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>
82 #include <sys/vnode.h>
83 #include <sys/vmmeter.h>
87 #include <vm/vm_param.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pageout.h>
93 #include <vm/vm_pager.h>
94 #include <vm/swap_pager.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_extern.h>
97 #include <vm/vm_radix.h>
98 #include <vm/vm_reserv.h>
101 static int old_msync;
102 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
103 "Use old (insecure) msync behavior");
105 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
106 int pagerflags, int flags, boolean_t *clearobjflags,
108 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
109 boolean_t *clearobjflags);
110 static void vm_object_qcollapse(vm_object_t object);
111 static void vm_object_vndeallocate(vm_object_t object);
114 * Virtual memory objects maintain the actual data
115 * associated with allocated virtual memory. A given
116 * page of memory exists within exactly one object.
118 * An object is only deallocated when all "references"
119 * are given up. Only one "reference" to a given
120 * region of an object should be writeable.
122 * Associated with each object is a list of all resident
123 * memory pages belonging to that object; this list is
124 * maintained by the "vm_page" module, and locked by the object's
127 * Each object also records a "pager" routine which is
128 * used to retrieve (and store) pages to the proper backing
129 * storage. In addition, objects may be backed by other
130 * objects from which they were virtual-copied.
132 * The only items within the object structure which are
133 * modified after time of creation are:
134 * reference count locked by object's lock
135 * pager routine locked by object's lock
139 struct object_q vm_object_list;
140 struct mtx vm_object_list_mtx; /* lock for object list and count */
142 struct vm_object kernel_object_store;
143 struct vm_object kmem_object_store;
145 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
148 static long object_collapses;
149 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
150 &object_collapses, 0, "VM object collapses");
152 static long object_bypasses;
153 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
154 &object_bypasses, 0, "VM object bypasses");
156 static uma_zone_t obj_zone;
158 static int vm_object_zinit(void *mem, int size, int flags);
161 static void vm_object_zdtor(void *mem, int size, void *arg);
164 vm_object_zdtor(void *mem, int size, void *arg)
168 object = (vm_object_t)mem;
169 KASSERT(TAILQ_EMPTY(&object->memq),
170 ("object %p has resident pages in its memq", object));
171 KASSERT(vm_radix_is_empty(&object->rtree),
172 ("object %p has resident pages in its trie", object));
173 #if VM_NRESERVLEVEL > 0
174 KASSERT(LIST_EMPTY(&object->rvq),
175 ("object %p has reservations",
178 KASSERT(vm_object_cache_is_empty(object),
179 ("object %p has cached pages",
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));
194 vm_object_zinit(void *mem, int size, int flags)
198 object = (vm_object_t)mem;
199 bzero(&object->lock, sizeof(object->lock));
200 rw_init_flags(&object->lock, "vm object", RW_DUPOK);
202 /* These are true for any object that has been freed */
203 object->rtree.rt_root = 0;
204 object->rtree.rt_flags = 0;
205 object->paging_in_progress = 0;
206 object->resident_page_count = 0;
207 object->shadow_count = 0;
208 object->cache.rt_root = 0;
209 object->cache.rt_flags = 0;
214 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
217 TAILQ_INIT(&object->memq);
218 LIST_INIT(&object->shadow_head);
223 panic("_vm_object_allocate: can't create OBJT_DEAD");
226 object->flags = OBJ_ONEMAPPING;
230 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
233 object->flags = OBJ_FICTITIOUS;
236 object->flags = OBJ_UNMANAGED;
242 panic("_vm_object_allocate: type %d is undefined", type);
245 object->generation = 1;
246 object->ref_count = 1;
247 object->memattr = VM_MEMATTR_DEFAULT;
250 object->handle = NULL;
251 object->backing_object = NULL;
252 object->backing_object_offset = (vm_ooffset_t) 0;
253 #if VM_NRESERVLEVEL > 0
254 LIST_INIT(&object->rvq);
257 mtx_lock(&vm_object_list_mtx);
258 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
259 mtx_unlock(&vm_object_list_mtx);
265 * Initialize the VM objects module.
270 TAILQ_INIT(&vm_object_list);
271 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
273 rw_init(&kernel_object->lock, "kernel vm object");
274 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
276 #if VM_NRESERVLEVEL > 0
277 kernel_object->flags |= OBJ_COLORED;
278 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
281 rw_init(&kmem_object->lock, "kmem vm object");
282 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
284 #if VM_NRESERVLEVEL > 0
285 kmem_object->flags |= OBJ_COLORED;
286 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
290 * The lock portion of struct vm_object must be type stable due
291 * to vm_pageout_fallback_object_lock locking a vm object
292 * without holding any references to it.
294 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
300 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
306 vm_object_clear_flag(vm_object_t object, u_short bits)
309 VM_OBJECT_ASSERT_WLOCKED(object);
310 object->flags &= ~bits;
314 * Sets the default memory attribute for the specified object. Pages
315 * that are allocated to this object are by default assigned this memory
318 * Presently, this function must be called before any pages are allocated
319 * to the object. In the future, this requirement may be relaxed for
320 * "default" and "swap" objects.
323 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
326 VM_OBJECT_ASSERT_WLOCKED(object);
327 switch (object->type) {
335 if (!TAILQ_EMPTY(&object->memq))
336 return (KERN_FAILURE);
339 return (KERN_INVALID_ARGUMENT);
341 panic("vm_object_set_memattr: object %p is of undefined type",
344 object->memattr = memattr;
345 return (KERN_SUCCESS);
349 vm_object_pip_add(vm_object_t object, short i)
352 VM_OBJECT_ASSERT_WLOCKED(object);
353 object->paging_in_progress += i;
357 vm_object_pip_subtract(vm_object_t object, short i)
360 VM_OBJECT_ASSERT_WLOCKED(object);
361 object->paging_in_progress -= i;
365 vm_object_pip_wakeup(vm_object_t object)
368 VM_OBJECT_ASSERT_WLOCKED(object);
369 object->paging_in_progress--;
370 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
371 vm_object_clear_flag(object, OBJ_PIPWNT);
377 vm_object_pip_wakeupn(vm_object_t object, short i)
380 VM_OBJECT_ASSERT_WLOCKED(object);
382 object->paging_in_progress -= i;
383 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
384 vm_object_clear_flag(object, OBJ_PIPWNT);
390 vm_object_pip_wait(vm_object_t object, char *waitid)
393 VM_OBJECT_ASSERT_WLOCKED(object);
394 while (object->paging_in_progress) {
395 object->flags |= OBJ_PIPWNT;
396 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
401 * vm_object_allocate:
403 * Returns a new object with the given size.
406 vm_object_allocate(objtype_t type, vm_pindex_t size)
410 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
411 _vm_object_allocate(type, size, object);
417 * vm_object_reference:
419 * Gets another reference to the given object. Note: OBJ_DEAD
420 * objects can be referenced during final cleaning.
