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, boolean_t *clearobjflags,
110 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
111 boolean_t *clearobjflags);
112 static void vm_object_qcollapse(vm_object_t object);
113 static void vm_object_vndeallocate(vm_object_t object);
116 * Virtual memory objects maintain the actual data
117 * associated with allocated virtual memory. A given
118 * page of memory exists within exactly one object.
120 * An object is only deallocated when all "references"
121 * are given up. Only one "reference" to a given
122 * region of an object should be writeable.
124 * Associated with each object is a list of all resident
125 * memory pages belonging to that object; this list is
126 * maintained by the "vm_page" module, and locked by the object's
129 * Each object also records a "pager" routine which is
130 * used to retrieve (and store) pages to the proper backing
131 * storage. In addition, objects may be backed by other
132 * objects from which they were virtual-copied.
134 * The only items within the object structure which are
135 * modified after time of creation are:
136 * reference count locked by object's lock
137 * pager routine locked by object's lock
141 struct object_q vm_object_list;
142 struct mtx vm_object_list_mtx; /* lock for object list and count */
144 struct vm_object kernel_object_store;
145 struct vm_object kmem_object_store;
147 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats");
149 static long object_collapses;
150 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
151 &object_collapses, 0, "VM object collapses");
153 static long object_bypasses;
154 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
155 &object_bypasses, 0, "VM object bypasses");
157 static uma_zone_t obj_zone;
159 static int vm_object_zinit(void *mem, int size, int flags);
162 static void vm_object_zdtor(void *mem, int size, void *arg);
165 vm_object_zdtor(void *mem, int size, void *arg)
169 object = (vm_object_t)mem;
170 KASSERT(TAILQ_EMPTY(&object->memq),
171 ("object %p has resident pages",
173 #if VM_NRESERVLEVEL > 0
174 KASSERT(LIST_EMPTY(&object->rvq),
175 ("object %p has reservations",
178 KASSERT(object->cache == NULL,
179 ("object %p has cached pages",
181 KASSERT(object->paging_in_progress == 0,
182 ("object %p paging_in_progress = %d",
183 object, object->paging_in_progress));
184 KASSERT(object->resident_page_count == 0,
185 ("object %p resident_page_count = %d",
186 object, object->resident_page_count));
187 KASSERT(object->shadow_count == 0,
188 ("object %p shadow_count = %d",
189 object, object->shadow_count));
194 vm_object_zinit(void *mem, int size, int flags)
198 object = (vm_object_t)mem;
199 bzero(&object->mtx, sizeof(object->mtx));
200 VM_OBJECT_LOCK_INIT(object, "standard object");
202 /* These are true for any object that has been freed */
203 object->paging_in_progress = 0;
204 object->resident_page_count = 0;
205 object->shadow_count = 0;
210 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
213 TAILQ_INIT(&object->memq);
214 LIST_INIT(&object->shadow_head);
219 object->generation = 1;
220 object->ref_count = 1;
221 object->memattr = VM_MEMATTR_DEFAULT;
225 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
226 object->flags = OBJ_ONEMAPPING;
227 object->pg_color = 0;
228 object->handle = NULL;
229 object->backing_object = NULL;
230 object->backing_object_offset = (vm_ooffset_t) 0;
231 #if VM_NRESERVLEVEL > 0
232 LIST_INIT(&object->rvq);
234 object->cache = NULL;
236 mtx_lock(&vm_object_list_mtx);
237 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
238 mtx_unlock(&vm_object_list_mtx);
244 * Initialize the VM objects module.
249 TAILQ_INIT(&vm_object_list);
250 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
252 VM_OBJECT_LOCK_INIT(kernel_object, "kernel object");
253 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
255 #if VM_NRESERVLEVEL > 0
256 kernel_object->flags |= OBJ_COLORED;
257 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
260 VM_OBJECT_LOCK_INIT(kmem_object, "kmem object");
261 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
263 #if VM_NRESERVLEVEL > 0
264 kmem_object->flags |= OBJ_COLORED;
265 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
269 * The lock portion of struct vm_object must be type stable due
270 * to vm_pageout_fallback_object_lock locking a vm object
271 * without holding any references to it.
273 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
279 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
283 vm_object_clear_flag(vm_object_t object, u_short bits)
286 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
287 object->flags &= ~bits;
291 * Sets the default memory attribute for the specified object. Pages
292 * that are allocated to this object are by default assigned this memory
295 * Presently, this function must be called before any pages are allocated
296 * to the object. In the future, this requirement may be relaxed for
297 * "default" and "swap" objects.
300 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
303 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
304 switch (object->type) {
311 if (!TAILQ_EMPTY(&object->memq))
312 return (KERN_FAILURE);
315 return (KERN_INVALID_ARGUMENT);
317 object->memattr = memattr;
318 return (KERN_SUCCESS);
322 vm_object_pip_add(vm_object_t object, short i)
325 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
326 object->paging_in_progress += i;
330 vm_object_pip_subtract(vm_object_t object, short i)
333 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
334 object->paging_in_progress -= i;
338 vm_object_pip_wakeup(vm_object_t object)
341 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
342 object->paging_in_progress--;
343 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
344 vm_object_clear_flag(object, OBJ_PIPWNT);
350 vm_object_pip_wakeupn(vm_object_t object, short i)
353 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
355 object->paging_in_progress -= i;
356 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
357 vm_object_clear_flag(object, OBJ_PIPWNT);
363 vm_object_pip_wait(vm_object_t object, char *waitid)
366 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
367 while (object->paging_in_progress) {
368 object->flags |= OBJ_PIPWNT;
369 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
374 * vm_object_allocate:
376 * Returns a new object with the given size.
379 vm_object_allocate(objtype_t type, vm_pindex_t size)
383 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
384 _vm_object_allocate(type, size, object);
390 * vm_object_reference:
392 * Gets another reference to the given object. Note: OBJ_DEAD
393 * objects can be referenced during final cleaning.
396 vm_object_reference(vm_object_t object)
400 VM_OBJECT_LOCK(object);
401 vm_object_reference_locked(object);
402 VM_OBJECT_UNLOCK(object);
406 * vm_object_reference_locked:
408 * Gets another reference to the given object.
410 * The object must be locked.
413 vm_object_reference_locked(vm_object_t object)
417 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
419 if (object->type == OBJT_VNODE) {
426 * Handle deallocating an object of type OBJT_VNODE.
