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,
109 static void vm_object_qcollapse(vm_object_t object);
110 static void vm_object_vndeallocate(vm_object_t object);
113 * Virtual memory objects maintain the actual data
114 * associated with allocated virtual memory. A given
115 * page of memory exists within exactly one object.
117 * An object is only deallocated when all "references"
118 * are given up. Only one "reference" to a given
119 * region of an object should be writeable.
121 * Associated with each object is a list of all resident
122 * memory pages belonging to that object; this list is
123 * maintained by the "vm_page" module, and locked by the object's
126 * Each object also records a "pager" routine which is
127 * used to retrieve (and store) pages to the proper backing
128 * storage. In addition, objects may be backed by other
129 * objects from which they were virtual-copied.
131 * The only items within the object structure which are
132 * modified after time of creation are:
133 * reference count locked by object's lock
134 * pager routine locked by object's lock
138 struct object_q vm_object_list;
139 struct mtx vm_object_list_mtx; /* lock for object list and count */
141 struct vm_object kernel_object_store;
142 struct vm_object kmem_object_store;
144 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats");
146 static long object_collapses;
147 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
148 &object_collapses, 0, "VM object collapses");
150 static long object_bypasses;
151 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
152 &object_bypasses, 0, "VM object bypasses");
154 static uma_zone_t obj_zone;
156 static int vm_object_zinit(void *mem, int size, int flags);
159 static void vm_object_zdtor(void *mem, int size, void *arg);
162 vm_object_zdtor(void *mem, int size, void *arg)
166 object = (vm_object_t)mem;
167 KASSERT(TAILQ_EMPTY(&object->memq),
168 ("object %p has resident pages",
170 #if VM_NRESERVLEVEL > 0
171 KASSERT(LIST_EMPTY(&object->rvq),
172 ("object %p has reservations",
175 KASSERT(object->cache == NULL,
176 ("object %p has cached pages",
178 KASSERT(object->paging_in_progress == 0,
179 ("object %p paging_in_progress = %d",
180 object, object->paging_in_progress));
181 KASSERT(object->resident_page_count == 0,
182 ("object %p resident_page_count = %d",
183 object, object->resident_page_count));
184 KASSERT(object->shadow_count == 0,
185 ("object %p shadow_count = %d",
186 object, object->shadow_count));
191 vm_object_zinit(void *mem, int size, int flags)
195 object = (vm_object_t)mem;
196 bzero(&object->mtx, sizeof(object->mtx));
197 VM_OBJECT_LOCK_INIT(object, "standard object");
199 /* These are true for any object that has been freed */
200 object->paging_in_progress = 0;
201 object->resident_page_count = 0;
202 object->shadow_count = 0;
207 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
210 TAILQ_INIT(&object->memq);
211 LIST_INIT(&object->shadow_head);
216 object->generation = 1;
217 object->ref_count = 1;
218 object->memattr = VM_MEMATTR_DEFAULT;
222 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
223 object->flags = OBJ_ONEMAPPING;
224 object->pg_color = 0;
225 object->handle = NULL;
226 object->backing_object = NULL;
227 object->backing_object_offset = (vm_ooffset_t) 0;
228 #if VM_NRESERVLEVEL > 0
229 LIST_INIT(&object->rvq);
231 object->cache = NULL;
233 mtx_lock(&vm_object_list_mtx);
234 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
235 mtx_unlock(&vm_object_list_mtx);
241 * Initialize the VM objects module.
246 TAILQ_INIT(&vm_object_list);
247 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
249 VM_OBJECT_LOCK_INIT(&kernel_object_store, "kernel object");
250 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
252 #if VM_NRESERVLEVEL > 0
253 kernel_object->flags |= OBJ_COLORED;
254 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
257 VM_OBJECT_LOCK_INIT(&kmem_object_store, "kmem object");
258 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
260 #if VM_NRESERVLEVEL > 0
261 kmem_object->flags |= OBJ_COLORED;
262 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
266 * The lock portion of struct vm_object must be type stable due
267 * to vm_pageout_fallback_object_lock locking a vm object
268 * without holding any references to it.
270 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
276 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
280 vm_object_clear_flag(vm_object_t object, u_short bits)
283 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
284 object->flags &= ~bits;
288 * Sets the default memory attribute for the specified object. Pages
289 * that are allocated to this object are by default assigned this memory
292 * Presently, this function must be called before any pages are allocated
293 * to the object. In the future, this requirement may be relaxed for
294 * "default" and "swap" objects.
297 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
300 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
301 switch (object->type) {
308 if (!TAILQ_EMPTY(&object->memq))
309 return (KERN_FAILURE);
312 return (KERN_INVALID_ARGUMENT);
314 object->memattr = memattr;
315 return (KERN_SUCCESS);
319 vm_object_pip_add(vm_object_t object, short i)
322 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
323 object->paging_in_progress += i;
327 vm_object_pip_subtract(vm_object_t object, short i)
330 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
331 object->paging_in_progress -= i;
335 vm_object_pip_wakeup(vm_object_t object)
338 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
339 object->paging_in_progress--;
340 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
341 vm_object_clear_flag(object, OBJ_PIPWNT);
347 vm_object_pip_wakeupn(vm_object_t object, short i)
350 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
352 object->paging_in_progress -= i;
353 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
354 vm_object_clear_flag(object, OBJ_PIPWNT);
360 vm_object_pip_wait(vm_object_t object, char *waitid)
363 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
364 while (object->paging_in_progress) {
365 object->flags |= OBJ_PIPWNT;
366 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
371 * vm_object_allocate:
373 * Returns a new object with the given size.
376 vm_object_allocate(objtype_t type, vm_pindex_t size)
380 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
381 _vm_object_allocate(type, size, object);
387 * vm_object_reference:
389 * Gets another reference to the given object. Note: OBJ_DEAD
390 * objects can be referenced during final cleaning.
393 vm_object_reference(vm_object_t object)
397 VM_OBJECT_LOCK(object);
398 vm_object_reference_locked(object);
399 VM_OBJECT_UNLOCK(object);
403 * vm_object_reference_locked:
405 * Gets another reference to the given object.
407 * The object must be locked.
