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/vnode.h>
82 #include <sys/vmmeter.h>
86 #include <vm/vm_param.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_pager.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_reserv.h>
100 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
101 "Use old (insecure) msync behavior");
103 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
104 int pagerflags, int flags, int *clearobjflags);
105 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
107 static void vm_object_qcollapse(vm_object_t object);
108 static void vm_object_vndeallocate(vm_object_t object);
111 * Virtual memory objects maintain the actual data
112 * associated with allocated virtual memory. A given
113 * page of memory exists within exactly one object.
115 * An object is only deallocated when all "references"
116 * are given up. Only one "reference" to a given
117 * region of an object should be writeable.
119 * Associated with each object is a list of all resident
120 * memory pages belonging to that object; this list is
121 * maintained by the "vm_page" module, and locked by the object's
124 * Each object also records a "pager" routine which is
125 * used to retrieve (and store) pages to the proper backing
126 * storage. In addition, objects may be backed by other
127 * objects from which they were virtual-copied.
129 * The only items within the object structure which are
130 * modified after time of creation are:
131 * reference count locked by object's lock
132 * pager routine locked by object's lock
136 struct object_q vm_object_list;
137 struct mtx vm_object_list_mtx; /* lock for object list and count */
139 struct vm_object kernel_object_store;
140 struct vm_object kmem_object_store;
142 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats");
144 static long object_collapses;
145 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
146 &object_collapses, 0, "VM object collapses");
148 static long object_bypasses;
149 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
150 &object_bypasses, 0, "VM object bypasses");
152 static uma_zone_t obj_zone;
154 static int vm_object_zinit(void *mem, int size, int flags);
157 static void vm_object_zdtor(void *mem, int size, void *arg);
160 vm_object_zdtor(void *mem, int size, void *arg)
164 object = (vm_object_t)mem;
165 KASSERT(TAILQ_EMPTY(&object->memq),
166 ("object %p has resident pages",
168 #if VM_NRESERVLEVEL > 0
169 KASSERT(LIST_EMPTY(&object->rvq),
170 ("object %p has reservations",
173 KASSERT(object->cache == NULL,
174 ("object %p has cached pages",
176 KASSERT(object->paging_in_progress == 0,
177 ("object %p paging_in_progress = %d",
178 object, object->paging_in_progress));
179 KASSERT(object->resident_page_count == 0,
180 ("object %p resident_page_count = %d",
181 object, object->resident_page_count));
182 KASSERT(object->shadow_count == 0,
183 ("object %p shadow_count = %d",
184 object, object->shadow_count));
189 vm_object_zinit(void *mem, int size, int flags)
193 object = (vm_object_t)mem;
194 bzero(&object->mtx, sizeof(object->mtx));
195 VM_OBJECT_LOCK_INIT(object, "standard object");
197 /* These are true for any object that has been freed */
198 object->paging_in_progress = 0;
199 object->resident_page_count = 0;
200 object->shadow_count = 0;
205 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
208 TAILQ_INIT(&object->memq);
209 LIST_INIT(&object->shadow_head);
214 object->generation = 1;
215 object->ref_count = 1;
216 object->memattr = VM_MEMATTR_DEFAULT;
220 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
221 object->flags = OBJ_ONEMAPPING;
222 object->pg_color = 0;
223 object->handle = NULL;
224 object->backing_object = NULL;
225 object->backing_object_offset = (vm_ooffset_t) 0;
226 #if VM_NRESERVLEVEL > 0
227 LIST_INIT(&object->rvq);
229 object->cache = NULL;
231 mtx_lock(&vm_object_list_mtx);
232 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
233 mtx_unlock(&vm_object_list_mtx);
239 * Initialize the VM objects module.
244 TAILQ_INIT(&vm_object_list);
245 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
247 VM_OBJECT_LOCK_INIT(kernel_object, "kernel object");
248 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
250 #if VM_NRESERVLEVEL > 0
251 kernel_object->flags |= OBJ_COLORED;
252 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
255 VM_OBJECT_LOCK_INIT(kmem_object, "kmem object");
256 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
258 #if VM_NRESERVLEVEL > 0
259 kmem_object->flags |= OBJ_COLORED;
260 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
264 * The lock portion of struct vm_object must be type stable due
265 * to vm_pageout_fallback_object_lock locking a vm object
266 * without holding any references to it.
268 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
274 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
278 vm_object_clear_flag(vm_object_t object, u_short bits)
281 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
282 object->flags &= ~bits;
286 * Sets the default memory attribute for the specified object. Pages
287 * that are allocated to this object are by default assigned this memory
290 * Presently, this function must be called before any pages are allocated
291 * to the object. In the future, this requirement may be relaxed for
292 * "default" and "swap" objects.
295 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
298 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
299 switch (object->type) {
306 if (!TAILQ_EMPTY(&object->memq))
307 return (KERN_FAILURE);
310 return (KERN_INVALID_ARGUMENT);
312 object->memattr = memattr;
313 return (KERN_SUCCESS);
317 vm_object_pip_add(vm_object_t object, short i)
320 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
321 object->paging_in_progress += i;
325 vm_object_pip_subtract(vm_object_t object, short i)
328 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
329 object->paging_in_progress -= i;
333 vm_object_pip_wakeup(vm_object_t object)
336 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
337 object->paging_in_progress--;
338 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
339 vm_object_clear_flag(object, OBJ_PIPWNT);
345 vm_object_pip_wakeupn(vm_object_t object, short i)
348 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
350 object->paging_in_progress -= i;
351 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
352 vm_object_clear_flag(object, OBJ_PIPWNT);
358 vm_object_pip_wait(vm_object_t object, char *waitid)
361 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
362 while (object->paging_in_progress) {
363 object->flags |= OBJ_PIPWNT;
364 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
369 * vm_object_allocate:
371 * Returns a new object with the given size.
374 vm_object_allocate(objtype_t type, vm_pindex_t size)
378 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
379 _vm_object_allocate(type, size, object);
385 * vm_object_reference:
387 * Gets another reference to the given object. Note: OBJ_DEAD
388 * objects can be referenced during final cleaning.
