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>
99 static int msync_flush_flags = 0;
100 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags, CTLFLAG_RW, &msync_flush_flags, 0,
101 "Does nothing; kept for backward compatibility");
103 static int old_msync;
104 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
105 "Use old (insecure) msync behavior");
107 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
108 int pagerflags, int flags, int *clearobjflags);
109 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
111 static void vm_object_qcollapse(vm_object_t object);
112 static void vm_object_vndeallocate(vm_object_t object);
115 * Virtual memory objects maintain the actual data
116 * associated with allocated virtual memory. A given
117 * page of memory exists within exactly one object.
119 * An object is only deallocated when all "references"
120 * are given up. Only one "reference" to a given
121 * region of an object should be writeable.
123 * Associated with each object is a list of all resident
124 * memory pages belonging to that object; this list is
125 * maintained by the "vm_page" module, and locked by the object's
128 * Each object also records a "pager" routine which is
129 * used to retrieve (and store) pages to the proper backing
130 * storage. In addition, objects may be backed by other
131 * objects from which they were virtual-copied.
133 * The only items within the object structure which are
134 * modified after time of creation are:
135 * reference count locked by object's lock
136 * pager routine locked by object's lock
140 struct object_q vm_object_list;
141 struct mtx vm_object_list_mtx; /* lock for object list and count */
143 struct vm_object kernel_object_store;
144 struct vm_object kmem_object_store;
146 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats");
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",
172 #if VM_NRESERVLEVEL > 0
173 KASSERT(LIST_EMPTY(&object->rvq),
174 ("object %p has reservations",
177 KASSERT(object->cache == NULL,
178 ("object %p has cached pages",
180 KASSERT(object->paging_in_progress == 0,
181 ("object %p paging_in_progress = %d",
182 object, object->paging_in_progress));
183 KASSERT(object->resident_page_count == 0,
184 ("object %p resident_page_count = %d",
185 object, object->resident_page_count));
186 KASSERT(object->shadow_count == 0,
187 ("object %p shadow_count = %d",
188 object, object->shadow_count));
193 vm_object_zinit(void *mem, int size, int flags)
197 object = (vm_object_t)mem;
198 bzero(&object->mtx, sizeof(object->mtx));
199 VM_OBJECT_LOCK_INIT(object, "standard object");
201 /* These are true for any object that has been freed */
202 object->paging_in_progress = 0;
203 object->resident_page_count = 0;
204 object->shadow_count = 0;
209 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
212 TAILQ_INIT(&object->memq);
213 LIST_INIT(&object->shadow_head);
218 object->generation = 1;
219 object->ref_count = 1;
220 object->memattr = VM_MEMATTR_DEFAULT;
224 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
225 object->flags = OBJ_ONEMAPPING;
226 object->pg_color = 0;
227 object->handle = NULL;
228 object->backing_object = NULL;
229 object->backing_object_offset = (vm_ooffset_t) 0;
230 #if VM_NRESERVLEVEL > 0
231 LIST_INIT(&object->rvq);
233 object->cache = NULL;
235 mtx_lock(&vm_object_list_mtx);
236 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
237 mtx_unlock(&vm_object_list_mtx);
243 * Initialize the VM objects module.
248 TAILQ_INIT(&vm_object_list);
249 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
251 VM_OBJECT_LOCK_INIT(kernel_object, "kernel object");
252 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
254 #if VM_NRESERVLEVEL > 0
255 kernel_object->flags |= OBJ_COLORED;
256 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
259 VM_OBJECT_LOCK_INIT(kmem_object, "kmem object");
260 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
262 #if VM_NRESERVLEVEL > 0
263 kmem_object->flags |= OBJ_COLORED;
264 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
268 * The lock portion of struct vm_object must be type stable due
269 * to vm_pageout_fallback_object_lock locking a vm object
270 * without holding any references to it.
272 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
278 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
282 vm_object_clear_flag(vm_object_t object, u_short bits)
285 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
286 object->flags &= ~bits;
290 * Sets the default memory attribute for the specified object. Pages
291 * that are allocated to this object are by default assigned this memory
294 * Presently, this function must be called before any pages are allocated
295 * to the object. In the future, this requirement may be relaxed for
296 * "default" and "swap" objects.
299 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
302 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
303 switch (object->type) {
310 if (!TAILQ_EMPTY(&object->memq))
311 return (KERN_FAILURE);
314 return (KERN_INVALID_ARGUMENT);
316 object->memattr = memattr;
317 return (KERN_SUCCESS);
321 vm_object_pip_add(vm_object_t object, short i)
324 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
325 object->paging_in_progress += i;
329 vm_object_pip_subtract(vm_object_t object, short i)
332 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
333 object->paging_in_progress -= i;
337 vm_object_pip_wakeup(vm_object_t object)
340 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
341 object->paging_in_progress--;
342 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
343 vm_object_clear_flag(object, OBJ_PIPWNT);
349 vm_object_pip_wakeupn(vm_object_t object, short i)
352 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
354 object->paging_in_progress -= i;
355 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
356 vm_object_clear_flag(object, OBJ_PIPWNT);
362 vm_object_pip_wait(vm_object_t object, char *waitid)
365 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
366 while (object->paging_in_progress) {
367 object->flags |= OBJ_PIPWNT;
368 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
373 * vm_object_allocate:
375 * Returns a new object with the given size.
378 vm_object_allocate(objtype_t type, vm_pindex_t size)
382 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
383 _vm_object_allocate(type, size, object);
389 * vm_object_reference:
391 * Gets another reference to the given object. Note: OBJ_DEAD
392 * objects can be referenced during final cleaning.
395 vm_object_reference(vm_object_t object)
399 VM_OBJECT_LOCK(object);
400 vm_object_reference_locked(object);
401 VM_OBJECT_UNLOCK(object);
405 * vm_object_reference_locked:
407 * Gets another reference to the given object.
409 * The object must be locked.
412 vm_object_reference_locked(vm_object_t object)
416 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
418 if (object->type == OBJT_VNODE) {
425 * Handle deallocating an object of type OBJT_VNODE.
