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->paging_in_progress = 0;
209 object->resident_page_count = 0;
210 object->shadow_count = 0;
212 mtx_lock(&vm_object_list_mtx);
213 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
214 mtx_unlock(&vm_object_list_mtx);
219 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
222 TAILQ_INIT(&object->memq);
223 LIST_INIT(&object->shadow_head);
228 panic("_vm_object_allocate: can't create OBJT_DEAD");
231 object->flags = OBJ_ONEMAPPING;
235 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
238 object->flags = OBJ_FICTITIOUS;
241 object->flags = OBJ_UNMANAGED;
247 panic("_vm_object_allocate: type %d is undefined", type);
250 object->generation = 1;
251 object->ref_count = 1;
252 object->memattr = VM_MEMATTR_DEFAULT;
255 object->handle = NULL;
256 object->backing_object = NULL;
257 object->backing_object_offset = (vm_ooffset_t) 0;
258 #if VM_NRESERVLEVEL > 0
259 LIST_INIT(&object->rvq);
261 umtx_shm_object_init(object);
267 * Initialize the VM objects module.
272 TAILQ_INIT(&vm_object_list);
273 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
275 rw_init(&kernel_object->lock, "kernel vm object");
276 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
278 #if VM_NRESERVLEVEL > 0
279 kernel_object->flags |= OBJ_COLORED;
280 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
283 rw_init(&kmem_object->lock, "kmem vm object");
284 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
286 #if VM_NRESERVLEVEL > 0
287 kmem_object->flags |= OBJ_COLORED;
288 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
292 * The lock portion of struct vm_object must be type stable due
293 * to vm_pageout_fallback_object_lock locking a vm object
294 * without holding any references to it.
296 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
302 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
308 vm_object_clear_flag(vm_object_t object, u_short bits)
311 VM_OBJECT_ASSERT_WLOCKED(object);
312 object->flags &= ~bits;
316 * Sets the default memory attribute for the specified object. Pages
317 * that are allocated to this object are by default assigned this memory
320 * Presently, this function must be called before any pages are allocated
321 * to the object. In the future, this requirement may be relaxed for
322 * "default" and "swap" objects.
325 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
328 VM_OBJECT_ASSERT_WLOCKED(object);
329 switch (object->type) {
337 if (!TAILQ_EMPTY(&object->memq))
338 return (KERN_FAILURE);
341 return (KERN_INVALID_ARGUMENT);
343 panic("vm_object_set_memattr: object %p is of undefined type",
346 object->memattr = memattr;
347 return (KERN_SUCCESS);
351 vm_object_pip_add(vm_object_t object, short i)
354 VM_OBJECT_ASSERT_WLOCKED(object);
355 object->paging_in_progress += i;
359 vm_object_pip_subtract(vm_object_t object, short i)
362 VM_OBJECT_ASSERT_WLOCKED(object);
363 object->paging_in_progress -= i;
367 vm_object_pip_wakeup(vm_object_t object)
370 VM_OBJECT_ASSERT_WLOCKED(object);
371 object->paging_in_progress--;
372 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
373 vm_object_clear_flag(object, OBJ_PIPWNT);
379 vm_object_pip_wakeupn(vm_object_t object, short i)
382 VM_OBJECT_ASSERT_WLOCKED(object);
384 object->paging_in_progress -= i;
385 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
386 vm_object_clear_flag(object, OBJ_PIPWNT);
392 vm_object_pip_wait(vm_object_t object, char *waitid)
395 VM_OBJECT_ASSERT_WLOCKED(object);
396 while (object->paging_in_progress) {
397 object->flags |= OBJ_PIPWNT;
398 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
403 * vm_object_allocate:
405 * Returns a new object with the given size.
408 vm_object_allocate(objtype_t type, vm_pindex_t size)
412 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
413 _vm_object_allocate(type, size, object);
419 * vm_object_reference:
421 * Gets another reference to the given object. Note: OBJ_DEAD
422 * objects can be referenced during final cleaning.
425 vm_object_reference(vm_object_t object)
429 VM_OBJECT_WLOCK(object);
430 vm_object_reference_locked(object);
431 VM_OBJECT_WUNLOCK(object);
435 * vm_object_reference_locked:
437 * Gets another reference to the given object.
439 * The object must be locked.
442 vm_object_reference_locked(vm_object_t object)
446 VM_OBJECT_ASSERT_WLOCKED(object);
448 if (object->type == OBJT_VNODE) {
455 * Handle deallocating an object of type OBJT_VNODE.
458 vm_object_vndeallocate(vm_object_t object)
460 struct vnode *vp = (struct vnode *) object->handle;
462 VM_OBJECT_ASSERT_WLOCKED(object);
463 KASSERT(object->type == OBJT_VNODE,
464 ("vm_object_vndeallocate: not a vnode object"));
465 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
467 if (object->ref_count == 0) {
468 vn_printf(vp, "vm_object_vndeallocate ");
469 panic("vm_object_vndeallocate: bad object reference count");
473 if (!umtx_shm_vnobj_persistent && object->ref_count == 1)
474 umtx_shm_object_terminated(object);
477 * The test for text of vp vnode does not need a bypass to
478 * reach right VV_TEXT there, since it is obtained from
481 if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) {
483 VM_OBJECT_WUNLOCK(object);
484 /* vrele may need the vnode lock. */
488 VM_OBJECT_WUNLOCK(object);
489 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
491 VM_OBJECT_WLOCK(object);
493 if (object->type == OBJT_DEAD) {
494 VM_OBJECT_WUNLOCK(object);
497 if (object->ref_count == 0)
499 VM_OBJECT_WUNLOCK(object);
506 * vm_object_deallocate:
508 * Release a reference to the specified object,
509 * gained either through a vm_object_allocate
510 * or a vm_object_reference call. When all references
511 * are gone, storage associated with this object
512 * may be relinquished.
514 * No object may be locked.
517 vm_object_deallocate(vm_object_t object)
522 while (object != NULL) {
523 VM_OBJECT_WLOCK(object);
524 if (object->type == OBJT_VNODE) {
525 vm_object_vndeallocate(object);
529 KASSERT(object->ref_count != 0,
530 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
533 * If the reference count goes to 0 we start calling
534 * vm_object_terminate() on the object chain.
535 * A ref count of 1 may be a special case depending on the
536 * shadow count being 0 or 1.
539 if (object->ref_count > 1) {
540 VM_OBJECT_WUNLOCK(object);
542 } else if (object->ref_count == 1) {
543 if (object->type == OBJT_SWAP &&
544 (object->flags & OBJ_TMPFS) != 0) {
545 vp = object->un_pager.swp.swp_tmpfs;
547 VM_OBJECT_WUNLOCK(object);
548 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
549 VM_OBJECT_WLOCK(object);
550 if (object->type == OBJT_DEAD ||
551 object->ref_count != 1) {
552 VM_OBJECT_WUNLOCK(object);
557 if ((object->flags & OBJ_TMPFS) != 0)
562 if (object->shadow_count == 0 &&
563 object->handle == NULL &&
564 (object->type == OBJT_DEFAULT ||
565 (object->type == OBJT_SWAP &&
566 (object->flags & OBJ_TMPFS_NODE) == 0))) {
567 vm_object_set_flag(object, OBJ_ONEMAPPING);
568 } else if ((object->shadow_count == 1) &&
569 (object->handle == NULL) &&
570 (object->type == OBJT_DEFAULT ||
571 object->type == OBJT_SWAP)) {
574 robject = LIST_FIRST(&object->shadow_head);
575 KASSERT(robject != NULL,
576 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
578 object->shadow_count));
579 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0,
580 ("shadowed tmpfs v_object %p", object));
581 if (!VM_OBJECT_TRYWLOCK(robject)) {
583 * Avoid a potential deadlock.