391 vm_object_reference(vm_object_t object)
395 VM_OBJECT_LOCK(object);
396 vm_object_reference_locked(object);
397 VM_OBJECT_UNLOCK(object);
401 * vm_object_reference_locked:
403 * Gets another reference to the given object.
405 * The object must be locked.
408 vm_object_reference_locked(vm_object_t object)
412 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
414 if (object->type == OBJT_VNODE) {
421 * Handle deallocating an object of type OBJT_VNODE.
424 vm_object_vndeallocate(vm_object_t object)
426 struct vnode *vp = (struct vnode *) object->handle;
428 VFS_ASSERT_GIANT(vp->v_mount);
429 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
430 KASSERT(object->type == OBJT_VNODE,
431 ("vm_object_vndeallocate: not a vnode object"));
432 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
434 if (object->ref_count == 0) {
435 vprint("vm_object_vndeallocate", vp);
436 panic("vm_object_vndeallocate: bad object reference count");
440 if (object->ref_count > 1) {
442 VM_OBJECT_UNLOCK(object);
443 /* vrele may need the vnode lock. */
447 VM_OBJECT_UNLOCK(object);
448 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
450 VM_OBJECT_LOCK(object);
452 if (object->type == OBJT_DEAD) {
453 VM_OBJECT_UNLOCK(object);
456 if (object->ref_count == 0)
457 vp->v_vflag &= ~VV_TEXT;
458 VM_OBJECT_UNLOCK(object);
465 * vm_object_deallocate:
467 * Release a reference to the specified object,
468 * gained either through a vm_object_allocate
469 * or a vm_object_reference call. When all references
470 * are gone, storage associated with this object
471 * may be relinquished.
473 * No object may be locked.
476 vm_object_deallocate(vm_object_t object)
480 while (object != NULL) {
485 VM_OBJECT_LOCK(object);
486 if (object->type == OBJT_VNODE) {
487 struct vnode *vp = (struct vnode *) object->handle;
490 * Conditionally acquire Giant for a vnode-backed
491 * object. We have to be careful since the type of
492 * a vnode object can change while the object is
495 if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
497 if (!mtx_trylock(&Giant)) {
498 VM_OBJECT_UNLOCK(object);
503 vm_object_vndeallocate(object);
504 VFS_UNLOCK_GIANT(vfslocked);
508 * This is to handle the case that the object
509 * changed type while we dropped its lock to
512 VFS_UNLOCK_GIANT(vfslocked);
514 KASSERT(object->ref_count != 0,
515 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
518 * If the reference count goes to 0 we start calling
519 * vm_object_terminate() on the object chain.
520 * A ref count of 1 may be a special case depending on the
521 * shadow count being 0 or 1.
524 if (object->ref_count > 1) {
525 VM_OBJECT_UNLOCK(object);
527 } else if (object->ref_count == 1) {
528 if (object->shadow_count == 0 &&
529 object->handle == NULL &&
530 (object->type == OBJT_DEFAULT ||
531 object->type == OBJT_SWAP)) {
532 vm_object_set_flag(object, OBJ_ONEMAPPING);
533 } else if ((object->shadow_count == 1) &&
534 (object->handle == NULL) &&
535 (object->type == OBJT_DEFAULT ||
536 object->type == OBJT_SWAP)) {
539 robject = LIST_FIRST(&object->shadow_head);
540 KASSERT(robject != NULL,
541 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
543 object->shadow_count));
544 if (!VM_OBJECT_TRYLOCK(robject)) {
546 * Avoid a potential deadlock.
549 VM_OBJECT_UNLOCK(object);
551 * More likely than not the thread
552 * holding robject's lock has lower
553 * priority than the current thread.
554 * Let the lower priority thread run.
560 * Collapse object into its shadow unless its
561 * shadow is dead. In that case, object will
562 * be deallocated by the thread that is
563 * deallocating its shadow.
565 if ((robject->flags & OBJ_DEAD) == 0 &&
566 (robject->handle == NULL) &&
567 (robject->type == OBJT_DEFAULT ||
568 robject->type == OBJT_SWAP)) {
570 robject->ref_count++;
572 if (robject->paging_in_progress) {
573 VM_OBJECT_UNLOCK(object);
574 vm_object_pip_wait(robject,
576 temp = robject->backing_object;
577 if (object == temp) {
578 VM_OBJECT_LOCK(object);
581 } else if (object->paging_in_progress) {
582 VM_OBJECT_UNLOCK(robject);
583 object->flags |= OBJ_PIPWNT;
585 VM_OBJECT_MTX(object),
586 PDROP | PVM, "objde2", 0);
587 VM_OBJECT_LOCK(robject);
588 temp = robject->backing_object;
589 if (object == temp) {
590 VM_OBJECT_LOCK(object);
594 VM_OBJECT_UNLOCK(object);
596 if (robject->ref_count == 1) {
597 robject->ref_count--;
602 vm_object_collapse(object);
603 VM_OBJECT_UNLOCK(object);
606 VM_OBJECT_UNLOCK(robject);
608 VM_OBJECT_UNLOCK(object);
612 temp = object->backing_object;
614 VM_OBJECT_LOCK(temp);
615 LIST_REMOVE(object, shadow_list);
616 temp->shadow_count--;
617 VM_OBJECT_UNLOCK(temp);
618 object->backing_object = NULL;
621 * Don't double-terminate, we could be in a termination
622 * recursion due to the terminate having to sync data
625 if ((object->flags & OBJ_DEAD) == 0)
626 vm_object_terminate(object);
628 VM_OBJECT_UNLOCK(object);
634 * vm_object_destroy removes the object from the global object list
635 * and frees the space for the object.
638 vm_object_destroy(vm_object_t object)
642 * Remove the object from the global object list.
644 mtx_lock(&vm_object_list_mtx);
645 TAILQ_REMOVE(&vm_object_list, object, object_list);
646 mtx_unlock(&vm_object_list_mtx);
649 * Release the allocation charge.
651 if (object->cred != NULL) {
652 KASSERT(object->type == OBJT_DEFAULT ||
653 object->type == OBJT_SWAP,
654 ("vm_object_terminate: non-swap obj %p has cred",
656 swap_release_by_cred(object->charge, object->cred);
658 crfree(object->cred);
663 * Free the space for the object.
