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));
211 vm_object_zinit(void *mem, int size, int flags)
215 object = (vm_object_t)mem;
216 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
218 /* These are true for any object that has been freed */
219 object->type = OBJT_DEAD;
220 vm_radix_init(&object->rtree);
221 refcount_init(&object->ref_count, 0);
222 blockcount_init(&object->paging_in_progress);
223 blockcount_init(&object->busy);
224 object->resident_page_count = 0;
225 atomic_store_int(&object->shadow_count, 0);
226 object->flags = OBJ_DEAD;
228 mtx_lock(&vm_object_list_mtx);
229 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
230 mtx_unlock(&vm_object_list_mtx);
235 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
236 vm_object_t object, void *handle)
239 TAILQ_INIT(&object->memq);
240 LIST_INIT(&object->shadow_head);
243 object->flags = flags;
244 if ((flags & OBJ_SWAP) != 0)
245 pctrie_init(&object->un_pager.swp.swp_blks);
248 * Ensure that swap_pager_swapoff() iteration over object_list
249 * sees up to date type and pctrie head if it observed
252 atomic_thread_fence_rel();
254 object->pg_color = 0;
256 object->domain.dr_policy = NULL;
257 object->generation = 1;
258 object->cleangeneration = 1;
259 refcount_init(&object->ref_count, 1);
260 object->memattr = VM_MEMATTR_DEFAULT;
263 object->handle = handle;
264 object->backing_object = NULL;
265 object->backing_object_offset = (vm_ooffset_t) 0;
266 #if VM_NRESERVLEVEL > 0
267 LIST_INIT(&object->rvq);
269 umtx_shm_object_init(object);
275 * Initialize the VM objects module.
280 TAILQ_INIT(&vm_object_list);
281 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
283 rw_init(&kernel_object->lock, "kernel vm object");
284 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
285 VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
286 #if VM_NRESERVLEVEL > 0
287 kernel_object->flags |= OBJ_COLORED;
288 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
290 kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
293 * The lock portion of struct vm_object must be type stable due
294 * to vm_pageout_fallback_object_lock locking a vm object
295 * without holding any references to it.
297 * paging_in_progress is valid always. Lockless references to
298 * the objects may acquire pip and then check OBJ_DEAD.
300 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
306 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
312 vm_object_clear_flag(vm_object_t object, u_short bits)
315 VM_OBJECT_ASSERT_WLOCKED(object);
316 object->flags &= ~bits;
320 * Sets the default memory attribute for the specified object. Pages
321 * that are allocated to this object are by default assigned this memory
324 * Presently, this function must be called before any pages are allocated
325 * to the object. In the future, this requirement may be relaxed for
326 * "default" and "swap" objects.
329 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
332 VM_OBJECT_ASSERT_WLOCKED(object);
334 if (object->type == OBJT_DEAD)
335 return (KERN_INVALID_ARGUMENT);
336 if (!TAILQ_EMPTY(&object->memq))
337 return (KERN_FAILURE);
339 object->memattr = memattr;
340 return (KERN_SUCCESS);
344 vm_object_pip_add(vm_object_t object, short i)
348 blockcount_acquire(&object->paging_in_progress, i);
352 vm_object_pip_wakeup(vm_object_t object)
355 vm_object_pip_wakeupn(object, 1);
359 vm_object_pip_wakeupn(vm_object_t object, short i)
363 blockcount_release(&object->paging_in_progress, i);
367 * Atomically drop the object lock and wait for pip to drain. This protects
368 * from sleep/wakeup races due to identity changes. The lock is not re-acquired
372 vm_object_pip_sleep(vm_object_t object, const char *waitid)
375 (void)blockcount_sleep(&object->paging_in_progress, &object->lock,
376 waitid, PVM | PDROP);
380 vm_object_pip_wait(vm_object_t object, const char *waitid)
383 VM_OBJECT_ASSERT_WLOCKED(object);
385 blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
390 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
393 VM_OBJECT_ASSERT_UNLOCKED(object);
395 blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
399 * vm_object_allocate:
401 * Returns a new object with the given size.
404 vm_object_allocate(objtype_t type, vm_pindex_t size)
411 panic("vm_object_allocate: can't create OBJT_DEAD");
416 flags = OBJ_COLORED | OBJ_SWAP;
420 flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
423 flags = OBJ_FICTITIOUS;
426 flags = OBJ_UNMANAGED;
432 panic("vm_object_allocate: type %d is undefined or dynamic",
435 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
436 _vm_object_allocate(type, size, flags, object, NULL);
442 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags)
446 MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */);
447 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
448 _vm_object_allocate(dyntype, size, flags, object, NULL);
454 * vm_object_allocate_anon:
456 * Returns a new default object of the given size and marked as
457 * anonymous memory for special split/collapse handling. Color
458 * to be initialized by the caller.
461 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
462 struct ucred *cred, vm_size_t charge)
464 vm_object_t handle, object;
466 if (backing_object == NULL)
468 else if ((backing_object->flags & OBJ_ANON) != 0)
469 handle = backing_object->handle;
471 handle = backing_object;
472 object = uma_zalloc(obj_zone, M_WAITOK);
473 _vm_object_allocate(OBJT_DEFAULT, size, OBJ_ANON | OBJ_ONEMAPPING,
476 object->charge = cred != NULL ? charge : 0;
481 vm_object_reference_vnode(vm_object_t object)
486 * vnode objects need the lock for the first reference
487 * to serialize with vnode_object_deallocate().
489 if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
490 VM_OBJECT_RLOCK(object);
491 old = refcount_acquire(&object->ref_count);
492 if (object->type == OBJT_VNODE && old == 0)
493 vref(object->handle);
494 VM_OBJECT_RUNLOCK(object);
499 * vm_object_reference:
501 * Acquires a reference to the given object.
504 vm_object_reference(vm_object_t object)
510 if (object->type == OBJT_VNODE)
511 vm_object_reference_vnode(object);
513 refcount_acquire(&object->ref_count);
514 KASSERT((object->flags & OBJ_DEAD) == 0,
515 ("vm_object_reference: Referenced dead object."));
519 * vm_object_reference_locked:
521 * Gets another reference to the given object.
523 * The object must be locked.
526 vm_object_reference_locked(vm_object_t object)
530 VM_OBJECT_ASSERT_LOCKED(object);
531 old = refcount_acquire(&object->ref_count);
532 if (object->type == OBJT_VNODE && old == 0)
533 vref(object->handle);
534 KASSERT((object->flags & OBJ_DEAD) == 0,
535 ("vm_object_reference: Referenced dead object."));
539 * Handle deallocating an object of type OBJT_VNODE.
542 vm_object_deallocate_vnode(vm_object_t object)
544 struct vnode *vp = (struct vnode *) object->handle;
547 KASSERT(object->type == OBJT_VNODE,
548 ("vm_object_deallocate_vnode: not a vnode object"));
549 KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
551 /* Object lock to protect handle lookup. */
552 last = refcount_release(&object->ref_count);
553 VM_OBJECT_RUNLOCK(object);
558 if (!umtx_shm_vnobj_persistent)
559 umtx_shm_object_terminated(object);
561 /* vrele may need the vnode lock. */
566 * We dropped a reference on an object and discovered that it had a
567 * single remaining shadow. This is a sibling of the reference we
568 * dropped. Attempt to collapse the sibling and backing object.
571 vm_object_deallocate_anon(vm_object_t backing_object)
575 /* Fetch the final shadow. */
576 object = LIST_FIRST(&backing_object->shadow_head);
577 KASSERT(object != NULL &&
578 atomic_load_int(&backing_object->shadow_count) == 1,
579 ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
580 backing_object->ref_count,
581 atomic_load_int(&backing_object->shadow_count)));
582 KASSERT((object->flags & OBJ_ANON) != 0,
583 ("invalid shadow object %p", object));
585 if (!VM_OBJECT_TRYWLOCK(object)) {
587 * Prevent object from disappearing since we do not have a
590 vm_object_pip_add(object, 1);
591 VM_OBJECT_WUNLOCK(backing_object);
592 VM_OBJECT_WLOCK(object);
593 vm_object_pip_wakeup(object);
595 VM_OBJECT_WUNLOCK(backing_object);
598 * Check for a collapse/terminate race with the last reference holder.
