2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94
35 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36 * All rights reserved.
38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
40 * Permission to use, copy, modify and distribute this software and
41 * its documentation is hereby granted, provided that both the copyright
42 * notice and this permission notice appear in all copies of the
43 * software, derivative works or modified versions, and any portions
44 * thereof, and that both notices appear in supporting documentation.
46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
50 * Carnegie Mellon requests users of this software to return to
52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
53 * School of Computer Science
54 * Carnegie Mellon University
55 * Pittsburgh PA 15213-3890
57 * any improvements or extensions that they make and grant Carnegie the
58 * rights to redistribute these changes.
62 * Virtual memory object module.
65 #include <sys/cdefs.h>
66 __FBSDID("$FreeBSD$");
70 #include <sys/param.h>
71 #include <sys/systm.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/mutex.h>
78 #include <sys/proc.h> /* for curproc, pageproc */
79 #include <sys/socket.h>
80 #include <sys/resourcevar.h>
81 #include <sys/rwlock.h>
82 #include <sys/vnode.h>
83 #include <sys/vmmeter.h>
87 #include <vm/vm_param.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pageout.h>
93 #include <vm/vm_pager.h>
94 #include <vm/swap_pager.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_extern.h>
97 #include <vm/vm_radix.h>
98 #include <vm/vm_reserv.h>
101 static int old_msync;
102 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
103 "Use old (insecure) msync behavior");
105 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
106 int pagerflags, int flags, boolean_t *clearobjflags,
108 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
109 boolean_t *clearobjflags);
110 static void vm_object_qcollapse(vm_object_t object);
111 static void vm_object_vndeallocate(vm_object_t object);
114 * Virtual memory objects maintain the actual data
115 * associated with allocated virtual memory. A given
116 * page of memory exists within exactly one object.
118 * An object is only deallocated when all "references"
119 * are given up. Only one "reference" to a given
120 * region of an object should be writeable.
122 * Associated with each object is a list of all resident
123 * memory pages belonging to that object; this list is
124 * maintained by the "vm_page" module, and locked by the object's
127 * Each object also records a "pager" routine which is
128 * used to retrieve (and store) pages to the proper backing
129 * storage. In addition, objects may be backed by other
130 * objects from which they were virtual-copied.
132 * The only items within the object structure which are
133 * modified after time of creation are:
134 * reference count locked by object's lock
135 * pager routine locked by object's lock
139 struct object_q vm_object_list;
140 struct mtx vm_object_list_mtx; /* lock for object list and count */
142 struct vm_object kernel_object_store;
143 struct vm_object kmem_object_store;
145 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
148 static long object_collapses;
149 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
150 &object_collapses, 0, "VM object collapses");
152 static long object_bypasses;
153 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
154 &object_bypasses, 0, "VM object bypasses");
156 static uma_zone_t obj_zone;
158 static int vm_object_zinit(void *mem, int size, int flags);
161 static void vm_object_zdtor(void *mem, int size, void *arg);
164 vm_object_zdtor(void *mem, int size, void *arg)
168 object = (vm_object_t)mem;
169 KASSERT(TAILQ_EMPTY(&object->memq),
170 ("object %p has resident pages in its memq", object));
171 KASSERT(vm_radix_is_empty(&object->rtree),
172 ("object %p has resident pages in its trie", object));
173 #if VM_NRESERVLEVEL > 0
174 KASSERT(LIST_EMPTY(&object->rvq),
175 ("object %p has reservations",
178 KASSERT(vm_object_cache_is_empty(object),
179 ("object %p has cached pages",
181 KASSERT(object->paging_in_progress == 0,
182 ("object %p paging_in_progress = %d",
183 object, object->paging_in_progress));
184 KASSERT(object->resident_page_count == 0,
185 ("object %p resident_page_count = %d",
186 object, object->resident_page_count));
187 KASSERT(object->shadow_count == 0,
188 ("object %p shadow_count = %d",
189 object, object->shadow_count));
194 vm_object_zinit(void *mem, int size, int flags)
198 object = (vm_object_t)mem;
199 bzero(&object->lock, sizeof(object->lock));
200 rw_init_flags(&object->lock, "vm object", RW_DUPOK);
202 /* These are true for any object that has been freed */
203 object->rtree.rt_root = 0;
204 object->rtree.rt_flags = 0;
205 object->paging_in_progress = 0;
206 object->resident_page_count = 0;
207 object->shadow_count = 0;
208 object->cache.rt_root = 0;
209 object->cache.rt_flags = 0;
214 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
217 TAILQ_INIT(&object->memq);
218 LIST_INIT(&object->shadow_head);
223 panic("_vm_object_allocate: can't create OBJT_DEAD");
226 object->flags = OBJ_ONEMAPPING;
230 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
233 object->flags = OBJ_FICTITIOUS;
236 object->flags = OBJ_UNMANAGED;
242 panic("_vm_object_allocate: type %d is undefined", type);
245 object->generation = 1;
246 object->ref_count = 1;
247 object->memattr = VM_MEMATTR_DEFAULT;
250 object->handle = NULL;
251 object->backing_object = NULL;
252 object->backing_object_offset = (vm_ooffset_t) 0;
253 #if VM_NRESERVLEVEL > 0
254 LIST_INIT(&object->rvq);
257 mtx_lock(&vm_object_list_mtx);
258 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
259 mtx_unlock(&vm_object_list_mtx);
265 * Initialize the VM objects module.
270 TAILQ_INIT(&vm_object_list);
271 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
273 rw_init(&kernel_object->lock, "kernel vm object");
274 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
276 #if VM_NRESERVLEVEL > 0
277 kernel_object->flags |= OBJ_COLORED;
278 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
281 rw_init(&kmem_object->lock, "kmem vm object");
282 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
284 #if VM_NRESERVLEVEL > 0
285 kmem_object->flags |= OBJ_COLORED;
286 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
290 * The lock portion of struct vm_object must be type stable due
291 * to vm_pageout_fallback_object_lock locking a vm object
292 * without holding any references to it.
294 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
300 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
306 vm_object_clear_flag(vm_object_t object, u_short bits)
309 VM_OBJECT_ASSERT_WLOCKED(object);
310 object->flags &= ~bits;
314 * Sets the default memory attribute for the specified object. Pages
315 * that are allocated to this object are by default assigned this memory
318 * Presently, this function must be called before any pages are allocated
319 * to the object. In the future, this requirement may be relaxed for
320 * "default" and "swap" objects.
