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, boolean_t *clearobjflags,
106 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
107 boolean_t *clearobjflags);
108 static void vm_object_qcollapse(vm_object_t object);
109 static void vm_object_vndeallocate(vm_object_t object);
112 * Virtual memory objects maintain the actual data
113 * associated with allocated virtual memory. A given
114 * page of memory exists within exactly one object.
116 * An object is only deallocated when all "references"
117 * are given up. Only one "reference" to a given
118 * region of an object should be writeable.
120 * Associated with each object is a list of all resident
121 * memory pages belonging to that object; this list is
122 * maintained by the "vm_page" module, and locked by the object's
125 * Each object also records a "pager" routine which is
126 * used to retrieve (and store) pages to the proper backing
127 * storage. In addition, objects may be backed by other
128 * objects from which they were virtual-copied.
130 * The only items within the object structure which are
131 * modified after time of creation are:
132 * reference count locked by object's lock
133 * pager routine locked by object's lock
137 struct object_q vm_object_list;
138 struct mtx vm_object_list_mtx; /* lock for object list and count */
140 struct vm_object kernel_object_store;
141 struct vm_object kmem_object_store;
143 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
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, "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, "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");
442 if (object->ref_count > 1) {
444 VM_OBJECT_UNLOCK(object);
445 /* vrele may need the vnode lock. */
449 VM_OBJECT_UNLOCK(object);
450 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
452 VM_OBJECT_LOCK(object);
454 if (object->type == OBJT_DEAD) {
455 VM_OBJECT_UNLOCK(object);
458 if (object->ref_count == 0)
460 VM_OBJECT_UNLOCK(object);
467 * vm_object_deallocate:
469 * Release a reference to the specified object,
470 * gained either through a vm_object_allocate
471 * or a vm_object_reference call. When all references
472 * are gone, storage associated with this object
473 * may be relinquished.
475 * No object may be locked.
478 vm_object_deallocate(vm_object_t object)
482 while (object != NULL) {
487 VM_OBJECT_LOCK(object);
488 if (object->type == OBJT_VNODE) {
489 struct vnode *vp = (struct vnode *) object->handle;
492 * Conditionally acquire Giant for a vnode-backed
493 * object. We have to be careful since the type of
494 * a vnode object can change while the object is
497 if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
499 if (!mtx_trylock(&Giant)) {
500 VM_OBJECT_UNLOCK(object);
505 vm_object_vndeallocate(object);
506 VFS_UNLOCK_GIANT(vfslocked);
510 * This is to handle the case that the object
511 * changed type while we dropped its lock to
514 VFS_UNLOCK_GIANT(vfslocked);
516 KASSERT(object->ref_count != 0,
517 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
520 * If the reference count goes to 0 we start calling
521 * vm_object_terminate() on the object chain.
522 * A ref count of 1 may be a special case depending on the
523 * shadow count being 0 or 1.
526 if (object->ref_count > 1) {
527 VM_OBJECT_UNLOCK(object);
529 } else if (object->ref_count == 1) {
530 if (object->shadow_count == 0 &&
531 object->handle == NULL &&
532 (object->type == OBJT_DEFAULT ||
533 object->type == OBJT_SWAP)) {
534 vm_object_set_flag(object, OBJ_ONEMAPPING);
535 } else if ((object->shadow_count == 1) &&
536 (object->handle == NULL) &&
537 (object->type == OBJT_DEFAULT ||
538 object->type == OBJT_SWAP)) {
541 robject = LIST_FIRST(&object->shadow_head);
542 KASSERT(robject != NULL,
543 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
545 object->shadow_count));
546 if (!VM_OBJECT_TRYLOCK(robject)) {
548 * Avoid a potential deadlock.
551 VM_OBJECT_UNLOCK(object);
553 * More likely than not the thread
554 * holding robject's lock has lower
555 * priority than the current thread.
556 * Let the lower priority thread run.
562 * Collapse object into its shadow unless its
563 * shadow is dead. In that case, object will
564 * be deallocated by the thread that is
565 * deallocating its shadow.
567 if ((robject->flags & OBJ_DEAD) == 0 &&
568 (robject->handle == NULL) &&
569 (robject->type == OBJT_DEFAULT ||
570 robject->type == OBJT_SWAP)) {
572 robject->ref_count++;
574 if (robject->paging_in_progress) {
575 VM_OBJECT_UNLOCK(object);
576 vm_object_pip_wait(robject,
578 temp = robject->backing_object;
579 if (object == temp) {
580 VM_OBJECT_LOCK(object);
583 } else if (object->paging_in_progress) {
584 VM_OBJECT_UNLOCK(robject);
585 object->flags |= OBJ_PIPWNT;
587 VM_OBJECT_MTX(object),
588 PDROP | PVM, "objde2", 0);
589 VM_OBJECT_LOCK(robject);
590 temp = robject->backing_object;
591 if (object == temp) {
592 VM_OBJECT_LOCK(object);
596 VM_OBJECT_UNLOCK(object);
598 if (robject->ref_count == 1) {
599 robject->ref_count--;
604 vm_object_collapse(object);
605 VM_OBJECT_UNLOCK(object);
608 VM_OBJECT_UNLOCK(robject);
610 VM_OBJECT_UNLOCK(object);
614 temp = object->backing_object;
616 VM_OBJECT_LOCK(temp);
617 LIST_REMOVE(object, shadow_list);
618 temp->shadow_count--;
619 VM_OBJECT_UNLOCK(temp);
620 object->backing_object = NULL;
623 * Don't double-terminate, we could be in a termination
624 * recursion due to the terminate having to sync data
627 if ((object->flags & OBJ_DEAD) == 0)
628 vm_object_terminate(object);
630 VM_OBJECT_UNLOCK(object);
636 * vm_object_destroy removes the object from the global object list
637 * and frees the space for the object.