423 vm_object_reference(vm_object_t object)
427 VM_OBJECT_WLOCK(object);
428 vm_object_reference_locked(object);
429 VM_OBJECT_WUNLOCK(object);
433 * vm_object_reference_locked:
435 * Gets another reference to the given object.
437 * The object must be locked.
440 vm_object_reference_locked(vm_object_t object)
444 VM_OBJECT_ASSERT_WLOCKED(object);
446 if (object->type == OBJT_VNODE) {
453 * Handle deallocating an object of type OBJT_VNODE.
456 vm_object_vndeallocate(vm_object_t object)
458 struct vnode *vp = (struct vnode *) object->handle;
460 VM_OBJECT_ASSERT_WLOCKED(object);
461 KASSERT(object->type == OBJT_VNODE,
462 ("vm_object_vndeallocate: not a vnode object"));
463 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
465 if (object->ref_count == 0) {
466 vprint("vm_object_vndeallocate", vp);
467 panic("vm_object_vndeallocate: bad object reference count");
472 * The test for text of vp vnode does not need a bypass to
473 * reach right VV_TEXT there, since it is obtained from
476 if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) {
478 VM_OBJECT_WUNLOCK(object);
479 /* vrele may need the vnode lock. */
483 VM_OBJECT_WUNLOCK(object);
484 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
486 VM_OBJECT_WLOCK(object);
488 if (object->type == OBJT_DEAD) {
489 VM_OBJECT_WUNLOCK(object);
492 if (object->ref_count == 0)
494 VM_OBJECT_WUNLOCK(object);
501 * vm_object_deallocate:
503 * Release a reference to the specified object,
504 * gained either through a vm_object_allocate
505 * or a vm_object_reference call. When all references
506 * are gone, storage associated with this object
507 * may be relinquished.
509 * No object may be locked.
512 vm_object_deallocate(vm_object_t object)
517 while (object != NULL) {
518 VM_OBJECT_WLOCK(object);
519 if (object->type == OBJT_VNODE) {
520 vm_object_vndeallocate(object);
524 KASSERT(object->ref_count != 0,
525 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
528 * If the reference count goes to 0 we start calling
529 * vm_object_terminate() on the object chain.
530 * A ref count of 1 may be a special case depending on the
531 * shadow count being 0 or 1.
534 if (object->ref_count > 1) {
535 VM_OBJECT_WUNLOCK(object);
537 } else if (object->ref_count == 1) {
538 if (object->type == OBJT_SWAP &&
539 (object->flags & OBJ_TMPFS) != 0) {
540 vp = object->un_pager.swp.swp_tmpfs;
542 VM_OBJECT_WUNLOCK(object);
543 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
544 VM_OBJECT_WLOCK(object);
545 if (object->type == OBJT_DEAD ||
546 object->ref_count != 1) {
547 VM_OBJECT_WUNLOCK(object);
552 if ((object->flags & OBJ_TMPFS) != 0)
557 if (object->shadow_count == 0 &&
558 object->handle == NULL &&
559 (object->type == OBJT_DEFAULT ||
560 (object->type == OBJT_SWAP &&
561 (object->flags & OBJ_TMPFS_NODE) == 0))) {
562 vm_object_set_flag(object, OBJ_ONEMAPPING);
563 } else if ((object->shadow_count == 1) &&
564 (object->handle == NULL) &&
565 (object->type == OBJT_DEFAULT ||
566 object->type == OBJT_SWAP)) {
569 robject = LIST_FIRST(&object->shadow_head);
570 KASSERT(robject != NULL,
571 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
573 object->shadow_count));
574 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0,
575 ("shadowed tmpfs v_object %p", object));
576 if (!VM_OBJECT_TRYWLOCK(robject)) {
578 * Avoid a potential deadlock.
581 VM_OBJECT_WUNLOCK(object);
583 * More likely than not the thread
584 * holding robject's lock has lower
585 * priority than the current thread.
586 * Let the lower priority thread run.
592 * Collapse object into its shadow unless its
593 * shadow is dead. In that case, object will
594 * be deallocated by the thread that is
595 * deallocating its shadow.
597 if ((robject->flags & OBJ_DEAD) == 0 &&
598 (robject->handle == NULL) &&
599 (robject->type == OBJT_DEFAULT ||
600 robject->type == OBJT_SWAP)) {
602 robject->ref_count++;
604 if (robject->paging_in_progress) {
605 VM_OBJECT_WUNLOCK(object);
606 vm_object_pip_wait(robject,
608 temp = robject->backing_object;
609 if (object == temp) {
610 VM_OBJECT_WLOCK(object);
613 } else if (object->paging_in_progress) {
614 VM_OBJECT_WUNLOCK(robject);
615 object->flags |= OBJ_PIPWNT;
616 VM_OBJECT_SLEEP(object, object,
617 PDROP | PVM, "objde2", 0);
618 VM_OBJECT_WLOCK(robject);
619 temp = robject->backing_object;
620 if (object == temp) {
621 VM_OBJECT_WLOCK(object);
625 VM_OBJECT_WUNLOCK(object);
627 if (robject->ref_count == 1) {
628 robject->ref_count--;
633 vm_object_collapse(object);
634 VM_OBJECT_WUNLOCK(object);
637 VM_OBJECT_WUNLOCK(robject);
639 VM_OBJECT_WUNLOCK(object);
643 temp = object->backing_object;
645 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
646 ("shadowed tmpfs v_object 2 %p", object));
647 VM_OBJECT_WLOCK(temp);
648 LIST_REMOVE(object, shadow_list);
649 temp->shadow_count--;
650 VM_OBJECT_WUNLOCK(temp);
651 object->backing_object = NULL;
654 * Don't double-terminate, we could be in a termination
655 * recursion due to the terminate having to sync data
658 if ((object->flags & OBJ_DEAD) == 0)
659 vm_object_terminate(object);
661 VM_OBJECT_WUNLOCK(object);
667 * vm_object_destroy removes the object from the global object list
668 * and frees the space for the object.
671 vm_object_destroy(vm_object_t object)
675 * Remove the object from the global object list.
677 mtx_lock(&vm_object_list_mtx);
678 TAILQ_REMOVE(&vm_object_list, object, object_list);
679 mtx_unlock(&vm_object_list_mtx);
682 * Release the allocation charge.
684 if (object->cred != NULL) {
685 KASSERT(object->type == OBJT_DEFAULT ||
686 object->type == OBJT_SWAP,
687 ("%s: non-swap obj %p has cred", __func__, object));
688 swap_release_by_cred(object->charge, object->cred);
690 crfree(object->cred);
695 * Free the space for the object.
697 uma_zfree(obj_zone, object);
701 * vm_object_terminate actually destroys the specified object, freeing
702 * up all previously used resources.
704 * The object must be locked.
705 * This routine may block.
708 vm_object_terminate(vm_object_t object)
712 VM_OBJECT_ASSERT_WLOCKED(object);
715 * Make sure no one uses us.