410 vm_object_reference_locked(vm_object_t object)
414 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
416 if (object->type == OBJT_VNODE) {
423 * Handle deallocating an object of type OBJT_VNODE.
426 vm_object_vndeallocate(vm_object_t object)
428 struct vnode *vp = (struct vnode *) object->handle;
430 VFS_ASSERT_GIANT(vp->v_mount);
431 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
432 KASSERT(object->type == OBJT_VNODE,
433 ("vm_object_vndeallocate: not a vnode object"));
434 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
436 if (object->ref_count == 0) {
437 vprint("vm_object_vndeallocate", vp);
438 panic("vm_object_vndeallocate: bad object reference count");
443 if (object->ref_count == 0) {
444 mp_fixme("Unlocked vflag access.");
445 vp->v_vflag &= ~VV_TEXT;
447 VM_OBJECT_UNLOCK(object);
449 * vrele may need a vop lock
455 * vm_object_deallocate:
457 * Release a reference to the specified object,
458 * gained either through a vm_object_allocate
459 * or a vm_object_reference call. When all references
460 * are gone, storage associated with this object
461 * may be relinquished.
463 * No object may be locked.
466 vm_object_deallocate(vm_object_t object)
470 while (object != NULL) {
475 VM_OBJECT_LOCK(object);
476 if (object->type == OBJT_VNODE) {
477 struct vnode *vp = (struct vnode *) object->handle;
480 * Conditionally acquire Giant for a vnode-backed
481 * object. We have to be careful since the type of
482 * a vnode object can change while the object is
485 if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
487 if (!mtx_trylock(&Giant)) {
488 VM_OBJECT_UNLOCK(object);
493 vm_object_vndeallocate(object);
494 VFS_UNLOCK_GIANT(vfslocked);
498 * This is to handle the case that the object
499 * changed type while we dropped its lock to
502 VFS_UNLOCK_GIANT(vfslocked);
504 KASSERT(object->ref_count != 0,
505 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
508 * If the reference count goes to 0 we start calling
509 * vm_object_terminate() on the object chain.
510 * A ref count of 1 may be a special case depending on the
511 * shadow count being 0 or 1.
514 if (object->ref_count > 1) {
515 VM_OBJECT_UNLOCK(object);
517 } else if (object->ref_count == 1) {
518 if (object->shadow_count == 0 &&
519 object->handle == NULL &&
520 (object->type == OBJT_DEFAULT ||
521 object->type == OBJT_SWAP)) {
522 vm_object_set_flag(object, OBJ_ONEMAPPING);
523 } else if ((object->shadow_count == 1) &&
524 (object->handle == NULL) &&
525 (object->type == OBJT_DEFAULT ||
526 object->type == OBJT_SWAP)) {
529 robject = LIST_FIRST(&object->shadow_head);
530 KASSERT(robject != NULL,
531 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
533 object->shadow_count));
534 if (!VM_OBJECT_TRYLOCK(robject)) {
536 * Avoid a potential deadlock.
539 VM_OBJECT_UNLOCK(object);
541 * More likely than not the thread
542 * holding robject's lock has lower
543 * priority than the current thread.
544 * Let the lower priority thread run.
550 * Collapse object into its shadow unless its
551 * shadow is dead. In that case, object will
552 * be deallocated by the thread that is
553 * deallocating its shadow.
555 if ((robject->flags & OBJ_DEAD) == 0 &&
556 (robject->handle == NULL) &&
557 (robject->type == OBJT_DEFAULT ||
558 robject->type == OBJT_SWAP)) {
560 robject->ref_count++;
562 if (robject->paging_in_progress) {
563 VM_OBJECT_UNLOCK(object);
564 vm_object_pip_wait(robject,
566 temp = robject->backing_object;
567 if (object == temp) {
568 VM_OBJECT_LOCK(object);
571 } else if (object->paging_in_progress) {
572 VM_OBJECT_UNLOCK(robject);
573 object->flags |= OBJ_PIPWNT;
575 VM_OBJECT_MTX(object),
576 PDROP | PVM, "objde2", 0);
577 VM_OBJECT_LOCK(robject);
578 temp = robject->backing_object;
579 if (object == temp) {
580 VM_OBJECT_LOCK(object);
584 VM_OBJECT_UNLOCK(object);
586 if (robject->ref_count == 1) {
587 robject->ref_count--;
592 vm_object_collapse(object);
593 VM_OBJECT_UNLOCK(object);
596 VM_OBJECT_UNLOCK(robject);
598 VM_OBJECT_UNLOCK(object);
602 temp = object->backing_object;
604 VM_OBJECT_LOCK(temp);
605 LIST_REMOVE(object, shadow_list);
606 temp->shadow_count--;
608 VM_OBJECT_UNLOCK(temp);
609 object->backing_object = NULL;
612 * Don't double-terminate, we could be in a termination
613 * recursion due to the terminate having to sync data
616 if ((object->flags & OBJ_DEAD) == 0)
617 vm_object_terminate(object);
619 VM_OBJECT_UNLOCK(object);
625 * vm_object_destroy removes the object from the global object list
626 * and frees the space for the object.
629 vm_object_destroy(vm_object_t object)
633 * Remove the object from the global object list.
635 mtx_lock(&vm_object_list_mtx);
636 TAILQ_REMOVE(&vm_object_list, object, object_list);
637 mtx_unlock(&vm_object_list_mtx);
640 * Release the allocation charge.
642 if (object->uip != NULL) {
643 KASSERT(object->type == OBJT_DEFAULT ||
644 object->type == OBJT_SWAP,
645 ("vm_object_terminate: non-swap obj %p has uip",
647 swap_release_by_uid(object->charge, object->uip);
654 * Free the space for the object.
656 uma_zfree(obj_zone, object);
660 * vm_object_terminate actually destroys the specified object, freeing
661 * up all previously used resources.
663 * The object must be locked.
664 * This routine may block.
667 vm_object_terminate(vm_object_t object)
671 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
674 * Make sure no one uses us.
676 vm_object_set_flag(object, OBJ_DEAD);
679 * wait for the pageout daemon to be done with the object
681 vm_object_pip_wait(object, "objtrm");
683 KASSERT(!object->paging_in_progress,
684 ("vm_object_terminate: pageout in progress"));
687 * Clean and free the pages, as appropriate. All references to the
688 * object are gone, so we don't need to lock it.