323 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
326 VM_OBJECT_ASSERT_WLOCKED(object);
327 switch (object->type) {
335 if (!TAILQ_EMPTY(&object->memq))
336 return (KERN_FAILURE);
339 return (KERN_INVALID_ARGUMENT);
341 panic("vm_object_set_memattr: object %p is of undefined type",
344 object->memattr = memattr;
345 return (KERN_SUCCESS);
349 vm_object_pip_add(vm_object_t object, short i)
352 VM_OBJECT_ASSERT_WLOCKED(object);
353 object->paging_in_progress += i;
357 vm_object_pip_subtract(vm_object_t object, short i)
360 VM_OBJECT_ASSERT_WLOCKED(object);
361 object->paging_in_progress -= i;
365 vm_object_pip_wakeup(vm_object_t object)
368 VM_OBJECT_ASSERT_WLOCKED(object);
369 object->paging_in_progress--;
370 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
371 vm_object_clear_flag(object, OBJ_PIPWNT);
377 vm_object_pip_wakeupn(vm_object_t object, short i)
380 VM_OBJECT_ASSERT_WLOCKED(object);
382 object->paging_in_progress -= i;
383 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
384 vm_object_clear_flag(object, OBJ_PIPWNT);
390 vm_object_pip_wait(vm_object_t object, char *waitid)
393 VM_OBJECT_ASSERT_WLOCKED(object);
394 while (object->paging_in_progress) {
395 object->flags |= OBJ_PIPWNT;
396 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
401 * vm_object_allocate:
403 * Returns a new object with the given size.
406 vm_object_allocate(objtype_t type, vm_pindex_t size)
410 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
411 _vm_object_allocate(type, size, object);
417 * vm_object_reference:
419 * Gets another reference to the given object. Note: OBJ_DEAD
420 * objects can be referenced during final cleaning.
423 vm_object_reference(vm_object_t object)
427 VM_OBJECT_WLOCK(object);
428 vm_object_reference_locked(object);
429 VM_OBJECT_WUNLOCK(object);
433 * vm_object_reference_locked:
435 * Gets another reference to the given object.
437 * The object must be locked.
440 vm_object_reference_locked(vm_object_t object)
444 VM_OBJECT_ASSERT_WLOCKED(object);
446 if (object->type == OBJT_VNODE) {
453 * Handle deallocating an object of type OBJT_VNODE.
456 vm_object_vndeallocate(vm_object_t object)
458 struct vnode *vp = (struct vnode *) object->handle;
460 VM_OBJECT_ASSERT_WLOCKED(object);
461 KASSERT(object->type == OBJT_VNODE,
462 ("vm_object_vndeallocate: not a vnode object"));
463 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
465 if (object->ref_count == 0) {
466 vprint("vm_object_vndeallocate", vp);
467 panic("vm_object_vndeallocate: bad object reference count");
471 if (object->ref_count > 1) {
473 VM_OBJECT_WUNLOCK(object);
474 /* vrele may need the vnode lock. */
478 VM_OBJECT_WUNLOCK(object);
479 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
481 VM_OBJECT_WLOCK(object);
483 if (object->type == OBJT_DEAD) {
484 VM_OBJECT_WUNLOCK(object);
487 if (object->ref_count == 0)
489 VM_OBJECT_WUNLOCK(object);
496 * vm_object_deallocate:
498 * Release a reference to the specified object,
499 * gained either through a vm_object_allocate
500 * or a vm_object_reference call. When all references
501 * are gone, storage associated with this object
502 * may be relinquished.
504 * No object may be locked.
507 vm_object_deallocate(vm_object_t object)
512 while (object != NULL) {
513 VM_OBJECT_WLOCK(object);
514 if (object->type == OBJT_VNODE) {
515 vm_object_vndeallocate(object);
519 KASSERT(object->ref_count != 0,
520 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
523 * If the reference count goes to 0 we start calling
524 * vm_object_terminate() on the object chain.
525 * A ref count of 1 may be a special case depending on the
526 * shadow count being 0 or 1.
529 if (object->ref_count > 1) {
530 VM_OBJECT_WUNLOCK(object);
532 } else if (object->ref_count == 1) {
533 if (object->type == OBJT_SWAP &&
534 (object->flags & OBJ_TMPFS) != 0) {
535 vp = object->un_pager.swp.swp_tmpfs;
537 VM_OBJECT_WUNLOCK(object);
538 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
540 VM_OBJECT_WLOCK(object);
541 if (object->type == OBJT_DEAD ||
542 object->ref_count != 1) {
543 VM_OBJECT_WUNLOCK(object);
547 if ((object->flags & OBJ_TMPFS) != 0)
551 if (object->shadow_count == 0 &&
552 object->handle == NULL &&
553 (object->type == OBJT_DEFAULT ||
554 (object->type == OBJT_SWAP &&
555 (object->flags & OBJ_TMPFS) == 0))) {
556 vm_object_set_flag(object, OBJ_ONEMAPPING);
557 } else if ((object->shadow_count == 1) &&
558 (object->handle == NULL) &&
559 (object->type == OBJT_DEFAULT ||
560 object->type == OBJT_SWAP)) {
561 KASSERT((object->flags & OBJ_TMPFS) == 0,
562 ("shadowed tmpfs v_object %p", object));
565 robject = LIST_FIRST(&object->shadow_head);
566 KASSERT(robject != NULL,
567 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
569 object->shadow_count));
570 if (!VM_OBJECT_TRYWLOCK(robject)) {
572 * Avoid a potential deadlock.
575 VM_OBJECT_WUNLOCK(object);
577 * More likely than not the thread
578 * holding robject's lock has lower
579 * priority than the current thread.
580 * Let the lower priority thread run.
586 * Collapse object into its shadow unless its
587 * shadow is dead. In that case, object will
588 * be deallocated by the thread that is
589 * deallocating its shadow.
591 if ((robject->flags & OBJ_DEAD) == 0 &&
592 (robject->handle == NULL) &&
593 (robject->type == OBJT_DEFAULT ||
594 robject->type == OBJT_SWAP)) {
596 robject->ref_count++;
598 if (robject->paging_in_progress) {
599 VM_OBJECT_WUNLOCK(object);
600 vm_object_pip_wait(robject,
602 temp = robject->backing_object;
603 if (object == temp) {
604 VM_OBJECT_WLOCK(object);
607 } else if (object->paging_in_progress) {
608 VM_OBJECT_WUNLOCK(robject);
609 object->flags |= OBJ_PIPWNT;
610 VM_OBJECT_SLEEP(object, object,
611 PDROP | PVM, "objde2", 0);
612 VM_OBJECT_WLOCK(robject);
613 temp = robject->backing_object;
614 if (object == temp) {
615 VM_OBJECT_WLOCK(object);
619 VM_OBJECT_WUNLOCK(object);
621 if (robject->ref_count == 1) {
622 robject->ref_count--;
627 vm_object_collapse(object);
628 VM_OBJECT_WUNLOCK(object);
631 VM_OBJECT_WUNLOCK(robject);
633 VM_OBJECT_WUNLOCK(object);
637 temp = object->backing_object;
639 VM_OBJECT_WLOCK(temp);
640 LIST_REMOVE(object, shadow_list);
641 temp->shadow_count--;
642 VM_OBJECT_WUNLOCK(temp);
643 object->backing_object = NULL;
646 * Don't double-terminate, we could be in a termination
647 * recursion due to the terminate having to sync data
650 if ((object->flags & OBJ_DEAD) == 0)
651 vm_object_terminate(object);
653 VM_OBJECT_WUNLOCK(object);
659 * vm_object_destroy removes the object from the global object list
660 * and frees the space for the object.