665 uma_zfree(obj_zone, object);
669 * vm_object_terminate actually destroys the specified object, freeing
670 * up all previously used resources.
672 * The object must be locked.
673 * This routine may block.
676 vm_object_terminate(vm_object_t object)
680 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
683 * Make sure no one uses us.
685 vm_object_set_flag(object, OBJ_DEAD);
688 * wait for the pageout daemon to be done with the object
690 vm_object_pip_wait(object, "objtrm");
692 KASSERT(!object->paging_in_progress,
693 ("vm_object_terminate: pageout in progress"));
696 * Clean and free the pages, as appropriate. All references to the
697 * object are gone, so we don't need to lock it.
699 if (object->type == OBJT_VNODE) {
700 struct vnode *vp = (struct vnode *)object->handle;
703 * Clean pages and flush buffers.
705 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
706 VM_OBJECT_UNLOCK(object);
708 vinvalbuf(vp, V_SAVE, 0, 0);
710 VM_OBJECT_LOCK(object);
713 KASSERT(object->ref_count == 0,
714 ("vm_object_terminate: object with references, ref_count=%d",
718 * Free any remaining pageable pages. This also removes them from the
719 * paging queues. However, don't free wired pages, just remove them
720 * from the object. Rather than incrementally removing each page from
721 * the object, the page and object are reset to any empty state.
723 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
724 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
725 ("vm_object_terminate: freeing busy page %p", p));
728 * Optimize the page's removal from the object by resetting
729 * its "object" field. Specifically, if the page is not
730 * wired, then the effect of this assignment is that
731 * vm_page_free()'s call to vm_page_remove() will return
732 * immediately without modifying the page or the object.
735 if (p->wire_count == 0) {
737 PCPU_INC(cnt.v_pfree);
742 * If the object contained any pages, then reset it to an empty state.
743 * None of the object's fields, including "resident_page_count", were
744 * modified by the preceding loop.
746 if (object->resident_page_count != 0) {
748 TAILQ_INIT(&object->memq);
749 object->resident_page_count = 0;
750 if (object->type == OBJT_VNODE)
751 vdrop(object->handle);
754 #if VM_NRESERVLEVEL > 0
755 if (__predict_false(!LIST_EMPTY(&object->rvq)))
756 vm_reserv_break_all(object);
758 if (__predict_false(object->cache != NULL))
759 vm_page_cache_free(object, 0, 0);
762 * Let the pager know object is dead.
764 vm_pager_deallocate(object);
765 VM_OBJECT_UNLOCK(object);
767 vm_object_destroy(object);
771 * Make the page read-only so that we can clear the object flags. However, if
772 * this is a nosync mmap then the object is likely to stay dirty so do not
773 * mess with the page and do not clear the object flags. Returns TRUE if the
774 * page should be flushed, and FALSE otherwise.
777 vm_object_page_remove_write(vm_page_t p, int flags, int *clearobjflags)
781 * If we have been asked to skip nosync pages and this is a
782 * nosync page, skip it. Note that the object flags were not
783 * cleared in this case so we do not have to set them.
785 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
789 pmap_remove_write(p);
790 return (p->dirty != 0);
795 * vm_object_page_clean
797 * Clean all dirty pages in the specified range of object. Leaves page
798 * on whatever queue it is currently on. If NOSYNC is set then do not
799 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
800 * leaving the object dirty.
802 * When stuffing pages asynchronously, allow clustering. XXX we need a
803 * synchronous clustering mode implementation.
805 * Odd semantics: if start == end, we clean everything.
807 * The object must be locked.
810 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
814 vm_pindex_t pi, tend, tstart;
815 int clearobjflags, curgeneration, n, pagerflags;
817 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
818 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
819 KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
820 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
821 object->resident_page_count == 0)
824 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
825 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
826 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
828 tstart = OFF_TO_IDX(start);
829 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
830 clearobjflags = tstart == 0 && tend >= object->size;
833 curgeneration = object->generation;
835 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
839 np = TAILQ_NEXT(p, listq);
842 if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
843 if (object->generation != curgeneration)
845 np = vm_page_find_least(object, pi);
848 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
851 n = vm_object_page_collect_flush(object, p, pagerflags,
852 flags, &clearobjflags);
853 if (object->generation != curgeneration)
857 * If the VOP_PUTPAGES() did a truncated write, so
858 * that even the first page of the run is not fully
859 * written, vm_pageout_flush() returns 0 as the run
860 * length. Since the condition that caused truncated
861 * write may be permanent, e.g. exhausted free space,
862 * accepting n == 0 would cause an infinite loop.
864 * Forwarding the iterator leaves the unwritten page
865 * behind, but there is not much we can do there if
866 * filesystem refuses to write it.
870 np = vm_page_find_least(object, pi + n);
873 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
877 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
881 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
882 int flags, int *clearobjflags)
884 vm_page_t ma[vm_pageout_page_count], p_first, tp;
885 int count, i, mreq, runlen;
887 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
888 vm_page_lock_assert(p, MA_NOTOWNED);
889 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
894 for (tp = p; count < vm_pageout_page_count; count++) {
895 tp = vm_page_next(tp);
896 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
898 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
902 for (p_first = p; count < vm_pageout_page_count; count++) {
903 tp = vm_page_prev(p_first);
904 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
906 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
912 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
915 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen);
920 * Note that there is absolutely no sense in writing out
921 * anonymous objects, so we track down the vnode object
923 * We invalidate (remove) all pages from the address space
924 * for semantic correctness.
926 * If the backing object is a device object with unmanaged pages, then any
927 * mappings to the specified range of pages must be removed before this
928 * function is called.
930 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
931 * may start out with a NULL object.
