2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94
35 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36 * All rights reserved.
38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
40 * Permission to use, copy, modify and distribute this software and
41 * its documentation is hereby granted, provided that both the copyright
42 * notice and this permission notice appear in all copies of the
43 * software, derivative works or modified versions, and any portions
44 * thereof, and that both notices appear in supporting documentation.
46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
50 * Carnegie Mellon requests users of this software to return to
52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
53 * School of Computer Science
54 * Carnegie Mellon University
55 * Pittsburgh PA 15213-3890
57 * any improvements or extensions that they make and grant Carnegie the
58 * rights to redistribute these changes.
62 * Virtual memory object module.
65 #include <sys/cdefs.h>
66 __FBSDID("$FreeBSD$");
70 #include <sys/param.h>
71 #include <sys/systm.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/mutex.h>
78 #include <sys/proc.h> /* for curproc, pageproc */
79 #include <sys/socket.h>
80 #include <sys/resourcevar.h>
81 #include <sys/rwlock.h>
82 #include <sys/vnode.h>
83 #include <sys/vmmeter.h>
87 #include <vm/vm_param.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pageout.h>
93 #include <vm/vm_pager.h>
94 #include <vm/swap_pager.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_extern.h>
97 #include <vm/vm_radix.h>
98 #include <vm/vm_reserv.h>
101 static int old_msync;
102 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
103 "Use old (insecure) msync behavior");
105 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
106 int pagerflags, int flags, boolean_t *clearobjflags,
108 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
109 boolean_t *clearobjflags);
110 static void vm_object_qcollapse(vm_object_t object);
111 static void vm_object_vndeallocate(vm_object_t object);
114 * Virtual memory objects maintain the actual data
115 * associated with allocated virtual memory. A given
116 * page of memory exists within exactly one object.
118 * An object is only deallocated when all "references"
119 * are given up. Only one "reference" to a given
120 * region of an object should be writeable.
122 * Associated with each object is a list of all resident
123 * memory pages belonging to that object; this list is
124 * maintained by the "vm_page" module, and locked by the object's
127 * Each object also records a "pager" routine which is
128 * used to retrieve (and store) pages to the proper backing
129 * storage. In addition, objects may be backed by other
130 * objects from which they were virtual-copied.
132 * The only items within the object structure which are
133 * modified after time of creation are:
134 * reference count locked by object's lock
135 * pager routine locked by object's lock
139 struct object_q vm_object_list;
140 struct mtx vm_object_list_mtx; /* lock for object list and count */
142 struct vm_object kernel_object_store;
143 struct vm_object kmem_object_store;
145 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
148 static long object_collapses;
149 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
150 &object_collapses, 0, "VM object collapses");
152 static long object_bypasses;
153 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
154 &object_bypasses, 0, "VM object bypasses");
156 static uma_zone_t obj_zone;
158 static int vm_object_zinit(void *mem, int size, int flags);
161 static void vm_object_zdtor(void *mem, int size, void *arg);
164 vm_object_zdtor(void *mem, int size, void *arg)
168 object = (vm_object_t)mem;
169 KASSERT(TAILQ_EMPTY(&object->memq),
170 ("object %p has resident pages in its memq", object));
171 KASSERT(vm_radix_is_empty(&object->rtree),
172 ("object %p has resident pages in its trie", object));
173 #if VM_NRESERVLEVEL > 0
174 KASSERT(LIST_EMPTY(&object->rvq),
175 ("object %p has reservations",
178 KASSERT(vm_object_cache_is_empty(object),
179 ("object %p has cached pages",
181 KASSERT(object->paging_in_progress == 0,
182 ("object %p paging_in_progress = %d",
183 object, object->paging_in_progress));
184 KASSERT(object->resident_page_count == 0,
185 ("object %p resident_page_count = %d",
186 object, object->resident_page_count));
187 KASSERT(object->shadow_count == 0,
188 ("object %p shadow_count = %d",
189 object, object->shadow_count));
194 vm_object_zinit(void *mem, int size, int flags)
198 object = (vm_object_t)mem;
199 bzero(&object->lock, sizeof(object->lock));
200 rw_init_flags(&object->lock, "vm object", RW_DUPOK);
202 /* These are true for any object that has been freed */
203 object->rtree.rt_root = 0;
204 object->paging_in_progress = 0;
205 object->resident_page_count = 0;
206 object->shadow_count = 0;
207 object->cache.rt_root = 0;
212 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
215 TAILQ_INIT(&object->memq);
216 LIST_INIT(&object->shadow_head);
221 panic("_vm_object_allocate: can't create OBJT_DEAD");
224 object->flags = OBJ_ONEMAPPING;
228 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
231 object->flags = OBJ_FICTITIOUS;
234 object->flags = OBJ_UNMANAGED;
240 panic("_vm_object_allocate: type %d is undefined", type);
243 object->generation = 1;
244 object->ref_count = 1;
245 object->memattr = VM_MEMATTR_DEFAULT;
248 object->handle = NULL;
249 object->backing_object = NULL;
250 object->backing_object_offset = (vm_ooffset_t) 0;
251 #if VM_NRESERVLEVEL > 0
252 LIST_INIT(&object->rvq);
255 mtx_lock(&vm_object_list_mtx);
256 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
257 mtx_unlock(&vm_object_list_mtx);
263 * Initialize the VM objects module.
268 TAILQ_INIT(&vm_object_list);
269 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
271 rw_init(&kernel_object->lock, "kernel vm object");
272 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
274 #if VM_NRESERVLEVEL > 0
275 kernel_object->flags |= OBJ_COLORED;
276 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
279 rw_init(&kmem_object->lock, "kmem vm object");
280 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
282 #if VM_NRESERVLEVEL > 0
283 kmem_object->flags |= OBJ_COLORED;
284 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
288 * The lock portion of struct vm_object must be type stable due
289 * to vm_pageout_fallback_object_lock locking a vm object
290 * without holding any references to it.
292 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
298 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
304 vm_object_clear_flag(vm_object_t object, u_short bits)
307 VM_OBJECT_ASSERT_WLOCKED(object);
308 object->flags &= ~bits;
312 * Sets the default memory attribute for the specified object. Pages
313 * that are allocated to this object are by default assigned this memory
316 * Presently, this function must be called before any pages are allocated
317 * to the object. In the future, this requirement may be relaxed for
318 * "default" and "swap" objects.