663 vm_object_destroy(vm_object_t object)
667 * Remove the object from the global object list.
669 mtx_lock(&vm_object_list_mtx);
670 TAILQ_REMOVE(&vm_object_list, object, object_list);
671 mtx_unlock(&vm_object_list_mtx);
674 * Release the allocation charge.
676 if (object->cred != NULL) {
677 KASSERT(object->type == OBJT_DEFAULT ||
678 object->type == OBJT_SWAP,
679 ("vm_object_terminate: non-swap obj %p has cred",
681 swap_release_by_cred(object->charge, object->cred);
683 crfree(object->cred);
688 * Free the space for the object.
690 uma_zfree(obj_zone, object);
694 * vm_object_terminate actually destroys the specified object, freeing
695 * up all previously used resources.
697 * The object must be locked.
698 * This routine may block.
701 vm_object_terminate(vm_object_t object)
705 VM_OBJECT_ASSERT_WLOCKED(object);
708 * Make sure no one uses us.
710 vm_object_set_flag(object, OBJ_DEAD);
713 * wait for the pageout daemon to be done with the object
715 vm_object_pip_wait(object, "objtrm");
717 KASSERT(!object->paging_in_progress,
718 ("vm_object_terminate: pageout in progress"));
721 * Clean and free the pages, as appropriate. All references to the
722 * object are gone, so we don't need to lock it.
724 if (object->type == OBJT_VNODE) {
725 struct vnode *vp = (struct vnode *)object->handle;
728 * Clean pages and flush buffers.
730 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
731 VM_OBJECT_WUNLOCK(object);
733 vinvalbuf(vp, V_SAVE, 0, 0);
735 VM_OBJECT_WLOCK(object);
738 KASSERT(object->ref_count == 0,
739 ("vm_object_terminate: object with references, ref_count=%d",
743 * Free any remaining pageable pages. This also removes them from the
744 * paging queues. However, don't free wired pages, just remove them
745 * from the object. Rather than incrementally removing each page from
746 * the object, the page and object are reset to any empty state.
748 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
749 vm_page_assert_unbusied(p);
752 * Optimize the page's removal from the object by resetting
753 * its "object" field. Specifically, if the page is not
754 * wired, then the effect of this assignment is that
755 * vm_page_free()'s call to vm_page_remove() will return
756 * immediately without modifying the page or the object.
759 if (p->wire_count == 0) {
761 PCPU_INC(cnt.v_pfree);
766 * If the object contained any pages, then reset it to an empty state.
767 * None of the object's fields, including "resident_page_count", were
768 * modified by the preceding loop.
770 if (object->resident_page_count != 0) {
771 vm_radix_reclaim_allnodes(&object->rtree);
772 TAILQ_INIT(&object->memq);
773 object->resident_page_count = 0;
774 if (object->type == OBJT_VNODE)
775 vdrop(object->handle);
778 #if VM_NRESERVLEVEL > 0
779 if (__predict_false(!LIST_EMPTY(&object->rvq)))
780 vm_reserv_break_all(object);
782 if (__predict_false(!vm_object_cache_is_empty(object)))
783 vm_page_cache_free(object, 0, 0);
786 * Let the pager know object is dead.
788 vm_pager_deallocate(object);
789 VM_OBJECT_WUNLOCK(object);
791 vm_object_destroy(object);
795 * Make the page read-only so that we can clear the object flags. However, if
796 * this is a nosync mmap then the object is likely to stay dirty so do not
797 * mess with the page and do not clear the object flags. Returns TRUE if the
798 * page should be flushed, and FALSE otherwise.
801 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
805 * If we have been asked to skip nosync pages and this is a
806 * nosync page, skip it. Note that the object flags were not
807 * cleared in this case so we do not have to set them.
809 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
810 *clearobjflags = FALSE;
813 pmap_remove_write(p);
814 return (p->dirty != 0);
819 * vm_object_page_clean
821 * Clean all dirty pages in the specified range of object. Leaves page
822 * on whatever queue it is currently on. If NOSYNC is set then do not
823 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
824 * leaving the object dirty.
826 * When stuffing pages asynchronously, allow clustering. XXX we need a
827 * synchronous clustering mode implementation.
829 * Odd semantics: if start == end, we clean everything.
831 * The object must be locked.
833 * Returns FALSE if some page from the range was not written, as
834 * reported by the pager, and TRUE otherwise.
837 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
841 vm_pindex_t pi, tend, tstart;
842 int curgeneration, n, pagerflags;
843 boolean_t clearobjflags, eio, res;
845 VM_OBJECT_ASSERT_WLOCKED(object);
848 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
849 * objects. The check below prevents the function from
850 * operating on non-vnode objects.
852 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
853 object->resident_page_count == 0)
856 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
857 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
858 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
860 tstart = OFF_TO_IDX(start);
861 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
862 clearobjflags = tstart == 0 && tend >= object->size;
866 curgeneration = object->generation;
868 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
872 np = TAILQ_NEXT(p, listq);
875 if (vm_page_sleep_if_busy(p, "vpcwai")) {
876 if (object->generation != curgeneration) {
877 if ((flags & OBJPC_SYNC) != 0)
880 clearobjflags = FALSE;
882 np = vm_page_find_least(object, pi);
885 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
888 n = vm_object_page_collect_flush(object, p, pagerflags,
889 flags, &clearobjflags, &eio);
892 clearobjflags = FALSE;
894 if (object->generation != curgeneration) {
895 if ((flags & OBJPC_SYNC) != 0)
898 clearobjflags = FALSE;
902 * If the VOP_PUTPAGES() did a truncated write, so
903 * that even the first page of the run is not fully
904 * written, vm_pageout_flush() returns 0 as the run
905 * length. Since the condition that caused truncated
906 * write may be permanent, e.g. exhausted free space,
907 * accepting n == 0 would cause an infinite loop.
909 * Forwarding the iterator leaves the unwritten page
910 * behind, but there is not much we can do there if
911 * filesystem refuses to write it.
915 clearobjflags = FALSE;
917 np = vm_page_find_least(object, pi + n);
920 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
924 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
929 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
930 int flags, boolean_t *clearobjflags, boolean_t *eio)
932 vm_page_t ma[vm_pageout_page_count], p_first, tp;
933 int count, i, mreq, runlen;
935 vm_page_lock_assert(p, MA_NOTOWNED);
936 VM_OBJECT_ASSERT_WLOCKED(object);
941 for (tp = p; count < vm_pageout_page_count; count++) {
942 tp = vm_page_next(tp);
943 if (tp == NULL || vm_page_busied(tp))
945 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
949 for (p_first = p; count < vm_pageout_page_count; count++) {
950 tp = vm_page_prev(p_first);
951 if (tp == NULL || vm_page_busied(tp))
953 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
959 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
962 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
967 * Note that there is absolutely no sense in writing out
968 * anonymous objects, so we track down the vnode object
970 * We invalidate (remove) all pages from the address space
971 * for semantic correctness.
