the object is in transition.
1600 vm_object_set_flag(orig_object, OBJ_SPLIT);
1606 m = vm_page_find_least(orig_object, offidxstart);
1607 KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1608 ("%s: object %p was repopulated", __func__, orig_object));
1609 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1611 m_next = TAILQ_NEXT(m, listq);
1614 * We must wait for pending I/O to complete before we can
1617 * We do not have to VM_PROT_NONE the page as mappings should
1618 * not be changed by this operation.
1620 if (vm_page_tryxbusy(m) == 0) {
1621 VM_OBJECT_WUNLOCK(new_object);
1622 if (vm_page_busy_sleep(m, "spltwt", 0))
1623 VM_OBJECT_WLOCK(orig_object);
1624 VM_OBJECT_WLOCK(new_object);
1629 * The page was left invalid. Likely placed there by
1630 * an incomplete fault. Just remove and ignore.
1632 if (vm_page_none_valid(m)) {
1633 if (vm_page_remove(m))
1638 /* vm_page_rename() will dirty the page. */
1639 if (vm_page_rename(m, new_object, idx)) {
1641 VM_OBJECT_WUNLOCK(new_object);
1642 VM_OBJECT_WUNLOCK(orig_object);
1644 VM_OBJECT_WLOCK(orig_object);
1645 VM_OBJECT_WLOCK(new_object);
1649 #if VM_NRESERVLEVEL > 0
1651 * If some of the reservation's allocated pages remain with
1652 * the original object, then transferring the reservation to
1653 * the new object is neither particularly beneficial nor
1654 * particularly harmful as compared to leaving the reservation
1655 * with the original object. If, however, all of the
1656 * reservation's allocated pages are transferred to the new
1657 * object, then transferring the reservation is typically
1658 * beneficial. Determining which of these two cases applies
1659 * would be more costly than unconditionally renaming the
1662 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1666 * orig_object's type may change while sleeping, so keep track
1667 * of the beginning of the busied range.
1669 if (orig_object->type != OBJT_SWAP)
1671 else if (m_busy == NULL)
1674 if ((orig_object->flags & OBJ_SWAP) != 0) {
1676 * swap_pager_copy() can sleep, in which case the orig_object's
1677 * and new_object's locks are released and reacquired.
1679 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1681 TAILQ_FOREACH_FROM(m_busy, &new_object->memq, listq)
1682 vm_page_xunbusy(m_busy);
1684 vm_object_clear_flag(orig_object, OBJ_SPLIT);
1685 VM_OBJECT_WUNLOCK(orig_object);
1686 VM_OBJECT_WUNLOCK(new_object);
1687 entry->object.vm_object = new_object;
1688 entry->offset = 0LL;
1689 vm_object_deallocate(orig_object);
1690 VM_OBJECT_WLOCK(new_object);
1694 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1696 vm_object_t backing_object;
1698 VM_OBJECT_ASSERT_WLOCKED(object);
1699 backing_object = object->backing_object;
1700 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1702 KASSERT(p == NULL || p->object == object || p->object == backing_object,
1703 ("invalid ownership %p %p %p", p, object, backing_object));
1704 /* The page is only NULL when rename fails. */
1706 VM_OBJECT_WUNLOCK(object);
1707 VM_OBJECT_WUNLOCK(backing_object);
1709 VM_OBJECT_WLOCK(object);
1710 } else if (p->object == object) {
1711 VM_OBJECT_WUNLOCK(backing_object);
1712 if (vm_page_busy_sleep(p, "vmocol", 0))
1713 VM_OBJECT_WLOCK(object);
1715 VM_OBJECT_WUNLOCK(object);
1716 if (!vm_page_busy_sleep(p, "vmocol", 0))
1717 VM_OBJECT_WUNLOCK(backing_object);
1718 VM_OBJECT_WLOCK(object);
1720 VM_OBJECT_WLOCK(backing_object);
1721 return (TAILQ_FIRST(&backing_object->memq));
1725 vm_object_scan_all_shadowed(vm_object_t object)
1727 vm_object_t backing_object;
1729 vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1731 VM_OBJECT_ASSERT_WLOCKED(object);
1732 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1734 backing_object = object->backing_object;
1736 if ((backing_object->flags & OBJ_ANON) == 0)
1739 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1740 p = vm_page_find_least(backing_object, pi);
1741 ps = swap_pager_find_least(backing_object, pi);
1744 * Only check pages inside the parent object's range and
1745 * inside the parent object's mapping of the backing object.
