2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94
35 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36 * All rights reserved.
38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
40 * Permission to use, copy, modify and distribute this software and
41 * its documentation is hereby granted, provided that both the copyright
42 * notice and this permission notice appear in all copies of the
43 * software, derivative works or modified versions, and any portions
44 * thereof, and that both notices appear in supporting documentation.
46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
50 * Carnegie Mellon requests users of this software to return to
52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
53 * School of Computer Science
54 * Carnegie Mellon University
55 * Pittsburgh PA 15213-3890
57 * any improvements or extensions that they make and grant Carnegie the
58 * rights to redistribute these changes.
62 * Virtual memory object module.
65 #include <sys/cdefs.h>
66 __FBSDID("$FreeBSD$");
70 #include <sys/param.h>
71 #include <sys/systm.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/mutex.h>
78 #include <sys/proc.h> /* for curproc, pageproc */
79 #include <sys/socket.h>
80 #include <sys/resourcevar.h>
81 #include <sys/vnode.h>
82 #include <sys/vmmeter.h>
86 #include <vm/vm_param.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_pager.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_reserv.h>
100 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
101 "Use old (insecure) msync behavior");
103 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
104 int pagerflags, int flags, int *clearobjflags);
105 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
107 static void vm_object_qcollapse(vm_object_t object);
108 static void vm_object_vndeallocate(vm_object_t object);
111 * Virtual memory objects maintain the actual data
112 * associated with allocated virtual memory. A given
113 * page of memory exists within exactly one object.
115 * An object is only deallocated when all "references"
116 * are given up. Only one "reference" to a given
117 * region of an object should be writeable.
119 * Associated with each object is a list of all resident
120 * memory pages belonging to that object; this list is
121 * maintained by the "vm_page" module, and locked by the object's
124 * Each object also records a "pager" routine which is
125 * used to retrieve (and store) pages to the proper backing
126 * storage. In addition, objects may be backed by other
127 * objects from which they were virtual-copied.
129 * The only items within the object structure which are
130 * modified after time of creation are:
131 * reference count locked by object's lock
132 * pager routine locked by object's lock
136 struct object_q vm_object_list;
137 struct mtx vm_object_list_mtx; /* lock for object list and count */
139 struct vm_object kernel_object_store;
140 struct vm_object kmem_object_store;
142 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
145 static long object_collapses;
146 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
147 &object_collapses, 0, "VM object collapses");
149 static long object_bypasses;
150 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
151 &object_bypasses, 0, "VM object bypasses");
153 static uma_zone_t obj_zone;
155 static int vm_object_zinit(void *mem, int size, int flags);
158 static void vm_object_zdtor(void *mem, int size, void *arg);
161 vm_object_zdtor(void *mem, int size, void *arg)
165 object = (vm_object_t)mem;
166 KASSERT(TAILQ_EMPTY(&object->memq),
167 ("object %p has resident pages",
169 #if VM_NRESERVLEVEL > 0
170 KASSERT(LIST_EMPTY(&object->rvq),
171 ("object %p has reservations",
174 KASSERT(object->cache == NULL,
175 ("object %p has cached pages",
177 KASSERT(object->paging_in_progress == 0,
178 ("object %p paging_in_progress = %d",
179 object, object->paging_in_progress));
180 KASSERT(object->resident_page_count == 0,
181 ("object %p resident_page_count = %d",
182 object, object->resident_page_count));
183 KASSERT(object->shadow_count == 0,
184 ("object %p shadow_count = %d",
185 object, object->shadow_count));
190 vm_object_zinit(void *mem, int size, int flags)
194 object = (vm_object_t)mem;
195 bzero(&object->mtx, sizeof(object->mtx));
196 VM_OBJECT_LOCK_INIT(object, "standard object");
198 /* These are true for any object that has been freed */
199 object->paging_in_progress = 0;
200 object->resident_page_count = 0;
201 object->shadow_count = 0;
206 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
209 TAILQ_INIT(&object->memq);
210 LIST_INIT(&object->shadow_head);
215 object->generation = 1;
216 object->ref_count = 1;
217 object->memattr = VM_MEMATTR_DEFAULT;
221 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
222 object->flags = OBJ_ONEMAPPING;
223 object->pg_color = 0;
224 object->handle = NULL;
225 object->backing_object = NULL;
226 object->backing_object_offset = (vm_ooffset_t) 0;
227 #if VM_NRESERVLEVEL > 0
228 LIST_INIT(&object->rvq);
230 object->cache = NULL;
232 mtx_lock(&vm_object_list_mtx);
233 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
234 mtx_unlock(&vm_object_list_mtx);
240 * Initialize the VM objects module.
245 TAILQ_INIT(&vm_object_list);
246 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
248 VM_OBJECT_LOCK_INIT(kernel_object, "kernel object");
249 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
251 #if VM_NRESERVLEVEL > 0
252 kernel_object->flags |= OBJ_COLORED;
253 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
256 VM_OBJECT_LOCK_INIT(kmem_object, "kmem object");
257 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
259 #if VM_NRESERVLEVEL > 0
260 kmem_object->flags |= OBJ_COLORED;
261 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
265 * The lock portion of struct vm_object must be type stable due
266 * to vm_pageout_fallback_object_lock locking a vm object
267 * without holding any references to it.
269 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
275 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
279 vm_object_clear_flag(vm_object_t object, u_short bits)
282 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
283 object->flags &= ~bits;
287 * Sets the default memory attribute for the specified object. Pages
288 * that are allocated to this object are by default assigned this memory
291 * Presently, this function must be called before any pages are allocated
292 * to the object. In the future, this requirement may be relaxed for
293 * "default" and "swap" objects.
296 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
299 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
300 switch (object->type) {
307 if (!TAILQ_EMPTY(&object->memq))
308 return (KERN_FAILURE);
311 return (KERN_INVALID_ARGUMENT);
313 object->memattr = memattr;
314 return (KERN_SUCCESS);
318 vm_object_pip_add(vm_object_t object, short i)
321 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
322 object->paging_in_progress += i;
326 vm_object_pip_subtract(vm_object_t object, short i)
329 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
330 object->paging_in_progress -= i;
334 vm_object_pip_wakeup(vm_object_t object)
337 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
338 object->paging_in_progress--;
339 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
340 vm_object_clear_flag(object, OBJ_PIPWNT);
346 vm_object_pip_wakeupn(vm_object_t object, short i)
349 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
351 object->paging_in_progress -= i;
352 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
353 vm_object_clear_flag(object, OBJ_PIPWNT);
359 vm_object_pip_wait(vm_object_t object, char *waitid)
362 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
363 while (object->paging_in_progress) {
364 object->flags |= OBJ_PIPWNT;
365 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
370 * vm_object_allocate:
372 * Returns a new object with the given size.