973 * If the backing object is a device object with unmanaged pages, then any
974 * mappings to the specified range of pages must be removed before this
975 * function is called.
977 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
978 * may start out with a NULL object.
981 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
982 boolean_t syncio, boolean_t invalidate)
984 vm_object_t backing_object;
987 int error, flags, fsync_after;
994 VM_OBJECT_WLOCK(object);
995 while ((backing_object = object->backing_object) != NULL) {
996 VM_OBJECT_WLOCK(backing_object);
997 offset += object->backing_object_offset;
998 VM_OBJECT_WUNLOCK(object);
999 object = backing_object;
1000 if (object->size < OFF_TO_IDX(offset + size))
1001 size = IDX_TO_OFF(object->size) - offset;
1004 * Flush pages if writing is allowed, invalidate them
1005 * if invalidation requested. Pages undergoing I/O
1006 * will be ignored by vm_object_page_remove().
1008 * We cannot lock the vnode and then wait for paging
1009 * to complete without deadlocking against vm_fault.
1010 * Instead we simply call vm_object_page_remove() and
1011 * allow it to block internally on a page-by-page
1012 * basis when it encounters pages undergoing async
1015 if (object->type == OBJT_VNODE &&
1016 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1017 vp = object->handle;
1018 VM_OBJECT_WUNLOCK(object);
1019 (void) vn_start_write(vp, &mp, V_WAIT);
1020 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1021 if (syncio && !invalidate && offset == 0 &&
1022 OFF_TO_IDX(size) == object->size) {
1024 * If syncing the whole mapping of the file,
1025 * it is faster to schedule all the writes in
1026 * async mode, also allowing the clustering,
1027 * and then wait for i/o to complete.
1032 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1033 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1034 fsync_after = FALSE;
1036 VM_OBJECT_WLOCK(object);
1037 res = vm_object_page_clean(object, offset, offset + size,
1039 VM_OBJECT_WUNLOCK(object);
1041 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1043 vn_finished_write(mp);
1046 VM_OBJECT_WLOCK(object);
1048 if ((object->type == OBJT_VNODE ||
1049 object->type == OBJT_DEVICE) && invalidate) {
1050 if (object->type == OBJT_DEVICE)
1052 * The option OBJPR_NOTMAPPED must be passed here
1053 * because vm_object_page_remove() cannot remove
1054 * unmanaged mappings.
1056 flags = OBJPR_NOTMAPPED;
1058 flags = OBJPR_NOTWIRED;
1060 flags = OBJPR_CLEANONLY | OBJPR_NOTWIRED;
1061 vm_object_page_remove(object, OFF_TO_IDX(offset),
1062 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1064 VM_OBJECT_WUNLOCK(object);
1069 * vm_object_madvise:
1071 * Implements the madvise function at the object/page level.
1073 * MADV_WILLNEED (any object)
1075 * Activate the specified pages if they are resident.
1077 * MADV_DONTNEED (any object)
1079 * Deactivate the specified pages if they are resident.
1081 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1082 * OBJ_ONEMAPPING only)
1084 * Deactivate and clean the specified pages if they are
1085 * resident. This permits the process to reuse the pages
1086 * without faulting or the kernel to reclaim the pages
1090 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1093 vm_pindex_t tpindex;
1094 vm_object_t backing_object, tobject;
1099 VM_OBJECT_WLOCK(object);
1101 * Locate and adjust resident pages
1103 for (; pindex < end; pindex += 1) {
1109 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1110 * and those pages must be OBJ_ONEMAPPING.
1112 if (advise == MADV_FREE) {
1113 if ((tobject->type != OBJT_DEFAULT &&
1114 tobject->type != OBJT_SWAP) ||
1115 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1116 goto unlock_tobject;
1118 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1119 goto unlock_tobject;
1120 m = vm_page_lookup(tobject, tpindex);
1121 if (m == NULL && advise == MADV_WILLNEED) {
1123 * If the page is cached, reactivate it.
1125 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1130 * There may be swap even if there is no backing page
1132 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1133 swap_pager_freespace(tobject, tpindex, 1);
1137 backing_object = tobject->backing_object;
1138 if (backing_object == NULL)
1139 goto unlock_tobject;
1140 VM_OBJECT_WLOCK(backing_object);
1141 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1142 if (tobject != object)
1143 VM_OBJECT_WUNLOCK(tobject);
1144 tobject = backing_object;
1146 } else if (m->valid != VM_PAGE_BITS_ALL)
1147 goto unlock_tobject;
1149 * If the page is not in a normal state, skip it.
1152 if (m->hold_count != 0 || m->wire_count != 0) {
1154 goto unlock_tobject;
1156 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1157 ("vm_object_madvise: page %p is fictitious", m));
1158 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1159 ("vm_object_madvise: page %p is not managed", m));
1160 if (vm_page_busied(m)) {
1161 if (advise == MADV_WILLNEED) {
1163 * Reference the page before unlocking and
1164 * sleeping so that the page daemon is less
1165 * likely to reclaim it.
1167 vm_page_aflag_set(m, PGA_REFERENCED);
1169 if (object != tobject)
1170 VM_OBJECT_WUNLOCK(object);
1171 VM_OBJECT_WUNLOCK(tobject);
1172 vm_page_busy_sleep(m, "madvpo");
1173 VM_OBJECT_WLOCK(object);
1176 if (advise == MADV_WILLNEED) {
1177 vm_page_activate(m);
1179 vm_page_advise(m, advise);
1182 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1183 swap_pager_freespace(tobject, tpindex, 1);
1185 if (tobject != object)
1186 VM_OBJECT_WUNLOCK(tobject);
1188 VM_OBJECT_WUNLOCK(object);
1194 * Create a new object which is backed by the
1195 * specified existing object range. The source
1196 * object reference is deallocated.
1198 * The new object and offset into that object
1199 * are returned in the source parameters.
1203 vm_object_t *object, /* IN/OUT */
1204 vm_ooffset_t *offset, /* IN/OUT */
1213 * Don't create the new object if the old object isn't shared.
1215 if (source != NULL) {
1216 VM_OBJECT_WLOCK(source);
1217 if (source->ref_count == 1 &&
1218 source->handle == NULL &&
1219 (source->type == OBJT_DEFAULT ||
1220 source->type == OBJT_SWAP)) {
1221 VM_OBJECT_WUNLOCK(source);
1224 VM_OBJECT_WUNLOCK(source);
1228 * Allocate a new object with the given length.
1230 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1233 * The new object shadows the source object, adding a reference to it.
1234 * Our caller changes his reference to point to the new object,
1235 * removing a reference to the source object. Net result: no change
1236 * of reference count.
