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_reserv.h>
100 static int old_msync;
101 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
102 "Use old (insecure) msync behavior");
104 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
105 int pagerflags, int flags, boolean_t *clearobjflags,
107 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
108 boolean_t *clearobjflags);
109 static void vm_object_qcollapse(vm_object_t object);
110 static void vm_object_vndeallocate(vm_object_t object);
113 * Virtual memory objects maintain the actual data
114 * associated with allocated virtual memory. A given
115 * page of memory exists within exactly one object.
117 * An object is only deallocated when all "references"
118 * are given up. Only one "reference" to a given
119 * region of an object should be writeable.
121 * Associated with each object is a list of all resident
122 * memory pages belonging to that object; this list is
123 * maintained by the "vm_page" module, and locked by the object's
126 * Each object also records a "pager" routine which is
127 * used to retrieve (and store) pages to the proper backing
128 * storage. In addition, objects may be backed by other
129 * objects from which they were virtual-copied.
131 * The only items within the object structure which are
132 * modified after time of creation are:
133 * reference count locked by object's lock
134 * pager routine locked by object's lock
138 struct object_q vm_object_list;
139 struct mtx vm_object_list_mtx; /* lock for object list and count */
141 struct vm_object kernel_object_store;
142 struct vm_object kmem_object_store;
144 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
147 static long object_collapses;
148 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
149 &object_collapses, 0, "VM object collapses");
151 static long object_bypasses;
152 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
153 &object_bypasses, 0, "VM object bypasses");
155 static uma_zone_t obj_zone;
157 static int vm_object_zinit(void *mem, int size, int flags);
160 static void vm_object_zdtor(void *mem, int size, void *arg);
163 vm_object_zdtor(void *mem, int size, void *arg)
167 object = (vm_object_t)mem;
168 KASSERT(TAILQ_EMPTY(&object->memq),
169 ("object %p has resident pages",
171 #if VM_NRESERVLEVEL > 0
172 KASSERT(LIST_EMPTY(&object->rvq),
173 ("object %p has reservations",
176 KASSERT(object->cache == NULL,
177 ("object %p has cached pages",
179 KASSERT(object->paging_in_progress == 0,
180 ("object %p paging_in_progress = %d",
181 object, object->paging_in_progress));
182 KASSERT(object->resident_page_count == 0,
183 ("object %p resident_page_count = %d",
184 object, object->resident_page_count));
185 KASSERT(object->shadow_count == 0,
186 ("object %p shadow_count = %d",
187 object, object->shadow_count));
192 vm_object_zinit(void *mem, int size, int flags)
196 object = (vm_object_t)mem;
197 bzero(&object->lock, sizeof(object->lock));
198 VM_OBJECT_LOCK_INIT(object, "standard vm object");
200 /* These are true for any object that has been freed */
201 object->paging_in_progress = 0;
202 object->resident_page_count = 0;
203 object->shadow_count = 0;
208 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
211 TAILQ_INIT(&object->memq);
212 LIST_INIT(&object->shadow_head);
218 panic("_vm_object_allocate: can't create OBJT_DEAD");
221 object->flags = OBJ_ONEMAPPING;
225 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
228 object->flags = OBJ_FICTITIOUS;
231 object->flags = OBJ_UNMANAGED;
237 panic("_vm_object_allocate: type %d is undefined", type);
240 object->generation = 1;
241 object->ref_count = 1;
242 object->memattr = VM_MEMATTR_DEFAULT;
245 object->pg_color = 0;
246 object->handle = NULL;
247 object->backing_object = NULL;
248 object->backing_object_offset = (vm_ooffset_t) 0;
249 #if VM_NRESERVLEVEL > 0
250 LIST_INIT(&object->rvq);
252 object->cache = NULL;
254 mtx_lock(&vm_object_list_mtx);
255 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
256 mtx_unlock(&vm_object_list_mtx);
262 * Initialize the VM objects module.
267 TAILQ_INIT(&vm_object_list);
268 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
270 VM_OBJECT_LOCK_INIT(kernel_object, "kernel vm object");
271 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
273 #if VM_NRESERVLEVEL > 0
274 kernel_object->flags |= OBJ_COLORED;
275 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
278 VM_OBJECT_LOCK_INIT(kmem_object, "kmem vm object");
279 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
281 #if VM_NRESERVLEVEL > 0
282 kmem_object->flags |= OBJ_COLORED;
283 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
287 * The lock portion of struct vm_object must be type stable due
288 * to vm_pageout_fallback_object_lock locking a vm object
289 * without holding any references to it.
291 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
297 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
301 vm_object_clear_flag(vm_object_t object, u_short bits)
304 VM_OBJECT_ASSERT_WLOCKED(object);
305 object->flags &= ~bits;
309 * Sets the default memory attribute for the specified object. Pages
310 * that are allocated to this object are by default assigned this memory
313 * Presently, this function must be called before any pages are allocated
314 * to the object. In the future, this requirement may be relaxed for
315 * "default" and "swap" objects.
318 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
321 VM_OBJECT_ASSERT_WLOCKED(object);
322 switch (object->type) {
330 if (!TAILQ_EMPTY(&object->memq))
331 return (KERN_FAILURE);
334 return (KERN_INVALID_ARGUMENT);
336 panic("vm_object_set_memattr: object %p is of undefined type",
339 object->memattr = memattr;
340 return (KERN_SUCCESS);
344 vm_object_pip_add(vm_object_t object, short i)
347 VM_OBJECT_ASSERT_WLOCKED(object);
348 object->paging_in_progress += i;
352 vm_object_pip_subtract(vm_object_t object, short i)
355 VM_OBJECT_ASSERT_WLOCKED(object);
356 object->paging_in_progress -= i;
360 vm_object_pip_wakeup(vm_object_t object)
363 VM_OBJECT_ASSERT_WLOCKED(object);
364 object->paging_in_progress--;
365 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
366 vm_object_clear_flag(object, OBJ_PIPWNT);
372 vm_object_pip_wakeupn(vm_object_t object, short i)
375 VM_OBJECT_ASSERT_WLOCKED(object);
377 object->paging_in_progress -= i;
378 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
379 vm_object_clear_flag(object, OBJ_PIPWNT);
385 vm_object_pip_wait(vm_object_t object, char *waitid)
388 VM_OBJECT_ASSERT_WLOCKED(object);
389 while (object->paging_in_progress) {
390 object->flags |= OBJ_PIPWNT;
391 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
396 * vm_object_allocate:
398 * Returns a new object with the given size.