1748 if (p != NULL && p->pindex < pi)
1749 p = TAILQ_NEXT(p, listq);
1751 ps = swap_pager_find_least(backing_object, pi);
1752 if (p == NULL && ps >= backing_object->size)
1757 pi = MIN(p->pindex, ps);
1759 new_pindex = pi - backing_offset_index;
1760 if (new_pindex >= object->size)
1765 * If the backing object page is busy a
1766 * grandparent or older page may still be
1767 * undergoing CoW. It is not safe to collapse
1768 * the backing object until it is quiesced.
1770 if (vm_page_tryxbusy(p) == 0)
1774 * We raced with the fault handler that left
1775 * newly allocated invalid page on the object
1776 * queue and retried.
1778 if (!vm_page_all_valid(p))
1783 * See if the parent has the page or if the parent's object
1784 * pager has the page. If the parent has the page but the page
1785 * is not valid, the parent's object pager must have the page.
1787 * If this fails, the parent does not completely shadow the
1788 * object and we might as well give up now.
1790 pp = vm_page_lookup(object, new_pindex);
1793 * The valid check here is stable due to object lock
1794 * being required to clear valid and initiate paging.
1795 * Busy of p disallows fault handler to validate pp.
1797 if ((pp == NULL || vm_page_none_valid(pp)) &&
1798 !vm_pager_has_page(object, new_pindex, NULL, NULL))
1812 vm_object_collapse_scan(vm_object_t object)
1814 vm_object_t backing_object;
1815 vm_page_t next, p, pp;
1816 vm_pindex_t backing_offset_index, new_pindex;
1818 VM_OBJECT_ASSERT_WLOCKED(object);
1819 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1821 backing_object = object->backing_object;
1822 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1827 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1828 next = TAILQ_NEXT(p, listq);
1829 new_pindex = p->pindex - backing_offset_index;
1832 * Check for busy page
1834 if (vm_page_tryxbusy(p) == 0) {
1835 next = vm_object_collapse_scan_wait(object, p);
1839 KASSERT(object->backing_object == backing_object,
1840 ("vm_object_collapse_scan: backing object mismatch %p != %p",
1841 object->backing_object, backing_object));
1842 KASSERT(p->object == backing_object,
1843 ("vm_object_collapse_scan: object mismatch %p != %p",
1844 p->object, backing_object));
1846 if (p->pindex < backing_offset_index ||
1847 new_pindex >= object->size) {
1848 vm_pager_freespace(backing_object, p->pindex, 1);
1850 KASSERT(!pmap_page_is_mapped(p),
1851 ("freeing mapped page %p", p));
1852 if (vm_page_remove(p))
1857 if (!vm_page_all_valid(p)) {
1858 KASSERT(!pmap_page_is_mapped(p),
1859 ("freeing mapped page %p", p));
1860 if (vm_page_remove(p))
1865 pp = vm_page_lookup(object, new_pindex);
1866 if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1869 * The page in the parent is busy and possibly not
1870 * (yet) valid. Until its state is finalized by the
1871 * busy bit owner, we can't tell whether it shadows the
1874 next = vm_object_collapse_scan_wait(object, pp);
1878 if (pp != NULL && vm_page_none_valid(pp)) {
1880 * The page was invalid in the parent. Likely placed
1881 * there by an incomplete fault. Just remove and
1882 * ignore. p can replace it.
1884 if (vm_page_remove(pp))
1889 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1892 * The page already exists in the parent OR swap exists
1893 * for this location in the parent. Leave the parent's
1894 * page alone. Destroy the original page from the
1897 vm_pager_freespace(backing_object, p->pindex, 1);
1898 KASSERT(!pmap_page_is_mapped(p),
1899 ("freeing mapped page %p", p));
1900 if (vm_page_remove(p))
1903 vm_page_xunbusy(pp);
1908 * Page does not exist in parent, rename the page from the
1909 * backing object to the main object.