429 vm_object_vndeallocate(vm_object_t object)
431 struct vnode *vp = (struct vnode *) object->handle;
433 VFS_ASSERT_GIANT(vp->v_mount);
434 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
435 KASSERT(object->type == OBJT_VNODE,
436 ("vm_object_vndeallocate: not a vnode object"));
437 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
439 if (object->ref_count == 0) {
440 vprint("vm_object_vndeallocate", vp);
441 panic("vm_object_vndeallocate: bad object reference count");
445 if (object->ref_count > 1) {
447 VM_OBJECT_UNLOCK(object);
448 /* vrele may need the vnode lock. */
452 VM_OBJECT_UNLOCK(object);
453 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
455 VM_OBJECT_LOCK(object);
457 if (object->type == OBJT_DEAD) {
458 VM_OBJECT_UNLOCK(object);
461 if (object->ref_count == 0)
462 vp->v_vflag &= ~VV_TEXT;
463 VM_OBJECT_UNLOCK(object);
470 * vm_object_deallocate:
472 * Release a reference to the specified object,
473 * gained either through a vm_object_allocate
474 * or a vm_object_reference call. When all references
475 * are gone, storage associated with this object
476 * may be relinquished.
478 * No object may be locked.
481 vm_object_deallocate(vm_object_t object)
485 while (object != NULL) {
490 VM_OBJECT_LOCK(object);
491 if (object->type == OBJT_VNODE) {
492 struct vnode *vp = (struct vnode *) object->handle;
495 * Conditionally acquire Giant for a vnode-backed
496 * object. We have to be careful since the type of
497 * a vnode object can change while the object is
500 if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
502 if (!mtx_trylock(&Giant)) {
503 VM_OBJECT_UNLOCK(object);
508 vm_object_vndeallocate(object);
509 VFS_UNLOCK_GIANT(vfslocked);
513 * This is to handle the case that the object
514 * changed type while we dropped its lock to
517 VFS_UNLOCK_GIANT(vfslocked);
519 KASSERT(object->ref_count != 0,
520 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
523 * If the reference count goes to 0 we start calling
524 * vm_object_terminate() on the object chain.
525 * A ref count of 1 may be a special case depending on the
526 * shadow count being 0 or 1.
529 if (object->ref_count > 1) {
530 VM_OBJECT_UNLOCK(object);
532 } else if (object->ref_count == 1) {
533 if (object->shadow_count == 0 &&
534 object->handle == NULL &&
535 (object->type == OBJT_DEFAULT ||
536 object->type == OBJT_SWAP)) {
537 vm_object_set_flag(object, OBJ_ONEMAPPING);
538 } else if ((object->shadow_count == 1) &&
539 (object->handle == NULL) &&
540 (object->type == OBJT_DEFAULT ||
541 object->type == OBJT_SWAP)) {
544 robject = LIST_FIRST(&object->shadow_head);
545 KASSERT(robject != NULL,
546 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
548 object->shadow_count));
549 if (!VM_OBJECT_TRYLOCK(robject)) {
551 * Avoid a potential deadlock.
554 VM_OBJECT_UNLOCK(object);
556 * More likely than not the thread
557 * holding robject's lock has lower
558 * priority than the current thread.
559 * Let the lower priority thread run.
565 * Collapse object into its shadow unless its
566 * shadow is dead. In that case, object will
567 * be deallocated by the thread that is
568 * deallocating its shadow.
570 if ((robject->flags & OBJ_DEAD) == 0 &&
571 (robject->handle == NULL) &&
572 (robject->type == OBJT_DEFAULT ||
573 robject->type == OBJT_SWAP)) {
575 robject->ref_count++;
577 if (robject->paging_in_progress) {
578 VM_OBJECT_UNLOCK(object);
579 vm_object_pip_wait(robject,
581 temp = robject->backing_object;
582 if (object == temp) {
583 VM_OBJECT_LOCK(object);
586 } else if (object->paging_in_progress) {
587 VM_OBJECT_UNLOCK(robject);
588 object->flags |= OBJ_PIPWNT;
590 VM_OBJECT_MTX(object),
591 PDROP | PVM, "objde2", 0);
592 VM_OBJECT_LOCK(robject);
593 temp = robject->backing_object;
594 if (object == temp) {
595 VM_OBJECT_LOCK(object);
599 VM_OBJECT_UNLOCK(object);
601 if (robject->ref_count == 1) {
602 robject->ref_count--;
607 vm_object_collapse(object);
608 VM_OBJECT_UNLOCK(object);
611 VM_OBJECT_UNLOCK(robject);
613 VM_OBJECT_UNLOCK(object);
617 temp = object->backing_object;
619 VM_OBJECT_LOCK(temp);
620 LIST_REMOVE(object, shadow_list);
621 temp->shadow_count--;
622 VM_OBJECT_UNLOCK(temp);
623 object->backing_object = NULL;
626 * Don't double-terminate, we could be in a termination
627 * recursion due to the terminate having to sync data
630 if ((object->flags & OBJ_DEAD) == 0)
631 vm_object_terminate(object);
633 VM_OBJECT_UNLOCK(object);
639 * vm_object_destroy removes the object from the global object list
640 * and frees the space for the object.
643 vm_object_destroy(vm_object_t object)
647 * Remove the object from the global object list.
649 mtx_lock(&vm_object_list_mtx);
650 TAILQ_REMOVE(&vm_object_list, object, object_list);
651 mtx_unlock(&vm_object_list_mtx);
654 * Release the allocation charge.
656 if (object->uip != NULL) {
657 KASSERT(object->type == OBJT_DEFAULT ||
658 object->type == OBJT_SWAP,
659 ("vm_object_terminate: non-swap obj %p has uip",
661 swap_release_by_uid(object->charge, object->uip);
668 * Free the space for the object.
670 uma_zfree(obj_zone, object);
674 * vm_object_terminate actually destroys the specified object, freeing
675 * up all previously used resources.
677 * The object must be locked.
678 * This routine may block.
681 vm_object_terminate(vm_object_t object)
685 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
688 * Make sure no one uses us.
690 vm_object_set_flag(object, OBJ_DEAD);
693 * wait for the pageout daemon to be done with the object
695 vm_object_pip_wait(object, "objtrm");
697 KASSERT(!object->paging_in_progress,
698 ("vm_object_terminate: pageout in progress"));
701 * Clean and free the pages, as appropriate. All references to the
702 * object are gone, so we don't need to lock it.