321 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
324 VM_OBJECT_ASSERT_WLOCKED(object);
325 switch (object->type) {
333 if (!TAILQ_EMPTY(&object->memq))
334 return (KERN_FAILURE);
337 return (KERN_INVALID_ARGUMENT);
339 panic("vm_object_set_memattr: object %p is of undefined type",
342 object->memattr = memattr;
343 return (KERN_SUCCESS);
347 vm_object_pip_add(vm_object_t object, short i)
350 VM_OBJECT_ASSERT_WLOCKED(object);
351 object->paging_in_progress += i;
355 vm_object_pip_subtract(vm_object_t object, short i)
358 VM_OBJECT_ASSERT_WLOCKED(object);
359 object->paging_in_progress -= i;
363 vm_object_pip_wakeup(vm_object_t object)
366 VM_OBJECT_ASSERT_WLOCKED(object);
367 object->paging_in_progress--;
368 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
369 vm_object_clear_flag(object, OBJ_PIPWNT);
375 vm_object_pip_wakeupn(vm_object_t object, short i)
378 VM_OBJECT_ASSERT_WLOCKED(object);
380 object->paging_in_progress -= i;
381 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
382 vm_object_clear_flag(object, OBJ_PIPWNT);
388 vm_object_pip_wait(vm_object_t object, char *waitid)
391 VM_OBJECT_ASSERT_WLOCKED(object);
392 while (object->paging_in_progress) {
393 object->flags |= OBJ_PIPWNT;
394 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
399 * vm_object_allocate:
401 * Returns a new object with the given size.
404 vm_object_allocate(objtype_t type, vm_pindex_t size)
408 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
409 _vm_object_allocate(type, size, object);
415 * vm_object_reference:
417 * Gets another reference to the given object. Note: OBJ_DEAD
418 * objects can be referenced during final cleaning.
421 vm_object_reference(vm_object_t object)
425 VM_OBJECT_WLOCK(object);
426 vm_object_reference_locked(object);
427 VM_OBJECT_WUNLOCK(object);
431 * vm_object_reference_locked:
433 * Gets another reference to the given object.
435 * The object must be locked.
438 vm_object_reference_locked(vm_object_t object)
442 VM_OBJECT_ASSERT_WLOCKED(object);
444 if (object->type == OBJT_VNODE) {
451 * Handle deallocating an object of type OBJT_VNODE.
454 vm_object_vndeallocate(vm_object_t object)
456 struct vnode *vp = (struct vnode *) object->handle;
458 VM_OBJECT_ASSERT_WLOCKED(object);
459 KASSERT(object->type == OBJT_VNODE,
460 ("vm_object_vndeallocate: not a vnode object"));
461 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
463 if (object->ref_count == 0) {
464 vprint("vm_object_vndeallocate", vp);
465 panic("vm_object_vndeallocate: bad object reference count");
469 if (object->ref_count > 1) {
471 VM_OBJECT_WUNLOCK(object);
472 /* vrele may need the vnode lock. */
476 VM_OBJECT_WUNLOCK(object);
477 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
479 VM_OBJECT_WLOCK(object);
481 if (object->type == OBJT_DEAD) {
482 VM_OBJECT_WUNLOCK(object);
485 if (object->ref_count == 0)
487 VM_OBJECT_WUNLOCK(object);
494 * vm_object_deallocate:
496 * Release a reference to the specified object,
497 * gained either through a vm_object_allocate
498 * or a vm_object_reference call. When all references
499 * are gone, storage associated with this object
500 * may be relinquished.
502 * No object may be locked.
505 vm_object_deallocate(vm_object_t object)
509 while (object != NULL) {
510 VM_OBJECT_WLOCK(object);
511 if (object->type == OBJT_VNODE) {
512 vm_object_vndeallocate(object);
516 KASSERT(object->ref_count != 0,
517 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
520 * If the reference count goes to 0 we start calling
521 * vm_object_terminate() on the object chain.
522 * A ref count of 1 may be a special case depending on the
523 * shadow count being 0 or 1.
526 if (object->ref_count > 1) {
527 VM_OBJECT_WUNLOCK(object);
529 } else if (object->ref_count == 1) {
530 if (object->shadow_count == 0 &&
531 object->handle == NULL &&
532 (object->type == OBJT_DEFAULT ||
533 object->type == OBJT_SWAP)) {
534 vm_object_set_flag(object, OBJ_ONEMAPPING);
535 } else if ((object->shadow_count == 1) &&
536 (object->handle == NULL) &&
537 (object->type == OBJT_DEFAULT ||
538 object->type == OBJT_SWAP)) {
541 robject = LIST_FIRST(&object->shadow_head);
542 KASSERT(robject != NULL,
543 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
545 object->shadow_count));
546 if (!VM_OBJECT_TRYWLOCK(robject)) {
548 * Avoid a potential deadlock.
551 VM_OBJECT_WUNLOCK(object);
553 * More likely than not the thread
554 * holding robject's lock has lower
555 * priority than the current thread.
556 * Let the lower priority thread run.
562 * Collapse object into its shadow unless its
563 * shadow is dead. In that case, object will
564 * be deallocated by the thread that is
565 * deallocating its shadow.
567 if ((robject->flags & OBJ_DEAD) == 0 &&
568 (robject->handle == NULL) &&
569 (robject->type == OBJT_DEFAULT ||
570 robject->type == OBJT_SWAP)) {
572 robject->ref_count++;
574 if (robject->paging_in_progress) {
575 VM_OBJECT_WUNLOCK(object);
576 vm_object_pip_wait(robject,
578 temp = robject->backing_object;
579 if (object == temp) {
580 VM_OBJECT_WLOCK(object);
583 } else if (object->paging_in_progress) {
584 VM_OBJECT_WUNLOCK(robject);
585 object->flags |= OBJ_PIPWNT;
586 VM_OBJECT_SLEEP(object, object,
587 PDROP | PVM, "objde2", 0);
588 VM_OBJECT_WLOCK(robject);
589 temp = robject->backing_object;
590 if (object == temp) {
591 VM_OBJECT_WLOCK(object);
595 VM_OBJECT_WUNLOCK(object);
597 if (robject->ref_count == 1) {
598 robject->ref_count--;
603 vm_object_collapse(object);
604 VM_OBJECT_WUNLOCK(object);
607 VM_OBJECT_WUNLOCK(robject);
609 VM_OBJECT_WUNLOCK(object);
613 temp = object->backing_object;
615 VM_OBJECT_WLOCK(temp);
616 LIST_REMOVE(object, shadow_list);
617 temp->shadow_count--;
618 VM_OBJECT_WUNLOCK(temp);
619 object->backing_object = NULL;
622 * Don't double-terminate, we could be in a termination
623 * recursion due to the terminate having to sync data
626 if ((object->flags & OBJ_DEAD) == 0)
627 vm_object_terminate(object);
629 VM_OBJECT_WUNLOCK(object);
635 * vm_object_destroy removes the object from the global object list
636 * and frees the space for the object.
639 vm_object_destroy(vm_object_t object)
643 * Remove the object from the global object list.
645 mtx_lock(&vm_object_list_mtx);
646 TAILQ_REMOVE(&vm_object_list, object, object_list);
647 mtx_unlock(&vm_object_list_mtx);
650 * Release the allocation charge.