1911 * If the page was mapped to a process, it can remain mapped
1912 * through the rename. vm_page_rename() will dirty the page.
1914 if (vm_page_rename(p, object, new_pindex)) {
1916 next = vm_object_collapse_scan_wait(object, NULL);
1920 /* Use the old pindex to free the right page. */
1921 vm_pager_freespace(backing_object, new_pindex +
1922 backing_offset_index, 1);
1924 #if VM_NRESERVLEVEL > 0
1926 * Rename the reservation.
1928 vm_reserv_rename(p, object, backing_object,
1929 backing_offset_index);
1937 * vm_object_collapse:
1939 * Collapse an object with the object backing it.
1940 * Pages in the backing object are moved into the
1941 * parent, and the backing object is deallocated.
1944 vm_object_collapse(vm_object_t object)
1946 vm_object_t backing_object, new_backing_object;
1948 VM_OBJECT_ASSERT_WLOCKED(object);
1951 KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1952 ("collapsing invalid object"));
1955 * Wait for the backing_object to finish any pending
1956 * collapse so that the caller sees the shortest possible
1959 backing_object = vm_object_backing_collapse_wait(object);
1960 if (backing_object == NULL)
1963 KASSERT(object->ref_count > 0 &&
1964 object->ref_count > atomic_load_int(&object->shadow_count),
1965 ("collapse with invalid ref %d or shadow %d count.",
1966 object->ref_count, atomic_load_int(&object->shadow_count)));
1967 KASSERT((backing_object->flags &
1968 (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1969 ("vm_object_collapse: Backing object already collapsing."));
1970 KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1971 ("vm_object_collapse: object is already collapsing."));
1974 * We know that we can either collapse the backing object if
1975 * the parent is the only reference to it, or (perhaps) have
1976 * the parent bypass the object if the parent happens to shadow
1977 * all the resident pages in the entire backing object.
1979 if (backing_object->ref_count == 1) {
1980 KASSERT(atomic_load_int(&backing_object->shadow_count)
1982 ("vm_object_collapse: shadow_count: %d",
1983 atomic_load_int(&backing_object->shadow_count)));
1984 vm_object_pip_add(object, 1);
1985 vm_object_set_flag(object, OBJ_COLLAPSING);
1986 vm_object_pip_add(backing_object, 1);
1987 vm_object_set_flag(backing_object, OBJ_DEAD);
1990 * If there is exactly one reference to the backing
1991 * object, we can collapse it into the parent.
1993 vm_object_collapse_scan(object);
1995 #if VM_NRESERVLEVEL > 0
1997 * Break any reservations from backing_object.
1999 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
2000 vm_reserv_break_all(backing_object);
2004 * Move the pager from backing_object to object.
2006 if ((backing_object->flags & OBJ_SWAP) != 0) {
2008 * swap_pager_copy() can sleep, in which case
2009 * the backing_object's and object's locks are
2010 * released and reacquired.
2011 * Since swap_pager_copy() is being asked to
2012 * destroy backing_object, it will change the
2013 * type to OBJT_DEFAULT.
2018 OFF_TO_IDX(object->backing_object_offset), TRUE);
2022 * Object now shadows whatever backing_object did.
2024 vm_object_clear_flag(object, OBJ_COLLAPSING);
2025 vm_object_backing_transfer(object, backing_object);
2026 object->backing_object_offset +=
2027 backing_object->backing_object_offset;
2028 VM_OBJECT_WUNLOCK(object);
2029 vm_object_pip_wakeup(object);
2032 * Discard backing_object.
2034 * Since the backing object has no pages, no pager left,
2035 * and no object references within it, all that is
2036 * necessary is to dispose of it.