704 if (object->type == OBJT_VNODE) {
705 struct vnode *vp = (struct vnode *)object->handle;
708 * Clean pages and flush buffers.
710 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
711 VM_OBJECT_UNLOCK(object);
713 vinvalbuf(vp, V_SAVE, 0, 0);
715 VM_OBJECT_LOCK(object);
718 KASSERT(object->ref_count == 0,
719 ("vm_object_terminate: object with references, ref_count=%d",
723 * Free any remaining pageable pages. This also removes them from the
724 * paging queues. However, don't free wired pages, just remove them
725 * from the object. Rather than incrementally removing each page from
726 * the object, the page and object are reset to any empty state.
728 vm_page_lock_queues();
729 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
730 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
731 ("vm_object_terminate: freeing busy page %p", p));
733 * Optimize the page's removal from the object by resetting
734 * its "object" field. Specifically, if the page is not
735 * wired, then the effect of this assignment is that
736 * vm_page_free()'s call to vm_page_remove() will return
737 * immediately without modifying the page or the object.
740 if (p->wire_count == 0) {
745 vm_page_unlock_queues();
747 * If the object contained any pages, then reset it to an empty state.
748 * None of the object's fields, including "resident_page_count", were
749 * modified by the preceding loop.
751 if (object->resident_page_count != 0) {
753 TAILQ_INIT(&object->memq);
754 object->resident_page_count = 0;
755 if (object->type == OBJT_VNODE)
756 vdrop(object->handle);
759 #if VM_NRESERVLEVEL > 0
760 if (__predict_false(!LIST_EMPTY(&object->rvq)))
761 vm_reserv_break_all(object);
763 if (__predict_false(object->cache != NULL))
764 vm_page_cache_free(object, 0, 0);
767 * Let the pager know object is dead.
769 vm_pager_deallocate(object);
770 VM_OBJECT_UNLOCK(object);
772 vm_object_destroy(object);
776 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
780 * If we have been asked to skip nosync pages and this is a
781 * nosync page, skip it. Note that the object flags were not
782 * cleared in this case so we do not have to set them.
784 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
785 *clearobjflags = FALSE;
788 pmap_remove_write(p);
789 return (p->dirty != 0);
794 * vm_object_page_clean
796 * Clean all dirty pages in the specified range of object. Leaves page
797 * on whatever queue it is currently on. If NOSYNC is set then do not
798 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
799 * leaving the object dirty.
801 * When stuffing pages asynchronously, allow clustering. XXX we need a
802 * synchronous clustering mode implementation.
804 * Odd semantics: if start == end, we clean everything.
806 * The object must be locked.
808 * Returns FALSE if some page from the range was not written, as
809 * reported by the pager, and TRUE otherwise.
812 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
816 vm_pindex_t pi, tend;
817 int curgeneration, n, pagerflags;
818 boolean_t clearobjflags, eio, res;
820 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
821 KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
822 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
823 object->resident_page_count == 0)
826 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
827 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
828 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
830 tend = (end == 0) ? object->size : end;
832 vm_object_set_flag(object, OBJ_CLEANING);
834 vm_page_lock_queues();
837 * Make the page read-only so we can then clear the object flags.
839 * However, if this is a nosync mmap then the object is likely to
840 * stay dirty so do not mess with the page and do not clear the
843 clearobjflags = TRUE;
847 curgeneration = object->generation;
849 for (p = vm_page_find_least(object, start); p != NULL; p = np) {
853 np = TAILQ_NEXT(p, listq);
856 if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
857 vm_page_lock_queues();
858 if (object->generation != curgeneration)
860 np = vm_page_find_least(object, pi);
863 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
866 n = vm_object_page_collect_flush(object, p, pagerflags,
867 flags, &clearobjflags, &eio);
870 clearobjflags = FALSE;
872 if (object->generation != curgeneration)
876 * If the VOP_PUTPAGES() did a truncated write, so
877 * that even the first page of the run is not fully
878 * written, vm_pageout_flush() returns 0 as the run
879 * length. Since the condition that caused truncated
880 * write may be permanent, e.g. exhausted free space,
881 * accepting n == 0 would cause an infinite loop.
883 * Forwarding the iterator leaves the unwritten page
884 * behind, but there is not much we can do there if
885 * filesystem refuses to write it.
889 clearobjflags = FALSE;
891 np = vm_page_find_least(object, pi + n);
893 vm_page_unlock_queues();
895 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
898 vm_object_clear_flag(object, OBJ_CLEANING);
899 if (clearobjflags && start == 0 && tend == object->size)
900 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
905 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
906 int flags, boolean_t *clearobjflags, boolean_t *eio)
908 vm_page_t ma[vm_pageout_page_count], p_first, tp;
909 int count, i, mreq, runlen;
911 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
916 for (tp = p; count < vm_pageout_page_count; count++) {
917 tp = vm_page_next(tp);
918 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
920 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
924 for (p_first = p; count < vm_pageout_page_count; count++) {
925 tp = vm_page_prev(p_first);
926 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
928 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
934 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
937 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
942 * Note that there is absolutely no sense in writing out
943 * anonymous objects, so we track down the vnode object
945 * We invalidate (remove) all pages from the address space
946 * for semantic correctness.
948 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
949 * may start out with a NULL object.
952 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
953 boolean_t syncio, boolean_t invalidate)
955 vm_object_t backing_object;
958 int error, flags, fsync_after;
965 VM_OBJECT_LOCK(object);
966 while ((backing_object = object->backing_object) != NULL) {
967 VM_OBJECT_LOCK(backing_object);
968 offset += object->backing_object_offset;
969 VM_OBJECT_UNLOCK(object);
970 object = backing_object;
971 if (object->size < OFF_TO_IDX(offset + size))
972 size = IDX_TO_OFF(object->size) - offset;
975 * Flush pages if writing is allowed, invalidate them
976 * if invalidation requested. Pages undergoing I/O
977 * will be ignored by vm_object_page_remove().
979 * We cannot lock the vnode and then wait for paging
980 * to complete without deadlocking against vm_fault.
981 * Instead we simply call vm_object_page_remove() and
982 * allow it to block internally on a page-by-page
983 * basis when it encounters pages undergoing async
986 if (object->type == OBJT_VNODE &&
987 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
990 VM_OBJECT_UNLOCK(object);
991 (void) vn_start_write(vp, &mp, V_WAIT);
992 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
993 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
994 if (syncio && !invalidate && offset == 0 &&
995 OFF_TO_IDX(size) == object->size) {
997 * If syncing the whole mapping of the file,
998 * it is faster to schedule all the writes in
999 * async mode, also allowing the clustering,
1000 * and then wait for i/o to complete.
