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(object->ref_count == 0,
170 ("object %p ref_count = %d", object, object->ref_count));
171 KASSERT(TAILQ_EMPTY(&object->memq),
172 ("object %p has resident pages in its memq", object));
173 KASSERT(vm_radix_is_empty(&object->rtree),
174 ("object %p has resident pages in its trie", object));
175 #if VM_NRESERVLEVEL > 0
176 KASSERT(LIST_EMPTY(&object->rvq),
177 ("object %p has reservations",
180 KASSERT(vm_object_cache_is_empty(object),
181 ("object %p has cached pages",
183 KASSERT(object->paging_in_progress == 0,
184 ("object %p paging_in_progress = %d",
185 object, object->paging_in_progress));
186 KASSERT(object->resident_page_count == 0,
187 ("object %p resident_page_count = %d",
188 object, object->resident_page_count));
189 KASSERT(object->shadow_count == 0,
190 ("object %p shadow_count = %d",
191 object, object->shadow_count));
192 KASSERT(object->type == OBJT_DEAD,
193 ("object %p has non-dead type %d",
194 object, object->type));
199 vm_object_zinit(void *mem, int size, int flags)
203 object = (vm_object_t)mem;
204 bzero(&object->lock, sizeof(object->lock));
205 rw_init_flags(&object->lock, "vm object", RW_DUPOK);
207 /* These are true for any object that has been freed */
208 object->type = OBJT_DEAD;
209 object->ref_count = 0;
210 object->rtree.rt_root = 0;
211 object->rtree.rt_flags = 0;
212 object->paging_in_progress = 0;
213 object->resident_page_count = 0;
214 object->shadow_count = 0;
215 object->cache.rt_root = 0;
216 object->cache.rt_flags = 0;
218 mtx_lock(&vm_object_list_mtx);
219 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
220 mtx_unlock(&vm_object_list_mtx);
225 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
228 TAILQ_INIT(&object->memq);
229 LIST_INIT(&object->shadow_head);
234 panic("_vm_object_allocate: can't create OBJT_DEAD");
237 object->flags = OBJ_ONEMAPPING;
241 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
244 object->flags = OBJ_FICTITIOUS;
247 object->flags = OBJ_UNMANAGED;
253 panic("_vm_object_allocate: type %d is undefined", type);
256 object->generation = 1;
257 object->ref_count = 1;
258 object->memattr = VM_MEMATTR_DEFAULT;
261 object->handle = NULL;
262 object->backing_object = NULL;
263 object->backing_object_offset = (vm_ooffset_t) 0;
264 #if VM_NRESERVLEVEL > 0
265 LIST_INIT(&object->rvq);
272 * Initialize the VM objects module.
277 TAILQ_INIT(&vm_object_list);
278 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
280 rw_init(&kernel_object->lock, "kernel vm object");
281 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
283 #if VM_NRESERVLEVEL > 0
284 kernel_object->flags |= OBJ_COLORED;
285 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
288 rw_init(&kmem_object->lock, "kmem vm object");
289 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
291 #if VM_NRESERVLEVEL > 0
292 kmem_object->flags |= OBJ_COLORED;
293 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
297 * The lock portion of struct vm_object must be type stable due
298 * to vm_pageout_fallback_object_lock locking a vm object
299 * without holding any references to it.
301 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
307 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
313 vm_object_clear_flag(vm_object_t object, u_short bits)
316 VM_OBJECT_ASSERT_WLOCKED(object);
317 object->flags &= ~bits;
321 * Sets the default memory attribute for the specified object. Pages
322 * that are allocated to this object are by default assigned this memory
325 * Presently, this function must be called before any pages are allocated
326 * to the object. In the future, this requirement may be relaxed for
327 * "default" and "swap" objects.
330 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
333 VM_OBJECT_ASSERT_WLOCKED(object);
334 switch (object->type) {
342 if (!TAILQ_EMPTY(&object->memq))
343 return (KERN_FAILURE);
346 return (KERN_INVALID_ARGUMENT);
348 panic("vm_object_set_memattr: object %p is of undefined type",
351 object->memattr = memattr;
352 return (KERN_SUCCESS);
356 vm_object_pip_add(vm_object_t object, short i)
359 VM_OBJECT_ASSERT_WLOCKED(object);
360 object->paging_in_progress += i;
364 vm_object_pip_subtract(vm_object_t object, short i)
367 VM_OBJECT_ASSERT_WLOCKED(object);
368 object->paging_in_progress -= i;
372 vm_object_pip_wakeup(vm_object_t object)
375 VM_OBJECT_ASSERT_WLOCKED(object);
376 object->paging_in_progress--;
377 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
378 vm_object_clear_flag(object, OBJ_PIPWNT);
384 vm_object_pip_wakeupn(vm_object_t object, short i)
387 VM_OBJECT_ASSERT_WLOCKED(object);
389 object->paging_in_progress -= i;
390 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
391 vm_object_clear_flag(object, OBJ_PIPWNT);
397 vm_object_pip_wait(vm_object_t object, char *waitid)
400 VM_OBJECT_ASSERT_WLOCKED(object);
401 while (object->paging_in_progress) {
402 object->flags |= OBJ_PIPWNT;
403 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
408 * vm_object_allocate:
410 * Returns a new object with the given size.
413 vm_object_allocate(objtype_t type, vm_pindex_t size)
417 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
418 _vm_object_allocate(type, size, object);
424 * vm_object_reference:
426 * Gets another reference to the given object. Note: OBJ_DEAD
427 * objects can be referenced during final cleaning.
430 vm_object_reference(vm_object_t object)
434 VM_OBJECT_WLOCK(object);
435 vm_object_reference_locked(object);
436 VM_OBJECT_WUNLOCK(object);
440 * vm_object_reference_locked:
442 * Gets another reference to the given object.
444 * The object must be locked.
447 vm_object_reference_locked(vm_object_t object)
451 VM_OBJECT_ASSERT_WLOCKED(object);
453 if (object->type == OBJT_VNODE) {
460 * Handle deallocating an object of type OBJT_VNODE.
463 vm_object_vndeallocate(vm_object_t object)
465 struct vnode *vp = (struct vnode *) object->handle;
467 VM_OBJECT_ASSERT_WLOCKED(object);
468 KASSERT(object->type == OBJT_VNODE,
469 ("vm_object_vndeallocate: not a vnode object"));
470 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
472 if (object->ref_count == 0) {
473 vprint("vm_object_vndeallocate", vp);
474 panic("vm_object_vndeallocate: bad object reference count");
479 * The test for text of vp vnode does not need a bypass to
480 * reach right VV_TEXT there, since it is obtained from
483 if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) {
485 VM_OBJECT_WUNLOCK(object);
486 /* vrele may need the vnode lock. */
490 VM_OBJECT_WUNLOCK(object);
491 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
493 VM_OBJECT_WLOCK(object);
495 if (object->type == OBJT_DEAD) {
496 VM_OBJECT_WUNLOCK(object);
499 if (object->ref_count == 0)
501 VM_OBJECT_WUNLOCK(object);
508 * vm_object_deallocate:
510 * Release a reference to the specified object,
511 * gained either through a vm_object_allocate
512 * or a vm_object_reference call. When all references
513 * are gone, storage associated with this object
514 * may be relinquished.
