2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94
35 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36 * All rights reserved.
38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
40 * Permission to use, copy, modify and distribute this software and
41 * its documentation is hereby granted, provided that both the copyright
42 * notice and this permission notice appear in all copies of the
43 * software, derivative works or modified versions, and any portions
44 * thereof, and that both notices appear in supporting documentation.
46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
50 * Carnegie Mellon requests users of this software to return to
52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
53 * School of Computer Science
54 * Carnegie Mellon University
55 * Pittsburgh PA 15213-3890
57 * any improvements or extensions that they make and grant Carnegie the
58 * rights to redistribute these changes.
62 * Virtual memory object module.
65 #include <sys/cdefs.h>
66 __FBSDID("$FreeBSD$");
70 #include <sys/param.h>
71 #include <sys/systm.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/mutex.h>
78 #include <sys/proc.h> /* for curproc, pageproc */
79 #include <sys/socket.h>
80 #include <sys/resourcevar.h>
81 #include <sys/rwlock.h>
83 #include <sys/vnode.h>
84 #include <sys/vmmeter.h>
88 #include <vm/vm_param.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_pageout.h>
94 #include <vm/vm_pager.h>
95 #include <vm/swap_pager.h>
96 #include <vm/vm_kern.h>
97 #include <vm/vm_extern.h>
98 #include <vm/vm_radix.h>
99 #include <vm/vm_reserv.h>
102 static int old_msync;
103 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
104 "Use old (insecure) msync behavior");
106 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
107 int pagerflags, int flags, boolean_t *clearobjflags,
109 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
110 boolean_t *clearobjflags);
111 static void vm_object_qcollapse(vm_object_t object);
112 static void vm_object_vndeallocate(vm_object_t object);
115 * Virtual memory objects maintain the actual data
116 * associated with allocated virtual memory. A given
117 * page of memory exists within exactly one object.
119 * An object is only deallocated when all "references"
120 * are given up. Only one "reference" to a given
121 * region of an object should be writeable.
123 * Associated with each object is a list of all resident
124 * memory pages belonging to that object; this list is
125 * maintained by the "vm_page" module, and locked by the object's
128 * Each object also records a "pager" routine which is
129 * used to retrieve (and store) pages to the proper backing
130 * storage. In addition, objects may be backed by other
131 * objects from which they were virtual-copied.
133 * The only items within the object structure which are
134 * modified after time of creation are:
135 * reference count locked by object's lock
136 * pager routine locked by object's lock
140 struct object_q vm_object_list;
141 struct mtx vm_object_list_mtx; /* lock for object list and count */
143 struct vm_object kernel_object_store;
144 struct vm_object kmem_object_store;
146 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
149 static long object_collapses;
150 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
151 &object_collapses, 0, "VM object collapses");
153 static long object_bypasses;
154 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
155 &object_bypasses, 0, "VM object bypasses");
157 static uma_zone_t obj_zone;
159 static int vm_object_zinit(void *mem, int size, int flags);
162 static void vm_object_zdtor(void *mem, int size, void *arg);
165 vm_object_zdtor(void *mem, int size, void *arg)
169 object = (vm_object_t)mem;
170 KASSERT(object->ref_count == 0,
171 ("object %p ref_count = %d", object, object->ref_count));
172 KASSERT(TAILQ_EMPTY(&object->memq),
173 ("object %p has resident pages in its memq", object));
174 KASSERT(vm_radix_is_empty(&object->rtree),
175 ("object %p has resident pages in its trie", object));
176 #if VM_NRESERVLEVEL > 0
177 KASSERT(LIST_EMPTY(&object->rvq),
178 ("object %p has reservations",
181 KASSERT(object->paging_in_progress == 0,
182 ("object %p paging_in_progress = %d",
183 object, object->paging_in_progress));
184 KASSERT(object->resident_page_count == 0,
185 ("object %p resident_page_count = %d",
186 object, object->resident_page_count));
187 KASSERT(object->shadow_count == 0,
188 ("object %p shadow_count = %d",
189 object, object->shadow_count));
190 KASSERT(object->type == OBJT_DEAD,
191 ("object %p has non-dead type %d",
192 object, object->type));
197 vm_object_zinit(void *mem, int size, int flags)
201 object = (vm_object_t)mem;
202 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
204 /* These are true for any object that has been freed */
205 object->type = OBJT_DEAD;
206 object->ref_count = 0;
207 object->rtree.rt_root = 0;
208 object->rtree.rt_flags = 0;
209 object->paging_in_progress = 0;
210 object->resident_page_count = 0;
211 object->shadow_count = 0;
213 mtx_lock(&vm_object_list_mtx);
214 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
215 mtx_unlock(&vm_object_list_mtx);
220 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
223 TAILQ_INIT(&object->memq);
224 LIST_INIT(&object->shadow_head);
229 panic("_vm_object_allocate: can't create OBJT_DEAD");
232 object->flags = OBJ_ONEMAPPING;
236 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
239 object->flags = OBJ_FICTITIOUS;
242 object->flags = OBJ_UNMANAGED;
248 panic("_vm_object_allocate: type %d is undefined", type);
251 object->generation = 1;
252 object->ref_count = 1;
253 object->memattr = VM_MEMATTR_DEFAULT;
256 object->handle = NULL;
257 object->backing_object = NULL;
258 object->backing_object_offset = (vm_ooffset_t) 0;
259 #if VM_NRESERVLEVEL > 0
260 LIST_INIT(&object->rvq);
262 umtx_shm_object_init(object);
268 * Initialize the VM objects module.
273 TAILQ_INIT(&vm_object_list);
274 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
276 rw_init(&kernel_object->lock, "kernel vm object");
277 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
279 #if VM_NRESERVLEVEL > 0
280 kernel_object->flags |= OBJ_COLORED;
281 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
284 rw_init(&kmem_object->lock, "kmem vm object");
285 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
287 #if VM_NRESERVLEVEL > 0
288 kmem_object->flags |= OBJ_COLORED;
289 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
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 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
303 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
309 vm_object_clear_flag(vm_object_t object, u_short bits)
312 VM_OBJECT_ASSERT_WLOCKED(object);
313 object->flags &= ~bits;
317 * Sets the default memory attribute for the specified object. Pages
318 * that are allocated to this object are by default assigned this memory
321 * Presently, this function must be called before any pages are allocated
322 * to the object. In the future, this requirement may be relaxed for
323 * "default" and "swap" objects.
326 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
329 VM_OBJECT_ASSERT_WLOCKED(object);
330 switch (object->type) {
338 if (!TAILQ_EMPTY(&object->memq))
339 return (KERN_FAILURE);
342 return (KERN_INVALID_ARGUMENT);
344 panic("vm_object_set_memattr: object %p is of undefined type",
347 object->memattr = memattr;
348 return (KERN_SUCCESS);
352 vm_object_pip_add(vm_object_t object, short i)
355 VM_OBJECT_ASSERT_WLOCKED(object);
356 object->paging_in_progress += i;
360 vm_object_pip_subtract(vm_object_t object, short i)
363 VM_OBJECT_ASSERT_WLOCKED(object);
364 object->paging_in_progress -= i;
368 vm_object_pip_wakeup(vm_object_t object)
371 VM_OBJECT_ASSERT_WLOCKED(object);
372 object->paging_in_progress--;
373 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
374 vm_object_clear_flag(object, OBJ_PIPWNT);
380 vm_object_pip_wakeupn(vm_object_t object, short i)
383 VM_OBJECT_ASSERT_WLOCKED(object);
385 object->paging_in_progress -= i;
386 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
387 vm_object_clear_flag(object, OBJ_PIPWNT);
393 vm_object_pip_wait(vm_object_t object, char *waitid)
396 VM_OBJECT_ASSERT_WLOCKED(object);
397 while (object->paging_in_progress) {
398 object->flags |= OBJ_PIPWNT;
399 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
404 * vm_object_allocate:
406 * Returns a new object with the given size.