1005 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1006 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1007 fsync_after = FALSE;
1009 VM_OBJECT_LOCK(object);
1010 res = vm_object_page_clean(object,
1012 OFF_TO_IDX(offset + size + PAGE_MASK),
1014 VM_OBJECT_UNLOCK(object);
1016 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1018 VFS_UNLOCK_GIANT(vfslocked);
1019 vn_finished_write(mp);
1022 VM_OBJECT_LOCK(object);
1024 if ((object->type == OBJT_VNODE ||
1025 object->type == OBJT_DEVICE) && invalidate) {
1027 purge = old_msync || (object->type == OBJT_DEVICE);
1028 vm_object_page_remove(object,
1030 OFF_TO_IDX(offset + size + PAGE_MASK),
1031 purge ? FALSE : TRUE);
1033 VM_OBJECT_UNLOCK(object);
1038 * vm_object_madvise:
1040 * Implements the madvise function at the object/page level.
1042 * MADV_WILLNEED (any object)
1044 * Activate the specified pages if they are resident.
1046 * MADV_DONTNEED (any object)
1048 * Deactivate the specified pages if they are resident.
1050 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1051 * OBJ_ONEMAPPING only)
1053 * Deactivate and clean the specified pages if they are
1054 * resident. This permits the process to reuse the pages
1055 * without faulting or the kernel to reclaim the pages
1059 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1062 vm_pindex_t tpindex;
1063 vm_object_t backing_object, tobject;
1068 VM_OBJECT_LOCK(object);
1070 * Locate and adjust resident pages
1072 for (; pindex < end; pindex += 1) {
1078 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1079 * and those pages must be OBJ_ONEMAPPING.
1081 if (advise == MADV_FREE) {
1082 if ((tobject->type != OBJT_DEFAULT &&
1083 tobject->type != OBJT_SWAP) ||
1084 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1085 goto unlock_tobject;
1087 } else if (tobject->type == OBJT_PHYS)
1088 goto unlock_tobject;
1089 m = vm_page_lookup(tobject, tpindex);
1090 if (m == NULL && advise == MADV_WILLNEED) {
1092 * If the page is cached, reactivate it.
1094 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1099 * There may be swap even if there is no backing page
1101 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1102 swap_pager_freespace(tobject, tpindex, 1);
1106 backing_object = tobject->backing_object;
1107 if (backing_object == NULL)
1108 goto unlock_tobject;
1109 VM_OBJECT_LOCK(backing_object);
1110 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1111 if (tobject != object)
1112 VM_OBJECT_UNLOCK(tobject);
1113 tobject = backing_object;
1115 } else if (m->valid != VM_PAGE_BITS_ALL)
1116 goto unlock_tobject;
1118 * If the page is not in a normal state, skip it.
1120 vm_page_lock_queues();
1121 if (m->hold_count != 0 || m->wire_count != 0) {
1122 vm_page_unlock_queues();
1123 goto unlock_tobject;
1125 if ((m->oflags & VPO_BUSY) || m->busy) {
1126 if (advise == MADV_WILLNEED)
1128 * Reference the page before unlocking and
1129 * sleeping so that the page daemon is less
1130 * likely to reclaim it.
1132 vm_page_flag_set(m, PG_REFERENCED);
1133 vm_page_unlock_queues();
1134 if (object != tobject)
1135 VM_OBJECT_UNLOCK(object);
1136 m->oflags |= VPO_WANTED;
1137 msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1139 VM_OBJECT_LOCK(object);
1142 if (advise == MADV_WILLNEED) {
1143 vm_page_activate(m);
1144 } else if (advise == MADV_DONTNEED) {
1145 vm_page_dontneed(m);
1146 } else if (advise == MADV_FREE) {
1148 * Mark the page clean. This will allow the page
1149 * to be freed up by the system. However, such pages
1150 * are often reused quickly by malloc()/free()
1151 * so we do not do anything that would cause
1152 * a page fault if we can help it.
1154 * Specifically, we do not try to actually free
1155 * the page now nor do we try to put it in the
1156 * cache (which would cause a page fault on reuse).
1158 * But we do make the page is freeable as we
1159 * can without actually taking the step of unmapping
1162 pmap_clear_modify(m);
1165 vm_page_dontneed(m);
1167 vm_page_unlock_queues();
1168 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1169 swap_pager_freespace(tobject, tpindex, 1);
1171 if (tobject != object)
1172 VM_OBJECT_UNLOCK(tobject);
1174 VM_OBJECT_UNLOCK(object);
1180 * Create a new object which is backed by the
1181 * specified existing object range. The source
1182 * object reference is deallocated.
1184 * The new object and offset into that object
1185 * are returned in the source parameters.
1189 vm_object_t *object, /* IN/OUT */
1190 vm_ooffset_t *offset, /* IN/OUT */
1199 * Don't create the new object if the old object isn't shared.
1201 if (source != NULL) {
1202 VM_OBJECT_LOCK(source);
1203 if (source->ref_count == 1 &&
1204 source->handle == NULL &&
1205 (source->type == OBJT_DEFAULT ||
1206 source->type == OBJT_SWAP)) {
1207 VM_OBJECT_UNLOCK(source);
1210 VM_OBJECT_UNLOCK(source);
1214 * Allocate a new object with the given length.
1216 result = vm_object_allocate(OBJT_DEFAULT, length);
1219 * The new object shadows the source object, adding a reference to it.
1220 * Our caller changes his reference to point to the new object,
1221 * removing a reference to the source object. Net result: no change
1222 * of reference count.
1224 * Try to optimize the result object's page color when shadowing
1225 * in order to maintain page coloring consistency in the combined
1228 result->backing_object = source;
1230 * Store the offset into the source object, and fix up the offset into
1233 result->backing_object_offset = *offset;
1234 if (source != NULL) {
1235 VM_OBJECT_LOCK(source);
1236 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1237 source->shadow_count++;
1238 #if VM_NRESERVLEVEL > 0
1239 result->flags |= source->flags & OBJ_COLORED;
1240 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1241 ((1 << (VM_NFREEORDER - 1)) - 1);
1243 VM_OBJECT_UNLOCK(source);
1248 * Return the new things
1257 * Split the pages in a map entry into a new object. This affords
1258 * easier removal of unused pages, and keeps object inheritance from
1259 * being a negative impact on memory usage.