690 if (object->type == OBJT_VNODE) {
691 struct vnode *vp = (struct vnode *)object->handle;
694 * Clean pages and flush buffers.
696 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
697 VM_OBJECT_UNLOCK(object);
699 vinvalbuf(vp, V_SAVE, 0, 0);
701 VM_OBJECT_LOCK(object);
704 KASSERT(object->ref_count == 0,
705 ("vm_object_terminate: object with references, ref_count=%d",
709 * Now free any remaining pages. For internal objects, this also
710 * removes them from paging queues. Don't free wired pages, just
711 * remove them from the object.
713 vm_page_lock_queues();
714 while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
715 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
716 ("vm_object_terminate: freeing busy page %p "
717 "p->busy = %d, p->oflags %x\n", p, p->busy, p->oflags));
718 if (p->wire_count == 0) {
725 vm_page_unlock_queues();
727 #if VM_NRESERVLEVEL > 0
728 if (__predict_false(!LIST_EMPTY(&object->rvq)))
729 vm_reserv_break_all(object);
731 if (__predict_false(object->cache != NULL))
732 vm_page_cache_free(object, 0, 0);
735 * Let the pager know object is dead.
737 vm_pager_deallocate(object);
738 VM_OBJECT_UNLOCK(object);
740 vm_object_destroy(object);
744 * vm_object_page_clean
746 * Clean all dirty pages in the specified range of object. Leaves page
747 * on whatever queue it is currently on. If NOSYNC is set then do not
748 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
749 * leaving the object dirty.
751 * When stuffing pages asynchronously, allow clustering. XXX we need a
752 * synchronous clustering mode implementation.
754 * Odd semantics: if start == end, we clean everything.
756 * The object must be locked.
759 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
763 vm_pindex_t pi, tend;
764 int clearobjflags, curgeneration, n, pagerflags;
766 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
767 KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
768 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
769 object->resident_page_count == 0)
772 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
773 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
774 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
776 tend = (end == 0) ? object->size : end;
778 vm_object_set_flag(object, OBJ_CLEANING);
780 vm_page_lock_queues();
783 * Make the page read-only so we can then clear the object flags.
785 * However, if this is a nosync mmap then the object is likely to
786 * stay dirty so do not mess with the page and do not clear the
790 for (p = vm_page_find_least(object, start);
791 p != NULL && p->pindex < tend; p = TAILQ_NEXT(p, listq)) {
792 if ((flags & OBJPC_NOSYNC) != 0 &&
793 (p->oflags & VPO_NOSYNC) != 0)
796 pmap_remove_write(p);
799 if (clearobjflags && (start == 0) && (tend == object->size))
800 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
803 curgeneration = object->generation;
805 for (p = vm_page_find_least(object, start); p != NULL; p = np) {
809 np = TAILQ_NEXT(p, listq);
812 while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
813 vm_page_lock_queues();
814 if (object->generation != curgeneration)
817 vm_page_test_dirty(p);
822 * If we have been asked to skip nosync pages and this is a
823 * nosync page, skip it. Note that the object flags were
824 * not cleared in this case so we do not have to set them.
826 if ((flags & OBJPC_NOSYNC) != 0 &&
827 (p->oflags & VPO_NOSYNC) != 0)
830 n = vm_object_page_collect_flush(object, p, pagerflags);
831 KASSERT(n > 0, ("vm_object_page_collect_flush failed"));
832 if (object->generation != curgeneration)
834 np = vm_page_find_least(object, pi + n);
836 vm_page_unlock_queues();
838 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
841 vm_object_clear_flag(object, OBJ_CLEANING);
845 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags)
853 vm_page_t maf[vm_pageout_page_count];
854 vm_page_t mab[vm_pageout_page_count];
855 vm_page_t ma[vm_pageout_page_count];
858 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
861 for (i = 1, p1 = p; i < vm_pageout_page_count; i++) {
862 tp = vm_page_next(p1);
863 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
865 vm_page_test_dirty(tp);
868 maf[i - 1] = p1 = tp;
873 chkb = vm_pageout_page_count - maxf;
874 for (i = 1, p1 = p; i < chkb; i++) {
875 tp = vm_page_prev(p1);
876 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
878 vm_page_test_dirty(tp);
881 mab[i - 1] = p1 = tp;
885 for (i = 0; i < maxb; i++) {
886 index = (maxb - i) - 1;
890 for (i = 0; i < maxf; i++) {
891 index = (maxb + i) + 1;
894 runlen = maxb + maxf + 1;
896 vm_pageout_flush(ma, runlen, pagerflags);
897 for (i = 0; i < runlen; i++) {
898 if (ma[i]->dirty != 0) {
899 KASSERT((ma[i]->flags & PG_WRITEABLE) == 0,
900 ("vm_object_page_collect_flush: page %p is not write protected",
904 for (i = 0; i < maxf; i++) {
905 if (ma[i + maxb + 1]->dirty != 0) {
907 * maxf will end up being the actual number of pages
908 * we wrote out contiguously, non-inclusive of the
909 * first page. We do not count look-behind pages.
921 * Note that there is absolutely no sense in writing out
922 * anonymous objects, so we track down the vnode object
924 * We invalidate (remove) all pages from the address space
925 * for semantic correctness.
927 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
928 * may start out with a NULL object.
931 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
932 boolean_t syncio, boolean_t invalidate)
934 vm_object_t backing_object;
941 VM_OBJECT_LOCK(object);
942 while ((backing_object = object->backing_object) != NULL) {
943 VM_OBJECT_LOCK(backing_object);
944 offset += object->backing_object_offset;
945 VM_OBJECT_UNLOCK(object);
946 object = backing_object;
947 if (object->size < OFF_TO_IDX(offset + size))
948 size = IDX_TO_OFF(object->size) - offset;
951 * Flush pages if writing is allowed, invalidate them
952 * if invalidation requested. Pages undergoing I/O
953 * will be ignored by vm_object_page_remove().
955 * We cannot lock the vnode and then wait for paging
956 * to complete without deadlocking against vm_fault.