409 vm_object_allocate(objtype_t type, vm_pindex_t size)
413 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
414 _vm_object_allocate(type, size, object);
420 * vm_object_reference:
422 * Gets another reference to the given object. Note: OBJ_DEAD
423 * objects can be referenced during final cleaning.
426 vm_object_reference(vm_object_t object)
430 VM_OBJECT_WLOCK(object);
431 vm_object_reference_locked(object);
432 VM_OBJECT_WUNLOCK(object);
436 * vm_object_reference_locked:
438 * Gets another reference to the given object.
440 * The object must be locked.
443 vm_object_reference_locked(vm_object_t object)
447 VM_OBJECT_ASSERT_WLOCKED(object);
449 if (object->type == OBJT_VNODE) {
456 * Handle deallocating an object of type OBJT_VNODE.
459 vm_object_vndeallocate(vm_object_t object)
461 struct vnode *vp = (struct vnode *) object->handle;
463 VM_OBJECT_ASSERT_WLOCKED(object);
464 KASSERT(object->type == OBJT_VNODE,
465 ("vm_object_vndeallocate: not a vnode object"));
466 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
468 if (object->ref_count == 0) {
469 vn_printf(vp, "vm_object_vndeallocate ");
470 panic("vm_object_vndeallocate: bad object reference count");
474 if (!umtx_shm_vnobj_persistent && object->ref_count == 1)
475 umtx_shm_object_terminated(object);
478 * The test for text of vp vnode does not need a bypass to
479 * reach right VV_TEXT there, since it is obtained from
482 if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) {
484 VM_OBJECT_WUNLOCK(object);
485 /* vrele may need the vnode lock. */
489 VM_OBJECT_WUNLOCK(object);
490 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
492 VM_OBJECT_WLOCK(object);
494 if (object->type == OBJT_DEAD) {
495 VM_OBJECT_WUNLOCK(object);
498 if (object->ref_count == 0)
500 VM_OBJECT_WUNLOCK(object);
507 * vm_object_deallocate:
509 * Release a reference to the specified object,
510 * gained either through a vm_object_allocate
511 * or a vm_object_reference call. When all references
512 * are gone, storage associated with this object
513 * may be relinquished.
515 * No object may be locked.
518 vm_object_deallocate(vm_object_t object)
523 while (object != NULL) {
524 VM_OBJECT_WLOCK(object);
525 if (object->type == OBJT_VNODE) {
526 vm_object_vndeallocate(object);
530 KASSERT(object->ref_count != 0,
531 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
534 * If the reference count goes to 0 we start calling
535 * vm_object_terminate() on the object chain.
536 * A ref count of 1 may be a special case depending on the
537 * shadow count being 0 or 1.
540 if (object->ref_count > 1) {
541 VM_OBJECT_WUNLOCK(object);
543 } else if (object->ref_count == 1) {
544 if (object->type == OBJT_SWAP &&
545 (object->flags & OBJ_TMPFS) != 0) {
546 vp = object->un_pager.swp.swp_tmpfs;
548 VM_OBJECT_WUNLOCK(object);
549 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
550 VM_OBJECT_WLOCK(object);
551 if (object->type == OBJT_DEAD ||
552 object->ref_count != 1) {
553 VM_OBJECT_WUNLOCK(object);
558 if ((object->flags & OBJ_TMPFS) != 0)
563 if (object->shadow_count == 0 &&
564 object->handle == NULL &&
565 (object->type == OBJT_DEFAULT ||
566 (object->type == OBJT_SWAP &&
567 (object->flags & OBJ_TMPFS_NODE) == 0))) {
568 vm_object_set_flag(object, OBJ_ONEMAPPING);
569 } else if ((object->shadow_count == 1) &&
570 (object->handle == NULL) &&
571 (object->type == OBJT_DEFAULT ||
572 object->type == OBJT_SWAP)) {
575 robject = LIST_FIRST(&object->shadow_head);
576 KASSERT(robject != NULL,
577 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
579 object->shadow_count));
580 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0,
581 ("shadowed tmpfs v_object %p", object));
582 if (!VM_OBJECT_TRYWLOCK(robject)) {
584 * Avoid a potential deadlock.
587 VM_OBJECT_WUNLOCK(object);
589 * More likely than not the thread
590 * holding robject's lock has lower
591 * priority than the current thread.
592 * Let the lower priority thread run.
598 * Collapse object into its shadow unless its
599 * shadow is dead. In that case, object will
600 * be deallocated by the thread that is
601 * deallocating its shadow.
603 if ((robject->flags & OBJ_DEAD) == 0 &&
604 (robject->handle == NULL) &&
605 (robject->type == OBJT_DEFAULT ||
606 robject->type == OBJT_SWAP)) {
608 robject->ref_count++;
610 if (robject->paging_in_progress) {
611 VM_OBJECT_WUNLOCK(object);
612 vm_object_pip_wait(robject,
614 temp = robject->backing_object;
615 if (object == temp) {
616 VM_OBJECT_WLOCK(object);
619 } else if (object->paging_in_progress) {
620 VM_OBJECT_WUNLOCK(robject);
621 object->flags |= OBJ_PIPWNT;
622 VM_OBJECT_SLEEP(object, object,
623 PDROP | PVM, "objde2", 0);
624 VM_OBJECT_WLOCK(robject);
625 temp = robject->backing_object;
626 if (object == temp) {
627 VM_OBJECT_WLOCK(object);
631 VM_OBJECT_WUNLOCK(object);
633 if (robject->ref_count == 1) {
634 robject->ref_count--;
639 vm_object_collapse(object);
640 VM_OBJECT_WUNLOCK(object);
643 VM_OBJECT_WUNLOCK(robject);
645 VM_OBJECT_WUNLOCK(object);
649 umtx_shm_object_terminated(object);
650 temp = object->backing_object;
652 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
653 ("shadowed tmpfs v_object 2 %p", object));
654 VM_OBJECT_WLOCK(temp);
655 LIST_REMOVE(object, shadow_list);
656 temp->shadow_count--;
657 VM_OBJECT_WUNLOCK(temp);
658 object->backing_object = NULL;
661 * Don't double-terminate, we could be in a termination
662 * recursion due to the terminate having to sync data
665 if ((object->flags & OBJ_DEAD) == 0)
666 vm_object_terminate(object);
668 VM_OBJECT_WUNLOCK(object);
674 * vm_object_destroy removes the object from the global object list
675 * and frees the space for the object.
678 vm_object_destroy(vm_object_t object)
682 * Release the allocation charge.
684 if (object->cred != NULL) {
685 swap_release_by_cred(object->charge, object->cred);
687 crfree(object->cred);
692 * Free the space for the object.
694 uma_zfree(obj_zone, object);
698 * vm_object_terminate actually destroys the specified object, freeing
699 * up all previously used resources.
701 * The object must be locked.
702 * This routine may block.
705 vm_object_terminate(vm_object_t object)
709 VM_OBJECT_ASSERT_WLOCKED(object);
712 * Make sure no one uses us.
714 vm_object_set_flag(object, OBJ_DEAD);
717 * wait for the pageout daemon to be done with the object
719 vm_object_pip_wait(object, "objtrm");
721 KASSERT(!object->paging_in_progress,
722 ("vm_object_terminate: pageout in progress"));
725 * Clean and free the pages, as appropriate. All references to the
726 * object are gone, so we don't need to lock it.