600 if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
601 !refcount_acquire_if_not_zero(&object->ref_count)) {
602 VM_OBJECT_WUNLOCK(object);
605 backing_object = object->backing_object;
606 if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
607 vm_object_collapse(object);
608 VM_OBJECT_WUNLOCK(object);
614 * vm_object_deallocate:
616 * Release a reference to the specified object,
617 * gained either through a vm_object_allocate
618 * or a vm_object_reference call. When all references
619 * are gone, storage associated with this object
620 * may be relinquished.
622 * No object may be locked.
625 vm_object_deallocate(vm_object_t object)
630 while (object != NULL) {
632 * If the reference count goes to 0 we start calling
633 * vm_object_terminate() on the object chain. A ref count
634 * of 1 may be a special case depending on the shadow count
635 * being 0 or 1. These cases require a write lock on the
638 if ((object->flags & OBJ_ANON) == 0)
639 released = refcount_release_if_gt(&object->ref_count, 1);
641 released = refcount_release_if_gt(&object->ref_count, 2);
645 if (object->type == OBJT_VNODE) {
646 VM_OBJECT_RLOCK(object);
647 if (object->type == OBJT_VNODE) {
648 vm_object_deallocate_vnode(object);
651 VM_OBJECT_RUNLOCK(object);
654 VM_OBJECT_WLOCK(object);
655 KASSERT(object->ref_count > 0,
656 ("vm_object_deallocate: object deallocated too many times: %d",
660 * If this is not the final reference to an anonymous
661 * object we may need to collapse the shadow chain.
663 if (!refcount_release(&object->ref_count)) {
664 if (object->ref_count > 1 ||
665 atomic_load_int(&object->shadow_count) == 0) {
666 if ((object->flags & OBJ_ANON) != 0 &&
667 object->ref_count == 1)
668 vm_object_set_flag(object,
670 VM_OBJECT_WUNLOCK(object);
674 /* Handle collapsing last ref on anonymous objects. */
675 object = vm_object_deallocate_anon(object);
680 * Handle the final reference to an object. We restart
681 * the loop with the backing object to avoid recursion.
683 umtx_shm_object_terminated(object);
684 temp = object->backing_object;
686 KASSERT(object->type == OBJT_DEFAULT ||
687 object->type == OBJT_SWAP,
688 ("shadowed tmpfs v_object 2 %p", object));
689 vm_object_backing_remove(object);
692 KASSERT((object->flags & OBJ_DEAD) == 0,
693 ("vm_object_deallocate: Terminating dead object."));
694 vm_object_set_flag(object, OBJ_DEAD);
695 vm_object_terminate(object);
701 * vm_object_destroy removes the object from the global object list
702 * and frees the space for the object.
705 vm_object_destroy(vm_object_t object)
709 * Release the allocation charge.
711 if (object->cred != NULL) {
712 swap_release_by_cred(object->charge, object->cred);
714 crfree(object->cred);
719 * Free the space for the object.
721 uma_zfree(obj_zone, object);
725 vm_object_sub_shadow(vm_object_t object)
727 KASSERT(object->shadow_count >= 1,
728 ("object %p sub_shadow count zero", object));
729 atomic_subtract_int(&object->shadow_count, 1);
733 vm_object_backing_remove_locked(vm_object_t object)
735 vm_object_t backing_object;
737 backing_object = object->backing_object;
738 VM_OBJECT_ASSERT_WLOCKED(object);
739 VM_OBJECT_ASSERT_WLOCKED(backing_object);
741 KASSERT((object->flags & OBJ_COLLAPSING) == 0,
742 ("vm_object_backing_remove: Removing collapsing object."));
744 vm_object_sub_shadow(backing_object);
745 if ((object->flags & OBJ_SHADOWLIST) != 0) {
746 LIST_REMOVE(object, shadow_list);
747 object->flags &= ~OBJ_SHADOWLIST;
749 object->backing_object = NULL;
753 vm_object_backing_remove(vm_object_t object)
755 vm_object_t backing_object;
757 VM_OBJECT_ASSERT_WLOCKED(object);
759 backing_object = object->backing_object;
760 if ((object->flags & OBJ_SHADOWLIST) != 0) {
761 VM_OBJECT_WLOCK(backing_object);
762 vm_object_backing_remove_locked(object);
763 VM_OBJECT_WUNLOCK(backing_object);
765 object->backing_object = NULL;
766 vm_object_sub_shadow(backing_object);
771 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
774 VM_OBJECT_ASSERT_WLOCKED(object);
776 atomic_add_int(&backing_object->shadow_count, 1);
777 if ((backing_object->flags & OBJ_ANON) != 0) {
778 VM_OBJECT_ASSERT_WLOCKED(backing_object);
779 LIST_INSERT_HEAD(&backing_object->shadow_head, object,
781 object->flags |= OBJ_SHADOWLIST;
783 object->backing_object = backing_object;
787 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
790 VM_OBJECT_ASSERT_WLOCKED(object);
792 if ((backing_object->flags & OBJ_ANON) != 0) {
793 VM_OBJECT_WLOCK(backing_object);
794 vm_object_backing_insert_locked(object, backing_object);
795 VM_OBJECT_WUNLOCK(backing_object);
797 object->backing_object = backing_object;
798 atomic_add_int(&backing_object->shadow_count, 1);
803 * Insert an object into a backing_object's shadow list with an additional
804 * reference to the backing_object added.
807 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
810 VM_OBJECT_ASSERT_WLOCKED(object);
812 if ((backing_object->flags & OBJ_ANON) != 0) {
813 VM_OBJECT_WLOCK(backing_object);
814 KASSERT((backing_object->flags & OBJ_DEAD) == 0,
815 ("shadowing dead anonymous object"));
816 vm_object_reference_locked(backing_object);
817 vm_object_backing_insert_locked(object, backing_object);
818 vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
819 VM_OBJECT_WUNLOCK(backing_object);
821 vm_object_reference(backing_object);
822 atomic_add_int(&backing_object->shadow_count, 1);
823 object->backing_object = backing_object;
828 * Transfer a backing reference from backing_object to object.
831 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
833 vm_object_t new_backing_object;
836 * Note that the reference to backing_object->backing_object
837 * moves from within backing_object to within object.
839 vm_object_backing_remove_locked(object);
840 new_backing_object = backing_object->backing_object;
841 if (new_backing_object == NULL)
843 if ((new_backing_object->flags & OBJ_ANON) != 0) {
844 VM_OBJECT_WLOCK(new_backing_object);
845 vm_object_backing_remove_locked(backing_object);
846 vm_object_backing_insert_locked(object, new_backing_object);
847 VM_OBJECT_WUNLOCK(new_backing_object);
850 * shadow_count for new_backing_object is left
851 * unchanged, its reference provided by backing_object
852 * is replaced by object.
854 object->backing_object = new_backing_object;
855 backing_object->backing_object = NULL;
860 * Wait for a concurrent collapse to settle.
863 vm_object_collapse_wait(vm_object_t object)
866 VM_OBJECT_ASSERT_WLOCKED(object);
868 while ((object->flags & OBJ_COLLAPSING) != 0) {
869 vm_object_pip_wait(object, "vmcolwait");
870 counter_u64_add(object_collapse_waits, 1);
875 * Waits for a backing object to clear a pending collapse and returns
876 * it locked if it is an ANON object.
879 vm_object_backing_collapse_wait(vm_object_t object)
881 vm_object_t backing_object;
883 VM_OBJECT_ASSERT_WLOCKED(object);
886 backing_object = object->backing_object;
887 if (backing_object == NULL ||
888 (backing_object->flags & OBJ_ANON) == 0)
890 VM_OBJECT_WLOCK(backing_object);
891 if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
893 VM_OBJECT_WUNLOCK(object);
894 vm_object_pip_sleep(backing_object, "vmbckwait");
895 counter_u64_add(object_collapse_waits, 1);
896 VM_OBJECT_WLOCK(object);
898 return (backing_object);
902 * vm_object_terminate_pages removes any remaining pageable pages
903 * from the object and resets the object to an empty state.