2038 KASSERT(backing_object->ref_count == 1, (
2039 "backing_object %p was somehow re-referenced during collapse!",
2041 vm_object_pip_wakeup(backing_object);
2042 (void)refcount_release(&backing_object->ref_count);
2043 vm_object_terminate(backing_object);
2044 counter_u64_add(object_collapses, 1);
2045 VM_OBJECT_WLOCK(object);
2048 * If we do not entirely shadow the backing object,
2049 * there is nothing we can do so we give up.
2051 * The object lock and backing_object lock must not
2052 * be dropped during this sequence.
2054 if (!vm_object_scan_all_shadowed(object)) {
2055 VM_OBJECT_WUNLOCK(backing_object);
2060 * Make the parent shadow the next object in the
2061 * chain. Deallocating backing_object will not remove
2062 * it, since its reference count is at least 2.
2064 vm_object_backing_remove_locked(object);
2065 new_backing_object = backing_object->backing_object;
2066 if (new_backing_object != NULL) {
2067 vm_object_backing_insert_ref(object,
2068 new_backing_object);
2069 object->backing_object_offset +=
2070 backing_object->backing_object_offset;
2074 * Drop the reference count on backing_object. Since
2075 * its ref_count was at least 2, it will not vanish.
2077 (void)refcount_release(&backing_object->ref_count);
2078 KASSERT(backing_object->ref_count >= 1, (
2079 "backing_object %p was somehow dereferenced during collapse!",
2081 VM_OBJECT_WUNLOCK(backing_object);
2082 counter_u64_add(object_bypasses, 1);
2086 * Try again with this object's new backing object.
2092 * vm_object_page_remove:
2094 * For the given object, either frees or invalidates each of the
2095 * specified pages. In general, a page is freed. However, if a page is
2096 * wired for any reason other than the existence of a managed, wired
2097 * mapping, then it may be invalidated but not removed from the object.
2098 * Pages are specified by the given range ["start", "end") and the option
2099 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
2100 * extends from "start" to the end of the object. If the option
2101 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2102 * specified range are affected. If the option OBJPR_NOTMAPPED is
2103 * specified, then the pages within the specified range must have no
2104 * mappings. Otherwise, if this option is not specified, any mappings to
2105 * the specified pages are removed before the pages are freed or
2108 * In general, this operation should only be performed on objects that
2109 * contain managed pages. There are, however, two exceptions. First, it
2110 * is performed on the kernel and kmem objects by vm_map_entry_delete().
2111 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2112 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
2113 * not be specified and the option OBJPR_NOTMAPPED must be specified.
2115 * The object must be locked.
2118 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2123 VM_OBJECT_ASSERT_WLOCKED(object);
2124 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2125 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2126 ("vm_object_page_remove: illegal options for object %p", object));
2127 if (object->resident_page_count == 0)
2129 vm_object_pip_add(object, 1);
2131 p = vm_page_find_least(object, start);
2134 * Here, the variable "p" is either (1) the page with the least pindex
2135 * greater than or equal to the parameter "start" or (2) NULL.
2137 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2138 next = TAILQ_NEXT(p, listq);
2141 * Skip invalid pages if asked to do so. Try to avoid acquiring
2142 * the busy lock, as some consumers rely on this to avoid
2145 * A thread may concurrently transition the page from invalid to
2146 * valid using only the busy lock, so the result of this check
2147 * is immediately stale. It is up to consumers to handle this,
2148 * for instance by ensuring that all invalid->valid transitions
2149 * happen with a mutex held, as may be possible for a
2152 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2156 * If the page is wired for any reason besides the existence
2157 * of managed, wired mappings, then it cannot be freed. For
2158 * example, fictitious pages, which represent device memory,
2159 * are inherently wired and cannot be freed. They can,
2160 * however, be invalidated if the option OBJPR_CLEANONLY is
2163 if (vm_page_tryxbusy(p) == 0) {
2164 if (vm_page_busy_sleep(p, "vmopar", 0))
2165 VM_OBJECT_WLOCK(object);
2168 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2172 if (vm_page_wired(p)) {
2174 if ((options & OBJPR_NOTMAPPED) == 0 &&
2175 object->ref_count != 0)
2177 if ((options & OBJPR_CLEANONLY) == 0) {
2184 KASSERT((p->flags & PG_FICTITIOUS) == 0,
2185 ("vm_object_page_remove: page %p is fictitious", p));
2186 if ((options & OBJPR_CLEANONLY) != 0 &&
2187 !vm_page_none_valid(p)) {
2188 if ((options & OBJPR_NOTMAPPED) == 0 &&
2189 object->ref_count != 0 &&
2190 !vm_page_try_remove_write(p))
2192 if (p->dirty != 0) {
2197 if ((options & OBJPR_NOTMAPPED) == 0 &&
2198 object->ref_count != 0 && !vm_page_try_remove_all(p))
2202 vm_object_pip_wakeup(object);
2204 vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2209 * vm_object_page_noreuse:
2211 * For the given object, attempt to move the specified pages to
2212 * the head of the inactive queue. This bypasses regular LRU
2213 * operation and allows the pages to be reused quickly under memory
2214 * pressure. If a page is wired for any reason, then it will not
2215 * be queued. Pages are specified by the range ["start", "end").