652 if (object->cred != NULL) {
653 KASSERT(object->type == OBJT_DEFAULT ||
654 object->type == OBJT_SWAP,
655 ("vm_object_terminate: non-swap obj %p has cred",
657 swap_release_by_cred(object->charge, object->cred);
659 crfree(object->cred);
664 * Free the space for the object.
666 uma_zfree(obj_zone, object);
670 * vm_object_terminate actually destroys the specified object, freeing
671 * up all previously used resources.
673 * The object must be locked.
674 * This routine may block.
677 vm_object_terminate(vm_object_t object)
681 VM_OBJECT_ASSERT_WLOCKED(object);
684 * Make sure no one uses us.
686 vm_object_set_flag(object, OBJ_DEAD);
689 * wait for the pageout daemon to be done with the object
691 vm_object_pip_wait(object, "objtrm");
693 KASSERT(!object->paging_in_progress,
694 ("vm_object_terminate: pageout in progress"));
697 * Clean and free the pages, as appropriate. All references to the
698 * object are gone, so we don't need to lock it.
700 if (object->type == OBJT_VNODE) {
701 struct vnode *vp = (struct vnode *)object->handle;
704 * Clean pages and flush buffers.
706 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
707 VM_OBJECT_WUNLOCK(object);
709 vinvalbuf(vp, V_SAVE, 0, 0);
711 VM_OBJECT_WLOCK(object);
714 KASSERT(object->ref_count == 0,
715 ("vm_object_terminate: object with references, ref_count=%d",
719 * Free any remaining pageable pages. This also removes them from the
720 * paging queues. However, don't free wired pages, just remove them
721 * from the object. Rather than incrementally removing each page from
722 * the object, the page and object are reset to any empty state.
724 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
725 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
726 ("vm_object_terminate: freeing busy page %p", p));
729 * Optimize the page's removal from the object by resetting
730 * its "object" field. Specifically, if the page is not
731 * wired, then the effect of this assignment is that
732 * vm_page_free()'s call to vm_page_remove() will return
733 * immediately without modifying the page or the object.
736 if (p->wire_count == 0) {
738 PCPU_INC(cnt.v_pfree);
743 * If the object contained any pages, then reset it to an empty state.
744 * None of the object's fields, including "resident_page_count", were
745 * modified by the preceding loop.
747 if (object->resident_page_count != 0) {
748 vm_radix_reclaim_allnodes(&object->rtree);
749 TAILQ_INIT(&object->memq);
750 object->resident_page_count = 0;
751 if (object->type == OBJT_VNODE)
752 vdrop(object->handle);
755 #if VM_NRESERVLEVEL > 0
756 if (__predict_false(!LIST_EMPTY(&object->rvq)))
757 vm_reserv_break_all(object);
759 if (__predict_false(!vm_object_cache_is_empty(object)))
760 vm_page_cache_free(object, 0, 0);
763 * Let the pager know object is dead.
765 vm_pager_deallocate(object);
766 VM_OBJECT_WUNLOCK(object);
768 vm_object_destroy(object);
772 * Make the page read-only so that we can clear the object flags. However, if
773 * this is a nosync mmap then the object is likely to stay dirty so do not
774 * mess with the page and do not clear the object flags. Returns TRUE if the
775 * page should be flushed, and FALSE otherwise.
778 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
782 * If we have been asked to skip nosync pages and this is a
783 * nosync page, skip it. Note that the object flags were not
784 * cleared in this case so we do not have to set them.
786 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
787 *clearobjflags = FALSE;
790 pmap_remove_write(p);
791 return (p->dirty != 0);
796 * vm_object_page_clean
798 * Clean all dirty pages in the specified range of object. Leaves page
799 * on whatever queue it is currently on. If NOSYNC is set then do not
800 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
801 * leaving the object dirty.
803 * When stuffing pages asynchronously, allow clustering. XXX we need a
804 * synchronous clustering mode implementation.
806 * Odd semantics: if start == end, we clean everything.
808 * The object must be locked.
810 * Returns FALSE if some page from the range was not written, as
811 * reported by the pager, and TRUE otherwise.
814 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
818 vm_pindex_t pi, tend, tstart;
819 int curgeneration, n, pagerflags;
820 boolean_t clearobjflags, eio, res;
822 VM_OBJECT_ASSERT_WLOCKED(object);
825 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
826 * objects. The check below prevents the function from
827 * operating on non-vnode objects.
829 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
830 object->resident_page_count == 0)
833 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
834 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
835 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
837 tstart = OFF_TO_IDX(start);
838 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
839 clearobjflags = tstart == 0 && tend >= object->size;
843 curgeneration = object->generation;
845 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
849 np = TAILQ_NEXT(p, listq);
852 if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
853 if (object->generation != curgeneration) {
854 if ((flags & OBJPC_SYNC) != 0)
857 clearobjflags = FALSE;
859 np = vm_page_find_least(object, pi);
862 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
865 n = vm_object_page_collect_flush(object, p, pagerflags,
866 flags, &clearobjflags, &eio);
869 clearobjflags = FALSE;
871 if (object->generation != curgeneration) {
872 if ((flags & OBJPC_SYNC) != 0)
875 clearobjflags = FALSE;
879 * If the VOP_PUTPAGES() did a truncated write, so
880 * that even the first page of the run is not fully
881 * written, vm_pageout_flush() returns 0 as the run
882 * length. Since the condition that caused truncated
883 * write may be permanent, e.g. exhausted free space,
884 * accepting n == 0 would cause an infinite loop.
886 * Forwarding the iterator leaves the unwritten page
887 * behind, but there is not much we can do there if
888 * filesystem refuses to write it.
892 clearobjflags = FALSE;
894 np = vm_page_find_least(object, pi + n);
897 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
901 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
906 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
907 int flags, boolean_t *clearobjflags, boolean_t *eio)
909 vm_page_t ma[vm_pageout_page_count], p_first, tp;
910 int count, i, mreq, runlen;
912 vm_page_lock_assert(p, MA_NOTOWNED);
913 VM_OBJECT_ASSERT_WLOCKED(object);
918 for (tp = p; count < vm_pageout_page_count; count++) {
919 tp = vm_page_next(tp);
920 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
922 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
926 for (p_first = p; count < vm_pageout_page_count; count++) {
927 tp = vm_page_prev(p_first);
928 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
930 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
936 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
939 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
944 * Note that there is absolutely no sense in writing out
945 * anonymous objects, so we track down the vnode object
947 * We invalidate (remove) all pages from the address space
948 * for semantic correctness.
950 * If the backing object is a device object with unmanaged pages, then any
951 * mappings to the specified range of pages must be removed before this
952 * function is called.
954 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
955 * may start out with a NULL object.