957 * Instead we simply call vm_object_page_remove() and
958 * allow it to block internally on a page-by-page
959 * basis when it encounters pages undergoing async
962 if (object->type == OBJT_VNODE &&
963 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
966 VM_OBJECT_UNLOCK(object);
967 (void) vn_start_write(vp, &mp, V_WAIT);
968 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
969 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
970 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
971 flags |= invalidate ? OBJPC_INVAL : 0;
972 VM_OBJECT_LOCK(object);
973 vm_object_page_clean(object,
975 OFF_TO_IDX(offset + size + PAGE_MASK),
977 VM_OBJECT_UNLOCK(object);
979 VFS_UNLOCK_GIANT(vfslocked);
980 vn_finished_write(mp);
981 VM_OBJECT_LOCK(object);
983 if ((object->type == OBJT_VNODE ||
984 object->type == OBJT_DEVICE) && invalidate) {
986 purge = old_msync || (object->type == OBJT_DEVICE);
987 vm_object_page_remove(object,
989 OFF_TO_IDX(offset + size + PAGE_MASK),
990 purge ? FALSE : TRUE);
992 VM_OBJECT_UNLOCK(object);
998 * Implements the madvise function at the object/page level.
1000 * MADV_WILLNEED (any object)
1002 * Activate the specified pages if they are resident.
1004 * MADV_DONTNEED (any object)
1006 * Deactivate the specified pages if they are resident.
1008 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1009 * OBJ_ONEMAPPING only)
1011 * Deactivate and clean the specified pages if they are
1012 * resident. This permits the process to reuse the pages
1013 * without faulting or the kernel to reclaim the pages
1017 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1019 vm_pindex_t end, tpindex;
1020 vm_object_t backing_object, tobject;
1025 VM_OBJECT_LOCK(object);
1026 end = pindex + count;
1028 * Locate and adjust resident pages
1030 for (; pindex < end; pindex += 1) {
1036 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1037 * and those pages must be OBJ_ONEMAPPING.
1039 if (advise == MADV_FREE) {
1040 if ((tobject->type != OBJT_DEFAULT &&
1041 tobject->type != OBJT_SWAP) ||
1042 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1043 goto unlock_tobject;
1045 } else if (tobject->type == OBJT_PHYS)
1046 goto unlock_tobject;
1047 m = vm_page_lookup(tobject, tpindex);
1048 if (m == NULL && advise == MADV_WILLNEED) {
1050 * If the page is cached, reactivate it.
1052 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1057 * There may be swap even if there is no backing page
1059 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1060 swap_pager_freespace(tobject, tpindex, 1);
1064 backing_object = tobject->backing_object;
1065 if (backing_object == NULL)
1066 goto unlock_tobject;
1067 VM_OBJECT_LOCK(backing_object);
1068 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1069 if (tobject != object)
1070 VM_OBJECT_UNLOCK(tobject);
1071 tobject = backing_object;
1073 } else if (m->valid != VM_PAGE_BITS_ALL)
1074 goto unlock_tobject;
1076 * If the page is not in a normal state, skip it.
1078 vm_page_lock_queues();
1079 if (m->hold_count != 0 || m->wire_count != 0) {
1080 vm_page_unlock_queues();
1081 goto unlock_tobject;
1083 if ((m->oflags & VPO_BUSY) || m->busy) {
1084 if (advise == MADV_WILLNEED)
1086 * Reference the page before unlocking and
1087 * sleeping so that the page daemon is less
1088 * likely to reclaim it.
1090 vm_page_flag_set(m, PG_REFERENCED);
1091 vm_page_unlock_queues();
1092 if (object != tobject)
1093 VM_OBJECT_UNLOCK(object);
1094 m->oflags |= VPO_WANTED;
1095 msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1097 VM_OBJECT_LOCK(object);
1100 if (advise == MADV_WILLNEED) {
1101 vm_page_activate(m);
1102 } else if (advise == MADV_DONTNEED) {
1103 vm_page_dontneed(m);
1104 } else if (advise == MADV_FREE) {
1106 * Mark the page clean. This will allow the page
1107 * to be freed up by the system. However, such pages
1108 * are often reused quickly by malloc()/free()
1109 * so we do not do anything that would cause
1110 * a page fault if we can help it.
1112 * Specifically, we do not try to actually free
1113 * the page now nor do we try to put it in the
1114 * cache (which would cause a page fault on reuse).
1116 * But we do make the page is freeable as we
1117 * can without actually taking the step of unmapping
1120 pmap_clear_modify(m);
1123 vm_page_dontneed(m);
1125 vm_page_unlock_queues();
1126 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1127 swap_pager_freespace(tobject, tpindex, 1);
1129 if (tobject != object)
1130 VM_OBJECT_UNLOCK(tobject);
1132 VM_OBJECT_UNLOCK(object);
1138 * Create a new object which is backed by the
1139 * specified existing object range. The source
1140 * object reference is deallocated.
1142 * The new object and offset into that object
1143 * are returned in the source parameters.
1147 vm_object_t *object, /* IN/OUT */
1148 vm_ooffset_t *offset, /* IN/OUT */
1157 * Don't create the new object if the old object isn't shared.
1159 if (source != NULL) {
1160 VM_OBJECT_LOCK(source);
1161 if (source->ref_count == 1 &&
1162 source->handle == NULL &&
1163 (source->type == OBJT_DEFAULT ||
1164 source->type == OBJT_SWAP)) {
1165 VM_OBJECT_UNLOCK(source);
1168 VM_OBJECT_UNLOCK(source);
1172 * Allocate a new object with the given length.
1174 result = vm_object_allocate(OBJT_DEFAULT, length);
1177 * The new object shadows the source object, adding a reference to it.
1178 * Our caller changes his reference to point to the new object,
1179 * removing a reference to the source object. Net result: no change
1180 * of reference count.