1262 vm_object_split(vm_map_entry_t entry)
1264 vm_page_t m, m_next;
1265 vm_object_t orig_object, new_object, source;
1266 vm_pindex_t idx, offidxstart;
1269 orig_object = entry->object.vm_object;
1270 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1272 if (orig_object->ref_count <= 1)
1274 VM_OBJECT_UNLOCK(orig_object);
1276 offidxstart = OFF_TO_IDX(entry->offset);
1277 size = atop(entry->end - entry->start);
1280 * If swap_pager_copy() is later called, it will convert new_object
1281 * into a swap object.
1283 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1286 * At this point, the new object is still private, so the order in
1287 * which the original and new objects are locked does not matter.
1289 VM_OBJECT_LOCK(new_object);
1290 VM_OBJECT_LOCK(orig_object);
1291 source = orig_object->backing_object;
1292 if (source != NULL) {
1293 VM_OBJECT_LOCK(source);
1294 if ((source->flags & OBJ_DEAD) != 0) {
1295 VM_OBJECT_UNLOCK(source);
1296 VM_OBJECT_UNLOCK(orig_object);
1297 VM_OBJECT_UNLOCK(new_object);
1298 vm_object_deallocate(new_object);
1299 VM_OBJECT_LOCK(orig_object);
1302 LIST_INSERT_HEAD(&source->shadow_head,
1303 new_object, shadow_list);
1304 source->shadow_count++;
1305 vm_object_reference_locked(source); /* for new_object */
1306 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1307 VM_OBJECT_UNLOCK(source);
1308 new_object->backing_object_offset =
1309 orig_object->backing_object_offset + entry->offset;
1310 new_object->backing_object = source;
1312 if (orig_object->uip != NULL) {
1313 new_object->uip = orig_object->uip;
1314 uihold(orig_object->uip);
1315 new_object->charge = ptoa(size);
1316 KASSERT(orig_object->charge >= ptoa(size),
1317 ("orig_object->charge < 0"));
1318 orig_object->charge -= ptoa(size);
1321 m = vm_page_find_least(orig_object, offidxstart);
1322 vm_page_lock_queues();
1323 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1325 m_next = TAILQ_NEXT(m, listq);
1328 * We must wait for pending I/O to complete before we can
1331 * We do not have to VM_PROT_NONE the page as mappings should
1332 * not be changed by this operation.
1334 if ((m->oflags & VPO_BUSY) || m->busy) {
1335 vm_page_unlock_queues();
1336 VM_OBJECT_UNLOCK(new_object);
1337 m->oflags |= VPO_WANTED;
1338 msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1339 VM_OBJECT_LOCK(new_object);
1342 vm_page_rename(m, new_object, idx);
1343 /* page automatically made dirty by rename and cache handled */
1346 vm_page_unlock_queues();
1347 if (orig_object->type == OBJT_SWAP) {
1349 * swap_pager_copy() can sleep, in which case the orig_object's
1350 * and new_object's locks are released and reacquired.
1352 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1355 * Transfer any cached pages from orig_object to new_object.
1357 if (__predict_false(orig_object->cache != NULL))
1358 vm_page_cache_transfer(orig_object, offidxstart,
1361 VM_OBJECT_UNLOCK(orig_object);
1362 TAILQ_FOREACH(m, &new_object->memq, listq)
1364 VM_OBJECT_UNLOCK(new_object);
1365 entry->object.vm_object = new_object;
1366 entry->offset = 0LL;
1367 vm_object_deallocate(orig_object);
1368 VM_OBJECT_LOCK(new_object);
1371 #define OBSC_TEST_ALL_SHADOWED 0x0001
1372 #define OBSC_COLLAPSE_NOWAIT 0x0002
1373 #define OBSC_COLLAPSE_WAIT 0x0004
1376 vm_object_backing_scan(vm_object_t object, int op)
1380 vm_object_t backing_object;
1381 vm_pindex_t backing_offset_index;
1383 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1384 VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1386 backing_object = object->backing_object;
1387 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1390 * Initial conditions
1392 if (op & OBSC_TEST_ALL_SHADOWED) {
1394 * We do not want to have to test for the existence of cache
1395 * or swap pages in the backing object. XXX but with the
1396 * new swapper this would be pretty easy to do.
1398 * XXX what about anonymous MAP_SHARED memory that hasn't
1399 * been ZFOD faulted yet? If we do not test for this, the
1400 * shadow test may succeed! XXX
1402 if (backing_object->type != OBJT_DEFAULT) {
1406 if (op & OBSC_COLLAPSE_WAIT) {
1407 vm_object_set_flag(backing_object, OBJ_DEAD);
1413 p = TAILQ_FIRST(&backing_object->memq);
1415 vm_page_t next = TAILQ_NEXT(p, listq);
1416 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1418 if (op & OBSC_TEST_ALL_SHADOWED) {
1422 * Ignore pages outside the parent object's range
1423 * and outside the parent object's mapping of the
1426 * note that we do not busy the backing object's
1430 p->pindex < backing_offset_index ||
1431 new_pindex >= object->size
1438 * See if the parent has the page or if the parent's
1439 * object pager has the page. If the parent has the
1440 * page but the page is not valid, the parent's
1441 * object pager must have the page.
1443 * If this fails, the parent does not completely shadow
1444 * the object and we might as well give up now.
1447 pp = vm_page_lookup(object, new_pindex);
1449 (pp == NULL || pp->valid == 0) &&
1450 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1458 * Check for busy page
1460 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1463 if (op & OBSC_COLLAPSE_NOWAIT) {
1464 if ((p->oflags & VPO_BUSY) ||
1470 } else if (op & OBSC_COLLAPSE_WAIT) {
1471 if ((p->oflags & VPO_BUSY) || p->busy) {
1472 VM_OBJECT_UNLOCK(object);
1473 p->oflags |= VPO_WANTED;
1474 msleep(p, VM_OBJECT_MTX(backing_object),
1475 PDROP | PVM, "vmocol", 0);
1476 VM_OBJECT_LOCK(object);
1477 VM_OBJECT_LOCK(backing_object);
1479 * If we slept, anything could have
1480 * happened. Since the object is
1481 * marked dead, the backing offset
1482 * should not have changed so we
1483 * just restart our scan.