401 vm_object_allocate(objtype_t type, vm_pindex_t size)
405 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
406 _vm_object_allocate(type, size, object);
412 * vm_object_reference:
414 * Gets another reference to the given object. Note: OBJ_DEAD
415 * objects can be referenced during final cleaning.
418 vm_object_reference(vm_object_t object)
422 VM_OBJECT_WLOCK(object);
423 vm_object_reference_locked(object);
424 VM_OBJECT_WUNLOCK(object);
428 * vm_object_reference_locked:
430 * Gets another reference to the given object.
432 * The object must be locked.
435 vm_object_reference_locked(vm_object_t object)
439 VM_OBJECT_ASSERT_WLOCKED(object);
441 if (object->type == OBJT_VNODE) {
448 * Handle deallocating an object of type OBJT_VNODE.
451 vm_object_vndeallocate(vm_object_t object)
453 struct vnode *vp = (struct vnode *) object->handle;
455 VM_OBJECT_ASSERT_WLOCKED(object);
456 KASSERT(object->type == OBJT_VNODE,
457 ("vm_object_vndeallocate: not a vnode object"));
458 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
460 if (object->ref_count == 0) {
461 vprint("vm_object_vndeallocate", vp);
462 panic("vm_object_vndeallocate: bad object reference count");
466 if (object->ref_count > 1) {
468 VM_OBJECT_WUNLOCK(object);
469 /* vrele may need the vnode lock. */
473 VM_OBJECT_WUNLOCK(object);
474 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
476 VM_OBJECT_WLOCK(object);
478 if (object->type == OBJT_DEAD) {
479 VM_OBJECT_WUNLOCK(object);
482 if (object->ref_count == 0)
484 VM_OBJECT_WUNLOCK(object);
491 * vm_object_deallocate:
493 * Release a reference to the specified object,
494 * gained either through a vm_object_allocate
495 * or a vm_object_reference call. When all references
496 * are gone, storage associated with this object
497 * may be relinquished.
499 * No object may be locked.
502 vm_object_deallocate(vm_object_t object)
506 while (object != NULL) {
507 VM_OBJECT_WLOCK(object);
508 if (object->type == OBJT_VNODE) {
509 vm_object_vndeallocate(object);
513 KASSERT(object->ref_count != 0,
514 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
517 * If the reference count goes to 0 we start calling
518 * vm_object_terminate() on the object chain.
519 * A ref count of 1 may be a special case depending on the
520 * shadow count being 0 or 1.
523 if (object->ref_count > 1) {
524 VM_OBJECT_WUNLOCK(object);
526 } else if (object->ref_count == 1) {
527 if (object->shadow_count == 0 &&
528 object->handle == NULL &&
529 (object->type == OBJT_DEFAULT ||
530 object->type == OBJT_SWAP)) {
531 vm_object_set_flag(object, OBJ_ONEMAPPING);
532 } else if ((object->shadow_count == 1) &&
533 (object->handle == NULL) &&
534 (object->type == OBJT_DEFAULT ||
535 object->type == OBJT_SWAP)) {
538 robject = LIST_FIRST(&object->shadow_head);
539 KASSERT(robject != NULL,
540 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
542 object->shadow_count));
543 if (!VM_OBJECT_TRYWLOCK(robject)) {
545 * Avoid a potential deadlock.
548 VM_OBJECT_WUNLOCK(object);
550 * More likely than not the thread
551 * holding robject's lock has lower
552 * priority than the current thread.
553 * Let the lower priority thread run.
559 * Collapse object into its shadow unless its
560 * shadow is dead. In that case, object will
561 * be deallocated by the thread that is
562 * deallocating its shadow.
564 if ((robject->flags & OBJ_DEAD) == 0 &&
565 (robject->handle == NULL) &&
566 (robject->type == OBJT_DEFAULT ||
567 robject->type == OBJT_SWAP)) {
569 robject->ref_count++;
571 if (robject->paging_in_progress) {
572 VM_OBJECT_WUNLOCK(object);
573 vm_object_pip_wait(robject,
575 temp = robject->backing_object;
576 if (object == temp) {
577 VM_OBJECT_WLOCK(object);
580 } else if (object->paging_in_progress) {
581 VM_OBJECT_WUNLOCK(robject);
582 object->flags |= OBJ_PIPWNT;
583 VM_OBJECT_SLEEP(object, object,
584 PDROP | PVM, "objde2", 0);
585 VM_OBJECT_WLOCK(robject);
586 temp = robject->backing_object;
587 if (object == temp) {
588 VM_OBJECT_WLOCK(object);
592 VM_OBJECT_WUNLOCK(object);
594 if (robject->ref_count == 1) {
595 robject->ref_count--;
600 vm_object_collapse(object);
601 VM_OBJECT_WUNLOCK(object);
604 VM_OBJECT_WUNLOCK(robject);
606 VM_OBJECT_WUNLOCK(object);
610 temp = object->backing_object;
612 VM_OBJECT_WLOCK(temp);
613 LIST_REMOVE(object, shadow_list);
614 temp->shadow_count--;
615 VM_OBJECT_WUNLOCK(temp);
616 object->backing_object = NULL;
619 * Don't double-terminate, we could be in a termination
620 * recursion due to the terminate having to sync data
623 if ((object->flags & OBJ_DEAD) == 0)
624 vm_object_terminate(object);
626 VM_OBJECT_WUNLOCK(object);
632 * vm_object_destroy removes the object from the global object list
633 * and frees the space for the object.
636 vm_object_destroy(vm_object_t object)
640 * Remove the object from the global object list.
642 mtx_lock(&vm_object_list_mtx);
643 TAILQ_REMOVE(&vm_object_list, object, object_list);
644 mtx_unlock(&vm_object_list_mtx);
647 * Release the allocation charge.
649 if (object->cred != NULL) {
650 KASSERT(object->type == OBJT_DEFAULT ||
651 object->type == OBJT_SWAP,
652 ("vm_object_terminate: non-swap obj %p has cred",
654 swap_release_by_cred(object->charge, object->cred);
656 crfree(object->cred);
661 * Free the space for the object.
663 uma_zfree(obj_zone, object);
667 * vm_object_terminate actually destroys the specified object, freeing
668 * up all previously used resources.
670 * The object must be locked.
671 * This routine may block.
674 vm_object_terminate(vm_object_t object)
678 VM_OBJECT_ASSERT_WLOCKED(object);
681 * Make sure no one uses us.
683 vm_object_set_flag(object, OBJ_DEAD);
686 * wait for the pageout daemon to be done with the object
688 vm_object_pip_wait(object, "objtrm");
690 KASSERT(!object->paging_in_progress,
691 ("vm_object_terminate: pageout in progress"));
694 * Clean and free the pages, as appropriate. All references to the
695 * object are gone, so we don't need to lock it.