428 vm_object_vndeallocate(vm_object_t object)
430 struct vnode *vp = (struct vnode *) object->handle;
432 VFS_ASSERT_GIANT(vp->v_mount);
433 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
434 KASSERT(object->type == OBJT_VNODE,
435 ("vm_object_vndeallocate: not a vnode object"));
436 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
438 if (object->ref_count == 0) {
439 vprint("vm_object_vndeallocate", vp);
440 panic("vm_object_vndeallocate: bad object reference count");
444 if (object->ref_count > 1) {
446 VM_OBJECT_UNLOCK(object);
447 /* vrele may need the vnode lock. */
451 VM_OBJECT_UNLOCK(object);
452 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
454 VM_OBJECT_LOCK(object);
456 if (object->type == OBJT_DEAD) {
457 VM_OBJECT_UNLOCK(object);
460 if (object->ref_count == 0)
461 vp->v_vflag &= ~VV_TEXT;
462 VM_OBJECT_UNLOCK(object);
469 * vm_object_deallocate:
471 * Release a reference to the specified object,
472 * gained either through a vm_object_allocate
473 * or a vm_object_reference call. When all references
474 * are gone, storage associated with this object
475 * may be relinquished.
477 * No object may be locked.
480 vm_object_deallocate(vm_object_t object)
484 while (object != NULL) {
489 VM_OBJECT_LOCK(object);
490 if (object->type == OBJT_VNODE) {
491 struct vnode *vp = (struct vnode *) object->handle;
494 * Conditionally acquire Giant for a vnode-backed
495 * object. We have to be careful since the type of
496 * a vnode object can change while the object is
499 if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
501 if (!mtx_trylock(&Giant)) {
502 VM_OBJECT_UNLOCK(object);
507 vm_object_vndeallocate(object);
508 VFS_UNLOCK_GIANT(vfslocked);
512 * This is to handle the case that the object
513 * changed type while we dropped its lock to
516 VFS_UNLOCK_GIANT(vfslocked);
518 KASSERT(object->ref_count != 0,
519 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
522 * If the reference count goes to 0 we start calling
523 * vm_object_terminate() on the object chain.
524 * A ref count of 1 may be a special case depending on the
525 * shadow count being 0 or 1.
528 if (object->ref_count > 1) {
529 VM_OBJECT_UNLOCK(object);
531 } else if (object->ref_count == 1) {
532 if (object->shadow_count == 0 &&
533 object->handle == NULL &&
534 (object->type == OBJT_DEFAULT ||
535 object->type == OBJT_SWAP)) {
536 vm_object_set_flag(object, OBJ_ONEMAPPING);
537 } else if ((object->shadow_count == 1) &&
538 (object->handle == NULL) &&
539 (object->type == OBJT_DEFAULT ||
540 object->type == OBJT_SWAP)) {
543 robject = LIST_FIRST(&object->shadow_head);
544 KASSERT(robject != NULL,
545 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
547 object->shadow_count));
548 if (!VM_OBJECT_TRYLOCK(robject)) {
550 * Avoid a potential deadlock.
553 VM_OBJECT_UNLOCK(object);
555 * More likely than not the thread
556 * holding robject's lock has lower
557 * priority than the current thread.
558 * Let the lower priority thread run.
564 * Collapse object into its shadow unless its
565 * shadow is dead. In that case, object will
566 * be deallocated by the thread that is
567 * deallocating its shadow.
569 if ((robject->flags & OBJ_DEAD) == 0 &&
570 (robject->handle == NULL) &&
571 (robject->type == OBJT_DEFAULT ||
572 robject->type == OBJT_SWAP)) {
574 robject->ref_count++;
576 if (robject->paging_in_progress) {
577 VM_OBJECT_UNLOCK(object);
578 vm_object_pip_wait(robject,
580 temp = robject->backing_object;
581 if (object == temp) {
582 VM_OBJECT_LOCK(object);
585 } else if (object->paging_in_progress) {
586 VM_OBJECT_UNLOCK(robject);
587 object->flags |= OBJ_PIPWNT;
589 VM_OBJECT_MTX(object),
590 PDROP | PVM, "objde2", 0);
591 VM_OBJECT_LOCK(robject);
592 temp = robject->backing_object;
593 if (object == temp) {
594 VM_OBJECT_LOCK(object);
598 VM_OBJECT_UNLOCK(object);
600 if (robject->ref_count == 1) {
601 robject->ref_count--;
606 vm_object_collapse(object);
607 VM_OBJECT_UNLOCK(object);
610 VM_OBJECT_UNLOCK(robject);
612 VM_OBJECT_UNLOCK(object);
616 temp = object->backing_object;
618 VM_OBJECT_LOCK(temp);
619 LIST_REMOVE(object, shadow_list);
620 temp->shadow_count--;
621 VM_OBJECT_UNLOCK(temp);
622 object->backing_object = NULL;
625 * Don't double-terminate, we could be in a termination
626 * recursion due to the terminate having to sync data
629 if ((object->flags & OBJ_DEAD) == 0)
630 vm_object_terminate(object);
632 VM_OBJECT_UNLOCK(object);
638 * vm_object_destroy removes the object from the global object list
639 * and frees the space for the object.
642 vm_object_destroy(vm_object_t object)
646 * Remove the object from the global object list.
648 mtx_lock(&vm_object_list_mtx);
649 TAILQ_REMOVE(&vm_object_list, object, object_list);
650 mtx_unlock(&vm_object_list_mtx);
653 * Release the allocation charge.
655 if (object->uip != NULL) {
656 KASSERT(object->type == OBJT_DEFAULT ||
657 object->type == OBJT_SWAP,
658 ("vm_object_terminate: non-swap obj %p has uip",
660 swap_release_by_uid(object->charge, object->uip);
667 * Free the space for the object.
669 uma_zfree(obj_zone, object);
673 * vm_object_terminate actually destroys the specified object, freeing
674 * up all previously used resources.
676 * The object must be locked.
677 * This routine may block.
680 vm_object_terminate(vm_object_t object)
684 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
687 * Make sure no one uses us.
689 vm_object_set_flag(object, OBJ_DEAD);
692 * wait for the pageout daemon to be done with the object
694 vm_object_pip_wait(object, "objtrm");
696 KASSERT(!object->paging_in_progress,
697 ("vm_object_terminate: pageout in progress"));
700 * Clean and free the pages, as appropriate. All references to the
701 * object are gone, so we don't need to lock it.