375 vm_object_allocate(objtype_t type, vm_pindex_t size)
379 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
380 _vm_object_allocate(type, size, object);
386 * vm_object_reference:
388 * Gets another reference to the given object. Note: OBJ_DEAD
389 * objects can be referenced during final cleaning.
392 vm_object_reference(vm_object_t object)
396 VM_OBJECT_LOCK(object);
397 vm_object_reference_locked(object);
398 VM_OBJECT_UNLOCK(object);
402 * vm_object_reference_locked:
404 * Gets another reference to the given object.
406 * The object must be locked.
409 vm_object_reference_locked(vm_object_t object)
413 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
415 if (object->type == OBJT_VNODE) {
422 * Handle deallocating an object of type OBJT_VNODE.
425 vm_object_vndeallocate(vm_object_t object)
427 struct vnode *vp = (struct vnode *) object->handle;
429 VFS_ASSERT_GIANT(vp->v_mount);
430 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
431 KASSERT(object->type == OBJT_VNODE,
432 ("vm_object_vndeallocate: not a vnode object"));
433 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
435 if (object->ref_count == 0) {
436 vprint("vm_object_vndeallocate", vp);
437 panic("vm_object_vndeallocate: bad object reference count");
441 if (object->ref_count > 1) {
443 VM_OBJECT_UNLOCK(object);
444 /* vrele may need the vnode lock. */
448 VM_OBJECT_UNLOCK(object);
449 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
451 VM_OBJECT_LOCK(object);
453 if (object->type == OBJT_DEAD) {
454 VM_OBJECT_UNLOCK(object);
457 if (object->ref_count == 0)
458 vp->v_vflag &= ~VV_TEXT;
459 VM_OBJECT_UNLOCK(object);
466 * vm_object_deallocate:
468 * Release a reference to the specified object,
469 * gained either through a vm_object_allocate
470 * or a vm_object_reference call. When all references
471 * are gone, storage associated with this object
472 * may be relinquished.
474 * No object may be locked.
477 vm_object_deallocate(vm_object_t object)
481 while (object != NULL) {
486 VM_OBJECT_LOCK(object);
487 if (object->type == OBJT_VNODE) {
488 struct vnode *vp = (struct vnode *) object->handle;
491 * Conditionally acquire Giant for a vnode-backed
492 * object. We have to be careful since the type of
493 * a vnode object can change while the object is
496 if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
498 if (!mtx_trylock(&Giant)) {
499 VM_OBJECT_UNLOCK(object);
504 vm_object_vndeallocate(object);
505 VFS_UNLOCK_GIANT(vfslocked);
509 * This is to handle the case that the object
510 * changed type while we dropped its lock to
513 VFS_UNLOCK_GIANT(vfslocked);
515 KASSERT(object->ref_count != 0,
516 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
519 * If the reference count goes to 0 we start calling
520 * vm_object_terminate() on the object chain.
521 * A ref count of 1 may be a special case depending on the
522 * shadow count being 0 or 1.
525 if (object->ref_count > 1) {
526 VM_OBJECT_UNLOCK(object);
528 } else if (object->ref_count == 1) {
529 if (object->shadow_count == 0 &&
530 object->handle == NULL &&
531 (object->type == OBJT_DEFAULT ||
532 object->type == OBJT_SWAP)) {
533 vm_object_set_flag(object, OBJ_ONEMAPPING);
534 } else if ((object->shadow_count == 1) &&
535 (object->handle == NULL) &&
536 (object->type == OBJT_DEFAULT ||
537 object->type == OBJT_SWAP)) {
540 robject = LIST_FIRST(&object->shadow_head);
541 KASSERT(robject != NULL,
542 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
544 object->shadow_count));
545 if (!VM_OBJECT_TRYLOCK(robject)) {
547 * Avoid a potential deadlock.
550 VM_OBJECT_UNLOCK(object);
552 * More likely than not the thread
553 * holding robject's lock has lower
554 * priority than the current thread.
555 * Let the lower priority thread run.
561 * Collapse object into its shadow unless its
562 * shadow is dead. In that case, object will
563 * be deallocated by the thread that is
564 * deallocating its shadow.
566 if ((robject->flags & OBJ_DEAD) == 0 &&
567 (robject->handle == NULL) &&
568 (robject->type == OBJT_DEFAULT ||
569 robject->type == OBJT_SWAP)) {
571 robject->ref_count++;
573 if (robject->paging_in_progress) {
574 VM_OBJECT_UNLOCK(object);
575 vm_object_pip_wait(robject,
577 temp = robject->backing_object;
578 if (object == temp) {
579 VM_OBJECT_LOCK(object);
582 } else if (object->paging_in_progress) {
583 VM_OBJECT_UNLOCK(robject);
584 object->flags |= OBJ_PIPWNT;
586 VM_OBJECT_MTX(object),
587 PDROP | PVM, "objde2", 0);
588 VM_OBJECT_LOCK(robject);
589 temp = robject->backing_object;
590 if (object == temp) {
591 VM_OBJECT_LOCK(object);
595 VM_OBJECT_UNLOCK(object);
597 if (robject->ref_count == 1) {
598 robject->ref_count--;
603 vm_object_collapse(object);
604 VM_OBJECT_UNLOCK(object);
607 VM_OBJECT_UNLOCK(robject);
609 VM_OBJECT_UNLOCK(object);
613 temp = object->backing_object;
615 VM_OBJECT_LOCK(temp);
616 LIST_REMOVE(object, shadow_list);
617 temp->shadow_count--;
618 VM_OBJECT_UNLOCK(temp);
619 object->backing_object = NULL;
622 * Don't double-terminate, we could be in a termination
623 * recursion due to the terminate having to sync data
626 if ((object->flags & OBJ_DEAD) == 0)
627 vm_object_terminate(object);
629 VM_OBJECT_UNLOCK(object);
635 * vm_object_destroy removes the object from the global object list
636 * and frees the space for the object.
639 vm_object_destroy(vm_object_t object)
643 * Remove the object from the global object list.
645 mtx_lock(&vm_object_list_mtx);
646 TAILQ_REMOVE(&vm_object_list, object, object_list);
647 mtx_unlock(&vm_object_list_mtx);
650 * Release the allocation charge.