697 if (object->type == OBJT_VNODE) {
698 struct vnode *vp = (struct vnode *)object->handle;
701 * Clean pages and flush buffers.
703 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
704 VM_OBJECT_WUNLOCK(object);
706 vinvalbuf(vp, V_SAVE, 0, 0);
708 VM_OBJECT_WLOCK(object);
711 KASSERT(object->ref_count == 0,
712 ("vm_object_terminate: object with references, ref_count=%d",
716 * Free any remaining pageable pages. This also removes them from the
717 * paging queues. However, don't free wired pages, just remove them
718 * from the object. Rather than incrementally removing each page from
719 * the object, the page and object are reset to any empty state.
721 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
722 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
723 ("vm_object_terminate: freeing busy page %p", p));
726 * Optimize the page's removal from the object by resetting
727 * its "object" field. Specifically, if the page is not
728 * wired, then the effect of this assignment is that
729 * vm_page_free()'s call to vm_page_remove() will return
730 * immediately without modifying the page or the object.
733 if (p->wire_count == 0) {
735 PCPU_INC(cnt.v_pfree);
740 * If the object contained any pages, then reset it to an empty state.
741 * None of the object's fields, including "resident_page_count", were
742 * modified by the preceding loop.
744 if (object->resident_page_count != 0) {
746 TAILQ_INIT(&object->memq);
747 object->resident_page_count = 0;
748 if (object->type == OBJT_VNODE)
749 vdrop(object->handle);
752 #if VM_NRESERVLEVEL > 0
753 if (__predict_false(!LIST_EMPTY(&object->rvq)))
754 vm_reserv_break_all(object);
756 if (__predict_false(object->cache != NULL))
757 vm_page_cache_free(object, 0, 0);
760 * Let the pager know object is dead.
762 vm_pager_deallocate(object);
763 VM_OBJECT_WUNLOCK(object);
765 vm_object_destroy(object);
769 * Make the page read-only so that we can clear the object flags. However, if
770 * this is a nosync mmap then the object is likely to stay dirty so do not
771 * mess with the page and do not clear the object flags. Returns TRUE if the
772 * page should be flushed, and FALSE otherwise.
775 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
779 * If we have been asked to skip nosync pages and this is a
780 * nosync page, skip it. Note that the object flags were not
781 * cleared in this case so we do not have to set them.
783 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
784 *clearobjflags = FALSE;
787 pmap_remove_write(p);
788 return (p->dirty != 0);
793 * vm_object_page_clean
795 * Clean all dirty pages in the specified range of object. Leaves page
796 * on whatever queue it is currently on. If NOSYNC is set then do not
797 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
798 * leaving the object dirty.
800 * When stuffing pages asynchronously, allow clustering. XXX we need a
801 * synchronous clustering mode implementation.
803 * Odd semantics: if start == end, we clean everything.
805 * The object must be locked.
807 * Returns FALSE if some page from the range was not written, as
808 * reported by the pager, and TRUE otherwise.
811 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
815 vm_pindex_t pi, tend, tstart;
816 int curgeneration, n, pagerflags;
817 boolean_t clearobjflags, eio, res;
819 VM_OBJECT_ASSERT_WLOCKED(object);
820 KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
821 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
822 object->resident_page_count == 0)
825 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
826 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
827 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
829 tstart = OFF_TO_IDX(start);
830 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
831 clearobjflags = tstart == 0 && tend >= object->size;
835 curgeneration = object->generation;
837 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
841 np = TAILQ_NEXT(p, listq);
844 if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
845 if (object->generation != curgeneration) {
846 if ((flags & OBJPC_SYNC) != 0)
849 clearobjflags = FALSE;
851 np = vm_page_find_least(object, pi);
854 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
857 n = vm_object_page_collect_flush(object, p, pagerflags,
858 flags, &clearobjflags, &eio);
861 clearobjflags = FALSE;
863 if (object->generation != curgeneration) {
864 if ((flags & OBJPC_SYNC) != 0)
867 clearobjflags = FALSE;
871 * If the VOP_PUTPAGES() did a truncated write, so
872 * that even the first page of the run is not fully
873 * written, vm_pageout_flush() returns 0 as the run
874 * length. Since the condition that caused truncated
875 * write may be permanent, e.g. exhausted free space,
876 * accepting n == 0 would cause an infinite loop.
878 * Forwarding the iterator leaves the unwritten page
879 * behind, but there is not much we can do there if
880 * filesystem refuses to write it.
884 clearobjflags = FALSE;
886 np = vm_page_find_least(object, pi + n);
889 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
893 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
898 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
899 int flags, boolean_t *clearobjflags, boolean_t *eio)
901 vm_page_t ma[vm_pageout_page_count], p_first, tp;
902 int count, i, mreq, runlen;
904 vm_page_lock_assert(p, MA_NOTOWNED);
905 VM_OBJECT_ASSERT_WLOCKED(object);
910 for (tp = p; count < vm_pageout_page_count; count++) {
911 tp = vm_page_next(tp);
912 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
914 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
918 for (p_first = p; count < vm_pageout_page_count; count++) {
919 tp = vm_page_prev(p_first);
920 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
922 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
928 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
931 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
936 * Note that there is absolutely no sense in writing out
937 * anonymous objects, so we track down the vnode object
939 * We invalidate (remove) all pages from the address space
940 * for semantic correctness.
942 * If the backing object is a device object with unmanaged pages, then any
943 * mappings to the specified range of pages must be removed before this
944 * function is called.
946 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
947 * may start out with a NULL object.
950 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
951 boolean_t syncio, boolean_t invalidate)
953 vm_object_t backing_object;
956 int error, flags, fsync_after;
963 VM_OBJECT_WLOCK(object);
964 while ((backing_object = object->backing_object) != NULL) {
965 VM_OBJECT_WLOCK(backing_object);
966 offset += object->backing_object_offset;
967 VM_OBJECT_WUNLOCK(object);
968 object = backing_object;
969 if (object->size < OFF_TO_IDX(offset + size))
970 size = IDX_TO_OFF(object->size) - offset;
973 * Flush pages if writing is allowed, invalidate them
974 * if invalidation requested. Pages undergoing I/O
975 * will be ignored by vm_object_page_remove().
977 * We cannot lock the vnode and then wait for paging
978 * to complete without deadlocking against vm_fault.