703 if (object->type == OBJT_VNODE) {
704 struct vnode *vp = (struct vnode *)object->handle;
707 * Clean pages and flush buffers.
709 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
710 VM_OBJECT_UNLOCK(object);
712 vinvalbuf(vp, V_SAVE, 0, 0);
714 VM_OBJECT_LOCK(object);
717 KASSERT(object->ref_count == 0,
718 ("vm_object_terminate: object with references, ref_count=%d",
722 * Free any remaining pageable pages. This also removes them from the
723 * paging queues. However, don't free wired pages, just remove them
724 * from the object. Rather than incrementally removing each page from
725 * the object, the page and object are reset to any empty state.
727 vm_page_lock_queues();
728 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
729 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
730 ("vm_object_terminate: freeing busy page %p", p));
732 * Optimize the page's removal from the object by resetting
733 * its "object" field. Specifically, if the page is not
734 * wired, then the effect of this assignment is that
735 * vm_page_free()'s call to vm_page_remove() will return
736 * immediately without modifying the page or the object.
739 if (p->wire_count == 0) {
744 vm_page_unlock_queues();
746 * If the object contained any pages, then reset it to an empty state.
747 * None of the object's fields, including "resident_page_count", were
748 * modified by the preceding loop.
750 if (object->resident_page_count != 0) {
752 TAILQ_INIT(&object->memq);
753 object->resident_page_count = 0;
754 if (object->type == OBJT_VNODE)
755 vdrop(object->handle);
758 #if VM_NRESERVLEVEL > 0
759 if (__predict_false(!LIST_EMPTY(&object->rvq)))
760 vm_reserv_break_all(object);
762 if (__predict_false(object->cache != NULL))
763 vm_page_cache_free(object, 0, 0);
766 * Let the pager know object is dead.
768 vm_pager_deallocate(object);
769 VM_OBJECT_UNLOCK(object);
771 vm_object_destroy(object);
775 vm_object_page_remove_write(vm_page_t p, int flags, int *clearobjflags)
779 * If we have been asked to skip nosync pages and this is a
780 * nosync page, skip it. Note that the object flags were not
781 * cleared in this case so we do not have to set them.
783 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
787 pmap_remove_write(p);
788 return (p->dirty != 0);
793 * vm_object_page_clean
795 * Clean all dirty pages in the specified range of object. Leaves page
796 * on whatever queue it is currently on. If NOSYNC is set then do not
797 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
798 * leaving the object dirty.
800 * When stuffing pages asynchronously, allow clustering. XXX we need a
801 * synchronous clustering mode implementation.
803 * Odd semantics: if start == end, we clean everything.
805 * The object must be locked.
808 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
812 vm_pindex_t pi, tend;
813 int clearobjflags, curgeneration, n, pagerflags;
815 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
816 KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
817 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
818 object->resident_page_count == 0)
821 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
822 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
823 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
825 tend = (end == 0) ? object->size : end;
827 vm_object_set_flag(object, OBJ_CLEANING);
829 vm_page_lock_queues();
832 * Make the page read-only so we can then clear the object flags.
834 * However, if this is a nosync mmap then the object is likely to
835 * stay dirty so do not mess with the page and do not clear the
841 curgeneration = object->generation;
843 for (p = vm_page_find_least(object, start); p != NULL; p = np) {
847 np = TAILQ_NEXT(p, listq);
850 if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
851 vm_page_lock_queues();
852 if (object->generation != curgeneration)
854 np = vm_page_find_least(object, pi);
857 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
860 n = vm_object_page_collect_flush(object, p, pagerflags,
861 flags, &clearobjflags);
862 if (object->generation != curgeneration)
866 * If the VOP_PUTPAGES() did a truncated write, so
867 * that even the first page of the run is not fully
868 * written, vm_pageout_flush() returns 0 as the run
869 * length. Since the condition that caused truncated
870 * write may be permanent, e.g. exhausted free space,
871 * accepting n == 0 would cause an infinite loop.
873 * Forwarding the iterator leaves the unwritten page
874 * behind, but there is not much we can do there if
875 * filesystem refuses to write it.
879 np = vm_page_find_least(object, pi + n);
881 vm_page_unlock_queues();
883 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
886 vm_object_clear_flag(object, OBJ_CLEANING);
887 if (clearobjflags && start == 0 && tend == object->size)
888 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
892 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
893 int flags, int *clearobjflags)
895 vm_page_t ma[vm_pageout_page_count], p_first, tp;
896 int count, i, mreq, runlen;
898 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
903 for (tp = p; count < vm_pageout_page_count; count++) {
904 tp = vm_page_next(tp);
905 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
907 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
911 for (p_first = p; count < vm_pageout_page_count; count++) {
912 tp = vm_page_prev(p_first);
913 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
915 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
921 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
924 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen);
929 * Note that there is absolutely no sense in writing out
930 * anonymous objects, so we track down the vnode object
932 * We invalidate (remove) all pages from the address space
933 * for semantic correctness.
935 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
936 * may start out with a NULL object.
939 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
940 boolean_t syncio, boolean_t invalidate)
942 vm_object_t backing_object;
945 int flags, fsync_after;
949 VM_OBJECT_LOCK(object);
950 while ((backing_object = object->backing_object) != NULL) {
951 VM_OBJECT_LOCK(backing_object);
952 offset += object->backing_object_offset;
953 VM_OBJECT_UNLOCK(object);
954 object = backing_object;
955 if (object->size < OFF_TO_IDX(offset + size))
956 size = IDX_TO_OFF(object->size) - offset;
959 * Flush pages if writing is allowed, invalidate them
960 * if invalidation requested. Pages undergoing I/O
961 * will be ignored by vm_object_page_remove().
963 * We cannot lock the vnode and then wait for paging
964 * to complete without deadlocking against vm_fault.