2216 * As a special case, if "end" is zero, then the range extends from
2217 * "start" to the end of the object.
2219 * This operation should only be performed on objects that
2220 * contain non-fictitious, managed pages.
2222 * The object must be locked.
2225 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2229 VM_OBJECT_ASSERT_LOCKED(object);
2230 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2231 ("vm_object_page_noreuse: illegal object %p", object));
2232 if (object->resident_page_count == 0)
2234 p = vm_page_find_least(object, start);
2237 * Here, the variable "p" is either (1) the page with the least pindex
2238 * greater than or equal to the parameter "start" or (2) NULL.
2240 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2241 next = TAILQ_NEXT(p, listq);
2242 vm_page_deactivate_noreuse(p);
2247 * Populate the specified range of the object with valid pages. Returns
2248 * TRUE if the range is successfully populated and FALSE otherwise.
2250 * Note: This function should be optimized to pass a larger array of
2251 * pages to vm_pager_get_pages() before it is applied to a non-
2252 * OBJT_DEVICE object.
2254 * The object must be locked.
2257 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2263 VM_OBJECT_ASSERT_WLOCKED(object);
2264 for (pindex = start; pindex < end; pindex++) {
2265 rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2266 if (rv != VM_PAGER_OK)
2270 * Keep "m" busy because a subsequent iteration may unlock
2274 if (pindex > start) {
2275 m = vm_page_lookup(object, start);
2276 while (m != NULL && m->pindex < pindex) {
2278 m = TAILQ_NEXT(m, listq);
2281 return (pindex == end);
2285 * Routine: vm_object_coalesce
2286 * Function: Coalesces two objects backing up adjoining
2287 * regions of memory into a single object.
2289 * returns TRUE if objects were combined.
2291 * NOTE: Only works at the moment if the second object is NULL -
2292 * if it's not, which object do we lock first?
2295 * prev_object First object to coalesce
2296 * prev_offset Offset into prev_object
2297 * prev_size Size of reference to prev_object
2298 * next_size Size of reference to the second object
2299 * reserved Indicator that extension region has
2300 * swap accounted for
2303 * The object must *not* be locked.
2306 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2307 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2309 vm_pindex_t next_pindex;
2311 if (prev_object == NULL)
2313 if ((prev_object->flags & OBJ_ANON) == 0)
2316 VM_OBJECT_WLOCK(prev_object);
2318 * Try to collapse the object first.
2320 vm_object_collapse(prev_object);
2323 * Can't coalesce if: . more than one reference . paged out . shadows
2324 * another object . has a copy elsewhere (any of which mean that the
2325 * pages not mapped to prev_entry may be in use anyway)
2327 if (prev_object->backing_object != NULL) {
2328 VM_OBJECT_WUNLOCK(prev_object);
2332 prev_size >>= PAGE_SHIFT;
2333 next_size >>= PAGE_SHIFT;
2334 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2336 if (prev_object->ref_count > 1 &&
2337 prev_object->size != next_pindex &&
2338 (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2339 VM_OBJECT_WUNLOCK(prev_object);
2344 * Account for the charge.