934 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
935 boolean_t syncio, boolean_t invalidate)
937 vm_object_t backing_object;
944 VM_OBJECT_LOCK(object);
945 while ((backing_object = object->backing_object) != NULL) {
946 VM_OBJECT_LOCK(backing_object);
947 offset += object->backing_object_offset;
948 VM_OBJECT_UNLOCK(object);
949 object = backing_object;
950 if (object->size < OFF_TO_IDX(offset + size))
951 size = IDX_TO_OFF(object->size) - offset;
954 * Flush pages if writing is allowed, invalidate them
955 * if invalidation requested. Pages undergoing I/O
956 * will be ignored by vm_object_page_remove().
958 * We cannot lock the vnode and then wait for paging
959 * to complete without deadlocking against vm_fault.
960 * Instead we simply call vm_object_page_remove() and
961 * allow it to block internally on a page-by-page
962 * basis when it encounters pages undergoing async
965 if (object->type == OBJT_VNODE &&
966 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
969 VM_OBJECT_UNLOCK(object);
970 (void) vn_start_write(vp, &mp, V_WAIT);
971 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
972 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
973 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
974 flags |= invalidate ? OBJPC_INVAL : 0;
975 VM_OBJECT_LOCK(object);
976 vm_object_page_clean(object, offset, offset + size, flags);
977 VM_OBJECT_UNLOCK(object);
979 VFS_UNLOCK_GIANT(vfslocked);
980 vn_finished_write(mp);
981 VM_OBJECT_LOCK(object);
983 if ((object->type == OBJT_VNODE ||
984 object->type == OBJT_DEVICE) && invalidate) {
985 if (object->type == OBJT_DEVICE)
987 * The option OBJPR_NOTMAPPED must be passed here
988 * because vm_object_page_remove() cannot remove
989 * unmanaged mappings.
991 flags = OBJPR_NOTMAPPED;
995 flags = OBJPR_CLEANONLY;
996 vm_object_page_remove(object, OFF_TO_IDX(offset),
997 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
999 VM_OBJECT_UNLOCK(object);
1003 * vm_object_madvise:
1005 * Implements the madvise function at the object/page level.
1007 * MADV_WILLNEED (any object)
1009 * Activate the specified pages if they are resident.
1011 * MADV_DONTNEED (any object)
1013 * Deactivate the specified pages if they are resident.
1015 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1016 * OBJ_ONEMAPPING only)
1018 * Deactivate and clean the specified pages if they are
1019 * resident. This permits the process to reuse the pages
1020 * without faulting or the kernel to reclaim the pages
1024 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1026 vm_pindex_t end, tpindex;
1027 vm_object_t backing_object, tobject;
1032 VM_OBJECT_LOCK(object);
1033 end = pindex + count;
1035 * Locate and adjust resident pages
1037 for (; pindex < end; pindex += 1) {
1043 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1044 * and those pages must be OBJ_ONEMAPPING.
1046 if (advise == MADV_FREE) {
1047 if ((tobject->type != OBJT_DEFAULT &&
1048 tobject->type != OBJT_SWAP) ||
1049 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1050 goto unlock_tobject;
1052 } else if (tobject->type == OBJT_PHYS)
1053 goto unlock_tobject;
1054 m = vm_page_lookup(tobject, tpindex);
1055 if (m == NULL && advise == MADV_WILLNEED) {
1057 * If the page is cached, reactivate it.
1059 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1064 * There may be swap even if there is no backing page
1066 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1067 swap_pager_freespace(tobject, tpindex, 1);
1071 backing_object = tobject->backing_object;
1072 if (backing_object == NULL)
1073 goto unlock_tobject;
1074 VM_OBJECT_LOCK(backing_object);
1075 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1076 if (tobject != object)
1077 VM_OBJECT_UNLOCK(tobject);
1078 tobject = backing_object;
1080 } else if (m->valid != VM_PAGE_BITS_ALL)
1081 goto unlock_tobject;
1083 * If the page is not in a normal state, skip it.
1086 if (m->hold_count != 0 || m->wire_count != 0) {
1088 goto unlock_tobject;
1090 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1091 ("vm_object_madvise: page %p is fictitious", m));
1092 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1093 ("vm_object_madvise: page %p is not managed", m));
1094 if ((m->oflags & VPO_BUSY) || m->busy) {
1095 if (advise == MADV_WILLNEED) {
1097 * Reference the page before unlocking and
1098 * sleeping so that the page daemon is less
1099 * likely to reclaim it.
1101 vm_page_aflag_set(m, PGA_REFERENCED);
1104 if (object != tobject)
1105 VM_OBJECT_UNLOCK(object);
1106 m->oflags |= VPO_WANTED;
1107 msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1109 VM_OBJECT_LOCK(object);
1112 if (advise == MADV_WILLNEED) {
1113 vm_page_activate(m);
1114 } else if (advise == MADV_DONTNEED) {
1115 vm_page_dontneed(m);
1116 } else if (advise == MADV_FREE) {
1118 * Mark the page clean. This will allow the page
1119 * to be freed up by the system. However, such pages
1120 * are often reused quickly by malloc()/free()
1121 * so we do not do anything that would cause
1122 * a page fault if we can help it.
1124 * Specifically, we do not try to actually free
1125 * the page now nor do we try to put it in the
1126 * cache (which would cause a page fault on reuse).
1128 * But we do make the page is freeable as we
1129 * can without actually taking the step of unmapping
1132 pmap_clear_modify(m);
1135 vm_page_dontneed(m);
1138 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1139 swap_pager_freespace(tobject, tpindex, 1);
1141 if (tobject != object)
1142 VM_OBJECT_UNLOCK(tobject);
1144 VM_OBJECT_UNLOCK(object);
1150 * Create a new object which is backed by the
1151 * specified existing object range. The source
1152 * object reference is deallocated.
1154 * The new object and offset into that object
1155 * are returned in the source parameters.
1159 vm_object_t *object, /* IN/OUT */
1160 vm_ooffset_t *offset, /* IN/OUT */
1169 * Don't create the new object if the old object isn't shared.
1171 if (source != NULL) {
1172 VM_OBJECT_LOCK(source);
1173 if (source->ref_count == 1 &&
1174 source->handle == NULL &&
1175 (source->type == OBJT_DEFAULT ||
1176 source->type == OBJT_SWAP)) {
1177 VM_OBJECT_UNLOCK(source);
1180 VM_OBJECT_UNLOCK(source);
1184 * Allocate a new object with the given length.