652 if (object->cred != NULL) {
653 KASSERT(object->type == OBJT_DEFAULT ||
654 object->type == OBJT_SWAP,
655 ("vm_object_terminate: non-swap obj %p has cred",
657 swap_release_by_cred(object->charge, object->cred);
659 crfree(object->cred);
664 * Free the space for the object.
666 uma_zfree(obj_zone, object);
670 * vm_object_terminate actually destroys the specified object, freeing
671 * up all previously used resources.
673 * The object must be locked.
674 * This routine may block.
677 vm_object_terminate(vm_object_t object)
681 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
684 * Make sure no one uses us.
686 vm_object_set_flag(object, OBJ_DEAD);
689 * wait for the pageout daemon to be done with the object
691 vm_object_pip_wait(object, "objtrm");
693 KASSERT(!object->paging_in_progress,
694 ("vm_object_terminate: pageout in progress"));
697 * Clean and free the pages, as appropriate. All references to the
698 * object are gone, so we don't need to lock it.
700 if (object->type == OBJT_VNODE) {
701 struct vnode *vp = (struct vnode *)object->handle;
704 * Clean pages and flush buffers.
706 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
707 VM_OBJECT_UNLOCK(object);
709 vinvalbuf(vp, V_SAVE, 0, 0);
711 VM_OBJECT_LOCK(object);
714 KASSERT(object->ref_count == 0,
715 ("vm_object_terminate: object with references, ref_count=%d",
719 * Free any remaining pageable pages. This also removes them from the
720 * paging queues. However, don't free wired pages, just remove them
721 * from the object. Rather than incrementally removing each page from
722 * the object, the page and object are reset to any empty state.
724 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
725 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
726 ("vm_object_terminate: freeing busy page %p", p));
729 * Optimize the page's removal from the object by resetting
730 * its "object" field. Specifically, if the page is not
731 * wired, then the effect of this assignment is that
732 * vm_page_free()'s call to vm_page_remove() will return
733 * immediately without modifying the page or the object.
736 if (p->wire_count == 0) {
738 PCPU_INC(cnt.v_pfree);
743 * If the object contained any pages, then reset it to an empty state.
744 * None of the object's fields, including "resident_page_count", were
745 * modified by the preceding loop.
747 if (object->resident_page_count != 0) {
749 TAILQ_INIT(&object->memq);
750 object->resident_page_count = 0;
751 if (object->type == OBJT_VNODE)
752 vdrop(object->handle);
755 #if VM_NRESERVLEVEL > 0
756 if (__predict_false(!LIST_EMPTY(&object->rvq)))
757 vm_reserv_break_all(object);
759 if (__predict_false(object->cache != NULL))
760 vm_page_cache_free(object, 0, 0);
763 * Let the pager know object is dead.
765 vm_pager_deallocate(object);
766 VM_OBJECT_UNLOCK(object);
768 vm_object_destroy(object);
772 * Make the page read-only so that we can clear the object flags. However, if
773 * this is a nosync mmap then the object is likely to stay dirty so do not
774 * mess with the page and do not clear the object flags. Returns TRUE if the
775 * page should be flushed, and FALSE otherwise.
778 vm_object_page_remove_write(vm_page_t p, int flags, int *clearobjflags)
782 * If we have been asked to skip nosync pages and this is a
783 * nosync page, skip it. Note that the object flags were not
784 * cleared in this case so we do not have to set them.
786 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
790 pmap_remove_write(p);
791 return (p->dirty != 0);
796 * vm_object_page_clean
798 * Clean all dirty pages in the specified range of object. Leaves page
799 * on whatever queue it is currently on. If NOSYNC is set then do not
800 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
801 * leaving the object dirty.
803 * When stuffing pages asynchronously, allow clustering. XXX we need a
804 * synchronous clustering mode implementation.
806 * Odd semantics: if start == end, we clean everything.
808 * The object must be locked.
811 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
815 vm_pindex_t pi, tend, tstart;
816 int clearobjflags, curgeneration, n, pagerflags;
818 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
819 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
820 KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
821 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
822 object->resident_page_count == 0)
825 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
826 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
827 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
829 tstart = OFF_TO_IDX(start);
830 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
831 clearobjflags = tstart == 0 && tend >= object->size;
834 curgeneration = object->generation;
836 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
840 np = TAILQ_NEXT(p, listq);
843 if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
844 if (object->generation != curgeneration) {
845 if ((flags & OBJPC_SYNC) != 0)
850 np = vm_page_find_least(object, pi);
853 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
856 n = vm_object_page_collect_flush(object, p, pagerflags,
857 flags, &clearobjflags);
858 if (object->generation != curgeneration) {
859 if ((flags & OBJPC_SYNC) != 0)
866 * If the VOP_PUTPAGES() did a truncated write, so
867 * that even the first page of the run is not fully
868 * written, vm_pageout_flush() returns 0 as the run
869 * length. Since the condition that caused truncated
870 * write may be permanent, e.g. exhausted free space,
871 * accepting n == 0 would cause an infinite loop.
873 * Forwarding the iterator leaves the unwritten page
874 * behind, but there is not much we can do there if
875 * filesystem refuses to write it.
879 np = vm_page_find_least(object, pi + n);
882 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
886 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
890 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
891 int flags, int *clearobjflags)
893 vm_page_t ma[vm_pageout_page_count], p_first, tp;
894 int count, i, mreq, runlen;
896 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
897 vm_page_lock_assert(p, MA_NOTOWNED);
898 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
903 for (tp = p; count < vm_pageout_page_count; count++) {
904 tp = vm_page_next(tp);
905 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
907 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
911 for (p_first = p; count < vm_pageout_page_count; count++) {
912 tp = vm_page_prev(p_first);
913 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
915 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
921 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
924 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen);
929 * Note that there is absolutely no sense in writing out
930 * anonymous objects, so we track down the vnode object
932 * We invalidate (remove) all pages from the address space
933 * for semantic correctness.
935 * If the backing object is a device object with unmanaged pages, then any
936 * mappings to the specified range of pages must be removed before this
937 * function is called.
939 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
940 * may start out with a NULL object.
943 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
944 boolean_t syncio, boolean_t invalidate)
946 vm_object_t backing_object;
949 int flags, fsync_after;
953 VM_OBJECT_LOCK(object);
954 while ((backing_object = object->backing_object) != NULL) {
955 VM_OBJECT_LOCK(backing_object);
956 offset += object->backing_object_offset;
957 VM_OBJECT_UNLOCK(object);
958 object = backing_object;
959 if (object->size < OFF_TO_IDX(offset + size))
960 size = IDX_TO_OFF(object->size) - offset;
963 * Flush pages if writing is allowed, invalidate them
964 * if invalidation requested. Pages undergoing I/O
965 * will be ignored by vm_object_page_remove().