640 vm_object_destroy(vm_object_t object)
644 * Remove the object from the global object list.
646 mtx_lock(&vm_object_list_mtx);
647 TAILQ_REMOVE(&vm_object_list, object, object_list);
648 mtx_unlock(&vm_object_list_mtx);
651 * Release the allocation charge.
653 if (object->cred != NULL) {
654 KASSERT(object->type == OBJT_DEFAULT ||
655 object->type == OBJT_SWAP,
656 ("vm_object_terminate: non-swap obj %p has cred",
658 swap_release_by_cred(object->charge, object->cred);
660 crfree(object->cred);
665 * Free the space for the object.
667 uma_zfree(obj_zone, object);
671 * vm_object_terminate actually destroys the specified object, freeing
672 * up all previously used resources.
674 * The object must be locked.
675 * This routine may block.
678 vm_object_terminate(vm_object_t object)
682 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
685 * Make sure no one uses us.
687 vm_object_set_flag(object, OBJ_DEAD);
690 * wait for the pageout daemon to be done with the object
692 vm_object_pip_wait(object, "objtrm");
694 KASSERT(!object->paging_in_progress,
695 ("vm_object_terminate: pageout in progress"));
698 * Clean and free the pages, as appropriate. All references to the
699 * object are gone, so we don't need to lock it.
701 if (object->type == OBJT_VNODE) {
702 struct vnode *vp = (struct vnode *)object->handle;
705 * Clean pages and flush buffers.
707 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
708 VM_OBJECT_UNLOCK(object);
710 vinvalbuf(vp, V_SAVE, 0, 0);
712 VM_OBJECT_LOCK(object);
715 KASSERT(object->ref_count == 0,
716 ("vm_object_terminate: object with references, ref_count=%d",
720 * Free any remaining pageable pages. This also removes them from the
721 * paging queues. However, don't free wired pages, just remove them
722 * from the object. Rather than incrementally removing each page from
723 * the object, the page and object are reset to any empty state.
725 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
726 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
727 ("vm_object_terminate: freeing busy page %p", p));
730 * Optimize the page's removal from the object by resetting
731 * its "object" field. Specifically, if the page is not
732 * wired, then the effect of this assignment is that
733 * vm_page_free()'s call to vm_page_remove() will return
734 * immediately without modifying the page or the object.
737 if (p->wire_count == 0) {
739 PCPU_INC(cnt.v_pfree);
744 * If the object contained any pages, then reset it to an empty state.
745 * None of the object's fields, including "resident_page_count", were
746 * modified by the preceding loop.
748 if (object->resident_page_count != 0) {
750 TAILQ_INIT(&object->memq);
751 object->resident_page_count = 0;
752 if (object->type == OBJT_VNODE)
753 vdrop(object->handle);
756 #if VM_NRESERVLEVEL > 0
757 if (__predict_false(!LIST_EMPTY(&object->rvq)))
758 vm_reserv_break_all(object);
760 if (__predict_false(object->cache != NULL))
761 vm_page_cache_free(object, 0, 0);
764 * Let the pager know object is dead.
766 vm_pager_deallocate(object);
767 VM_OBJECT_UNLOCK(object);
769 vm_object_destroy(object);
773 * Make the page read-only so that we can clear the object flags. However, if
774 * this is a nosync mmap then the object is likely to stay dirty so do not
775 * mess with the page and do not clear the object flags. Returns TRUE if the
776 * page should be flushed, and FALSE otherwise.
779 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
783 * If we have been asked to skip nosync pages and this is a
784 * nosync page, skip it. Note that the object flags were not
785 * cleared in this case so we do not have to set them.
787 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
788 *clearobjflags = FALSE;
791 pmap_remove_write(p);
792 return (p->dirty != 0);
797 * vm_object_page_clean
799 * Clean all dirty pages in the specified range of object. Leaves page
800 * on whatever queue it is currently on. If NOSYNC is set then do not
801 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
802 * leaving the object dirty.
804 * When stuffing pages asynchronously, allow clustering. XXX we need a
805 * synchronous clustering mode implementation.
807 * Odd semantics: if start == end, we clean everything.
809 * The object must be locked.
811 * Returns FALSE if some page from the range was not written, as
812 * reported by the pager, and TRUE otherwise.
815 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
819 vm_pindex_t pi, tend, tstart;
820 int curgeneration, n, pagerflags;
821 boolean_t clearobjflags, eio, res;
823 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
824 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
825 KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
826 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
827 object->resident_page_count == 0)
830 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
831 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
832 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
834 tstart = OFF_TO_IDX(start);
835 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
836 clearobjflags = tstart == 0 && tend >= object->size;
840 curgeneration = object->generation;
842 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
846 np = TAILQ_NEXT(p, listq);
849 if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
850 if (object->generation != curgeneration) {
851 if ((flags & OBJPC_SYNC) != 0)
854 clearobjflags = FALSE;
856 np = vm_page_find_least(object, pi);
859 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
862 n = vm_object_page_collect_flush(object, p, pagerflags,
863 flags, &clearobjflags, &eio);
866 clearobjflags = FALSE;
868 if (object->generation != curgeneration) {
869 if ((flags & OBJPC_SYNC) != 0)
872 clearobjflags = FALSE;
876 * If the VOP_PUTPAGES() did a truncated write, so
877 * that even the first page of the run is not fully
878 * written, vm_pageout_flush() returns 0 as the run
879 * length. Since the condition that caused truncated
880 * write may be permanent, e.g. exhausted free space,
881 * accepting n == 0 would cause an infinite loop.