965 * Instead we simply call vm_object_page_remove() and
966 * allow it to block internally on a page-by-page
967 * basis when it encounters pages undergoing async
970 if (object->type == OBJT_VNODE &&
971 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
974 VM_OBJECT_UNLOCK(object);
975 (void) vn_start_write(vp, &mp, V_WAIT);
976 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
977 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
978 if (syncio && !invalidate && offset == 0 &&
979 OFF_TO_IDX(size) == object->size) {
981 * If syncing the whole mapping of the file,
982 * it is faster to schedule all the writes in
983 * async mode, also allowing the clustering,
984 * and then wait for i/o to complete.
989 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
990 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
993 VM_OBJECT_LOCK(object);
994 vm_object_page_clean(object,
996 OFF_TO_IDX(offset + size + PAGE_MASK),
998 VM_OBJECT_UNLOCK(object);
1000 (void) VOP_FSYNC(vp, MNT_WAIT, curthread);
1002 VFS_UNLOCK_GIANT(vfslocked);
1003 vn_finished_write(mp);
1004 VM_OBJECT_LOCK(object);
1006 if ((object->type == OBJT_VNODE ||
1007 object->type == OBJT_DEVICE) && invalidate) {
1009 purge = old_msync || (object->type == OBJT_DEVICE);
1010 vm_object_page_remove(object,
1012 OFF_TO_IDX(offset + size + PAGE_MASK),
1013 purge ? FALSE : TRUE);
1015 VM_OBJECT_UNLOCK(object);
1019 * vm_object_madvise:
1021 * Implements the madvise function at the object/page level.
1023 * MADV_WILLNEED (any object)
1025 * Activate the specified pages if they are resident.
1027 * MADV_DONTNEED (any object)
1029 * Deactivate the specified pages if they are resident.
1031 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1032 * OBJ_ONEMAPPING only)
1034 * Deactivate and clean the specified pages if they are
1035 * resident. This permits the process to reuse the pages
1036 * without faulting or the kernel to reclaim the pages
1040 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1042 vm_pindex_t end, tpindex;
1043 vm_object_t backing_object, tobject;
1048 VM_OBJECT_LOCK(object);
1049 end = pindex + count;
1051 * Locate and adjust resident pages
1053 for (; pindex < end; pindex += 1) {
1059 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1060 * and those pages must be OBJ_ONEMAPPING.
1062 if (advise == MADV_FREE) {
1063 if ((tobject->type != OBJT_DEFAULT &&
1064 tobject->type != OBJT_SWAP) ||
1065 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1066 goto unlock_tobject;
1068 } else if (tobject->type == OBJT_PHYS)
1069 goto unlock_tobject;
1070 m = vm_page_lookup(tobject, tpindex);
1071 if (m == NULL && advise == MADV_WILLNEED) {
1073 * If the page is cached, reactivate it.
1075 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1080 * There may be swap even if there is no backing page
1082 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1083 swap_pager_freespace(tobject, tpindex, 1);
1087 backing_object = tobject->backing_object;
1088 if (backing_object == NULL)
1089 goto unlock_tobject;
1090 VM_OBJECT_LOCK(backing_object);
1091 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1092 if (tobject != object)
1093 VM_OBJECT_UNLOCK(tobject);
1094 tobject = backing_object;
1096 } else if (m->valid != VM_PAGE_BITS_ALL)
1097 goto unlock_tobject;
1099 * If the page is not in a normal state, skip it.
1101 vm_page_lock_queues();
1102 if (m->hold_count != 0 || m->wire_count != 0) {
1103 vm_page_unlock_queues();
1104 goto unlock_tobject;
1106 if ((m->oflags & VPO_BUSY) || m->busy) {
1107 if (advise == MADV_WILLNEED)
1109 * Reference the page before unlocking and
1110 * sleeping so that the page daemon is less
1111 * likely to reclaim it.
1113 vm_page_flag_set(m, PG_REFERENCED);
1114 vm_page_unlock_queues();
1115 if (object != tobject)
1116 VM_OBJECT_UNLOCK(object);
1117 m->oflags |= VPO_WANTED;
1118 msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1120 VM_OBJECT_LOCK(object);
1123 if (advise == MADV_WILLNEED) {
1124 vm_page_activate(m);
1125 } else if (advise == MADV_DONTNEED) {
1126 vm_page_dontneed(m);
1127 } else if (advise == MADV_FREE) {
1129 * Mark the page clean. This will allow the page
1130 * to be freed up by the system. However, such pages
1131 * are often reused quickly by malloc()/free()
1132 * so we do not do anything that would cause
1133 * a page fault if we can help it.
1135 * Specifically, we do not try to actually free
1136 * the page now nor do we try to put it in the
1137 * cache (which would cause a page fault on reuse).
1139 * But we do make the page is freeable as we
1140 * can without actually taking the step of unmapping
1143 pmap_clear_modify(m);
1146 vm_page_dontneed(m);
1148 vm_page_unlock_queues();
1149 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1150 swap_pager_freespace(tobject, tpindex, 1);
1152 if (tobject != object)
1153 VM_OBJECT_UNLOCK(tobject);
1155 VM_OBJECT_UNLOCK(object);
1161 * Create a new object which is backed by the
1162 * specified existing object range. The source
1163 * object reference is deallocated.
1165 * The new object and offset into that object
1166 * are returned in the source parameters.
1170 vm_object_t *object, /* IN/OUT */
1171 vm_ooffset_t *offset, /* IN/OUT */
1180 * Don't create the new object if the old object isn't shared.
1182 if (source != NULL) {
1183 VM_OBJECT_LOCK(source);
1184 if (source->ref_count == 1 &&
1185 source->handle == NULL &&
1186 (source->type == OBJT_DEFAULT ||
1187 source->type == OBJT_SWAP)) {
1188 VM_OBJECT_UNLOCK(source);
1191 VM_OBJECT_UNLOCK(source);
1195 * Allocate a new object with the given length.
1197 result = vm_object_allocate(OBJT_DEFAULT, length);
1200 * The new object shadows the source object, adding a reference to it.
1201 * Our caller changes his reference to point to the new object,
1202 * removing a reference to the source object. Net result: no change
1203 * of reference count.