1485 p = TAILQ_FIRST(&backing_object->memq);
1491 p->object == backing_object,
1492 ("vm_object_backing_scan: object mismatch")
1496 * Destroy any associated swap
1498 if (backing_object->type == OBJT_SWAP) {
1499 swap_pager_freespace(
1507 p->pindex < backing_offset_index ||
1508 new_pindex >= object->size
1511 * Page is out of the parent object's range, we
1512 * can simply destroy it.
1514 vm_page_lock_queues();
1515 KASSERT(!pmap_page_is_mapped(p),
1516 ("freeing mapped page %p", p));
1517 if (p->wire_count == 0)
1521 vm_page_unlock_queues();
1526 pp = vm_page_lookup(object, new_pindex);
1528 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1529 (pp != NULL && pp->valid == 0)
1532 * The page in the parent is not (yet) valid.
1533 * We don't know anything about the state of
1534 * the original page. It might be mapped,
1535 * so we must avoid the next if here.
1537 * This is due to a race in vm_fault() where
1538 * we must unbusy the original (backing_obj)
1539 * page before we can (re)lock the parent.
1540 * Hence we can get here.
1547 vm_pager_has_page(object, new_pindex, NULL, NULL)
1550 * page already exists in parent OR swap exists
1551 * for this location in the parent. Destroy
1552 * the original page from the backing object.
1554 * Leave the parent's page alone
1556 vm_page_lock_queues();
1557 KASSERT(!pmap_page_is_mapped(p),
1558 ("freeing mapped page %p", p));
1559 if (p->wire_count == 0)
1563 vm_page_unlock_queues();
1568 #if VM_NRESERVLEVEL > 0
1570 * Rename the reservation.
1572 vm_reserv_rename(p, object, backing_object,
1573 backing_offset_index);
1577 * Page does not exist in parent, rename the
1578 * page from the backing object to the main object.
1580 * If the page was mapped to a process, it can remain
1581 * mapped through the rename.
1583 vm_page_lock_queues();
1584 vm_page_rename(p, object, new_pindex);
1585 vm_page_unlock_queues();
1586 /* page automatically made dirty by rename */
1595 * this version of collapse allows the operation to occur earlier and
1596 * when paging_in_progress is true for an object... This is not a complete
1597 * operation, but should plug 99.9% of the rest of the leaks.
1600 vm_object_qcollapse(vm_object_t object)
1602 vm_object_t backing_object = object->backing_object;
1604 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1605 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1607 if (backing_object->ref_count != 1)
1610 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1614 * vm_object_collapse:
1616 * Collapse an object with the object backing it.
1617 * Pages in the backing object are moved into the
1618 * parent, and the backing object is deallocated.
1621 vm_object_collapse(vm_object_t object)
1623 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1626 vm_object_t backing_object;
1629 * Verify that the conditions are right for collapse:
1631 * The object exists and the backing object exists.
1633 if ((backing_object = object->backing_object) == NULL)
1637 * we check the backing object first, because it is most likely
1640 VM_OBJECT_LOCK(backing_object);
1641 if (backing_object->handle != NULL ||
1642 (backing_object->type != OBJT_DEFAULT &&
1643 backing_object->type != OBJT_SWAP) ||
1644 (backing_object->flags & OBJ_DEAD) ||
1645 object->handle != NULL ||
1646 (object->type != OBJT_DEFAULT &&
1647 object->type != OBJT_SWAP) ||
1648 (object->flags & OBJ_DEAD)) {
1649 VM_OBJECT_UNLOCK(backing_object);
1654 object->paging_in_progress != 0 ||
1655 backing_object->paging_in_progress != 0
1657 vm_object_qcollapse(object);
1658 VM_OBJECT_UNLOCK(backing_object);
1662 * We know that we can either collapse the backing object (if
1663 * the parent is the only reference to it) or (perhaps) have
1664 * the parent bypass the object if the parent happens to shadow
1665 * all the resident pages in the entire backing object.
1667 * This is ignoring pager-backed pages such as swap pages.
1668 * vm_object_backing_scan fails the shadowing test in this
1671 if (backing_object->ref_count == 1) {
1673 * If there is exactly one reference to the backing
1674 * object, we can collapse it into the parent.
1676 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1678 #if VM_NRESERVLEVEL > 0
1680 * Break any reservations from backing_object.
1682 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1683 vm_reserv_break_all(backing_object);
1687 * Move the pager from backing_object to object.
1689 if (backing_object->type == OBJT_SWAP) {
1691 * swap_pager_copy() can sleep, in which case
1692 * the backing_object's and object's locks are
1693 * released and reacquired.
1698 OFF_TO_IDX(object->backing_object_offset), TRUE);
1701 * Free any cached pages from backing_object.
1703 if (__predict_false(backing_object->cache != NULL))
1704 vm_page_cache_free(backing_object, 0, 0);
1707 * Object now shadows whatever backing_object did.
1708 * Note that the reference to
1709 * backing_object->backing_object moves from within
1710 * backing_object to within object.
1712 LIST_REMOVE(object, shadow_list);
1713 backing_object->shadow_count--;
1714 if (backing_object->backing_object) {
1715 VM_OBJECT_LOCK(backing_object->backing_object);
1716 LIST_REMOVE(backing_object, shadow_list);
1718 &backing_object->backing_object->shadow_head,
1719 object, shadow_list);
1721 * The shadow_count has not changed.
1723 VM_OBJECT_UNLOCK(backing_object->backing_object);
1725 object->backing_object = backing_object->backing_object;
1726 object->backing_object_offset +=
1727 backing_object->backing_object_offset;
1730 * Discard backing_object.
1732 * Since the backing object has no pages, no pager left,
1733 * and no object references within it, all that is
1734 * necessary is to dispose of it.
1736 KASSERT(backing_object->ref_count == 1, (
1737 "backing_object %p was somehow re-referenced during collapse!",
1739 VM_OBJECT_UNLOCK(backing_object);
1740 vm_object_destroy(backing_object);
1744 vm_object_t new_backing_object;
1747 * If we do not entirely shadow the backing object,
1748 * there is nothing we can do so we give up.
1750 if (object->resident_page_count != object->size &&
1751 vm_object_backing_scan(object,
1752 OBSC_TEST_ALL_SHADOWED) == 0) {
1753 VM_OBJECT_UNLOCK(backing_object);
1758 * Make the parent shadow the next object in the
1759 * chain. Deallocating backing_object will not remove
1760 * it, since its reference count is at least 2.