586 VM_OBJECT_WUNLOCK(object);
588 * More likely than not the thread
589 * holding robject's lock has lower
590 * priority than the current thread.
591 * Let the lower priority thread run.
597 * Collapse object into its shadow unless its
598 * shadow is dead. In that case, object will
599 * be deallocated by the thread that is
600 * deallocating its shadow.
602 if ((robject->flags & OBJ_DEAD) == 0 &&
603 (robject->handle == NULL) &&
604 (robject->type == OBJT_DEFAULT ||
605 robject->type == OBJT_SWAP)) {
607 robject->ref_count++;
609 if (robject->paging_in_progress) {
610 VM_OBJECT_WUNLOCK(object);
611 vm_object_pip_wait(robject,
613 temp = robject->backing_object;
614 if (object == temp) {
615 VM_OBJECT_WLOCK(object);
618 } else if (object->paging_in_progress) {
619 VM_OBJECT_WUNLOCK(robject);
620 object->flags |= OBJ_PIPWNT;
621 VM_OBJECT_SLEEP(object, object,
622 PDROP | PVM, "objde2", 0);
623 VM_OBJECT_WLOCK(robject);
624 temp = robject->backing_object;
625 if (object == temp) {
626 VM_OBJECT_WLOCK(object);
630 VM_OBJECT_WUNLOCK(object);
632 if (robject->ref_count == 1) {
633 robject->ref_count--;
638 vm_object_collapse(object);
639 VM_OBJECT_WUNLOCK(object);
642 VM_OBJECT_WUNLOCK(robject);
644 VM_OBJECT_WUNLOCK(object);
648 umtx_shm_object_terminated(object);
649 temp = object->backing_object;
651 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
652 ("shadowed tmpfs v_object 2 %p", object));
653 VM_OBJECT_WLOCK(temp);
654 LIST_REMOVE(object, shadow_list);
655 temp->shadow_count--;
656 VM_OBJECT_WUNLOCK(temp);
657 object->backing_object = NULL;
660 * Don't double-terminate, we could be in a termination
661 * recursion due to the terminate having to sync data
664 if ((object->flags & OBJ_DEAD) == 0)
665 vm_object_terminate(object);
667 VM_OBJECT_WUNLOCK(object);
673 * vm_object_destroy removes the object from the global object list
674 * and frees the space for the object.
677 vm_object_destroy(vm_object_t object)
681 * Release the allocation charge.
683 if (object->cred != NULL) {
684 swap_release_by_cred(object->charge, object->cred);
686 crfree(object->cred);
691 * Free the space for the object.
693 uma_zfree(obj_zone, object);
697 * vm_object_terminate actually destroys the specified object, freeing
698 * up all previously used resources.
700 * The object must be locked.
701 * This routine may block.
704 vm_object_terminate(vm_object_t object)
708 VM_OBJECT_ASSERT_WLOCKED(object);
711 * Make sure no one uses us.
713 vm_object_set_flag(object, OBJ_DEAD);
716 * wait for the pageout daemon to be done with the object
718 vm_object_pip_wait(object, "objtrm");
720 KASSERT(!object->paging_in_progress,
721 ("vm_object_terminate: pageout in progress"));
724 * Clean and free the pages, as appropriate. All references to the
725 * object are gone, so we don't need to lock it.
727 if (object->type == OBJT_VNODE) {
728 struct vnode *vp = (struct vnode *)object->handle;
731 * Clean pages and flush buffers.
733 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
734 VM_OBJECT_WUNLOCK(object);
736 vinvalbuf(vp, V_SAVE, 0, 0);
738 BO_LOCK(&vp->v_bufobj);
739 vp->v_bufobj.bo_flag |= BO_DEAD;
740 BO_UNLOCK(&vp->v_bufobj);
742 VM_OBJECT_WLOCK(object);
745 KASSERT(object->ref_count == 0,
746 ("vm_object_terminate: object with references, ref_count=%d",
750 * Free any remaining pageable pages. This also removes them from the
751 * paging queues. However, don't free wired pages, just remove them
752 * from the object. Rather than incrementally removing each page from
753 * the object, the page and object are reset to any empty state.
755 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
756 vm_page_assert_unbusied(p);
759 * Optimize the page's removal from the object by resetting
760 * its "object" field. Specifically, if the page is not
761 * wired, then the effect of this assignment is that
762 * vm_page_free()'s call to vm_page_remove() will return
763 * immediately without modifying the page or the object.
766 if (p->wire_count == 0) {
768 PCPU_INC(cnt.v_pfree);
773 * If the object contained any pages, then reset it to an empty state.
774 * None of the object's fields, including "resident_page_count", were
775 * modified by the preceding loop.
777 if (object->resident_page_count != 0) {
778 vm_radix_reclaim_allnodes(&object->rtree);
779 TAILQ_INIT(&object->memq);
780 object->resident_page_count = 0;
781 if (object->type == OBJT_VNODE)
782 vdrop(object->handle);
785 #if VM_NRESERVLEVEL > 0
786 if (__predict_false(!LIST_EMPTY(&object->rvq)))
787 vm_reserv_break_all(object);
790 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
791 object->type == OBJT_SWAP,
792 ("%s: non-swap obj %p has cred", __func__, object));
795 * Let the pager know object is dead.
797 vm_pager_deallocate(object);
798 VM_OBJECT_WUNLOCK(object);
800 vm_object_destroy(object);
804 * Make the page read-only so that we can clear the object flags. However, if
805 * this is a nosync mmap then the object is likely to stay dirty so do not
806 * mess with the page and do not clear the object flags. Returns TRUE if the
807 * page should be flushed, and FALSE otherwise.
810 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
814 * If we have been asked to skip nosync pages and this is a
815 * nosync page, skip it. Note that the object flags were not
816 * cleared in this case so we do not have to set them.
818 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
819 *clearobjflags = FALSE;
822 pmap_remove_write(p);
823 return (p->dirty != 0);
828 * vm_object_page_clean
830 * Clean all dirty pages in the specified range of object. Leaves page
831 * on whatever queue it is currently on. If NOSYNC is set then do not
832 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
833 * leaving the object dirty.
835 * When stuffing pages asynchronously, allow clustering. XXX we need a
836 * synchronous clustering mode implementation.
838 * Odd semantics: if start == end, we clean everything.
840 * The object must be locked.
842 * Returns FALSE if some page from the range was not written, as
843 * reported by the pager, and TRUE otherwise.