1238 * Try to optimize the result object's page color when shadowing
1239 * in order to maintain page coloring consistency in the combined
1242 result->backing_object = source;
1244 * Store the offset into the source object, and fix up the offset into
1247 result->backing_object_offset = *offset;
1248 if (source != NULL) {
1249 VM_OBJECT_WLOCK(source);
1250 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1251 source->shadow_count++;
1252 #if VM_NRESERVLEVEL > 0
1253 result->flags |= source->flags & OBJ_COLORED;
1254 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1255 ((1 << (VM_NFREEORDER - 1)) - 1);
1257 VM_OBJECT_WUNLOCK(source);
1262 * Return the new things
1271 * Split the pages in a map entry into a new object. This affords
1272 * easier removal of unused pages, and keeps object inheritance from
1273 * being a negative impact on memory usage.
1276 vm_object_split(vm_map_entry_t entry)
1278 vm_page_t m, m_next;
1279 vm_object_t orig_object, new_object, source;
1280 vm_pindex_t idx, offidxstart;
1283 orig_object = entry->object.vm_object;
1284 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1286 if (orig_object->ref_count <= 1)
1288 VM_OBJECT_WUNLOCK(orig_object);
1290 offidxstart = OFF_TO_IDX(entry->offset);
1291 size = atop(entry->end - entry->start);
1294 * If swap_pager_copy() is later called, it will convert new_object
1295 * into a swap object.
1297 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1300 * At this point, the new object is still private, so the order in
1301 * which the original and new objects are locked does not matter.
1303 VM_OBJECT_WLOCK(new_object);
1304 VM_OBJECT_WLOCK(orig_object);
1305 source = orig_object->backing_object;
1306 if (source != NULL) {
1307 VM_OBJECT_WLOCK(source);
1308 if ((source->flags & OBJ_DEAD) != 0) {
1309 VM_OBJECT_WUNLOCK(source);
1310 VM_OBJECT_WUNLOCK(orig_object);
1311 VM_OBJECT_WUNLOCK(new_object);
1312 vm_object_deallocate(new_object);
1313 VM_OBJECT_WLOCK(orig_object);
1316 LIST_INSERT_HEAD(&source->shadow_head,
1317 new_object, shadow_list);
1318 source->shadow_count++;
1319 vm_object_reference_locked(source); /* for new_object */
1320 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1321 VM_OBJECT_WUNLOCK(source);
1322 new_object->backing_object_offset =
1323 orig_object->backing_object_offset + entry->offset;
1324 new_object->backing_object = source;
1326 if (orig_object->cred != NULL) {
1327 new_object->cred = orig_object->cred;
1328 crhold(orig_object->cred);
1329 new_object->charge = ptoa(size);
1330 KASSERT(orig_object->charge >= ptoa(size),
1331 ("orig_object->charge < 0"));
1332 orig_object->charge -= ptoa(size);
1335 m = vm_page_find_least(orig_object, offidxstart);
1336 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1338 m_next = TAILQ_NEXT(m, listq);
1341 * We must wait for pending I/O to complete before we can
1344 * We do not have to VM_PROT_NONE the page as mappings should
1345 * not be changed by this operation.
1347 if (vm_page_busied(m)) {
1348 VM_OBJECT_WUNLOCK(new_object);
1350 VM_OBJECT_WUNLOCK(orig_object);
1351 vm_page_busy_sleep(m, "spltwt");
1352 VM_OBJECT_WLOCK(orig_object);
1353 VM_OBJECT_WLOCK(new_object);
1357 /* vm_page_rename() will handle dirty and cache. */
1358 if (vm_page_rename(m, new_object, idx)) {
1359 VM_OBJECT_WUNLOCK(new_object);
1360 VM_OBJECT_WUNLOCK(orig_object);
1362 VM_OBJECT_WLOCK(orig_object);
1363 VM_OBJECT_WLOCK(new_object);
1366 #if VM_NRESERVLEVEL > 0
1368 * If some of the reservation's allocated pages remain with
1369 * the original object, then transferring the reservation to
1370 * the new object is neither particularly beneficial nor
1371 * particularly harmful as compared to leaving the reservation
1372 * with the original object. If, however, all of the
1373 * reservation's allocated pages are transferred to the new
1374 * object, then transferring the reservation is typically
1375 * beneficial. Determining which of these two cases applies
1376 * would be more costly than unconditionally renaming the
1379 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1381 if (orig_object->type == OBJT_SWAP)
1384 if (orig_object->type == OBJT_SWAP) {
1386 * swap_pager_copy() can sleep, in which case the orig_object's
1387 * and new_object's locks are released and reacquired.
1389 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1390 TAILQ_FOREACH(m, &new_object->memq, listq)
1394 * Transfer any cached pages from orig_object to new_object.
1395 * If swap_pager_copy() found swapped out pages within the
1396 * specified range of orig_object, then it changed
1397 * new_object's type to OBJT_SWAP when it transferred those
1398 * pages to new_object. Otherwise, new_object's type
1399 * should still be OBJT_DEFAULT and orig_object should not
1400 * contain any cached pages within the specified range.
1402 if (__predict_false(!vm_object_cache_is_empty(orig_object)))
1403 vm_page_cache_transfer(orig_object, offidxstart,
1406 VM_OBJECT_WUNLOCK(orig_object);
1407 VM_OBJECT_WUNLOCK(new_object);
1408 entry->object.vm_object = new_object;
1409 entry->offset = 0LL;
1410 vm_object_deallocate(orig_object);
1411 VM_OBJECT_WLOCK(new_object);
1414 #define OBSC_TEST_ALL_SHADOWED 0x0001
1415 #define OBSC_COLLAPSE_NOWAIT 0x0002
1416 #define OBSC_COLLAPSE_WAIT 0x0004
1419 vm_object_backing_scan(vm_object_t object, int op)
1423 vm_object_t backing_object;
1424 vm_pindex_t backing_offset_index;
1426 VM_OBJECT_ASSERT_WLOCKED(object);
1427 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1429 backing_object = object->backing_object;
1430 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1433 * Initial conditions
1435 if (op & OBSC_TEST_ALL_SHADOWED) {
1437 * We do not want to have to test for the existence of cache
1438 * or swap pages in the backing object. XXX but with the
1439 * new swapper this would be pretty easy to do.
1441 * XXX what about anonymous MAP_SHARED memory that hasn't
1442 * been ZFOD faulted yet? If we do not test for this, the
1443 * shadow test may succeed! XXX
1445 if (backing_object->type != OBJT_DEFAULT) {
1449 if (op & OBSC_COLLAPSE_WAIT) {
1450 vm_object_set_flag(backing_object, OBJ_DEAD);
1456 p = TAILQ_FIRST(&backing_object->memq);
1458 vm_page_t next = TAILQ_NEXT(p, listq);
1459 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1461 if (op & OBSC_TEST_ALL_SHADOWED) {
1465 * Ignore pages outside the parent object's range
1466 * and outside the parent object's mapping of the
1469 * note that we do not busy the backing object's
1473 p->pindex < backing_offset_index ||
1474 new_pindex >= object->size
1481 * See if the parent has the page or if the parent's
1482 * object pager has the page. If the parent has the
1483 * page but the page is not valid, the parent's
1484 * object pager must have the page.