958 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
959 boolean_t syncio, boolean_t invalidate)
961 vm_object_t backing_object;
964 int error, flags, fsync_after;
971 VM_OBJECT_WLOCK(object);
972 while ((backing_object = object->backing_object) != NULL) {
973 VM_OBJECT_WLOCK(backing_object);
974 offset += object->backing_object_offset;
975 VM_OBJECT_WUNLOCK(object);
976 object = backing_object;
977 if (object->size < OFF_TO_IDX(offset + size))
978 size = IDX_TO_OFF(object->size) - offset;
981 * Flush pages if writing is allowed, invalidate them
982 * if invalidation requested. Pages undergoing I/O
983 * will be ignored by vm_object_page_remove().
985 * We cannot lock the vnode and then wait for paging
986 * to complete without deadlocking against vm_fault.
987 * Instead we simply call vm_object_page_remove() and
988 * allow it to block internally on a page-by-page
989 * basis when it encounters pages undergoing async
992 if (object->type == OBJT_VNODE &&
993 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
995 VM_OBJECT_WUNLOCK(object);
996 (void) vn_start_write(vp, &mp, V_WAIT);
997 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
998 if (syncio && !invalidate && offset == 0 &&
999 OFF_TO_IDX(size) == object->size) {
1001 * If syncing the whole mapping of the file,
1002 * it is faster to schedule all the writes in
1003 * async mode, also allowing the clustering,
1004 * and then wait for i/o to complete.
1009 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1010 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1011 fsync_after = FALSE;
1013 VM_OBJECT_WLOCK(object);
1014 res = vm_object_page_clean(object, offset, offset + size,
1016 VM_OBJECT_WUNLOCK(object);
1018 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1020 vn_finished_write(mp);
1023 VM_OBJECT_WLOCK(object);
1025 if ((object->type == OBJT_VNODE ||
1026 object->type == OBJT_DEVICE) && invalidate) {
1027 if (object->type == OBJT_DEVICE)
1029 * The option OBJPR_NOTMAPPED must be passed here
1030 * because vm_object_page_remove() cannot remove
1031 * unmanaged mappings.
1033 flags = OBJPR_NOTMAPPED;
1037 flags = OBJPR_CLEANONLY;
1038 vm_object_page_remove(object, OFF_TO_IDX(offset),
1039 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1041 VM_OBJECT_WUNLOCK(object);
1046 * vm_object_madvise:
1048 * Implements the madvise function at the object/page level.
1050 * MADV_WILLNEED (any object)
1052 * Activate the specified pages if they are resident.
1054 * MADV_DONTNEED (any object)
1056 * Deactivate the specified pages if they are resident.
1058 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1059 * OBJ_ONEMAPPING only)
1061 * Deactivate and clean the specified pages if they are
1062 * resident. This permits the process to reuse the pages
1063 * without faulting or the kernel to reclaim the pages
1067 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1070 vm_pindex_t tpindex;
1071 vm_object_t backing_object, tobject;
1076 VM_OBJECT_WLOCK(object);
1078 * Locate and adjust resident pages
1080 for (; pindex < end; pindex += 1) {
1086 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1087 * and those pages must be OBJ_ONEMAPPING.
1089 if (advise == MADV_FREE) {
1090 if ((tobject->type != OBJT_DEFAULT &&
1091 tobject->type != OBJT_SWAP) ||
1092 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1093 goto unlock_tobject;
1095 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1096 goto unlock_tobject;
1097 m = vm_page_lookup(tobject, tpindex);
1098 if (m == NULL && advise == MADV_WILLNEED) {
1100 * If the page is cached, reactivate it.
1102 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1107 * There may be swap even if there is no backing page
1109 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1110 swap_pager_freespace(tobject, tpindex, 1);
1114 backing_object = tobject->backing_object;
1115 if (backing_object == NULL)
1116 goto unlock_tobject;
1117 VM_OBJECT_WLOCK(backing_object);
1118 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1119 if (tobject != object)
1120 VM_OBJECT_WUNLOCK(tobject);
1121 tobject = backing_object;
1123 } else if (m->valid != VM_PAGE_BITS_ALL)
1124 goto unlock_tobject;
1126 * If the page is not in a normal state, skip it.
1129 if (m->hold_count != 0 || m->wire_count != 0) {
1131 goto unlock_tobject;
1133 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1134 ("vm_object_madvise: page %p is fictitious", m));
1135 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1136 ("vm_object_madvise: page %p is not managed", m));
1137 if ((m->oflags & VPO_BUSY) || m->busy) {
1138 if (advise == MADV_WILLNEED) {
1140 * Reference the page before unlocking and
1141 * sleeping so that the page daemon is less
1142 * likely to reclaim it.
1144 vm_page_aflag_set(m, PGA_REFERENCED);
1147 if (object != tobject)
1148 VM_OBJECT_WUNLOCK(object);
1149 m->oflags |= VPO_WANTED;
1150 VM_OBJECT_SLEEP(tobject, m, PDROP | PVM, "madvpo", 0);
1151 VM_OBJECT_WLOCK(object);
1154 if (advise == MADV_WILLNEED) {
1155 vm_page_activate(m);
1156 } else if (advise == MADV_DONTNEED) {
1157 vm_page_dontneed(m);
1158 } else if (advise == MADV_FREE) {
1160 * Mark the page clean. This will allow the page
1161 * to be freed up by the system. However, such pages
1162 * are often reused quickly by malloc()/free()
1163 * so we do not do anything that would cause
1164 * a page fault if we can help it.
1166 * Specifically, we do not try to actually free
1167 * the page now nor do we try to put it in the
1168 * cache (which would cause a page fault on reuse).
1170 * But we do make the page is freeable as we
1171 * can without actually taking the step of unmapping
1174 pmap_clear_modify(m);
1177 vm_page_dontneed(m);
1180 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1181 swap_pager_freespace(tobject, tpindex, 1);
1183 if (tobject != object)
1184 VM_OBJECT_WUNLOCK(tobject);
1186 VM_OBJECT_WUNLOCK(object);
1192 * Create a new object which is backed by the
1193 * specified existing object range. The source
1194 * object reference is deallocated.
1196 * The new object and offset into that object
1197 * are returned in the source parameters.
1201 vm_object_t *object, /* IN/OUT */
1202 vm_ooffset_t *offset, /* IN/OUT */
1211 * Don't create the new object if the old object isn't shared.
1213 if (source != NULL) {
1214 VM_OBJECT_WLOCK(source);
1215 if (source->ref_count == 1 &&
1216 source->handle == NULL &&
1217 (source->type == OBJT_DEFAULT ||
1218 source->type == OBJT_SWAP)) {
1219 VM_OBJECT_WUNLOCK(source);
1222 VM_OBJECT_WUNLOCK(source);
1226 * Allocate a new object with the given length.
1228 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1231 * The new object shadows the source object, adding a reference to it.