846 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
850 vm_pindex_t pi, tend, tstart;
851 int curgeneration, n, pagerflags;
852 boolean_t clearobjflags, eio, res;
854 VM_OBJECT_ASSERT_WLOCKED(object);
857 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
858 * objects. The check below prevents the function from
859 * operating on non-vnode objects.
861 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
862 object->resident_page_count == 0)
865 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
866 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
867 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
869 tstart = OFF_TO_IDX(start);
870 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
871 clearobjflags = tstart == 0 && tend >= object->size;
875 curgeneration = object->generation;
877 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
881 np = TAILQ_NEXT(p, listq);
884 if (vm_page_sleep_if_busy(p, "vpcwai")) {
885 if (object->generation != curgeneration) {
886 if ((flags & OBJPC_SYNC) != 0)
889 clearobjflags = FALSE;
891 np = vm_page_find_least(object, pi);
894 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
897 n = vm_object_page_collect_flush(object, p, pagerflags,
898 flags, &clearobjflags, &eio);
901 clearobjflags = FALSE;
903 if (object->generation != curgeneration) {
904 if ((flags & OBJPC_SYNC) != 0)
907 clearobjflags = FALSE;
911 * If the VOP_PUTPAGES() did a truncated write, so
912 * that even the first page of the run is not fully
913 * written, vm_pageout_flush() returns 0 as the run
914 * length. Since the condition that caused truncated
915 * write may be permanent, e.g. exhausted free space,
916 * accepting n == 0 would cause an infinite loop.
918 * Forwarding the iterator leaves the unwritten page
919 * behind, but there is not much we can do there if
920 * filesystem refuses to write it.
924 clearobjflags = FALSE;
926 np = vm_page_find_least(object, pi + n);
929 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
933 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
938 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
939 int flags, boolean_t *clearobjflags, boolean_t *eio)
941 vm_page_t ma[vm_pageout_page_count], p_first, tp;
942 int count, i, mreq, runlen;
944 vm_page_lock_assert(p, MA_NOTOWNED);
945 VM_OBJECT_ASSERT_WLOCKED(object);
950 for (tp = p; count < vm_pageout_page_count; count++) {
951 tp = vm_page_next(tp);
952 if (tp == NULL || vm_page_busied(tp))
954 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
958 for (p_first = p; count < vm_pageout_page_count; count++) {
959 tp = vm_page_prev(p_first);
960 if (tp == NULL || vm_page_busied(tp))
962 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
968 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
971 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
976 * Note that there is absolutely no sense in writing out
977 * anonymous objects, so we track down the vnode object
979 * We invalidate (remove) all pages from the address space
980 * for semantic correctness.
982 * If the backing object is a device object with unmanaged pages, then any
983 * mappings to the specified range of pages must be removed before this
984 * function is called.
986 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
987 * may start out with a NULL object.
990 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
991 boolean_t syncio, boolean_t invalidate)
993 vm_object_t backing_object;
996 int error, flags, fsync_after;
1003 VM_OBJECT_WLOCK(object);
1004 while ((backing_object = object->backing_object) != NULL) {
1005 VM_OBJECT_WLOCK(backing_object);
1006 offset += object->backing_object_offset;
1007 VM_OBJECT_WUNLOCK(object);
1008 object = backing_object;
1009 if (object->size < OFF_TO_IDX(offset + size))
1010 size = IDX_TO_OFF(object->size) - offset;
1013 * Flush pages if writing is allowed, invalidate them
1014 * if invalidation requested. Pages undergoing I/O
1015 * will be ignored by vm_object_page_remove().
1017 * We cannot lock the vnode and then wait for paging
1018 * to complete without deadlocking against vm_fault.
1019 * Instead we simply call vm_object_page_remove() and
1020 * allow it to block internally on a page-by-page
1021 * basis when it encounters pages undergoing async
1024 if (object->type == OBJT_VNODE &&
1025 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1026 vp = object->handle;
1027 VM_OBJECT_WUNLOCK(object);
1028 (void) vn_start_write(vp, &mp, V_WAIT);
1029 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1030 if (syncio && !invalidate && offset == 0 &&
1031 OFF_TO_IDX(size) == object->size) {
1033 * If syncing the whole mapping of the file,
1034 * it is faster to schedule all the writes in
1035 * async mode, also allowing the clustering,
1036 * and then wait for i/o to complete.
1041 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1042 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1043 fsync_after = FALSE;
1045 VM_OBJECT_WLOCK(object);
1046 res = vm_object_page_clean(object, offset, offset + size,
1048 VM_OBJECT_WUNLOCK(object);
1050 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1052 vn_finished_write(mp);
1055 VM_OBJECT_WLOCK(object);
1057 if ((object->type == OBJT_VNODE ||
1058 object->type == OBJT_DEVICE) && invalidate) {
1059 if (object->type == OBJT_DEVICE)
1061 * The option OBJPR_NOTMAPPED must be passed here
1062 * because vm_object_page_remove() cannot remove
1063 * unmanaged mappings.
1065 flags = OBJPR_NOTMAPPED;
1069 flags = OBJPR_CLEANONLY;
1070 vm_object_page_remove(object, OFF_TO_IDX(offset),
1071 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1073 VM_OBJECT_WUNLOCK(object);
1078 * vm_object_madvise:
1080 * Implements the madvise function at the object/page level.
1082 * MADV_WILLNEED (any object)
1084 * Activate the specified pages if they are resident.
1086 * MADV_DONTNEED (any object)
1088 * Deactivate the specified pages if they are resident.
1090 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1091 * OBJ_ONEMAPPING only)
1093 * Deactivate and clean the specified pages if they are
1094 * resident. This permits the process to reuse the pages
1095 * without faulting or the kernel to reclaim the pages
1099 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1102 vm_pindex_t tpindex;
1103 vm_object_t backing_object, tobject;
1108 VM_OBJECT_WLOCK(object);
1110 * Locate and adjust resident pages
1112 for (; pindex < end; pindex += 1) {
1118 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1119 * and those pages must be OBJ_ONEMAPPING.
1121 if (advise == MADV_FREE) {
1122 if ((tobject->type != OBJT_DEFAULT &&
1123 tobject->type != OBJT_SWAP) ||
1124 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1125 goto unlock_tobject;
1127 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1128 goto unlock_tobject;
1129 m = vm_page_lookup(tobject, tpindex);
1132 * There may be swap even if there is no backing page
1134 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1135 swap_pager_freespace(tobject, tpindex, 1);
1139 backing_object = tobject->backing_object;
1140 if (backing_object == NULL)
1141 goto unlock_tobject;
1142 VM_OBJECT_WLOCK(backing_object);
1143 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1144 if (tobject != object)
1145 VM_OBJECT_WUNLOCK(tobject);
1146 tobject = backing_object;
1148 } else if (m->valid != VM_PAGE_BITS_ALL)
1149 goto unlock_tobject;
1151 * If the page is not in a normal state, skip it.