883 * Forwarding the iterator leaves the unwritten page
884 * behind, but there is not much we can do there if
885 * filesystem refuses to write it.
889 clearobjflags = FALSE;
891 np = vm_page_find_least(object, pi + n);
894 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
898 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
903 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
904 int flags, boolean_t *clearobjflags, boolean_t *eio)
906 vm_page_t ma[vm_pageout_page_count], p_first, tp;
907 int count, i, mreq, runlen;
909 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
910 vm_page_lock_assert(p, MA_NOTOWNED);
911 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
916 for (tp = p; count < vm_pageout_page_count; count++) {
917 tp = vm_page_next(tp);
918 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
920 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
924 for (p_first = p; count < vm_pageout_page_count; count++) {
925 tp = vm_page_prev(p_first);
926 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
928 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
934 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
937 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
942 * Note that there is absolutely no sense in writing out
943 * anonymous objects, so we track down the vnode object
945 * We invalidate (remove) all pages from the address space
946 * for semantic correctness.
948 * If the backing object is a device object with unmanaged pages, then any
949 * mappings to the specified range of pages must be removed before this
950 * function is called.
952 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
953 * may start out with a NULL object.
956 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
957 boolean_t syncio, boolean_t invalidate)
959 vm_object_t backing_object;
962 int error, flags, fsync_after;
969 VM_OBJECT_LOCK(object);
970 while ((backing_object = object->backing_object) != NULL) {
971 VM_OBJECT_LOCK(backing_object);
972 offset += object->backing_object_offset;
973 VM_OBJECT_UNLOCK(object);
974 object = backing_object;
975 if (object->size < OFF_TO_IDX(offset + size))
976 size = IDX_TO_OFF(object->size) - offset;
979 * Flush pages if writing is allowed, invalidate them
980 * if invalidation requested. Pages undergoing I/O
981 * will be ignored by vm_object_page_remove().
983 * We cannot lock the vnode and then wait for paging
984 * to complete without deadlocking against vm_fault.
985 * Instead we simply call vm_object_page_remove() and
986 * allow it to block internally on a page-by-page
987 * basis when it encounters pages undergoing async
990 if (object->type == OBJT_VNODE &&
991 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
994 VM_OBJECT_UNLOCK(object);
995 (void) vn_start_write(vp, &mp, V_WAIT);
996 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
997 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
998 if (syncio && !invalidate && offset == 0 &&
999 OFF_TO_IDX(size) == object->size) {
1001 * If syncing the whole mapping of the file,
1002 * it is faster to schedule all the writes in
1003 * async mode, also allowing the clustering,
1004 * and then wait for i/o to complete.
1009 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1010 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1011 fsync_after = FALSE;
1013 VM_OBJECT_LOCK(object);
1014 res = vm_object_page_clean(object, offset, offset + size,
1016 VM_OBJECT_UNLOCK(object);
1018 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1020 VFS_UNLOCK_GIANT(vfslocked);
1021 vn_finished_write(mp);
1024 VM_OBJECT_LOCK(object);
1026 if ((object->type == OBJT_VNODE ||
1027 object->type == OBJT_DEVICE) && invalidate) {
1028 if (object->type == OBJT_DEVICE)
1030 * The option OBJPR_NOTMAPPED must be passed here
1031 * because vm_object_page_remove() cannot remove
1032 * unmanaged mappings.
1034 flags = OBJPR_NOTMAPPED;
1036 flags = OBJPR_NOTWIRED;
1038 flags = OBJPR_CLEANONLY | OBJPR_NOTWIRED;
1039 vm_object_page_remove(object, OFF_TO_IDX(offset),
1040 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1042 VM_OBJECT_UNLOCK(object);
1047 * vm_object_madvise:
1049 * Implements the madvise function at the object/page level.
1051 * MADV_WILLNEED (any object)
1053 * Activate the specified pages if they are resident.
1055 * MADV_DONTNEED (any object)
1057 * Deactivate the specified pages if they are resident.
1059 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1060 * OBJ_ONEMAPPING only)
1062 * Deactivate and clean the specified pages if they are
1063 * resident. This permits the process to reuse the pages
1064 * without faulting or the kernel to reclaim the pages
1068 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1071 vm_pindex_t tpindex;
1072 vm_object_t backing_object, tobject;
1077 VM_OBJECT_LOCK(object);
1079 * Locate and adjust resident pages
1081 for (; pindex < end; pindex += 1) {
1087 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1088 * and those pages must be OBJ_ONEMAPPING.
1090 if (advise == MADV_FREE) {
1091 if ((tobject->type != OBJT_DEFAULT &&
1092 tobject->type != OBJT_SWAP) ||
1093 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1094 goto unlock_tobject;
1096 } else if (tobject->type == OBJT_PHYS)
1097 goto unlock_tobject;
1098 m = vm_page_lookup(tobject, tpindex);
1099 if (m == NULL && advise == MADV_WILLNEED) {
1101 * If the page is cached, reactivate it.
1103 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1108 * There may be swap even if there is no backing page
1110 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1111 swap_pager_freespace(tobject, tpindex, 1);
1115 backing_object = tobject->backing_object;
1116 if (backing_object == NULL)
1117 goto unlock_tobject;
1118 VM_OBJECT_LOCK(backing_object);
1119 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1120 if (tobject != object)
1121 VM_OBJECT_UNLOCK(tobject);
1122 tobject = backing_object;
1124 } else if (m->valid != VM_PAGE_BITS_ALL)
1125 goto unlock_tobject;
1127 * If the page is not in a normal state, skip it.