728 if (object->type == OBJT_VNODE) {
729 struct vnode *vp = (struct vnode *)object->handle;
732 * Clean pages and flush buffers.
734 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
735 VM_OBJECT_WUNLOCK(object);
737 vinvalbuf(vp, V_SAVE, 0, 0);
739 BO_LOCK(&vp->v_bufobj);
740 vp->v_bufobj.bo_flag |= BO_DEAD;
741 BO_UNLOCK(&vp->v_bufobj);
743 VM_OBJECT_WLOCK(object);
746 KASSERT(object->ref_count == 0,
747 ("vm_object_terminate: object with references, ref_count=%d",
751 * Free any remaining pageable pages. This also removes them from the
752 * paging queues. However, don't free wired pages, just remove them
753 * from the object. Rather than incrementally removing each page from
754 * the object, the page and object are reset to any empty state.
756 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
757 vm_page_assert_unbusied(p);
760 * Optimize the page's removal from the object by resetting
761 * its "object" field. Specifically, if the page is not
762 * wired, then the effect of this assignment is that
763 * vm_page_free()'s call to vm_page_remove() will return
764 * immediately without modifying the page or the object.
767 if (p->wire_count == 0) {
769 PCPU_INC(cnt.v_pfree);
774 * If the object contained any pages, then reset it to an empty state.
775 * None of the object's fields, including "resident_page_count", were
776 * modified by the preceding loop.
778 if (object->resident_page_count != 0) {
779 vm_radix_reclaim_allnodes(&object->rtree);
780 TAILQ_INIT(&object->memq);
781 object->resident_page_count = 0;
782 if (object->type == OBJT_VNODE)
783 vdrop(object->handle);
786 #if VM_NRESERVLEVEL > 0
787 if (__predict_false(!LIST_EMPTY(&object->rvq)))
788 vm_reserv_break_all(object);
791 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
792 object->type == OBJT_SWAP,
793 ("%s: non-swap obj %p has cred", __func__, object));
796 * Let the pager know object is dead.
798 vm_pager_deallocate(object);
799 VM_OBJECT_WUNLOCK(object);
801 vm_object_destroy(object);
805 * Make the page read-only so that we can clear the object flags. However, if
806 * this is a nosync mmap then the object is likely to stay dirty so do not
807 * mess with the page and do not clear the object flags. Returns TRUE if the
808 * page should be flushed, and FALSE otherwise.
811 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
815 * If we have been asked to skip nosync pages and this is a
816 * nosync page, skip it. Note that the object flags were not
817 * cleared in this case so we do not have to set them.
819 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
820 *clearobjflags = FALSE;
823 pmap_remove_write(p);
824 return (p->dirty != 0);
829 * vm_object_page_clean
831 * Clean all dirty pages in the specified range of object. Leaves page
832 * on whatever queue it is currently on. If NOSYNC is set then do not
833 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
834 * leaving the object dirty.
836 * When stuffing pages asynchronously, allow clustering. XXX we need a
837 * synchronous clustering mode implementation.
839 * Odd semantics: if start == end, we clean everything.
841 * The object must be locked.
843 * Returns FALSE if some page from the range was not written, as
844 * reported by the pager, and TRUE otherwise.
847 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
851 vm_pindex_t pi, tend, tstart;
852 int curgeneration, n, pagerflags;
853 boolean_t clearobjflags, eio, res;
855 VM_OBJECT_ASSERT_WLOCKED(object);
858 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
859 * objects. The check below prevents the function from
860 * operating on non-vnode objects.
862 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
863 object->resident_page_count == 0)
866 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
867 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
868 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
870 tstart = OFF_TO_IDX(start);
871 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
872 clearobjflags = tstart == 0 && tend >= object->size;
876 curgeneration = object->generation;
878 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
882 np = TAILQ_NEXT(p, listq);
885 if (vm_page_sleep_if_busy(p, "vpcwai")) {
886 if (object->generation != curgeneration) {
887 if ((flags & OBJPC_SYNC) != 0)
890 clearobjflags = FALSE;
892 np = vm_page_find_least(object, pi);
895 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
898 n = vm_object_page_collect_flush(object, p, pagerflags,
899 flags, &clearobjflags, &eio);
902 clearobjflags = FALSE;
904 if (object->generation != curgeneration) {
905 if ((flags & OBJPC_SYNC) != 0)
908 clearobjflags = FALSE;
912 * If the VOP_PUTPAGES() did a truncated write, so
913 * that even the first page of the run is not fully
914 * written, vm_pageout_flush() returns 0 as the run
915 * length. Since the condition that caused truncated
916 * write may be permanent, e.g. exhausted free space,
917 * accepting n == 0 would cause an infinite loop.
919 * Forwarding the iterator leaves the unwritten page
920 * behind, but there is not much we can do there if
921 * filesystem refuses to write it.
925 clearobjflags = FALSE;
927 np = vm_page_find_least(object, pi + n);
930 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
934 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
939 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
940 int flags, boolean_t *clearobjflags, boolean_t *eio)
942 vm_page_t ma[vm_pageout_page_count], p_first, tp;
943 int count, i, mreq, runlen;
945 vm_page_lock_assert(p, MA_NOTOWNED);
946 VM_OBJECT_ASSERT_WLOCKED(object);
951 for (tp = p; count < vm_pageout_page_count; count++) {
952 tp = vm_page_next(tp);
953 if (tp == NULL || vm_page_busied(tp))
955 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
959 for (p_first = p; count < vm_pageout_page_count; count++) {
960 tp = vm_page_prev(p_first);
961 if (tp == NULL || vm_page_busied(tp))
963 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
969 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
972 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
977 * Note that there is absolutely no sense in writing out
978 * anonymous objects, so we track down the vnode object
980 * We invalidate (remove) all pages from the address space
981 * for semantic correctness.
983 * If the backing object is a device object with unmanaged pages, then any
984 * mappings to the specified range of pages must be removed before this
985 * function is called.
987 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
988 * may start out with a NULL object.
991 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
992 boolean_t syncio, boolean_t invalidate)
994 vm_object_t backing_object;
997 int error, flags, fsync_after;
1004 VM_OBJECT_WLOCK(object);
1005 while ((backing_object = object->backing_object) != NULL) {
1006 VM_OBJECT_WLOCK(backing_object);
1007 offset += object->backing_object_offset;
1008 VM_OBJECT_WUNLOCK(object);
1009 object = backing_object;
1010 if (object->size < OFF_TO_IDX(offset + size))
1011 size = IDX_TO_OFF(object->size) - offset;
1014 * Flush pages if writing is allowed, invalidate them
1015 * if invalidation requested. Pages undergoing I/O
1016 * will be ignored by vm_object_page_remove().
1018 * We cannot lock the vnode and then wait for paging
1019 * to complete without deadlocking against vm_fault.
1020 * Instead we simply call vm_object_page_remove() and
1021 * allow it to block internally on a page-by-page
1022 * basis when it encounters pages undergoing async
1025 if (object->type == OBJT_VNODE &&
1026 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1027 vp = object->handle;
1028 VM_OBJECT_WUNLOCK(object);
1029 (void) vn_start_write(vp, &mp, V_WAIT);
1030 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1031 if (syncio && !invalidate && offset == 0 &&
1032 OFF_TO_IDX(size) == object->size) {
1034 * If syncing the whole mapping of the file,
1035 * it is faster to schedule all the writes in
1036 * async mode, also allowing the clustering,
1037 * and then wait for i/o to complete.