1486 * If this fails, the parent does not completely shadow
1487 * the object and we might as well give up now.
1490 pp = vm_page_lookup(object, new_pindex);
1492 (pp == NULL || pp->valid == 0) &&
1493 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1501 * Check for busy page
1503 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1506 if (op & OBSC_COLLAPSE_NOWAIT) {
1507 if (!p->valid || vm_page_busied(p)) {
1511 } else if (op & OBSC_COLLAPSE_WAIT) {
1512 if (vm_page_busied(p)) {
1513 VM_OBJECT_WUNLOCK(object);
1515 VM_OBJECT_WUNLOCK(backing_object);
1516 vm_page_busy_sleep(p, "vmocol");
1517 VM_OBJECT_WLOCK(object);
1518 VM_OBJECT_WLOCK(backing_object);
1520 * If we slept, anything could have
1521 * happened. Since the object is
1522 * marked dead, the backing offset
1523 * should not have changed so we
1524 * just restart our scan.
1526 p = TAILQ_FIRST(&backing_object->memq);
1532 p->object == backing_object,
1533 ("vm_object_backing_scan: object mismatch")
1537 p->pindex < backing_offset_index ||
1538 new_pindex >= object->size
1540 if (backing_object->type == OBJT_SWAP)
1541 swap_pager_freespace(backing_object,
1545 * Page is out of the parent object's range, we
1546 * can simply destroy it.
1549 KASSERT(!pmap_page_is_mapped(p),
1550 ("freeing mapped page %p", p));
1551 if (p->wire_count == 0)
1560 pp = vm_page_lookup(object, new_pindex);
1562 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1563 (pp != NULL && pp->valid == 0)
1565 if (backing_object->type == OBJT_SWAP)
1566 swap_pager_freespace(backing_object,
1570 * The page in the parent is not (yet) valid.
1571 * We don't know anything about the state of
1572 * the original page. It might be mapped,
1573 * so we must avoid the next if here.
1575 * This is due to a race in vm_fault() where
1576 * we must unbusy the original (backing_obj)
1577 * page before we can (re)lock the parent.
1578 * Hence we can get here.
1585 vm_pager_has_page(object, new_pindex, NULL, NULL)
1587 if (backing_object->type == OBJT_SWAP)
1588 swap_pager_freespace(backing_object,
1592 * page already exists in parent OR swap exists
1593 * for this location in the parent. Destroy
1594 * the original page from the backing object.
1596 * Leave the parent's page alone
1599 KASSERT(!pmap_page_is_mapped(p),
1600 ("freeing mapped page %p", p));
1601 if (p->wire_count == 0)
1611 * Page does not exist in parent, rename the
1612 * page from the backing object to the main object.
1614 * If the page was mapped to a process, it can remain
1615 * mapped through the rename.
1616 * vm_page_rename() will handle dirty and cache.
1618 if (vm_page_rename(p, object, new_pindex)) {
1619 if (op & OBSC_COLLAPSE_NOWAIT) {
1623 VM_OBJECT_WLOCK(backing_object);
1624 VM_OBJECT_WUNLOCK(object);
1626 VM_OBJECT_WLOCK(object);
1627 VM_OBJECT_WLOCK(backing_object);
1628 p = TAILQ_FIRST(&backing_object->memq);
1631 if (backing_object->type == OBJT_SWAP)
1632 swap_pager_freespace(backing_object, p->pindex,
1635 #if VM_NRESERVLEVEL > 0
1637 * Rename the reservation.
1639 vm_reserv_rename(p, object, backing_object,
1640 backing_offset_index);
1650 * this version of collapse allows the operation to occur earlier and
1651 * when paging_in_progress is true for an object... This is not a complete
1652 * operation, but should plug 99.9% of the rest of the leaks.
1655 vm_object_qcollapse(vm_object_t object)
1657 vm_object_t backing_object = object->backing_object;
1659 VM_OBJECT_ASSERT_WLOCKED(object);
1660 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1662 if (backing_object->ref_count != 1)
1665 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1669 * vm_object_collapse:
1671 * Collapse an object with the object backing it.
1672 * Pages in the backing object are moved into the
1673 * parent, and the backing object is deallocated.
1676 vm_object_collapse(vm_object_t object)
1678 VM_OBJECT_ASSERT_WLOCKED(object);
1681 vm_object_t backing_object;
1684 * Verify that the conditions are right for collapse:
1686 * The object exists and the backing object exists.
1688 if ((backing_object = object->backing_object) == NULL)
1692 * we check the backing object first, because it is most likely
1695 VM_OBJECT_WLOCK(backing_object);
1696 if (backing_object->handle != NULL ||
1697 (backing_object->type != OBJT_DEFAULT &&
1698 backing_object->type != OBJT_SWAP) ||
1699 (backing_object->flags & OBJ_DEAD) ||
1700 object->handle != NULL ||
1701 (object->type != OBJT_DEFAULT &&
1702 object->type != OBJT_SWAP) ||
1703 (object->flags & OBJ_DEAD)) {
1704 VM_OBJECT_WUNLOCK(backing_object);
1709 object->paging_in_progress != 0 ||
1710 backing_object->paging_in_progress != 0
1712 vm_object_qcollapse(object);
1713 VM_OBJECT_WUNLOCK(backing_object);
1717 * We know that we can either collapse the backing object (if
1718 * the parent is the only reference to it) or (perhaps) have
1719 * the parent bypass the object if the parent happens to shadow
1720 * all the resident pages in the entire backing object.
1722 * This is ignoring pager-backed pages such as swap pages.
1723 * vm_object_backing_scan fails the shadowing test in this
1726 if (backing_object->ref_count == 1) {
1728 * If there is exactly one reference to the backing
1729 * object, we can collapse it into the parent.
1731 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1733 #if VM_NRESERVLEVEL > 0
1735 * Break any reservations from backing_object.
1737 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1738 vm_reserv_break_all(backing_object);
1742 * Move the pager from backing_object to object.
1744 if (backing_object->type == OBJT_SWAP) {
1746 * swap_pager_copy() can sleep, in which case
1747 * the backing_object's and object's locks are
1748 * released and reacquired.
1749 * Since swap_pager_copy() is being asked to
1750 * destroy the source, it will change the
1751 * backing_object's type to OBJT_DEFAULT.
1756 OFF_TO_IDX(object->backing_object_offset), TRUE);
1759 * Free any cached pages from backing_object.
1761 if (__predict_false(
1762 !vm_object_cache_is_empty(backing_object)))
1763 vm_page_cache_free(backing_object, 0, 0);
1766 * Object now shadows whatever backing_object did.
1767 * Note that the reference to
1768 * backing_object->backing_object moves from within
1769 * backing_object to within object.