1205 * Try to optimize the result object's page color when shadowing
1206 * in order to maintain page coloring consistency in the combined
1209 result->backing_object = source;
1211 * Store the offset into the source object, and fix up the offset into
1214 result->backing_object_offset = *offset;
1215 if (source != NULL) {
1216 VM_OBJECT_LOCK(source);
1217 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1218 source->shadow_count++;
1219 #if VM_NRESERVLEVEL > 0
1220 result->flags |= source->flags & OBJ_COLORED;
1221 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1222 ((1 << (VM_NFREEORDER - 1)) - 1);
1224 VM_OBJECT_UNLOCK(source);
1229 * Return the new things
1238 * Split the pages in a map entry into a new object. This affords
1239 * easier removal of unused pages, and keeps object inheritance from
1240 * being a negative impact on memory usage.
1243 vm_object_split(vm_map_entry_t entry)
1245 vm_page_t m, m_next;
1246 vm_object_t orig_object, new_object, source;
1247 vm_pindex_t idx, offidxstart;
1250 orig_object = entry->object.vm_object;
1251 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1253 if (orig_object->ref_count <= 1)
1255 VM_OBJECT_UNLOCK(orig_object);
1257 offidxstart = OFF_TO_IDX(entry->offset);
1258 size = atop(entry->end - entry->start);
1261 * If swap_pager_copy() is later called, it will convert new_object
1262 * into a swap object.
1264 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1267 * At this point, the new object is still private, so the order in
1268 * which the original and new objects are locked does not matter.
1270 VM_OBJECT_LOCK(new_object);
1271 VM_OBJECT_LOCK(orig_object);
1272 source = orig_object->backing_object;
1273 if (source != NULL) {
1274 VM_OBJECT_LOCK(source);
1275 if ((source->flags & OBJ_DEAD) != 0) {
1276 VM_OBJECT_UNLOCK(source);
1277 VM_OBJECT_UNLOCK(orig_object);
1278 VM_OBJECT_UNLOCK(new_object);
1279 vm_object_deallocate(new_object);
1280 VM_OBJECT_LOCK(orig_object);
1283 LIST_INSERT_HEAD(&source->shadow_head,
1284 new_object, shadow_list);
1285 source->shadow_count++;
1286 vm_object_reference_locked(source); /* for new_object */
1287 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1288 VM_OBJECT_UNLOCK(source);
1289 new_object->backing_object_offset =
1290 orig_object->backing_object_offset + entry->offset;
1291 new_object->backing_object = source;
1293 if (orig_object->uip != NULL) {
1294 new_object->uip = orig_object->uip;
1295 uihold(orig_object->uip);
1296 new_object->charge = ptoa(size);
1297 KASSERT(orig_object->charge >= ptoa(size),
1298 ("orig_object->charge < 0"));
1299 orig_object->charge -= ptoa(size);
1302 m = vm_page_find_least(orig_object, offidxstart);
1303 vm_page_lock_queues();
1304 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1306 m_next = TAILQ_NEXT(m, listq);
1309 * We must wait for pending I/O to complete before we can
1312 * We do not have to VM_PROT_NONE the page as mappings should
1313 * not be changed by this operation.
1315 if ((m->oflags & VPO_BUSY) || m->busy) {
1316 vm_page_unlock_queues();
1317 VM_OBJECT_UNLOCK(new_object);
1318 m->oflags |= VPO_WANTED;
1319 msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1320 VM_OBJECT_LOCK(new_object);
1323 vm_page_rename(m, new_object, idx);
1324 /* page automatically made dirty by rename and cache handled */
1327 vm_page_unlock_queues();
1328 if (orig_object->type == OBJT_SWAP) {
1330 * swap_pager_copy() can sleep, in which case the orig_object's
1331 * and new_object's locks are released and reacquired.
1333 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1336 * Transfer any cached pages from orig_object to new_object.
1338 if (__predict_false(orig_object->cache != NULL))
1339 vm_page_cache_transfer(orig_object, offidxstart,
1342 VM_OBJECT_UNLOCK(orig_object);
1343 TAILQ_FOREACH(m, &new_object->memq, listq)
1345 VM_OBJECT_UNLOCK(new_object);
1346 entry->object.vm_object = new_object;
1347 entry->offset = 0LL;
1348 vm_object_deallocate(orig_object);
1349 VM_OBJECT_LOCK(new_object);
1352 #define OBSC_TEST_ALL_SHADOWED 0x0001
1353 #define OBSC_COLLAPSE_NOWAIT 0x0002
1354 #define OBSC_COLLAPSE_WAIT 0x0004
1357 vm_object_backing_scan(vm_object_t object, int op)
1361 vm_object_t backing_object;
1362 vm_pindex_t backing_offset_index;
1364 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1365 VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1367 backing_object = object->backing_object;
1368 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1371 * Initial conditions
1373 if (op & OBSC_TEST_ALL_SHADOWED) {
1375 * We do not want to have to test for the existence of cache
1376 * or swap pages in the backing object. XXX but with the
1377 * new swapper this would be pretty easy to do.
1379 * XXX what about anonymous MAP_SHARED memory that hasn't
1380 * been ZFOD faulted yet? If we do not test for this, the
1381 * shadow test may succeed! XXX
1383 if (backing_object->type != OBJT_DEFAULT) {
1387 if (op & OBSC_COLLAPSE_WAIT) {
1388 vm_object_set_flag(backing_object, OBJ_DEAD);
1394 p = TAILQ_FIRST(&backing_object->memq);
1396 vm_page_t next = TAILQ_NEXT(p, listq);
1397 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1399 if (op & OBSC_TEST_ALL_SHADOWED) {
1403 * Ignore pages outside the parent object's range
1404 * and outside the parent object's mapping of the
1407 * note that we do not busy the backing object's
1411 p->pindex < backing_offset_index ||
1412 new_pindex >= object->size
1419 * See if the parent has the page or if the parent's
1420 * object pager has the page. If the parent has the
1421 * page but the page is not valid, the parent's
1422 * object pager must have the page.
1424 * If this fails, the parent does not completely shadow
1425 * the object and we might as well give up now.