1186 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1189 * The new object shadows the source object, adding a reference to it.
1190 * Our caller changes his reference to point to the new object,
1191 * removing a reference to the source object. Net result: no change
1192 * of reference count.
1194 * Try to optimize the result object's page color when shadowing
1195 * in order to maintain page coloring consistency in the combined
1198 result->backing_object = source;
1200 * Store the offset into the source object, and fix up the offset into
1203 result->backing_object_offset = *offset;
1204 if (source != NULL) {
1205 VM_OBJECT_LOCK(source);
1206 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1207 source->shadow_count++;
1208 #if VM_NRESERVLEVEL > 0
1209 result->flags |= source->flags & OBJ_COLORED;
1210 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1211 ((1 << (VM_NFREEORDER - 1)) - 1);
1213 VM_OBJECT_UNLOCK(source);
1218 * Return the new things
1227 * Split the pages in a map entry into a new object. This affords
1228 * easier removal of unused pages, and keeps object inheritance from
1229 * being a negative impact on memory usage.
1232 vm_object_split(vm_map_entry_t entry)
1234 vm_page_t m, m_next;
1235 vm_object_t orig_object, new_object, source;
1236 vm_pindex_t idx, offidxstart;
1239 orig_object = entry->object.vm_object;
1240 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1242 if (orig_object->ref_count <= 1)
1244 VM_OBJECT_UNLOCK(orig_object);
1246 offidxstart = OFF_TO_IDX(entry->offset);
1247 size = atop(entry->end - entry->start);
1250 * If swap_pager_copy() is later called, it will convert new_object
1251 * into a swap object.
1253 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1256 * At this point, the new object is still private, so the order in
1257 * which the original and new objects are locked does not matter.
1259 VM_OBJECT_LOCK(new_object);
1260 VM_OBJECT_LOCK(orig_object);
1261 source = orig_object->backing_object;
1262 if (source != NULL) {
1263 VM_OBJECT_LOCK(source);
1264 if ((source->flags & OBJ_DEAD) != 0) {
1265 VM_OBJECT_UNLOCK(source);
1266 VM_OBJECT_UNLOCK(orig_object);
1267 VM_OBJECT_UNLOCK(new_object);
1268 vm_object_deallocate(new_object);
1269 VM_OBJECT_LOCK(orig_object);
1272 LIST_INSERT_HEAD(&source->shadow_head,
1273 new_object, shadow_list);
1274 source->shadow_count++;
1275 vm_object_reference_locked(source); /* for new_object */
1276 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1277 VM_OBJECT_UNLOCK(source);
1278 new_object->backing_object_offset =
1279 orig_object->backing_object_offset + entry->offset;
1280 new_object->backing_object = source;
1282 if (orig_object->cred != NULL) {
1283 new_object->cred = orig_object->cred;
1284 crhold(orig_object->cred);
1285 new_object->charge = ptoa(size);
1286 KASSERT(orig_object->charge >= ptoa(size),
1287 ("orig_object->charge < 0"));
1288 orig_object->charge -= ptoa(size);
1291 m = vm_page_find_least(orig_object, offidxstart);
1292 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1294 m_next = TAILQ_NEXT(m, listq);
1297 * We must wait for pending I/O to complete before we can
1300 * We do not have to VM_PROT_NONE the page as mappings should
1301 * not be changed by this operation.
1303 if ((m->oflags & VPO_BUSY) || m->busy) {
1304 VM_OBJECT_UNLOCK(new_object);
1305 m->oflags |= VPO_WANTED;
1306 msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1307 VM_OBJECT_LOCK(new_object);
1311 vm_page_rename(m, new_object, idx);
1313 /* page automatically made dirty by rename and cache handled */
1316 if (orig_object->type == OBJT_SWAP) {
1318 * swap_pager_copy() can sleep, in which case the orig_object's
1319 * and new_object's locks are released and reacquired.
1321 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1324 * Transfer any cached pages from orig_object to new_object.
1326 if (__predict_false(orig_object->cache != NULL))
1327 vm_page_cache_transfer(orig_object, offidxstart,
1330 VM_OBJECT_UNLOCK(orig_object);
1331 TAILQ_FOREACH(m, &new_object->memq, listq)
1333 VM_OBJECT_UNLOCK(new_object);
1334 entry->object.vm_object = new_object;
1335 entry->offset = 0LL;
1336 vm_object_deallocate(orig_object);
1337 VM_OBJECT_LOCK(new_object);
1340 #define OBSC_TEST_ALL_SHADOWED 0x0001
1341 #define OBSC_COLLAPSE_NOWAIT 0x0002
1342 #define OBSC_COLLAPSE_WAIT 0x0004
1345 vm_object_backing_scan(vm_object_t object, int op)
1349 vm_object_t backing_object;
1350 vm_pindex_t backing_offset_index;
1352 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1353 VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1355 backing_object = object->backing_object;
1356 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1359 * Initial conditions
1361 if (op & OBSC_TEST_ALL_SHADOWED) {
1363 * We do not want to have to test for the existence of cache
1364 * or swap pages in the backing object. XXX but with the
1365 * new swapper this would be pretty easy to do.
1367 * XXX what about anonymous MAP_SHARED memory that hasn't
1368 * been ZFOD faulted yet? If we do not test for this, the
1369 * shadow test may succeed! XXX
1371 if (backing_object->type != OBJT_DEFAULT) {
1375 if (op & OBSC_COLLAPSE_WAIT) {
1376 vm_object_set_flag(backing_object, OBJ_DEAD);
1382 p = TAILQ_FIRST(&backing_object->memq);
1384 vm_page_t next = TAILQ_NEXT(p, listq);
1385 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1387 if (op & OBSC_TEST_ALL_SHADOWED) {
1391 * Ignore pages outside the parent object's range
1392 * and outside the parent object's mapping of the
1395 * note that we do not busy the backing object's
1399 p->pindex < backing_offset_index ||
1400 new_pindex >= object->size
1407 * See if the parent has the page or if the parent's
1408 * object pager has the page. If the parent has the
1409 * page but the page is not valid, the parent's
1410 * object pager must have the page.