516 * No object may be locked.
519 vm_object_deallocate(vm_object_t object)
524 while (object != NULL) {
525 VM_OBJECT_WLOCK(object);
526 if (object->type == OBJT_VNODE) {
527 vm_object_vndeallocate(object);
531 KASSERT(object->ref_count != 0,
532 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
535 * If the reference count goes to 0 we start calling
536 * vm_object_terminate() on the object chain.
537 * A ref count of 1 may be a special case depending on the
538 * shadow count being 0 or 1.
541 if (object->ref_count > 1) {
542 VM_OBJECT_WUNLOCK(object);
544 } else if (object->ref_count == 1) {
545 if (object->type == OBJT_SWAP &&
546 (object->flags & OBJ_TMPFS) != 0) {
547 vp = object->un_pager.swp.swp_tmpfs;
549 VM_OBJECT_WUNLOCK(object);
550 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
551 VM_OBJECT_WLOCK(object);
552 if (object->type == OBJT_DEAD ||
553 object->ref_count != 1) {
554 VM_OBJECT_WUNLOCK(object);
559 if ((object->flags & OBJ_TMPFS) != 0)
564 if (object->shadow_count == 0 &&
565 object->handle == NULL &&
566 (object->type == OBJT_DEFAULT ||
567 (object->type == OBJT_SWAP &&
568 (object->flags & OBJ_TMPFS_NODE) == 0))) {
569 vm_object_set_flag(object, OBJ_ONEMAPPING);
570 } else if ((object->shadow_count == 1) &&
571 (object->handle == NULL) &&
572 (object->type == OBJT_DEFAULT ||
573 object->type == OBJT_SWAP)) {
576 robject = LIST_FIRST(&object->shadow_head);
577 KASSERT(robject != NULL,
578 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
580 object->shadow_count));
581 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0,
582 ("shadowed tmpfs v_object %p", object));
583 if (!VM_OBJECT_TRYWLOCK(robject)) {
585 * Avoid a potential deadlock.
588 VM_OBJECT_WUNLOCK(object);
590 * More likely than not the thread
591 * holding robject's lock has lower
592 * priority than the current thread.
593 * Let the lower priority thread run.
599 * Collapse object into its shadow unless its
600 * shadow is dead. In that case, object will
601 * be deallocated by the thread that is
602 * deallocating its shadow.
604 if ((robject->flags & OBJ_DEAD) == 0 &&
605 (robject->handle == NULL) &&
606 (robject->type == OBJT_DEFAULT ||
607 robject->type == OBJT_SWAP)) {
609 robject->ref_count++;
611 if (robject->paging_in_progress) {
612 VM_OBJECT_WUNLOCK(object);
613 vm_object_pip_wait(robject,
615 temp = robject->backing_object;
616 if (object == temp) {
617 VM_OBJECT_WLOCK(object);
620 } else if (object->paging_in_progress) {
621 VM_OBJECT_WUNLOCK(robject);
622 object->flags |= OBJ_PIPWNT;
623 VM_OBJECT_SLEEP(object, object,
624 PDROP | PVM, "objde2", 0);
625 VM_OBJECT_WLOCK(robject);
626 temp = robject->backing_object;
627 if (object == temp) {
628 VM_OBJECT_WLOCK(object);
632 VM_OBJECT_WUNLOCK(object);
634 if (robject->ref_count == 1) {
635 robject->ref_count--;
640 vm_object_collapse(object);
641 VM_OBJECT_WUNLOCK(object);
644 VM_OBJECT_WUNLOCK(robject);
646 VM_OBJECT_WUNLOCK(object);
650 temp = object->backing_object;
652 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
653 ("shadowed tmpfs v_object 2 %p", object));
654 VM_OBJECT_WLOCK(temp);
655 LIST_REMOVE(object, shadow_list);
656 temp->shadow_count--;
657 VM_OBJECT_WUNLOCK(temp);
658 object->backing_object = NULL;
661 * Don't double-terminate, we could be in a termination
662 * recursion due to the terminate having to sync data
665 if ((object->flags & OBJ_DEAD) == 0)
666 vm_object_terminate(object);
668 VM_OBJECT_WUNLOCK(object);
674 * vm_object_destroy removes the object from the global object list
675 * and frees the space for the object.
678 vm_object_destroy(vm_object_t object)
682 * Release the allocation charge.
684 if (object->cred != NULL) {
685 swap_release_by_cred(object->charge, object->cred);
687 crfree(object->cred);
692 * Free the space for the object.
694 uma_zfree(obj_zone, object);
698 * vm_object_terminate actually destroys the specified object, freeing
699 * up all previously used resources.
701 * The object must be locked.
702 * This routine may block.
705 vm_object_terminate(vm_object_t object)
709 VM_OBJECT_ASSERT_WLOCKED(object);
712 * Make sure no one uses us.
714 vm_object_set_flag(object, OBJ_DEAD);
717 * wait for the pageout daemon to be done with the object
719 vm_object_pip_wait(object, "objtrm");
721 KASSERT(!object->paging_in_progress,
722 ("vm_object_terminate: pageout in progress"));
725 * Clean and free the pages, as appropriate. All references to the
726 * object are gone, so we don't need to lock it.
728 if (object->type == OBJT_VNODE) {
729 struct vnode *vp = (struct vnode *)object->handle;
732 * Clean pages and flush buffers.
734 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
735 VM_OBJECT_WUNLOCK(object);
737 vinvalbuf(vp, V_SAVE, 0, 0);
739 VM_OBJECT_WLOCK(object);
742 KASSERT(object->ref_count == 0,
743 ("vm_object_terminate: object with references, ref_count=%d",
747 * Free any remaining pageable pages. This also removes them from the
748 * paging queues. However, don't free wired pages, just remove them
749 * from the object. Rather than incrementally removing each page from
750 * the object, the page and object are reset to any empty state.
752 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
753 vm_page_assert_unbusied(p);
756 * Optimize the page's removal from the object by resetting
757 * its "object" field. Specifically, if the page is not
758 * wired, then the effect of this assignment is that
759 * vm_page_free()'s call to vm_page_remove() will return
760 * immediately without modifying the page or the object.
763 if (p->wire_count == 0) {
765 PCPU_INC(cnt.v_pfree);
770 * If the object contained any pages, then reset it to an empty state.
771 * None of the object's fields, including "resident_page_count", were
772 * modified by the preceding loop.