1130 if (m->hold_count != 0 || m->wire_count != 0) {
1132 goto unlock_tobject;
1134 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1135 ("vm_object_madvise: page %p is fictitious", m));
1136 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1137 ("vm_object_madvise: page %p is not managed", m));
1138 if ((m->oflags & VPO_BUSY) || m->busy) {
1139 if (advise == MADV_WILLNEED) {
1141 * Reference the page before unlocking and
1142 * sleeping so that the page daemon is less
1143 * likely to reclaim it.
1145 vm_page_aflag_set(m, PGA_REFERENCED);
1148 if (object != tobject)
1149 VM_OBJECT_UNLOCK(object);
1150 m->oflags |= VPO_WANTED;
1151 msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1153 VM_OBJECT_LOCK(object);
1156 if (advise == MADV_WILLNEED) {
1157 vm_page_activate(m);
1158 } else if (advise == MADV_DONTNEED) {
1159 vm_page_dontneed(m);
1160 } else if (advise == MADV_FREE) {
1162 * Mark the page clean. This will allow the page
1163 * to be freed up by the system. However, such pages
1164 * are often reused quickly by malloc()/free()
1165 * so we do not do anything that would cause
1166 * a page fault if we can help it.
1168 * Specifically, we do not try to actually free
1169 * the page now nor do we try to put it in the
1170 * cache (which would cause a page fault on reuse).
1172 * But we do make the page is freeable as we
1173 * can without actually taking the step of unmapping
1176 pmap_clear_modify(m);
1179 vm_page_dontneed(m);
1182 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1183 swap_pager_freespace(tobject, tpindex, 1);
1185 if (tobject != object)
1186 VM_OBJECT_UNLOCK(tobject);
1188 VM_OBJECT_UNLOCK(object);
1194 * Create a new object which is backed by the
1195 * specified existing object range. The source
1196 * object reference is deallocated.
1198 * The new object and offset into that object
1199 * are returned in the source parameters.
1203 vm_object_t *object, /* IN/OUT */
1204 vm_ooffset_t *offset, /* IN/OUT */
1213 * Don't create the new object if the old object isn't shared.
1215 if (source != NULL) {
1216 VM_OBJECT_LOCK(source);
1217 if (source->ref_count == 1 &&
1218 source->handle == NULL &&
1219 (source->type == OBJT_DEFAULT ||
1220 source->type == OBJT_SWAP)) {
1221 VM_OBJECT_UNLOCK(source);
1224 VM_OBJECT_UNLOCK(source);
1228 * Allocate a new object with the given length.
1230 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1233 * The new object shadows the source object, adding a reference to it.
1234 * Our caller changes his reference to point to the new object,
1235 * removing a reference to the source object. Net result: no change
1236 * of reference count.
1238 * Try to optimize the result object's page color when shadowing
1239 * in order to maintain page coloring consistency in the combined
1242 result->backing_object = source;
1244 * Store the offset into the source object, and fix up the offset into
1247 result->backing_object_offset = *offset;
1248 if (source != NULL) {
1249 VM_OBJECT_LOCK(source);
1250 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1251 source->shadow_count++;
1252 #if VM_NRESERVLEVEL > 0
1253 result->flags |= source->flags & OBJ_COLORED;
1254 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1255 ((1 << (VM_NFREEORDER - 1)) - 1);
1257 VM_OBJECT_UNLOCK(source);
1262 * Return the new things
1271 * Split the pages in a map entry into a new object. This affords
1272 * easier removal of unused pages, and keeps object inheritance from
1273 * being a negative impact on memory usage.
1276 vm_object_split(vm_map_entry_t entry)
1278 vm_page_t m, m_next;
1279 vm_object_t orig_object, new_object, source;
1280 vm_pindex_t idx, offidxstart;
1283 orig_object = entry->object.vm_object;
1284 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1286 if (orig_object->ref_count <= 1)
1288 VM_OBJECT_UNLOCK(orig_object);
1290 offidxstart = OFF_TO_IDX(entry->offset);
1291 size = atop(entry->end - entry->start);
1294 * If swap_pager_copy() is later called, it will convert new_object
1295 * into a swap object.
1297 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1300 * At this point, the new object is still private, so the order in
1301 * which the original and new objects are locked does not matter.
1303 VM_OBJECT_LOCK(new_object);
1304 VM_OBJECT_LOCK(orig_object);
1305 source = orig_object->backing_object;
1306 if (source != NULL) {
1307 VM_OBJECT_LOCK(source);
1308 if ((source->flags & OBJ_DEAD) != 0) {
1309 VM_OBJECT_UNLOCK(source);
1310 VM_OBJECT_UNLOCK(orig_object);
1311 VM_OBJECT_UNLOCK(new_object);
1312 vm_object_deallocate(new_object);
1313 VM_OBJECT_LOCK(orig_object);
1316 LIST_INSERT_HEAD(&source->shadow_head,
1317 new_object, shadow_list);
1318 source->shadow_count++;
1319 vm_object_reference_locked(source); /* for new_object */
1320 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1321 VM_OBJECT_UNLOCK(source);
1322 new_object->backing_object_offset =
1323 orig_object->backing_object_offset + entry->offset;
1324 new_object->backing_object = source;
1326 if (orig_object->cred != NULL) {
1327 new_object->cred = orig_object->cred;
1328 crhold(orig_object->cred);
1329 new_object->charge = ptoa(size);
1330 KASSERT(orig_object->charge >= ptoa(size),
1331 ("orig_object->charge < 0"));
1332 orig_object->charge -= ptoa(size);
1335 m = vm_page_find_least(orig_object, offidxstart);
1336 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1338 m_next = TAILQ_NEXT(m, listq);
1341 * We must wait for pending I/O to complete before we can
1344 * We do not have to VM_PROT_NONE the page as mappings should
1345 * not be changed by this operation.