906 vm_object_terminate_pages(vm_object_t object)
910 VM_OBJECT_ASSERT_WLOCKED(object);
913 * Free any remaining pageable pages. This also removes them from the
914 * paging queues. However, don't free wired pages, just remove them
915 * from the object. Rather than incrementally removing each page from
916 * the object, the page and object are reset to any empty state.
918 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
919 vm_page_assert_unbusied(p);
920 KASSERT(p->object == object &&
921 (p->ref_count & VPRC_OBJREF) != 0,
922 ("vm_object_terminate_pages: page %p is inconsistent", p));
925 if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
932 * If the object contained any pages, then reset it to an empty state.
933 * None of the object's fields, including "resident_page_count", were
934 * modified by the preceding loop.
936 if (object->resident_page_count != 0) {
937 vm_radix_reclaim_allnodes(&object->rtree);
938 TAILQ_INIT(&object->memq);
939 object->resident_page_count = 0;
940 if (object->type == OBJT_VNODE)
941 vdrop(object->handle);
946 * vm_object_terminate actually destroys the specified object, freeing
947 * up all previously used resources.
949 * The object must be locked.
950 * This routine may block.
953 vm_object_terminate(vm_object_t object)
956 VM_OBJECT_ASSERT_WLOCKED(object);
957 KASSERT((object->flags & OBJ_DEAD) != 0,
958 ("terminating non-dead obj %p", object));
959 KASSERT((object->flags & OBJ_COLLAPSING) == 0,
960 ("terminating collapsing obj %p", object));
961 KASSERT(object->backing_object == NULL,
962 ("terminating shadow obj %p", object));
965 * Wait for the pageout daemon and other current users to be
966 * done with the object. Note that new paging_in_progress
967 * users can come after this wait, but they must check
968 * OBJ_DEAD flag set (without unlocking the object), and avoid
969 * the object being terminated.
971 vm_object_pip_wait(object, "objtrm");
973 KASSERT(object->ref_count == 0,
974 ("vm_object_terminate: object with references, ref_count=%d",
977 if ((object->flags & OBJ_PG_DTOR) == 0)
978 vm_object_terminate_pages(object);
980 #if VM_NRESERVLEVEL > 0
981 if (__predict_false(!LIST_EMPTY(&object->rvq)))
982 vm_reserv_break_all(object);
985 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
986 (object->flags & OBJ_SWAP) != 0,
987 ("%s: non-swap obj %p has cred", __func__, object));
990 * Let the pager know object is dead.
992 vm_pager_deallocate(object);
993 VM_OBJECT_WUNLOCK(object);
995 vm_object_destroy(object);
999 * Make the page read-only so that we can clear the object flags. However, if
1000 * this is a nosync mmap then the object is likely to stay dirty so do not
1001 * mess with the page and do not clear the object flags. Returns TRUE if the
1002 * page should be flushed, and FALSE otherwise.
1005 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
1008 vm_page_assert_busied(p);
1011 * If we have been asked to skip nosync pages and this is a
1012 * nosync page, skip it. Note that the object flags were not
1013 * cleared in this case so we do not have to set them.
1015 if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
1019 pmap_remove_write(p);
1020 return (p->dirty != 0);
1025 * vm_object_page_clean
1027 * Clean all dirty pages in the specified range of object. Leaves page
1028 * on whatever queue it is currently on. If NOSYNC is set then do not
1029 * write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1030 * leaving the object dirty.
1032 * For swap objects backing tmpfs regular files, do not flush anything,
1033 * but remove write protection on the mapped pages to update mtime through
1036 * When stuffing pages asynchronously, allow clustering. XXX we need a
1037 * synchronous clustering mode implementation.
1039 * Odd semantics: if start == end, we clean everything.
1041 * The object must be locked.
1043 * Returns FALSE if some page from the range was not written, as
1044 * reported by the pager, and TRUE otherwise.
1047 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1051 vm_pindex_t pi, tend, tstart;
1052 int curgeneration, n, pagerflags;
1053 boolean_t eio, res, allclean;
1055 VM_OBJECT_ASSERT_WLOCKED(object);
1057 if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1060 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1061 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1062 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1064 tstart = OFF_TO_IDX(start);
1065 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1066 allclean = tstart == 0 && tend >= object->size;
1070 curgeneration = object->generation;
1072 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1076 np = TAILQ_NEXT(p, listq);
1077 if (vm_page_none_valid(p))
1079 if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1080 if (object->generation != curgeneration &&
1081 (flags & OBJPC_SYNC) != 0)
1083 np = vm_page_find_least(object, pi);
1086 if (!vm_object_page_remove_write(p, flags, &allclean)) {
1090 if (object->type == OBJT_VNODE) {
1091 n = vm_object_page_collect_flush(object, p, pagerflags,
1092 flags, &allclean, &eio);
1097 if (object->generation != curgeneration &&
1098 (flags & OBJPC_SYNC) != 0)
1102 * If the VOP_PUTPAGES() did a truncated write, so
1103 * that even the first page of the run is not fully
1104 * written, vm_pageout_flush() returns 0 as the run
1105 * length. Since the condition that caused truncated
1106 * write may be permanent, e.g. exhausted free space,
1107 * accepting n == 0 would cause an infinite loop.
1109 * Forwarding the iterator leaves the unwritten page
1110 * behind, but there is not much we can do there if
1111 * filesystem refuses to write it.
1121 np = vm_page_find_least(object, pi + n);
1124 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1128 * Leave updating cleangeneration for tmpfs objects to tmpfs
1129 * scan. It needs to update mtime, which happens for other
1130 * filesystems during page writeouts.
1132 if (allclean && object->type == OBJT_VNODE)
1133 object->cleangeneration = curgeneration;
1138 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1139 int flags, boolean_t *allclean, boolean_t *eio)
1141 vm_page_t ma[vm_pageout_page_count], p_first, tp;
1142 int count, i, mreq, runlen;
1144 vm_page_lock_assert(p, MA_NOTOWNED);
1145 vm_page_assert_xbusied(p);
1146 VM_OBJECT_ASSERT_WLOCKED(object);
1151 for (tp = p; count < vm_pageout_page_count; count++) {
1152 tp = vm_page_next(tp);
1153 if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1155 if (!vm_object_page_remove_write(tp, flags, allclean)) {
1156 vm_page_xunbusy(tp);
1161 for (p_first = p; count < vm_pageout_page_count; count++) {
1162 tp = vm_page_prev(p_first);
1163 if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1165 if (!vm_object_page_remove_write(tp, flags, allclean)) {
1166 vm_page_xunbusy(tp);
1173 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1176 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1181 * Note that there is absolutely no sense in writing out
1182 * anonymous objects, so we track down the vnode object
1184 * We invalidate (remove) all pages from the address space
1185 * for semantic correctness.
1187 * If the backing object is a device object with unmanaged pages, then any
1188 * mappings to the specified range of pages must be removed before this
1189 * function is called.
1191 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1192 * may start out with a NULL object.
1195 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1196 boolean_t syncio, boolean_t invalidate)
1198 vm_object_t backing_object;
1201 int error, flags, fsync_after;
1208 VM_OBJECT_WLOCK(object);
1209 while ((backing_object = object->backing_object) != NULL) {
1210 VM_OBJECT_WLOCK(backing_object);
1211 offset += object->backing_object_offset;
1212 VM_OBJECT_WUNLOCK(object);
1213 object = backing_object;
1214 if (object->size < OFF_TO_IDX(offset + size))
1215 size = IDX_TO_OFF(object->size) - offset;
1218 * Flush pages if writing is allowed, invalidate them
1219 * if invalidation requested. Pages undergoing I/O
1220 * will be ignored by vm_object_page_remove().
1222 * We cannot lock the vnode and then wait for paging
1223 * to complete without deadlocking against vm_fault.
1224 * Instead we simply call vm_object_page_remove() and
1225 * allow it to block internally on a page-by-page
1226 * basis when it encounters pages undergoing async
1229 if (object->type == OBJT_VNODE &&
1230 vm_object_mightbedirty(object) != 0 &&
1231 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1232 VM_OBJECT_WUNLOCK(object);
1233 (void) vn_start_write(vp, &mp, V_WAIT);
1234 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1235 if (syncio && !invalidate && offset == 0 &&
1236 atop(size) == object->size) {
1238 * If syncing the whole mapping of the file,
1239 * it is faster to schedule all the writes in
1240 * async mode, also allowing the clustering,
1241 * and then wait for i/o to complete.