967 * We cannot lock the vnode and then wait for paging
968 * to complete without deadlocking against vm_fault.
969 * Instead we simply call vm_object_page_remove() and
970 * allow it to block internally on a page-by-page
971 * basis when it encounters pages undergoing async
974 if (object->type == OBJT_VNODE &&
975 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
978 VM_OBJECT_UNLOCK(object);
979 (void) vn_start_write(vp, &mp, V_WAIT);
980 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
981 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
982 if (syncio && !invalidate && offset == 0 &&
983 OFF_TO_IDX(size) == object->size) {
985 * If syncing the whole mapping of the file,
986 * it is faster to schedule all the writes in
987 * async mode, also allowing the clustering,
988 * and then wait for i/o to complete.
993 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
994 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
997 VM_OBJECT_LOCK(object);
998 vm_object_page_clean(object, offset, offset + size, flags);
999 VM_OBJECT_UNLOCK(object);
1001 (void) VOP_FSYNC(vp, MNT_WAIT, curthread);
1003 VFS_UNLOCK_GIANT(vfslocked);
1004 vn_finished_write(mp);
1005 VM_OBJECT_LOCK(object);
1007 if ((object->type == OBJT_VNODE ||
1008 object->type == OBJT_DEVICE) && invalidate) {
1009 if (object->type == OBJT_DEVICE)
1011 * The option OBJPR_NOTMAPPED must be passed here
1012 * because vm_object_page_remove() cannot remove
1013 * unmanaged mappings.
1015 flags = OBJPR_NOTMAPPED;
1019 flags = OBJPR_CLEANONLY;
1020 vm_object_page_remove(object, OFF_TO_IDX(offset),
1021 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1023 VM_OBJECT_UNLOCK(object);
1027 * vm_object_madvise:
1029 * Implements the madvise function at the object/page level.
1031 * MADV_WILLNEED (any object)
1033 * Activate the specified pages if they are resident.
1035 * MADV_DONTNEED (any object)
1037 * Deactivate the specified pages if they are resident.
1039 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1040 * OBJ_ONEMAPPING only)
1042 * Deactivate and clean the specified pages if they are
1043 * resident. This permits the process to reuse the pages
1044 * without faulting or the kernel to reclaim the pages
1048 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1050 vm_pindex_t end, tpindex;
1051 vm_object_t backing_object, tobject;
1056 VM_OBJECT_LOCK(object);
1057 end = pindex + count;
1059 * Locate and adjust resident pages
1061 for (; pindex < end; pindex += 1) {
1067 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1068 * and those pages must be OBJ_ONEMAPPING.
1070 if (advise == MADV_FREE) {
1071 if ((tobject->type != OBJT_DEFAULT &&
1072 tobject->type != OBJT_SWAP) ||
1073 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1074 goto unlock_tobject;
1076 } else if (tobject->type == OBJT_PHYS)
1077 goto unlock_tobject;
1078 m = vm_page_lookup(tobject, tpindex);
1079 if (m == NULL && advise == MADV_WILLNEED) {
1081 * If the page is cached, reactivate it.
1083 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1088 * There may be swap even if there is no backing page
1090 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1091 swap_pager_freespace(tobject, tpindex, 1);
1095 backing_object = tobject->backing_object;
1096 if (backing_object == NULL)
1097 goto unlock_tobject;
1098 VM_OBJECT_LOCK(backing_object);
1099 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1100 if (tobject != object)
1101 VM_OBJECT_UNLOCK(tobject);
1102 tobject = backing_object;
1104 } else if (m->valid != VM_PAGE_BITS_ALL)
1105 goto unlock_tobject;
1107 * If the page is not in a normal state, skip it.
1110 if (m->hold_count != 0 || m->wire_count != 0) {
1112 goto unlock_tobject;
1114 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1115 ("vm_object_madvise: page %p is fictitious", m));
1116 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1117 ("vm_object_madvise: page %p is not managed", m));
1118 if ((m->oflags & VPO_BUSY) || m->busy) {
1119 if (advise == MADV_WILLNEED) {
1121 * Reference the page before unlocking and
1122 * sleeping so that the page daemon is less
1123 * likely to reclaim it.
1125 vm_page_aflag_set(m, PGA_REFERENCED);
1128 if (object != tobject)
1129 VM_OBJECT_UNLOCK(object);
1130 m->oflags |= VPO_WANTED;
1131 msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1133 VM_OBJECT_LOCK(object);
1136 if (advise == MADV_WILLNEED) {
1137 vm_page_activate(m);
1138 } else if (advise == MADV_DONTNEED) {
1139 vm_page_dontneed(m);
1140 } else if (advise == MADV_FREE) {
1142 * Mark the page clean. This will allow the page
1143 * to be freed up by the system. However, such pages
1144 * are often reused quickly by malloc()/free()
1145 * so we do not do anything that would cause
1146 * a page fault if we can help it.
1148 * Specifically, we do not try to actually free
1149 * the page now nor do we try to put it in the
1150 * cache (which would cause a page fault on reuse).
1152 * But we do make the page is freeable as we
1153 * can without actually taking the step of unmapping
1156 pmap_clear_modify(m);
1159 vm_page_dontneed(m);
1162 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1163 swap_pager_freespace(tobject, tpindex, 1);
1165 if (tobject != object)
1166 VM_OBJECT_UNLOCK(tobject);
1168 VM_OBJECT_UNLOCK(object);
1174 * Create a new object which is backed by the
1175 * specified existing object range. The source
1176 * object reference is deallocated.
1178 * The new object and offset into that object
1179 * are returned in the source parameters.
1183 vm_object_t *object, /* IN/OUT */
1184 vm_ooffset_t *offset, /* IN/OUT */
1193 * Don't create the new object if the old object isn't shared.
1195 if (source != NULL) {
1196 VM_OBJECT_LOCK(source);
1197 if (source->ref_count == 1 &&
1198 source->handle == NULL &&
1199 (source->type == OBJT_DEFAULT ||
1200 source->type == OBJT_SWAP)) {
1201 VM_OBJECT_UNLOCK(source);
1204 VM_OBJECT_UNLOCK(source);
1208 * Allocate a new object with the given length.
1210 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1213 * The new object shadows the source object, adding a reference to it.