1154 if (m->hold_count != 0 || m->wire_count != 0) {
1156 goto unlock_tobject;
1158 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1159 ("vm_object_madvise: page %p is fictitious", m));
1160 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1161 ("vm_object_madvise: page %p is not managed", m));
1162 if (vm_page_busied(m)) {
1163 if (advise == MADV_WILLNEED) {
1165 * Reference the page before unlocking and
1166 * sleeping so that the page daemon is less
1167 * likely to reclaim it.
1169 vm_page_aflag_set(m, PGA_REFERENCED);
1171 if (object != tobject)
1172 VM_OBJECT_WUNLOCK(object);
1173 VM_OBJECT_WUNLOCK(tobject);
1174 vm_page_busy_sleep(m, "madvpo", false);
1175 VM_OBJECT_WLOCK(object);
1178 if (advise == MADV_WILLNEED) {
1179 vm_page_activate(m);
1181 vm_page_advise(m, advise);
1184 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1185 swap_pager_freespace(tobject, tpindex, 1);
1187 if (tobject != object)
1188 VM_OBJECT_WUNLOCK(tobject);
1190 VM_OBJECT_WUNLOCK(object);
1196 * Create a new object which is backed by the
1197 * specified existing object range. The source
1198 * object reference is deallocated.
1200 * The new object and offset into that object
1201 * are returned in the source parameters.
1205 vm_object_t *object, /* IN/OUT */
1206 vm_ooffset_t *offset, /* IN/OUT */
1215 * Don't create the new object if the old object isn't shared.
1217 if (source != NULL) {
1218 VM_OBJECT_WLOCK(source);
1219 if (source->ref_count == 1 &&
1220 source->handle == NULL &&
1221 (source->type == OBJT_DEFAULT ||
1222 source->type == OBJT_SWAP)) {
1223 VM_OBJECT_WUNLOCK(source);
1226 VM_OBJECT_WUNLOCK(source);
1230 * Allocate a new object with the given length.
1232 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1235 * The new object shadows the source object, adding a reference to it.
1236 * Our caller changes his reference to point to the new object,
1237 * removing a reference to the source object. Net result: no change
1238 * of reference count.
1240 * Try to optimize the result object's page color when shadowing
1241 * in order to maintain page coloring consistency in the combined
1244 result->backing_object = source;
1246 * Store the offset into the source object, and fix up the offset into
1249 result->backing_object_offset = *offset;
1250 if (source != NULL) {
1251 VM_OBJECT_WLOCK(source);
1252 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1253 source->shadow_count++;
1254 #if VM_NRESERVLEVEL > 0
1255 result->flags |= source->flags & OBJ_COLORED;
1256 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1257 ((1 << (VM_NFREEORDER - 1)) - 1);
1259 VM_OBJECT_WUNLOCK(source);
1264 * Return the new things
1273 * Split the pages in a map entry into a new object. This affords
1274 * easier removal of unused pages, and keeps object inheritance from
1275 * being a negative impact on memory usage.
1278 vm_object_split(vm_map_entry_t entry)
1280 vm_page_t m, m_next;
1281 vm_object_t orig_object, new_object, source;
1282 vm_pindex_t idx, offidxstart;
1285 orig_object = entry->object.vm_object;
1286 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1288 if (orig_object->ref_count <= 1)
1290 VM_OBJECT_WUNLOCK(orig_object);
1292 offidxstart = OFF_TO_IDX(entry->offset);
1293 size = atop(entry->end - entry->start);
1296 * If swap_pager_copy() is later called, it will convert new_object
1297 * into a swap object.
1299 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1302 * At this point, the new object is still private, so the order in
1303 * which the original and new objects are locked does not matter.
1305 VM_OBJECT_WLOCK(new_object);
1306 VM_OBJECT_WLOCK(orig_object);
1307 source = orig_object->backing_object;
1308 if (source != NULL) {
1309 VM_OBJECT_WLOCK(source);
1310 if ((source->flags & OBJ_DEAD) != 0) {
1311 VM_OBJECT_WUNLOCK(source);
1312 VM_OBJECT_WUNLOCK(orig_object);
1313 VM_OBJECT_WUNLOCK(new_object);
1314 vm_object_deallocate(new_object);
1315 VM_OBJECT_WLOCK(orig_object);
1318 LIST_INSERT_HEAD(&source->shadow_head,
1319 new_object, shadow_list);
1320 source->shadow_count++;
1321 vm_object_reference_locked(source); /* for new_object */
1322 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1323 VM_OBJECT_WUNLOCK(source);
1324 new_object->backing_object_offset =
1325 orig_object->backing_object_offset + entry->offset;
1326 new_object->backing_object = source;
1328 if (orig_object->cred != NULL) {
1329 new_object->cred = orig_object->cred;
1330 crhold(orig_object->cred);
1331 new_object->charge = ptoa(size);
1332 KASSERT(orig_object->charge >= ptoa(size),
1333 ("orig_object->charge < 0"));
1334 orig_object->charge -= ptoa(size);
1337 m = vm_page_find_least(orig_object, offidxstart);
1338 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1340 m_next = TAILQ_NEXT(m, listq);
1343 * We must wait for pending I/O to complete before we can
1346 * We do not have to VM_PROT_NONE the page as mappings should
1347 * not be changed by this operation.
1349 if (vm_page_busied(m)) {
1350 VM_OBJECT_WUNLOCK(new_object);
1352 VM_OBJECT_WUNLOCK(orig_object);
1353 vm_page_busy_sleep(m, "spltwt", false);
1354 VM_OBJECT_WLOCK(orig_object);
1355 VM_OBJECT_WLOCK(new_object);
1359 /* vm_page_rename() will dirty the page. */
1360 if (vm_page_rename(m, new_object, idx)) {
1361 VM_OBJECT_WUNLOCK(new_object);
1362 VM_OBJECT_WUNLOCK(orig_object);
1364 VM_OBJECT_WLOCK(orig_object);
1365 VM_OBJECT_WLOCK(new_object);
1368 #if VM_NRESERVLEVEL > 0
1370 * If some of the reservation's allocated pages remain with
1371 * the original object, then transferring the reservation to
1372 * the new object is neither particularly beneficial nor
1373 * particularly harmful as compared to leaving the reservation
1374 * with the original object. If, however, all of the
1375 * reservation's allocated pages are transferred to the new
1376 * object, then transferring the reservation is typically
1377 * beneficial. Determining which of these two cases applies
1378 * would be more costly than unconditionally renaming the
1381 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1383 if (orig_object->type == OBJT_SWAP)
1386 if (orig_object->type == OBJT_SWAP) {
1388 * swap_pager_copy() can sleep, in which case the orig_object's
1389 * and new_object's locks are released and reacquired.