1246 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1247 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1248 fsync_after = FALSE;
1250 VM_OBJECT_WLOCK(object);
1251 res = vm_object_page_clean(object, offset, offset + size,
1253 VM_OBJECT_WUNLOCK(object);
1255 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1257 vn_finished_write(mp);
1260 VM_OBJECT_WLOCK(object);
1262 if ((object->type == OBJT_VNODE ||
1263 object->type == OBJT_DEVICE) && invalidate) {
1264 if (object->type == OBJT_DEVICE)
1266 * The option OBJPR_NOTMAPPED must be passed here
1267 * because vm_object_page_remove() cannot remove
1268 * unmanaged mappings.
1270 flags = OBJPR_NOTMAPPED;
1274 flags = OBJPR_CLEANONLY;
1275 vm_object_page_remove(object, OFF_TO_IDX(offset),
1276 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1278 VM_OBJECT_WUNLOCK(object);
1283 * Determine whether the given advice can be applied to the object. Advice is
1284 * not applied to unmanaged pages since they never belong to page queues, and
1285 * since MADV_FREE is destructive, it can apply only to anonymous pages that
1286 * have been mapped at most once.
1289 vm_object_advice_applies(vm_object_t object, int advice)
1292 if ((object->flags & OBJ_UNMANAGED) != 0)
1294 if (advice != MADV_FREE)
1296 return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1297 (OBJ_ONEMAPPING | OBJ_ANON));
1301 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1305 if (advice == MADV_FREE)
1306 vm_pager_freespace(object, pindex, size);
1310 * vm_object_madvise:
1312 * Implements the madvise function at the object/page level.
1314 * MADV_WILLNEED (any object)
1316 * Activate the specified pages if they are resident.
1318 * MADV_DONTNEED (any object)
1320 * Deactivate the specified pages if they are resident.
1322 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1323 * OBJ_ONEMAPPING only)
1325 * Deactivate and clean the specified pages if they are
1326 * resident. This permits the process to reuse the pages
1327 * without faulting or the kernel to reclaim the pages
1331 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1334 vm_pindex_t tpindex;
1335 vm_object_t backing_object, tobject;
1342 VM_OBJECT_WLOCK(object);
1343 if (!vm_object_advice_applies(object, advice)) {
1344 VM_OBJECT_WUNLOCK(object);
1347 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1351 * If the next page isn't resident in the top-level object, we
1352 * need to search the shadow chain. When applying MADV_FREE, we
1353 * take care to release any swap space used to store
1354 * non-resident pages.
1356 if (m == NULL || pindex < m->pindex) {
1358 * Optimize a common case: if the top-level object has
1359 * no backing object, we can skip over the non-resident
1360 * range in constant time.
1362 if (object->backing_object == NULL) {
1363 tpindex = (m != NULL && m->pindex < end) ?
1365 vm_object_madvise_freespace(object, advice,
1366 pindex, tpindex - pindex);
1367 if ((pindex = tpindex) == end)
1374 vm_object_madvise_freespace(tobject, advice,
1377 * Prepare to search the next object in the
1380 backing_object = tobject->backing_object;
1381 if (backing_object == NULL)
1383 VM_OBJECT_WLOCK(backing_object);
1385 OFF_TO_IDX(tobject->backing_object_offset);
1386 if (tobject != object)
1387 VM_OBJECT_WUNLOCK(tobject);
1388 tobject = backing_object;
1389 if (!vm_object_advice_applies(tobject, advice))
1391 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
1396 m = TAILQ_NEXT(m, listq);
1400 * If the page is not in a normal state, skip it. The page
1401 * can not be invalidated while the object lock is held.
1403 if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1405 KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1406 ("vm_object_madvise: page %p is fictitious", tm));
1407 KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1408 ("vm_object_madvise: page %p is not managed", tm));
1409 if (vm_page_tryxbusy(tm) == 0) {
1410 if (object != tobject)
1411 VM_OBJECT_WUNLOCK(object);
1412 if (advice == MADV_WILLNEED) {
1414 * Reference the page before unlocking and
1415 * sleeping so that the page daemon is less
1416 * likely to reclaim it.
1418 vm_page_aflag_set(tm, PGA_REFERENCED);
1420 if (!vm_page_busy_sleep(tm, "madvpo", 0))
1421 VM_OBJECT_WUNLOCK(tobject);
1424 vm_page_advise(tm, advice);
1425 vm_page_xunbusy(tm);
1426 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1428 if (tobject != object)
1429 VM_OBJECT_WUNLOCK(tobject);
1431 VM_OBJECT_WUNLOCK(object);
1437 * Create a new object which is backed by the
1438 * specified existing object range. The source
1439 * object reference is deallocated.
1441 * The new object and offset into that object
1442 * are returned in the source parameters.
1445 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1446 struct ucred *cred, bool shared)
1454 * Don't create the new object if the old object isn't shared.
1456 * If we hold the only reference we can guarantee that it won't
1457 * increase while we have the map locked. Otherwise the race is
1458 * harmless and we will end up with an extra shadow object that
1459 * will be collapsed later.
1461 if (source != NULL && source->ref_count == 1 &&
1462 (source->flags & OBJ_ANON) != 0)
1466 * Allocate a new object with the given length.
1468 result = vm_object_allocate_anon(atop(length), source, cred, length);
1471 * Store the offset into the source object, and fix up the offset into
1474 result->backing_object_offset = *offset;
1476 if (shared || source != NULL) {
1477 VM_OBJECT_WLOCK(result);
1480 * The new object shadows the source object, adding a
1481 * reference to it. Our caller changes his reference
1482 * to point to the new object, removing a reference to
1483 * the source object. Net result: no change of
1484 * reference count, unless the caller needs to add one
1485 * more reference due to forking a shared map entry.
1488 vm_object_reference_locked(result);
1489 vm_object_clear_flag(result, OBJ_ONEMAPPING);
1493 * Try to optimize the result object's page color when
1494 * shadowing in order to maintain page coloring
1495 * consistency in the combined shadowed object.
1497 if (source != NULL) {
1498 vm_object_backing_insert(result, source);
1499 result->domain = source->domain;
1500 #if VM_NRESERVLEVEL > 0
1501 result->flags |= source->flags & OBJ_COLORED;
1502 result->pg_color = (source->pg_color +
1503 OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1507 VM_OBJECT_WUNLOCK(result);
1511 * Return the new things
1520 * Split the pages in a map entry into a new object. This affords
1521 * easier removal of unused pages, and keeps object inheritance from
1522 * being a negative impact on memory usage.
1525 vm_object_split(vm_map_entry_t entry)
1527 vm_page_t m, m_busy, m_next;
1528 vm_object_t orig_object, new_object, backing_object;
1529 vm_pindex_t idx, offidxstart;
1532 orig_object = entry->object.vm_object;
1533 KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1534 ("vm_object_split: Splitting object with multiple mappings."));
1535 if ((orig_object->flags & OBJ_ANON) == 0)
1537 if (orig_object->ref_count <= 1)
1539 VM_OBJECT_WUNLOCK(orig_object);
1541 offidxstart = OFF_TO_IDX(entry->offset);
1542 size = atop(entry->end - entry->start);
1545 * If swap_pager_copy() is later called, it will convert new_object
1546 * into a swap object.
1548 new_object = vm_object_allocate_anon(size, orig_object,
1549 orig_object->cred, ptoa(size));
1552 * We must wait for the orig_object to complete any in-progress
1553 * collapse so that the swap blocks are stable below. The
1554 * additional reference on backing_object by new object will
1555 * prevent further collapse operations until split completes.
1557 VM_OBJECT_WLOCK(orig_object);
1558 vm_object_collapse_wait(orig_object);
1561 * At this point, the new object is still private, so the order in
1562 * which the original and new objects are locked does not matter.