1042 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1043 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1044 fsync_after = FALSE;
1046 VM_OBJECT_WLOCK(object);
1047 res = vm_object_page_clean(object, offset, offset + size,
1049 VM_OBJECT_WUNLOCK(object);
1051 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1053 vn_finished_write(mp);
1056 VM_OBJECT_WLOCK(object);
1058 if ((object->type == OBJT_VNODE ||
1059 object->type == OBJT_DEVICE) && invalidate) {
1060 if (object->type == OBJT_DEVICE)
1062 * The option OBJPR_NOTMAPPED must be passed here
1063 * because vm_object_page_remove() cannot remove
1064 * unmanaged mappings.
1066 flags = OBJPR_NOTMAPPED;
1070 flags = OBJPR_CLEANONLY;
1071 vm_object_page_remove(object, OFF_TO_IDX(offset),
1072 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1074 VM_OBJECT_WUNLOCK(object);
1079 * vm_object_madvise:
1081 * Implements the madvise function at the object/page level.
1083 * MADV_WILLNEED (any object)
1085 * Activate the specified pages if they are resident.
1087 * MADV_DONTNEED (any object)
1089 * Deactivate the specified pages if they are resident.
1091 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1092 * OBJ_ONEMAPPING only)
1094 * Deactivate and clean the specified pages if they are
1095 * resident. This permits the process to reuse the pages
1096 * without faulting or the kernel to reclaim the pages
1100 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1103 vm_pindex_t tpindex;
1104 vm_object_t backing_object, tobject;
1109 VM_OBJECT_WLOCK(object);
1111 * Locate and adjust resident pages
1113 for (; pindex < end; pindex += 1) {
1119 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1120 * and those pages must be OBJ_ONEMAPPING.
1122 if (advise == MADV_FREE) {
1123 if ((tobject->type != OBJT_DEFAULT &&
1124 tobject->type != OBJT_SWAP) ||
1125 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1126 goto unlock_tobject;
1128 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1129 goto unlock_tobject;
1130 m = vm_page_lookup(tobject, tpindex);
1133 * There may be swap even if there is no backing page
1135 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1136 swap_pager_freespace(tobject, tpindex, 1);
1140 backing_object = tobject->backing_object;
1141 if (backing_object == NULL)
1142 goto unlock_tobject;
1143 VM_OBJECT_WLOCK(backing_object);
1144 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1145 if (tobject != object)
1146 VM_OBJECT_WUNLOCK(tobject);
1147 tobject = backing_object;
1149 } else if (m->valid != VM_PAGE_BITS_ALL)
1150 goto unlock_tobject;
1152 * If the page is not in a normal state, skip it.
1155 if (m->hold_count != 0 || m->wire_count != 0) {
1157 goto unlock_tobject;
1159 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1160 ("vm_object_madvise: page %p is fictitious", m));
1161 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1162 ("vm_object_madvise: page %p is not managed", m));
1163 if (vm_page_busied(m)) {
1164 if (advise == MADV_WILLNEED) {
1166 * Reference the page before unlocking and
1167 * sleeping so that the page daemon is less
1168 * likely to reclaim it.
1170 vm_page_aflag_set(m, PGA_REFERENCED);
1172 if (object != tobject)
1173 VM_OBJECT_WUNLOCK(object);
1174 VM_OBJECT_WUNLOCK(tobject);
1175 vm_page_busy_sleep(m, "madvpo", false);
1176 VM_OBJECT_WLOCK(object);
1179 if (advise == MADV_WILLNEED) {
1180 vm_page_activate(m);
1182 vm_page_advise(m, advise);
1185 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1186 swap_pager_freespace(tobject, tpindex, 1);
1188 if (tobject != object)
1189 VM_OBJECT_WUNLOCK(tobject);
1191 VM_OBJECT_WUNLOCK(object);
1197 * Create a new object which is backed by the
1198 * specified existing object range. The source
1199 * object reference is deallocated.
1201 * The new object and offset into that object
1202 * are returned in the source parameters.
1206 vm_object_t *object, /* IN/OUT */
1207 vm_ooffset_t *offset, /* IN/OUT */
1216 * Don't create the new object if the old object isn't shared.
1218 if (source != NULL) {
1219 VM_OBJECT_WLOCK(source);
1220 if (source->ref_count == 1 &&
1221 source->handle == NULL &&
1222 (source->type == OBJT_DEFAULT ||
1223 source->type == OBJT_SWAP)) {
1224 VM_OBJECT_WUNLOCK(source);
1227 VM_OBJECT_WUNLOCK(source);
1231 * Allocate a new object with the given length.
1233 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1236 * The new object shadows the source object, adding a reference to it.
1237 * Our caller changes his reference to point to the new object,
1238 * removing a reference to the source object. Net result: no change
1239 * of reference count.
1241 * Try to optimize the result object's page color when shadowing
1242 * in order to maintain page coloring consistency in the combined
1245 result->backing_object = source;
1247 * Store the offset into the source object, and fix up the offset into
1250 result->backing_object_offset = *offset;
1251 if (source != NULL) {
1252 VM_OBJECT_WLOCK(source);
1253 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1254 source->shadow_count++;
1255 #if VM_NRESERVLEVEL > 0
1256 result->flags |= source->flags & OBJ_COLORED;
1257 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1258 ((1 << (VM_NFREEORDER - 1)) - 1);
1260 VM_OBJECT_WUNLOCK(source);
1265 * Return the new things
1274 * Split the pages in a map entry into a new object. This affords
1275 * easier removal of unused pages, and keeps object inheritance from
1276 * being a negative impact on memory usage.
1279 vm_object_split(vm_map_entry_t entry)
1281 vm_page_t m, m_next;
1282 vm_object_t orig_object, new_object, source;
1283 vm_pindex_t idx, offidxstart;
1286 orig_object = entry->object.vm_object;
1287 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1289 if (orig_object->ref_count <= 1)
1291 VM_OBJECT_WUNLOCK(orig_object);
1293 offidxstart = OFF_TO_IDX(entry->offset);
1294 size = atop(entry->end - entry->start);
1297 * If swap_pager_copy() is later called, it will convert new_object
1298 * into a swap object.
1300 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1303 * At this point, the new object is still private, so the order in
1304 * which the original and new objects are locked does not matter.
1306 VM_OBJECT_WLOCK(new_object);
1307 VM_OBJECT_WLOCK(orig_object);
1308 source = orig_object->backing_object;
1309 if (source != NULL) {
1310 VM_OBJECT_WLOCK(source);
1311 if ((source->flags & OBJ_DEAD) != 0) {
1312 VM_OBJECT_WUNLOCK(source);
1313 VM_OBJECT_WUNLOCK(orig_object);
1314 VM_OBJECT_WUNLOCK(new_object);
1315 vm_object_deallocate(new_object);
1316 VM_OBJECT_WLOCK(orig_object);
1319 LIST_INSERT_HEAD(&source->shadow_head,
1320 new_object, shadow_list);
1321 source->shadow_count++;
1322 vm_object_reference_locked(source); /* for new_object */
1323 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1324 VM_OBJECT_WUNLOCK(source);
1325 new_object->backing_object_offset =
1326 orig_object->backing_object_offset + entry->offset;
1327 new_object->backing_object = source;
1329 if (orig_object->cred != NULL) {
1330 new_object->cred = orig_object->cred;
1331 crhold(orig_object->cred);
1332 new_object->charge = ptoa(size);
1333 KASSERT(orig_object->charge >= ptoa(size),
1334 ("orig_object->charge < 0"));
1335 orig_object->charge -= ptoa(size);
1338 m = vm_page_find_least(orig_object, offidxstart);
1339 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1341 m_next = TAILQ_NEXT(m, listq);
1344 * We must wait for pending I/O to complete before we can
1347 * We do not have to VM_PROT_NONE the page as mappings should
1348 * not be changed by this operation.