774 if (object->resident_page_count != 0) {
775 vm_radix_reclaim_allnodes(&object->rtree);
776 TAILQ_INIT(&object->memq);
777 object->resident_page_count = 0;
778 if (object->type == OBJT_VNODE)
779 vdrop(object->handle);
782 #if VM_NRESERVLEVEL > 0
783 if (__predict_false(!LIST_EMPTY(&object->rvq)))
784 vm_reserv_break_all(object);
786 if (__predict_false(!vm_object_cache_is_empty(object)))
787 vm_page_cache_free(object, 0, 0);
789 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
790 object->type == OBJT_SWAP,
791 ("%s: non-swap obj %p has cred", __func__, object));
794 * Let the pager know object is dead.
796 vm_pager_deallocate(object);
797 VM_OBJECT_WUNLOCK(object);
799 vm_object_destroy(object);
803 * Make the page read-only so that we can clear the object flags. However, if
804 * this is a nosync mmap then the object is likely to stay dirty so do not
805 * mess with the page and do not clear the object flags. Returns TRUE if the
806 * page should be flushed, and FALSE otherwise.
809 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
813 * If we have been asked to skip nosync pages and this is a
814 * nosync page, skip it. Note that the object flags were not
815 * cleared in this case so we do not have to set them.
817 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
818 *clearobjflags = FALSE;
821 pmap_remove_write(p);
822 return (p->dirty != 0);
827 * vm_object_page_clean
829 * Clean all dirty pages in the specified range of object. Leaves page
830 * on whatever queue it is currently on. If NOSYNC is set then do not
831 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
832 * leaving the object dirty.
834 * When stuffing pages asynchronously, allow clustering. XXX we need a
835 * synchronous clustering mode implementation.
837 * Odd semantics: if start == end, we clean everything.
839 * The object must be locked.
841 * Returns FALSE if some page from the range was not written, as
842 * reported by the pager, and TRUE otherwise.
845 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
849 vm_pindex_t pi, tend, tstart;
850 int curgeneration, n, pagerflags;
851 boolean_t clearobjflags, eio, res;
853 VM_OBJECT_ASSERT_WLOCKED(object);
856 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
857 * objects. The check below prevents the function from
858 * operating on non-vnode objects.
860 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
861 object->resident_page_count == 0)
864 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
865 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
866 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
868 tstart = OFF_TO_IDX(start);
869 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
870 clearobjflags = tstart == 0 && tend >= object->size;
874 curgeneration = object->generation;
876 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
880 np = TAILQ_NEXT(p, listq);
883 if (vm_page_sleep_if_busy(p, "vpcwai")) {
884 if (object->generation != curgeneration) {
885 if ((flags & OBJPC_SYNC) != 0)
888 clearobjflags = FALSE;
890 np = vm_page_find_least(object, pi);
893 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
896 n = vm_object_page_collect_flush(object, p, pagerflags,
897 flags, &clearobjflags, &eio);
900 clearobjflags = FALSE;
902 if (object->generation != curgeneration) {
903 if ((flags & OBJPC_SYNC) != 0)
906 clearobjflags = FALSE;
910 * If the VOP_PUTPAGES() did a truncated write, so
911 * that even the first page of the run is not fully
912 * written, vm_pageout_flush() returns 0 as the run
913 * length. Since the condition that caused truncated
914 * write may be permanent, e.g. exhausted free space,
915 * accepting n == 0 would cause an infinite loop.
917 * Forwarding the iterator leaves the unwritten page
918 * behind, but there is not much we can do there if
919 * filesystem refuses to write it.
923 clearobjflags = FALSE;
925 np = vm_page_find_least(object, pi + n);
928 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
932 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
937 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
938 int flags, boolean_t *clearobjflags, boolean_t *eio)
940 vm_page_t ma[vm_pageout_page_count], p_first, tp;
941 int count, i, mreq, runlen;
943 vm_page_lock_assert(p, MA_NOTOWNED);
944 VM_OBJECT_ASSERT_WLOCKED(object);
949 for (tp = p; count < vm_pageout_page_count; count++) {
950 tp = vm_page_next(tp);
951 if (tp == NULL || vm_page_busied(tp))
953 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
957 for (p_first = p; count < vm_pageout_page_count; count++) {
958 tp = vm_page_prev(p_first);
959 if (tp == NULL || vm_page_busied(tp))
961 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
967 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
970 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
975 * Note that there is absolutely no sense in writing out
976 * anonymous objects, so we track down the vnode object
978 * We invalidate (remove) all pages from the address space
979 * for semantic correctness.
981 * If the backing object is a device object with unmanaged pages, then any
982 * mappings to the specified range of pages must be removed before this
983 * function is called.
985 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
986 * may start out with a NULL object.
989 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
990 boolean_t syncio, boolean_t invalidate)
992 vm_object_t backing_object;
995 int error, flags, fsync_after;
1002 VM_OBJECT_WLOCK(object);
1003 while ((backing_object = object->backing_object) != NULL) {
1004 VM_OBJECT_WLOCK(backing_object);
1005 offset += object->backing_object_offset;
1006 VM_OBJECT_WUNLOCK(object);
1007 object = backing_object;
1008 if (object->size < OFF_TO_IDX(offset + size))
1009 size = IDX_TO_OFF(object->size) - offset;
1012 * Flush pages if writing is allowed, invalidate them
1013 * if invalidation requested. Pages undergoing I/O
1014 * will be ignored by vm_object_page_remove().
1016 * We cannot lock the vnode and then wait for paging
1017 * to complete without deadlocking against vm_fault.
1018 * Instead we simply call vm_object_page_remove() and
1019 * allow it to block internally on a page-by-page
1020 * basis when it encounters pages undergoing async
1023 if (object->type == OBJT_VNODE &&
1024 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1025 vp = object->handle;
1026 VM_OBJECT_WUNLOCK(object);
1027 (void) vn_start_write(vp, &mp, V_WAIT);
1028 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1029 if (syncio && !invalidate && offset == 0 &&
1030 OFF_TO_IDX(size) == object->size) {
1032 * If syncing the whole mapping of the file,
1033 * it is faster to schedule all the writes in
1034 * async mode, also allowing the clustering,
1035 * and then wait for i/o to complete.
1040 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1041 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1042 fsync_after = FALSE;
1044 VM_OBJECT_WLOCK(object);
1045 res = vm_object_page_clean(object, offset, offset + size,
1047 VM_OBJECT_WUNLOCK(object);
1049 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1051 vn_finished_write(mp);
1054 VM_OBJECT_WLOCK(object);
1056 if ((object->type == OBJT_VNODE ||
1057 object->type == OBJT_DEVICE) && invalidate) {
1058 if (object->type == OBJT_DEVICE)
1060 * The option OBJPR_NOTMAPPED must be passed here
1061 * because vm_object_page_remove() cannot remove
1062 * unmanaged mappings.