1232 * Our caller changes his reference to point to the new object,
1233 * removing a reference to the source object. Net result: no change
1234 * of reference count.
1236 * Try to optimize the result object's page color when shadowing
1237 * in order to maintain page coloring consistency in the combined
1240 result->backing_object = source;
1242 * Store the offset into the source object, and fix up the offset into
1245 result->backing_object_offset = *offset;
1246 if (source != NULL) {
1247 VM_OBJECT_WLOCK(source);
1248 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1249 source->shadow_count++;
1250 #if VM_NRESERVLEVEL > 0
1251 result->flags |= source->flags & OBJ_COLORED;
1252 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1253 ((1 << (VM_NFREEORDER - 1)) - 1);
1255 VM_OBJECT_WUNLOCK(source);
1260 * Return the new things
1269 * Split the pages in a map entry into a new object. This affords
1270 * easier removal of unused pages, and keeps object inheritance from
1271 * being a negative impact on memory usage.
1274 vm_object_split(vm_map_entry_t entry)
1276 vm_page_t m, m_next;
1277 vm_object_t orig_object, new_object, source;
1278 vm_pindex_t idx, offidxstart;
1281 orig_object = entry->object.vm_object;
1282 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1284 if (orig_object->ref_count <= 1)
1286 VM_OBJECT_WUNLOCK(orig_object);
1288 offidxstart = OFF_TO_IDX(entry->offset);
1289 size = atop(entry->end - entry->start);
1292 * If swap_pager_copy() is later called, it will convert new_object
1293 * into a swap object.
1295 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1298 * At this point, the new object is still private, so the order in
1299 * which the original and new objects are locked does not matter.
1301 VM_OBJECT_WLOCK(new_object);
1302 VM_OBJECT_WLOCK(orig_object);
1303 source = orig_object->backing_object;
1304 if (source != NULL) {
1305 VM_OBJECT_WLOCK(source);
1306 if ((source->flags & OBJ_DEAD) != 0) {
1307 VM_OBJECT_WUNLOCK(source);
1308 VM_OBJECT_WUNLOCK(orig_object);
1309 VM_OBJECT_WUNLOCK(new_object);
1310 vm_object_deallocate(new_object);
1311 VM_OBJECT_WLOCK(orig_object);
1314 LIST_INSERT_HEAD(&source->shadow_head,
1315 new_object, shadow_list);
1316 source->shadow_count++;
1317 vm_object_reference_locked(source); /* for new_object */
1318 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1319 VM_OBJECT_WUNLOCK(source);
1320 new_object->backing_object_offset =
1321 orig_object->backing_object_offset + entry->offset;
1322 new_object->backing_object = source;
1324 if (orig_object->cred != NULL) {
1325 new_object->cred = orig_object->cred;
1326 crhold(orig_object->cred);
1327 new_object->charge = ptoa(size);
1328 KASSERT(orig_object->charge >= ptoa(size),
1329 ("orig_object->charge < 0"));
1330 orig_object->charge -= ptoa(size);
1333 m = vm_page_find_least(orig_object, offidxstart);
1334 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1336 m_next = TAILQ_NEXT(m, listq);
1339 * We must wait for pending I/O to complete before we can
1342 * We do not have to VM_PROT_NONE the page as mappings should
1343 * not be changed by this operation.
1345 if ((m->oflags & VPO_BUSY) || m->busy) {
1346 VM_OBJECT_WUNLOCK(new_object);
1347 m->oflags |= VPO_WANTED;
1348 VM_OBJECT_SLEEP(orig_object, m, PVM, "spltwt", 0);
1349 VM_OBJECT_WLOCK(new_object);
1352 #if VM_NRESERVLEVEL > 0
1354 * If some of the reservation's allocated pages remain with
1355 * the original object, then transferring the reservation to
1356 * the new object is neither particularly beneficial nor
1357 * particularly harmful as compared to leaving the reservation
1358 * with the original object. If, however, all of the
1359 * reservation's allocated pages are transferred to the new
1360 * object, then transferring the reservation is typically
1361 * beneficial. Determining which of these two cases applies
1362 * would be more costly than unconditionally renaming the
1365 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1368 vm_page_rename(m, new_object, idx);
1370 /* page automatically made dirty by rename and cache handled */
1373 if (orig_object->type == OBJT_SWAP) {
1375 * swap_pager_copy() can sleep, in which case the orig_object's
1376 * and new_object's locks are released and reacquired.
1378 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1381 * Transfer any cached pages from orig_object to new_object.
1382 * If swap_pager_copy() found swapped out pages within the
1383 * specified range of orig_object, then it changed
1384 * new_object's type to OBJT_SWAP when it transferred those
1385 * pages to new_object. Otherwise, new_object's type
1386 * should still be OBJT_DEFAULT and orig_object should not
1387 * contain any cached pages within the specified range.
1389 if (__predict_false(!vm_object_cache_is_empty(orig_object)))
1390 vm_page_cache_transfer(orig_object, offidxstart,
1393 VM_OBJECT_WUNLOCK(orig_object);
1394 TAILQ_FOREACH(m, &new_object->memq, listq)
1396 VM_OBJECT_WUNLOCK(new_object);
1397 entry->object.vm_object = new_object;
1398 entry->offset = 0LL;
1399 vm_object_deallocate(orig_object);
1400 VM_OBJECT_WLOCK(new_object);
1403 #define OBSC_TEST_ALL_SHADOWED 0x0001
1404 #define OBSC_COLLAPSE_NOWAIT 0x0002
1405 #define OBSC_COLLAPSE_WAIT 0x0004
1408 vm_object_backing_scan(vm_object_t object, int op)
1412 vm_object_t backing_object;
1413 vm_pindex_t backing_offset_index;
1415 VM_OBJECT_ASSERT_WLOCKED(object);
1416 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1418 backing_object = object->backing_object;
1419 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1422 * Initial conditions
1424 if (op & OBSC_TEST_ALL_SHADOWED) {
1426 * We do not want to have to test for the existence of cache
1427 * or swap pages in the backing object. XXX but with the
1428 * new swapper this would be pretty easy to do.
1430 * XXX what about anonymous MAP_SHARED memory that hasn't
1431 * been ZFOD faulted yet? If we do not test for this, the
1432 * shadow test may succeed! XXX
1434 if (backing_object->type != OBJT_DEFAULT) {
1438 if (op & OBSC_COLLAPSE_WAIT) {
1439 vm_object_set_flag(backing_object, OBJ_DEAD);
1445 p = TAILQ_FIRST(&backing_object->memq);
1447 vm_page_t next = TAILQ_NEXT(p, listq);
1448 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1450 if (op & OBSC_TEST_ALL_SHADOWED) {
1454 * Ignore pages outside the parent object's range
1455 * and outside the parent object's mapping of the
1458 * note that we do not busy the backing object's
1462 p->pindex < backing_offset_index ||
1463 new_pindex >= object->size
1470 * See if the parent has the page or if the parent's
1471 * object pager has the page. If the parent has the
1472 * page but the page is not valid, the parent's
1473 * object pager must have the page.