1182 * Try to optimize the result object's page color when shadowing
1183 * in order to maintain page coloring consistency in the combined
1186 result->backing_object = source;
1188 * Store the offset into the source object, and fix up the offset into
1191 result->backing_object_offset = *offset;
1192 if (source != NULL) {
1193 VM_OBJECT_LOCK(source);
1194 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1195 source->shadow_count++;
1196 source->generation++;
1197 #if VM_NRESERVLEVEL > 0
1198 result->flags |= source->flags & OBJ_COLORED;
1199 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1200 ((1 << (VM_NFREEORDER - 1)) - 1);
1202 VM_OBJECT_UNLOCK(source);
1207 * Return the new things
1216 * Split the pages in a map entry into a new object. This affords
1217 * easier removal of unused pages, and keeps object inheritance from
1218 * being a negative impact on memory usage.
1221 vm_object_split(vm_map_entry_t entry)
1223 vm_page_t m, m_next;
1224 vm_object_t orig_object, new_object, source;
1225 vm_pindex_t idx, offidxstart;
1228 orig_object = entry->object.vm_object;
1229 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1231 if (orig_object->ref_count <= 1)
1233 VM_OBJECT_UNLOCK(orig_object);
1235 offidxstart = OFF_TO_IDX(entry->offset);
1236 size = atop(entry->end - entry->start);
1239 * If swap_pager_copy() is later called, it will convert new_object
1240 * into a swap object.
1242 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1245 * At this point, the new object is still private, so the order in
1246 * which the original and new objects are locked does not matter.
1248 VM_OBJECT_LOCK(new_object);
1249 VM_OBJECT_LOCK(orig_object);
1250 source = orig_object->backing_object;
1251 if (source != NULL) {
1252 VM_OBJECT_LOCK(source);
1253 if ((source->flags & OBJ_DEAD) != 0) {
1254 VM_OBJECT_UNLOCK(source);
1255 VM_OBJECT_UNLOCK(orig_object);
1256 VM_OBJECT_UNLOCK(new_object);
1257 vm_object_deallocate(new_object);
1258 VM_OBJECT_LOCK(orig_object);
1261 LIST_INSERT_HEAD(&source->shadow_head,
1262 new_object, shadow_list);
1263 source->shadow_count++;
1264 source->generation++;
1265 vm_object_reference_locked(source); /* for new_object */
1266 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1267 VM_OBJECT_UNLOCK(source);
1268 new_object->backing_object_offset =
1269 orig_object->backing_object_offset + entry->offset;
1270 new_object->backing_object = source;
1272 if (orig_object->uip != NULL) {
1273 new_object->uip = orig_object->uip;
1274 uihold(orig_object->uip);
1275 new_object->charge = ptoa(size);
1276 KASSERT(orig_object->charge >= ptoa(size),
1277 ("orig_object->charge < 0"));
1278 orig_object->charge -= ptoa(size);
1281 m = vm_page_find_least(orig_object, offidxstart);
1282 vm_page_lock_queues();
1283 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1285 m_next = TAILQ_NEXT(m, listq);
1288 * We must wait for pending I/O to complete before we can
1291 * We do not have to VM_PROT_NONE the page as mappings should
1292 * not be changed by this operation.
1294 if ((m->oflags & VPO_BUSY) || m->busy) {
1295 vm_page_unlock_queues();
1296 VM_OBJECT_UNLOCK(new_object);
1297 m->oflags |= VPO_WANTED;
1298 msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1299 VM_OBJECT_LOCK(new_object);
1302 vm_page_rename(m, new_object, idx);
1303 /* page automatically made dirty by rename and cache handled */
1306 vm_page_unlock_queues();
1307 if (orig_object->type == OBJT_SWAP) {
1309 * swap_pager_copy() can sleep, in which case the orig_object's
1310 * and new_object's locks are released and reacquired.
1312 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1315 * Transfer any cached pages from orig_object to new_object.
1317 if (__predict_false(orig_object->cache != NULL))
1318 vm_page_cache_transfer(orig_object, offidxstart,
1321 VM_OBJECT_UNLOCK(orig_object);
1322 TAILQ_FOREACH(m, &new_object->memq, listq)
1324 VM_OBJECT_UNLOCK(new_object);
1325 entry->object.vm_object = new_object;
1326 entry->offset = 0LL;
1327 vm_object_deallocate(orig_object);
1328 VM_OBJECT_LOCK(new_object);
1331 #define OBSC_TEST_ALL_SHADOWED 0x0001
1332 #define OBSC_COLLAPSE_NOWAIT 0x0002
1333 #define OBSC_COLLAPSE_WAIT 0x0004
1336 vm_object_backing_scan(vm_object_t object, int op)
1340 vm_object_t backing_object;
1341 vm_pindex_t backing_offset_index;
1343 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1344 VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1346 backing_object = object->backing_object;
1347 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1350 * Initial conditions
1352 if (op & OBSC_TEST_ALL_SHADOWED) {
1354 * We do not want to have to test for the existence of cache
1355 * or swap pages in the backing object. XXX but with the
1356 * new swapper this would be pretty easy to do.
1358 * XXX what about anonymous MAP_SHARED memory that hasn't
1359 * been ZFOD faulted yet? If we do not test for this, the
1360 * shadow test may succeed! XXX
1362 if (backing_object->type != OBJT_DEFAULT) {
1366 if (op & OBSC_COLLAPSE_WAIT) {
1367 vm_object_set_flag(backing_object, OBJ_DEAD);
1373 p = TAILQ_FIRST(&backing_object->memq);
1375 vm_page_t next = TAILQ_NEXT(p, listq);
1376 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1378 if (op & OBSC_TEST_ALL_SHADOWED) {
1382 * Ignore pages outside the parent object's range
1383 * and outside the parent object's mapping of the
1386 * note that we do not busy the backing object's
1390 p->pindex < backing_offset_index ||
1391 new_pindex >= object->size
1398 * See if the parent has the page or if the parent's
1399 * object pager has the page. If the parent has the
1400 * page but the page is not valid, the parent's
1401 * object pager must have the page.
1403 * If this fails, the parent does not completely shadow
1404 * the object and we might as well give up now.