1412 * If this fails, the parent does not completely shadow
1413 * the object and we might as well give up now.
1416 pp = vm_page_lookup(object, new_pindex);
1418 (pp == NULL || pp->valid == 0) &&
1419 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1427 * Check for busy page
1429 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1432 if (op & OBSC_COLLAPSE_NOWAIT) {
1433 if ((p->oflags & VPO_BUSY) ||
1439 } else if (op & OBSC_COLLAPSE_WAIT) {
1440 if ((p->oflags & VPO_BUSY) || p->busy) {
1441 VM_OBJECT_UNLOCK(object);
1442 p->oflags |= VPO_WANTED;
1443 msleep(p, VM_OBJECT_MTX(backing_object),
1444 PDROP | PVM, "vmocol", 0);
1445 VM_OBJECT_LOCK(object);
1446 VM_OBJECT_LOCK(backing_object);
1448 * If we slept, anything could have
1449 * happened. Since the object is
1450 * marked dead, the backing offset
1451 * should not have changed so we
1452 * just restart our scan.
1454 p = TAILQ_FIRST(&backing_object->memq);
1460 p->object == backing_object,
1461 ("vm_object_backing_scan: object mismatch")
1465 * Destroy any associated swap
1467 if (backing_object->type == OBJT_SWAP) {
1468 swap_pager_freespace(
1476 p->pindex < backing_offset_index ||
1477 new_pindex >= object->size
1480 * Page is out of the parent object's range, we
1481 * can simply destroy it.
1484 KASSERT(!pmap_page_is_mapped(p),
1485 ("freeing mapped page %p", p));
1486 if (p->wire_count == 0)
1495 pp = vm_page_lookup(object, new_pindex);
1497 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1498 (pp != NULL && pp->valid == 0)
1501 * The page in the parent is not (yet) valid.
1502 * We don't know anything about the state of
1503 * the original page. It might be mapped,
1504 * so we must avoid the next if here.
1506 * This is due to a race in vm_fault() where
1507 * we must unbusy the original (backing_obj)
1508 * page before we can (re)lock the parent.
1509 * Hence we can get here.
1516 vm_pager_has_page(object, new_pindex, NULL, NULL)
1519 * page already exists in parent OR swap exists
1520 * for this location in the parent. Destroy
1521 * the original page from the backing object.
1523 * Leave the parent's page alone
1526 KASSERT(!pmap_page_is_mapped(p),
1527 ("freeing mapped page %p", p));
1528 if (p->wire_count == 0)
1537 #if VM_NRESERVLEVEL > 0
1539 * Rename the reservation.
1541 vm_reserv_rename(p, object, backing_object,
1542 backing_offset_index);
1546 * Page does not exist in parent, rename the
1547 * page from the backing object to the main object.
1549 * If the page was mapped to a process, it can remain
1550 * mapped through the rename.
1553 vm_page_rename(p, object, new_pindex);
1555 /* page automatically made dirty by rename */
1564 * this version of collapse allows the operation to occur earlier and
1565 * when paging_in_progress is true for an object... This is not a complete
1566 * operation, but should plug 99.9% of the rest of the leaks.
1569 vm_object_qcollapse(vm_object_t object)
1571 vm_object_t backing_object = object->backing_object;
1573 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1574 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1576 if (backing_object->ref_count != 1)
1579 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1583 * vm_object_collapse:
1585 * Collapse an object with the object backing it.
1586 * Pages in the backing object are moved into the
1587 * parent, and the backing object is deallocated.
1590 vm_object_collapse(vm_object_t object)
1592 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1595 vm_object_t backing_object;
1598 * Verify that the conditions are right for collapse:
1600 * The object exists and the backing object exists.
1602 if ((backing_object = object->backing_object) == NULL)
1606 * we check the backing object first, because it is most likely
1609 VM_OBJECT_LOCK(backing_object);
1610 if (backing_object->handle != NULL ||
1611 (backing_object->type != OBJT_DEFAULT &&
1612 backing_object->type != OBJT_SWAP) ||
1613 (backing_object->flags & OBJ_DEAD) ||
1614 object->handle != NULL ||
1615 (object->type != OBJT_DEFAULT &&
1616 object->type != OBJT_SWAP) ||
1617 (object->flags & OBJ_DEAD)) {
1618 VM_OBJECT_UNLOCK(backing_object);
1623 object->paging_in_progress != 0 ||
1624 backing_object->paging_in_progress != 0
1626 vm_object_qcollapse(object);
1627 VM_OBJECT_UNLOCK(backing_object);
1631 * We know that we can either collapse the backing object (if
1632 * the parent is the only reference to it) or (perhaps) have
1633 * the parent bypass the object if the parent happens to shadow
1634 * all the resident pages in the entire backing object.
1636 * This is ignoring pager-backed pages such as swap pages.
1637 * vm_object_backing_scan fails the shadowing test in this
1640 if (backing_object->ref_count == 1) {
1642 * If there is exactly one reference to the backing
1643 * object, we can collapse it into the parent.
1645 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1647 #if VM_NRESERVLEVEL > 0
1649 * Break any reservations from backing_object.
1651 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1652 vm_reserv_break_all(backing_object);
1656 * Move the pager from backing_object to object.
1658 if (backing_object->type == OBJT_SWAP) {
1660 * swap_pager_copy() can sleep, in which case
1661 * the backing_object's and object's locks are
1662 * released and reacquired.
1667 OFF_TO_IDX(object->backing_object_offset), TRUE);
1670 * Free any cached pages from backing_object.
1672 if (__predict_false(backing_object->cache != NULL))
1673 vm_page_cache_free(backing_object, 0, 0);
1676 * Object now shadows whatever backing_object did.
1677 * Note that the reference to
1678 * backing_object->backing_object moves from within
1679 * backing_object to within object.
1681 LIST_REMOVE(object, shadow_list);
1682 backing_object->shadow_count--;
1683 if (backing_object->backing_object) {
1684 VM_OBJECT_LOCK(backing_object->backing_object);
1685 LIST_REMOVE(backing_object, shadow_list);
1687 &backing_object->backing_object->shadow_head,
1688 object, shadow_list);
1690 * The shadow_count has not changed.