1391 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1392 TAILQ_FOREACH(m, &new_object->memq, listq)
1395 VM_OBJECT_WUNLOCK(orig_object);
1396 VM_OBJECT_WUNLOCK(new_object);
1397 entry->object.vm_object = new_object;
1398 entry->offset = 0LL;
1399 vm_object_deallocate(orig_object);
1400 VM_OBJECT_WLOCK(new_object);
1403 #define OBSC_COLLAPSE_NOWAIT 0x0002
1404 #define OBSC_COLLAPSE_WAIT 0x0004
1407 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next,
1410 vm_object_t backing_object;
1412 VM_OBJECT_ASSERT_WLOCKED(object);
1413 backing_object = object->backing_object;
1414 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1416 KASSERT(p == NULL || vm_page_busied(p), ("unbusy page %p", p));
1417 KASSERT(p == NULL || p->object == object || p->object == backing_object,
1418 ("invalid ownership %p %p %p", p, object, backing_object));
1419 if ((op & OBSC_COLLAPSE_NOWAIT) != 0)
1423 VM_OBJECT_WUNLOCK(object);
1424 VM_OBJECT_WUNLOCK(backing_object);
1428 vm_page_busy_sleep(p, "vmocol", false);
1429 VM_OBJECT_WLOCK(object);
1430 VM_OBJECT_WLOCK(backing_object);
1431 return (TAILQ_FIRST(&backing_object->memq));
1435 vm_object_scan_all_shadowed(vm_object_t object)
1437 vm_object_t backing_object;
1439 vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1441 VM_OBJECT_ASSERT_WLOCKED(object);
1442 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1444 backing_object = object->backing_object;
1446 if (backing_object->type != OBJT_DEFAULT &&
1447 backing_object->type != OBJT_SWAP)
1450 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1451 p = vm_page_find_least(backing_object, pi);
1452 ps = swap_pager_find_least(backing_object, pi);
1455 * Only check pages inside the parent object's range and
1456 * inside the parent object's mapping of the backing object.
1459 if (p != NULL && p->pindex < pi)
1460 p = TAILQ_NEXT(p, listq);
1462 ps = swap_pager_find_least(backing_object, pi);
1463 if (p == NULL && ps >= backing_object->size)
1468 pi = MIN(p->pindex, ps);
1470 new_pindex = pi - backing_offset_index;
1471 if (new_pindex >= object->size)
1475 * See if the parent has the page or if the parent's object
1476 * pager has the page. If the parent has the page but the page
1477 * is not valid, the parent's object pager must have the page.
1479 * If this fails, the parent does not completely shadow the
1480 * object and we might as well give up now.
1482 pp = vm_page_lookup(object, new_pindex);
1483 if ((pp == NULL || pp->valid == 0) &&
1484 !vm_pager_has_page(object, new_pindex, NULL, NULL))
1491 vm_object_collapse_scan(vm_object_t object, int op)
1493 vm_object_t backing_object;
1494 vm_page_t next, p, pp;
1495 vm_pindex_t backing_offset_index, new_pindex;
1497 VM_OBJECT_ASSERT_WLOCKED(object);
1498 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1500 backing_object = object->backing_object;
1501 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1504 * Initial conditions
1506 if ((op & OBSC_COLLAPSE_WAIT) != 0)
1507 vm_object_set_flag(backing_object, OBJ_DEAD);
1512 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1513 next = TAILQ_NEXT(p, listq);
1514 new_pindex = p->pindex - backing_offset_index;
1517 * Check for busy page
1519 if (vm_page_busied(p)) {
1520 next = vm_object_collapse_scan_wait(object, p, next, op);
1524 KASSERT(p->object == backing_object,
1525 ("vm_object_collapse_scan: object mismatch"));
1527 if (p->pindex < backing_offset_index ||
1528 new_pindex >= object->size) {
1529 if (backing_object->type == OBJT_SWAP)
1530 swap_pager_freespace(backing_object, p->pindex,
1534 * Page is out of the parent object's range, we can
1535 * simply destroy it.
1538 KASSERT(!pmap_page_is_mapped(p),
1539 ("freeing mapped page %p", p));
1540 if (p->wire_count == 0)
1548 pp = vm_page_lookup(object, new_pindex);
1549 if (pp != NULL && vm_page_busied(pp)) {
1551 * The page in the parent is busy and possibly not
1552 * (yet) valid. Until its state is finalized by the
1553 * busy bit owner, we can't tell whether it shadows the
1554 * original page. Therefore, we must either skip it
1555 * and the original (backing_object) page or wait for
1556 * its state to be finalized.
1558 * This is due to a race with vm_fault() where we must
1559 * unbusy the original (backing_obj) page before we can
1560 * (re)lock the parent. Hence we can get here.
1562 next = vm_object_collapse_scan_wait(object, pp, next,
1567 KASSERT(pp == NULL || pp->valid != 0,
1568 ("unbusy invalid page %p", pp));
1570 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1573 * The page already exists in the parent OR swap exists
1574 * for this location in the parent. Leave the parent's
1575 * page alone. Destroy the original page from the
1578 if (backing_object->type == OBJT_SWAP)
1579 swap_pager_freespace(backing_object, p->pindex,
1582 KASSERT(!pmap_page_is_mapped(p),
1583 ("freeing mapped page %p", p));
1584 if (p->wire_count == 0)
1593 * Page does not exist in parent, rename the page from the
1594 * backing object to the main object.
1596 * If the page was mapped to a process, it can remain mapped
1597 * through the rename. vm_page_rename() will dirty the page.
1599 if (vm_page_rename(p, object, new_pindex)) {
1600 next = vm_object_collapse_scan_wait(object, NULL, next,
1605 /* Use the old pindex to free the right page. */
1606 if (backing_object->type == OBJT_SWAP)
1607 swap_pager_freespace(backing_object,
1608 new_pindex + backing_offset_index, 1);
1610 #if VM_NRESERVLEVEL > 0
1612 * Rename the reservation.
1614 vm_reserv_rename(p, object, backing_object,
1615 backing_offset_index);
1623 * this version of collapse allows the operation to occur earlier and
1624 * when paging_in_progress is true for an object... This is not a complete
1625 * operation, but should plug 99.9% of the rest of the leaks.
1628 vm_object_qcollapse(vm_object_t object)
1630 vm_object_t backing_object = object->backing_object;
1632 VM_OBJECT_ASSERT_WLOCKED(object);
1633 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1635 if (backing_object->ref_count != 1)
1638 vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT);
1642 * vm_object_collapse:
1644 * Collapse an object with the object backing it.
1645 * Pages in the backing object are moved into the
1646 * parent, and the backing object is deallocated.
1649 vm_object_collapse(vm_object_t object)
1651 vm_object_t backing_object, new_backing_object;
1653 VM_OBJECT_ASSERT_WLOCKED(object);
1657 * Verify that the conditions are right for collapse:
1659 * The object exists and the backing object exists.
1661 if ((backing_object = object->backing_object) == NULL)
1665 * we check the backing object first, because it is most likely
1668 VM_OBJECT_WLOCK(backing_object);
1669 if (backing_object->handle != NULL ||
1670 (backing_object->type != OBJT_DEFAULT &&
1671 backing_object->type != OBJT_SWAP) ||
1672 (backing_object->flags & OBJ_DEAD) ||
1673 object->handle != NULL ||
1674 (object->type != OBJT_DEFAULT &&
1675 object->type != OBJT_SWAP) ||
1676 (object->flags & OBJ_DEAD)) {
1677 VM_OBJECT_WUNLOCK(backing_object);
1681 if (object->paging_in_progress != 0 ||
1682 backing_object->paging_in_progress != 0) {
1683 vm_object_qcollapse(object);
1684 VM_OBJECT_WUNLOCK(backing_object);
1689 * We know that we can either collapse the backing object (if
1690 * the parent is the only reference to it) or (perhaps) have
1691 * the parent bypass the object if the parent happens to shadow
1692 * all the resident pages in the entire backing object.