717 vm_object_set_flag(object, OBJ_DEAD);
720 * wait for the pageout daemon to be done with the object
722 vm_object_pip_wait(object, "objtrm");
724 KASSERT(!object->paging_in_progress,
725 ("vm_object_terminate: pageout in progress"));
728 * Clean and free the pages, as appropriate. All references to the
729 * object are gone, so we don't need to lock it.
731 if (object->type == OBJT_VNODE) {
732 struct vnode *vp = (struct vnode *)object->handle;
735 * Clean pages and flush buffers.
737 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
738 VM_OBJECT_WUNLOCK(object);
740 vinvalbuf(vp, V_SAVE, 0, 0);
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);
789 if (__predict_false(!vm_object_cache_is_empty(object)))
790 vm_page_cache_free(object, 0, 0);
793 * Let the pager know object is dead.
795 vm_pager_deallocate(object);
796 VM_OBJECT_WUNLOCK(object);
798 vm_object_destroy(object);
802 * Make the page read-only so that we can clear the object flags. However, if
803 * this is a nosync mmap then the object is likely to stay dirty so do not
804 * mess with the page and do not clear the object flags. Returns TRUE if the
805 * page should be flushed, and FALSE otherwise.
808 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
812 * If we have been asked to skip nosync pages and this is a
813 * nosync page, skip it. Note that the object flags were not
814 * cleared in this case so we do not have to set them.
816 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
817 *clearobjflags = FALSE;
820 pmap_remove_write(p);
821 return (p->dirty != 0);
826 * vm_object_page_clean
828 * Clean all dirty pages in the specified range of object. Leaves page
829 * on whatever queue it is currently on. If NOSYNC is set then do not
830 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
831 * leaving the object dirty.
833 * When stuffing pages asynchronously, allow clustering. XXX we need a
834 * synchronous clustering mode implementation.
836 * Odd semantics: if start == end, we clean everything.
838 * The object must be locked.
840 * Returns FALSE if some page from the range was not written, as
841 * reported by the pager, and TRUE otherwise.
844 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
848 vm_pindex_t pi, tend, tstart;
849 int curgeneration, n, pagerflags;
850 boolean_t clearobjflags, eio, res;
852 VM_OBJECT_ASSERT_WLOCKED(object);
855 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
856 * objects. The check below prevents the function from
857 * operating on non-vnode objects.
859 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
860 object->resident_page_count == 0)
863 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
864 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
865 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
867 tstart = OFF_TO_IDX(start);
868 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
869 clearobjflags = tstart == 0 && tend >= object->size;
873 curgeneration = object->generation;
875 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
879 np = TAILQ_NEXT(p, listq);
882 if (vm_page_sleep_if_busy(p, "vpcwai")) {
883 if (object->generation != curgeneration) {
884 if ((flags & OBJPC_SYNC) != 0)
887 clearobjflags = FALSE;
889 np = vm_page_find_least(object, pi);
892 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
895 n = vm_object_page_collect_flush(object, p, pagerflags,
896 flags, &clearobjflags, &eio);
899 clearobjflags = FALSE;
901 if (object->generation != curgeneration) {
902 if ((flags & OBJPC_SYNC) != 0)
905 clearobjflags = FALSE;
909 * If the VOP_PUTPAGES() did a truncated write, so
910 * that even the first page of the run is not fully
911 * written, vm_pageout_flush() returns 0 as the run
912 * length. Since the condition that caused truncated
913 * write may be permanent, e.g. exhausted free space,
914 * accepting n == 0 would cause an infinite loop.
916 * Forwarding the iterator leaves the unwritten page
917 * behind, but there is not much we can do there if
918 * filesystem refuses to write it.
922 clearobjflags = FALSE;
924 np = vm_page_find_least(object, pi + n);
927 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
931 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
936 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
937 int flags, boolean_t *clearobjflags, boolean_t *eio)
939 vm_page_t ma[vm_pageout_page_count], p_first, tp;
940 int count, i, mreq, runlen;
942 vm_page_lock_assert(p, MA_NOTOWNED);
943 VM_OBJECT_ASSERT_WLOCKED(object);
948 for (tp = p; count < vm_pageout_page_count; count++) {
949 tp = vm_page_next(tp);
950 if (tp == NULL || vm_page_busied(tp))
952 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
956 for (p_first = p; count < vm_pageout_page_count; count++) {
957 tp = vm_page_prev(p_first);
958 if (tp == NULL || vm_page_busied(tp))
960 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
966 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
969 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
974 * Note that there is absolutely no sense in writing out
975 * anonymous objects, so we track down the vnode object
977 * We invalidate (remove) all pages from the address space
978 * for semantic correctness.
980 * If the backing object is a device object with unmanaged pages, then any
981 * mappings to the specified range of pages must be removed before this
982 * function is called.
984 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
985 * may start out with a NULL object.
988 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
989 boolean_t syncio, boolean_t invalidate)
991 vm_object_t backing_object;
994 int error, flags, fsync_after;
1001 VM_OBJECT_WLOCK(object);
1002 while ((backing_object = object->backing_object) != NULL) {
1003 VM_OBJECT_WLOCK(backing_object);
1004 offset += object->backing_object_offset;
1005 VM_OBJECT_WUNLOCK(object);
1006 object = backing_object;
1007 if (object->size < OFF_TO_IDX(offset + size))
1008 size = IDX_TO_OFF(object->size) - offset;
1011 * Flush pages if writing is allowed, invalidate them
1012 * if invalidation requested. Pages undergoing I/O
1013 * will be ignored by vm_object_page_remove().
1015 * We cannot lock the vnode and then wait for paging
1016 * to complete without deadlocking against vm_fault.
1017 * Instead we simply call vm_object_page_remove() and
1018 * allow it to block internally on a page-by-page
1019 * basis when it encounters pages undergoing async
1022 if (object->type == OBJT_VNODE &&
1023 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1024 vp = object->handle;
1025 VM_OBJECT_WUNLOCK(object);
1026 (void) vn_start_write(vp, &mp, V_WAIT);
1027 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1028 if (syncio && !invalidate && offset == 0 &&
1029 OFF_TO_IDX(size) == object->size) {
1031 * If syncing the whole mapping of the file,
1032 * it is faster to schedule all the writes in
1033 * async mode, also allowing the clustering,
1034 * and then wait for i/o to complete.
1039 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1040 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1041 fsync_after = FALSE;
1043 VM_OBJECT_WLOCK(object);
1044 res = vm_object_page_clean(object, offset, offset + size,
1046 VM_OBJECT_WUNLOCK(object);
1048 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1050 vn_finished_write(mp);
1053 VM_OBJECT_WLOCK(object);
1055 if ((object->type == OBJT_VNODE ||
1056 object->type == OBJT_DEVICE) && invalidate) {
1057 if (object->type == OBJT_DEVICE)
1059 * The option OBJPR_NOTMAPPED must be passed here
1060 * because vm_object_page_remove() cannot remove
1061 * unmanaged mappings.