1347 if ((m->oflags & VPO_BUSY) || m->busy) {
1348 VM_OBJECT_UNLOCK(new_object);
1349 m->oflags |= VPO_WANTED;
1350 msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1351 VM_OBJECT_LOCK(new_object);
1355 vm_page_rename(m, new_object, idx);
1357 /* page automatically made dirty by rename and cache handled */
1360 if (orig_object->type == OBJT_SWAP) {
1362 * swap_pager_copy() can sleep, in which case the orig_object's
1363 * and new_object's locks are released and reacquired.
1365 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1368 * Transfer any cached pages from orig_object to new_object.
1370 if (__predict_false(orig_object->cache != NULL))
1371 vm_page_cache_transfer(orig_object, offidxstart,
1374 VM_OBJECT_UNLOCK(orig_object);
1375 TAILQ_FOREACH(m, &new_object->memq, listq)
1377 VM_OBJECT_UNLOCK(new_object);
1378 entry->object.vm_object = new_object;
1379 entry->offset = 0LL;
1380 vm_object_deallocate(orig_object);
1381 VM_OBJECT_LOCK(new_object);
1384 #define OBSC_TEST_ALL_SHADOWED 0x0001
1385 #define OBSC_COLLAPSE_NOWAIT 0x0002
1386 #define OBSC_COLLAPSE_WAIT 0x0004
1389 vm_object_backing_scan(vm_object_t object, int op)
1393 vm_object_t backing_object;
1394 vm_pindex_t backing_offset_index;
1396 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1397 VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1399 backing_object = object->backing_object;
1400 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1403 * Initial conditions
1405 if (op & OBSC_TEST_ALL_SHADOWED) {
1407 * We do not want to have to test for the existence of cache
1408 * or swap pages in the backing object. XXX but with the
1409 * new swapper this would be pretty easy to do.
1411 * XXX what about anonymous MAP_SHARED memory that hasn't
1412 * been ZFOD faulted yet? If we do not test for this, the
1413 * shadow test may succeed! XXX
1415 if (backing_object->type != OBJT_DEFAULT) {
1419 if (op & OBSC_COLLAPSE_WAIT) {
1420 vm_object_set_flag(backing_object, OBJ_DEAD);
1426 p = TAILQ_FIRST(&backing_object->memq);
1428 vm_page_t next = TAILQ_NEXT(p, listq);
1429 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1431 if (op & OBSC_TEST_ALL_SHADOWED) {
1435 * Ignore pages outside the parent object's range
1436 * and outside the parent object's mapping of the
1439 * note that we do not busy the backing object's
1443 p->pindex < backing_offset_index ||
1444 new_pindex >= object->size
1451 * See if the parent has the page or if the parent's
1452 * object pager has the page. If the parent has the
1453 * page but the page is not valid, the parent's
1454 * object pager must have the page.
1456 * If this fails, the parent does not completely shadow
1457 * the object and we might as well give up now.
1460 pp = vm_page_lookup(object, new_pindex);
1462 (pp == NULL || pp->valid == 0) &&
1463 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1471 * Check for busy page
1473 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1476 if (op & OBSC_COLLAPSE_NOWAIT) {
1477 if ((p->oflags & VPO_BUSY) ||
1483 } else if (op & OBSC_COLLAPSE_WAIT) {
1484 if ((p->oflags & VPO_BUSY) || p->busy) {
1485 VM_OBJECT_UNLOCK(object);
1486 p->oflags |= VPO_WANTED;
1487 msleep(p, VM_OBJECT_MTX(backing_object),
1488 PDROP | PVM, "vmocol", 0);
1489 VM_OBJECT_LOCK(object);
1490 VM_OBJECT_LOCK(backing_object);
1492 * If we slept, anything could have
1493 * happened. Since the object is
1494 * marked dead, the backing offset
1495 * should not have changed so we
1496 * just restart our scan.
1498 p = TAILQ_FIRST(&backing_object->memq);
1504 p->object == backing_object,
1505 ("vm_object_backing_scan: object mismatch")
1509 * Destroy any associated swap
1511 if (backing_object->type == OBJT_SWAP) {
1512 swap_pager_freespace(
1520 p->pindex < backing_offset_index ||
1521 new_pindex >= object->size
1524 * Page is out of the parent object's range, we
1525 * can simply destroy it.
1528 KASSERT(!pmap_page_is_mapped(p),
1529 ("freeing mapped page %p", p));
1530 if (p->wire_count == 0)
1539 pp = vm_page_lookup(object, new_pindex);
1541 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1542 (pp != NULL && pp->valid == 0)
1545 * The page in the parent is not (yet) valid.
1546 * We don't know anything about the state of
1547 * the original page. It might be mapped,
1548 * so we must avoid the next if here.
1550 * This is due to a race in vm_fault() where
1551 * we must unbusy the original (backing_obj)
1552 * page before we can (re)lock the parent.
1553 * Hence we can get here.
1560 vm_pager_has_page(object, new_pindex, NULL, NULL)
1563 * page already exists in parent OR swap exists
1564 * for this location in the parent. Destroy
1565 * the original page from the backing object.
1567 * Leave the parent's page alone
1570 KASSERT(!pmap_page_is_mapped(p),
1571 ("freeing mapped page %p", p));
1572 if (p->wire_count == 0)
1581 #if VM_NRESERVLEVEL > 0
1583 * Rename the reservation.
1585 vm_reserv_rename(p, object, backing_object,
1586 backing_offset_index);
1590 * Page does not exist in parent, rename the
1591 * page from the backing object to the main object.
1593 * If the page was mapped to a process, it can remain
1594 * mapped through the rename.