1564 VM_OBJECT_WLOCK(new_object);
1565 new_object->domain = orig_object->domain;
1566 backing_object = orig_object->backing_object;
1567 if (backing_object != NULL) {
1568 vm_object_backing_insert_ref(new_object, backing_object);
1569 new_object->backing_object_offset =
1570 orig_object->backing_object_offset + entry->offset;
1572 if (orig_object->cred != NULL) {
1573 crhold(orig_object->cred);
1574 KASSERT(orig_object->charge >= ptoa(size),
1575 ("orig_object->charge < 0"));
1576 orig_object->charge -= ptoa(size);
1580 * Mark the split operation so that swap_pager_getpages() knows
1581 * that the object is in transition.
1583 vm_object_set_flag(orig_object, OBJ_SPLIT);
1589 m = vm_page_find_least(orig_object, offidxstart);
1590 KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1591 ("%s: object %p was repopulated", __func__, orig_object));
1592 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1594 m_next = TAILQ_NEXT(m, listq);
1597 * We must wait for pending I/O to complete before we can
1600 * We do not have to VM_PROT_NONE the page as mappings should
1601 * not be changed by this operation.
1603 if (vm_page_tryxbusy(m) == 0) {
1604 VM_OBJECT_WUNLOCK(new_object);
1605 if (vm_page_busy_sleep(m, "spltwt", 0))
1606 VM_OBJECT_WLOCK(orig_object);
1607 VM_OBJECT_WLOCK(new_object);
1612 * The page was left invalid. Likely placed there by
1613 * an incomplete fault. Just remove and ignore.
1615 if (vm_page_none_valid(m)) {
1616 if (vm_page_remove(m))
1621 /* vm_page_rename() will dirty the page. */
1622 if (vm_page_rename(m, new_object, idx)) {
1624 VM_OBJECT_WUNLOCK(new_object);
1625 VM_OBJECT_WUNLOCK(orig_object);
1627 VM_OBJECT_WLOCK(orig_object);
1628 VM_OBJECT_WLOCK(new_object);
1632 #if VM_NRESERVLEVEL > 0
1634 * If some of the reservation's allocated pages remain with
1635 * the original object, then transferring the reservation to
1636 * the new object is neither particularly beneficial nor
1637 * particularly harmful as compared to leaving the reservation
1638 * with the original object. If, however, all of the
1639 * reservation's allocated pages are transferred to the new
1640 * object, then transferring the reservation is typically
1641 * beneficial. Determining which of these two cases applies
1642 * would be more costly than unconditionally renaming the
1645 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1649 * orig_object's type may change while sleeping, so keep track
1650 * of the beginning of the busied range.
1652 if (orig_object->type != OBJT_SWAP)
1654 else if (m_busy == NULL)
1657 if ((orig_object->flags & OBJ_SWAP) != 0) {
1659 * swap_pager_copy() can sleep, in which case the orig_object's
1660 * and new_object's locks are released and reacquired.
1662 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1664 TAILQ_FOREACH_FROM(m_busy, &new_object->memq, listq)
1665 vm_page_xunbusy(m_busy);
1667 vm_object_clear_flag(orig_object, OBJ_SPLIT);
1668 VM_OBJECT_WUNLOCK(orig_object);
1669 VM_OBJECT_WUNLOCK(new_object);
1670 entry->object.vm_object = new_object;
1671 entry->offset = 0LL;
1672 vm_object_deallocate(orig_object);
1673 VM_OBJECT_WLOCK(new_object);
1677 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1679 vm_object_t backing_object;
1681 VM_OBJECT_ASSERT_WLOCKED(object);
1682 backing_object = object->backing_object;
1683 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1685 KASSERT(p == NULL || p->object == object || p->object == backing_object,
1686 ("invalid ownership %p %p %p", p, object, backing_object));
1687 /* The page is only NULL when rename fails. */
1689 VM_OBJECT_WUNLOCK(object);
1690 VM_OBJECT_WUNLOCK(backing_object);
1692 VM_OBJECT_WLOCK(object);
1693 } else if (p->object == object) {
1694 VM_OBJECT_WUNLOCK(backing_object);
1695 if (vm_page_busy_sleep(p, "vmocol", 0))
1696 VM_OBJECT_WLOCK(object);
1698 VM_OBJECT_WUNLOCK(object);
1699 if (!vm_page_busy_sleep(p, "vmocol", 0))
1700 VM_OBJECT_WUNLOCK(backing_object);
1701 VM_OBJECT_WLOCK(object);
1703 VM_OBJECT_WLOCK(backing_object);
1704 return (TAILQ_FIRST(&backing_object->memq));
1708 vm_object_scan_all_shadowed(vm_object_t object)
1710 vm_object_t backing_object;
1712 vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1714 VM_OBJECT_ASSERT_WLOCKED(object);
1715 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1717 backing_object = object->backing_object;
1719 if ((backing_object->flags & OBJ_ANON) == 0)
1722 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1723 p = vm_page_find_least(backing_object, pi);
1724 ps = swap_pager_find_least(backing_object, pi);
1727 * Only check pages inside the parent object's range and
1728 * inside the parent object's mapping of the backing object.
1731 if (p != NULL && p->pindex < pi)
1732 p = TAILQ_NEXT(p, listq);
1734 ps = swap_pager_find_least(backing_object, pi);
1735 if (p == NULL && ps >= backing_object->size)
1740 pi = MIN(p->pindex, ps);
1742 new_pindex = pi - backing_offset_index;
1743 if (new_pindex >= object->size)
1748 * If the backing object page is busy a
1749 * grandparent or older page may still be
1750 * undergoing CoW. It is not safe to collapse
1751 * the backing object until it is quiesced.
1753 if (vm_page_tryxbusy(p) == 0)
1757 * We raced with the fault handler that left
1758 * newly allocated invalid page on the object
1759 * queue and retried.
1761 if (!vm_page_all_valid(p))
1766 * See if the parent has the page or if the parent's object
1767 * pager has the page. If the parent has the page but the page
1768 * is not valid, the parent's object pager must have the page.
1770 * If this fails, the parent does not completely shadow the
1771 * object and we might as well give up now.
1773 pp = vm_page_lookup(object, new_pindex);
1776 * The valid check here is stable due to object lock
1777 * being required to clear valid and initiate paging.
1778 * Busy of p disallows fault handler to validate pp.
1780 if ((pp == NULL || vm_page_none_valid(pp)) &&
1781 !vm_pager_has_page(object, new_pindex, NULL, NULL))
1795 vm_object_collapse_scan(vm_object_t object)
1797 vm_object_t backing_object;
1798 vm_page_t next, p, pp;
1799 vm_pindex_t backing_offset_index, new_pindex;
1801 VM_OBJECT_ASSERT_WLOCKED(object);
1802 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1804 backing_object = object->backing_object;
1805 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1810 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1811 next = TAILQ_NEXT(p, listq);
1812 new_pindex = p->pindex - backing_offset_index;
1815 * Check for busy page
1817 if (vm_page_tryxbusy(p) == 0) {
1818 next = vm_object_collapse_scan_wait(object, p);
1822 KASSERT(object->backing_object == backing_object,
1823 ("vm_object_collapse_scan: backing object mismatch %p != %p",
1824 object->backing_object, backing_object));
1825 KASSERT(p->object == backing_object,
1826 ("vm_object_collapse_scan: object mismatch %p != %p",
1827 p->object, backing_object));
1829 if (p->pindex < backing_offset_index ||
1830 new_pindex >= object->size) {
1831 vm_pager_freespace(backing_object, p->pindex, 1);
1833 KASSERT(!pmap_page_is_mapped(p),
1834 ("freeing mapped page %p", p));
1835 if (vm_page_remove(p))
1840 if (!vm_page_all_valid(p)) {
1841 KASSERT(!pmap_page_is_mapped(p),
1842 ("freeing mapped page %p", p));
1843 if (vm_page_remove(p))
1848 pp = vm_page_lookup(object, new_pindex);
1849 if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1852 * The page in the parent is busy and possibly not
1853 * (yet) valid. Until its state is finalized by the
1854 * busy bit owner, we can't tell whether it shadows the
1857 next = vm_object_collapse_scan_wait(object, pp);
1861 if (pp != NULL && vm_page_none_valid(pp)) {
1863 * The page was invalid in the parent. Likely placed
1864 * there by an incomplete fault. Just remove and
1865 * ignore. p can replace it.