1428 pp = vm_page_lookup(object, new_pindex);
1430 (pp == NULL || pp->valid == 0) &&
1431 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1439 * Check for busy page
1441 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1444 if (op & OBSC_COLLAPSE_NOWAIT) {
1445 if ((p->oflags & VPO_BUSY) ||
1451 } else if (op & OBSC_COLLAPSE_WAIT) {
1452 if ((p->oflags & VPO_BUSY) || p->busy) {
1453 VM_OBJECT_UNLOCK(object);
1454 p->oflags |= VPO_WANTED;
1455 msleep(p, VM_OBJECT_MTX(backing_object),
1456 PDROP | PVM, "vmocol", 0);
1457 VM_OBJECT_LOCK(object);
1458 VM_OBJECT_LOCK(backing_object);
1460 * If we slept, anything could have
1461 * happened. Since the object is
1462 * marked dead, the backing offset
1463 * should not have changed so we
1464 * just restart our scan.
1466 p = TAILQ_FIRST(&backing_object->memq);
1472 p->object == backing_object,
1473 ("vm_object_backing_scan: object mismatch")
1477 * Destroy any associated swap
1479 if (backing_object->type == OBJT_SWAP) {
1480 swap_pager_freespace(
1488 p->pindex < backing_offset_index ||
1489 new_pindex >= object->size
1492 * Page is out of the parent object's range, we
1493 * can simply destroy it.
1495 vm_page_lock_queues();
1496 KASSERT(!pmap_page_is_mapped(p),
1497 ("freeing mapped page %p", p));
1498 if (p->wire_count == 0)
1502 vm_page_unlock_queues();
1507 pp = vm_page_lookup(object, new_pindex);
1509 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1510 (pp != NULL && pp->valid == 0)
1513 * The page in the parent is not (yet) valid.
1514 * We don't know anything about the state of
1515 * the original page. It might be mapped,
1516 * so we must avoid the next if here.
1518 * This is due to a race in vm_fault() where
1519 * we must unbusy the original (backing_obj)
1520 * page before we can (re)lock the parent.
1521 * Hence we can get here.
1528 vm_pager_has_page(object, new_pindex, NULL, NULL)
1531 * page already exists in parent OR swap exists
1532 * for this location in the parent. Destroy
1533 * the original page from the backing object.
1535 * Leave the parent's page alone
1537 vm_page_lock_queues();
1538 KASSERT(!pmap_page_is_mapped(p),
1539 ("freeing mapped page %p", p));
1540 if (p->wire_count == 0)
1544 vm_page_unlock_queues();
1549 #if VM_NRESERVLEVEL > 0
1551 * Rename the reservation.
1553 vm_reserv_rename(p, object, backing_object,
1554 backing_offset_index);
1558 * Page does not exist in parent, rename the
1559 * page from the backing object to the main object.
1561 * If the page was mapped to a process, it can remain
1562 * mapped through the rename.
1564 vm_page_lock_queues();
1565 vm_page_rename(p, object, new_pindex);
1566 vm_page_unlock_queues();
1567 /* page automatically made dirty by rename */
1576 * this version of collapse allows the operation to occur earlier and
1577 * when paging_in_progress is true for an object... This is not a complete
1578 * operation, but should plug 99.9% of the rest of the leaks.
1581 vm_object_qcollapse(vm_object_t object)
1583 vm_object_t backing_object = object->backing_object;
1585 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1586 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1588 if (backing_object->ref_count != 1)
1591 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1595 * vm_object_collapse:
1597 * Collapse an object with the object backing it.
1598 * Pages in the backing object are moved into the
1599 * parent, and the backing object is deallocated.
1602 vm_object_collapse(vm_object_t object)
1604 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1607 vm_object_t backing_object;
1610 * Verify that the conditions are right for collapse:
1612 * The object exists and the backing object exists.
1614 if ((backing_object = object->backing_object) == NULL)
1618 * we check the backing object first, because it is most likely
1621 VM_OBJECT_LOCK(backing_object);
1622 if (backing_object->handle != NULL ||
1623 (backing_object->type != OBJT_DEFAULT &&
1624 backing_object->type != OBJT_SWAP) ||
1625 (backing_object->flags & OBJ_DEAD) ||
1626 object->handle != NULL ||
1627 (object->type != OBJT_DEFAULT &&
1628 object->type != OBJT_SWAP) ||
1629 (object->flags & OBJ_DEAD)) {
1630 VM_OBJECT_UNLOCK(backing_object);
1635 object->paging_in_progress != 0 ||
1636 backing_object->paging_in_progress != 0
1638 vm_object_qcollapse(object);
1639 VM_OBJECT_UNLOCK(backing_object);
1643 * We know that we can either collapse the backing object (if
1644 * the parent is the only reference to it) or (perhaps) have
1645 * the parent bypass the object if the parent happens to shadow
1646 * all the resident pages in the entire backing object.
1648 * This is ignoring pager-backed pages such as swap pages.
1649 * vm_object_backing_scan fails the shadowing test in this
1652 if (backing_object->ref_count == 1) {
1654 * If there is exactly one reference to the backing
1655 * object, we can collapse it into the parent.
1657 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1659 #if VM_NRESERVLEVEL > 0
1661 * Break any reservations from backing_object.
1663 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1664 vm_reserv_break_all(backing_object);
1668 * Move the pager from backing_object to object.
1670 if (backing_object->type == OBJT_SWAP) {
1672 * swap_pager_copy() can sleep, in which case
1673 * the backing_object's and object's locks are
1674 * released and reacquired.
1679 OFF_TO_IDX(object->backing_object_offset), TRUE);
1682 * Free any cached pages from backing_object.
1684 if (__predict_false(backing_object->cache != NULL))
1685 vm_page_cache_free(backing_object, 0, 0);
1688 * Object now shadows whatever backing_object did.
1689 * Note that the reference to
1690 * backing_object->backing_object moves from within
1691 * backing_object to within object.
1693 LIST_REMOVE(object, shadow_list);
1694 backing_object->shadow_count--;
1695 if (backing_object->backing_object) {
1696 VM_OBJECT_LOCK(backing_object->backing_object);
1697 LIST_REMOVE(backing_object, shadow_list);
1699 &backing_object->backing_object->shadow_head,
1700 object, shadow_list);
1702 * The shadow_count has not changed.
1704 VM_OBJECT_UNLOCK(backing_object->backing_object);
1706 object->backing_object = backing_object->backing_object;
1707 object->backing_object_offset +=
1708 backing_object->backing_object_offset;
1711 * Discard backing_object.