1063 flags = OBJPR_NOTMAPPED;
1065 flags = OBJPR_NOTWIRED;
1067 flags = OBJPR_CLEANONLY | OBJPR_NOTWIRED;
1068 vm_object_page_remove(object, OFF_TO_IDX(offset),
1069 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1071 VM_OBJECT_WUNLOCK(object);
1076 * vm_object_madvise:
1078 * Implements the madvise function at the object/page level.
1080 * MADV_WILLNEED (any object)
1082 * Activate the specified pages if they are resident.
1084 * MADV_DONTNEED (any object)
1086 * Deactivate the specified pages if they are resident.
1088 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1089 * OBJ_ONEMAPPING only)
1091 * Deactivate and clean the specified pages if they are
1092 * resident. This permits the process to reuse the pages
1093 * without faulting or the kernel to reclaim the pages
1097 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1100 vm_pindex_t tpindex;
1101 vm_object_t backing_object, tobject;
1106 VM_OBJECT_WLOCK(object);
1108 * Locate and adjust resident pages
1110 for (; pindex < end; pindex += 1) {
1116 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1117 * and those pages must be OBJ_ONEMAPPING.
1119 if (advise == MADV_FREE) {
1120 if ((tobject->type != OBJT_DEFAULT &&
1121 tobject->type != OBJT_SWAP) ||
1122 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1123 goto unlock_tobject;
1125 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1126 goto unlock_tobject;
1127 m = vm_page_lookup(tobject, tpindex);
1128 if (m == NULL && advise == MADV_WILLNEED) {
1130 * If the page is cached, reactivate it.
1132 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1137 * There may be swap even if there is no backing page
1139 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1140 swap_pager_freespace(tobject, tpindex, 1);
1144 backing_object = tobject->backing_object;
1145 if (backing_object == NULL)
1146 goto unlock_tobject;
1147 VM_OBJECT_WLOCK(backing_object);
1148 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1149 if (tobject != object)
1150 VM_OBJECT_WUNLOCK(tobject);
1151 tobject = backing_object;
1153 } else if (m->valid != VM_PAGE_BITS_ALL)
1154 goto unlock_tobject;
1156 * If the page is not in a normal state, skip it.
1159 if (m->hold_count != 0 || m->wire_count != 0) {
1161 goto unlock_tobject;
1163 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1164 ("vm_object_madvise: page %p is fictitious", m));
1165 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1166 ("vm_object_madvise: page %p is not managed", m));
1167 if (vm_page_busied(m)) {
1168 if (advise == MADV_WILLNEED) {
1170 * Reference the page before unlocking and
1171 * sleeping so that the page daemon is less
1172 * likely to reclaim it.
1174 vm_page_aflag_set(m, PGA_REFERENCED);
1176 if (object != tobject)
1177 VM_OBJECT_WUNLOCK(object);
1178 VM_OBJECT_WUNLOCK(tobject);
1179 vm_page_busy_sleep(m, "madvpo");
1180 VM_OBJECT_WLOCK(object);
1183 if (advise == MADV_WILLNEED) {
1184 vm_page_activate(m);
1186 vm_page_advise(m, advise);
1189 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1190 swap_pager_freespace(tobject, tpindex, 1);
1192 if (tobject != object)
1193 VM_OBJECT_WUNLOCK(tobject);
1195 VM_OBJECT_WUNLOCK(object);
1201 * Create a new object which is backed by the
1202 * specified existing object range. The source
1203 * object reference is deallocated.
1205 * The new object and offset into that object
1206 * are returned in the source parameters.
1210 vm_object_t *object, /* IN/OUT */
1211 vm_ooffset_t *offset, /* IN/OUT */
1220 * Don't create the new object if the old object isn't shared.
1222 if (source != NULL) {
1223 VM_OBJECT_WLOCK(source);
1224 if (source->ref_count == 1 &&
1225 source->handle == NULL &&
1226 (source->type == OBJT_DEFAULT ||
1227 source->type == OBJT_SWAP)) {
1228 VM_OBJECT_WUNLOCK(source);
1231 VM_OBJECT_WUNLOCK(source);
1235 * Allocate a new object with the given length.
1237 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1240 * The new object shadows the source object, adding a reference to it.
1241 * Our caller changes his reference to point to the new object,
1242 * removing a reference to the source object. Net result: no change
1243 * of reference count.
1245 * Try to optimize the result object's page color when shadowing
1246 * in order to maintain page coloring consistency in the combined
1249 result->backing_object = source;
1251 * Store the offset into the source object, and fix up the offset into
1254 result->backing_object_offset = *offset;
1255 if (source != NULL) {
1256 VM_OBJECT_WLOCK(source);
1257 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1258 source->shadow_count++;
1259 #if VM_NRESERVLEVEL > 0
1260 result->flags |= source->flags & OBJ_COLORED;
1261 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1262 ((1 << (VM_NFREEORDER - 1)) - 1);
1264 VM_OBJECT_WUNLOCK(source);
1269 * Return the new things
1278 * Split the pages in a map entry into a new object. This affords
1279 * easier removal of unused pages, and keeps object inheritance from
1280 * being a negative impact on memory usage.
1283 vm_object_split(vm_map_entry_t entry)
1285 vm_page_t m, m_next;
1286 vm_object_t orig_object, new_object, source;
1287 vm_pindex_t idx, offidxstart;
1290 orig_object = entry->object.vm_object;
1291 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1293 if (orig_object->ref_count <= 1)
1295 VM_OBJECT_WUNLOCK(orig_object);
1297 offidxstart = OFF_TO_IDX(entry->offset);
1298 size = atop(entry->end - entry->start);
1301 * If swap_pager_copy() is later called, it will convert new_object
1302 * into a swap object.
1304 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1307 * At this point, the new object is still private, so the order in
1308 * which the original and new objects are locked does not matter.
1310 VM_OBJECT_WLOCK(new_object);
1311 VM_OBJECT_WLOCK(orig_object);
1312 source = orig_object->backing_object;
1313 if (source != NULL) {
1314 VM_OBJECT_WLOCK(source);
1315 if ((source->flags & OBJ_DEAD) != 0) {
1316 VM_OBJECT_WUNLOCK(source);
1317 VM_OBJECT_WUNLOCK(orig_object);
1318 VM_OBJECT_WUNLOCK(new_object);
1319 vm_object_deallocate(new_object);
1320 VM_OBJECT_WLOCK(orig_object);
1323 LIST_INSERT_HEAD(&source->shadow_head,
1324 new_object, shadow_list);
1325 source->shadow_count++;
1326 vm_object_reference_locked(source); /* for new_object */
1327 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1328 VM_OBJECT_WUNLOCK(source);
1329 new_object->backing_object_offset =
1330 orig_object->backing_object_offset + entry->offset;
1331 new_object->backing_object = source;
1333 if (orig_object->cred != NULL) {
1334 new_object->cred = orig_object->cred;
1335 crhold(orig_object->cred);
1336 new_object->charge = ptoa(size);
1337 KASSERT(orig_object->charge >= ptoa(size),
1338 ("orig_object->charge < 0"));
1339 orig_object->charge -= ptoa(size);
1342 m = vm_page_find_least(orig_object, offidxstart);
1343 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1345 m_next = TAILQ_NEXT(m, listq);
1348 * We must wait for pending I/O to complete before we can
1351 * We do not have to VM_PROT_NONE the page as mappings should
1352 * not be changed by this operation.