1407 pp = vm_page_lookup(object, new_pindex);
1409 (pp == NULL || pp->valid == 0) &&
1410 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1418 * Check for busy page
1420 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1423 if (op & OBSC_COLLAPSE_NOWAIT) {
1424 if ((p->oflags & VPO_BUSY) ||
1430 } else if (op & OBSC_COLLAPSE_WAIT) {
1431 if ((p->oflags & VPO_BUSY) || p->busy) {
1432 VM_OBJECT_UNLOCK(object);
1433 p->oflags |= VPO_WANTED;
1434 msleep(p, VM_OBJECT_MTX(backing_object),
1435 PDROP | PVM, "vmocol", 0);
1436 VM_OBJECT_LOCK(object);
1437 VM_OBJECT_LOCK(backing_object);
1439 * If we slept, anything could have
1440 * happened. Since the object is
1441 * marked dead, the backing offset
1442 * should not have changed so we
1443 * just restart our scan.
1445 p = TAILQ_FIRST(&backing_object->memq);
1451 p->object == backing_object,
1452 ("vm_object_backing_scan: object mismatch")
1456 * Destroy any associated swap
1458 if (backing_object->type == OBJT_SWAP) {
1459 swap_pager_freespace(
1467 p->pindex < backing_offset_index ||
1468 new_pindex >= object->size
1471 * Page is out of the parent object's range, we
1472 * can simply destroy it.
1474 vm_page_lock_queues();
1475 KASSERT(!pmap_page_is_mapped(p),
1476 ("freeing mapped page %p", p));
1477 if (p->wire_count == 0)
1481 vm_page_unlock_queues();
1486 pp = vm_page_lookup(object, new_pindex);
1489 vm_pager_has_page(object, new_pindex, NULL, NULL)
1492 * page already exists in parent OR swap exists
1493 * for this location in the parent. Destroy
1494 * the original page from the backing object.
1496 * Leave the parent's page alone
1498 vm_page_lock_queues();
1499 KASSERT(!pmap_page_is_mapped(p),
1500 ("freeing mapped page %p", p));
1501 if (p->wire_count == 0)
1505 vm_page_unlock_queues();
1510 #if VM_NRESERVLEVEL > 0
1512 * Rename the reservation.
1514 vm_reserv_rename(p, object, backing_object,
1515 backing_offset_index);
1519 * Page does not exist in parent, rename the
1520 * page from the backing object to the main object.
1522 * If the page was mapped to a process, it can remain
1523 * mapped through the rename.
1525 vm_page_lock_queues();
1526 vm_page_rename(p, object, new_pindex);
1527 vm_page_unlock_queues();
1528 /* page automatically made dirty by rename */
1537 * this version of collapse allows the operation to occur earlier and
1538 * when paging_in_progress is true for an object... This is not a complete
1539 * operation, but should plug 99.9% of the rest of the leaks.
1542 vm_object_qcollapse(vm_object_t object)
1544 vm_object_t backing_object = object->backing_object;
1546 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1547 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1549 if (backing_object->ref_count != 1)
1552 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1556 * vm_object_collapse:
1558 * Collapse an object with the object backing it.
1559 * Pages in the backing object are moved into the
1560 * parent, and the backing object is deallocated.
1563 vm_object_collapse(vm_object_t object)
1565 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1568 vm_object_t backing_object;
1571 * Verify that the conditions are right for collapse:
1573 * The object exists and the backing object exists.
1575 if ((backing_object = object->backing_object) == NULL)
1579 * we check the backing object first, because it is most likely
1582 VM_OBJECT_LOCK(backing_object);
1583 if (backing_object->handle != NULL ||
1584 (backing_object->type != OBJT_DEFAULT &&
1585 backing_object->type != OBJT_SWAP) ||
1586 (backing_object->flags & OBJ_DEAD) ||
1587 object->handle != NULL ||
1588 (object->type != OBJT_DEFAULT &&
1589 object->type != OBJT_SWAP) ||
1590 (object->flags & OBJ_DEAD)) {
1591 VM_OBJECT_UNLOCK(backing_object);
1596 object->paging_in_progress != 0 ||
1597 backing_object->paging_in_progress != 0
1599 vm_object_qcollapse(object);
1600 VM_OBJECT_UNLOCK(backing_object);
1604 * We know that we can either collapse the backing object (if
1605 * the parent is the only reference to it) or (perhaps) have
1606 * the parent bypass the object if the parent happens to shadow
1607 * all the resident pages in the entire backing object.
1609 * This is ignoring pager-backed pages such as swap pages.
1610 * vm_object_backing_scan fails the shadowing test in this
1613 if (backing_object->ref_count == 1) {
1615 * If there is exactly one reference to the backing
1616 * object, we can collapse it into the parent.
1618 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1620 #if VM_NRESERVLEVEL > 0
1622 * Break any reservations from backing_object.
1624 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1625 vm_reserv_break_all(backing_object);
1629 * Move the pager from backing_object to object.
1631 if (backing_object->type == OBJT_SWAP) {
1633 * swap_pager_copy() can sleep, in which case
1634 * the backing_object's and object's locks are
1635 * released and reacquired.
1640 OFF_TO_IDX(object->backing_object_offset), TRUE);
1643 * Free any cached pages from backing_object.
1645 if (__predict_false(backing_object->cache != NULL))
1646 vm_page_cache_free(backing_object, 0, 0);
1649 * Object now shadows whatever backing_object did.
1650 * Note that the reference to
1651 * backing_object->backing_object moves from within
1652 * backing_object to within object.
1654 LIST_REMOVE(object, shadow_list);
1655 backing_object->shadow_count--;
1656 backing_object->generation++;
1657 if (backing_object->backing_object) {
1658 VM_OBJECT_LOCK(backing_object->backing_object);
1659 LIST_REMOVE(backing_object, shadow_list);
1661 &backing_object->backing_object->shadow_head,
1662 object, shadow_list);
1664 * The shadow_count has not changed.
1666 backing_object->backing_object->generation++;
1667 VM_OBJECT_UNLOCK(backing_object->backing_object);
1669 object->backing_object = backing_object->backing_object;
1670 object->backing_object_offset +=
1671 backing_object->backing_object_offset;
1674 * Discard backing_object.
1676 * Since the backing object has no pages, no pager left,
1677 * and no object references within it, all that is
1678 * necessary is to dispose of it.