1694 * This is ignoring pager-backed pages such as swap pages.
1695 * vm_object_collapse_scan fails the shadowing test in this
1698 if (backing_object->ref_count == 1) {
1699 vm_object_pip_add(object, 1);
1700 vm_object_pip_add(backing_object, 1);
1703 * If there is exactly one reference to the backing
1704 * object, we can collapse it into the parent.
1706 vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT);
1708 #if VM_NRESERVLEVEL > 0
1710 * Break any reservations from backing_object.
1712 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1713 vm_reserv_break_all(backing_object);
1717 * Move the pager from backing_object to object.
1719 if (backing_object->type == OBJT_SWAP) {
1721 * swap_pager_copy() can sleep, in which case
1722 * the backing_object's and object's locks are
1723 * released and reacquired.
1724 * Since swap_pager_copy() is being asked to
1725 * destroy the source, it will change the
1726 * backing_object's type to OBJT_DEFAULT.
1731 OFF_TO_IDX(object->backing_object_offset), TRUE);
1734 * Object now shadows whatever backing_object did.
1735 * Note that the reference to
1736 * backing_object->backing_object moves from within
1737 * backing_object to within object.
1739 LIST_REMOVE(object, shadow_list);
1740 backing_object->shadow_count--;
1741 if (backing_object->backing_object) {
1742 VM_OBJECT_WLOCK(backing_object->backing_object);
1743 LIST_REMOVE(backing_object, shadow_list);
1745 &backing_object->backing_object->shadow_head,
1746 object, shadow_list);
1748 * The shadow_count has not changed.
1750 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1752 object->backing_object = backing_object->backing_object;
1753 object->backing_object_offset +=
1754 backing_object->backing_object_offset;
1757 * Discard backing_object.
1759 * Since the backing object has no pages, no pager left,
1760 * and no object references within it, all that is
1761 * necessary is to dispose of it.
1763 KASSERT(backing_object->ref_count == 1, (
1764 "backing_object %p was somehow re-referenced during collapse!",
1766 vm_object_pip_wakeup(backing_object);
1767 backing_object->type = OBJT_DEAD;
1768 backing_object->ref_count = 0;
1769 VM_OBJECT_WUNLOCK(backing_object);
1770 vm_object_destroy(backing_object);
1772 vm_object_pip_wakeup(object);
1776 * If we do not entirely shadow the backing object,
1777 * there is nothing we can do so we give up.
1779 if (object->resident_page_count != object->size &&
1780 !vm_object_scan_all_shadowed(object)) {
1781 VM_OBJECT_WUNLOCK(backing_object);
1786 * Make the parent shadow the next object in the
1787 * chain. Deallocating backing_object will not remove
1788 * it, since its reference count is at least 2.
1790 LIST_REMOVE(object, shadow_list);
1791 backing_object->shadow_count--;
1793 new_backing_object = backing_object->backing_object;
1794 if ((object->backing_object = new_backing_object) != NULL) {
1795 VM_OBJECT_WLOCK(new_backing_object);
1797 &new_backing_object->shadow_head,
1801 new_backing_object->shadow_count++;
1802 vm_object_reference_locked(new_backing_object);
1803 VM_OBJECT_WUNLOCK(new_backing_object);
1804 object->backing_object_offset +=
1805 backing_object->backing_object_offset;
1809 * Drop the reference count on backing_object. Since
1810 * its ref_count was at least 2, it will not vanish.
1812 backing_object->ref_count--;
1813 VM_OBJECT_WUNLOCK(backing_object);
1818 * Try again with this object's new backing object.
1824 * vm_object_page_remove:
1826 * For the given object, either frees or invalidates each of the
1827 * specified pages. In general, a page is freed. However, if a page is
1828 * wired for any reason other than the existence of a managed, wired
1829 * mapping, then it may be invalidated but not removed from the object.
1830 * Pages are specified by the given range ["start", "end") and the option
1831 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1832 * extends from "start" to the end of the object. If the option
1833 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1834 * specified range are affected. If the option OBJPR_NOTMAPPED is
1835 * specified, then the pages within the specified range must have no
1836 * mappings. Otherwise, if this option is not specified, any mappings to
1837 * the specified pages are removed before the pages are freed or
1840 * In general, this operation should only be performed on objects that
1841 * contain managed pages. There are, however, two exceptions. First, it
1842 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1843 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1844 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1845 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1847 * The object must be locked.
1850 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1855 VM_OBJECT_ASSERT_WLOCKED(object);
1856 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1857 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1858 ("vm_object_page_remove: illegal options for object %p", object));
1859 if (object->resident_page_count == 0)
1861 vm_object_pip_add(object, 1);
1863 p = vm_page_find_least(object, start);
1866 * Here, the variable "p" is either (1) the page with the least pindex
1867 * greater than or equal to the parameter "start" or (2) NULL.
1869 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1870 next = TAILQ_NEXT(p, listq);
1873 * If the page is wired for any reason besides the existence
1874 * of managed, wired mappings, then it cannot be freed. For
1875 * example, fictitious pages, which represent device memory,
1876 * are inherently wired and cannot be freed. They can,
1877 * however, be invalidated if the option OBJPR_CLEANONLY is
1881 if (vm_page_xbusied(p)) {
1882 VM_OBJECT_WUNLOCK(object);
1883 vm_page_busy_sleep(p, "vmopax", true);
1884 VM_OBJECT_WLOCK(object);
1887 if (p->wire_count != 0) {
1888 if ((options & OBJPR_NOTMAPPED) == 0)
1890 if ((options & OBJPR_CLEANONLY) == 0) {
1896 if (vm_page_busied(p)) {
1897 VM_OBJECT_WUNLOCK(object);
1898 vm_page_busy_sleep(p, "vmopar", false);
1899 VM_OBJECT_WLOCK(object);
1902 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1903 ("vm_object_page_remove: page %p is fictitious", p));
1904 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1905 if ((options & OBJPR_NOTMAPPED) == 0)
1906 pmap_remove_write(p);
1910 if ((options & OBJPR_NOTMAPPED) == 0)
1916 vm_object_pip_wakeup(object);
1920 * vm_object_page_noreuse:
1922 * For the given object, attempt to move the specified pages to
1923 * the head of the inactive queue. This bypasses regular LRU
1924 * operation and allows the pages to be reused quickly under memory
1925 * pressure. If a page is wired for any reason, then it will not
1926 * be queued. Pages are specified by the range ["start", "end").
1927 * As a special case, if "end" is zero, then the range extends from
1928 * "start" to the end of the object.
1930 * This operation should only be performed on objects that
1931 * contain non-fictitious, managed pages.
1933 * The object must be locked.