1064 flags = OBJPR_NOTMAPPED;
1068 flags = OBJPR_CLEANONLY;
1069 vm_object_page_remove(object, OFF_TO_IDX(offset),
1070 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1072 VM_OBJECT_WUNLOCK(object);
1077 * vm_object_madvise:
1079 * Implements the madvise function at the object/page level.
1081 * MADV_WILLNEED (any object)
1083 * Activate the specified pages if they are resident.
1085 * MADV_DONTNEED (any object)
1087 * Deactivate the specified pages if they are resident.
1089 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1090 * OBJ_ONEMAPPING only)
1092 * Deactivate and clean the specified pages if they are
1093 * resident. This permits the process to reuse the pages
1094 * without faulting or the kernel to reclaim the pages
1098 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1101 vm_pindex_t tpindex;
1102 vm_object_t backing_object, tobject;
1107 VM_OBJECT_WLOCK(object);
1109 * Locate and adjust resident pages
1111 for (; pindex < end; pindex += 1) {
1117 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1118 * and those pages must be OBJ_ONEMAPPING.
1120 if (advise == MADV_FREE) {
1121 if ((tobject->type != OBJT_DEFAULT &&
1122 tobject->type != OBJT_SWAP) ||
1123 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1124 goto unlock_tobject;
1126 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1127 goto unlock_tobject;
1128 m = vm_page_lookup(tobject, tpindex);
1129 if (m == NULL && advise == MADV_WILLNEED) {
1131 * If the page is cached, reactivate it.
1133 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1138 * There may be swap even if there is no backing page
1140 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1141 swap_pager_freespace(tobject, tpindex, 1);
1145 backing_object = tobject->backing_object;
1146 if (backing_object == NULL)
1147 goto unlock_tobject;
1148 VM_OBJECT_WLOCK(backing_object);
1149 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1150 if (tobject != object)
1151 VM_OBJECT_WUNLOCK(tobject);
1152 tobject = backing_object;
1154 } else if (m->valid != VM_PAGE_BITS_ALL)
1155 goto unlock_tobject;
1157 * If the page is not in a normal state, skip it.
1160 if (m->hold_count != 0 || m->wire_count != 0) {
1162 goto unlock_tobject;
1164 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1165 ("vm_object_madvise: page %p is fictitious", m));
1166 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1167 ("vm_object_madvise: page %p is not managed", m));
1168 if (vm_page_busied(m)) {
1169 if (advise == MADV_WILLNEED) {
1171 * Reference the page before unlocking and
1172 * sleeping so that the page daemon is less
1173 * likely to reclaim it.
1175 vm_page_aflag_set(m, PGA_REFERENCED);
1177 if (object != tobject)
1178 VM_OBJECT_WUNLOCK(object);
1179 VM_OBJECT_WUNLOCK(tobject);
1180 vm_page_busy_sleep(m, "madvpo");
1181 VM_OBJECT_WLOCK(object);
1184 if (advise == MADV_WILLNEED) {
1185 vm_page_activate(m);
1187 vm_page_advise(m, advise);
1190 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1191 swap_pager_freespace(tobject, tpindex, 1);
1193 if (tobject != object)
1194 VM_OBJECT_WUNLOCK(tobject);
1196 VM_OBJECT_WUNLOCK(object);
1202 * Create a new object which is backed by the
1203 * specified existing object range. The source
1204 * object reference is deallocated.
1206 * The new object and offset into that object
1207 * are returned in the source parameters.
1211 vm_object_t *object, /* IN/OUT */
1212 vm_ooffset_t *offset, /* IN/OUT */
1221 * Don't create the new object if the old object isn't shared.
1223 if (source != NULL) {
1224 VM_OBJECT_WLOCK(source);
1225 if (source->ref_count == 1 &&
1226 source->handle == NULL &&
1227 (source->type == OBJT_DEFAULT ||
1228 source->type == OBJT_SWAP)) {
1229 VM_OBJECT_WUNLOCK(source);
1232 VM_OBJECT_WUNLOCK(source);
1236 * Allocate a new object with the given length.
1238 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1241 * The new object shadows the source object, adding a reference to it.
1242 * Our caller changes his reference to point to the new object,
1243 * removing a reference to the source object. Net result: no change
1244 * of reference count.
1246 * Try to optimize the result object's page color when shadowing
1247 * in order to maintain page coloring consistency in the combined
1250 result->backing_object = source;
1252 * Store the offset into the source object, and fix up the offset into
1255 result->backing_object_offset = *offset;
1256 if (source != NULL) {
1257 VM_OBJECT_WLOCK(source);
1258 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1259 source->shadow_count++;
1260 #if VM_NRESERVLEVEL > 0
1261 result->flags |= source->flags & OBJ_COLORED;
1262 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1263 ((1 << (VM_NFREEORDER - 1)) - 1);
1265 VM_OBJECT_WUNLOCK(source);
1270 * Return the new things
1279 * Split the pages in a map entry into a new object. This affords
1280 * easier removal of unused pages, and keeps object inheritance from
1281 * being a negative impact on memory usage.
1284 vm_object_split(vm_map_entry_t entry)
1286 vm_page_t m, m_next;
1287 vm_object_t orig_object, new_object, source;
1288 vm_pindex_t idx, offidxstart;
1291 orig_object = entry->object.vm_object;
1292 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1294 if (orig_object->ref_count <= 1)
1296 VM_OBJECT_WUNLOCK(orig_object);
1298 offidxstart = OFF_TO_IDX(entry->offset);
1299 size = atop(entry->end - entry->start);
1302 * If swap_pager_copy() is later called, it will convert new_object
1303 * into a swap object.
1305 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1308 * At this point, the new object is still private, so the order in
1309 * which the original and new objects are locked does not matter.
1311 VM_OBJECT_WLOCK(new_object);
1312 VM_OBJECT_WLOCK(orig_object);
1313 source = orig_object->backing_object;
1314 if (source != NULL) {
1315 VM_OBJECT_WLOCK(source);
1316 if ((source->flags & OBJ_DEAD) != 0) {
1317 VM_OBJECT_WUNLOCK(source);
1318 VM_OBJECT_WUNLOCK(orig_object);
1319 VM_OBJECT_WUNLOCK(new_object);
1320 vm_object_deallocate(new_object);
1321 VM_OBJECT_WLOCK(orig_object);
1324 LIST_INSERT_HEAD(&source->shadow_head,
1325 new_object, shadow_list);
1326 source->shadow_count++;
1327 vm_object_reference_locked(source); /* for new_object */
1328 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1329 VM_OBJECT_WUNLOCK(source);
1330 new_object->backing_object_offset =
1331 orig_object->backing_object_offset + entry->offset;
1332 new_object->backing_object = source;
1334 if (orig_object->cred != NULL) {
1335 new_object->cred = orig_object->cred;
1336 crhold(orig_object->cred);
1337 new_object->charge = ptoa(size);
1338 KASSERT(orig_object->charge >= ptoa(size),
1339 ("orig_object->charge < 0"));
1340 orig_object->charge -= ptoa(size);
1343 m = vm_page_find_least(orig_object, offidxstart);
1344 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1346 m_next = TAILQ_NEXT(m, listq);
1349 * We must wait for pending I/O to complete before we can
1352 * We do not have to VM_PROT_NONE the page as mappings should
1353 * not be changed by this operation.