1680 KASSERT(backing_object->ref_count == 1, (
1681 "backing_object %p was somehow re-referenced during collapse!",
1683 VM_OBJECT_UNLOCK(backing_object);
1684 vm_object_destroy(backing_object);
1688 vm_object_t new_backing_object;
1691 * If we do not entirely shadow the backing object,
1692 * there is nothing we can do so we give up.
1694 if (object->resident_page_count != object->size &&
1695 vm_object_backing_scan(object,
1696 OBSC_TEST_ALL_SHADOWED) == 0) {
1697 VM_OBJECT_UNLOCK(backing_object);
1702 * Make the parent shadow the next object in the
1703 * chain. Deallocating backing_object will not remove
1704 * it, since its reference count is at least 2.
1706 LIST_REMOVE(object, shadow_list);
1707 backing_object->shadow_count--;
1708 backing_object->generation++;
1710 new_backing_object = backing_object->backing_object;
1711 if ((object->backing_object = new_backing_object) != NULL) {
1712 VM_OBJECT_LOCK(new_backing_object);
1714 &new_backing_object->shadow_head,
1718 new_backing_object->shadow_count++;
1719 new_backing_object->generation++;
1720 vm_object_reference_locked(new_backing_object);
1721 VM_OBJECT_UNLOCK(new_backing_object);
1722 object->backing_object_offset +=
1723 backing_object->backing_object_offset;
1727 * Drop the reference count on backing_object. Since
1728 * its ref_count was at least 2, it will not vanish.
1730 backing_object->ref_count--;
1731 VM_OBJECT_UNLOCK(backing_object);
1736 * Try again with this object's new backing object.
1742 * vm_object_page_remove:
1744 * For the given object, either frees or invalidates each of the
1745 * specified pages. In general, a page is freed. However, if a
1746 * page is wired for any reason other than the existence of a
1747 * managed, wired mapping, then it may be invalidated but not
1748 * removed from the object. Pages are specified by the given
1749 * range ["start", "end") and Boolean "clean_only". As a
1750 * special case, if "end" is zero, then the range extends from
1751 * "start" to the end of the object. If "clean_only" is TRUE,
1752 * then only the non-dirty pages within the specified range are
1755 * In general, this operation should only be performed on objects
1756 * that contain managed pages. There are two exceptions. First,
1757 * it may be performed on the kernel and kmem objects. Second,
1758 * it may be used by msync(..., MS_INVALIDATE) to invalidate
1759 * device-backed pages. In both of these cases, "clean_only"
1762 * The object must be locked.
1765 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1766 boolean_t clean_only)
1771 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1772 if (object->resident_page_count == 0)
1776 * Since physically-backed objects do not use managed pages, we can't
1777 * remove pages from the object (we must instead remove the page
1778 * references, and then destroy the object).
1780 KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1781 object == kmem_object,
1782 ("attempt to remove pages from a physical object"));
1784 vm_object_pip_add(object, 1);
1786 p = vm_page_find_least(object, start);
1787 vm_page_lock_queues();
1789 * Assert: the variable p is either (1) the page with the
1790 * least pindex greater than or equal to the parameter pindex
1794 p != NULL && (p->pindex < end || end == 0);
1796 next = TAILQ_NEXT(p, listq);
1799 * If the page is wired for any reason besides the
1800 * existence of managed, wired mappings, then it cannot
1801 * be freed. For example, fictitious pages, which
1802 * represent device memory, are inherently wired and
1803 * cannot be freed. They can, however, be invalidated
1804 * if "clean_only" is FALSE.
1806 if ((wirings = p->wire_count) != 0 &&
1807 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1808 /* Fictitious pages do not have managed mappings. */
1809 if ((p->flags & PG_FICTITIOUS) == 0)
1811 /* Account for removal of managed, wired mappings. */
1812 p->wire_count -= wirings;
1819 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1821 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1822 ("vm_object_page_remove: page %p is fictitious", p));
1823 if (clean_only && p->valid) {
1824 pmap_remove_write(p);
1829 /* Account for removal of managed, wired mappings. */
1831 p->wire_count -= wirings;
1834 vm_page_unlock_queues();
1835 vm_object_pip_wakeup(object);
1837 if (__predict_false(object->cache != NULL))
1838 vm_page_cache_free(object, start, end);
1842 * Populate the specified range of the object with valid pages. Returns
1843 * TRUE if the range is successfully populated and FALSE otherwise.
1845 * Note: This function should be optimized to pass a larger array of
1846 * pages to vm_pager_get_pages() before it is applied to a non-
1847 * OBJT_DEVICE object.
1849 * The object must be locked.
1852 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1858 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1859 for (pindex = start; pindex < end; pindex++) {
1860 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1862 if (m->valid != VM_PAGE_BITS_ALL) {
1864 rv = vm_pager_get_pages(object, ma, 1, 0);
1865 m = vm_page_lookup(object, pindex);
1868 if (rv != VM_PAGER_OK) {
1869 vm_page_lock_queues();
1871 vm_page_unlock_queues();
1876 * Keep "m" busy because a subsequent iteration may unlock
1880 if (pindex > start) {
1881 m = vm_page_lookup(object, start);
1882 while (m != NULL && m->pindex < pindex) {
1884 m = TAILQ_NEXT(m, listq);
1887 return (pindex == end);
1891 * Routine: vm_object_coalesce
1892 * Function: Coalesces two objects backing up adjoining
1893 * regions of memory into a single object.
1895 * returns TRUE if objects were combined.
1897 * NOTE: Only works at the moment if the second object is NULL -
1898 * if it's not, which object do we lock first?
1901 * prev_object First object to coalesce
1902 * prev_offset Offset into prev_object
1903 * prev_size Size of reference to prev_object
1904 * next_size Size of reference to the second object
1905 * reserved Indicator that extension region has
1906 * swap accounted for
1909 * The object must *not* be locked.