979 * Instead we simply call vm_object_page_remove() and
980 * allow it to block internally on a page-by-page
981 * basis when it encounters pages undergoing async
984 if (object->type == OBJT_VNODE &&
985 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
987 VM_OBJECT_WUNLOCK(object);
988 (void) vn_start_write(vp, &mp, V_WAIT);
989 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
990 if (syncio && !invalidate && offset == 0 &&
991 OFF_TO_IDX(size) == object->size) {
993 * If syncing the whole mapping of the file,
994 * it is faster to schedule all the writes in
995 * async mode, also allowing the clustering,
996 * and then wait for i/o to complete.
1001 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1002 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1003 fsync_after = FALSE;
1005 VM_OBJECT_WLOCK(object);
1006 res = vm_object_page_clean(object, offset, offset + size,
1008 VM_OBJECT_WUNLOCK(object);
1010 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1012 vn_finished_write(mp);
1015 VM_OBJECT_WLOCK(object);
1017 if ((object->type == OBJT_VNODE ||
1018 object->type == OBJT_DEVICE) && invalidate) {
1019 if (object->type == OBJT_DEVICE)
1021 * The option OBJPR_NOTMAPPED must be passed here
1022 * because vm_object_page_remove() cannot remove
1023 * unmanaged mappings.
1025 flags = OBJPR_NOTMAPPED;
1029 flags = OBJPR_CLEANONLY;
1030 vm_object_page_remove(object, OFF_TO_IDX(offset),
1031 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1033 VM_OBJECT_WUNLOCK(object);
1038 * vm_object_madvise:
1040 * Implements the madvise function at the object/page level.
1042 * MADV_WILLNEED (any object)
1044 * Activate the specified pages if they are resident.
1046 * MADV_DONTNEED (any object)
1048 * Deactivate the specified pages if they are resident.
1050 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1051 * OBJ_ONEMAPPING only)
1053 * Deactivate and clean the specified pages if they are
1054 * resident. This permits the process to reuse the pages
1055 * without faulting or the kernel to reclaim the pages
1059 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1062 vm_pindex_t tpindex;
1063 vm_object_t backing_object, tobject;
1068 VM_OBJECT_WLOCK(object);
1070 * Locate and adjust resident pages
1072 for (; pindex < end; pindex += 1) {
1078 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1079 * and those pages must be OBJ_ONEMAPPING.
1081 if (advise == MADV_FREE) {
1082 if ((tobject->type != OBJT_DEFAULT &&
1083 tobject->type != OBJT_SWAP) ||
1084 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1085 goto unlock_tobject;
1087 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1088 goto unlock_tobject;
1089 m = vm_page_lookup(tobject, tpindex);
1090 if (m == NULL && advise == MADV_WILLNEED) {
1092 * If the page is cached, reactivate it.
1094 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1099 * There may be swap even if there is no backing page
1101 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1102 swap_pager_freespace(tobject, tpindex, 1);
1106 backing_object = tobject->backing_object;
1107 if (backing_object == NULL)
1108 goto unlock_tobject;
1109 VM_OBJECT_WLOCK(backing_object);
1110 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1111 if (tobject != object)
1112 VM_OBJECT_WUNLOCK(tobject);
1113 tobject = backing_object;
1115 } else if (m->valid != VM_PAGE_BITS_ALL)
1116 goto unlock_tobject;
1118 * If the page is not in a normal state, skip it.
1121 if (m->hold_count != 0 || m->wire_count != 0) {
1123 goto unlock_tobject;
1125 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1126 ("vm_object_madvise: page %p is fictitious", m));
1127 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1128 ("vm_object_madvise: page %p is not managed", m));
1129 if ((m->oflags & VPO_BUSY) || m->busy) {
1130 if (advise == MADV_WILLNEED) {
1132 * Reference the page before unlocking and
1133 * sleeping so that the page daemon is less
1134 * likely to reclaim it.
1136 vm_page_aflag_set(m, PGA_REFERENCED);
1139 if (object != tobject)
1140 VM_OBJECT_WUNLOCK(object);
1141 m->oflags |= VPO_WANTED;
1142 VM_OBJECT_SLEEP(tobject, m, PDROP | PVM, "madvpo", 0);
1143 VM_OBJECT_WLOCK(object);
1146 if (advise == MADV_WILLNEED) {
1147 vm_page_activate(m);
1148 } else if (advise == MADV_DONTNEED) {
1149 vm_page_dontneed(m);
1150 } else if (advise == MADV_FREE) {
1152 * Mark the page clean. This will allow the page
1153 * to be freed up by the system. However, such pages
1154 * are often reused quickly by malloc()/free()
1155 * so we do not do anything that would cause
1156 * a page fault if we can help it.
1158 * Specifically, we do not try to actually free
1159 * the page now nor do we try to put it in the
1160 * cache (which would cause a page fault on reuse).
1162 * But we do make the page is freeable as we
1163 * can without actually taking the step of unmapping
1166 pmap_clear_modify(m);
1169 vm_page_dontneed(m);
1172 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1173 swap_pager_freespace(tobject, tpindex, 1);
1175 if (tobject != object)
1176 VM_OBJECT_WUNLOCK(tobject);
1178 VM_OBJECT_WUNLOCK(object);
1184 * Create a new object which is backed by the
1185 * specified existing object range. The source
1186 * object reference is deallocated.
1188 * The new object and offset into that object
1189 * are returned in the source parameters.
1193 vm_object_t *object, /* IN/OUT */
1194 vm_ooffset_t *offset, /* IN/OUT */
1203 * Don't create the new object if the old object isn't shared.
1205 if (source != NULL) {
1206 VM_OBJECT_WLOCK(source);
1207 if (source->ref_count == 1 &&
1208 source->handle == NULL &&
1209 (source->type == OBJT_DEFAULT ||
1210 source->type == OBJT_SWAP)) {
1211 VM_OBJECT_WUNLOCK(source);
1214 VM_OBJECT_WUNLOCK(source);
1218 * Allocate a new object with the given length.
1220 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1223 * The new object shadows the source object, adding a reference to it.
1224 * Our caller changes his reference to point to the new object,
1225 * removing a reference to the source object. Net result: no change
1226 * of reference count.
1228 * Try to optimize the result object's page color when shadowing
1229 * in order to maintain page coloring consistency in the combined
1232 result->backing_object = source;
1234 * Store the offset into the source object, and fix up the offset into
1237 result->backing_object_offset = *offset;
1238 if (source != NULL) {
1239 VM_OBJECT_WLOCK(source);
1240 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1241 source->shadow_count++;
1242 #if VM_NRESERVLEVEL > 0
1243 result->flags |= source->flags & OBJ_COLORED;
1244 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1245 ((1 << (VM_NFREEORDER - 1)) - 1);
1247 VM_OBJECT_WUNLOCK(source);
1252 * Return the new things
1261 * Split the pages in a map entry into a new object. This affords
1262 * easier removal of unused pages, and keeps object inheritance from
1263 * being a negative impact on memory usage.