1713 * Since the backing object has no pages, no pager left,
1714 * and no object references within it, all that is
1715 * necessary is to dispose of it.
1717 KASSERT(backing_object->ref_count == 1, (
1718 "backing_object %p was somehow re-referenced during collapse!",
1720 VM_OBJECT_UNLOCK(backing_object);
1721 vm_object_destroy(backing_object);
1725 vm_object_t new_backing_object;
1728 * If we do not entirely shadow the backing object,
1729 * there is nothing we can do so we give up.
1731 if (object->resident_page_count != object->size &&
1732 vm_object_backing_scan(object,
1733 OBSC_TEST_ALL_SHADOWED) == 0) {
1734 VM_OBJECT_UNLOCK(backing_object);
1739 * Make the parent shadow the next object in the
1740 * chain. Deallocating backing_object will not remove
1741 * it, since its reference count is at least 2.
1743 LIST_REMOVE(object, shadow_list);
1744 backing_object->shadow_count--;
1746 new_backing_object = backing_object->backing_object;
1747 if ((object->backing_object = new_backing_object) != NULL) {
1748 VM_OBJECT_LOCK(new_backing_object);
1750 &new_backing_object->shadow_head,
1754 new_backing_object->shadow_count++;
1755 vm_object_reference_locked(new_backing_object);
1756 VM_OBJECT_UNLOCK(new_backing_object);
1757 object->backing_object_offset +=
1758 backing_object->backing_object_offset;
1762 * Drop the reference count on backing_object. Since
1763 * its ref_count was at least 2, it will not vanish.
1765 backing_object->ref_count--;
1766 VM_OBJECT_UNLOCK(backing_object);
1771 * Try again with this object's new backing object.
1777 * vm_object_page_remove:
1779 * For the given object, either frees or invalidates each of the
1780 * specified pages. In general, a page is freed. However, if a
1781 * page is wired for any reason other than the existence of a
1782 * managed, wired mapping, then it may be invalidated but not
1783 * removed from the object. Pages are specified by the given
1784 * range ["start", "end") and Boolean "clean_only". As a
1785 * special case, if "end" is zero, then the range extends from
1786 * "start" to the end of the object. If "clean_only" is TRUE,
1787 * then only the non-dirty pages within the specified range are
1790 * In general, this operation should only be performed on objects
1791 * that contain managed pages. There are two exceptions. First,
1792 * it may be performed on the kernel and kmem objects. Second,
1793 * it may be used by msync(..., MS_INVALIDATE) to invalidate
1794 * device-backed pages. In both of these cases, "clean_only"
1797 * The object must be locked.
1800 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1801 boolean_t clean_only)
1806 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1807 if (object->resident_page_count == 0)
1811 * Since physically-backed objects do not use managed pages, we can't
1812 * remove pages from the object (we must instead remove the page
1813 * references, and then destroy the object).
1815 KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1816 object == kmem_object,
1817 ("attempt to remove pages from a physical object"));
1819 vm_object_pip_add(object, 1);
1821 p = vm_page_find_least(object, start);
1822 vm_page_lock_queues();
1824 * Assert: the variable p is either (1) the page with the
1825 * least pindex greater than or equal to the parameter pindex
1829 p != NULL && (p->pindex < end || end == 0);
1831 next = TAILQ_NEXT(p, listq);
1834 * If the page is wired for any reason besides the
1835 * existence of managed, wired mappings, then it cannot
1836 * be freed. For example, fictitious pages, which
1837 * represent device memory, are inherently wired and
1838 * cannot be freed. They can, however, be invalidated
1839 * if "clean_only" is FALSE.
1841 if ((wirings = p->wire_count) != 0 &&
1842 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1843 /* Fictitious pages do not have managed mappings. */
1844 if ((p->flags & PG_FICTITIOUS) == 0)
1846 /* Account for removal of managed, wired mappings. */
1847 p->wire_count -= wirings;
1854 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1856 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1857 ("vm_object_page_remove: page %p is fictitious", p));
1858 if (clean_only && p->valid) {
1859 pmap_remove_write(p);
1864 /* Account for removal of managed, wired mappings. */
1866 p->wire_count -= wirings;
1869 vm_page_unlock_queues();
1870 vm_object_pip_wakeup(object);
1872 if (__predict_false(object->cache != NULL))
1873 vm_page_cache_free(object, start, end);
1877 * Populate the specified range of the object with valid pages. Returns
1878 * TRUE if the range is successfully populated and FALSE otherwise.
1880 * Note: This function should be optimized to pass a larger array of
1881 * pages to vm_pager_get_pages() before it is applied to a non-
1882 * OBJT_DEVICE object.
1884 * The object must be locked.
1887 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1893 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1894 for (pindex = start; pindex < end; pindex++) {
1895 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1897 if (m->valid != VM_PAGE_BITS_ALL) {
1899 rv = vm_pager_get_pages(object, ma, 1, 0);
1900 m = vm_page_lookup(object, pindex);
1903 if (rv != VM_PAGER_OK) {
1904 vm_page_lock_queues();
1906 vm_page_unlock_queues();
1911 * Keep "m" busy because a subsequent iteration may unlock
1915 if (pindex > start) {
1916 m = vm_page_lookup(object, start);
1917 while (m != NULL && m->pindex < pindex) {
1919 m = TAILQ_NEXT(m, listq);
1922 return (pindex == end);
1926 * Routine: vm_object_coalesce
1927 * Function: Coalesces two objects backing up adjoining
1928 * regions of memory into a single object.
1930 * returns TRUE if objects were combined.
1932 * NOTE: Only works at the moment if the second object is NULL -
1933 * if it's not, which object do we lock first?
1936 * prev_object First object to coalesce
1937 * prev_offset Offset into prev_object
1938 * prev_size Size of reference to prev_object
1939 * next_size Size of reference to the second object
1940 * reserved Indicator that extension region has
1941 * swap accounted for
1944 * The object must *not* be locked.