1912 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
1913 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
1915 vm_pindex_t next_pindex;
1917 if (prev_object == NULL)
1919 VM_OBJECT_LOCK(prev_object);
1920 if (prev_object->type != OBJT_DEFAULT &&
1921 prev_object->type != OBJT_SWAP) {
1922 VM_OBJECT_UNLOCK(prev_object);
1927 * Try to collapse the object first
1929 vm_object_collapse(prev_object);
1932 * Can't coalesce if: . more than one reference . paged out . shadows
1933 * another object . has a copy elsewhere (any of which mean that the
1934 * pages not mapped to prev_entry may be in use anyway)
1936 if (prev_object->backing_object != NULL) {
1937 VM_OBJECT_UNLOCK(prev_object);
1941 prev_size >>= PAGE_SHIFT;
1942 next_size >>= PAGE_SHIFT;
1943 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
1945 if ((prev_object->ref_count > 1) &&
1946 (prev_object->size != next_pindex)) {
1947 VM_OBJECT_UNLOCK(prev_object);
1952 * Account for the charge.
1954 if (prev_object->uip != NULL) {
1957 * If prev_object was charged, then this mapping,
1958 * althought not charged now, may become writable
1959 * later. Non-NULL uip in the object would prevent
1960 * swap reservation during enabling of the write
1961 * access, so reserve swap now. Failed reservation
1962 * cause allocation of the separate object for the map
1963 * entry, and swap reservation for this entry is
1964 * managed in appropriate time.
1966 if (!reserved && !swap_reserve_by_uid(ptoa(next_size),
1967 prev_object->uip)) {
1970 prev_object->charge += ptoa(next_size);
1974 * Remove any pages that may still be in the object from a previous
1977 if (next_pindex < prev_object->size) {
1978 vm_object_page_remove(prev_object,
1980 next_pindex + next_size, FALSE);
1981 if (prev_object->type == OBJT_SWAP)
1982 swap_pager_freespace(prev_object,
1983 next_pindex, next_size);
1985 if (prev_object->uip != NULL) {
1986 KASSERT(prev_object->charge >=
1987 ptoa(prev_object->size - next_pindex),
1988 ("object %p overcharged 1 %jx %jx", prev_object,
1989 (uintmax_t)next_pindex, (uintmax_t)next_size));
1990 prev_object->charge -= ptoa(prev_object->size -
1997 * Extend the object if necessary.
1999 if (next_pindex + next_size > prev_object->size)
2000 prev_object->size = next_pindex + next_size;
2002 VM_OBJECT_UNLOCK(prev_object);
2007 vm_object_set_writeable_dirty(vm_object_t object)
2010 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2011 if (object->type != OBJT_VNODE ||
2012 (object->flags & OBJ_MIGHTBEDIRTY) != 0)
2014 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2017 #include "opt_ddb.h"
2019 #include <sys/kernel.h>
2021 #include <sys/cons.h>
2023 #include <ddb/ddb.h>
2026 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2029 vm_map_entry_t tmpe;
2037 tmpe = map->header.next;
2038 entcount = map->nentries;
2039 while (entcount-- && (tmpe != &map->header)) {
2040 if (_vm_object_in_map(map, object, tmpe)) {
2045 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2046 tmpm = entry->object.sub_map;
2047 tmpe = tmpm->header.next;
2048 entcount = tmpm->nentries;
2049 while (entcount-- && tmpe != &tmpm->header) {
2050 if (_vm_object_in_map(tmpm, object, tmpe)) {
2055 } else if ((obj = entry->object.vm_object) != NULL) {
2056 for (; obj; obj = obj->backing_object)
2057 if (obj == object) {
2065 vm_object_in_map(vm_object_t object)
2069 /* sx_slock(&allproc_lock); */
2070 FOREACH_PROC_IN_SYSTEM(p) {
2071 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2073 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2074 /* sx_sunlock(&allproc_lock); */
2078 /* sx_sunlock(&allproc_lock); */
2079 if (_vm_object_in_map(kernel_map, object, 0))
2081 if (_vm_object_in_map(kmem_map, object, 0))
2083 if (_vm_object_in_map(pager_map, object, 0))
2085 if (_vm_object_in_map(buffer_map, object, 0))
2090 DB_SHOW_COMMAND(vmochk, vm_object_check)
2095 * make sure that internal objs are in a map somewhere
2096 * and none have zero ref counts.
2098 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2099 if (object->handle == NULL &&
2100 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2101 if (object->ref_count == 0) {
2102 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2103 (long)object->size);
2105 if (!vm_object_in_map(object)) {
2107 "vmochk: internal obj is not in a map: "
2108 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2109 object->ref_count, (u_long)object->size,
2110 (u_long)object->size,
2111 (void *)object->backing_object);
2118 * vm_object_print: [ debug ]
2120 DB_SHOW_COMMAND(object, vm_object_print_static)
2122 /* XXX convert args. */
2123 vm_object_t object = (vm_object_t)addr;
2124 boolean_t full = have_addr;
2128 /* XXX count is an (unused) arg. Avoid shadowing it. */
2129 #define count was_count
2137 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x uip %d charge %jx\n",
2138 object, (int)object->type, (uintmax_t)object->size,
2139 object->resident_page_count, object->ref_count, object->flags,
2140 object->uip ? object->uip->ui_uid : -1, (uintmax_t)object->charge);
2141 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2142 object->shadow_count,
2143 object->backing_object ? object->backing_object->ref_count : 0,
2144 object->backing_object, (uintmax_t)object->backing_object_offset);
2151 TAILQ_FOREACH(p, &object->memq, listq) {
2153 db_iprintf("memory:=");
2154 else if (count == 6) {
2162 db_printf("(off=0x%jx,page=0x%jx)",
2163 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2173 /* XXX need this non-static entry for calling from vm_map_print. */
2176 /* db_expr_t */ long addr,
2177 boolean_t have_addr,
2178 /* db_expr_t */ long count,
2181 vm_object_print_static(addr, have_addr, count, modif);
2184 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2189 vm_page_t m, prev_m;
2193 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2194 db_printf("new object: %p\n", (void *)object);
2205 TAILQ_FOREACH(m, &object->memq, listq) {
2206 if (m->pindex > 128)
2208 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2209 prev_m->pindex + 1 != m->pindex) {
2211 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2212 (long)fidx, rcount, (long)pa);
2224 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2229 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2230 (long)fidx, rcount, (long)pa);
2240 pa = VM_PAGE_TO_PHYS(m);
2244 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2245 (long)fidx, rcount, (long)pa);