1936 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1938 struct mtx *mtx, *new_mtx;
1941 VM_OBJECT_ASSERT_WLOCKED(object);
1942 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1943 ("vm_object_page_noreuse: illegal object %p", object));
1944 if (object->resident_page_count == 0)
1946 p = vm_page_find_least(object, start);
1949 * Here, the variable "p" is either (1) the page with the least pindex
1950 * greater than or equal to the parameter "start" or (2) NULL.
1953 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1954 next = TAILQ_NEXT(p, listq);
1957 * Avoid releasing and reacquiring the same page lock.
1959 new_mtx = vm_page_lockptr(p);
1960 if (mtx != new_mtx) {
1966 vm_page_deactivate_noreuse(p);
1973 * Populate the specified range of the object with valid pages. Returns
1974 * TRUE if the range is successfully populated and FALSE otherwise.
1976 * Note: This function should be optimized to pass a larger array of
1977 * pages to vm_pager_get_pages() before it is applied to a non-
1978 * OBJT_DEVICE object.
1980 * The object must be locked.
1983 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1989 VM_OBJECT_ASSERT_WLOCKED(object);
1990 for (pindex = start; pindex < end; pindex++) {
1991 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1992 if (m->valid != VM_PAGE_BITS_ALL) {
1993 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
1994 if (rv != VM_PAGER_OK) {
2002 * Keep "m" busy because a subsequent iteration may unlock
2006 if (pindex > start) {
2007 m = vm_page_lookup(object, start);
2008 while (m != NULL && m->pindex < pindex) {
2010 m = TAILQ_NEXT(m, listq);
2013 return (pindex == end);
2017 * Routine: vm_object_coalesce
2018 * Function: Coalesces two objects backing up adjoining
2019 * regions of memory into a single object.
2021 * returns TRUE if objects were combined.
2023 * NOTE: Only works at the moment if the second object is NULL -
2024 * if it's not, which object do we lock first?
2027 * prev_object First object to coalesce
2028 * prev_offset Offset into prev_object
2029 * prev_size Size of reference to prev_object
2030 * next_size Size of reference to the second object
2031 * reserved Indicator that extension region has
2032 * swap accounted for
2035 * The object must *not* be locked.
2038 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2039 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2041 vm_pindex_t next_pindex;
2043 if (prev_object == NULL)
2045 VM_OBJECT_WLOCK(prev_object);
2046 if ((prev_object->type != OBJT_DEFAULT &&
2047 prev_object->type != OBJT_SWAP) ||
2048 (prev_object->flags & OBJ_TMPFS_NODE) != 0) {
2049 VM_OBJECT_WUNLOCK(prev_object);
2054 * Try to collapse the object first
2056 vm_object_collapse(prev_object);
2059 * Can't coalesce if: . more than one reference . paged out . shadows
2060 * another object . has a copy elsewhere (any of which mean that the
2061 * pages not mapped to prev_entry may be in use anyway)
2063 if (prev_object->backing_object != NULL) {
2064 VM_OBJECT_WUNLOCK(prev_object);
2068 prev_size >>= PAGE_SHIFT;
2069 next_size >>= PAGE_SHIFT;
2070 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2072 if ((prev_object->ref_count > 1) &&
2073 (prev_object->size != next_pindex)) {
2074 VM_OBJECT_WUNLOCK(prev_object);
2079 * Account for the charge.
2081 if (prev_object->cred != NULL) {
2084 * If prev_object was charged, then this mapping,
2085 * although not charged now, may become writable
2086 * later. Non-NULL cred in the object would prevent
2087 * swap reservation during enabling of the write
2088 * access, so reserve swap now. Failed reservation
2089 * cause allocation of the separate object for the map
2090 * entry, and swap reservation for this entry is
2091 * managed in appropriate time.
2093 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2094 prev_object->cred)) {
2095 VM_OBJECT_WUNLOCK(prev_object);
2098 prev_object->charge += ptoa(next_size);
2102 * Remove any pages that may still be in the object from a previous
2105 if (next_pindex < prev_object->size) {
2106 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2108 if (prev_object->type == OBJT_SWAP)
2109 swap_pager_freespace(prev_object,
2110 next_pindex, next_size);
2112 if (prev_object->cred != NULL) {
2113 KASSERT(prev_object->charge >=
2114 ptoa(prev_object->size - next_pindex),
2115 ("object %p overcharged 1 %jx %jx", prev_object,
2116 (uintmax_t)next_pindex, (uintmax_t)next_size));
2117 prev_object->charge -= ptoa(prev_object->size -
2124 * Extend the object if necessary.
2126 if (next_pindex + next_size > prev_object->size)
2127 prev_object->size = next_pindex + next_size;
2129 VM_OBJECT_WUNLOCK(prev_object);
2134 vm_object_set_writeable_dirty(vm_object_t object)
2137 VM_OBJECT_ASSERT_WLOCKED(object);
2138 if (object->type != OBJT_VNODE) {
2139 if ((object->flags & OBJ_TMPFS_NODE) != 0) {
2140 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs"));
2141 vm_object_set_flag(object, OBJ_TMPFS_DIRTY);
2145 object->generation++;
2146 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2148 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2154 * For each page offset within the specified range of the given object,
2155 * find the highest-level page in the shadow chain and unwire it. A page
2156 * must exist at every page offset, and the highest-level page must be
2160 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2163 vm_object_t tobject;
2165 vm_pindex_t end_pindex, pindex, tpindex;
2166 int depth, locked_depth;
2168 KASSERT((offset & PAGE_MASK) == 0,
2169 ("vm_object_unwire: offset is not page aligned"));
2170 KASSERT((length & PAGE_MASK) == 0,
2171 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2172 /* The wired count of a fictitious page never changes. */
2173 if ((object->flags & OBJ_FICTITIOUS) != 0)
2175 pindex = OFF_TO_IDX(offset);
2176 end_pindex = pindex + atop(length);
2178 VM_OBJECT_RLOCK(object);
2179 m = vm_page_find_least(object, pindex);
2180 while (pindex < end_pindex) {
2181 if (m == NULL || pindex < m->pindex) {
2183 * The first object in the shadow chain doesn't
2184 * contain a page at the current index. Therefore,
2185 * the page must exist in a backing object.
2192 OFF_TO_IDX(tobject->backing_object_offset);
2193 tobject = tobject->backing_object;
2194 KASSERT(tobject != NULL,
2195 ("vm_object_unwire: missing page"));
2196 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2199 if (depth == locked_depth) {
2201 VM_OBJECT_RLOCK(tobject);
2203 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2207 m = TAILQ_NEXT(m, listq);
2210 vm_page_unwire(tm, queue);
2215 /* Release the accumulated object locks. */
2216 for (depth = 0; depth < locked_depth; depth++) {
2217 tobject = object->backing_object;
2218 VM_OBJECT_RUNLOCK(object);
2224 vm_object_vnode(vm_object_t object)
2227 VM_OBJECT_ASSERT_LOCKED(object);
2228 if (object->type == OBJT_VNODE)
2229 return (object->handle);
2230 if (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS) != 0)
2231 return (object->un_pager.swp.swp_tmpfs);
2236 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2238 struct kinfo_vmobject kvo;
2239 char *fullpath, *freepath;
2246 if (req->oldptr == NULL) {
2248 * If an old buffer has not been provided, generate an
2249 * estimate of the space needed for a subsequent call.