1354 if (vm_page_busied(m)) {
1355 VM_OBJECT_WUNLOCK(new_object);
1357 VM_OBJECT_WUNLOCK(orig_object);
1358 vm_page_busy_sleep(m, "spltwt");
1359 VM_OBJECT_WLOCK(orig_object);
1360 VM_OBJECT_WLOCK(new_object);
1364 /* vm_page_rename() will handle dirty and cache. */
1365 if (vm_page_rename(m, new_object, idx)) {
1366 VM_OBJECT_WUNLOCK(new_object);
1367 VM_OBJECT_WUNLOCK(orig_object);
1369 VM_OBJECT_WLOCK(orig_object);
1370 VM_OBJECT_WLOCK(new_object);
1373 #if VM_NRESERVLEVEL > 0
1375 * If some of the reservation's allocated pages remain with
1376 * the original object, then transferring the reservation to
1377 * the new object is neither particularly beneficial nor
1378 * particularly harmful as compared to leaving the reservation
1379 * with the original object. If, however, all of the
1380 * reservation's allocated pages are transferred to the new
1381 * object, then transferring the reservation is typically
1382 * beneficial. Determining which of these two cases applies
1383 * would be more costly than unconditionally renaming the
1386 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1388 if (orig_object->type == OBJT_SWAP)
1391 if (orig_object->type == OBJT_SWAP) {
1393 * swap_pager_copy() can sleep, in which case the orig_object's
1394 * and new_object's locks are released and reacquired.
1396 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1397 TAILQ_FOREACH(m, &new_object->memq, listq)
1401 * Transfer any cached pages from orig_object to new_object.
1402 * If swap_pager_copy() found swapped out pages within the
1403 * specified range of orig_object, then it changed
1404 * new_object's type to OBJT_SWAP when it transferred those
1405 * pages to new_object. Otherwise, new_object's type
1406 * should still be OBJT_DEFAULT and orig_object should not
1407 * contain any cached pages within the specified range.
1409 if (__predict_false(!vm_object_cache_is_empty(orig_object)))
1410 vm_page_cache_transfer(orig_object, offidxstart,
1413 VM_OBJECT_WUNLOCK(orig_object);
1414 VM_OBJECT_WUNLOCK(new_object);
1415 entry->object.vm_object = new_object;
1416 entry->offset = 0LL;
1417 vm_object_deallocate(orig_object);
1418 VM_OBJECT_WLOCK(new_object);
1421 #define OBSC_TEST_ALL_SHADOWED 0x0001
1422 #define OBSC_COLLAPSE_NOWAIT 0x0002
1423 #define OBSC_COLLAPSE_WAIT 0x0004
1426 vm_object_backing_scan(vm_object_t object, int op)
1430 vm_object_t backing_object;
1431 vm_pindex_t backing_offset_index;
1433 VM_OBJECT_ASSERT_WLOCKED(object);
1434 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1436 backing_object = object->backing_object;
1437 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1440 * Initial conditions
1442 if (op & OBSC_TEST_ALL_SHADOWED) {
1444 * We do not want to have to test for the existence of cache
1445 * or swap pages in the backing object. XXX but with the
1446 * new swapper this would be pretty easy to do.
1448 * XXX what about anonymous MAP_SHARED memory that hasn't
1449 * been ZFOD faulted yet? If we do not test for this, the
1450 * shadow test may succeed! XXX
1452 if (backing_object->type != OBJT_DEFAULT) {
1456 if (op & OBSC_COLLAPSE_WAIT) {
1457 vm_object_set_flag(backing_object, OBJ_DEAD);
1463 p = TAILQ_FIRST(&backing_object->memq);
1465 vm_page_t next = TAILQ_NEXT(p, listq);
1466 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1468 if (op & OBSC_TEST_ALL_SHADOWED) {
1472 * Ignore pages outside the parent object's range
1473 * and outside the parent object's mapping of the
1476 * note that we do not busy the backing object's
1480 p->pindex < backing_offset_index ||
1481 new_pindex >= object->size
1488 * See if the parent has the page or if the parent's
1489 * object pager has the page. If the parent has the
1490 * page but the page is not valid, the parent's
1491 * object pager must have the page.
1493 * If this fails, the parent does not completely shadow
1494 * the object and we might as well give up now.
1497 pp = vm_page_lookup(object, new_pindex);
1499 (pp == NULL || pp->valid == 0) &&
1500 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1508 * Check for busy page
1510 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1513 if (op & OBSC_COLLAPSE_NOWAIT) {
1514 if (!p->valid || vm_page_busied(p)) {
1518 } else if (op & OBSC_COLLAPSE_WAIT) {
1519 if (vm_page_busied(p)) {
1520 VM_OBJECT_WUNLOCK(object);
1522 VM_OBJECT_WUNLOCK(backing_object);
1523 vm_page_busy_sleep(p, "vmocol");
1524 VM_OBJECT_WLOCK(object);
1525 VM_OBJECT_WLOCK(backing_object);
1527 * If we slept, anything could have
1528 * happened. Since the object is
1529 * marked dead, the backing offset
1530 * should not have changed so we
1531 * just restart our scan.
1533 p = TAILQ_FIRST(&backing_object->memq);
1539 p->object == backing_object,
1540 ("vm_object_backing_scan: object mismatch")
1544 p->pindex < backing_offset_index ||
1545 new_pindex >= object->size
1547 if (backing_object->type == OBJT_SWAP)
1548 swap_pager_freespace(backing_object,
1552 * Page is out of the parent object's range, we
1553 * can simply destroy it.
1556 KASSERT(!pmap_page_is_mapped(p),
1557 ("freeing mapped page %p", p));
1558 if (p->wire_count == 0)
1567 pp = vm_page_lookup(object, new_pindex);
1569 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1570 (pp != NULL && pp->valid == 0)
1572 if (backing_object->type == OBJT_SWAP)
1573 swap_pager_freespace(backing_object,
1577 * The page in the parent is not (yet) valid.
1578 * We don't know anything about the state of
1579 * the original page. It might be mapped,
1580 * so we must avoid the next if here.
1582 * This is due to a race in vm_fault() where
1583 * we must unbusy the original (backing_obj)
1584 * page before we can (re)lock the parent.
1585 * Hence we can get here.