2346 if (prev_object->cred != NULL) {
2348 * If prev_object was charged, then this mapping,
2349 * although not charged now, may become writable
2350 * later. Non-NULL cred in the object would prevent
2351 * swap reservation during enabling of the write
2352 * access, so reserve swap now. Failed reservation
2353 * cause allocation of the separate object for the map
2354 * entry, and swap reservation for this entry is
2355 * managed in appropriate time.
2357 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2358 prev_object->cred)) {
2359 VM_OBJECT_WUNLOCK(prev_object);
2362 prev_object->charge += ptoa(next_size);
2366 * Remove any pages that may still be in the object from a previous
2369 if (next_pindex < prev_object->size) {
2370 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2373 if (prev_object->cred != NULL) {
2374 KASSERT(prev_object->charge >=
2375 ptoa(prev_object->size - next_pindex),
2376 ("object %p overcharged 1 %jx %jx", prev_object,
2377 (uintmax_t)next_pindex, (uintmax_t)next_size));
2378 prev_object->charge -= ptoa(prev_object->size -
2385 * Extend the object if necessary.
2387 if (next_pindex + next_size > prev_object->size)
2388 prev_object->size = next_pindex + next_size;
2390 VM_OBJECT_WUNLOCK(prev_object);
2395 vm_object_set_writeable_dirty_(vm_object_t object)
2397 atomic_add_int(&object->generation, 1);
2401 vm_object_mightbedirty_(vm_object_t object)
2403 return (object->generation != object->cleangeneration);
2409 * For each page offset within the specified range of the given object,
2410 * find the highest-level page in the shadow chain and unwire it. A page
2411 * must exist at every page offset, and the highest-level page must be
2415 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2418 vm_object_t tobject, t1object;
2420 vm_pindex_t end_pindex, pindex, tpindex;
2421 int depth, locked_depth;
2423 KASSERT((offset & PAGE_MASK) == 0,
2424 ("vm_object_unwire: offset is not page aligned"));
2425 KASSERT((length & PAGE_MASK) == 0,
2426 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2427 /* The wired count of a fictitious page never changes. */
2428 if ((object->flags & OBJ_FICTITIOUS) != 0)
2430 pindex = OFF_TO_IDX(offset);
2431 end_pindex = pindex + atop(length);
2434 VM_OBJECT_RLOCK(object);
2435 m = vm_page_find_least(object, pindex);
2436 while (pindex < end_pindex) {
2437 if (m == NULL || pindex < m->pindex) {
2439 * The first object in the shadow chain doesn't
2440 * contain a page at the current index. Therefore,
2441 * the page must exist in a backing object.
2448 OFF_TO_IDX(tobject->backing_object_offset);
2449 tobject = tobject->backing_object;
2450 KASSERT(tobject != NULL,
2451 ("vm_object_unwire: missing page"));
2452 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2455 if (depth == locked_depth) {
2457 VM_OBJECT_RLOCK(tobject);
2459 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2463 m = TAILQ_NEXT(m, listq);
2465 if (vm_page_trysbusy(tm) == 0) {
2466 for (tobject = object; locked_depth >= 1;
2468 t1object = tobject->backing_object;
2469 if (tm->object != tobject)
2470 VM_OBJECT_RUNLOCK(tobject);
2473 tobject = tm->object;
2474 if (!vm_page_busy_sleep(tm, "unwbo",
2475 VM_ALLOC_IGN_SBUSY))
2476 VM_OBJECT_RUNLOCK(tobject);
2479 vm_page_unwire(tm, queue);
2480 vm_page_sunbusy(tm);
2484 /* Release the accumulated object locks. */
2485 for (tobject = object; locked_depth >= 1; locked_depth--) {
2486 t1object = tobject->backing_object;
2487 VM_OBJECT_RUNLOCK(tobject);
2493 * Return the vnode for the given object, or NULL if none exists.
2494 * For tmpfs objects, the function may return NULL if there is
2495 * no vnode allocated at the time of the call.
2498 vm_object_vnode(vm_object_t object)
2502 VM_OBJECT_ASSERT_LOCKED(object);
2503 vm_pager_getvp(object, &vp, NULL);
2508 * Busy the vm object. This prevents new pages belonging to the object from
2509 * becoming busy. Existing pages persist as busy. Callers are responsible
2510 * for checking page state before proceeding.