1266 vm_object_split(vm_map_entry_t entry)
1268 vm_page_t m, m_next;
1269 vm_object_t orig_object, new_object, source;
1270 vm_pindex_t idx, offidxstart;
1273 orig_object = entry->object.vm_object;
1274 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1276 if (orig_object->ref_count <= 1)
1278 VM_OBJECT_WUNLOCK(orig_object);
1280 offidxstart = OFF_TO_IDX(entry->offset);
1281 size = atop(entry->end - entry->start);
1284 * If swap_pager_copy() is later called, it will convert new_object
1285 * into a swap object.
1287 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1290 * At this point, the new object is still private, so the order in
1291 * which the original and new objects are locked does not matter.
1293 VM_OBJECT_WLOCK(new_object);
1294 VM_OBJECT_WLOCK(orig_object);
1295 source = orig_object->backing_object;
1296 if (source != NULL) {
1297 VM_OBJECT_WLOCK(source);
1298 if ((source->flags & OBJ_DEAD) != 0) {
1299 VM_OBJECT_WUNLOCK(source);
1300 VM_OBJECT_WUNLOCK(orig_object);
1301 VM_OBJECT_WUNLOCK(new_object);
1302 vm_object_deallocate(new_object);
1303 VM_OBJECT_WLOCK(orig_object);
1306 LIST_INSERT_HEAD(&source->shadow_head,
1307 new_object, shadow_list);
1308 source->shadow_count++;
1309 vm_object_reference_locked(source); /* for new_object */
1310 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1311 VM_OBJECT_WUNLOCK(source);
1312 new_object->backing_object_offset =
1313 orig_object->backing_object_offset + entry->offset;
1314 new_object->backing_object = source;
1316 if (orig_object->cred != NULL) {
1317 new_object->cred = orig_object->cred;
1318 crhold(orig_object->cred);
1319 new_object->charge = ptoa(size);
1320 KASSERT(orig_object->charge >= ptoa(size),
1321 ("orig_object->charge < 0"));
1322 orig_object->charge -= ptoa(size);
1325 m = vm_page_find_least(orig_object, offidxstart);
1326 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1328 m_next = TAILQ_NEXT(m, listq);
1331 * We must wait for pending I/O to complete before we can
1334 * We do not have to VM_PROT_NONE the page as mappings should
1335 * not be changed by this operation.
1337 if ((m->oflags & VPO_BUSY) || m->busy) {
1338 VM_OBJECT_WUNLOCK(new_object);
1339 m->oflags |= VPO_WANTED;
1340 VM_OBJECT_SLEEP(orig_object, m, PVM, "spltwt", 0);
1341 VM_OBJECT_WLOCK(new_object);
1344 #if VM_NRESERVLEVEL > 0
1346 * If some of the reservation's allocated pages remain with
1347 * the original object, then transferring the reservation to
1348 * the new object is neither particularly beneficial nor
1349 * particularly harmful as compared to leaving the reservation
1350 * with the original object. If, however, all of the
1351 * reservation's allocated pages are transferred to the new
1352 * object, then transferring the reservation is typically
1353 * beneficial. Determining which of these two cases applies
1354 * would be more costly than unconditionally renaming the
1357 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1360 vm_page_rename(m, new_object, idx);
1362 /* page automatically made dirty by rename and cache handled */
1365 if (orig_object->type == OBJT_SWAP) {
1367 * swap_pager_copy() can sleep, in which case the orig_object's
1368 * and new_object's locks are released and reacquired.
1370 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1373 * Transfer any cached pages from orig_object to new_object.
1374 * If swap_pager_copy() found swapped out pages within the
1375 * specified range of orig_object, then it changed
1376 * new_object's type to OBJT_SWAP when it transferred those
1377 * pages to new_object. Otherwise, new_object's type
1378 * should still be OBJT_DEFAULT and orig_object should not
1379 * contain any cached pages within the specified range.
1381 if (__predict_false(orig_object->cache != NULL))
1382 vm_page_cache_transfer(orig_object, offidxstart,
1385 VM_OBJECT_WUNLOCK(orig_object);
1386 TAILQ_FOREACH(m, &new_object->memq, listq)
1388 VM_OBJECT_WUNLOCK(new_object);
1389 entry->object.vm_object = new_object;
1390 entry->offset = 0LL;
1391 vm_object_deallocate(orig_object);
1392 VM_OBJECT_WLOCK(new_object);
1395 #define OBSC_TEST_ALL_SHADOWED 0x0001
1396 #define OBSC_COLLAPSE_NOWAIT 0x0002
1397 #define OBSC_COLLAPSE_WAIT 0x0004
1400 vm_object_backing_scan(vm_object_t object, int op)
1404 vm_object_t backing_object;
1405 vm_pindex_t backing_offset_index;
1407 VM_OBJECT_ASSERT_WLOCKED(object);
1408 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1410 backing_object = object->backing_object;
1411 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1414 * Initial conditions
1416 if (op & OBSC_TEST_ALL_SHADOWED) {
1418 * We do not want to have to test for the existence of cache
1419 * or swap pages in the backing object. XXX but with the
1420 * new swapper this would be pretty easy to do.
1422 * XXX what about anonymous MAP_SHARED memory that hasn't
1423 * been ZFOD faulted yet? If we do not test for this, the
1424 * shadow test may succeed! XXX
1426 if (backing_object->type != OBJT_DEFAULT) {
1430 if (op & OBSC_COLLAPSE_WAIT) {
1431 vm_object_set_flag(backing_object, OBJ_DEAD);
1437 p = TAILQ_FIRST(&backing_object->memq);
1439 vm_page_t next = TAILQ_NEXT(p, listq);
1440 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1442 if (op & OBSC_TEST_ALL_SHADOWED) {
1446 * Ignore pages outside the parent object's range
1447 * and outside the parent object's mapping of the
1450 * note that we do not busy the backing object's
1454 p->pindex < backing_offset_index ||
1455 new_pindex >= object->size
1462 * See if the parent has the page or if the parent's
1463 * object pager has the page. If the parent has the
1464 * page but the page is not valid, the parent's
1465 * object pager must have the page.
1467 * If this fails, the parent does not completely shadow
1468 * the object and we might as well give up now.