1592 vm_pager_has_page(object, new_pindex, NULL, NULL)
1594 if (backing_object->type == OBJT_SWAP)
1595 swap_pager_freespace(backing_object,
1599 * page already exists in parent OR swap exists
1600 * for this location in the parent. Destroy
1601 * the original page from the backing object.
1603 * Leave the parent's page alone
1606 KASSERT(!pmap_page_is_mapped(p),
1607 ("freeing mapped page %p", p));
1608 if (p->wire_count == 0)
1618 * Page does not exist in parent, rename the
1619 * page from the backing object to the main object.
1621 * If the page was mapped to a process, it can remain
1622 * mapped through the rename.
1623 * vm_page_rename() will handle dirty and cache.
1625 if (vm_page_rename(p, object, new_pindex)) {
1626 if (op & OBSC_COLLAPSE_NOWAIT) {
1630 VM_OBJECT_WLOCK(backing_object);
1631 VM_OBJECT_WUNLOCK(object);
1633 VM_OBJECT_WLOCK(object);
1634 VM_OBJECT_WLOCK(backing_object);
1635 p = TAILQ_FIRST(&backing_object->memq);
1639 /* Use the old pindex to free the right page. */
1640 if (backing_object->type == OBJT_SWAP)
1641 swap_pager_freespace(backing_object,
1642 new_pindex + backing_offset_index, 1);
1644 #if VM_NRESERVLEVEL > 0
1646 * Rename the reservation.
1648 vm_reserv_rename(p, object, backing_object,
1649 backing_offset_index);
1659 * this version of collapse allows the operation to occur earlier and
1660 * when paging_in_progress is true for an object... This is not a complete
1661 * operation, but should plug 99.9% of the rest of the leaks.
1664 vm_object_qcollapse(vm_object_t object)
1666 vm_object_t backing_object = object->backing_object;
1668 VM_OBJECT_ASSERT_WLOCKED(object);
1669 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1671 if (backing_object->ref_count != 1)
1674 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1678 * vm_object_collapse:
1680 * Collapse an object with the object backing it.
1681 * Pages in the backing object are moved into the
1682 * parent, and the backing object is deallocated.
1685 vm_object_collapse(vm_object_t object)
1687 VM_OBJECT_ASSERT_WLOCKED(object);
1690 vm_object_t backing_object;
1693 * Verify that the conditions are right for collapse:
1695 * The object exists and the backing object exists.
1697 if ((backing_object = object->backing_object) == NULL)
1701 * we check the backing object first, because it is most likely
1704 VM_OBJECT_WLOCK(backing_object);
1705 if (backing_object->handle != NULL ||
1706 (backing_object->type != OBJT_DEFAULT &&
1707 backing_object->type != OBJT_SWAP) ||
1708 (backing_object->flags & OBJ_DEAD) ||
1709 object->handle != NULL ||
1710 (object->type != OBJT_DEFAULT &&
1711 object->type != OBJT_SWAP) ||
1712 (object->flags & OBJ_DEAD)) {
1713 VM_OBJECT_WUNLOCK(backing_object);
1718 object->paging_in_progress != 0 ||
1719 backing_object->paging_in_progress != 0
1721 vm_object_qcollapse(object);
1722 VM_OBJECT_WUNLOCK(backing_object);
1726 * We know that we can either collapse the backing object (if
1727 * the parent is the only reference to it) or (perhaps) have
1728 * the parent bypass the object if the parent happens to shadow
1729 * all the resident pages in the entire backing object.
1731 * This is ignoring pager-backed pages such as swap pages.
1732 * vm_object_backing_scan fails the shadowing test in this
1735 if (backing_object->ref_count == 1) {
1737 * If there is exactly one reference to the backing
1738 * object, we can collapse it into the parent.
1740 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1742 #if VM_NRESERVLEVEL > 0
1744 * Break any reservations from backing_object.
1746 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1747 vm_reserv_break_all(backing_object);
1751 * Move the pager from backing_object to object.
1753 if (backing_object->type == OBJT_SWAP) {
1755 * swap_pager_copy() can sleep, in which case
1756 * the backing_object's and object's locks are
1757 * released and reacquired.
1758 * Since swap_pager_copy() is being asked to
1759 * destroy the source, it will change the
1760 * backing_object's type to OBJT_DEFAULT.
1765 OFF_TO_IDX(object->backing_object_offset), TRUE);
1768 * Free any cached pages from backing_object.
1770 if (__predict_false(
1771 !vm_object_cache_is_empty(backing_object)))
1772 vm_page_cache_free(backing_object, 0, 0);
1775 * Object now shadows whatever backing_object did.
1776 * Note that the reference to
1777 * backing_object->backing_object moves from within
1778 * backing_object to within object.
1780 LIST_REMOVE(object, shadow_list);
1781 backing_object->shadow_count--;
1782 if (backing_object->backing_object) {
1783 VM_OBJECT_WLOCK(backing_object->backing_object);
1784 LIST_REMOVE(backing_object, shadow_list);
1786 &backing_object->backing_object->shadow_head,
1787 object, shadow_list);
1789 * The shadow_count has not changed.
1791 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1793 object->backing_object = backing_object->backing_object;
1794 object->backing_object_offset +=
1795 backing_object->backing_object_offset;
1798 * Discard backing_object.
1800 * Since the backing object has no pages, no pager left,
1801 * and no object references within it, all that is
1802 * necessary is to dispose of it.
1804 KASSERT(backing_object->ref_count == 1, (
1805 "backing_object %p was somehow re-referenced during collapse!",
1807 VM_OBJECT_WUNLOCK(backing_object);
1808 vm_object_destroy(backing_object);
1812 vm_object_t new_backing_object;
1815 * If we do not entirely shadow the backing object,
1816 * there is nothing we can do so we give up.
1818 if (object->resident_page_count != object->size &&
1819 vm_object_backing_scan(object,
1820 OBSC_TEST_ALL_SHADOWED) == 0) {
1821 VM_OBJECT_WUNLOCK(backing_object);
1826 * Make the parent shadow the next object in the
1827 * chain. Deallocating backing_object will not remove
1828 * it, since its reference count is at least 2.
1830 LIST_REMOVE(object, shadow_list);
1831 backing_object->shadow_count--;
1833 new_backing_object = backing_object->backing_object;
1834 if ((object->backing_object = new_backing_object) != NULL) {
1835 VM_OBJECT_WLOCK(new_backing_object);
1837 &new_backing_object->shadow_head,
1841 new_backing_object->shadow_count++;
1842 vm_object_reference_locked(new_backing_object);
1843 VM_OBJECT_WUNLOCK(new_backing_object);
1844 object->backing_object_offset +=
1845 backing_object->backing_object_offset;
1849 * Drop the reference count on backing_object. Since
1850 * its ref_count was at least 2, it will not vanish.
1852 backing_object->ref_count--;
1853 VM_OBJECT_WUNLOCK(backing_object);
1858 * Try again with this object's new backing object.