1762 LIST_REMOVE(object, shadow_list);
1763 backing_object->shadow_count--;
1765 new_backing_object = backing_object->backing_object;
1766 if ((object->backing_object = new_backing_object) != NULL) {
1767 VM_OBJECT_LOCK(new_backing_object);
1769 &new_backing_object->shadow_head,
1773 new_backing_object->shadow_count++;
1774 vm_object_reference_locked(new_backing_object);
1775 VM_OBJECT_UNLOCK(new_backing_object);
1776 object->backing_object_offset +=
1777 backing_object->backing_object_offset;
1781 * Drop the reference count on backing_object. Since
1782 * its ref_count was at least 2, it will not vanish.
1784 backing_object->ref_count--;
1785 VM_OBJECT_UNLOCK(backing_object);
1790 * Try again with this object's new backing object.
1796 * vm_object_page_remove:
1798 * For the given object, either frees or invalidates each of the
1799 * specified pages. In general, a page is freed. However, if a
1800 * page is wired for any reason other than the existence of a
1801 * managed, wired mapping, then it may be invalidated but not
1802 * removed from the object. Pages are specified by the given
1803 * range ["start", "end") and Boolean "clean_only". As a
1804 * special case, if "end" is zero, then the range extends from
1805 * "start" to the end of the object. If "clean_only" is TRUE,
1806 * then only the non-dirty pages within the specified range are
1809 * In general, this operation should only be performed on objects
1810 * that contain managed pages. There are two exceptions. First,
1811 * it may be performed on the kernel and kmem objects. Second,
1812 * it may be used by msync(..., MS_INVALIDATE) to invalidate
1813 * device-backed pages. In both of these cases, "clean_only"
1816 * The object must be locked.
1819 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1820 boolean_t clean_only)
1825 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1826 if (object->resident_page_count == 0)
1830 * Since physically-backed objects do not use managed pages, we can't
1831 * remove pages from the object (we must instead remove the page
1832 * references, and then destroy the object).
1834 KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1835 object == kmem_object,
1836 ("attempt to remove pages from a physical object"));
1838 vm_object_pip_add(object, 1);
1840 p = vm_page_find_least(object, start);
1841 vm_page_lock_queues();
1843 * Assert: the variable p is either (1) the page with the
1844 * least pindex greater than or equal to the parameter pindex
1848 p != NULL && (p->pindex < end || end == 0);
1850 next = TAILQ_NEXT(p, listq);
1853 * If the page is wired for any reason besides the
1854 * existence of managed, wired mappings, then it cannot
1855 * be freed. For example, fictitious pages, which
1856 * represent device memory, are inherently wired and
1857 * cannot be freed. They can, however, be invalidated
1858 * if "clean_only" is FALSE.
1860 if ((wirings = p->wire_count) != 0 &&
1861 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1862 /* Fictitious pages do not have managed mappings. */
1863 if ((p->flags & PG_FICTITIOUS) == 0)
1865 /* Account for removal of managed, wired mappings. */
1866 p->wire_count -= wirings;
1873 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1875 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1876 ("vm_object_page_remove: page %p is fictitious", p));
1877 if (clean_only && p->valid) {
1878 pmap_remove_write(p);
1883 /* Account for removal of managed, wired mappings. */
1885 p->wire_count -= wirings;
1888 vm_page_unlock_queues();
1889 vm_object_pip_wakeup(object);
1891 if (__predict_false(object->cache != NULL))
1892 vm_page_cache_free(object, start, end);
1896 * vm_object_page_cache:
1898 * For the given object, attempt to move the specified clean
1899 * pages to the cache queue. If a page is wired for any reason,
1900 * then it will not be changed. Pages are specified by the given
1901 * range ["start", "end"). As a special case, if "end" is zero,
1902 * then the range extends from "start" to the end of the object.
1903 * Any mappings to the specified pages are removed before the
1904 * pages are moved to the cache queue.
1906 * This operation should only be performed on objects that
1907 * contain managed pages.
1909 * The object must be locked.
1912 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1916 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1917 KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_SG &&
1918 object->type != OBJT_PHYS),
1919 ("vm_object_page_cache: illegal object %p", object));
1920 if (object->resident_page_count == 0)
1922 p = vm_page_find_least(object, start);
1925 * Here, the variable "p" is either (1) the page with the least pindex
1926 * greater than or equal to the parameter "start" or (2) NULL.
1928 vm_page_lock_queues();
1929 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1930 next = TAILQ_NEXT(p, listq);
1932 vm_page_try_to_cache(p);
1934 vm_page_unlock_queues();
1938 * Populate the specified range of the object with valid pages. Returns
1939 * TRUE if the range is successfully populated and FALSE otherwise.
1941 * Note: This function should be optimized to pass a larger array of
1942 * pages to vm_pager_get_pages() before it is applied to a non-
1943 * OBJT_DEVICE object.
1945 * The object must be locked.
1948 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1954 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1955 for (pindex = start; pindex < end; pindex++) {
1956 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1958 if (m->valid != VM_PAGE_BITS_ALL) {
1960 rv = vm_pager_get_pages(object, ma, 1, 0);
1961 m = vm_page_lookup(object, pindex);
1964 if (rv != VM_PAGER_OK) {
1965 vm_page_lock_queues();
1967 vm_page_unlock_queues();
1972 * Keep "m" busy because a subsequent iteration may unlock
1976 if (pindex > start) {
1977 m = vm_page_lookup(object, start);
1978 while (m != NULL && m->pindex < pindex) {
1980 m = TAILQ_NEXT(m, listq);
1983 return (pindex == end);
1987 * Routine: vm_object_coalesce
1988 * Function: Coalesces two objects backing up adjoining
1989 * regions of memory into a single object.
1991 * returns TRUE if objects were combined.
1993 * NOTE: Only works at the moment if the second object is NULL -
1994 * if it's not, which object do we lock first?
1997 * prev_object First object to coalesce
1998 * prev_offset Offset into prev_object
1999 * prev_size Size of reference to prev_object
2000 * next_size Size of reference to the second object
2001 * reserved Indicator that extension region has
2002 * swap accounted for
2005 * The object must *not* be locked.