1597 vm_page_rename(p, object, new_pindex);
1599 /* page automatically made dirty by rename */
1608 * this version of collapse allows the operation to occur earlier and
1609 * when paging_in_progress is true for an object... This is not a complete
1610 * operation, but should plug 99.9% of the rest of the leaks.
1613 vm_object_qcollapse(vm_object_t object)
1615 vm_object_t backing_object = object->backing_object;
1617 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1618 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1620 if (backing_object->ref_count != 1)
1623 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1627 * vm_object_collapse:
1629 * Collapse an object with the object backing it.
1630 * Pages in the backing object are moved into the
1631 * parent, and the backing object is deallocated.
1634 vm_object_collapse(vm_object_t object)
1636 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1639 vm_object_t backing_object;
1642 * Verify that the conditions are right for collapse:
1644 * The object exists and the backing object exists.
1646 if ((backing_object = object->backing_object) == NULL)
1650 * we check the backing object first, because it is most likely
1653 VM_OBJECT_LOCK(backing_object);
1654 if (backing_object->handle != NULL ||
1655 (backing_object->type != OBJT_DEFAULT &&
1656 backing_object->type != OBJT_SWAP) ||
1657 (backing_object->flags & OBJ_DEAD) ||
1658 object->handle != NULL ||
1659 (object->type != OBJT_DEFAULT &&
1660 object->type != OBJT_SWAP) ||
1661 (object->flags & OBJ_DEAD)) {
1662 VM_OBJECT_UNLOCK(backing_object);
1667 object->paging_in_progress != 0 ||
1668 backing_object->paging_in_progress != 0
1670 vm_object_qcollapse(object);
1671 VM_OBJECT_UNLOCK(backing_object);
1675 * We know that we can either collapse the backing object (if
1676 * the parent is the only reference to it) or (perhaps) have
1677 * the parent bypass the object if the parent happens to shadow
1678 * all the resident pages in the entire backing object.
1680 * This is ignoring pager-backed pages such as swap pages.
1681 * vm_object_backing_scan fails the shadowing test in this
1684 if (backing_object->ref_count == 1) {
1686 * If there is exactly one reference to the backing
1687 * object, we can collapse it into the parent.
1689 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1691 #if VM_NRESERVLEVEL > 0
1693 * Break any reservations from backing_object.
1695 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1696 vm_reserv_break_all(backing_object);
1700 * Move the pager from backing_object to object.
1702 if (backing_object->type == OBJT_SWAP) {
1704 * swap_pager_copy() can sleep, in which case
1705 * the backing_object's and object's locks are
1706 * released and reacquired.
1711 OFF_TO_IDX(object->backing_object_offset), TRUE);
1714 * Free any cached pages from backing_object.
1716 if (__predict_false(backing_object->cache != NULL))
1717 vm_page_cache_free(backing_object, 0, 0);
1720 * Object now shadows whatever backing_object did.
1721 * Note that the reference to
1722 * backing_object->backing_object moves from within
1723 * backing_object to within object.
1725 LIST_REMOVE(object, shadow_list);
1726 backing_object->shadow_count--;
1727 if (backing_object->backing_object) {
1728 VM_OBJECT_LOCK(backing_object->backing_object);
1729 LIST_REMOVE(backing_object, shadow_list);
1731 &backing_object->backing_object->shadow_head,
1732 object, shadow_list);
1734 * The shadow_count has not changed.
1736 VM_OBJECT_UNLOCK(backing_object->backing_object);
1738 object->backing_object = backing_object->backing_object;
1739 object->backing_object_offset +=
1740 backing_object->backing_object_offset;
1743 * Discard backing_object.
1745 * Since the backing object has no pages, no pager left,
1746 * and no object references within it, all that is
1747 * necessary is to dispose of it.
1749 KASSERT(backing_object->ref_count == 1, (
1750 "backing_object %p was somehow re-referenced during collapse!",
1752 VM_OBJECT_UNLOCK(backing_object);
1753 vm_object_destroy(backing_object);
1757 vm_object_t new_backing_object;
1760 * If we do not entirely shadow the backing object,
1761 * there is nothing we can do so we give up.
1763 if (object->resident_page_count != object->size &&
1764 vm_object_backing_scan(object,
1765 OBSC_TEST_ALL_SHADOWED) == 0) {
1766 VM_OBJECT_UNLOCK(backing_object);
1771 * Make the parent shadow the next object in the
1772 * chain. Deallocating backing_object will not remove
1773 * it, since its reference count is at least 2.
1775 LIST_REMOVE(object, shadow_list);
1776 backing_object->shadow_count--;
1778 new_backing_object = backing_object->backing_object;
1779 if ((object->backing_object = new_backing_object) != NULL) {
1780 VM_OBJECT_LOCK(new_backing_object);
1782 &new_backing_object->shadow_head,
1786 new_backing_object->shadow_count++;
1787 vm_object_reference_locked(new_backing_object);
1788 VM_OBJECT_UNLOCK(new_backing_object);
1789 object->backing_object_offset +=
1790 backing_object->backing_object_offset;
1794 * Drop the reference count on backing_object. Since
1795 * its ref_count was at least 2, it will not vanish.
1797 backing_object->ref_count--;
1798 VM_OBJECT_UNLOCK(backing_object);
1803 * Try again with this object's new backing object.
1809 * vm_object_page_remove:
1811 * For the given object, either frees or invalidates each of the
1812 * specified pages. In general, a page is freed. However, if a page is
1813 * wired for any reason other than the existence of a managed, wired
1814 * mapping, then it may be invalidated but not removed from the object.