1864 * vm_object_page_remove:
1866 * For the given object, either frees or invalidates each of the
1867 * specified pages. In general, a page is freed. However, if a page is
1868 * wired for any reason other than the existence of a managed, wired
1869 * mapping, then it may be invalidated but not removed from the object.
1870 * Pages are specified by the given range ["start", "end") and the option
1871 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1872 * extends from "start" to the end of the object. If the option
1873 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1874 * specified range are affected. If the option OBJPR_NOTMAPPED is
1875 * specified, then the pages within the specified range must have no
1876 * mappings. Otherwise, if this option is not specified, any mappings to
1877 * the specified pages are removed before the pages are freed or
1880 * In general, this operation should only be performed on objects that
1881 * contain managed pages. There are, however, two exceptions. First, it
1882 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1883 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1884 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1885 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1887 * The object must be locked.
1890 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1896 VM_OBJECT_ASSERT_WLOCKED(object);
1897 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1898 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1899 ("vm_object_page_remove: illegal options for object %p", object));
1900 if (object->resident_page_count == 0)
1902 vm_object_pip_add(object, 1);
1904 p = vm_page_find_least(object, start);
1907 * Here, the variable "p" is either (1) the page with the least pindex
1908 * greater than or equal to the parameter "start" or (2) NULL.
1910 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1911 next = TAILQ_NEXT(p, listq);
1914 * If the page is wired for any reason besides the existence
1915 * of managed, wired mappings, then it cannot be freed. For
1916 * example, fictitious pages, which represent device memory,
1917 * are inherently wired and cannot be freed. They can,
1918 * however, be invalidated if the option OBJPR_CLEANONLY is
1922 if (vm_page_xbusied(p)) {
1923 VM_OBJECT_WUNLOCK(object);
1924 vm_page_busy_sleep(p, "vmopax");
1925 VM_OBJECT_WLOCK(object);
1928 if ((wirings = p->wire_count) != 0 &&
1929 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1930 if ((options & (OBJPR_NOTWIRED | OBJPR_NOTMAPPED)) ==
1933 /* Account for removal of wired mappings. */
1935 p->wire_count -= wirings;
1937 if ((options & OBJPR_CLEANONLY) == 0) {
1943 if (vm_page_busied(p)) {
1944 VM_OBJECT_WUNLOCK(object);
1945 vm_page_busy_sleep(p, "vmopar");
1946 VM_OBJECT_WLOCK(object);
1949 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1950 ("vm_object_page_remove: page %p is fictitious", p));
1951 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1952 if ((options & OBJPR_NOTMAPPED) == 0)
1953 pmap_remove_write(p);
1957 if ((options & OBJPR_NOTMAPPED) == 0) {
1958 if ((options & OBJPR_NOTWIRED) != 0 && wirings != 0)
1961 /* Account for removal of wired mappings. */
1963 KASSERT(p->wire_count == wirings,
1964 ("inconsistent wire count %d %d %p",
1965 p->wire_count, wirings, p));
1967 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1974 vm_object_pip_wakeup(object);
1976 if (__predict_false(!vm_object_cache_is_empty(object)))
1977 vm_page_cache_free(object, start, end);
1981 * vm_object_page_cache:
1983 * For the given object, attempt to move the specified clean
1984 * pages to the cache queue. If a page is wired for any reason,
1985 * then it will not be changed. Pages are specified by the given
1986 * range ["start", "end"). As a special case, if "end" is zero,
1987 * then the range extends from "start" to the end of the object.
1988 * Any mappings to the specified pages are removed before the
1989 * pages are moved to the cache queue.
1991 * This operation should only be performed on objects that
1992 * contain non-fictitious, managed pages.
1994 * The object must be locked.
1997 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1999 struct mtx *mtx, *new_mtx;
2002 VM_OBJECT_ASSERT_WLOCKED(object);
2003 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2004 ("vm_object_page_cache: illegal object %p", object));
2005 if (object->resident_page_count == 0)
2007 p = vm_page_find_least(object, start);
2010 * Here, the variable "p" is either (1) the page with the least pindex
2011 * greater than or equal to the parameter "start" or (2) NULL.
2014 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2015 next = TAILQ_NEXT(p, listq);
2018 * Avoid releasing and reacquiring the same page lock.
2020 new_mtx = vm_page_lockptr(p);
2021 if (mtx != new_mtx) {
2027 vm_page_try_to_cache(p);
2034 * Populate the specified range of the object with valid pages. Returns
2035 * TRUE if the range is successfully populated and FALSE otherwise.
2037 * Note: This function should be optimized to pass a larger array of
2038 * pages to vm_pager_get_pages() before it is applied to a non-
2039 * OBJT_DEVICE object.
2041 * The object must be locked.
2044 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2050 VM_OBJECT_ASSERT_WLOCKED(object);
2051 for (pindex = start; pindex < end; pindex++) {
2052 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
2053 if (m->valid != VM_PAGE_BITS_ALL) {
2055 rv = vm_pager_get_pages(object, ma, 1, 0);
2056 m = vm_page_lookup(object, pindex);
2059 if (rv != VM_PAGER_OK) {
2067 * Keep "m" busy because a subsequent iteration may unlock
2071 if (pindex > start) {
2072 m = vm_page_lookup(object, start);
2073 while (m != NULL && m->pindex < pindex) {
2075 m = TAILQ_NEXT(m, listq);
2078 return (pindex == end);
2082 * Routine: vm_object_coalesce
2083 * Function: Coalesces two objects backing up adjoining
2084 * regions of memory into a single object.
2086 * returns TRUE if objects were combined.
2088 * NOTE: Only works at the moment if the second object is NULL -
2089 * if it's not, which object do we lock first?
2092 * prev_object First object to coalesce
2093 * prev_offset Offset into prev_object
2094 * prev_size Size of reference to prev_object
2095 * next_size Size of reference to the second object
2096 * reserved Indicator that extension region has
2097 * swap accounted for
2100 * The object must *not* be locked.
2103 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2104 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2106 vm_pindex_t next_pindex;
2108 if (prev_object == NULL)
2110 VM_OBJECT_WLOCK(prev_object);
2111 if ((prev_object->type != OBJT_DEFAULT &&
2112 prev_object->type != OBJT_SWAP) ||
2113 (prev_object->flags & OBJ_TMPFS_NODE) != 0) {
2114 VM_OBJECT_WUNLOCK(prev_object);
2119 * Try to collapse the object first
2121 vm_object_collapse(prev_object);
2124 * Can't coalesce if: . more than one reference . paged out . shadows
2125 * another object . has a copy elsewhere (any of which mean that the
2126 * pages not mapped to prev_entry may be in use anyway)
2128 if (prev_object->backing_object != NULL) {
2129 VM_OBJECT_WUNLOCK(prev_object);
2133 prev_size >>= PAGE_SHIFT;
2134 next_size >>= PAGE_SHIFT;
2135 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2137 if ((prev_object->ref_count > 1) &&
2138 (prev_object->size != next_pindex)) {
2139 VM_OBJECT_WUNLOCK(prev_object);
2144 * Account for the charge.