1214 * Our caller changes his reference to point to the new object,
1215 * removing a reference to the source object. Net result: no change
1216 * of reference count.
1218 * Try to optimize the result object's page color when shadowing
1219 * in order to maintain page coloring consistency in the combined
1222 result->backing_object = source;
1224 * Store the offset into the source object, and fix up the offset into
1227 result->backing_object_offset = *offset;
1228 if (source != NULL) {
1229 VM_OBJECT_LOCK(source);
1230 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1231 source->shadow_count++;
1232 #if VM_NRESERVLEVEL > 0
1233 result->flags |= source->flags & OBJ_COLORED;
1234 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1235 ((1 << (VM_NFREEORDER - 1)) - 1);
1237 VM_OBJECT_UNLOCK(source);
1242 * Return the new things
1251 * Split the pages in a map entry into a new object. This affords
1252 * easier removal of unused pages, and keeps object inheritance from
1253 * being a negative impact on memory usage.
1256 vm_object_split(vm_map_entry_t entry)
1258 vm_page_t m, m_next;
1259 vm_object_t orig_object, new_object, source;
1260 vm_pindex_t idx, offidxstart;
1263 orig_object = entry->object.vm_object;
1264 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1266 if (orig_object->ref_count <= 1)
1268 VM_OBJECT_UNLOCK(orig_object);
1270 offidxstart = OFF_TO_IDX(entry->offset);
1271 size = atop(entry->end - entry->start);
1274 * If swap_pager_copy() is later called, it will convert new_object
1275 * into a swap object.
1277 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1280 * At this point, the new object is still private, so the order in
1281 * which the original and new objects are locked does not matter.
1283 VM_OBJECT_LOCK(new_object);
1284 VM_OBJECT_LOCK(orig_object);
1285 source = orig_object->backing_object;
1286 if (source != NULL) {
1287 VM_OBJECT_LOCK(source);
1288 if ((source->flags & OBJ_DEAD) != 0) {
1289 VM_OBJECT_UNLOCK(source);
1290 VM_OBJECT_UNLOCK(orig_object);
1291 VM_OBJECT_UNLOCK(new_object);
1292 vm_object_deallocate(new_object);
1293 VM_OBJECT_LOCK(orig_object);
1296 LIST_INSERT_HEAD(&source->shadow_head,
1297 new_object, shadow_list);
1298 source->shadow_count++;
1299 vm_object_reference_locked(source); /* for new_object */
1300 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1301 VM_OBJECT_UNLOCK(source);
1302 new_object->backing_object_offset =
1303 orig_object->backing_object_offset + entry->offset;
1304 new_object->backing_object = source;
1306 if (orig_object->cred != NULL) {
1307 new_object->cred = orig_object->cred;
1308 crhold(orig_object->cred);
1309 new_object->charge = ptoa(size);
1310 KASSERT(orig_object->charge >= ptoa(size),
1311 ("orig_object->charge < 0"));
1312 orig_object->charge -= ptoa(size);
1315 m = vm_page_find_least(orig_object, offidxstart);
1316 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1318 m_next = TAILQ_NEXT(m, listq);
1321 * We must wait for pending I/O to complete before we can
1324 * We do not have to VM_PROT_NONE the page as mappings should
1325 * not be changed by this operation.
1327 if ((m->oflags & VPO_BUSY) || m->busy) {
1328 VM_OBJECT_UNLOCK(new_object);
1329 m->oflags |= VPO_WANTED;
1330 msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1331 VM_OBJECT_LOCK(new_object);
1334 #if VM_NRESERVLEVEL > 0
1336 * If some of the reservation's allocated pages remain with
1337 * the original object, then transferring the reservation to
1338 * the new object is neither particularly beneficial nor
1339 * particularly harmful as compared to leaving the reservation
1340 * with the original object. If, however, all of the
1341 * reservation's allocated pages are transferred to the new
1342 * object, then transferring the reservation is typically
1343 * beneficial. Determining which of these two cases applies
1344 * would be more costly than unconditionally renaming the
1347 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1350 vm_page_rename(m, new_object, idx);
1352 /* page automatically made dirty by rename and cache handled */
1355 if (orig_object->type == OBJT_SWAP) {
1357 * swap_pager_copy() can sleep, in which case the orig_object's
1358 * and new_object's locks are released and reacquired.
1360 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1363 * Transfer any cached pages from orig_object to new_object.
1365 if (__predict_false(orig_object->cache != NULL))
1366 vm_page_cache_transfer(orig_object, offidxstart,
1369 VM_OBJECT_UNLOCK(orig_object);
1370 TAILQ_FOREACH(m, &new_object->memq, listq)
1372 VM_OBJECT_UNLOCK(new_object);
1373 entry->object.vm_object = new_object;
1374 entry->offset = 0LL;
1375 vm_object_deallocate(orig_object);
1376 VM_OBJECT_LOCK(new_object);
1379 #define OBSC_TEST_ALL_SHADOWED 0x0001
1380 #define OBSC_COLLAPSE_NOWAIT 0x0002
1381 #define OBSC_COLLAPSE_WAIT 0x0004
1384 vm_object_backing_scan(vm_object_t object, int op)
1388 vm_object_t backing_object;
1389 vm_pindex_t backing_offset_index;
1391 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1392 VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1394 backing_object = object->backing_object;
1395 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1398 * Initial conditions
1400 if (op & OBSC_TEST_ALL_SHADOWED) {
1402 * We do not want to have to test for the existence of cache
1403 * or swap pages in the backing object. XXX but with the
1404 * new swapper this would be pretty easy to do.
1406 * XXX what about anonymous MAP_SHARED memory that hasn't
1407 * been ZFOD faulted yet? If we do not test for this, the
1408 * shadow test may succeed! XXX
1410 if (backing_object->type != OBJT_DEFAULT) {
1414 if (op & OBSC_COLLAPSE_WAIT) {
1415 vm_object_set_flag(backing_object, OBJ_DEAD);
1421 p = TAILQ_FIRST(&backing_object->memq);
1423 vm_page_t next = TAILQ_NEXT(p, listq);
1424 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1426 if (op & OBSC_TEST_ALL_SHADOWED) {
1430 * Ignore pages outside the parent object's range
1431 * and outside the parent object's mapping of the
1434 * note that we do not busy the backing object's
1438 p->pindex < backing_offset_index ||
1439 new_pindex >= object->size
1446 * See if the parent has the page or if the parent's
1447 * object pager has the page. If the parent has the
1448 * page but the page is not valid, the parent's
1449 * object pager must have the page.