1350 if (vm_page_busied(m)) {
1351 VM_OBJECT_WUNLOCK(new_object);
1353 VM_OBJECT_WUNLOCK(orig_object);
1354 vm_page_busy_sleep(m, "spltwt", false);
1355 VM_OBJECT_WLOCK(orig_object);
1356 VM_OBJECT_WLOCK(new_object);
1360 /* vm_page_rename() will handle dirty and cache. */
1361 if (vm_page_rename(m, new_object, idx)) {
1362 VM_OBJECT_WUNLOCK(new_object);
1363 VM_OBJECT_WUNLOCK(orig_object);
1365 VM_OBJECT_WLOCK(orig_object);
1366 VM_OBJECT_WLOCK(new_object);
1369 #if VM_NRESERVLEVEL > 0
1371 * If some of the reservation's allocated pages remain with
1372 * the original object, then transferring the reservation to
1373 * the new object is neither particularly beneficial nor
1374 * particularly harmful as compared to leaving the reservation
1375 * with the original object. If, however, all of the
1376 * reservation's allocated pages are transferred to the new
1377 * object, then transferring the reservation is typically
1378 * beneficial. Determining which of these two cases applies
1379 * would be more costly than unconditionally renaming the
1382 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1384 if (orig_object->type == OBJT_SWAP)
1387 if (orig_object->type == OBJT_SWAP) {
1389 * swap_pager_copy() can sleep, in which case the orig_object's
1390 * and new_object's locks are released and reacquired.
1392 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1393 TAILQ_FOREACH(m, &new_object->memq, listq)
1396 VM_OBJECT_WUNLOCK(orig_object);
1397 VM_OBJECT_WUNLOCK(new_object);
1398 entry->object.vm_object = new_object;
1399 entry->offset = 0LL;
1400 vm_object_deallocate(orig_object);
1401 VM_OBJECT_WLOCK(new_object);
1404 #define OBSC_COLLAPSE_NOWAIT 0x0002
1405 #define OBSC_COLLAPSE_WAIT 0x0004
1408 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next,
1411 vm_object_t backing_object;
1413 VM_OBJECT_ASSERT_WLOCKED(object);
1414 backing_object = object->backing_object;
1415 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1417 KASSERT(p == NULL || vm_page_busied(p), ("unbusy page %p", p));
1418 KASSERT(p == NULL || p->object == object || p->object == backing_object,
1419 ("invalid ownership %p %p %p", p, object, backing_object));
1420 if ((op & OBSC_COLLAPSE_NOWAIT) != 0)
1424 VM_OBJECT_WUNLOCK(object);
1425 VM_OBJECT_WUNLOCK(backing_object);
1429 vm_page_busy_sleep(p, "vmocol", false);
1430 VM_OBJECT_WLOCK(object);
1431 VM_OBJECT_WLOCK(backing_object);
1432 return (TAILQ_FIRST(&backing_object->memq));
1436 vm_object_scan_all_shadowed(vm_object_t object)
1438 vm_object_t backing_object;
1440 vm_pindex_t backing_offset_index, new_pindex;
1442 VM_OBJECT_ASSERT_WLOCKED(object);
1443 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1445 backing_object = object->backing_object;
1448 * Initial conditions:
1450 * We do not want to have to test for the existence of cache or swap
1451 * pages in the backing object. XXX but with the new swapper this
1452 * would be pretty easy to do.
1454 if (backing_object->type != OBJT_DEFAULT)
1457 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1459 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL;
1460 p = TAILQ_NEXT(p, listq)) {
1461 new_pindex = p->pindex - backing_offset_index;
1464 * Ignore pages outside the parent object's range and outside
1465 * the parent object's mapping of the backing object.
1467 if (p->pindex < backing_offset_index ||
1468 new_pindex >= object->size)
1472 * See if the parent has the page or if the parent's object
1473 * pager has the page. If the parent has the page but the page
1474 * is not valid, the parent's object pager must have the page.
1476 * If this fails, the parent does not completely shadow the
1477 * object and we might as well give up now.
1479 pp = vm_page_lookup(object, new_pindex);
1480 if ((pp == NULL || pp->valid == 0) &&
1481 !vm_pager_has_page(object, new_pindex, NULL, NULL))
1488 vm_object_collapse_scan(vm_object_t object, int op)
1490 vm_object_t backing_object;
1491 vm_page_t next, p, pp;
1492 vm_pindex_t backing_offset_index, new_pindex;
1494 VM_OBJECT_ASSERT_WLOCKED(object);
1495 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1497 backing_object = object->backing_object;
1498 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1501 * Initial conditions
1503 if ((op & OBSC_COLLAPSE_WAIT) != 0)
1504 vm_object_set_flag(backing_object, OBJ_DEAD);
1509 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1510 next = TAILQ_NEXT(p, listq);
1511 new_pindex = p->pindex - backing_offset_index;
1514 * Check for busy page
1516 if (vm_page_busied(p)) {
1517 next = vm_object_collapse_scan_wait(object, p, next, op);
1521 KASSERT(p->object == backing_object,
1522 ("vm_object_collapse_scan: object mismatch"));
1524 if (p->pindex < backing_offset_index ||
1525 new_pindex >= object->size) {
1526 if (backing_object->type == OBJT_SWAP)
1527 swap_pager_freespace(backing_object, p->pindex,
1531 * Page is out of the parent object's range, we can
1532 * simply destroy it.
1535 KASSERT(!pmap_page_is_mapped(p),
1536 ("freeing mapped page %p", p));
1537 if (p->wire_count == 0)
1545 pp = vm_page_lookup(object, new_pindex);
1546 if (pp != NULL && vm_page_busied(pp)) {
1548 * The page in the parent is busy and possibly not
1549 * (yet) valid. Until its state is finalized by the
1550 * busy bit owner, we can't tell whether it shadows the
1551 * original page. Therefore, we must either skip it
1552 * and the original (backing_object) page or wait for
1553 * its state to be finalized.
1555 * This is due to a race with vm_fault() where we must
1556 * unbusy the original (backing_obj) page before we can
1557 * (re)lock the parent. Hence we can get here.
1559 next = vm_object_collapse_scan_wait(object, pp, next,
1564 KASSERT(pp == NULL || pp->valid != 0,
1565 ("unbusy invalid page %p", pp));
1567 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1570 * The page already exists in the parent OR swap exists
1571 * for this location in the parent. Leave the parent's
1572 * page alone. Destroy the original page from the
1575 if (backing_object->type == OBJT_SWAP)
1576 swap_pager_freespace(backing_object, p->pindex,
1579 KASSERT(!pmap_page_is_mapped(p),
1580 ("freeing mapped page %p", p));
1581 if (p->wire_count == 0)
1590 * Page does not exist in parent, rename the page from the
1591 * backing object to the main object.
1593 * If the page was mapped to a process, it can remain mapped
1594 * through the rename. vm_page_rename() will handle dirty and
1597 if (vm_page_rename(p, object, new_pindex)) {
1598 next = vm_object_collapse_scan_wait(object, NULL, next,
1603 /* Use the old pindex to free the right page. */
1604 if (backing_object->type == OBJT_SWAP)
1605 swap_pager_freespace(backing_object,
1606 new_pindex + backing_offset_index, 1);
1608 #if VM_NRESERVLEVEL > 0
1610 * Rename the reservation.