2146 if (prev_object->cred != NULL) {
2149 * If prev_object was charged, then this mapping,
2150 * althought not charged now, may become writable
2151 * later. Non-NULL cred in the object would prevent
2152 * swap reservation during enabling of the write
2153 * access, so reserve swap now. Failed reservation
2154 * cause allocation of the separate object for the map
2155 * entry, and swap reservation for this entry is
2156 * managed in appropriate time.
2158 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2159 prev_object->cred)) {
2162 prev_object->charge += ptoa(next_size);
2166 * Remove any pages that may still be in the object from a previous
2169 if (next_pindex < prev_object->size) {
2170 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2172 if (prev_object->type == OBJT_SWAP)
2173 swap_pager_freespace(prev_object,
2174 next_pindex, next_size);
2176 if (prev_object->cred != NULL) {
2177 KASSERT(prev_object->charge >=
2178 ptoa(prev_object->size - next_pindex),
2179 ("object %p overcharged 1 %jx %jx", prev_object,
2180 (uintmax_t)next_pindex, (uintmax_t)next_size));
2181 prev_object->charge -= ptoa(prev_object->size -
2188 * Extend the object if necessary.
2190 if (next_pindex + next_size > prev_object->size)
2191 prev_object->size = next_pindex + next_size;
2193 VM_OBJECT_WUNLOCK(prev_object);
2198 vm_object_set_writeable_dirty(vm_object_t object)
2201 VM_OBJECT_ASSERT_WLOCKED(object);
2202 if (object->type != OBJT_VNODE)
2204 object->generation++;
2205 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2207 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2213 * For each page offset within the specified range of the given object,
2214 * find the highest-level page in the shadow chain and unwire it. A page
2215 * must exist at every page offset, and the highest-level page must be
2219 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2222 vm_object_t tobject;
2224 vm_pindex_t end_pindex, pindex, tpindex;
2225 int depth, locked_depth;
2227 KASSERT((offset & PAGE_MASK) == 0,
2228 ("vm_object_unwire: offset is not page aligned"));
2229 KASSERT((length & PAGE_MASK) == 0,
2230 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2231 /* The wired count of a fictitious page never changes. */
2232 if ((object->flags & OBJ_FICTITIOUS) != 0)
2234 pindex = OFF_TO_IDX(offset);
2235 end_pindex = pindex + atop(length);
2237 VM_OBJECT_RLOCK(object);
2238 m = vm_page_find_least(object, pindex);
2239 while (pindex < end_pindex) {
2240 if (m == NULL || pindex < m->pindex) {
2242 * The first object in the shadow chain doesn't
2243 * contain a page at the current index. Therefore,
2244 * the page must exist in a backing object.
2251 OFF_TO_IDX(tobject->backing_object_offset);
2252 tobject = tobject->backing_object;
2253 KASSERT(tobject != NULL,
2254 ("vm_object_unwire: missing page"));
2255 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2258 if (depth == locked_depth) {
2260 VM_OBJECT_RLOCK(tobject);
2262 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2266 m = TAILQ_NEXT(m, listq);
2269 vm_page_unwire(tm, queue);
2274 /* Release the accumulated object locks. */
2275 for (depth = 0; depth < locked_depth; depth++) {
2276 tobject = object->backing_object;
2277 VM_OBJECT_RUNLOCK(object);
2282 #include "opt_ddb.h"
2284 #include <sys/kernel.h>
2286 #include <sys/cons.h>
2288 #include <ddb/ddb.h>
2291 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2294 vm_map_entry_t tmpe;
2302 tmpe = map->header.next;
2303 entcount = map->nentries;
2304 while (entcount-- && (tmpe != &map->header)) {
2305 if (_vm_object_in_map(map, object, tmpe)) {
2310 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2311 tmpm = entry->object.sub_map;
2312 tmpe = tmpm->header.next;
2313 entcount = tmpm->nentries;
2314 while (entcount-- && tmpe != &tmpm->header) {
2315 if (_vm_object_in_map(tmpm, object, tmpe)) {
2320 } else if ((obj = entry->object.vm_object) != NULL) {
2321 for (; obj; obj = obj->backing_object)
2322 if (obj == object) {
2330 vm_object_in_map(vm_object_t object)
2334 /* sx_slock(&allproc_lock); */
2335 FOREACH_PROC_IN_SYSTEM(p) {
2336 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2338 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2339 /* sx_sunlock(&allproc_lock); */
2343 /* sx_sunlock(&allproc_lock); */
2344 if (_vm_object_in_map(kernel_map, object, 0))
2349 DB_SHOW_COMMAND(vmochk, vm_object_check)
2354 * make sure that internal objs are in a map somewhere
2355 * and none have zero ref counts.
2357 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2358 if (object->handle == NULL &&
2359 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2360 if (object->ref_count == 0) {
2361 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2362 (long)object->size);
2364 if (!vm_object_in_map(object)) {
2366 "vmochk: internal obj is not in a map: "
2367 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2368 object->ref_count, (u_long)object->size,
2369 (u_long)object->size,
2370 (void *)object->backing_object);
2377 * vm_object_print: [ debug ]
2379 DB_SHOW_COMMAND(object, vm_object_print_static)
2381 /* XXX convert args. */
2382 vm_object_t object = (vm_object_t)addr;
2383 boolean_t full = have_addr;
2387 /* XXX count is an (unused) arg. Avoid shadowing it. */
2388 #define count was_count
2396 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2397 object, (int)object->type, (uintmax_t)object->size,
2398 object->resident_page_count, object->ref_count, object->flags,
2399 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2400 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2401 object->shadow_count,
2402 object->backing_object ? object->backing_object->ref_count : 0,
2403 object->backing_object, (uintmax_t)object->backing_object_offset);
2410 TAILQ_FOREACH(p, &object->memq, listq) {
2412 db_iprintf("memory:=");
2413 else if (count == 6) {
2421 db_printf("(off=0x%jx,page=0x%jx)",
2422 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2432 /* XXX need this non-static entry for calling from vm_map_print. */
2435 /* db_expr_t */ long addr,
2436 boolean_t have_addr,
2437 /* db_expr_t */ long count,
2440 vm_object_print_static(addr, have_addr, count, modif);
2443 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2448 vm_page_t m, prev_m;
2452 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2453 db_printf("new object: %p\n", (void *)object);
2464 TAILQ_FOREACH(m, &object->memq, listq) {
2465 if (m->pindex > 128)
2467 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2468 prev_m->pindex + 1 != m->pindex) {
2470 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2471 (long)fidx, rcount, (long)pa);
2483 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2488 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2489 (long)fidx, rcount, (long)pa);
2499 pa = VM_PAGE_TO_PHYS(m);
2503 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2504 (long)fidx, rcount, (long)pa);