1471 pp = vm_page_lookup(object, new_pindex);
1473 (pp == NULL || pp->valid == 0) &&
1474 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1482 * Check for busy page
1484 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1487 if (op & OBSC_COLLAPSE_NOWAIT) {
1488 if ((p->oflags & VPO_BUSY) ||
1494 } else if (op & OBSC_COLLAPSE_WAIT) {
1495 if ((p->oflags & VPO_BUSY) || p->busy) {
1496 VM_OBJECT_WUNLOCK(object);
1497 p->oflags |= VPO_WANTED;
1498 VM_OBJECT_SLEEP(backing_object, p,
1499 PDROP | PVM, "vmocol", 0);
1500 VM_OBJECT_WLOCK(object);
1501 VM_OBJECT_WLOCK(backing_object);
1503 * If we slept, anything could have
1504 * happened. Since the object is
1505 * marked dead, the backing offset
1506 * should not have changed so we
1507 * just restart our scan.
1509 p = TAILQ_FIRST(&backing_object->memq);
1515 p->object == backing_object,
1516 ("vm_object_backing_scan: object mismatch")
1520 * Destroy any associated swap
1522 if (backing_object->type == OBJT_SWAP) {
1523 swap_pager_freespace(
1531 p->pindex < backing_offset_index ||
1532 new_pindex >= object->size
1535 * Page is out of the parent object's range, we
1536 * can simply destroy it.
1539 KASSERT(!pmap_page_is_mapped(p),
1540 ("freeing mapped page %p", p));
1541 if (p->wire_count == 0)
1550 pp = vm_page_lookup(object, new_pindex);
1552 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1553 (pp != NULL && pp->valid == 0)
1556 * The page in the parent is not (yet) valid.
1557 * We don't know anything about the state of
1558 * the original page. It might be mapped,
1559 * so we must avoid the next if here.
1561 * This is due to a race in vm_fault() where
1562 * we must unbusy the original (backing_obj)
1563 * page before we can (re)lock the parent.
1564 * Hence we can get here.
1571 vm_pager_has_page(object, new_pindex, NULL, NULL)
1574 * page already exists in parent OR swap exists
1575 * for this location in the parent. Destroy
1576 * the original page from the backing object.
1578 * Leave the parent's page alone
1581 KASSERT(!pmap_page_is_mapped(p),
1582 ("freeing mapped page %p", p));
1583 if (p->wire_count == 0)
1592 #if VM_NRESERVLEVEL > 0
1594 * Rename the reservation.
1596 vm_reserv_rename(p, object, backing_object,
1597 backing_offset_index);
1601 * Page does not exist in parent, rename the
1602 * page from the backing object to the main object.
1604 * If the page was mapped to a process, it can remain
1605 * mapped through the rename.
1608 vm_page_rename(p, object, new_pindex);
1610 /* page automatically made dirty by rename */
1619 * this version of collapse allows the operation to occur earlier and
1620 * when paging_in_progress is true for an object... This is not a complete
1621 * operation, but should plug 99.9% of the rest of the leaks.
1624 vm_object_qcollapse(vm_object_t object)
1626 vm_object_t backing_object = object->backing_object;
1628 VM_OBJECT_ASSERT_WLOCKED(object);
1629 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1631 if (backing_object->ref_count != 1)
1634 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1638 * vm_object_collapse:
1640 * Collapse an object with the object backing it.
1641 * Pages in the backing object are moved into the
1642 * parent, and the backing object is deallocated.
1645 vm_object_collapse(vm_object_t object)
1647 VM_OBJECT_ASSERT_WLOCKED(object);
1650 vm_object_t backing_object;
1653 * Verify that the conditions are right for collapse:
1655 * The object exists and the backing object exists.
1657 if ((backing_object = object->backing_object) == NULL)
1661 * we check the backing object first, because it is most likely
1664 VM_OBJECT_WLOCK(backing_object);
1665 if (backing_object->handle != NULL ||
1666 (backing_object->type != OBJT_DEFAULT &&
1667 backing_object->type != OBJT_SWAP) ||
1668 (backing_object->flags & OBJ_DEAD) ||
1669 object->handle != NULL ||
1670 (object->type != OBJT_DEFAULT &&
1671 object->type != OBJT_SWAP) ||
1672 (object->flags & OBJ_DEAD)) {
1673 VM_OBJECT_WUNLOCK(backing_object);
1678 object->paging_in_progress != 0 ||
1679 backing_object->paging_in_progress != 0
1681 vm_object_qcollapse(object);
1682 VM_OBJECT_WUNLOCK(backing_object);
1686 * We know that we can either collapse the backing object (if
1687 * the parent is the only reference to it) or (perhaps) have
1688 * the parent bypass the object if the parent happens to shadow
1689 * all the resident pages in the entire backing object.
1691 * This is ignoring pager-backed pages such as swap pages.
1692 * vm_object_backing_scan fails the shadowing test in this
1695 if (backing_object->ref_count == 1) {
1697 * If there is exactly one reference to the backing
1698 * object, we can collapse it into the parent.
1700 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1702 #if VM_NRESERVLEVEL > 0
1704 * Break any reservations from backing_object.
1706 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1707 vm_reserv_break_all(backing_object);
1711 * Move the pager from backing_object to object.
1713 if (backing_object->type == OBJT_SWAP) {
1715 * swap_pager_copy() can sleep, in which case
1716 * the backing_object's and object's locks are
1717 * released and reacquired.
1718 * Since swap_pager_copy() is being asked to
1719 * destroy the source, it will change the
1720 * backing_object's type to OBJT_DEFAULT.
1725 OFF_TO_IDX(object->backing_object_offset), TRUE);
1728 * Free any cached pages from backing_object.
1730 if (__predict_false(backing_object->cache != NULL))
1731 vm_page_cache_free(backing_object, 0, 0);
1734 * Object now shadows whatever backing_object did.
1735 * Note that the reference to
1736 * backing_object->backing_object moves from within
1737 * backing_object to within object.
1739 LIST_REMOVE(object, shadow_list);
1740 backing_object->shadow_count--;
1741 if (backing_object->backing_object) {
1742 VM_OBJECT_WLOCK(backing_object->backing_object);
1743 LIST_REMOVE(backing_object, shadow_list);
1745 &backing_object->backing_object->shadow_head,
1746 object, shadow_list);
1748 * The shadow_count has not changed.
1750 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1752 object->backing_object = backing_object->backing_object;
1753 object->backing_object_offset +=
1754 backing_object->backing_object_offset;
1757 * Discard backing_object.
1759 * Since the backing object has no pages, no pager left,
1760 * and no object references within it, all that is
1761 * necessary is to dispose of it.