1692 VM_OBJECT_UNLOCK(backing_object->backing_object);
1694 object->backing_object = backing_object->backing_object;
1695 object->backing_object_offset +=
1696 backing_object->backing_object_offset;
1699 * Discard backing_object.
1701 * Since the backing object has no pages, no pager left,
1702 * and no object references within it, all that is
1703 * necessary is to dispose of it.
1705 KASSERT(backing_object->ref_count == 1, (
1706 "backing_object %p was somehow re-referenced during collapse!",
1708 VM_OBJECT_UNLOCK(backing_object);
1709 vm_object_destroy(backing_object);
1713 vm_object_t new_backing_object;
1716 * If we do not entirely shadow the backing object,
1717 * there is nothing we can do so we give up.
1719 if (object->resident_page_count != object->size &&
1720 vm_object_backing_scan(object,
1721 OBSC_TEST_ALL_SHADOWED) == 0) {
1722 VM_OBJECT_UNLOCK(backing_object);
1727 * Make the parent shadow the next object in the
1728 * chain. Deallocating backing_object will not remove
1729 * it, since its reference count is at least 2.
1731 LIST_REMOVE(object, shadow_list);
1732 backing_object->shadow_count--;
1734 new_backing_object = backing_object->backing_object;
1735 if ((object->backing_object = new_backing_object) != NULL) {
1736 VM_OBJECT_LOCK(new_backing_object);
1738 &new_backing_object->shadow_head,
1742 new_backing_object->shadow_count++;
1743 vm_object_reference_locked(new_backing_object);
1744 VM_OBJECT_UNLOCK(new_backing_object);
1745 object->backing_object_offset +=
1746 backing_object->backing_object_offset;
1750 * Drop the reference count on backing_object. Since
1751 * its ref_count was at least 2, it will not vanish.
1753 backing_object->ref_count--;
1754 VM_OBJECT_UNLOCK(backing_object);
1759 * Try again with this object's new backing object.
1765 * vm_object_page_remove:
1767 * For the given object, either frees or invalidates each of the
1768 * specified pages. In general, a page is freed. However, if a page is
1769 * wired for any reason other than the existence of a managed, wired
1770 * mapping, then it may be invalidated but not removed from the object.
1771 * Pages are specified by the given range ["start", "end") and the option
1772 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1773 * extends from "start" to the end of the object. If the option
1774 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1775 * specified range are affected. If the option OBJPR_NOTMAPPED is
1776 * specified, then the pages within the specified range must have no
1777 * mappings. Otherwise, if this option is not specified, any mappings to
1778 * the specified pages are removed before the pages are freed or
1781 * In general, this operation should only be performed on objects that
1782 * contain managed pages. There are, however, two exceptions. First, it
1783 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1784 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1785 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1786 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1788 * The object must be locked.
1791 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1797 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1798 KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_PHYS) ||
1799 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1800 ("vm_object_page_remove: illegal options for object %p", object));
1801 if (object->resident_page_count == 0)
1803 vm_object_pip_add(object, 1);
1805 p = vm_page_find_least(object, start);
1808 * Here, the variable "p" is either (1) the page with the least pindex
1809 * greater than or equal to the parameter "start" or (2) NULL.
1811 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1812 next = TAILQ_NEXT(p, listq);
1815 * If the page is wired for any reason besides the existence
1816 * of managed, wired mappings, then it cannot be freed. For
1817 * example, fictitious pages, which represent device memory,
1818 * are inherently wired and cannot be freed. They can,
1819 * however, be invalidated if the option OBJPR_CLEANONLY is
1823 if ((wirings = p->wire_count) != 0 &&
1824 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1825 if ((options & OBJPR_NOTMAPPED) == 0) {
1827 /* Account for removal of wired mappings. */
1829 p->wire_count -= wirings;
1831 if ((options & OBJPR_CLEANONLY) == 0) {
1838 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1840 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1841 ("vm_object_page_remove: page %p is fictitious", p));
1842 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1843 if ((options & OBJPR_NOTMAPPED) == 0)
1844 pmap_remove_write(p);
1850 if ((options & OBJPR_NOTMAPPED) == 0) {
1852 /* Account for removal of wired mappings. */
1854 p->wire_count -= wirings;
1859 vm_object_pip_wakeup(object);
1861 if (__predict_false(object->cache != NULL))
1862 vm_page_cache_free(object, start, end);
1866 * Populate the specified range of the object with valid pages. Returns
1867 * TRUE if the range is successfully populated and FALSE otherwise.
1869 * Note: This function should be optimized to pass a larger array of
1870 * pages to vm_pager_get_pages() before it is applied to a non-
1871 * OBJT_DEVICE object.
1873 * The object must be locked.
1876 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1882 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1883 for (pindex = start; pindex < end; pindex++) {
1884 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1886 if (m->valid != VM_PAGE_BITS_ALL) {
1888 rv = vm_pager_get_pages(object, ma, 1, 0);
1889 m = vm_page_lookup(object, pindex);
1892 if (rv != VM_PAGER_OK) {
1900 * Keep "m" busy because a subsequent iteration may unlock
1904 if (pindex > start) {
1905 m = vm_page_lookup(object, start);
1906 while (m != NULL && m->pindex < pindex) {
1908 m = TAILQ_NEXT(m, listq);
1911 return (pindex == end);
1915 * Routine: vm_object_coalesce
1916 * Function: Coalesces two objects backing up adjoining
1917 * regions of memory into a single object.
1919 * returns TRUE if objects were combined.
1921 * NOTE: Only works at the moment if the second object is NULL -
1922 * if it's not, which object do we lock first?
1925 * prev_object First object to coalesce
1926 * prev_offset Offset into prev_object
1927 * prev_size Size of reference to prev_object
1928 * next_size Size of reference to the second object
1929 * reserved Indicator that extension region has
1930 * swap accounted for
1933 * The object must *not* be locked.