1355 if (vm_page_busied(m)) {
1356 VM_OBJECT_WUNLOCK(new_object);
1358 VM_OBJECT_WUNLOCK(orig_object);
1359 vm_page_busy_sleep(m, "spltwt");
1360 VM_OBJECT_WLOCK(orig_object);
1361 VM_OBJECT_WLOCK(new_object);
1365 /* vm_page_rename() will handle dirty and cache. */
1366 if (vm_page_rename(m, new_object, idx)) {
1367 VM_OBJECT_WUNLOCK(new_object);
1368 VM_OBJECT_WUNLOCK(orig_object);
1370 VM_OBJECT_WLOCK(orig_object);
1371 VM_OBJECT_WLOCK(new_object);
1374 #if VM_NRESERVLEVEL > 0
1376 * If some of the reservation's allocated pages remain with
1377 * the original object, then transferring the reservation to
1378 * the new object is neither particularly beneficial nor
1379 * particularly harmful as compared to leaving the reservation
1380 * with the original object. If, however, all of the
1381 * reservation's allocated pages are transferred to the new
1382 * object, then transferring the reservation is typically
1383 * beneficial. Determining which of these two cases applies
1384 * would be more costly than unconditionally renaming the
1387 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1389 if (orig_object->type == OBJT_SWAP)
1392 if (orig_object->type == OBJT_SWAP) {
1394 * swap_pager_copy() can sleep, in which case the orig_object's
1395 * and new_object's locks are released and reacquired.
1397 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1398 TAILQ_FOREACH(m, &new_object->memq, listq)
1402 * Transfer any cached pages from orig_object to new_object.
1403 * If swap_pager_copy() found swapped out pages within the
1404 * specified range of orig_object, then it changed
1405 * new_object's type to OBJT_SWAP when it transferred those
1406 * pages to new_object. Otherwise, new_object's type
1407 * should still be OBJT_DEFAULT and orig_object should not
1408 * contain any cached pages within the specified range.
1410 if (__predict_false(!vm_object_cache_is_empty(orig_object)))
1411 vm_page_cache_transfer(orig_object, offidxstart,
1414 VM_OBJECT_WUNLOCK(orig_object);
1415 VM_OBJECT_WUNLOCK(new_object);
1416 entry->object.vm_object = new_object;
1417 entry->offset = 0LL;
1418 vm_object_deallocate(orig_object);
1419 VM_OBJECT_WLOCK(new_object);
1422 #define OBSC_TEST_ALL_SHADOWED 0x0001
1423 #define OBSC_COLLAPSE_NOWAIT 0x0002
1424 #define OBSC_COLLAPSE_WAIT 0x0004
1427 vm_object_backing_scan(vm_object_t object, int op)
1431 vm_object_t backing_object;
1432 vm_pindex_t backing_offset_index;
1434 VM_OBJECT_ASSERT_WLOCKED(object);
1435 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1437 backing_object = object->backing_object;
1438 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1441 * Initial conditions
1443 if (op & OBSC_TEST_ALL_SHADOWED) {
1445 * We do not want to have to test for the existence of cache
1446 * or swap pages in the backing object. XXX but with the
1447 * new swapper this would be pretty easy to do.
1449 * XXX what about anonymous MAP_SHARED memory that hasn't
1450 * been ZFOD faulted yet? If we do not test for this, the
1451 * shadow test may succeed! XXX
1453 if (backing_object->type != OBJT_DEFAULT) {
1457 if (op & OBSC_COLLAPSE_WAIT) {
1458 vm_object_set_flag(backing_object, OBJ_DEAD);
1464 p = TAILQ_FIRST(&backing_object->memq);
1466 vm_page_t next = TAILQ_NEXT(p, listq);
1467 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1469 if (op & OBSC_TEST_ALL_SHADOWED) {
1473 * Ignore pages outside the parent object's range
1474 * and outside the parent object's mapping of the
1477 * note that we do not busy the backing object's
1481 p->pindex < backing_offset_index ||
1482 new_pindex >= object->size
1489 * See if the parent has the page or if the parent's
1490 * object pager has the page. If the parent has the
1491 * page but the page is not valid, the parent's
1492 * object pager must have the page.
1494 * If this fails, the parent does not completely shadow
1495 * the object and we might as well give up now.
1498 pp = vm_page_lookup(object, new_pindex);
1500 (pp == NULL || pp->valid == 0) &&
1501 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1509 * Check for busy page
1511 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1514 if (op & OBSC_COLLAPSE_NOWAIT) {
1515 if (!p->valid || vm_page_busied(p)) {
1519 } else if (op & OBSC_COLLAPSE_WAIT) {
1520 if (vm_page_busied(p)) {
1521 VM_OBJECT_WUNLOCK(object);
1523 VM_OBJECT_WUNLOCK(backing_object);
1524 vm_page_busy_sleep(p, "vmocol");
1525 VM_OBJECT_WLOCK(object);
1526 VM_OBJECT_WLOCK(backing_object);
1528 * If we slept, anything could have
1529 * happened. Since the object is
1530 * marked dead, the backing offset
1531 * should not have changed so we
1532 * just restart our scan.
1534 p = TAILQ_FIRST(&backing_object->memq);
1540 p->object == backing_object,
1541 ("vm_object_backing_scan: object mismatch")
1545 p->pindex < backing_offset_index ||
1546 new_pindex >= object->size
1548 if (backing_object->type == OBJT_SWAP)
1549 swap_pager_freespace(backing_object,
1553 * Page is out of the parent object's range, we
1554 * can simply destroy it.
1557 KASSERT(!pmap_page_is_mapped(p),
1558 ("freeing mapped page %p", p));
1559 if (p->wire_count == 0)
1568 pp = vm_page_lookup(object, new_pindex);
1570 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1571 (pp != NULL && pp->valid == 0)
1573 if (backing_object->type == OBJT_SWAP)
1574 swap_pager_freespace(backing_object,
1578 * The page in the parent is not (yet) valid.
1579 * We don't know anything about the state of
1580 * the original page. It might be mapped,
1581 * so we must avoid the next if here.
1583 * This is due to a race in vm_fault() where
1584 * we must unbusy the original (backing_obj)
1585 * page before we can (re)lock the parent.
1586 * Hence we can get here.