1612 vm_reserv_rename(p, object, backing_object,
1613 backing_offset_index);
1621 * this version of collapse allows the operation to occur earlier and
1622 * when paging_in_progress is true for an object... This is not a complete
1623 * operation, but should plug 99.9% of the rest of the leaks.
1626 vm_object_qcollapse(vm_object_t object)
1628 vm_object_t backing_object = object->backing_object;
1630 VM_OBJECT_ASSERT_WLOCKED(object);
1631 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1633 if (backing_object->ref_count != 1)
1636 vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT);
1640 * vm_object_collapse:
1642 * Collapse an object with the object backing it.
1643 * Pages in the backing object are moved into the
1644 * parent, and the backing object is deallocated.
1647 vm_object_collapse(vm_object_t object)
1649 vm_object_t backing_object, new_backing_object;
1651 VM_OBJECT_ASSERT_WLOCKED(object);
1655 * Verify that the conditions are right for collapse:
1657 * The object exists and the backing object exists.
1659 if ((backing_object = object->backing_object) == NULL)
1663 * we check the backing object first, because it is most likely
1666 VM_OBJECT_WLOCK(backing_object);
1667 if (backing_object->handle != NULL ||
1668 (backing_object->type != OBJT_DEFAULT &&
1669 backing_object->type != OBJT_SWAP) ||
1670 (backing_object->flags & OBJ_DEAD) ||
1671 object->handle != NULL ||
1672 (object->type != OBJT_DEFAULT &&
1673 object->type != OBJT_SWAP) ||
1674 (object->flags & OBJ_DEAD)) {
1675 VM_OBJECT_WUNLOCK(backing_object);
1679 if (object->paging_in_progress != 0 ||
1680 backing_object->paging_in_progress != 0) {
1681 vm_object_qcollapse(object);
1682 VM_OBJECT_WUNLOCK(backing_object);
1687 * We know that we can either collapse the backing object (if
1688 * the parent is the only reference to it) or (perhaps) have
1689 * the parent bypass the object if the parent happens to shadow
1690 * all the resident pages in the entire backing object.
1692 * This is ignoring pager-backed pages such as swap pages.
1693 * vm_object_collapse_scan fails the shadowing test in this
1696 if (backing_object->ref_count == 1) {
1697 vm_object_pip_add(object, 1);
1698 vm_object_pip_add(backing_object, 1);
1701 * If there is exactly one reference to the backing
1702 * object, we can collapse it into the parent.
1704 vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT);
1706 #if VM_NRESERVLEVEL > 0
1708 * Break any reservations from backing_object.
1710 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1711 vm_reserv_break_all(backing_object);
1715 * Move the pager from backing_object to object.
1717 if (backing_object->type == OBJT_SWAP) {
1719 * swap_pager_copy() can sleep, in which case
1720 * the backing_object's and object's locks are
1721 * released and reacquired.
1722 * Since swap_pager_copy() is being asked to
1723 * destroy the source, it will change the
1724 * backing_object's type to OBJT_DEFAULT.
1729 OFF_TO_IDX(object->backing_object_offset), TRUE);
1732 * Object now shadows whatever backing_object did.
1733 * Note that the reference to
1734 * backing_object->backing_object moves from within
1735 * backing_object to within object.
1737 LIST_REMOVE(object, shadow_list);
1738 backing_object->shadow_count--;
1739 if (backing_object->backing_object) {
1740 VM_OBJECT_WLOCK(backing_object->backing_object);
1741 LIST_REMOVE(backing_object, shadow_list);
1743 &backing_object->backing_object->shadow_head,
1744 object, shadow_list);
1746 * The shadow_count has not changed.
1748 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1750 object->backing_object = backing_object->backing_object;
1751 object->backing_object_offset +=
1752 backing_object->backing_object_offset;
1755 * Discard backing_object.
1757 * Since the backing object has no pages, no pager left,
1758 * and no object references within it, all that is
1759 * necessary is to dispose of it.
1761 KASSERT(backing_object->ref_count == 1, (
1762 "backing_object %p was somehow re-referenced during collapse!",
1764 vm_object_pip_wakeup(backing_object);
1765 backing_object->type = OBJT_DEAD;
1766 backing_object->ref_count = 0;
1767 VM_OBJECT_WUNLOCK(backing_object);
1768 vm_object_destroy(backing_object);
1770 vm_object_pip_wakeup(object);
1774 * If we do not entirely shadow the backing object,
1775 * there is nothing we can do so we give up.
1777 if (object->resident_page_count != object->size &&
1778 !vm_object_scan_all_shadowed(object)) {
1779 VM_OBJECT_WUNLOCK(backing_object);
1784 * Make the parent shadow the next object in the
1785 * chain. Deallocating backing_object will not remove
1786 * it, since its reference count is at least 2.
1788 LIST_REMOVE(object, shadow_list);
1789 backing_object->shadow_count--;
1791 new_backing_object = backing_object->backing_object;
1792 if ((object->backing_object = new_backing_object) != NULL) {
1793 VM_OBJECT_WLOCK(new_backing_object);
1795 &new_backing_object->shadow_head,
1799 new_backing_object->shadow_count++;
1800 vm_object_reference_locked(new_backing_object);
1801 VM_OBJECT_WUNLOCK(new_backing_object);
1802 object->backing_object_offset +=
1803 backing_object->backing_object_offset;
1807 * Drop the reference count on backing_object. Since
1808 * its ref_count was at least 2, it will not vanish.
1810 backing_object->ref_count--;
1811 VM_OBJECT_WUNLOCK(backing_object);
1816 * Try again with this object's new backing object.
1822 * vm_object_page_remove:
1824 * For the given object, either frees or invalidates each of the
1825 * specified pages. In general, a page is freed. However, if a page is
1826 * wired for any reason other than the existence of a managed, wired
1827 * mapping, then it may be invalidated but not removed from the object.
1828 * Pages are specified by the given range ["start", "end") and the option
1829 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1830 * extends from "start" to the end of the object. If the option
1831 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1832 * specified range are affected. If the option OBJPR_NOTMAPPED is
1833 * specified, then the pages within the specified range must have no
1834 * mappings. Otherwise, if this option is not specified, any mappings to
1835 * the specified pages are removed before the pages are freed or
1838 * In general, this operation should only be performed on objects that
1839 * contain managed pages. There are, however, two exceptions. First, it
1840 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1841 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1842 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1843 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1845 * The object must be locked.
1848 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1853 VM_OBJECT_ASSERT_WLOCKED(object);
1854 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1855 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1856 ("vm_object_page_remove: illegal options for object %p", object));
1857 if (object->resident_page_count == 0)
1859 vm_object_pip_add(object, 1);
1861 p = vm_page_find_least(object, start);
1864 * Here, the variable "p" is either (1) the page with the least pindex
1865 * greater than or equal to the parameter "start" or (2) NULL.