1867 if (vm_page_remove(pp))
1872 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1875 * The page already exists in the parent OR swap exists
1876 * for this location in the parent. Leave the parent's
1877 * page alone. Destroy the original page from the
1880 vm_pager_freespace(backing_object, p->pindex, 1);
1881 KASSERT(!pmap_page_is_mapped(p),
1882 ("freeing mapped page %p", p));
1883 if (vm_page_remove(p))
1886 vm_page_xunbusy(pp);
1891 * Page does not exist in parent, rename the page from the
1892 * backing object to the main object.
1894 * If the page was mapped to a process, it can remain mapped
1895 * through the rename. vm_page_rename() will dirty the page.
1897 if (vm_page_rename(p, object, new_pindex)) {
1899 next = vm_object_collapse_scan_wait(object, NULL);
1903 /* Use the old pindex to free the right page. */
1904 vm_pager_freespace(backing_object, new_pindex +
1905 backing_offset_index, 1);
1907 #if VM_NRESERVLEVEL > 0
1909 * Rename the reservation.
1911 vm_reserv_rename(p, object, backing_object,
1912 backing_offset_index);
1920 * vm_object_collapse:
1922 * Collapse an object with the object backing it.
1923 * Pages in the backing object are moved into the
1924 * parent, and the backing object is deallocated.
1927 vm_object_collapse(vm_object_t object)
1929 vm_object_t backing_object, new_backing_object;
1931 VM_OBJECT_ASSERT_WLOCKED(object);
1934 KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1935 ("collapsing invalid object"));
1938 * Wait for the backing_object to finish any pending
1939 * collapse so that the caller sees the shortest possible
1942 backing_object = vm_object_backing_collapse_wait(object);
1943 if (backing_object == NULL)
1946 KASSERT(object->ref_count > 0 &&
1947 object->ref_count > atomic_load_int(&object->shadow_count),
1948 ("collapse with invalid ref %d or shadow %d count.",
1949 object->ref_count, atomic_load_int(&object->shadow_count)));
1950 KASSERT((backing_object->flags &
1951 (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1952 ("vm_object_collapse: Backing object already collapsing."));
1953 KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1954 ("vm_object_collapse: object is already collapsing."));
1957 * We know that we can either collapse the backing object if
1958 * the parent is the only reference to it, or (perhaps) have
1959 * the parent bypass the object if the parent happens to shadow
1960 * all the resident pages in the entire backing object.
1962 if (backing_object->ref_count == 1) {
1963 KASSERT(atomic_load_int(&backing_object->shadow_count)
1965 ("vm_object_collapse: shadow_count: %d",
1966 atomic_load_int(&backing_object->shadow_count)));
1967 vm_object_pip_add(object, 1);
1968 vm_object_set_flag(object, OBJ_COLLAPSING);
1969 vm_object_pip_add(backing_object, 1);
1970 vm_object_set_flag(backing_object, OBJ_DEAD);
1973 * If there is exactly one reference to the backing
1974 * object, we can collapse it into the parent.
1976 vm_object_collapse_scan(object);
1978 #if VM_NRESERVLEVEL > 0
1980 * Break any reservations from backing_object.
1982 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1983 vm_reserv_break_all(backing_object);
1987 * Move the pager from backing_object to object.
1989 if ((backing_object->flags & OBJ_SWAP) != 0) {
1991 * swap_pager_copy() can sleep, in which case
1992 * the backing_object's and object's locks are
1993 * released and reacquired.
1994 * Since swap_pager_copy() is being asked to
1995 * destroy backing_object, it will change the
1996 * type to OBJT_DEFAULT.
2001 OFF_TO_IDX(object->backing_object_offset), TRUE);
2005 * Object now shadows whatever backing_object did.
2007 vm_object_clear_flag(object, OBJ_COLLAPSING);
2008 vm_object_backing_transfer(object, backing_object);
2009 object->backing_object_offset +=
2010 backing_object->backing_object_offset;
2011 VM_OBJECT_WUNLOCK(object);
2012 vm_object_pip_wakeup(object);
2015 * Discard backing_object.
2017 * Since the backing object has no pages, no pager left,
2018 * and no object references within it, all that is
2019 * necessary is to dispose of it.
2021 KASSERT(backing_object->ref_count == 1, (
2022 "backing_object %p was somehow re-referenced during collapse!",
2024 vm_object_pip_wakeup(backing_object);
2025 (void)refcount_release(&backing_object->ref_count);
2026 vm_object_terminate(backing_object);
2027 counter_u64_add(object_collapses, 1);
2028 VM_OBJECT_WLOCK(object);
2031 * If we do not entirely shadow the backing object,
2032 * there is nothing we can do so we give up.
2034 * The object lock and backing_object lock must not
2035 * be dropped during this sequence.
2037 if (!vm_object_scan_all_shadowed(object)) {
2038 VM_OBJECT_WUNLOCK(backing_object);
2043 * Make the parent shadow the next object in the
2044 * chain. Deallocating backing_object will not remove
2045 * it, since its reference count is at least 2.
2047 vm_object_backing_remove_locked(object);
2048 new_backing_object = backing_object->backing_object;
2049 if (new_backing_object != NULL) {
2050 vm_object_backing_insert_ref(object,
2051 new_backing_object);
2052 object->backing_object_offset +=
2053 backing_object->backing_object_offset;
2057 * Drop the reference count on backing_object. Since
2058 * its ref_count was at least 2, it will not vanish.
2060 (void)refcount_release(&backing_object->ref_count);
2061 KASSERT(backing_object->ref_count >= 1, (
2062 "backing_object %p was somehow dereferenced during collapse!",
2064 VM_OBJECT_WUNLOCK(backing_object);
2065 counter_u64_add(object_bypasses, 1);
2069 * Try again with this object's new backing object.
2075 * vm_object_page_remove:
2077 * For the given object, either frees or invalidates each of the
2078 * specified pages. In general, a page is freed. However, if a page is
2079 * wired for any reason other than the existence of a managed, wired
2080 * mapping, then it may be invalidated but not removed from the object.
2081 * Pages are specified by the given range ["start", "end") and the option
2082 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
2083 * extends from "start" to the end of the object. If the option
2084 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2085 * specified range are affected. If the option OBJPR_NOTMAPPED is
2086 * specified, then the pages within the specified range must have no
2087 * mappings. Otherwise, if this option is not specified, any mappings to
2088 * the specified pages are removed before the pages are freed or
2091 * In general, this operation should only be performed on objects that
2092 * contain managed pages. There are, however, two exceptions. First, it
2093 * is performed on the kernel and kmem objects by vm_map_entry_delete().
2094 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2095 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
2096 * not be specified and the option OBJPR_NOTMAPPED must be specified.
2098 * The object must be locked.
2101 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2106 VM_OBJECT_ASSERT_WLOCKED(object);
2107 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2108 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2109 ("vm_object_page_remove: illegal options for object %p", object));
2110 if (object->resident_page_count == 0)
2112 vm_object_pip_add(object, 1);
2114 p = vm_page_find_least(object, start);
2117 * Here, the variable "p" is either (1) the page with the least pindex
2118 * greater than or equal to the parameter "start" or (2) NULL.