1451 * If this fails, the parent does not completely shadow
1452 * the object and we might as well give up now.
1455 pp = vm_page_lookup(object, new_pindex);
1457 (pp == NULL || pp->valid == 0) &&
1458 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1466 * Check for busy page
1468 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1471 if (op & OBSC_COLLAPSE_NOWAIT) {
1472 if ((p->oflags & VPO_BUSY) ||
1478 } else if (op & OBSC_COLLAPSE_WAIT) {
1479 if ((p->oflags & VPO_BUSY) || p->busy) {
1480 VM_OBJECT_UNLOCK(object);
1481 p->oflags |= VPO_WANTED;
1482 msleep(p, VM_OBJECT_MTX(backing_object),
1483 PDROP | PVM, "vmocol", 0);
1484 VM_OBJECT_LOCK(object);
1485 VM_OBJECT_LOCK(backing_object);
1487 * If we slept, anything could have
1488 * happened. Since the object is
1489 * marked dead, the backing offset
1490 * should not have changed so we
1491 * just restart our scan.
1493 p = TAILQ_FIRST(&backing_object->memq);
1499 p->object == backing_object,
1500 ("vm_object_backing_scan: object mismatch")
1504 * Destroy any associated swap
1506 if (backing_object->type == OBJT_SWAP) {
1507 swap_pager_freespace(
1515 p->pindex < backing_offset_index ||
1516 new_pindex >= object->size
1519 * Page is out of the parent object's range, we
1520 * can simply destroy it.
1523 KASSERT(!pmap_page_is_mapped(p),
1524 ("freeing mapped page %p", p));
1525 if (p->wire_count == 0)
1534 pp = vm_page_lookup(object, new_pindex);
1536 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1537 (pp != NULL && pp->valid == 0)
1540 * The page in the parent is not (yet) valid.
1541 * We don't know anything about the state of
1542 * the original page. It might be mapped,
1543 * so we must avoid the next if here.
1545 * This is due to a race in vm_fault() where
1546 * we must unbusy the original (backing_obj)
1547 * page before we can (re)lock the parent.
1548 * Hence we can get here.
1555 vm_pager_has_page(object, new_pindex, NULL, NULL)
1558 * page already exists in parent OR swap exists
1559 * for this location in the parent. Destroy
1560 * the original page from the backing object.
1562 * Leave the parent's page alone
1565 KASSERT(!pmap_page_is_mapped(p),
1566 ("freeing mapped page %p", p));
1567 if (p->wire_count == 0)
1576 #if VM_NRESERVLEVEL > 0
1578 * Rename the reservation.
1580 vm_reserv_rename(p, object, backing_object,
1581 backing_offset_index);
1585 * Page does not exist in parent, rename the
1586 * page from the backing object to the main object.
1588 * If the page was mapped to a process, it can remain
1589 * mapped through the rename.
1592 vm_page_rename(p, object, new_pindex);
1594 /* page automatically made dirty by rename */
1603 * this version of collapse allows the operation to occur earlier and
1604 * when paging_in_progress is true for an object... This is not a complete
1605 * operation, but should plug 99.9% of the rest of the leaks.
1608 vm_object_qcollapse(vm_object_t object)
1610 vm_object_t backing_object = object->backing_object;
1612 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1613 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1615 if (backing_object->ref_count != 1)
1618 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1622 * vm_object_collapse:
1624 * Collapse an object with the object backing it.
1625 * Pages in the backing object are moved into the
1626 * parent, and the backing object is deallocated.
1629 vm_object_collapse(vm_object_t object)
1631 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1634 vm_object_t backing_object;
1637 * Verify that the conditions are right for collapse:
1639 * The object exists and the backing object exists.
1641 if ((backing_object = object->backing_object) == NULL)
1645 * we check the backing object first, because it is most likely
1648 VM_OBJECT_LOCK(backing_object);
1649 if (backing_object->handle != NULL ||
1650 (backing_object->type != OBJT_DEFAULT &&
1651 backing_object->type != OBJT_SWAP) ||
1652 (backing_object->flags & OBJ_DEAD) ||
1653 object->handle != NULL ||
1654 (object->type != OBJT_DEFAULT &&
1655 object->type != OBJT_SWAP) ||
1656 (object->flags & OBJ_DEAD)) {
1657 VM_OBJECT_UNLOCK(backing_object);
1662 object->paging_in_progress != 0 ||
1663 backing_object->paging_in_progress != 0
1665 vm_object_qcollapse(object);
1666 VM_OBJECT_UNLOCK(backing_object);
1670 * We know that we can either collapse the backing object (if
1671 * the parent is the only reference to it) or (perhaps) have
1672 * the parent bypass the object if the parent happens to shadow
1673 * all the resident pages in the entire backing object.
1675 * This is ignoring pager-backed pages such as swap pages.
1676 * vm_object_backing_scan fails the shadowing test in this
1679 if (backing_object->ref_count == 1) {
1681 * If there is exactly one reference to the backing
1682 * object, we can collapse it into the parent.
1684 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1686 #if VM_NRESERVLEVEL > 0
1688 * Break any reservations from backing_object.
1690 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1691 vm_reserv_break_all(backing_object);
1695 * Move the pager from backing_object to object.
1697 if (backing_object->type == OBJT_SWAP) {
1699 * swap_pager_copy() can sleep, in which case
1700 * the backing_object's and object's locks are
1701 * released and reacquired.
1706 OFF_TO_IDX(object->backing_object_offset), TRUE);
1709 * Free any cached pages from backing_object.
1711 if (__predict_false(backing_object->cache != NULL))
1712 vm_page_cache_free(backing_object, 0, 0);
1715 * Object now shadows whatever backing_object did.
1716 * Note that the reference to
1717 * backing_object->backing_object moves from within
1718 * backing_object to within object.
1720 LIST_REMOVE(object, shadow_list);
1721 backing_object->shadow_count--;
1722 if (backing_object->backing_object) {
1723 VM_OBJECT_LOCK(backing_object->backing_object);
1724 LIST_REMOVE(backing_object, shadow_list);
1726 &backing_object->backing_object->shadow_head,
1727 object, shadow_list);
1729 * The shadow_count has not changed.
1731 VM_OBJECT_UNLOCK(backing_object->backing_object);
1733 object->backing_object = backing_object->backing_object;
1734 object->backing_object_offset +=
1735 backing_object->backing_object_offset;
1738 * Discard backing_object.