1867 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1868 next = TAILQ_NEXT(p, listq);
1871 * If the page is wired for any reason besides the existence
1872 * of managed, wired mappings, then it cannot be freed. For
1873 * example, fictitious pages, which represent device memory,
1874 * are inherently wired and cannot be freed. They can,
1875 * however, be invalidated if the option OBJPR_CLEANONLY is
1879 if (vm_page_xbusied(p)) {
1880 VM_OBJECT_WUNLOCK(object);
1881 vm_page_busy_sleep(p, "vmopax", true);
1882 VM_OBJECT_WLOCK(object);
1885 if (p->wire_count != 0) {
1886 if ((options & OBJPR_NOTMAPPED) == 0)
1888 if ((options & OBJPR_CLEANONLY) == 0) {
1894 if (vm_page_busied(p)) {
1895 VM_OBJECT_WUNLOCK(object);
1896 vm_page_busy_sleep(p, "vmopar", false);
1897 VM_OBJECT_WLOCK(object);
1900 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1901 ("vm_object_page_remove: page %p is fictitious", p));
1902 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1903 if ((options & OBJPR_NOTMAPPED) == 0)
1904 pmap_remove_write(p);
1908 if ((options & OBJPR_NOTMAPPED) == 0)
1914 vm_object_pip_wakeup(object);
1918 * vm_object_page_noreuse:
1920 * For the given object, attempt to move the specified pages to
1921 * the head of the inactive queue. This bypasses regular LRU
1922 * operation and allows the pages to be reused quickly under memory
1923 * pressure. If a page is wired for any reason, then it will not
1924 * be queued. Pages are specified by the range ["start", "end").
1925 * As a special case, if "end" is zero, then the range extends from
1926 * "start" to the end of the object.
1928 * This operation should only be performed on objects that
1929 * contain non-fictitious, managed pages.
1931 * The object must be locked.
1934 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1936 struct mtx *mtx, *new_mtx;
1939 VM_OBJECT_ASSERT_WLOCKED(object);
1940 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1941 ("vm_object_page_noreuse: illegal object %p", object));
1942 if (object->resident_page_count == 0)
1944 p = vm_page_find_least(object, start);
1947 * Here, the variable "p" is either (1) the page with the least pindex
1948 * greater than or equal to the parameter "start" or (2) NULL.
1951 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1952 next = TAILQ_NEXT(p, listq);
1955 * Avoid releasing and reacquiring the same page lock.
1957 new_mtx = vm_page_lockptr(p);
1958 if (mtx != new_mtx) {
1964 vm_page_deactivate_noreuse(p);
1971 * Populate the specified range of the object with valid pages. Returns
1972 * TRUE if the range is successfully populated and FALSE otherwise.
1974 * Note: This function should be optimized to pass a larger array of
1975 * pages to vm_pager_get_pages() before it is applied to a non-
1976 * OBJT_DEVICE object.
1978 * The object must be locked.
1981 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1987 VM_OBJECT_ASSERT_WLOCKED(object);
1988 for (pindex = start; pindex < end; pindex++) {
1989 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1990 if (m->valid != VM_PAGE_BITS_ALL) {
1991 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
1992 if (rv != VM_PAGER_OK) {
2000 * Keep "m" busy because a subsequent iteration may unlock
2004 if (pindex > start) {
2005 m = vm_page_lookup(object, start);
2006 while (m != NULL && m->pindex < pindex) {
2008 m = TAILQ_NEXT(m, listq);
2011 return (pindex == end);
2015 * Routine: vm_object_coalesce
2016 * Function: Coalesces two objects backing up adjoining
2017 * regions of memory into a single object.
2019 * returns TRUE if objects were combined.
2021 * NOTE: Only works at the moment if the second object is NULL -
2022 * if it's not, which object do we lock first?
2025 * prev_object First object to coalesce
2026 * prev_offset Offset into prev_object
2027 * prev_size Size of reference to prev_object
2028 * next_size Size of reference to the second object
2029 * reserved Indicator that extension region has
2030 * swap accounted for
2033 * The object must *not* be locked.
2036 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2037 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2039 vm_pindex_t next_pindex;
2041 if (prev_object == NULL)
2043 VM_OBJECT_WLOCK(prev_object);
2044 if ((prev_object->type != OBJT_DEFAULT &&
2045 prev_object->type != OBJT_SWAP) ||
2046 (prev_object->flags & OBJ_TMPFS_NODE) != 0) {
2047 VM_OBJECT_WUNLOCK(prev_object);
2052 * Try to collapse the object first
2054 vm_object_collapse(prev_object);
2057 * Can't coalesce if: . more than one reference . paged out . shadows
2058 * another object . has a copy elsewhere (any of which mean that the
2059 * pages not mapped to prev_entry may be in use anyway)
2061 if (prev_object->backing_object != NULL) {
2062 VM_OBJECT_WUNLOCK(prev_object);
2066 prev_size >>= PAGE_SHIFT;
2067 next_size >>= PAGE_SHIFT;
2068 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2070 if ((prev_object->ref_count > 1) &&
2071 (prev_object->size != next_pindex)) {
2072 VM_OBJECT_WUNLOCK(prev_object);
2077 * Account for the charge.
2079 if (prev_object->cred != NULL) {
2082 * If prev_object was charged, then this mapping,
2083 * although not charged now, may become writable
2084 * later. Non-NULL cred in the object would prevent
2085 * swap reservation during enabling of the write
2086 * access, so reserve swap now. Failed reservation
2087 * cause allocation of the separate object for the map
2088 * entry, and swap reservation for this entry is
2089 * managed in appropriate time.
2091 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2092 prev_object->cred)) {
2093 VM_OBJECT_WUNLOCK(prev_object);
2096 prev_object->charge += ptoa(next_size);
2100 * Remove any pages that may still be in the object from a previous
2103 if (next_pindex < prev_object->size) {
2104 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2106 if (prev_object->type == OBJT_SWAP)
2107 swap_pager_freespace(prev_object,
2108 next_pindex, next_size);
2110 if (prev_object->cred != NULL) {
2111 KASSERT(prev_object->charge >=
2112 ptoa(prev_object->size - next_pindex),
2113 ("object %p overcharged 1 %jx %jx", prev_object,
2114 (uintmax_t)next_pindex, (uintmax_t)next_size));
2115 prev_object->charge -= ptoa(prev_object->size -
2122 * Extend the object if necessary.
2124 if (next_pindex + next_size > prev_object->size)
2125 prev_object->size = next_pindex + next_size;
2127 VM_OBJECT_WUNLOCK(prev_object);
2132 vm_object_set_writeable_dirty(vm_object_t object)
2135 VM_OBJECT_ASSERT_WLOCKED(object);
2136 if (object->type != OBJT_VNODE) {
2137 if ((object->flags & OBJ_TMPFS_NODE) != 0) {
2138 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs"));
2139 vm_object_set_flag(object, OBJ_TMPFS_DIRTY);
2143 object->generation++;
2144 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2146 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2152 * For each page offset within the specified range of the given object,
2153 * find the highest-level page in the shadow chain and unwire it. A page
2154 * must exist at every page offset, and the highest-level page must be
2158 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2161 vm_object_t tobject;
2163 vm_pindex_t end_pindex, pindex, tpindex;
2164 int depth, locked_depth;
2166 KASSERT((offset & PAGE_MASK) == 0,
2167 ("vm_object_unwire: offset is not page aligned"));
2168 KASSERT((length & PAGE_MASK) == 0,
2169 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2170 /* The wired count of a fictitious page never changes. */
2171 if ((object->flags & OBJ_FICTITIOUS) != 0)
2173 pindex = OFF_TO_IDX(offset);
2174 end_pindex = pindex + atop(length);
2176 VM_OBJECT_RLOCK(object);
2177 m = vm_page_find_least(object, pindex);
2178 while (pindex < end_pindex) {
2179 if (m == NULL || pindex < m->pindex) {
2181 * The first object in the shadow chain doesn't
2182 * contain a page at the current index. Therefore,
2183 * the page must exist in a backing object.