1593 vm_pager_has_page(object, new_pindex, NULL, NULL)
1595 if (backing_object->type == OBJT_SWAP)
1596 swap_pager_freespace(backing_object,
1600 * page already exists in parent OR swap exists
1601 * for this location in the parent. Destroy
1602 * the original page from the backing object.
1604 * Leave the parent's page alone
1607 KASSERT(!pmap_page_is_mapped(p),
1608 ("freeing mapped page %p", p));
1609 if (p->wire_count == 0)
1619 * Page does not exist in parent, rename the
1620 * page from the backing object to the main object.
1622 * If the page was mapped to a process, it can remain
1623 * mapped through the rename.
1624 * vm_page_rename() will handle dirty and cache.
1626 if (vm_page_rename(p, object, new_pindex)) {
1627 if (op & OBSC_COLLAPSE_NOWAIT) {
1631 VM_OBJECT_WUNLOCK(backing_object);
1632 VM_OBJECT_WUNLOCK(object);
1634 VM_OBJECT_WLOCK(object);
1635 VM_OBJECT_WLOCK(backing_object);
1636 p = TAILQ_FIRST(&backing_object->memq);
1640 /* Use the old pindex to free the right page. */
1641 if (backing_object->type == OBJT_SWAP)
1642 swap_pager_freespace(backing_object,
1643 new_pindex + backing_offset_index, 1);
1645 #if VM_NRESERVLEVEL > 0
1647 * Rename the reservation.
1649 vm_reserv_rename(p, object, backing_object,
1650 backing_offset_index);
1660 * this version of collapse allows the operation to occur earlier and
1661 * when paging_in_progress is true for an object... This is not a complete
1662 * operation, but should plug 99.9% of the rest of the leaks.
1665 vm_object_qcollapse(vm_object_t object)
1667 vm_object_t backing_object = object->backing_object;
1669 VM_OBJECT_ASSERT_WLOCKED(object);
1670 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1672 if (backing_object->ref_count != 1)
1675 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1679 * vm_object_collapse:
1681 * Collapse an object with the object backing it.
1682 * Pages in the backing object are moved into the
1683 * parent, and the backing object is deallocated.
1686 vm_object_collapse(vm_object_t object)
1688 VM_OBJECT_ASSERT_WLOCKED(object);
1691 vm_object_t backing_object;
1694 * Verify that the conditions are right for collapse:
1696 * The object exists and the backing object exists.
1698 if ((backing_object = object->backing_object) == NULL)
1702 * we check the backing object first, because it is most likely
1705 VM_OBJECT_WLOCK(backing_object);
1706 if (backing_object->handle != NULL ||
1707 (backing_object->type != OBJT_DEFAULT &&
1708 backing_object->type != OBJT_SWAP) ||
1709 (backing_object->flags & OBJ_DEAD) ||
1710 object->handle != NULL ||
1711 (object->type != OBJT_DEFAULT &&
1712 object->type != OBJT_SWAP) ||
1713 (object->flags & OBJ_DEAD)) {
1714 VM_OBJECT_WUNLOCK(backing_object);
1719 object->paging_in_progress != 0 ||
1720 backing_object->paging_in_progress != 0
1722 vm_object_qcollapse(object);
1723 VM_OBJECT_WUNLOCK(backing_object);
1727 * We know that we can either collapse the backing object (if
1728 * the parent is the only reference to it) or (perhaps) have
1729 * the parent bypass the object if the parent happens to shadow
1730 * all the resident pages in the entire backing object.
1732 * This is ignoring pager-backed pages such as swap pages.
1733 * vm_object_backing_scan fails the shadowing test in this
1736 if (backing_object->ref_count == 1) {
1738 * If there is exactly one reference to the backing
1739 * object, we can collapse it into the parent.
1741 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1743 #if VM_NRESERVLEVEL > 0
1745 * Break any reservations from backing_object.
1747 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1748 vm_reserv_break_all(backing_object);
1752 * Move the pager from backing_object to object.
1754 if (backing_object->type == OBJT_SWAP) {
1756 * swap_pager_copy() can sleep, in which case
1757 * the backing_object's and object's locks are
1758 * released and reacquired.
1759 * Since swap_pager_copy() is being asked to
1760 * destroy the source, it will change the
1761 * backing_object's type to OBJT_DEFAULT.
1766 OFF_TO_IDX(object->backing_object_offset), TRUE);
1769 * Free any cached pages from backing_object.
1771 if (__predict_false(
1772 !vm_object_cache_is_empty(backing_object)))
1773 vm_page_cache_free(backing_object, 0, 0);
1776 * Object now shadows whatever backing_object did.
1777 * Note that the reference to
1778 * backing_object->backing_object moves from within
1779 * backing_object to within object.
1781 LIST_REMOVE(object, shadow_list);
1782 backing_object->shadow_count--;
1783 if (backing_object->backing_object) {
1784 VM_OBJECT_WLOCK(backing_object->backing_object);
1785 LIST_REMOVE(backing_object, shadow_list);
1787 &backing_object->backing_object->shadow_head,
1788 object, shadow_list);
1790 * The shadow_count has not changed.
1792 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1794 object->backing_object = backing_object->backing_object;
1795 object->backing_object_offset +=
1796 backing_object->backing_object_offset;
1799 * Discard backing_object.
1801 * Since the backing object has no pages, no pager left,
1802 * and no object references within it, all that is
1803 * necessary is to dispose of it.
1805 KASSERT(backing_object->ref_count == 1, (
1806 "backing_object %p was somehow re-referenced during collapse!",
1808 backing_object->type = OBJT_DEAD;
1809 backing_object->ref_count = 0;
1810 VM_OBJECT_WUNLOCK(backing_object);
1811 vm_object_destroy(backing_object);
1815 vm_object_t new_backing_object;
1818 * If we do not entirely shadow the backing object,
1819 * there is nothing we can do so we give up.
1821 if (object->resident_page_count != object->size &&
1822 vm_object_backing_scan(object,
1823 OBSC_TEST_ALL_SHADOWED) == 0) {
1824 VM_OBJECT_WUNLOCK(backing_object);
1829 * Make the parent shadow the next object in the
1830 * chain. Deallocating backing_object will not remove
1831 * it, since its reference count is at least 2.
1833 LIST_REMOVE(object, shadow_list);
1834 backing_object->shadow_count--;
1836 new_backing_object = backing_object->backing_object;
1837 if ((object->backing_object = new_backing_object) != NULL) {
1838 VM_OBJECT_WLOCK(new_backing_object);
1840 &new_backing_object->shadow_head,
1844 new_backing_object->shadow_count++;
1845 vm_object_reference_locked(new_backing_object);
1846 VM_OBJECT_WUNLOCK(new_backing_object);
1847 object->backing_object_offset +=
1848 backing_object->backing_object_offset;
1852 * Drop the reference count on backing_object. Since
1853 * its ref_count was at least 2, it will not vanish.