1740 * Since the backing object has no pages, no pager left,
1741 * and no object references within it, all that is
1742 * necessary is to dispose of it.
1744 KASSERT(backing_object->ref_count == 1, (
1745 "backing_object %p was somehow re-referenced during collapse!",
1747 VM_OBJECT_UNLOCK(backing_object);
1748 vm_object_destroy(backing_object);
1752 vm_object_t new_backing_object;
1755 * If we do not entirely shadow the backing object,
1756 * there is nothing we can do so we give up.
1758 if (object->resident_page_count != object->size &&
1759 vm_object_backing_scan(object,
1760 OBSC_TEST_ALL_SHADOWED) == 0) {
1761 VM_OBJECT_UNLOCK(backing_object);
1766 * Make the parent shadow the next object in the
1767 * chain. Deallocating backing_object will not remove
1768 * it, since its reference count is at least 2.
1770 LIST_REMOVE(object, shadow_list);
1771 backing_object->shadow_count--;
1773 new_backing_object = backing_object->backing_object;
1774 if ((object->backing_object = new_backing_object) != NULL) {
1775 VM_OBJECT_LOCK(new_backing_object);
1777 &new_backing_object->shadow_head,
1781 new_backing_object->shadow_count++;
1782 vm_object_reference_locked(new_backing_object);
1783 VM_OBJECT_UNLOCK(new_backing_object);
1784 object->backing_object_offset +=
1785 backing_object->backing_object_offset;
1789 * Drop the reference count on backing_object. Since
1790 * its ref_count was at least 2, it will not vanish.
1792 backing_object->ref_count--;
1793 VM_OBJECT_UNLOCK(backing_object);
1798 * Try again with this object's new backing object.
1804 * vm_object_page_remove:
1806 * For the given object, either frees or invalidates each of the
1807 * specified pages. In general, a page is freed. However, if a page is
1808 * wired for any reason other than the existence of a managed, wired
1809 * mapping, then it may be invalidated but not removed from the object.
1810 * Pages are specified by the given range ["start", "end") and the option
1811 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1812 * extends from "start" to the end of the object. If the option
1813 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1814 * specified range are affected. If the option OBJPR_NOTMAPPED is
1815 * specified, then the pages within the specified range must have no
1816 * mappings. Otherwise, if this option is not specified, any mappings to
1817 * the specified pages are removed before the pages are freed or
1820 * In general, this operation should only be performed on objects that
1821 * contain managed pages. There are, however, two exceptions. First, it
1822 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1823 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1824 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1825 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1827 * The object must be locked.
1830 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1836 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1837 KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_PHYS) ||
1838 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1839 ("vm_object_page_remove: illegal options for object %p", object));
1840 if (object->resident_page_count == 0)
1842 vm_object_pip_add(object, 1);
1844 p = vm_page_find_least(object, start);
1847 * Here, the variable "p" is either (1) the page with the least pindex
1848 * greater than or equal to the parameter "start" or (2) NULL.
1850 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1851 next = TAILQ_NEXT(p, listq);
1854 * If the page is wired for any reason besides the existence
1855 * of managed, wired mappings, then it cannot be freed. For
1856 * example, fictitious pages, which represent device memory,
1857 * are inherently wired and cannot be freed. They can,
1858 * however, be invalidated if the option OBJPR_CLEANONLY is
1862 if ((wirings = p->wire_count) != 0 &&
1863 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1864 if ((options & OBJPR_NOTMAPPED) == 0) {
1866 /* Account for removal of wired mappings. */
1868 p->wire_count -= wirings;
1870 if ((options & OBJPR_CLEANONLY) == 0) {
1877 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1879 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1880 ("vm_object_page_remove: page %p is fictitious", p));
1881 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1882 if ((options & OBJPR_NOTMAPPED) == 0)
1883 pmap_remove_write(p);
1889 if ((options & OBJPR_NOTMAPPED) == 0) {
1891 /* Account for removal of wired mappings. */
1893 p->wire_count -= wirings;
1898 vm_object_pip_wakeup(object);
1900 if (__predict_false(object->cache != NULL))
1901 vm_page_cache_free(object, start, end);
1905 * vm_object_page_cache:
1907 * For the given object, attempt to move the specified clean
1908 * pages to the cache queue. If a page is wired for any reason,
1909 * then it will not be changed. Pages are specified by the given
1910 * range ["start", "end"). As a special case, if "end" is zero,
1911 * then the range extends from "start" to the end of the object.
1912 * Any mappings to the specified pages are removed before the
1913 * pages are moved to the cache queue.
1915 * This operation should only be performed on objects that
1916 * contain managed pages.
1918 * The object must be locked.
1921 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1923 struct mtx *mtx, *new_mtx;
1926 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1927 KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_SG &&
1928 object->type != OBJT_PHYS),
1929 ("vm_object_page_cache: illegal object %p", object));
1930 if (object->resident_page_count == 0)
1932 p = vm_page_find_least(object, start);
1935 * Here, the variable "p" is either (1) the page with the least pindex
1936 * greater than or equal to the parameter "start" or (2) NULL.
1939 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1940 next = TAILQ_NEXT(p, listq);
1943 * Avoid releasing and reacquiring the same page lock.
1945 new_mtx = vm_page_lockptr(p);
1946 if (mtx != new_mtx) {
1952 vm_page_try_to_cache(p);
1959 * Populate the specified range of the object with valid pages. Returns
1960 * TRUE if the range is successfully populated and FALSE otherwise.
1962 * Note: This function should be optimized to pass a larger array of
1963 * pages to vm_pager_get_pages() before it is applied to a non-
1964 * OBJT_DEVICE object.
1966 * The object must be locked.
1969 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1975 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1976 for (pindex = start; pindex < end; pindex++) {
1977 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1979 if (m->valid != VM_PAGE_BITS_ALL) {
1981 rv = vm_pager_get_pages(object, ma, 1, 0);
1982 m = vm_page_lookup(object, pindex);
1985 if (rv != VM_PAGER_OK) {
1993 * Keep "m" busy because a subsequent iteration may unlock
1997 if (pindex > start) {
1998 m = vm_page_lookup(object, start);
1999 while (m != NULL && m->pindex < pindex) {
2001 m = TAILQ_NEXT(m, listq);
2004 return (pindex == end);
2008 * Routine: vm_object_coalesce
2009 * Function: Coalesces two objects backing up adjoining
2010 * regions of memory into a single object.
2012 * returns TRUE if objects were combined.
2014 * NOTE: Only works at the moment if the second object is NULL -
2015 * if it's not, which object do we lock first?