1763 KASSERT(backing_object->ref_count == 1, (
1764 "backing_object %p was somehow re-referenced during collapse!",
1766 VM_OBJECT_WUNLOCK(backing_object);
1767 vm_object_destroy(backing_object);
1771 vm_object_t new_backing_object;
1774 * If we do not entirely shadow the backing object,
1775 * there is nothing we can do so we give up.
1777 if (object->resident_page_count != object->size &&
1778 vm_object_backing_scan(object,
1779 OBSC_TEST_ALL_SHADOWED) == 0) {
1780 VM_OBJECT_WUNLOCK(backing_object);
1785 * Make the parent shadow the next object in the
1786 * chain. Deallocating backing_object will not remove
1787 * it, since its reference count is at least 2.
1789 LIST_REMOVE(object, shadow_list);
1790 backing_object->shadow_count--;
1792 new_backing_object = backing_object->backing_object;
1793 if ((object->backing_object = new_backing_object) != NULL) {
1794 VM_OBJECT_WLOCK(new_backing_object);
1796 &new_backing_object->shadow_head,
1800 new_backing_object->shadow_count++;
1801 vm_object_reference_locked(new_backing_object);
1802 VM_OBJECT_WUNLOCK(new_backing_object);
1803 object->backing_object_offset +=
1804 backing_object->backing_object_offset;
1808 * Drop the reference count on backing_object. Since
1809 * its ref_count was at least 2, it will not vanish.
1811 backing_object->ref_count--;
1812 VM_OBJECT_WUNLOCK(backing_object);
1817 * Try again with this object's new backing object.
1823 * vm_object_page_remove:
1825 * For the given object, either frees or invalidates each of the
1826 * specified pages. In general, a page is freed. However, if a page is
1827 * wired for any reason other than the existence of a managed, wired
1828 * mapping, then it may be invalidated but not removed from the object.
1829 * Pages are specified by the given range ["start", "end") and the option
1830 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1831 * extends from "start" to the end of the object. If the option
1832 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1833 * specified range are affected. If the option OBJPR_NOTMAPPED is
1834 * specified, then the pages within the specified range must have no
1835 * mappings. Otherwise, if this option is not specified, any mappings to
1836 * the specified pages are removed before the pages are freed or
1839 * In general, this operation should only be performed on objects that
1840 * contain managed pages. There are, however, two exceptions. First, it
1841 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1842 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1843 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1844 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1846 * The object must be locked.
1849 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1855 VM_OBJECT_ASSERT_WLOCKED(object);
1856 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1857 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1858 ("vm_object_page_remove: illegal options for object %p", object));
1859 if (object->resident_page_count == 0)
1861 vm_object_pip_add(object, 1);
1863 p = vm_page_find_least(object, start);
1866 * Here, the variable "p" is either (1) the page with the least pindex
1867 * greater than or equal to the parameter "start" or (2) NULL.
1869 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1870 next = TAILQ_NEXT(p, listq);
1873 * If the page is wired for any reason besides the existence
1874 * of managed, wired mappings, then it cannot be freed. For
1875 * example, fictitious pages, which represent device memory,
1876 * are inherently wired and cannot be freed. They can,
1877 * however, be invalidated if the option OBJPR_CLEANONLY is
1881 if ((wirings = p->wire_count) != 0 &&
1882 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1883 if ((options & OBJPR_NOTMAPPED) == 0) {
1885 /* Account for removal of wired mappings. */
1887 p->wire_count -= wirings;
1889 if ((options & OBJPR_CLEANONLY) == 0) {
1896 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1898 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1899 ("vm_object_page_remove: page %p is fictitious", p));
1900 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1901 if ((options & OBJPR_NOTMAPPED) == 0)
1902 pmap_remove_write(p);
1908 if ((options & OBJPR_NOTMAPPED) == 0) {
1910 /* Account for removal of wired mappings. */
1912 KASSERT(p->wire_count == wirings,
1913 ("inconsistent wire count %d %d %p",
1914 p->wire_count, wirings, p));
1916 atomic_subtract_int(&cnt.v_wire_count, 1);
1922 vm_object_pip_wakeup(object);
1924 if (__predict_false(object->cache != NULL))
1925 vm_page_cache_free(object, start, end);
1929 * vm_object_page_cache:
1931 * For the given object, attempt to move the specified clean
1932 * pages to the cache queue. If a page is wired for any reason,
1933 * then it will not be changed. Pages are specified by the given
1934 * range ["start", "end"). As a special case, if "end" is zero,
1935 * then the range extends from "start" to the end of the object.
1936 * Any mappings to the specified pages are removed before the
1937 * pages are moved to the cache queue.
1939 * This operation should only be performed on objects that
1940 * contain non-fictitious, managed pages.
1942 * The object must be locked.
1945 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1947 struct mtx *mtx, *new_mtx;
1950 VM_OBJECT_ASSERT_WLOCKED(object);
1951 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1952 ("vm_object_page_cache: illegal object %p", object));
1953 if (object->resident_page_count == 0)
1955 p = vm_page_find_least(object, start);
1958 * Here, the variable "p" is either (1) the page with the least pindex
1959 * greater than or equal to the parameter "start" or (2) NULL.
1962 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1963 next = TAILQ_NEXT(p, listq);
1966 * Avoid releasing and reacquiring the same page lock.
1968 new_mtx = vm_page_lockptr(p);
1969 if (mtx != new_mtx) {
1975 vm_page_try_to_cache(p);
1982 * Populate the specified range of the object with valid pages. Returns
1983 * TRUE if the range is successfully populated and FALSE otherwise.
1985 * Note: This function should be optimized to pass a larger array of
1986 * pages to vm_pager_get_pages() before it is applied to a non-
1987 * OBJT_DEVICE object.
1989 * The object must be locked.
1992 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1998 VM_OBJECT_ASSERT_WLOCKED(object);
1999 for (pindex = start; pindex < end; pindex++) {
2000 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
2002 if (m->valid != VM_PAGE_BITS_ALL) {
2004 rv = vm_pager_get_pages(object, ma, 1, 0);
2005 m = vm_page_lookup(object, pindex);
2008 if (rv != VM_PAGER_OK) {
2016 * Keep "m" busy because a subsequent iteration may unlock
2020 if (pindex > start) {
2021 m = vm_page_lookup(object, start);
2022 while (m != NULL && m->pindex < pindex) {
2024 m = TAILQ_NEXT(m, listq);
2027 return (pindex == end);
2031 * Routine: vm_object_coalesce
2032 * Function: Coalesces two objects backing up adjoining
2033 * regions of memory into a single object.
2035 * returns TRUE if objects were combined.
2037 * NOTE: Only works at the moment if the second object is NULL -
2038 * if it's not, which object do we lock first?