1475 * If this fails, the parent does not completely shadow
1476 * the object and we might as well give up now.
1479 pp = vm_page_lookup(object, new_pindex);
1481 (pp == NULL || pp->valid == 0) &&
1482 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1490 * Check for busy page
1492 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1495 if (op & OBSC_COLLAPSE_NOWAIT) {
1496 if ((p->oflags & VPO_BUSY) ||
1502 } else if (op & OBSC_COLLAPSE_WAIT) {
1503 if ((p->oflags & VPO_BUSY) || p->busy) {
1504 VM_OBJECT_WUNLOCK(object);
1505 p->oflags |= VPO_WANTED;
1506 VM_OBJECT_SLEEP(backing_object, p,
1507 PDROP | PVM, "vmocol", 0);
1508 VM_OBJECT_WLOCK(object);
1509 VM_OBJECT_WLOCK(backing_object);
1511 * If we slept, anything could have
1512 * happened. Since the object is
1513 * marked dead, the backing offset
1514 * should not have changed so we
1515 * just restart our scan.
1517 p = TAILQ_FIRST(&backing_object->memq);
1523 p->object == backing_object,
1524 ("vm_object_backing_scan: object mismatch")
1528 * Destroy any associated swap
1530 if (backing_object->type == OBJT_SWAP) {
1531 swap_pager_freespace(
1539 p->pindex < backing_offset_index ||
1540 new_pindex >= object->size
1543 * Page is out of the parent object's range, we
1544 * can simply destroy it.
1547 KASSERT(!pmap_page_is_mapped(p),
1548 ("freeing mapped page %p", p));
1549 if (p->wire_count == 0)
1558 pp = vm_page_lookup(object, new_pindex);
1560 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1561 (pp != NULL && pp->valid == 0)
1564 * The page in the parent is not (yet) valid.
1565 * We don't know anything about the state of
1566 * the original page. It might be mapped,
1567 * so we must avoid the next if here.
1569 * This is due to a race in vm_fault() where
1570 * we must unbusy the original (backing_obj)
1571 * page before we can (re)lock the parent.
1572 * Hence we can get here.
1579 vm_pager_has_page(object, new_pindex, NULL, NULL)
1582 * page already exists in parent OR swap exists
1583 * for this location in the parent. Destroy
1584 * the original page from the backing object.
1586 * Leave the parent's page alone
1589 KASSERT(!pmap_page_is_mapped(p),
1590 ("freeing mapped page %p", p));
1591 if (p->wire_count == 0)
1600 #if VM_NRESERVLEVEL > 0
1602 * Rename the reservation.
1604 vm_reserv_rename(p, object, backing_object,
1605 backing_offset_index);
1609 * Page does not exist in parent, rename the
1610 * page from the backing object to the main object.
1612 * If the page was mapped to a process, it can remain
1613 * mapped through the rename.
1616 vm_page_rename(p, object, new_pindex);
1618 /* page automatically made dirty by rename */
1627 * this version of collapse allows the operation to occur earlier and
1628 * when paging_in_progress is true for an object... This is not a complete
1629 * operation, but should plug 99.9% of the rest of the leaks.
1632 vm_object_qcollapse(vm_object_t object)
1634 vm_object_t backing_object = object->backing_object;
1636 VM_OBJECT_ASSERT_WLOCKED(object);
1637 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1639 if (backing_object->ref_count != 1)
1642 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1646 * vm_object_collapse:
1648 * Collapse an object with the object backing it.
1649 * Pages in the backing object are moved into the
1650 * parent, and the backing object is deallocated.
1653 vm_object_collapse(vm_object_t object)
1655 VM_OBJECT_ASSERT_WLOCKED(object);
1658 vm_object_t backing_object;
1661 * Verify that the conditions are right for collapse:
1663 * The object exists and the backing object exists.
1665 if ((backing_object = object->backing_object) == NULL)
1669 * we check the backing object first, because it is most likely
1672 VM_OBJECT_WLOCK(backing_object);
1673 if (backing_object->handle != NULL ||
1674 (backing_object->type != OBJT_DEFAULT &&
1675 backing_object->type != OBJT_SWAP) ||
1676 (backing_object->flags & OBJ_DEAD) ||
1677 object->handle != NULL ||
1678 (object->type != OBJT_DEFAULT &&
1679 object->type != OBJT_SWAP) ||
1680 (object->flags & OBJ_DEAD)) {
1681 VM_OBJECT_WUNLOCK(backing_object);
1686 object->paging_in_progress != 0 ||
1687 backing_object->paging_in_progress != 0
1689 vm_object_qcollapse(object);
1690 VM_OBJECT_WUNLOCK(backing_object);
1694 * We know that we can either collapse the backing object (if
1695 * the parent is the only reference to it) or (perhaps) have
1696 * the parent bypass the object if the parent happens to shadow
1697 * all the resident pages in the entire backing object.
1699 * This is ignoring pager-backed pages such as swap pages.
1700 * vm_object_backing_scan fails the shadowing test in this
1703 if (backing_object->ref_count == 1) {
1705 * If there is exactly one reference to the backing
1706 * object, we can collapse it into the parent.
1708 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1710 #if VM_NRESERVLEVEL > 0
1712 * Break any reservations from backing_object.
1714 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1715 vm_reserv_break_all(backing_object);
1719 * Move the pager from backing_object to object.
1721 if (backing_object->type == OBJT_SWAP) {
1723 * swap_pager_copy() can sleep, in which case
1724 * the backing_object's and object's locks are
1725 * released and reacquired.
1726 * Since swap_pager_copy() is being asked to
1727 * destroy the source, it will change the
1728 * backing_object's type to OBJT_DEFAULT.
1733 OFF_TO_IDX(object->backing_object_offset), TRUE);
1736 * Free any cached pages from backing_object.
1738 if (__predict_false(
1739 !vm_object_cache_is_empty(backing_object)))
1740 vm_page_cache_free(backing_object, 0, 0);
1743 * Object now shadows whatever backing_object did.
1744 * Note that the reference to
1745 * backing_object->backing_object moves from within
1746 * backing_object to within object.
1748 LIST_REMOVE(object, shadow_list);
1749 backing_object->shadow_count--;
1750 if (backing_object->backing_object) {
1751 VM_OBJECT_WLOCK(backing_object->backing_object);
1752 LIST_REMOVE(backing_object, shadow_list);
1754 &backing_object->backing_object->shadow_head,
1755 object, shadow_list);
1757 * The shadow_count has not changed.