1855 backing_object->ref_count--;
1856 VM_OBJECT_WUNLOCK(backing_object);
1861 * Try again with this object's new backing object.
1867 * vm_object_page_remove:
1869 * For the given object, either frees or invalidates each of the
1870 * specified pages. In general, a page is freed. However, if a page is
1871 * wired for any reason other than the existence of a managed, wired
1872 * mapping, then it may be invalidated but not removed from the object.
1873 * Pages are specified by the given range ["start", "end") and the option
1874 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1875 * extends from "start" to the end of the object. If the option
1876 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1877 * specified range are affected. If the option OBJPR_NOTMAPPED is
1878 * specified, then the pages within the specified range must have no
1879 * mappings. Otherwise, if this option is not specified, any mappings to
1880 * the specified pages are removed before the pages are freed or
1883 * In general, this operation should only be performed on objects that
1884 * contain managed pages. There are, however, two exceptions. First, it
1885 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1886 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1887 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1888 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1890 * The object must be locked.
1893 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1898 VM_OBJECT_ASSERT_WLOCKED(object);
1899 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1900 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1901 ("vm_object_page_remove: illegal options for object %p", object));
1902 if (object->resident_page_count == 0)
1904 vm_object_pip_add(object, 1);
1906 p = vm_page_find_least(object, start);
1909 * Here, the variable "p" is either (1) the page with the least pindex
1910 * greater than or equal to the parameter "start" or (2) NULL.
1912 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1913 next = TAILQ_NEXT(p, listq);
1916 * If the page is wired for any reason besides the existence
1917 * of managed, wired mappings, then it cannot be freed. For
1918 * example, fictitious pages, which represent device memory,
1919 * are inherently wired and cannot be freed. They can,
1920 * however, be invalidated if the option OBJPR_CLEANONLY is
1924 if (vm_page_xbusied(p)) {
1925 VM_OBJECT_WUNLOCK(object);
1926 vm_page_busy_sleep(p, "vmopax");
1927 VM_OBJECT_WLOCK(object);
1930 if (p->wire_count != 0) {
1931 if ((options & OBJPR_NOTMAPPED) == 0)
1933 if ((options & OBJPR_CLEANONLY) == 0) {
1939 if (vm_page_busied(p)) {
1940 VM_OBJECT_WUNLOCK(object);
1941 vm_page_busy_sleep(p, "vmopar");
1942 VM_OBJECT_WLOCK(object);
1945 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1946 ("vm_object_page_remove: page %p is fictitious", p));
1947 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1948 if ((options & OBJPR_NOTMAPPED) == 0)
1949 pmap_remove_write(p);
1953 if ((options & OBJPR_NOTMAPPED) == 0)
1959 vm_object_pip_wakeup(object);
1961 if (__predict_false(!vm_object_cache_is_empty(object)))
1962 vm_page_cache_free(object, start, end);
1966 * vm_object_page_cache:
1968 * For the given object, attempt to move the specified clean
1969 * pages to the cache queue. If a page is wired for any reason,
1970 * then it will not be changed. Pages are specified by the given
1971 * range ["start", "end"). As a special case, if "end" is zero,
1972 * then the range extends from "start" to the end of the object.
1973 * Any mappings to the specified pages are removed before the
1974 * pages are moved to the cache queue.
1976 * This operation should only be performed on objects that
1977 * contain non-fictitious, managed pages.
1979 * The object must be locked.
1982 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1984 struct mtx *mtx, *new_mtx;
1987 VM_OBJECT_ASSERT_WLOCKED(object);
1988 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1989 ("vm_object_page_cache: illegal object %p", object));
1990 if (object->resident_page_count == 0)
1992 p = vm_page_find_least(object, start);
1995 * Here, the variable "p" is either (1) the page with the least pindex
1996 * greater than or equal to the parameter "start" or (2) NULL.
1999 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2000 next = TAILQ_NEXT(p, listq);
2003 * Avoid releasing and reacquiring the same page lock.
2005 new_mtx = vm_page_lockptr(p);
2006 if (mtx != new_mtx) {
2012 vm_page_try_to_cache(p);
2019 * Populate the specified range of the object with valid pages. Returns
2020 * TRUE if the range is successfully populated and FALSE otherwise.
2022 * Note: This function should be optimized to pass a larger array of
2023 * pages to vm_pager_get_pages() before it is applied to a non-
2024 * OBJT_DEVICE object.
2026 * The object must be locked.
2029 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2035 VM_OBJECT_ASSERT_WLOCKED(object);
2036 for (pindex = start; pindex < end; pindex++) {
2037 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
2038 if (m->valid != VM_PAGE_BITS_ALL) {
2040 rv = vm_pager_get_pages(object, ma, 1, 0);
2041 m = vm_page_lookup(object, pindex);
2044 if (rv != VM_PAGER_OK) {
2052 * Keep "m" busy because a subsequent iteration may unlock
2056 if (pindex > start) {
2057 m = vm_page_lookup(object, start);
2058 while (m != NULL && m->pindex < pindex) {
2060 m = TAILQ_NEXT(m, listq);
2063 return (pindex == end);
2067 * Routine: vm_object_coalesce
2068 * Function: Coalesces two objects backing up adjoining
2069 * regions of memory into a single object.
2071 * returns TRUE if objects were combined.
2073 * NOTE: Only works at the moment if the second object is NULL -
2074 * if it's not, which object do we lock first?
2077 * prev_object First object to coalesce
2078 * prev_offset Offset into prev_object
2079 * prev_size Size of reference to prev_object
2080 * next_size Size of reference to the second object
2081 * reserved Indicator that extension region has
2082 * swap accounted for
2085 * The object must *not* be locked.
2088 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2089 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2091 vm_pindex_t next_pindex;
2093 if (prev_object == NULL)
2095 VM_OBJECT_WLOCK(prev_object);
2096 if ((prev_object->type != OBJT_DEFAULT &&
2097 prev_object->type != OBJT_SWAP) ||
2098 (prev_object->flags & OBJ_TMPFS_NODE) != 0) {
2099 VM_OBJECT_WUNLOCK(prev_object);
2104 * Try to collapse the object first
2106 vm_object_collapse(prev_object);
2109 * Can't coalesce if: . more than one reference . paged out . shadows
2110 * another object . has a copy elsewhere (any of which mean that the
2111 * pages not mapped to prev_entry may be in use anyway)
2113 if (prev_object->backing_object != NULL) {
2114 VM_OBJECT_WUNLOCK(prev_object);
2118 prev_size >>= PAGE_SHIFT;
2119 next_size >>= PAGE_SHIFT;
2120 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2122 if ((prev_object->ref_count > 1) &&
2123 (prev_object->size != next_pindex)) {
2124 VM_OBJECT_WUNLOCK(prev_object);
2129 * Account for the charge.