2008 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2009 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2011 vm_pindex_t next_pindex;
2013 if (prev_object == NULL)
2015 VM_OBJECT_LOCK(prev_object);
2016 if (prev_object->type != OBJT_DEFAULT &&
2017 prev_object->type != OBJT_SWAP) {
2018 VM_OBJECT_UNLOCK(prev_object);
2023 * Try to collapse the object first
2025 vm_object_collapse(prev_object);
2028 * Can't coalesce if: . more than one reference . paged out . shadows
2029 * another object . has a copy elsewhere (any of which mean that the
2030 * pages not mapped to prev_entry may be in use anyway)
2032 if (prev_object->backing_object != NULL) {
2033 VM_OBJECT_UNLOCK(prev_object);
2037 prev_size >>= PAGE_SHIFT;
2038 next_size >>= PAGE_SHIFT;
2039 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2041 if ((prev_object->ref_count > 1) &&
2042 (prev_object->size != next_pindex)) {
2043 VM_OBJECT_UNLOCK(prev_object);
2048 * Account for the charge.
2050 if (prev_object->uip != NULL) {
2053 * If prev_object was charged, then this mapping,
2054 * althought not charged now, may become writable
2055 * later. Non-NULL uip in the object would prevent
2056 * swap reservation during enabling of the write
2057 * access, so reserve swap now. Failed reservation
2058 * cause allocation of the separate object for the map
2059 * entry, and swap reservation for this entry is
2060 * managed in appropriate time.
2062 if (!reserved && !swap_reserve_by_uid(ptoa(next_size),
2063 prev_object->uip)) {
2066 prev_object->charge += ptoa(next_size);
2070 * Remove any pages that may still be in the object from a previous
2073 if (next_pindex < prev_object->size) {
2074 vm_object_page_remove(prev_object,
2076 next_pindex + next_size, FALSE);
2077 if (prev_object->type == OBJT_SWAP)
2078 swap_pager_freespace(prev_object,
2079 next_pindex, next_size);
2081 if (prev_object->uip != NULL) {
2082 KASSERT(prev_object->charge >=
2083 ptoa(prev_object->size - next_pindex),
2084 ("object %p overcharged 1 %jx %jx", prev_object,
2085 (uintmax_t)next_pindex, (uintmax_t)next_size));
2086 prev_object->charge -= ptoa(prev_object->size -
2093 * Extend the object if necessary.
2095 if (next_pindex + next_size > prev_object->size)
2096 prev_object->size = next_pindex + next_size;
2098 VM_OBJECT_UNLOCK(prev_object);
2103 vm_object_set_writeable_dirty(vm_object_t object)
2106 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2107 if (object->type != OBJT_VNODE)
2109 object->generation++;
2110 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2112 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2115 #include "opt_ddb.h"
2117 #include <sys/kernel.h>
2119 #include <sys/cons.h>
2121 #include <ddb/ddb.h>
2124 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2127 vm_map_entry_t tmpe;
2135 tmpe = map->header.next;
2136 entcount = map->nentries;
2137 while (entcount-- && (tmpe != &map->header)) {
2138 if (_vm_object_in_map(map, object, tmpe)) {
2143 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2144 tmpm = entry->object.sub_map;
2145 tmpe = tmpm->header.next;
2146 entcount = tmpm->nentries;
2147 while (entcount-- && tmpe != &tmpm->header) {
2148 if (_vm_object_in_map(tmpm, object, tmpe)) {
2153 } else if ((obj = entry->object.vm_object) != NULL) {
2154 for (; obj; obj = obj->backing_object)
2155 if (obj == object) {
2163 vm_object_in_map(vm_object_t object)
2167 /* sx_slock(&allproc_lock); */
2168 FOREACH_PROC_IN_SYSTEM(p) {
2169 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2171 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2172 /* sx_sunlock(&allproc_lock); */
2176 /* sx_sunlock(&allproc_lock); */
2177 if (_vm_object_in_map(kernel_map, object, 0))
2179 if (_vm_object_in_map(kmem_map, object, 0))
2181 if (_vm_object_in_map(pager_map, object, 0))
2183 if (_vm_object_in_map(buffer_map, object, 0))
2188 DB_SHOW_COMMAND(vmochk, vm_object_check)
2193 * make sure that internal objs are in a map somewhere
2194 * and none have zero ref counts.
2196 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2197 if (object->handle == NULL &&
2198 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2199 if (object->ref_count == 0) {
2200 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2201 (long)object->size);
2203 if (!vm_object_in_map(object)) {
2205 "vmochk: internal obj is not in a map: "
2206 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2207 object->ref_count, (u_long)object->size,
2208 (u_long)object->size,
2209 (void *)object->backing_object);
2216 * vm_object_print: [ debug ]
2218 DB_SHOW_COMMAND(object, vm_object_print_static)
2220 /* XXX convert args. */
2221 vm_object_t object = (vm_object_t)addr;
2222 boolean_t full = have_addr;
2226 /* XXX count is an (unused) arg. Avoid shadowing it. */
2227 #define count was_count
2235 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x uip %d charge %jx\n",
2236 object, (int)object->type, (uintmax_t)object->size,
2237 object->resident_page_count, object->ref_count, object->flags,
2238 object->uip ? object->uip->ui_uid : -1, (uintmax_t)object->charge);
2239 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2240 object->shadow_count,
2241 object->backing_object ? object->backing_object->ref_count : 0,
2242 object->backing_object, (uintmax_t)object->backing_object_offset);
2249 TAILQ_FOREACH(p, &object->memq, listq) {
2251 db_iprintf("memory:=");
2252 else if (count == 6) {
2260 db_printf("(off=0x%jx,page=0x%jx)",
2261 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2271 /* XXX need this non-static entry for calling from vm_map_print. */
2274 /* db_expr_t */ long addr,
2275 boolean_t have_addr,
2276 /* db_expr_t */ long count,
2279 vm_object_print_static(addr, have_addr, count, modif);
2282 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2287 vm_page_t m, prev_m;
2291 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2292 db_printf("new object: %p\n", (void *)object);
2303 TAILQ_FOREACH(m, &object->memq, listq) {
2304 if (m->pindex > 128)
2306 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2307 prev_m->pindex + 1 != m->pindex) {
2309 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2310 (long)fidx, rcount, (long)pa);
2322 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2327 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2328 (long)fidx, rcount, (long)pa);
2338 pa = VM_PAGE_TO_PHYS(m);
2342 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2343 (long)fidx, rcount, (long)pa);