1936 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
1937 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
1939 vm_pindex_t next_pindex;
1941 if (prev_object == NULL)
1943 VM_OBJECT_LOCK(prev_object);
1944 if (prev_object->type != OBJT_DEFAULT &&
1945 prev_object->type != OBJT_SWAP) {
1946 VM_OBJECT_UNLOCK(prev_object);
1951 * Try to collapse the object first
1953 vm_object_collapse(prev_object);
1956 * Can't coalesce if: . more than one reference . paged out . shadows
1957 * another object . has a copy elsewhere (any of which mean that the
1958 * pages not mapped to prev_entry may be in use anyway)
1960 if (prev_object->backing_object != NULL) {
1961 VM_OBJECT_UNLOCK(prev_object);
1965 prev_size >>= PAGE_SHIFT;
1966 next_size >>= PAGE_SHIFT;
1967 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
1969 if ((prev_object->ref_count > 1) &&
1970 (prev_object->size != next_pindex)) {
1971 VM_OBJECT_UNLOCK(prev_object);
1976 * Account for the charge.
1978 if (prev_object->cred != NULL) {
1981 * If prev_object was charged, then this mapping,
1982 * althought not charged now, may become writable
1983 * later. Non-NULL cred in the object would prevent
1984 * swap reservation during enabling of the write
1985 * access, so reserve swap now. Failed reservation
1986 * cause allocation of the separate object for the map
1987 * entry, and swap reservation for this entry is
1988 * managed in appropriate time.
1990 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
1991 prev_object->cred)) {
1994 prev_object->charge += ptoa(next_size);
1998 * Remove any pages that may still be in the object from a previous
2001 if (next_pindex < prev_object->size) {
2002 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2004 if (prev_object->type == OBJT_SWAP)
2005 swap_pager_freespace(prev_object,
2006 next_pindex, next_size);
2008 if (prev_object->cred != NULL) {
2009 KASSERT(prev_object->charge >=
2010 ptoa(prev_object->size - next_pindex),
2011 ("object %p overcharged 1 %jx %jx", prev_object,
2012 (uintmax_t)next_pindex, (uintmax_t)next_size));
2013 prev_object->charge -= ptoa(prev_object->size -
2020 * Extend the object if necessary.
2022 if (next_pindex + next_size > prev_object->size)
2023 prev_object->size = next_pindex + next_size;
2025 VM_OBJECT_UNLOCK(prev_object);
2030 vm_object_set_writeable_dirty(vm_object_t object)
2033 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2034 if (object->type != OBJT_VNODE)
2036 object->generation++;
2037 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2039 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2042 #include "opt_ddb.h"
2044 #include <sys/kernel.h>
2046 #include <sys/cons.h>
2048 #include <ddb/ddb.h>
2051 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2054 vm_map_entry_t tmpe;
2062 tmpe = map->header.next;
2063 entcount = map->nentries;
2064 while (entcount-- && (tmpe != &map->header)) {
2065 if (_vm_object_in_map(map, object, tmpe)) {
2070 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2071 tmpm = entry->object.sub_map;
2072 tmpe = tmpm->header.next;
2073 entcount = tmpm->nentries;
2074 while (entcount-- && tmpe != &tmpm->header) {
2075 if (_vm_object_in_map(tmpm, object, tmpe)) {
2080 } else if ((obj = entry->object.vm_object) != NULL) {
2081 for (; obj; obj = obj->backing_object)
2082 if (obj == object) {
2090 vm_object_in_map(vm_object_t object)
2094 /* sx_slock(&allproc_lock); */
2095 FOREACH_PROC_IN_SYSTEM(p) {
2096 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2098 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2099 /* sx_sunlock(&allproc_lock); */
2103 /* sx_sunlock(&allproc_lock); */
2104 if (_vm_object_in_map(kernel_map, object, 0))
2106 if (_vm_object_in_map(kmem_map, object, 0))
2108 if (_vm_object_in_map(pager_map, object, 0))
2110 if (_vm_object_in_map(buffer_map, object, 0))
2115 DB_SHOW_COMMAND(vmochk, vm_object_check)
2120 * make sure that internal objs are in a map somewhere
2121 * and none have zero ref counts.
2123 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2124 if (object->handle == NULL &&
2125 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2126 if (object->ref_count == 0) {
2127 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2128 (long)object->size);
2130 if (!vm_object_in_map(object)) {
2132 "vmochk: internal obj is not in a map: "
2133 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2134 object->ref_count, (u_long)object->size,
2135 (u_long)object->size,
2136 (void *)object->backing_object);
2143 * vm_object_print: [ debug ]
2145 DB_SHOW_COMMAND(object, vm_object_print_static)
2147 /* XXX convert args. */
2148 vm_object_t object = (vm_object_t)addr;
2149 boolean_t full = have_addr;
2153 /* XXX count is an (unused) arg. Avoid shadowing it. */
2154 #define count was_count
2162 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2163 object, (int)object->type, (uintmax_t)object->size,
2164 object->resident_page_count, object->ref_count, object->flags,
2165 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2166 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2167 object->shadow_count,
2168 object->backing_object ? object->backing_object->ref_count : 0,
2169 object->backing_object, (uintmax_t)object->backing_object_offset);
2176 TAILQ_FOREACH(p, &object->memq, listq) {
2178 db_iprintf("memory:=");
2179 else if (count == 6) {
2187 db_printf("(off=0x%jx,page=0x%jx)",
2188 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2198 /* XXX need this non-static entry for calling from vm_map_print. */
2201 /* db_expr_t */ long addr,
2202 boolean_t have_addr,
2203 /* db_expr_t */ long count,
2206 vm_object_print_static(addr, have_addr, count, modif);
2209 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2214 vm_page_t m, prev_m;
2218 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2219 db_printf("new object: %p\n", (void *)object);
2230 TAILQ_FOREACH(m, &object->memq, listq) {
2231 if (m->pindex > 128)
2233 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2234 prev_m->pindex + 1 != m->pindex) {
2236 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2237 (long)fidx, rcount, (long)pa);
2249 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2254 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2255 (long)fidx, rcount, (long)pa);
2265 pa = VM_PAGE_TO_PHYS(m);
2269 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2270 (long)fidx, rcount, (long)pa);