2131 if (prev_object->cred != NULL) {
2134 * If prev_object was charged, then this mapping,
2135 * althought not charged now, may become writable
2136 * later. Non-NULL cred in the object would prevent
2137 * swap reservation during enabling of the write
2138 * access, so reserve swap now. Failed reservation
2139 * cause allocation of the separate object for the map
2140 * entry, and swap reservation for this entry is
2141 * managed in appropriate time.
2143 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2144 prev_object->cred)) {
2147 prev_object->charge += ptoa(next_size);
2151 * Remove any pages that may still be in the object from a previous
2154 if (next_pindex < prev_object->size) {
2155 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2157 if (prev_object->type == OBJT_SWAP)
2158 swap_pager_freespace(prev_object,
2159 next_pindex, next_size);
2161 if (prev_object->cred != NULL) {
2162 KASSERT(prev_object->charge >=
2163 ptoa(prev_object->size - next_pindex),
2164 ("object %p overcharged 1 %jx %jx", prev_object,
2165 (uintmax_t)next_pindex, (uintmax_t)next_size));
2166 prev_object->charge -= ptoa(prev_object->size -
2173 * Extend the object if necessary.
2175 if (next_pindex + next_size > prev_object->size)
2176 prev_object->size = next_pindex + next_size;
2178 VM_OBJECT_WUNLOCK(prev_object);
2183 vm_object_set_writeable_dirty(vm_object_t object)
2186 VM_OBJECT_ASSERT_WLOCKED(object);
2187 if (object->type != OBJT_VNODE) {
2188 if ((object->flags & OBJ_TMPFS_NODE) != 0) {
2189 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs"));
2190 vm_object_set_flag(object, OBJ_TMPFS_DIRTY);
2194 object->generation++;
2195 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2197 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2203 * For each page offset within the specified range of the given object,
2204 * find the highest-level page in the shadow chain and unwire it. A page
2205 * must exist at every page offset, and the highest-level page must be
2209 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2212 vm_object_t tobject;
2214 vm_pindex_t end_pindex, pindex, tpindex;
2215 int depth, locked_depth;
2217 KASSERT((offset & PAGE_MASK) == 0,
2218 ("vm_object_unwire: offset is not page aligned"));
2219 KASSERT((length & PAGE_MASK) == 0,
2220 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2221 /* The wired count of a fictitious page never changes. */
2222 if ((object->flags & OBJ_FICTITIOUS) != 0)
2224 pindex = OFF_TO_IDX(offset);
2225 end_pindex = pindex + atop(length);
2227 VM_OBJECT_RLOCK(object);
2228 m = vm_page_find_least(object, pindex);
2229 while (pindex < end_pindex) {
2230 if (m == NULL || pindex < m->pindex) {
2232 * The first object in the shadow chain doesn't
2233 * contain a page at the current index. Therefore,
2234 * the page must exist in a backing object.
2241 OFF_TO_IDX(tobject->backing_object_offset);
2242 tobject = tobject->backing_object;
2243 KASSERT(tobject != NULL,
2244 ("vm_object_unwire: missing page"));
2245 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2248 if (depth == locked_depth) {
2250 VM_OBJECT_RLOCK(tobject);
2252 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2256 m = TAILQ_NEXT(m, listq);
2259 vm_page_unwire(tm, queue);
2264 /* Release the accumulated object locks. */
2265 for (depth = 0; depth < locked_depth; depth++) {
2266 tobject = object->backing_object;
2267 VM_OBJECT_RUNLOCK(object);
2272 #include "opt_ddb.h"
2274 #include <sys/kernel.h>
2276 #include <sys/cons.h>
2278 #include <ddb/ddb.h>
2281 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2284 vm_map_entry_t tmpe;
2292 tmpe = map->header.next;
2293 entcount = map->nentries;
2294 while (entcount-- && (tmpe != &map->header)) {
2295 if (_vm_object_in_map(map, object, tmpe)) {
2300 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2301 tmpm = entry->object.sub_map;
2302 tmpe = tmpm->header.next;
2303 entcount = tmpm->nentries;
2304 while (entcount-- && tmpe != &tmpm->header) {
2305 if (_vm_object_in_map(tmpm, object, tmpe)) {
2310 } else if ((obj = entry->object.vm_object) != NULL) {
2311 for (; obj; obj = obj->backing_object)
2312 if (obj == object) {
2320 vm_object_in_map(vm_object_t object)
2324 /* sx_slock(&allproc_lock); */
2325 FOREACH_PROC_IN_SYSTEM(p) {
2326 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2328 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2329 /* sx_sunlock(&allproc_lock); */
2333 /* sx_sunlock(&allproc_lock); */
2334 if (_vm_object_in_map(kernel_map, object, 0))
2339 DB_SHOW_COMMAND(vmochk, vm_object_check)
2344 * make sure that internal objs are in a map somewhere
2345 * and none have zero ref counts.
2347 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2348 if (object->handle == NULL &&
2349 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2350 if (object->ref_count == 0) {
2351 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2352 (long)object->size);
2354 if (!vm_object_in_map(object)) {
2356 "vmochk: internal obj is not in a map: "
2357 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2358 object->ref_count, (u_long)object->size,
2359 (u_long)object->size,
2360 (void *)object->backing_object);
2367 * vm_object_print: [ debug ]
2369 DB_SHOW_COMMAND(object, vm_object_print_static)
2371 /* XXX convert args. */
2372 vm_object_t object = (vm_object_t)addr;
2373 boolean_t full = have_addr;
2377 /* XXX count is an (unused) arg. Avoid shadowing it. */
2378 #define count was_count
2386 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2387 object, (int)object->type, (uintmax_t)object->size,
2388 object->resident_page_count, object->ref_count, object->flags,
2389 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2390 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2391 object->shadow_count,
2392 object->backing_object ? object->backing_object->ref_count : 0,
2393 object->backing_object, (uintmax_t)object->backing_object_offset);
2400 TAILQ_FOREACH(p, &object->memq, listq) {
2402 db_iprintf("memory:=");
2403 else if (count == 6) {
2411 db_printf("(off=0x%jx,page=0x%jx)",
2412 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2422 /* XXX need this non-static entry for calling from vm_map_print. */
2425 /* db_expr_t */ long addr,
2426 boolean_t have_addr,
2427 /* db_expr_t */ long count,
2430 vm_object_print_static(addr, have_addr, count, modif);
2433 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2438 vm_page_t m, prev_m;
2442 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2443 db_printf("new object: %p\n", (void *)object);
2454 TAILQ_FOREACH(m, &object->memq, listq) {
2455 if (m->pindex > 128)
2457 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2458 prev_m->pindex + 1 != m->pindex) {
2460 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2461 (long)fidx, rcount, (long)pa);
2473 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2478 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2479 (long)fidx, rcount, (long)pa);
2489 pa = VM_PAGE_TO_PHYS(m);
2493 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2494 (long)fidx, rcount, (long)pa);