2190 OFF_TO_IDX(tobject->backing_object_offset);
2191 tobject = tobject->backing_object;
2192 KASSERT(tobject != NULL,
2193 ("vm_object_unwire: missing page"));
2194 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2197 if (depth == locked_depth) {
2199 VM_OBJECT_RLOCK(tobject);
2201 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2205 m = TAILQ_NEXT(m, listq);
2208 vm_page_unwire(tm, queue);
2213 /* Release the accumulated object locks. */
2214 for (depth = 0; depth < locked_depth; depth++) {
2215 tobject = object->backing_object;
2216 VM_OBJECT_RUNLOCK(object);
2222 vm_object_vnode(vm_object_t object)
2225 VM_OBJECT_ASSERT_LOCKED(object);
2226 if (object->type == OBJT_VNODE)
2227 return (object->handle);
2228 if (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS) != 0)
2229 return (object->un_pager.swp.swp_tmpfs);
2234 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2236 struct kinfo_vmobject kvo;
2237 char *fullpath, *freepath;
2244 if (req->oldptr == NULL) {
2246 * If an old buffer has not been provided, generate an
2247 * estimate of the space needed for a subsequent call.
2249 mtx_lock(&vm_object_list_mtx);
2251 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2252 if (obj->type == OBJT_DEAD)
2256 mtx_unlock(&vm_object_list_mtx);
2257 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2264 * VM objects are type stable and are never removed from the
2265 * list once added. This allows us to safely read obj->object_list
2266 * after reacquiring the VM object lock.
2268 mtx_lock(&vm_object_list_mtx);
2269 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2270 if (obj->type == OBJT_DEAD)
2272 VM_OBJECT_RLOCK(obj);
2273 if (obj->type == OBJT_DEAD) {
2274 VM_OBJECT_RUNLOCK(obj);
2277 mtx_unlock(&vm_object_list_mtx);
2278 kvo.kvo_size = ptoa(obj->size);
2279 kvo.kvo_resident = obj->resident_page_count;
2280 kvo.kvo_ref_count = obj->ref_count;
2281 kvo.kvo_shadow_count = obj->shadow_count;
2282 kvo.kvo_memattr = obj->memattr;
2284 kvo.kvo_inactive = 0;
2285 TAILQ_FOREACH(m, &obj->memq, listq) {
2287 * A page may belong to the object but be
2288 * dequeued and set to PQ_NONE while the
2289 * object lock is not held. This makes the
2290 * reads of m->queue below racy, and we do not
2291 * count pages set to PQ_NONE. However, this
2292 * sysctl is only meant to give an
2293 * approximation of the system anyway.
2295 if (vm_page_active(m))
2297 else if (vm_page_inactive(m))
2301 kvo.kvo_vn_fileid = 0;
2302 kvo.kvo_vn_fsid = 0;
2306 switch (obj->type) {
2308 kvo.kvo_type = KVME_TYPE_DEFAULT;
2311 kvo.kvo_type = KVME_TYPE_VNODE;
2316 kvo.kvo_type = KVME_TYPE_SWAP;
2319 kvo.kvo_type = KVME_TYPE_DEVICE;
2322 kvo.kvo_type = KVME_TYPE_PHYS;
2325 kvo.kvo_type = KVME_TYPE_DEAD;
2328 kvo.kvo_type = KVME_TYPE_SG;
2330 case OBJT_MGTDEVICE:
2331 kvo.kvo_type = KVME_TYPE_MGTDEVICE;
2334 kvo.kvo_type = KVME_TYPE_UNKNOWN;
2337 VM_OBJECT_RUNLOCK(obj);
2339 vn_fullpath(curthread, vp, &fullpath, &freepath);
2340 vn_lock(vp, LK_SHARED | LK_RETRY);
2341 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2342 kvo.kvo_vn_fileid = va.va_fileid;
2343 kvo.kvo_vn_fsid = va.va_fsid;
2348 strlcpy(kvo.kvo_path, fullpath, sizeof(kvo.kvo_path));
2349 if (freepath != NULL)
2350 free(freepath, M_TEMP);
2352 /* Pack record size down */
2353 kvo.kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) +
2354 strlen(kvo.kvo_path) + 1;
2355 kvo.kvo_structsize = roundup(kvo.kvo_structsize,
2357 error = SYSCTL_OUT(req, &kvo, kvo.kvo_structsize);
2358 mtx_lock(&vm_object_list_mtx);
2362 mtx_unlock(&vm_object_list_mtx);
2365 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2366 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2367 "List of VM objects");
2369 #include "opt_ddb.h"
2371 #include <sys/kernel.h>
2373 #include <sys/cons.h>
2375 #include <ddb/ddb.h>
2378 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2381 vm_map_entry_t tmpe;
2389 tmpe = map->header.next;
2390 entcount = map->nentries;
2391 while (entcount-- && (tmpe != &map->header)) {
2392 if (_vm_object_in_map(map, object, tmpe)) {
2397 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2398 tmpm = entry->object.sub_map;
2399 tmpe = tmpm->header.next;
2400 entcount = tmpm->nentries;
2401 while (entcount-- && tmpe != &tmpm->header) {
2402 if (_vm_object_in_map(tmpm, object, tmpe)) {
2407 } else if ((obj = entry->object.vm_object) != NULL) {
2408 for (; obj; obj = obj->backing_object)
2409 if (obj == object) {
2417 vm_object_in_map(vm_object_t object)
2421 /* sx_slock(&allproc_lock); */
2422 FOREACH_PROC_IN_SYSTEM(p) {
2423 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2425 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2426 /* sx_sunlock(&allproc_lock); */
2430 /* sx_sunlock(&allproc_lock); */
2431 if (_vm_object_in_map(kernel_map, object, 0))
2436 DB_SHOW_COMMAND(vmochk, vm_object_check)
2441 * make sure that internal objs are in a map somewhere
2442 * and none have zero ref counts.
2444 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2445 if (object->handle == NULL &&
2446 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2447 if (object->ref_count == 0) {
2448 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2449 (long)object->size);
2451 if (!vm_object_in_map(object)) {
2453 "vmochk: internal obj is not in a map: "
2454 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2455 object->ref_count, (u_long)object->size,
2456 (u_long)object->size,
2457 (void *)object->backing_object);
2464 * vm_object_print: [ debug ]
2466 DB_SHOW_COMMAND(object, vm_object_print_static)
2468 /* XXX convert args. */
2469 vm_object_t object = (vm_object_t)addr;
2470 boolean_t full = have_addr;
2474 /* XXX count is an (unused) arg. Avoid shadowing it. */
2475 #define count was_count
2483 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2484 object, (int)object->type, (uintmax_t)object->size,
2485 object->resident_page_count, object->ref_count, object->flags,
2486 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2487 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2488 object->shadow_count,
2489 object->backing_object ? object->backing_object->ref_count : 0,
2490 object->backing_object, (uintmax_t)object->backing_object_offset);
2497 TAILQ_FOREACH(p, &object->memq, listq) {
2499 db_iprintf("memory:=");
2500 else if (count == 6) {
2508 db_printf("(off=0x%jx,page=0x%jx)",
2509 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2519 /* XXX need this non-static entry for calling from vm_map_print. */
2522 /* db_expr_t */ long addr,
2523 boolean_t have_addr,
2524 /* db_expr_t */ long count,
2527 vm_object_print_static(addr, have_addr, count, modif);
2530 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2535 vm_page_t m, prev_m;
2539 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2540 db_printf("new object: %p\n", (void *)object);
2551 TAILQ_FOREACH(m, &object->memq, listq) {
2552 if (m->pindex > 128)
2554 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2555 prev_m->pindex + 1 != m->pindex) {
2557 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2558 (long)fidx, rcount, (long)pa);
2570 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2575 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2576 (long)fidx, rcount, (long)pa);
2586 pa = VM_PAGE_TO_PHYS(m);
2590 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2591 (long)fidx, rcount, (long)pa);