1815 * Pages are specified by the given range ["start", "end") and the option
1816 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1817 * extends from "start" to the end of the object. If the option
1818 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1819 * specified range are affected. If the option OBJPR_NOTMAPPED is
1820 * specified, then the pages within the specified range must have no
1821 * mappings. Otherwise, if this option is not specified, any mappings to
1822 * the specified pages are removed before the pages are freed or
1825 * In general, this operation should only be performed on objects that
1826 * contain managed pages. There are, however, two exceptions. First, it
1827 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1828 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1829 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1830 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1832 * The object must be locked.
1835 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1841 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1842 KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_PHYS) ||
1843 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1844 ("vm_object_page_remove: illegal options for object %p", object));
1845 if (object->resident_page_count == 0)
1847 vm_object_pip_add(object, 1);
1849 p = vm_page_find_least(object, start);
1852 * Here, the variable "p" is either (1) the page with the least pindex
1853 * greater than or equal to the parameter "start" or (2) NULL.
1855 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1856 next = TAILQ_NEXT(p, listq);
1859 * If the page is wired for any reason besides the existence
1860 * of managed, wired mappings, then it cannot be freed. For
1861 * example, fictitious pages, which represent device memory,
1862 * are inherently wired and cannot be freed. They can,
1863 * however, be invalidated if the option OBJPR_CLEANONLY is
1867 if ((wirings = p->wire_count) != 0 &&
1868 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1869 if ((options & (OBJPR_NOTWIRED | OBJPR_NOTMAPPED)) ==
1872 /* Account for removal of wired mappings. */
1874 p->wire_count -= wirings;
1876 if ((options & OBJPR_CLEANONLY) == 0) {
1882 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1884 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1885 ("vm_object_page_remove: page %p is fictitious", p));
1886 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1887 if ((options & OBJPR_NOTMAPPED) == 0)
1888 pmap_remove_write(p);
1892 if ((options & OBJPR_NOTMAPPED) == 0) {
1893 if ((options & OBJPR_NOTWIRED) != 0 && wirings != 0)
1896 /* Account for removal of wired mappings. */
1898 KASSERT(p->wire_count == wirings,
1899 ("inconsistent wire count %d %d %p",
1900 p->wire_count, wirings, p));
1902 atomic_subtract_int(&cnt.v_wire_count, 1);
1909 vm_object_pip_wakeup(object);
1911 if (__predict_false(object->cache != NULL))
1912 vm_page_cache_free(object, start, end);
1916 * vm_object_page_cache:
1918 * For the given object, attempt to move the specified clean
1919 * pages to the cache queue. If a page is wired for any reason,
1920 * then it will not be changed. Pages are specified by the given
1921 * range ["start", "end"). As a special case, if "end" is zero,
1922 * then the range extends from "start" to the end of the object.
1923 * Any mappings to the specified pages are removed before the
1924 * pages are moved to the cache queue.
1926 * This operation should only be performed on objects that
1927 * contain managed pages.
1929 * The object must be locked.
1932 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1934 struct mtx *mtx, *new_mtx;
1937 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1938 KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_SG &&
1939 object->type != OBJT_PHYS),
1940 ("vm_object_page_cache: illegal object %p", object));
1941 if (object->resident_page_count == 0)
1943 p = vm_page_find_least(object, start);
1946 * Here, the variable "p" is either (1) the page with the least pindex
1947 * greater than or equal to the parameter "start" or (2) NULL.
1950 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1951 next = TAILQ_NEXT(p, listq);
1954 * Avoid releasing and reacquiring the same page lock.
1956 new_mtx = vm_page_lockptr(p);
1957 if (mtx != new_mtx) {
1963 vm_page_try_to_cache(p);
1970 * Populate the specified range of the object with valid pages. Returns
1971 * TRUE if the range is successfully populated and FALSE otherwise.
1973 * Note: This function should be optimized to pass a larger array of
1974 * pages to vm_pager_get_pages() before it is applied to a non-
1975 * OBJT_DEVICE object.
1977 * The object must be locked.
1980 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1986 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1987 for (pindex = start; pindex < end; pindex++) {
1988 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1990 if (m->valid != VM_PAGE_BITS_ALL) {
1992 rv = vm_pager_get_pages(object, ma, 1, 0);
1993 m = vm_page_lookup(object, pindex);
1996 if (rv != VM_PAGER_OK) {
2004 * Keep "m" busy because a subsequent iteration may unlock
2008 if (pindex > start) {
2009 m = vm_page_lookup(object, start);
2010 while (m != NULL && m->pindex < pindex) {
2012 m = TAILQ_NEXT(m, listq);
2015 return (pindex == end);
2019 * Routine: vm_object_coalesce
2020 * Function: Coalesces two objects backing up adjoining
2021 * regions of memory into a single object.
2023 * returns TRUE if objects were combined.
2025 * NOTE: Only works at the moment if the second object is NULL -
2026 * if it's not, which object do we lock first?
2029 * prev_object First object to coalesce
2030 * prev_offset Offset into prev_object
2031 * prev_size Size of reference to prev_object
2032 * next_size Size of reference to the second object
2033 * reserved Indicator that extension region has
2034 * swap accounted for
2037 * The object must *not* be locked.
2040 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2041 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2043 vm_pindex_t next_pindex;
2045 if (prev_object == NULL)
2047 VM_OBJECT_LOCK(prev_object);
2048 if (prev_object->type != OBJT_DEFAULT &&
2049 prev_object->type != OBJT_SWAP) {
2050 VM_OBJECT_UNLOCK(prev_object);
2055 * Try to collapse the object first
2057 vm_object_collapse(prev_object);
2060 * Can't coalesce if: . more than one reference . paged out . shadows
2061 * another object . has a copy elsewhere (any of which mean that the
2062 * pages not mapped to prev_entry may be in use anyway)
2064 if (prev_object->backing_object != NULL) {
2065 VM_OBJECT_UNLOCK(prev_object);
2069 prev_size >>= PAGE_SHIFT;
2070 next_size >>= PAGE_SHIFT;
2071 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2073 if ((prev_object->ref_count > 1) &&
2074 (prev_object->size != next_pindex)) {
2075 VM_OBJECT_UNLOCK(prev_object);
2080 * Account for the charge.