1947 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
1948 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
1950 vm_pindex_t next_pindex;
1952 if (prev_object == NULL)
1954 VM_OBJECT_LOCK(prev_object);
1955 if (prev_object->type != OBJT_DEFAULT &&
1956 prev_object->type != OBJT_SWAP) {
1957 VM_OBJECT_UNLOCK(prev_object);
1962 * Try to collapse the object first
1964 vm_object_collapse(prev_object);
1967 * Can't coalesce if: . more than one reference . paged out . shadows
1968 * another object . has a copy elsewhere (any of which mean that the
1969 * pages not mapped to prev_entry may be in use anyway)
1971 if (prev_object->backing_object != NULL) {
1972 VM_OBJECT_UNLOCK(prev_object);
1976 prev_size >>= PAGE_SHIFT;
1977 next_size >>= PAGE_SHIFT;
1978 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
1980 if ((prev_object->ref_count > 1) &&
1981 (prev_object->size != next_pindex)) {
1982 VM_OBJECT_UNLOCK(prev_object);
1987 * Account for the charge.
1989 if (prev_object->uip != NULL) {
1992 * If prev_object was charged, then this mapping,
1993 * althought not charged now, may become writable
1994 * later. Non-NULL uip in the object would prevent
1995 * swap reservation during enabling of the write
1996 * access, so reserve swap now. Failed reservation
1997 * cause allocation of the separate object for the map
1998 * entry, and swap reservation for this entry is
1999 * managed in appropriate time.
2001 if (!reserved && !swap_reserve_by_uid(ptoa(next_size),
2002 prev_object->uip)) {
2005 prev_object->charge += ptoa(next_size);
2009 * Remove any pages that may still be in the object from a previous
2012 if (next_pindex < prev_object->size) {
2013 vm_object_page_remove(prev_object,
2015 next_pindex + next_size, FALSE);
2016 if (prev_object->type == OBJT_SWAP)
2017 swap_pager_freespace(prev_object,
2018 next_pindex, next_size);
2020 if (prev_object->uip != NULL) {
2021 KASSERT(prev_object->charge >=
2022 ptoa(prev_object->size - next_pindex),
2023 ("object %p overcharged 1 %jx %jx", prev_object,
2024 (uintmax_t)next_pindex, (uintmax_t)next_size));
2025 prev_object->charge -= ptoa(prev_object->size -
2032 * Extend the object if necessary.
2034 if (next_pindex + next_size > prev_object->size)
2035 prev_object->size = next_pindex + next_size;
2037 VM_OBJECT_UNLOCK(prev_object);
2042 vm_object_set_writeable_dirty(vm_object_t object)
2045 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2046 if (object->type != OBJT_VNODE)
2048 object->generation++;
2049 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2051 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2054 #include "opt_ddb.h"
2056 #include <sys/kernel.h>
2058 #include <sys/cons.h>
2060 #include <ddb/ddb.h>
2063 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2066 vm_map_entry_t tmpe;
2074 tmpe = map->header.next;
2075 entcount = map->nentries;
2076 while (entcount-- && (tmpe != &map->header)) {
2077 if (_vm_object_in_map(map, object, tmpe)) {
2082 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2083 tmpm = entry->object.sub_map;
2084 tmpe = tmpm->header.next;
2085 entcount = tmpm->nentries;
2086 while (entcount-- && tmpe != &tmpm->header) {
2087 if (_vm_object_in_map(tmpm, object, tmpe)) {
2092 } else if ((obj = entry->object.vm_object) != NULL) {
2093 for (; obj; obj = obj->backing_object)
2094 if (obj == object) {
2102 vm_object_in_map(vm_object_t object)
2106 /* sx_slock(&allproc_lock); */
2107 FOREACH_PROC_IN_SYSTEM(p) {
2108 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2110 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2111 /* sx_sunlock(&allproc_lock); */
2115 /* sx_sunlock(&allproc_lock); */
2116 if (_vm_object_in_map(kernel_map, object, 0))
2118 if (_vm_object_in_map(kmem_map, object, 0))
2120 if (_vm_object_in_map(pager_map, object, 0))
2122 if (_vm_object_in_map(buffer_map, object, 0))
2127 DB_SHOW_COMMAND(vmochk, vm_object_check)
2132 * make sure that internal objs are in a map somewhere
2133 * and none have zero ref counts.
2135 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2136 if (object->handle == NULL &&
2137 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2138 if (object->ref_count == 0) {
2139 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2140 (long)object->size);
2142 if (!vm_object_in_map(object)) {
2144 "vmochk: internal obj is not in a map: "
2145 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2146 object->ref_count, (u_long)object->size,
2147 (u_long)object->size,
2148 (void *)object->backing_object);
2155 * vm_object_print: [ debug ]
2157 DB_SHOW_COMMAND(object, vm_object_print_static)
2159 /* XXX convert args. */
2160 vm_object_t object = (vm_object_t)addr;
2161 boolean_t full = have_addr;
2165 /* XXX count is an (unused) arg. Avoid shadowing it. */
2166 #define count was_count
2174 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x uip %d charge %jx\n",
2175 object, (int)object->type, (uintmax_t)object->size,
2176 object->resident_page_count, object->ref_count, object->flags,
2177 object->uip ? object->uip->ui_uid : -1, (uintmax_t)object->charge);
2178 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2179 object->shadow_count,
2180 object->backing_object ? object->backing_object->ref_count : 0,
2181 object->backing_object, (uintmax_t)object->backing_object_offset);
2188 TAILQ_FOREACH(p, &object->memq, listq) {
2190 db_iprintf("memory:=");
2191 else if (count == 6) {
2199 db_printf("(off=0x%jx,page=0x%jx)",
2200 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2210 /* XXX need this non-static entry for calling from vm_map_print. */
2213 /* db_expr_t */ long addr,
2214 boolean_t have_addr,
2215 /* db_expr_t */ long count,
2218 vm_object_print_static(addr, have_addr, count, modif);
2221 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2226 vm_page_t m, prev_m;
2230 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2231 db_printf("new object: %p\n", (void *)object);
2242 TAILQ_FOREACH(m, &object->memq, listq) {
2243 if (m->pindex > 128)
2245 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2246 prev_m->pindex + 1 != m->pindex) {
2248 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2249 (long)fidx, rcount, (long)pa);
2261 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2266 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2267 (long)fidx, rcount, (long)pa);
2277 pa = VM_PAGE_TO_PHYS(m);
2281 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2282 (long)fidx, rcount, (long)pa);