2513 vm_object_busy(vm_object_t obj)
2516 VM_OBJECT_ASSERT_LOCKED(obj);
2518 blockcount_acquire(&obj->busy, 1);
2519 /* The fence is required to order loads of page busy. */
2520 atomic_thread_fence_acq_rel();
2524 vm_object_unbusy(vm_object_t obj)
2527 blockcount_release(&obj->busy, 1);
2531 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2534 VM_OBJECT_ASSERT_UNLOCKED(obj);
2536 (void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2540 * This function aims to determine if the object is mapped,
2541 * specifically, if it is referenced by a vm_map_entry. Because
2542 * objects occasionally acquire transient references that do not
2543 * represent a mapping, the method used here is inexact. However, it
2544 * has very low overhead and is good enough for the advisory
2545 * vm.vmtotal sysctl.
2548 vm_object_is_active(vm_object_t obj)
2551 return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2555 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2557 struct kinfo_vmobject *kvo;
2558 char *fullpath, *freepath;
2566 if (req->oldptr == NULL) {
2568 * If an old buffer has not been provided, generate an
2569 * estimate of the space needed for a subsequent call.
2571 mtx_lock(&vm_object_list_mtx);
2573 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2574 if (obj->type == OBJT_DEAD)
2578 mtx_unlock(&vm_object_list_mtx);
2579 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2583 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO);
2587 * VM objects are type stable and are never removed from the
2588 * list once added. This allows us to safely read obj->object_list
2589 * after reacquiring the VM object lock.
2591 mtx_lock(&vm_object_list_mtx);
2592 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2593 if (obj->type == OBJT_DEAD ||
2594 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2596 VM_OBJECT_RLOCK(obj);
2597 if (obj->type == OBJT_DEAD ||
2598 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2599 VM_OBJECT_RUNLOCK(obj);
2602 mtx_unlock(&vm_object_list_mtx);
2603 kvo->kvo_size = ptoa(obj->size);
2604 kvo->kvo_resident = obj->resident_page_count;
2605 kvo->kvo_ref_count = obj->ref_count;
2606 kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2607 kvo->kvo_memattr = obj->memattr;
2608 kvo->kvo_active = 0;
2609 kvo->kvo_inactive = 0;
2611 TAILQ_FOREACH(m, &obj->memq, listq) {
2613 * A page may belong to the object but be
2614 * dequeued and set to PQ_NONE while the
2615 * object lock is not held. This makes the
2616 * reads of m->queue below racy, and we do not
2617 * count pages set to PQ_NONE. However, this
2618 * sysctl is only meant to give an
2619 * approximation of the system anyway.
2621 if (m->a.queue == PQ_ACTIVE)
2623 else if (m->a.queue == PQ_INACTIVE)
2624 kvo->kvo_inactive++;
2628 kvo->kvo_vn_fileid = 0;
2629 kvo->kvo_vn_fsid = 0;
2630 kvo->kvo_vn_fsid_freebsd11 = 0;
2634 kvo->kvo_type = vm_object_kvme_type(obj, swap_only ? NULL : &vp);
2637 } else if ((obj->flags & OBJ_ANON) != 0) {
2638 MPASS(kvo->kvo_type == KVME_TYPE_DEFAULT ||
2639 kvo->kvo_type == KVME_TYPE_SWAP);
2640 kvo->kvo_me = (uintptr_t)obj;
2641 /* tmpfs objs are reported as vnodes */
2642 kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2643 sp = swap_pager_swapped_pages(obj);
2644 kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2646 VM_OBJECT_RUNLOCK(obj);
2648 vn_fullpath(vp, &fullpath, &freepath);
2649 vn_lock(vp, LK_SHARED | LK_RETRY);
2650 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2651 kvo->kvo_vn_fileid = va.va_fileid;
2652 kvo->kvo_vn_fsid = va.va_fsid;
2653 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2659 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2660 if (freepath != NULL)
2661 free(freepath, M_TEMP);
2663 /* Pack record size down */
2664 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2665 + strlen(kvo->kvo_path) + 1;
2666 kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2668 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2670 mtx_lock(&vm_object_list_mtx);
2674 mtx_unlock(&vm_object_list_mtx);
2680 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2682 return (vm_object_list_handler(req, false));
2685 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2686 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2687 "List of VM objects");
2690 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2692 return (vm_object_list_handler(req, true));
2696 * This sysctl returns list of the anonymous or swap objects. Intent
2697 * is to provide stripped optimized list useful to analyze swap use.