1759 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1761 object->backing_object = backing_object->backing_object;
1762 object->backing_object_offset +=
1763 backing_object->backing_object_offset;
1766 * Discard backing_object.
1768 * Since the backing object has no pages, no pager left,
1769 * and no object references within it, all that is
1770 * necessary is to dispose of it.
1772 KASSERT(backing_object->ref_count == 1, (
1773 "backing_object %p was somehow re-referenced during collapse!",
1775 VM_OBJECT_WUNLOCK(backing_object);
1776 vm_object_destroy(backing_object);
1780 vm_object_t new_backing_object;
1783 * If we do not entirely shadow the backing object,
1784 * there is nothing we can do so we give up.
1786 if (object->resident_page_count != object->size &&
1787 vm_object_backing_scan(object,
1788 OBSC_TEST_ALL_SHADOWED) == 0) {
1789 VM_OBJECT_WUNLOCK(backing_object);
1794 * Make the parent shadow the next object in the
1795 * chain. Deallocating backing_object will not remove
1796 * it, since its reference count is at least 2.
1798 LIST_REMOVE(object, shadow_list);
1799 backing_object->shadow_count--;
1801 new_backing_object = backing_object->backing_object;
1802 if ((object->backing_object = new_backing_object) != NULL) {
1803 VM_OBJECT_WLOCK(new_backing_object);
1805 &new_backing_object->shadow_head,
1809 new_backing_object->shadow_count++;
1810 vm_object_reference_locked(new_backing_object);
1811 VM_OBJECT_WUNLOCK(new_backing_object);
1812 object->backing_object_offset +=
1813 backing_object->backing_object_offset;
1817 * Drop the reference count on backing_object. Since
1818 * its ref_count was at least 2, it will not vanish.
1820 backing_object->ref_count--;
1821 VM_OBJECT_WUNLOCK(backing_object);
1826 * Try again with this object's new backing object.
1832 * vm_object_page_remove:
1834 * For the given object, either frees or invalidates each of the
1835 * specified pages. In general, a page is freed. However, if a page is
1836 * wired for any reason other than the existence of a managed, wired
1837 * mapping, then it may be invalidated but not removed from the object.
1838 * Pages are specified by the given range ["start", "end") and the option
1839 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1840 * extends from "start" to the end of the object. If the option
1841 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1842 * specified range are affected. If the option OBJPR_NOTMAPPED is
1843 * specified, then the pages within the specified range must have no
1844 * mappings. Otherwise, if this option is not specified, any mappings to
1845 * the specified pages are removed before the pages are freed or
1848 * In general, this operation should only be performed on objects that
1849 * contain managed pages. There are, however, two exceptions. First, it
1850 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1851 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1852 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1853 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1855 * The object must be locked.
1858 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1864 VM_OBJECT_ASSERT_WLOCKED(object);
1865 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1866 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1867 ("vm_object_page_remove: illegal options for object %p", object));
1868 if (object->resident_page_count == 0)
1870 vm_object_pip_add(object, 1);
1872 p = vm_page_find_least(object, start);
1875 * Here, the variable "p" is either (1) the page with the least pindex
1876 * greater than or equal to the parameter "start" or (2) NULL.
1878 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1879 next = TAILQ_NEXT(p, listq);
1882 * If the page is wired for any reason besides the existence
1883 * of managed, wired mappings, then it cannot be freed. For
1884 * example, fictitious pages, which represent device memory,
1885 * are inherently wired and cannot be freed. They can,
1886 * however, be invalidated if the option OBJPR_CLEANONLY is
1890 if ((wirings = p->wire_count) != 0 &&
1891 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1892 if ((options & OBJPR_NOTMAPPED) == 0) {
1894 /* Account for removal of wired mappings. */
1896 p->wire_count -= wirings;
1898 if ((options & OBJPR_CLEANONLY) == 0) {
1905 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1907 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1908 ("vm_object_page_remove: page %p is fictitious", p));
1909 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1910 if ((options & OBJPR_NOTMAPPED) == 0)
1911 pmap_remove_write(p);
1917 if ((options & OBJPR_NOTMAPPED) == 0) {
1919 /* Account for removal of wired mappings. */
1921 KASSERT(p->wire_count == wirings,
1922 ("inconsistent wire count %d %d %p",
1923 p->wire_count, wirings, p));
1925 atomic_subtract_int(&cnt.v_wire_count, 1);
1931 vm_object_pip_wakeup(object);
1933 if (__predict_false(!vm_object_cache_is_empty(object)))
1934 vm_page_cache_free(object, start, end);
1938 * vm_object_page_cache:
1940 * For the given object, attempt to move the specified clean
1941 * pages to the cache queue. If a page is wired for any reason,
1942 * then it will not be changed. Pages are specified by the given
1943 * range ["start", "end"). As a special case, if "end" is zero,
1944 * then the range extends from "start" to the end of the object.
1945 * Any mappings to the specified pages are removed before the
1946 * pages are moved to the cache queue.
1948 * This operation should only be performed on objects that
1949 * contain non-fictitious, managed pages.
1951 * The object must be locked.
1954 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1956 struct mtx *mtx, *new_mtx;
1959 VM_OBJECT_ASSERT_WLOCKED(object);
1960 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1961 ("vm_object_page_cache: illegal object %p", object));
1962 if (object->resident_page_count == 0)
1964 p = vm_page_find_least(object, start);
1967 * Here, the variable "p" is either (1) the page with the least pindex
1968 * greater than or equal to the parameter "start" or (2) NULL.
1971 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1972 next = TAILQ_NEXT(p, listq);
1975 * Avoid releasing and reacquiring the same page lock.
1977 new_mtx = vm_page_lockptr(p);
1978 if (mtx != new_mtx) {
1984 vm_page_try_to_cache(p);
1991 * Populate the specified range of the object with valid pages. Returns
1992 * TRUE if the range is successfully populated and FALSE otherwise.
1994 * Note: This function should be optimized to pass a larger array of
1995 * pages to vm_pager_get_pages() before it is applied to a non-
1996 * OBJT_DEVICE object.
1998 * The object must be locked.
2001 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2007 VM_OBJECT_ASSERT_WLOCKED(object);
2008 for (pindex = start; pindex < end; pindex++) {
2009 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
2011 if (m->valid != VM_PAGE_BITS_ALL) {
2013 rv = vm_pager_get_pages(object, ma, 1, 0);
2014 m = vm_page_lookup(object, pindex);
2017 if (rv != VM_PAGER_OK) {
2025 * Keep "m" busy because a subsequent iteration may unlock
2029 if (pindex > start) {
2030 m = vm_page_lookup(object, start);
2031 while (m != NULL && m->pindex < pindex) {
2033 m = TAILQ_NEXT(m, listq);
2036 return (pindex == end);
2040 * Routine: vm_object_coalesce
2041 * Function: Coalesces two objects backing up adjoining
2042 * regions of memory into a single object.