2120 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2121 next = TAILQ_NEXT(p, listq);
2124 * Skip invalid pages if asked to do so. Try to avoid acquiring
2125 * the busy lock, as some consumers rely on this to avoid
2128 * A thread may concurrently transition the page from invalid to
2129 * valid using only the busy lock, so the result of this check
2130 * is immediately stale. It is up to consumers to handle this,
2131 * for instance by ensuring that all invalid->valid transitions
2132 * happen with a mutex held, as may be possible for a
2135 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2139 * If the page is wired for any reason besides the existence
2140 * of managed, wired mappings, then it cannot be freed. For
2141 * example, fictitious pages, which represent device memory,
2142 * are inherently wired and cannot be freed. They can,
2143 * however, be invalidated if the option OBJPR_CLEANONLY is
2146 if (vm_page_tryxbusy(p) == 0) {
2147 if (vm_page_busy_sleep(p, "vmopar", 0))
2148 VM_OBJECT_WLOCK(object);
2151 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2155 if (vm_page_wired(p)) {
2157 if ((options & OBJPR_NOTMAPPED) == 0 &&
2158 object->ref_count != 0)
2160 if ((options & OBJPR_CLEANONLY) == 0) {
2167 KASSERT((p->flags & PG_FICTITIOUS) == 0,
2168 ("vm_object_page_remove: page %p is fictitious", p));
2169 if ((options & OBJPR_CLEANONLY) != 0 &&
2170 !vm_page_none_valid(p)) {
2171 if ((options & OBJPR_NOTMAPPED) == 0 &&
2172 object->ref_count != 0 &&
2173 !vm_page_try_remove_write(p))
2175 if (p->dirty != 0) {
2180 if ((options & OBJPR_NOTMAPPED) == 0 &&
2181 object->ref_count != 0 && !vm_page_try_remove_all(p))
2185 vm_object_pip_wakeup(object);
2187 vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2192 * vm_object_page_noreuse:
2194 * For the given object, attempt to move the specified pages to
2195 * the head of the inactive queue. This bypasses regular LRU
2196 * operation and allows the pages to be reused quickly under memory
2197 * pressure. If a page is wired for any reason, then it will not
2198 * be queued. Pages are specified by the range ["start", "end").
2199 * As a special case, if "end" is zero, then the range extends from
2200 * "start" to the end of the object.
2202 * This operation should only be performed on objects that
2203 * contain non-fictitious, managed pages.
2205 * The object must be locked.
2208 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2212 VM_OBJECT_ASSERT_LOCKED(object);
2213 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2214 ("vm_object_page_noreuse: illegal object %p", object));
2215 if (object->resident_page_count == 0)
2217 p = vm_page_find_least(object, start);
2220 * Here, the variable "p" is either (1) the page with the least pindex
2221 * greater than or equal to the parameter "start" or (2) NULL.
2223 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2224 next = TAILQ_NEXT(p, listq);
2225 vm_page_deactivate_noreuse(p);
2230 * Populate the specified range of the object with valid pages. Returns
2231 * TRUE if the range is successfully populated and FALSE otherwise.
2233 * Note: This function should be optimized to pass a larger array of
2234 * pages to vm_pager_get_pages() before it is applied to a non-
2235 * OBJT_DEVICE object.
2237 * The object must be locked.
2240 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2246 VM_OBJECT_ASSERT_WLOCKED(object);
2247 for (pindex = start; pindex < end; pindex++) {
2248 rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2249 if (rv != VM_PAGER_OK)
2253 * Keep "m" busy because a subsequent iteration may unlock
2257 if (pindex > start) {
2258 m = vm_page_lookup(object, start);
2259 while (m != NULL && m->pindex < pindex) {
2261 m = TAILQ_NEXT(m, listq);
2264 return (pindex == end);
2268 * Routine: vm_object_coalesce
2269 * Function: Coalesces two objects backing up adjoining
2270 * regions of memory into a single object.
2272 * returns TRUE if objects were combined.
2274 * NOTE: Only works at the moment if the second object is NULL -
2275 * if it's not, which object do we lock first?
2278 * prev_object First object to coalesce
2279 * prev_offset Offset into prev_object
2280 * prev_size Size of reference to prev_object
2281 * next_size Size of reference to the second object
2282 * reserved Indicator that extension region has
2283 * swap accounted for
2286 * The object must *not* be locked.
2289 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2290 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2292 vm_pindex_t next_pindex;
2294 if (prev_object == NULL)
2296 if ((prev_object->flags & OBJ_ANON) == 0)
2299 VM_OBJECT_WLOCK(prev_object);
2301 * Try to collapse the object first.
2303 vm_object_collapse(prev_object);
2306 * Can't coalesce if: . more than one reference . paged out . shadows
2307 * another object . has a copy elsewhere (any of which mean that the
2308 * pages not mapped to prev_entry may be in use anyway)
2310 if (prev_object->backing_object != NULL) {
2311 VM_OBJECT_WUNLOCK(prev_object);
2315 prev_size >>= PAGE_SHIFT;
2316 next_size >>= PAGE_SHIFT;
2317 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2319 if (prev_object->ref_count > 1 &&
2320 prev_object->size != next_pindex &&
2321 (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2322 VM_OBJECT_WUNLOCK(prev_object);
2327 * Account for the charge.
2329 if (prev_object->cred != NULL) {
2331 * If prev_object was charged, then this mapping,
2332 * although not charged now, may become writable
2333 * later. Non-NULL cred in the object would prevent
2334 * swap reservation during enabling of the write
2335 * access, so reserve swap now. Failed reservation
2336 * cause allocation of the separate object for the map
2337 * entry, and swap reservation for this entry is
2338 * managed in appropriate time.
2340 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2341 prev_object->cred)) {
2342 VM_OBJECT_WUNLOCK(prev_object);
2345 prev_object->charge += ptoa(next_size);
2349 * Remove any pages that may still be in the object from a previous
2352 if (next_pindex < prev_object->size) {
2353 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2356 if (prev_object->cred != NULL) {
2357 KASSERT(prev_object->charge >=
2358 ptoa(prev_object->size - next_pindex),
2359 ("object %p overcharged 1 %jx %jx", prev_object,
2360 (uintmax_t)next_pindex, (uintmax_t)next_size));
2361 prev_object->charge -= ptoa(prev_object->size -
2368 * Extend the object if necessary.
2370 if (next_pindex + next_size > prev_object->size)
2371 prev_object->size = next_pindex + next_size;
2373 VM_OBJECT_WUNLOCK(prev_object);
2378 vm_object_set_writeable_dirty_(vm_object_t object)
2380 atomic_add_int(&object->generation, 1);
2384 vm_object_mightbedirty_(vm_object_t object)
2386 return (object->generation != object->cleangeneration);
2392 * For each page offset within the specified range of the given object,
2393 * find the highest-level page in the shadow chain and unwire it. A page
2394 * must exist at every page offset, and the highest-level page must be
2398 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2401 vm_object_t tobject, t1object;
2403 vm_pindex_t end_pindex, pindex, tpindex;
2404 int depth, locked_depth;
2406 KASSERT((offset & PAGE_MASK) == 0,
2407 ("vm_object_unwire: offset is not page aligned"));
2408 KASSERT((length & PAGE_MASK) == 0,
2409 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2410 /* The wired count of a fictitious page never changes. */
2411 if ((object->flags & OBJ_FICTITIOUS) != 0)
2413 pindex = OFF_TO_IDX(offset);
2414 end_pindex = pindex + atop(length);
2417 VM_OBJECT_RLOCK(object);
2418 m = vm_page_find_least(object, pindex);
2419 while (pindex < end_pindex) {
2420 if (m == NULL || pindex < m->pindex) {
2422 * The first object in the shadow chain doesn't
2423 * contain a page at the current index. Therefore,
2424 * the page must exist in a backing object.
2431 OFF_TO_IDX(tobject->backing_object_offset);
2432 tobject = tobject->backing_object;
2433 KASSERT(tobject != NULL,
2434 ("vm_object_unwire: missing page"));
2435 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2438 if (depth == locked_depth) {
2440 VM_OBJECT_RLOCK(tobject);
2442 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2446 m = TAILQ_NEXT(m, listq);
2448 if (vm_page_trysbusy(tm) == 0) {
2449 for (tobject = object; locked_depth >= 1;
2451 t1object = tobject->backing_object;
2452 if (tm->object != tobject)
2453 VM_OBJECT_RUNLOCK(tobject);
2456 tobject = tm->object;
2457 if (!vm_page_busy_sleep(tm, "unwbo",
2458 VM_ALLOC_IGN_SBUSY))
2459 VM_OBJECT_RUNLOCK(tobject);
2462 vm_page_unwire(tm, queue);
2463 vm_page_sunbusy(tm);
2467 /* Release the accumulated object locks. */
2468 for (tobject = object; locked_depth >= 1; locked_depth--) {
2469 t1object = tobject->backing_object;
2470 VM_OBJECT_RUNLOCK(tobject);
2476 * Return the vnode for the given object, or NULL if none exists.
2477 * For tmpfs objects, the function may return NULL if there is
2478 * no vnode allocated at the time of the call.