2251 mtx_lock(&vm_object_list_mtx);
2253 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2254 if (obj->type == OBJT_DEAD)
2258 mtx_unlock(&vm_object_list_mtx);
2259 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2266 * VM objects are type stable and are never removed from the
2267 * list once added. This allows us to safely read obj->object_list
2268 * after reacquiring the VM object lock.
2270 mtx_lock(&vm_object_list_mtx);
2271 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2272 if (obj->type == OBJT_DEAD)
2274 VM_OBJECT_RLOCK(obj);
2275 if (obj->type == OBJT_DEAD) {
2276 VM_OBJECT_RUNLOCK(obj);
2279 mtx_unlock(&vm_object_list_mtx);
2280 kvo.kvo_size = ptoa(obj->size);
2281 kvo.kvo_resident = obj->resident_page_count;
2282 kvo.kvo_ref_count = obj->ref_count;
2283 kvo.kvo_shadow_count = obj->shadow_count;
2284 kvo.kvo_memattr = obj->memattr;
2286 kvo.kvo_inactive = 0;
2287 TAILQ_FOREACH(m, &obj->memq, listq) {
2289 * A page may belong to the object but be
2290 * dequeued and set to PQ_NONE while the
2291 * object lock is not held. This makes the
2292 * reads of m->queue below racy, and we do not
2293 * count pages set to PQ_NONE. However, this
2294 * sysctl is only meant to give an
2295 * approximation of the system anyway.
2297 if (vm_page_active(m))
2299 else if (vm_page_inactive(m))
2303 kvo.kvo_vn_fileid = 0;
2304 kvo.kvo_vn_fsid = 0;
2308 switch (obj->type) {
2310 kvo.kvo_type = KVME_TYPE_DEFAULT;
2313 kvo.kvo_type = KVME_TYPE_VNODE;
2318 kvo.kvo_type = KVME_TYPE_SWAP;
2321 kvo.kvo_type = KVME_TYPE_DEVICE;
2324 kvo.kvo_type = KVME_TYPE_PHYS;
2327 kvo.kvo_type = KVME_TYPE_DEAD;
2330 kvo.kvo_type = KVME_TYPE_SG;
2332 case OBJT_MGTDEVICE:
2333 kvo.kvo_type = KVME_TYPE_MGTDEVICE;
2336 kvo.kvo_type = KVME_TYPE_UNKNOWN;
2339 VM_OBJECT_RUNLOCK(obj);
2341 vn_fullpath(curthread, vp, &fullpath, &freepath);
2342 vn_lock(vp, LK_SHARED | LK_RETRY);
2343 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2344 kvo.kvo_vn_fileid = va.va_fileid;
2345 kvo.kvo_vn_fsid = va.va_fsid;
2350 strlcpy(kvo.kvo_path, fullpath, sizeof(kvo.kvo_path));
2351 if (freepath != NULL)
2352 free(freepath, M_TEMP);
2354 /* Pack record size down */
2355 kvo.kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) +
2356 strlen(kvo.kvo_path) + 1;
2357 kvo.kvo_structsize = roundup(kvo.kvo_structsize,
2359 error = SYSCTL_OUT(req, &kvo, kvo.kvo_structsize);
2360 mtx_lock(&vm_object_list_mtx);
2364 mtx_unlock(&vm_object_list_mtx);
2367 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2368 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2369 "List of VM objects");
2371 #include "opt_ddb.h"
2373 #include <sys/kernel.h>
2375 #include <sys/cons.h>
2377 #include <ddb/ddb.h>
2380 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2383 vm_map_entry_t tmpe;
2391 tmpe = map->header.next;
2392 entcount = map->nentries;
2393 while (entcount-- && (tmpe != &map->header)) {
2394 if (_vm_object_in_map(map, object, tmpe)) {
2399 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2400 tmpm = entry->object.sub_map;
2401 tmpe = tmpm->header.next;
2402 entcount = tmpm->nentries;
2403 while (entcount-- && tmpe != &tmpm->header) {
2404 if (_vm_object_in_map(tmpm, object, tmpe)) {
2409 } else if ((obj = entry->object.vm_object) != NULL) {
2410 for (; obj; obj = obj->backing_object)
2411 if (obj == object) {
2419 vm_object_in_map(vm_object_t object)
2423 /* sx_slock(&allproc_lock); */
2424 FOREACH_PROC_IN_SYSTEM(p) {
2425 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2427 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2428 /* sx_sunlock(&allproc_lock); */
2432 /* sx_sunlock(&allproc_lock); */
2433 if (_vm_object_in_map(kernel_map, object, 0))
2438 DB_SHOW_COMMAND(vmochk, vm_object_check)
2443 * make sure that internal objs are in a map somewhere
2444 * and none have zero ref counts.
2446 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2447 if (object->handle == NULL &&
2448 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2449 if (object->ref_count == 0) {
2450 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2451 (long)object->size);
2453 if (!vm_object_in_map(object)) {
2455 "vmochk: internal obj is not in a map: "
2456 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2457 object->ref_count, (u_long)object->size,
2458 (u_long)object->size,
2459 (void *)object->backing_object);
2466 * vm_object_print: [ debug ]
2468 DB_SHOW_COMMAND(object, vm_object_print_static)
2470 /* XXX convert args. */
2471 vm_object_t object = (vm_object_t)addr;
2472 boolean_t full = have_addr;
2476 /* XXX count is an (unused) arg. Avoid shadowing it. */
2477 #define count was_count
2485 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2486 object, (int)object->type, (uintmax_t)object->size,
2487 object->resident_page_count, object->ref_count, object->flags,
2488 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2489 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2490 object->shadow_count,
2491 object->backing_object ? object->backing_object->ref_count : 0,
2492 object->backing_object, (uintmax_t)object->backing_object_offset);
2499 TAILQ_FOREACH(p, &object->memq, listq) {
2501 db_iprintf("memory:=");
2502 else if (count == 6) {
2510 db_printf("(off=0x%jx,page=0x%jx)",
2511 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2521 /* XXX need this non-static entry for calling from vm_map_print. */
2524 /* db_expr_t */ long addr,
2525 boolean_t have_addr,
2526 /* db_expr_t */ long count,
2529 vm_object_print_static(addr, have_addr, count, modif);
2532 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2537 vm_page_t m, prev_m;
2541 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2542 db_printf("new object: %p\n", (void *)object);
2553 TAILQ_FOREACH(m, &object->memq, listq) {
2554 if (m->pindex > 128)
2556 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2557 prev_m->pindex + 1 != m->pindex) {
2559 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2560 (long)fidx, rcount, (long)pa);
2572 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2577 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2578 (long)fidx, rcount, (long)pa);
2588 pa = VM_PAGE_TO_PHYS(m);
2592 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2593 (long)fidx, rcount, (long)pa);