2044 * returns TRUE if objects were combined.
2046 * NOTE: Only works at the moment if the second object is NULL -
2047 * if it's not, which object do we lock first?
2050 * prev_object First object to coalesce
2051 * prev_offset Offset into prev_object
2052 * prev_size Size of reference to prev_object
2053 * next_size Size of reference to the second object
2054 * reserved Indicator that extension region has
2055 * swap accounted for
2058 * The object must *not* be locked.
2061 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2062 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2064 vm_pindex_t next_pindex;
2066 if (prev_object == NULL)
2068 VM_OBJECT_WLOCK(prev_object);
2069 if (prev_object->type != OBJT_DEFAULT &&
2070 prev_object->type != OBJT_SWAP) {
2071 VM_OBJECT_WUNLOCK(prev_object);
2076 * Try to collapse the object first
2078 vm_object_collapse(prev_object);
2081 * Can't coalesce if: . more than one reference . paged out . shadows
2082 * another object . has a copy elsewhere (any of which mean that the
2083 * pages not mapped to prev_entry may be in use anyway)
2085 if (prev_object->backing_object != NULL) {
2086 VM_OBJECT_WUNLOCK(prev_object);
2090 prev_size >>= PAGE_SHIFT;
2091 next_size >>= PAGE_SHIFT;
2092 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2094 if ((prev_object->ref_count > 1) &&
2095 (prev_object->size != next_pindex)) {
2096 VM_OBJECT_WUNLOCK(prev_object);
2101 * Account for the charge.
2103 if (prev_object->cred != NULL) {
2106 * If prev_object was charged, then this mapping,
2107 * althought not charged now, may become writable
2108 * later. Non-NULL cred in the object would prevent
2109 * swap reservation during enabling of the write
2110 * access, so reserve swap now. Failed reservation
2111 * cause allocation of the separate object for the map
2112 * entry, and swap reservation for this entry is
2113 * managed in appropriate time.
2115 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2116 prev_object->cred)) {
2119 prev_object->charge += ptoa(next_size);
2123 * Remove any pages that may still be in the object from a previous
2126 if (next_pindex < prev_object->size) {
2127 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2129 if (prev_object->type == OBJT_SWAP)
2130 swap_pager_freespace(prev_object,
2131 next_pindex, next_size);
2133 if (prev_object->cred != NULL) {
2134 KASSERT(prev_object->charge >=
2135 ptoa(prev_object->size - next_pindex),
2136 ("object %p overcharged 1 %jx %jx", prev_object,
2137 (uintmax_t)next_pindex, (uintmax_t)next_size));
2138 prev_object->charge -= ptoa(prev_object->size -
2145 * Extend the object if necessary.
2147 if (next_pindex + next_size > prev_object->size)
2148 prev_object->size = next_pindex + next_size;
2150 VM_OBJECT_WUNLOCK(prev_object);
2155 vm_object_set_writeable_dirty(vm_object_t object)
2158 VM_OBJECT_ASSERT_WLOCKED(object);
2159 if (object->type != OBJT_VNODE)
2161 object->generation++;
2162 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2164 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2167 #include "opt_ddb.h"
2169 #include <sys/kernel.h>
2171 #include <sys/cons.h>
2173 #include <ddb/ddb.h>
2176 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2179 vm_map_entry_t tmpe;
2187 tmpe = map->header.next;
2188 entcount = map->nentries;
2189 while (entcount-- && (tmpe != &map->header)) {
2190 if (_vm_object_in_map(map, object, tmpe)) {
2195 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2196 tmpm = entry->object.sub_map;
2197 tmpe = tmpm->header.next;
2198 entcount = tmpm->nentries;
2199 while (entcount-- && tmpe != &tmpm->header) {
2200 if (_vm_object_in_map(tmpm, object, tmpe)) {
2205 } else if ((obj = entry->object.vm_object) != NULL) {
2206 for (; obj; obj = obj->backing_object)
2207 if (obj == object) {
2215 vm_object_in_map(vm_object_t object)
2219 /* sx_slock(&allproc_lock); */
2220 FOREACH_PROC_IN_SYSTEM(p) {
2221 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2223 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2224 /* sx_sunlock(&allproc_lock); */
2228 /* sx_sunlock(&allproc_lock); */
2229 if (_vm_object_in_map(kernel_map, object, 0))
2231 if (_vm_object_in_map(kmem_map, object, 0))
2233 if (_vm_object_in_map(pager_map, object, 0))
2235 if (_vm_object_in_map(buffer_map, object, 0))
2240 DB_SHOW_COMMAND(vmochk, vm_object_check)
2245 * make sure that internal objs are in a map somewhere
2246 * and none have zero ref counts.
2248 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2249 if (object->handle == NULL &&
2250 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2251 if (object->ref_count == 0) {
2252 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2253 (long)object->size);
2255 if (!vm_object_in_map(object)) {
2257 "vmochk: internal obj is not in a map: "
2258 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2259 object->ref_count, (u_long)object->size,
2260 (u_long)object->size,
2261 (void *)object->backing_object);
2268 * vm_object_print: [ debug ]
2270 DB_SHOW_COMMAND(object, vm_object_print_static)
2272 /* XXX convert args. */
2273 vm_object_t object = (vm_object_t)addr;
2274 boolean_t full = have_addr;
2278 /* XXX count is an (unused) arg. Avoid shadowing it. */
2279 #define count was_count
2287 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2288 object, (int)object->type, (uintmax_t)object->size,
2289 object->resident_page_count, object->ref_count, object->flags,
2290 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2291 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2292 object->shadow_count,
2293 object->backing_object ? object->backing_object->ref_count : 0,
2294 object->backing_object, (uintmax_t)object->backing_object_offset);
2301 TAILQ_FOREACH(p, &object->memq, listq) {
2303 db_iprintf("memory:=");
2304 else if (count == 6) {
2312 db_printf("(off=0x%jx,page=0x%jx)",
2313 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2323 /* XXX need this non-static entry for calling from vm_map_print. */
2326 /* db_expr_t */ long addr,
2327 boolean_t have_addr,
2328 /* db_expr_t */ long count,
2331 vm_object_print_static(addr, have_addr, count, modif);
2334 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2339 vm_page_t m, prev_m;
2343 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2344 db_printf("new object: %p\n", (void *)object);
2355 TAILQ_FOREACH(m, &object->memq, listq) {
2356 if (m->pindex > 128)
2358 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2359 prev_m->pindex + 1 != m->pindex) {
2361 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2362 (long)fidx, rcount, (long)pa);
2374 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2379 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2380 (long)fidx, rcount, (long)pa);
2390 pa = VM_PAGE_TO_PHYS(m);
2394 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2395 (long)fidx, rcount, (long)pa);