2481 vm_object_vnode(vm_object_t object)
2485 VM_OBJECT_ASSERT_LOCKED(object);
2486 vm_pager_getvp(object, &vp, NULL);
2491 * Busy the vm object. This prevents new pages belonging to the object from
2492 * becoming busy. Existing pages persist as busy. Callers are responsible
2493 * for checking page state before proceeding.
2496 vm_object_busy(vm_object_t obj)
2499 VM_OBJECT_ASSERT_LOCKED(obj);
2501 blockcount_acquire(&obj->busy, 1);
2502 /* The fence is required to order loads of page busy. */
2503 atomic_thread_fence_acq_rel();
2507 vm_object_unbusy(vm_object_t obj)
2510 blockcount_release(&obj->busy, 1);
2514 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2517 VM_OBJECT_ASSERT_UNLOCKED(obj);
2519 (void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2523 * This function aims to determine if the object is mapped,
2524 * specifically, if it is referenced by a vm_map_entry. Because
2525 * objects occasionally acquire transient references that do not
2526 * represent a mapping, the method used here is inexact. However, it
2527 * has very low overhead and is good enough for the advisory
2528 * vm.vmtotal sysctl.
2531 vm_object_is_active(vm_object_t obj)
2534 return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2538 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2540 struct kinfo_vmobject *kvo;
2541 char *fullpath, *freepath;
2549 if (req->oldptr == NULL) {
2551 * If an old buffer has not been provided, generate an
2552 * estimate of the space needed for a subsequent call.
2554 mtx_lock(&vm_object_list_mtx);
2556 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2557 if (obj->type == OBJT_DEAD)
2561 mtx_unlock(&vm_object_list_mtx);
2562 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2566 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK);
2570 * VM objects are type stable and are never removed from the
2571 * list once added. This allows us to safely read obj->object_list
2572 * after reacquiring the VM object lock.
2574 mtx_lock(&vm_object_list_mtx);
2575 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2576 if (obj->type == OBJT_DEAD ||
2577 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2579 VM_OBJECT_RLOCK(obj);
2580 if (obj->type == OBJT_DEAD ||
2581 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2582 VM_OBJECT_RUNLOCK(obj);
2585 mtx_unlock(&vm_object_list_mtx);
2586 kvo->kvo_size = ptoa(obj->size);
2587 kvo->kvo_resident = obj->resident_page_count;
2588 kvo->kvo_ref_count = obj->ref_count;
2589 kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2590 kvo->kvo_memattr = obj->memattr;
2591 kvo->kvo_active = 0;
2592 kvo->kvo_inactive = 0;
2594 TAILQ_FOREACH(m, &obj->memq, listq) {
2596 * A page may belong to the object but be
2597 * dequeued and set to PQ_NONE while the
2598 * object lock is not held. This makes the
2599 * reads of m->queue below racy, and we do not
2600 * count pages set to PQ_NONE. However, this
2601 * sysctl is only meant to give an
2602 * approximation of the system anyway.
2604 if (m->a.queue == PQ_ACTIVE)
2606 else if (m->a.queue == PQ_INACTIVE)
2607 kvo->kvo_inactive++;
2611 kvo->kvo_vn_fileid = 0;
2612 kvo->kvo_vn_fsid = 0;
2613 kvo->kvo_vn_fsid_freebsd11 = 0;
2617 kvo->kvo_type = vm_object_kvme_type(obj, swap_only ? NULL : &vp);
2620 } else if ((obj->flags & OBJ_ANON) != 0) {
2621 MPASS(kvo->kvo_type == KVME_TYPE_DEFAULT ||
2622 kvo->kvo_type == KVME_TYPE_SWAP);
2623 kvo->kvo_me = (uintptr_t)obj;
2624 /* tmpfs objs are reported as vnodes */
2625 kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2626 sp = swap_pager_swapped_pages(obj);
2627 kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2629 VM_OBJECT_RUNLOCK(obj);
2631 vn_fullpath(vp, &fullpath, &freepath);
2632 vn_lock(vp, LK_SHARED | LK_RETRY);
2633 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2634 kvo->kvo_vn_fileid = va.va_fileid;
2635 kvo->kvo_vn_fsid = va.va_fsid;
2636 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2642 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2643 if (freepath != NULL)
2644 free(freepath, M_TEMP);
2646 /* Pack record size down */
2647 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2648 + strlen(kvo->kvo_path) + 1;
2649 kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2651 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2653 mtx_lock(&vm_object_list_mtx);
2657 mtx_unlock(&vm_object_list_mtx);
2663 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2665 return (vm_object_list_handler(req, false));
2668 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2669 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2670 "List of VM objects");
2673 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2675 return (vm_object_list_handler(req, true));
2679 * This sysctl returns list of the anonymous or swap objects. Intent
2680 * is to provide stripped optimized list useful to analyze swap use.
2681 * Since technically non-swap (default) objects participate in the
2682 * shadow chains, and are converted to swap type as needed by swap
2683 * pager, we must report them.
2685 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2686 CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2687 sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2688 "List of swap VM objects");
2690 #include "opt_ddb.h"
2692 #include <sys/kernel.h>
2694 #include <sys/cons.h>
2696 #include <ddb/ddb.h>
2699 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2702 vm_map_entry_t tmpe;
2709 VM_MAP_ENTRY_FOREACH(tmpe, map) {
2710 if (_vm_object_in_map(map, object, tmpe)) {
2714 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2715 tmpm = entry->object.sub_map;
2716 VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2717 if (_vm_object_in_map(tmpm, object, tmpe)) {
2721 } else if ((obj = entry->object.vm_object) != NULL) {
2722 for (; obj; obj = obj->backing_object)
2723 if (obj == object) {
2731 vm_object_in_map(vm_object_t object)
2735 /* sx_slock(&allproc_lock); */
2736 FOREACH_PROC_IN_SYSTEM(p) {
2737 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2739 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2740 /* sx_sunlock(&allproc_lock); */
2744 /* sx_sunlock(&allproc_lock); */
2745 if (_vm_object_in_map(kernel_map, object, 0))
2750 DB_SHOW_COMMAND(vmochk, vm_object_check)
2755 * make sure that internal objs are in a map somewhere
2756 * and none have zero ref counts.
2758 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2759 if ((object->flags & OBJ_ANON) != 0) {
2760 if (object->ref_count == 0) {
2761 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2762 (long)object->size);
2764 if (!vm_object_in_map(object)) {
2766 "vmochk: internal obj is not in a map: "
2767 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2768 object->ref_count, (u_long)object->size,
2769 (u_long)object->size,
2770 (void *)object->backing_object);
2779 * vm_object_print: [ debug ]
2781 DB_SHOW_COMMAND(object, vm_object_print_static)
2783 /* XXX convert args. */
2784 vm_object_t object = (vm_object_t)addr;
2785 boolean_t full = have_addr;
2789 /* XXX count is an (unused) arg. Avoid shadowing it. */
2790 #define count was_count
2798 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2799 object, (int)object->type, (uintmax_t)object->size,
2800 object->resident_page_count, object->ref_count, object->flags,
2801 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2802 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2803 atomic_load_int(&object->shadow_count),
2804 object->backing_object ? object->backing_object->ref_count : 0,
2805 object->backing_object, (uintmax_t)object->backing_object_offset);
2812 TAILQ_FOREACH(p, &object->memq, listq) {
2814 db_iprintf("memory:=");
2815 else if (count == 6) {
2823 db_printf("(off=0x%jx,page=0x%jx)",
2824 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2837 /* XXX need this non-static entry for calling from vm_map_print. */
2840 /* db_expr_t */ long addr,
2841 boolean_t have_addr,
2842 /* db_expr_t */ long count,
2845 vm_object_print_static(addr, have_addr, count, modif);
2848 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2853 vm_page_t m, prev_m;
2856 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2857 db_printf("new object: %p\n", (void *)object);
2864 TAILQ_FOREACH(m, &object->memq, listq) {
2865 if (m->pindex > 128)
2867 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2868 prev_m->pindex + 1 != m->pindex) {
2870 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2871 (long)fidx, rcount, (long)pa);
2878 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2883 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2884 (long)fidx, rcount, (long)pa);
2889 pa = VM_PAGE_TO_PHYS(m);
2893 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2894 (long)fidx, rcount, (long)pa);