2018 * prev_object First object to coalesce
2019 * prev_offset Offset into prev_object
2020 * prev_size Size of reference to prev_object
2021 * next_size Size of reference to the second object
2022 * reserved Indicator that extension region has
2023 * swap accounted for
2026 * The object must *not* be locked.
2029 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2030 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2032 vm_pindex_t next_pindex;
2034 if (prev_object == NULL)
2036 VM_OBJECT_LOCK(prev_object);
2037 if (prev_object->type != OBJT_DEFAULT &&
2038 prev_object->type != OBJT_SWAP) {
2039 VM_OBJECT_UNLOCK(prev_object);
2044 * Try to collapse the object first
2046 vm_object_collapse(prev_object);
2049 * Can't coalesce if: . more than one reference . paged out . shadows
2050 * another object . has a copy elsewhere (any of which mean that the
2051 * pages not mapped to prev_entry may be in use anyway)
2053 if (prev_object->backing_object != NULL) {
2054 VM_OBJECT_UNLOCK(prev_object);
2058 prev_size >>= PAGE_SHIFT;
2059 next_size >>= PAGE_SHIFT;
2060 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2062 if ((prev_object->ref_count > 1) &&
2063 (prev_object->size != next_pindex)) {
2064 VM_OBJECT_UNLOCK(prev_object);
2069 * Account for the charge.
2071 if (prev_object->cred != NULL) {
2074 * If prev_object was charged, then this mapping,
2075 * althought not charged now, may become writable
2076 * later. Non-NULL cred in the object would prevent
2077 * swap reservation during enabling of the write
2078 * access, so reserve swap now. Failed reservation
2079 * cause allocation of the separate object for the map
2080 * entry, and swap reservation for this entry is
2081 * managed in appropriate time.
2083 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2084 prev_object->cred)) {
2087 prev_object->charge += ptoa(next_size);
2091 * Remove any pages that may still be in the object from a previous
2094 if (next_pindex < prev_object->size) {
2095 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2097 if (prev_object->type == OBJT_SWAP)
2098 swap_pager_freespace(prev_object,
2099 next_pindex, next_size);
2101 if (prev_object->cred != NULL) {
2102 KASSERT(prev_object->charge >=
2103 ptoa(prev_object->size - next_pindex),
2104 ("object %p overcharged 1 %jx %jx", prev_object,
2105 (uintmax_t)next_pindex, (uintmax_t)next_size));
2106 prev_object->charge -= ptoa(prev_object->size -
2113 * Extend the object if necessary.
2115 if (next_pindex + next_size > prev_object->size)
2116 prev_object->size = next_pindex + next_size;
2118 VM_OBJECT_UNLOCK(prev_object);
2123 vm_object_set_writeable_dirty(vm_object_t object)
2126 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2127 if (object->type != OBJT_VNODE)
2129 object->generation++;
2130 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2132 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2135 #include "opt_ddb.h"
2137 #include <sys/kernel.h>
2139 #include <sys/cons.h>
2141 #include <ddb/ddb.h>
2144 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2147 vm_map_entry_t tmpe;
2155 tmpe = map->header.next;
2156 entcount = map->nentries;
2157 while (entcount-- && (tmpe != &map->header)) {
2158 if (_vm_object_in_map(map, object, tmpe)) {
2163 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2164 tmpm = entry->object.sub_map;
2165 tmpe = tmpm->header.next;
2166 entcount = tmpm->nentries;
2167 while (entcount-- && tmpe != &tmpm->header) {
2168 if (_vm_object_in_map(tmpm, object, tmpe)) {
2173 } else if ((obj = entry->object.vm_object) != NULL) {
2174 for (; obj; obj = obj->backing_object)
2175 if (obj == object) {
2183 vm_object_in_map(vm_object_t object)
2187 /* sx_slock(&allproc_lock); */
2188 FOREACH_PROC_IN_SYSTEM(p) {
2189 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2191 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2192 /* sx_sunlock(&allproc_lock); */
2196 /* sx_sunlock(&allproc_lock); */
2197 if (_vm_object_in_map(kernel_map, object, 0))
2199 if (_vm_object_in_map(kmem_map, object, 0))
2201 if (_vm_object_in_map(pager_map, object, 0))
2203 if (_vm_object_in_map(buffer_map, object, 0))
2208 DB_SHOW_COMMAND(vmochk, vm_object_check)
2213 * make sure that internal objs are in a map somewhere
2214 * and none have zero ref counts.
2216 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2217 if (object->handle == NULL &&
2218 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2219 if (object->ref_count == 0) {
2220 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2221 (long)object->size);
2223 if (!vm_object_in_map(object)) {
2225 "vmochk: internal obj is not in a map: "
2226 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2227 object->ref_count, (u_long)object->size,
2228 (u_long)object->size,
2229 (void *)object->backing_object);
2236 * vm_object_print: [ debug ]
2238 DB_SHOW_COMMAND(object, vm_object_print_static)
2240 /* XXX convert args. */
2241 vm_object_t object = (vm_object_t)addr;
2242 boolean_t full = have_addr;
2246 /* XXX count is an (unused) arg. Avoid shadowing it. */
2247 #define count was_count
2255 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2256 object, (int)object->type, (uintmax_t)object->size,
2257 object->resident_page_count, object->ref_count, object->flags,
2258 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2259 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2260 object->shadow_count,
2261 object->backing_object ? object->backing_object->ref_count : 0,
2262 object->backing_object, (uintmax_t)object->backing_object_offset);
2269 TAILQ_FOREACH(p, &object->memq, listq) {
2271 db_iprintf("memory:=");
2272 else if (count == 6) {
2280 db_printf("(off=0x%jx,page=0x%jx)",
2281 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2291 /* XXX need this non-static entry for calling from vm_map_print. */
2294 /* db_expr_t */ long addr,
2295 boolean_t have_addr,
2296 /* db_expr_t */ long count,
2299 vm_object_print_static(addr, have_addr, count, modif);
2302 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2307 vm_page_t m, prev_m;
2311 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2312 db_printf("new object: %p\n", (void *)object);
2323 TAILQ_FOREACH(m, &object->memq, listq) {
2324 if (m->pindex > 128)
2326 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2327 prev_m->pindex + 1 != m->pindex) {
2329 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2330 (long)fidx, rcount, (long)pa);
2342 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2347 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2348 (long)fidx, rcount, (long)pa);
2358 pa = VM_PAGE_TO_PHYS(m);
2362 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2363 (long)fidx, rcount, (long)pa);