1771 LIST_REMOVE(object, shadow_list);
1772 backing_object->shadow_count--;
1773 if (backing_object->backing_object) {
1774 VM_OBJECT_WLOCK(backing_object->backing_object);
1775 LIST_REMOVE(backing_object, shadow_list);
1777 &backing_object->backing_object->shadow_head,
1778 object, shadow_list);
1780 * The shadow_count has not changed.
1782 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1784 object->backing_object = backing_object->backing_object;
1785 object->backing_object_offset +=
1786 backing_object->backing_object_offset;
1789 * Discard backing_object.
1791 * Since the backing object has no pages, no pager left,
1792 * and no object references within it, all that is
1793 * necessary is to dispose of it.
1795 KASSERT(backing_object->ref_count == 1, (
1796 "backing_object %p was somehow re-referenced during collapse!",
1798 VM_OBJECT_WUNLOCK(backing_object);
1799 vm_object_destroy(backing_object);
1803 vm_object_t new_backing_object;
1806 * If we do not entirely shadow the backing object,
1807 * there is nothing we can do so we give up.
1809 if (object->resident_page_count != object->size &&
1810 vm_object_backing_scan(object,
1811 OBSC_TEST_ALL_SHADOWED) == 0) {
1812 VM_OBJECT_WUNLOCK(backing_object);
1817 * Make the parent shadow the next object in the
1818 * chain. Deallocating backing_object will not remove
1819 * it, since its reference count is at least 2.
1821 LIST_REMOVE(object, shadow_list);
1822 backing_object->shadow_count--;
1824 new_backing_object = backing_object->backing_object;
1825 if ((object->backing_object = new_backing_object) != NULL) {
1826 VM_OBJECT_WLOCK(new_backing_object);
1828 &new_backing_object->shadow_head,
1832 new_backing_object->shadow_count++;
1833 vm_object_reference_locked(new_backing_object);
1834 VM_OBJECT_WUNLOCK(new_backing_object);
1835 object->backing_object_offset +=
1836 backing_object->backing_object_offset;
1840 * Drop the reference count on backing_object. Since
1841 * its ref_count was at least 2, it will not vanish.
1843 backing_object->ref_count--;
1844 VM_OBJECT_WUNLOCK(backing_object);
1849 * Try again with this object's new backing object.
1855 * vm_object_page_remove:
1857 * For the given object, either frees or invalidates each of the
1858 * specified pages. In general, a page is freed. However, if a page is
1859 * wired for any reason other than the existence of a managed, wired
1860 * mapping, then it may be invalidated but not removed from the object.
1861 * Pages are specified by the given range ["start", "end") and the option
1862 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1863 * extends from "start" to the end of the object. If the option
1864 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1865 * specified range are affected. If the option OBJPR_NOTMAPPED is
1866 * specified, then the pages within the specified range must have no
1867 * mappings. Otherwise, if this option is not specified, any mappings to
1868 * the specified pages are removed before the pages are freed or
1871 * In general, this operation should only be performed on objects that
1872 * contain managed pages. There are, however, two exceptions. First, it
1873 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1874 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1875 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1876 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1878 * The object must be locked.
1881 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1887 VM_OBJECT_ASSERT_WLOCKED(object);
1888 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1889 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1890 ("vm_object_page_remove: illegal options for object %p", object));
1891 if (object->resident_page_count == 0)
1893 vm_object_pip_add(object, 1);
1895 p = vm_page_find_least(object, start);
1898 * Here, the variable "p" is either (1) the page with the least pindex
1899 * greater than or equal to the parameter "start" or (2) NULL.
1901 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1902 next = TAILQ_NEXT(p, listq);
1905 * If the page is wired for any reason besides the existence
1906 * of managed, wired mappings, then it cannot be freed. For
1907 * example, fictitious pages, which represent device memory,
1908 * are inherently wired and cannot be freed. They can,
1909 * however, be invalidated if the option OBJPR_CLEANONLY is
1913 if ((wirings = p->wire_count) != 0 &&
1914 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1915 if ((options & (OBJPR_NOTWIRED | OBJPR_NOTMAPPED)) ==
1918 /* Account for removal of wired mappings. */
1920 p->wire_count -= wirings;
1922 if ((options & OBJPR_CLEANONLY) == 0) {
1928 if (vm_page_busied(p)) {
1929 VM_OBJECT_WUNLOCK(object);
1930 vm_page_busy_sleep(p, "vmopar");
1931 VM_OBJECT_WLOCK(object);
1934 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1935 ("vm_object_page_remove: page %p is fictitious", p));
1936 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1937 if ((options & OBJPR_NOTMAPPED) == 0)
1938 pmap_remove_write(p);
1942 if ((options & OBJPR_NOTMAPPED) == 0) {
1943 if ((options & OBJPR_NOTWIRED) != 0 && wirings != 0)
1946 /* Account for removal of wired mappings. */
1948 KASSERT(p->wire_count == wirings,
1949 ("inconsistent wire count %d %d %p",
1950 p->wire_count, wirings, p));
1952 atomic_subtract_int(&cnt.v_wire_count, 1);
1959 vm_object_pip_wakeup(object);
1961 if (__predict_false(!vm_object_cache_is_empty(object)))
1962 vm_page_cache_free(object, start, end);
1966 * vm_object_page_cache:
1968 * For the given object, attempt to move the specified clean
1969 * pages to the cache queue. If a page is wired for any reason,
1970 * then it will not be changed. Pages are specified by the given
1971 * range ["start", "end"). As a special case, if "end" is zero,
1972 * then the range extends from "start" to the end of the object.
1973 * Any mappings to the specified pages are removed before the
1974 * pages are moved to the cache queue.
1976 * This operation should only be performed on objects that
1977 * contain non-fictitious, managed pages.
1979 * The object must be locked.
1982 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1984 struct mtx *mtx, *new_mtx;
1987 VM_OBJECT_ASSERT_WLOCKED(object);
1988 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1989 ("vm_object_page_cache: illegal object %p", object));
1990 if (object->resident_page_count == 0)
1992 p = vm_page_find_least(object, start);
1995 * Here, the variable "p" is either (1) the page with the least pindex
1996 * greater than or equal to the parameter "start" or (2) NULL.
1999 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2000 next = TAILQ_NEXT(p, listq);
2003 * Avoid releasing and reacquiring the same page lock.
2005 new_mtx = vm_page_lockptr(p);
2006 if (mtx != new_mtx) {
2012 vm_page_try_to_cache(p);
2019 * Populate the specified range of the object with valid pages. Returns
2020 * TRUE if the range is successfully populated and FALSE otherwise.
2022 * Note: This function should be optimized to pass a larger array of
2023 * pages to vm_pager_get_pages() before it is applied to a non-
2024 * OBJT_DEVICE object.
2026 * The object must be locked.