2082 if (prev_object->cred != NULL) {
2085 * If prev_object was charged, then this mapping,
2086 * althought not charged now, may become writable
2087 * later. Non-NULL cred in the object would prevent
2088 * swap reservation during enabling of the write
2089 * access, so reserve swap now. Failed reservation
2090 * cause allocation of the separate object for the map
2091 * entry, and swap reservation for this entry is
2092 * managed in appropriate time.
2094 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2095 prev_object->cred)) {
2098 prev_object->charge += ptoa(next_size);
2102 * Remove any pages that may still be in the object from a previous
2105 if (next_pindex < prev_object->size) {
2106 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2108 if (prev_object->type == OBJT_SWAP)
2109 swap_pager_freespace(prev_object,
2110 next_pindex, next_size);
2112 if (prev_object->cred != NULL) {
2113 KASSERT(prev_object->charge >=
2114 ptoa(prev_object->size - next_pindex),
2115 ("object %p overcharged 1 %jx %jx", prev_object,
2116 (uintmax_t)next_pindex, (uintmax_t)next_size));
2117 prev_object->charge -= ptoa(prev_object->size -
2124 * Extend the object if necessary.
2126 if (next_pindex + next_size > prev_object->size)
2127 prev_object->size = next_pindex + next_size;
2129 VM_OBJECT_UNLOCK(prev_object);
2134 vm_object_set_writeable_dirty(vm_object_t object)
2137 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2138 if (object->type != OBJT_VNODE)
2140 object->generation++;
2141 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2143 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2146 #include "opt_ddb.h"
2148 #include <sys/kernel.h>
2150 #include <sys/cons.h>
2152 #include <ddb/ddb.h>
2155 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2158 vm_map_entry_t tmpe;
2166 tmpe = map->header.next;
2167 entcount = map->nentries;
2168 while (entcount-- && (tmpe != &map->header)) {
2169 if (_vm_object_in_map(map, object, tmpe)) {
2174 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2175 tmpm = entry->object.sub_map;
2176 tmpe = tmpm->header.next;
2177 entcount = tmpm->nentries;
2178 while (entcount-- && tmpe != &tmpm->header) {
2179 if (_vm_object_in_map(tmpm, object, tmpe)) {
2184 } else if ((obj = entry->object.vm_object) != NULL) {
2185 for (; obj; obj = obj->backing_object)
2186 if (obj == object) {
2194 vm_object_in_map(vm_object_t object)
2198 /* sx_slock(&allproc_lock); */
2199 FOREACH_PROC_IN_SYSTEM(p) {
2200 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2202 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2203 /* sx_sunlock(&allproc_lock); */
2207 /* sx_sunlock(&allproc_lock); */
2208 if (_vm_object_in_map(kernel_map, object, 0))
2210 if (_vm_object_in_map(kmem_map, object, 0))
2212 if (_vm_object_in_map(pager_map, object, 0))
2214 if (_vm_object_in_map(buffer_map, object, 0))
2219 DB_SHOW_COMMAND(vmochk, vm_object_check)
2224 * make sure that internal objs are in a map somewhere
2225 * and none have zero ref counts.
2227 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2228 if (object->handle == NULL &&
2229 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2230 if (object->ref_count == 0) {
2231 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2232 (long)object->size);
2234 if (!vm_object_in_map(object)) {
2236 "vmochk: internal obj is not in a map: "
2237 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2238 object->ref_count, (u_long)object->size,
2239 (u_long)object->size,
2240 (void *)object->backing_object);
2247 * vm_object_print: [ debug ]
2249 DB_SHOW_COMMAND(object, vm_object_print_static)
2251 /* XXX convert args. */
2252 vm_object_t object = (vm_object_t)addr;
2253 boolean_t full = have_addr;
2257 /* XXX count is an (unused) arg. Avoid shadowing it. */
2258 #define count was_count
2266 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2267 object, (int)object->type, (uintmax_t)object->size,
2268 object->resident_page_count, object->ref_count, object->flags,
2269 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2270 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2271 object->shadow_count,
2272 object->backing_object ? object->backing_object->ref_count : 0,
2273 object->backing_object, (uintmax_t)object->backing_object_offset);
2280 TAILQ_FOREACH(p, &object->memq, listq) {
2282 db_iprintf("memory:=");
2283 else if (count == 6) {
2291 db_printf("(off=0x%jx,page=0x%jx)",
2292 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2302 /* XXX need this non-static entry for calling from vm_map_print. */
2305 /* db_expr_t */ long addr,
2306 boolean_t have_addr,
2307 /* db_expr_t */ long count,
2310 vm_object_print_static(addr, have_addr, count, modif);
2313 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2318 vm_page_t m, prev_m;
2322 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2323 db_printf("new object: %p\n", (void *)object);
2334 TAILQ_FOREACH(m, &object->memq, listq) {
2335 if (m->pindex > 128)
2337 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2338 prev_m->pindex + 1 != m->pindex) {
2340 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2341 (long)fidx, rcount, (long)pa);
2353 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2358 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2359 (long)fidx, rcount, (long)pa);
2369 pa = VM_PAGE_TO_PHYS(m);
2373 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2374 (long)fidx, rcount, (long)pa);