2041 * prev_object First object to coalesce
2042 * prev_offset Offset into prev_object
2043 * prev_size Size of reference to prev_object
2044 * next_size Size of reference to the second object
2045 * reserved Indicator that extension region has
2046 * swap accounted for
2049 * The object must *not* be locked.
2052 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2053 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2055 vm_pindex_t next_pindex;
2057 if (prev_object == NULL)
2059 VM_OBJECT_WLOCK(prev_object);
2060 if (prev_object->type != OBJT_DEFAULT &&
2061 prev_object->type != OBJT_SWAP) {
2062 VM_OBJECT_WUNLOCK(prev_object);
2067 * Try to collapse the object first
2069 vm_object_collapse(prev_object);
2072 * Can't coalesce if: . more than one reference . paged out . shadows
2073 * another object . has a copy elsewhere (any of which mean that the
2074 * pages not mapped to prev_entry may be in use anyway)
2076 if (prev_object->backing_object != NULL) {
2077 VM_OBJECT_WUNLOCK(prev_object);
2081 prev_size >>= PAGE_SHIFT;
2082 next_size >>= PAGE_SHIFT;
2083 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2085 if ((prev_object->ref_count > 1) &&
2086 (prev_object->size != next_pindex)) {
2087 VM_OBJECT_WUNLOCK(prev_object);
2092 * Account for the charge.
2094 if (prev_object->cred != NULL) {
2097 * If prev_object was charged, then this mapping,
2098 * althought not charged now, may become writable
2099 * later. Non-NULL cred in the object would prevent
2100 * swap reservation during enabling of the write
2101 * access, so reserve swap now. Failed reservation
2102 * cause allocation of the separate object for the map
2103 * entry, and swap reservation for this entry is
2104 * managed in appropriate time.
2106 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2107 prev_object->cred)) {
2110 prev_object->charge += ptoa(next_size);
2114 * Remove any pages that may still be in the object from a previous
2117 if (next_pindex < prev_object->size) {
2118 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2120 if (prev_object->type == OBJT_SWAP)
2121 swap_pager_freespace(prev_object,
2122 next_pindex, next_size);
2124 if (prev_object->cred != NULL) {
2125 KASSERT(prev_object->charge >=
2126 ptoa(prev_object->size - next_pindex),
2127 ("object %p overcharged 1 %jx %jx", prev_object,
2128 (uintmax_t)next_pindex, (uintmax_t)next_size));
2129 prev_object->charge -= ptoa(prev_object->size -
2136 * Extend the object if necessary.
2138 if (next_pindex + next_size > prev_object->size)
2139 prev_object->size = next_pindex + next_size;
2141 VM_OBJECT_WUNLOCK(prev_object);
2146 vm_object_set_writeable_dirty(vm_object_t object)
2149 VM_OBJECT_ASSERT_WLOCKED(object);
2150 if (object->type != OBJT_VNODE)
2152 object->generation++;
2153 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2155 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2158 #include "opt_ddb.h"
2160 #include <sys/kernel.h>
2162 #include <sys/cons.h>
2164 #include <ddb/ddb.h>
2167 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2170 vm_map_entry_t tmpe;
2178 tmpe = map->header.next;
2179 entcount = map->nentries;
2180 while (entcount-- && (tmpe != &map->header)) {
2181 if (_vm_object_in_map(map, object, tmpe)) {
2186 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2187 tmpm = entry->object.sub_map;
2188 tmpe = tmpm->header.next;
2189 entcount = tmpm->nentries;
2190 while (entcount-- && tmpe != &tmpm->header) {
2191 if (_vm_object_in_map(tmpm, object, tmpe)) {
2196 } else if ((obj = entry->object.vm_object) != NULL) {
2197 for (; obj; obj = obj->backing_object)
2198 if (obj == object) {
2206 vm_object_in_map(vm_object_t object)
2210 /* sx_slock(&allproc_lock); */
2211 FOREACH_PROC_IN_SYSTEM(p) {
2212 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2214 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2215 /* sx_sunlock(&allproc_lock); */
2219 /* sx_sunlock(&allproc_lock); */
2220 if (_vm_object_in_map(kernel_map, object, 0))
2222 if (_vm_object_in_map(kmem_map, object, 0))
2224 if (_vm_object_in_map(pager_map, object, 0))
2226 if (_vm_object_in_map(buffer_map, object, 0))
2231 DB_SHOW_COMMAND(vmochk, vm_object_check)
2236 * make sure that internal objs are in a map somewhere
2237 * and none have zero ref counts.
2239 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2240 if (object->handle == NULL &&
2241 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2242 if (object->ref_count == 0) {
2243 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2244 (long)object->size);
2246 if (!vm_object_in_map(object)) {
2248 "vmochk: internal obj is not in a map: "
2249 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2250 object->ref_count, (u_long)object->size,
2251 (u_long)object->size,
2252 (void *)object->backing_object);
2259 * vm_object_print: [ debug ]
2261 DB_SHOW_COMMAND(object, vm_object_print_static)
2263 /* XXX convert args. */
2264 vm_object_t object = (vm_object_t)addr;
2265 boolean_t full = have_addr;
2269 /* XXX count is an (unused) arg. Avoid shadowing it. */
2270 #define count was_count
2278 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2279 object, (int)object->type, (uintmax_t)object->size,
2280 object->resident_page_count, object->ref_count, object->flags,
2281 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2282 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2283 object->shadow_count,
2284 object->backing_object ? object->backing_object->ref_count : 0,
2285 object->backing_object, (uintmax_t)object->backing_object_offset);
2292 TAILQ_FOREACH(p, &object->memq, listq) {
2294 db_iprintf("memory:=");
2295 else if (count == 6) {
2303 db_printf("(off=0x%jx,page=0x%jx)",
2304 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2314 /* XXX need this non-static entry for calling from vm_map_print. */
2317 /* db_expr_t */ long addr,
2318 boolean_t have_addr,
2319 /* db_expr_t */ long count,
2322 vm_object_print_static(addr, have_addr, count, modif);
2325 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2330 vm_page_t m, prev_m;
2334 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2335 db_printf("new object: %p\n", (void *)object);
2346 TAILQ_FOREACH(m, &object->memq, listq) {
2347 if (m->pindex > 128)
2349 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2350 prev_m->pindex + 1 != m->pindex) {
2352 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2353 (long)fidx, rcount, (long)pa);
2365 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2370 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2371 (long)fidx, rcount, (long)pa);
2381 pa = VM_PAGE_TO_PHYS(m);
2385 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2386 (long)fidx, rcount, (long)pa);