2029 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2035 VM_OBJECT_ASSERT_WLOCKED(object);
2036 for (pindex = start; pindex < end; pindex++) {
2037 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
2038 if (m->valid != VM_PAGE_BITS_ALL) {
2040 rv = vm_pager_get_pages(object, ma, 1, 0);
2041 m = vm_page_lookup(object, pindex);
2044 if (rv != VM_PAGER_OK) {
2052 * Keep "m" busy because a subsequent iteration may unlock
2056 if (pindex > start) {
2057 m = vm_page_lookup(object, start);
2058 while (m != NULL && m->pindex < pindex) {
2060 m = TAILQ_NEXT(m, listq);
2063 return (pindex == end);
2067 * Routine: vm_object_coalesce
2068 * Function: Coalesces two objects backing up adjoining
2069 * regions of memory into a single object.
2071 * returns TRUE if objects were combined.
2073 * NOTE: Only works at the moment if the second object is NULL -
2074 * if it's not, which object do we lock first?
2077 * prev_object First object to coalesce
2078 * prev_offset Offset into prev_object
2079 * prev_size Size of reference to prev_object
2080 * next_size Size of reference to the second object
2081 * reserved Indicator that extension region has
2082 * swap accounted for
2085 * The object must *not* be locked.
2088 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2089 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2091 vm_pindex_t next_pindex;
2093 if (prev_object == NULL)
2095 VM_OBJECT_WLOCK(prev_object);
2096 if (prev_object->type != OBJT_DEFAULT &&
2097 prev_object->type != OBJT_SWAP) {
2098 VM_OBJECT_WUNLOCK(prev_object);
2103 * Try to collapse the object first
2105 vm_object_collapse(prev_object);
2108 * Can't coalesce if: . more than one reference . paged out . shadows
2109 * another object . has a copy elsewhere (any of which mean that the
2110 * pages not mapped to prev_entry may be in use anyway)
2112 if (prev_object->backing_object != NULL) {
2113 VM_OBJECT_WUNLOCK(prev_object);
2117 prev_size >>= PAGE_SHIFT;
2118 next_size >>= PAGE_SHIFT;
2119 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2121 if ((prev_object->ref_count > 1) &&
2122 (prev_object->size != next_pindex)) {
2123 VM_OBJECT_WUNLOCK(prev_object);
2128 * Account for the charge.
2130 if (prev_object->cred != NULL) {
2133 * If prev_object was charged, then this mapping,
2134 * althought not charged now, may become writable
2135 * later. Non-NULL cred in the object would prevent
2136 * swap reservation during enabling of the write
2137 * access, so reserve swap now. Failed reservation
2138 * cause allocation of the separate object for the map
2139 * entry, and swap reservation for this entry is
2140 * managed in appropriate time.
2142 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2143 prev_object->cred)) {
2146 prev_object->charge += ptoa(next_size);
2150 * Remove any pages that may still be in the object from a previous
2153 if (next_pindex < prev_object->size) {
2154 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2156 if (prev_object->type == OBJT_SWAP)
2157 swap_pager_freespace(prev_object,
2158 next_pindex, next_size);
2160 if (prev_object->cred != NULL) {
2161 KASSERT(prev_object->charge >=
2162 ptoa(prev_object->size - next_pindex),
2163 ("object %p overcharged 1 %jx %jx", prev_object,
2164 (uintmax_t)next_pindex, (uintmax_t)next_size));
2165 prev_object->charge -= ptoa(prev_object->size -
2172 * Extend the object if necessary.
2174 if (next_pindex + next_size > prev_object->size)
2175 prev_object->size = next_pindex + next_size;
2177 VM_OBJECT_WUNLOCK(prev_object);
2182 vm_object_set_writeable_dirty(vm_object_t object)
2185 VM_OBJECT_ASSERT_WLOCKED(object);
2186 if (object->type != OBJT_VNODE)
2188 object->generation++;
2189 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2191 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2194 #include "opt_ddb.h"
2196 #include <sys/kernel.h>
2198 #include <sys/cons.h>
2200 #include <ddb/ddb.h>
2203 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2206 vm_map_entry_t tmpe;
2214 tmpe = map->header.next;
2215 entcount = map->nentries;
2216 while (entcount-- && (tmpe != &map->header)) {
2217 if (_vm_object_in_map(map, object, tmpe)) {
2222 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2223 tmpm = entry->object.sub_map;
2224 tmpe = tmpm->header.next;
2225 entcount = tmpm->nentries;
2226 while (entcount-- && tmpe != &tmpm->header) {
2227 if (_vm_object_in_map(tmpm, object, tmpe)) {
2232 } else if ((obj = entry->object.vm_object) != NULL) {
2233 for (; obj; obj = obj->backing_object)
2234 if (obj == object) {
2242 vm_object_in_map(vm_object_t object)
2246 /* sx_slock(&allproc_lock); */
2247 FOREACH_PROC_IN_SYSTEM(p) {
2248 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2250 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2251 /* sx_sunlock(&allproc_lock); */
2255 /* sx_sunlock(&allproc_lock); */
2256 if (_vm_object_in_map(kernel_map, object, 0))
2261 DB_SHOW_COMMAND(vmochk, vm_object_check)
2266 * make sure that internal objs are in a map somewhere
2267 * and none have zero ref counts.
2269 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2270 if (object->handle == NULL &&
2271 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2272 if (object->ref_count == 0) {
2273 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2274 (long)object->size);
2276 if (!vm_object_in_map(object)) {
2278 "vmochk: internal obj is not in a map: "
2279 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2280 object->ref_count, (u_long)object->size,
2281 (u_long)object->size,
2282 (void *)object->backing_object);
2289 * vm_object_print: [ debug ]
2291 DB_SHOW_COMMAND(object, vm_object_print_static)
2293 /* XXX convert args. */
2294 vm_object_t object = (vm_object_t)addr;
2295 boolean_t full = have_addr;
2299 /* XXX count is an (unused) arg. Avoid shadowing it. */
2300 #define count was_count
2308 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2309 object, (int)object->type, (uintmax_t)object->size,
2310 object->resident_page_count, object->ref_count, object->flags,
2311 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2312 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2313 object->shadow_count,
2314 object->backing_object ? object->backing_object->ref_count : 0,
2315 object->backing_object, (uintmax_t)object->backing_object_offset);
2322 TAILQ_FOREACH(p, &object->memq, listq) {
2324 db_iprintf("memory:=");
2325 else if (count == 6) {
2333 db_printf("(off=0x%jx,page=0x%jx)",
2334 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2344 /* XXX need this non-static entry for calling from vm_map_print. */
2347 /* db_expr_t */ long addr,
2348 boolean_t have_addr,
2349 /* db_expr_t */ long count,
2352 vm_object_print_static(addr, have_addr, count, modif);
2355 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2360 vm_page_t m, prev_m;
2364 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2365 db_printf("new object: %p\n", (void *)object);
2376 TAILQ_FOREACH(m, &object->memq, listq) {
2377 if (m->pindex > 128)
2379 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2380 prev_m->pindex + 1 != m->pindex) {
2382 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2383 (long)fidx, rcount, (long)pa);
2395 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2400 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2401 (long)fidx, rcount, (long)pa);
2411 pa = VM_PAGE_TO_PHYS(m);
2415 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2416 (long)fidx, rcount, (long)pa);