2698 * Since technically non-swap (default) objects participate in the
2699 * shadow chains, and are converted to swap type as needed by swap
2700 * pager, we must report them.
2702 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2703 CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2704 sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2705 "List of swap VM objects");
2707 #include "opt_ddb.h"
2709 #include <sys/kernel.h>
2711 #include <sys/cons.h>
2713 #include <ddb/ddb.h>
2716 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2719 vm_map_entry_t tmpe;
2726 VM_MAP_ENTRY_FOREACH(tmpe, map) {
2727 if (_vm_object_in_map(map, object, tmpe)) {
2731 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2732 tmpm = entry->object.sub_map;
2733 VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2734 if (_vm_object_in_map(tmpm, object, tmpe)) {
2738 } else if ((obj = entry->object.vm_object) != NULL) {
2739 for (; obj; obj = obj->backing_object)
2740 if (obj == object) {
2748 vm_object_in_map(vm_object_t object)
2752 /* sx_slock(&allproc_lock); */
2753 FOREACH_PROC_IN_SYSTEM(p) {
2754 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2756 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2757 /* sx_sunlock(&allproc_lock); */
2761 /* sx_sunlock(&allproc_lock); */
2762 if (_vm_object_in_map(kernel_map, object, 0))
2767 DB_SHOW_COMMAND(vmochk, vm_object_check)
2772 * make sure that internal objs are in a map somewhere
2773 * and none have zero ref counts.
2775 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2776 if ((object->flags & OBJ_ANON) != 0) {
2777 if (object->ref_count == 0) {
2778 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2779 (long)object->size);
2781 if (!vm_object_in_map(object)) {
2783 "vmochk: internal obj is not in a map: "
2784 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2785 object->ref_count, (u_long)object->size,
2786 (u_long)object->size,
2787 (void *)object->backing_object);
2796 * vm_object_print: [ debug ]
2798 DB_SHOW_COMMAND(object, vm_object_print_static)
2800 /* XXX convert args. */
2801 vm_object_t object = (vm_object_t)addr;
2802 boolean_t full = have_addr;
2806 /* XXX count is an (unused) arg. Avoid shadowing it. */
2807 #define count was_count
2815 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2816 object, (int)object->type, (uintmax_t)object->size,
2817 object->resident_page_count, object->ref_count, object->flags,
2818 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2819 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2820 atomic_load_int(&object->shadow_count),
2821 object->backing_object ? object->backing_object->ref_count : 0,
2822 object->backing_object, (uintmax_t)object->backing_object_offset);
2829 TAILQ_FOREACH(p, &object->memq, listq) {
2831 db_iprintf("memory:=");
2832 else if (count == 6) {
2840 db_printf("(off=0x%jx,page=0x%jx)",
2841 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2854 /* XXX need this non-static entry for calling from vm_map_print. */
2857 /* db_expr_t */ long addr,
2858 boolean_t have_addr,
2859 /* db_expr_t */ long count,
2862 vm_object_print_static(addr, have_addr, count, modif);
2865 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2870 vm_page_t m, prev_m;
2873 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2874 db_printf("new object: %p\n", (void *)object);
2881 TAILQ_FOREACH(m, &object->memq, listq) {
2882 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2883 prev_m->pindex + 1 != m->pindex) {
2885 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2886 (long)fidx, rcount, (long)pa);
2893 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2898 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2899 (long)fidx, rcount, (long)pa);
2904 pa = VM_PAGE_TO_PHYS(m);
2908 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2909 (long)fidx, rcount, (long)pa);