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/rwlock.h>
82 #include <sys/vnode.h>
83 #include <sys/vmmeter.h>
87 #include <vm/vm_param.h>
89 #include <vm/vm_map.h>
90 #include <vm/vm_object.h>
91 #include <vm/vm_page.h>
92 #include <vm/vm_pageout.h>
93 #include <vm/vm_pager.h>
94 #include <vm/swap_pager.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_extern.h>
97 #include <vm/vm_radix.h>
98 #include <vm/vm_reserv.h>
101 static int old_msync;
102 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
103 "Use old (insecure) msync behavior");
105 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
106 int pagerflags, int flags, boolean_t *clearobjflags,
108 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
109 boolean_t *clearobjflags);
110 static void vm_object_qcollapse(vm_object_t object);
111 static void vm_object_vndeallocate(vm_object_t object);
114 * Virtual memory objects maintain the actual data
115 * associated with allocated virtual memory. A given
116 * page of memory exists within exactly one object.
118 * An object is only deallocated when all "references"
119 * are given up. Only one "reference" to a given
120 * region of an object should be writeable.
122 * Associated with each object is a list of all resident
123 * memory pages belonging to that object; this list is
124 * maintained by the "vm_page" module, and locked by the object's
127 * Each object also records a "pager" routine which is
128 * used to retrieve (and store) pages to the proper backing
129 * storage. In addition, objects may be backed by other
130 * objects from which they were virtual-copied.
132 * The only items within the object structure which are
133 * modified after time of creation are:
134 * reference count locked by object's lock
135 * pager routine locked by object's lock
139 struct object_q vm_object_list;
140 struct mtx vm_object_list_mtx; /* lock for object list and count */
142 struct vm_object kernel_object_store;
143 struct vm_object kmem_object_store;
145 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
148 static long object_collapses;
149 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
150 &object_collapses, 0, "VM object collapses");
152 static long object_bypasses;
153 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
154 &object_bypasses, 0, "VM object bypasses");
156 static uma_zone_t obj_zone;
158 static int vm_object_zinit(void *mem, int size, int flags);
161 static void vm_object_zdtor(void *mem, int size, void *arg);
164 vm_object_zdtor(void *mem, int size, void *arg)
168 object = (vm_object_t)mem;
169 KASSERT(TAILQ_EMPTY(&object->memq),
170 ("object %p has resident pages in its memq", object));
171 KASSERT(vm_radix_is_empty(&object->rtree),
172 ("object %p has resident pages in its trie", object));
173 #if VM_NRESERVLEVEL > 0
174 KASSERT(LIST_EMPTY(&object->rvq),
175 ("object %p has reservations",
178 KASSERT(vm_object_cache_is_empty(object),
179 ("object %p has cached pages",
181 KASSERT(object->paging_in_progress == 0,
182 ("object %p paging_in_progress = %d",
183 object, object->paging_in_progress));
184 KASSERT(object->resident_page_count == 0,
185 ("object %p resident_page_count = %d",
186 object, object->resident_page_count));
187 KASSERT(object->shadow_count == 0,
188 ("object %p shadow_count = %d",
189 object, object->shadow_count));
194 vm_object_zinit(void *mem, int size, int flags)
198 object = (vm_object_t)mem;
199 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
201 /* These are true for any object that has been freed */
202 object->rtree.rt_root = 0;
203 object->rtree.rt_flags = 0;
204 object->paging_in_progress = 0;
205 object->resident_page_count = 0;
206 object->shadow_count = 0;
207 object->cache.rt_root = 0;
208 object->cache.rt_flags = 0;
213 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
216 TAILQ_INIT(&object->memq);
217 LIST_INIT(&object->shadow_head);
222 panic("_vm_object_allocate: can't create OBJT_DEAD");
225 object->flags = OBJ_ONEMAPPING;
229 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
232 object->flags = OBJ_FICTITIOUS;
235 object->flags = OBJ_UNMANAGED;
241 panic("_vm_object_allocate: type %d is undefined", type);
244 object->generation = 1;
245 object->ref_count = 1;
246 object->memattr = VM_MEMATTR_DEFAULT;
249 object->handle = NULL;
250 object->backing_object = NULL;
251 object->backing_object_offset = (vm_ooffset_t) 0;
252 #if VM_NRESERVLEVEL > 0
253 LIST_INIT(&object->rvq);
256 mtx_lock(&vm_object_list_mtx);
257 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
258 mtx_unlock(&vm_object_list_mtx);
264 * Initialize the VM objects module.
269 TAILQ_INIT(&vm_object_list);
270 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
272 rw_init(&kernel_object->lock, "kernel vm object");
273 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
275 #if VM_NRESERVLEVEL > 0
276 kernel_object->flags |= OBJ_COLORED;
277 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
280 rw_init(&kmem_object->lock, "kmem vm object");
281 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
283 #if VM_NRESERVLEVEL > 0
284 kmem_object->flags |= OBJ_COLORED;
285 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
289 * The lock portion of struct vm_object must be type stable due
290 * to vm_pageout_fallback_object_lock locking a vm object
291 * without holding any references to it.
293 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
299 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
305 vm_object_clear_flag(vm_object_t object, u_short bits)
308 VM_OBJECT_ASSERT_WLOCKED(object);
309 object->flags &= ~bits;
313 * Sets the default memory attribute for the specified object. Pages
314 * that are allocated to this object are by default assigned this memory
317 * Presently, this function must be called before any pages are allocated
318 * to the object. In the future, this requirement may be relaxed for
319 * "default" and "swap" objects.
322 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
325 VM_OBJECT_ASSERT_WLOCKED(object);
326 switch (object->type) {
334 if (!TAILQ_EMPTY(&object->memq))
335 return (KERN_FAILURE);
338 return (KERN_INVALID_ARGUMENT);
340 panic("vm_object_set_memattr: object %p is of undefined type",
343 object->memattr = memattr;
344 return (KERN_SUCCESS);
348 vm_object_pip_add(vm_object_t object, short i)
351 VM_OBJECT_ASSERT_WLOCKED(object);
352 object->paging_in_progress += i;
356 vm_object_pip_subtract(vm_object_t object, short i)
359 VM_OBJECT_ASSERT_WLOCKED(object);
360 object->paging_in_progress -= i;
364 vm_object_pip_wakeup(vm_object_t object)
367 VM_OBJECT_ASSERT_WLOCKED(object);
368 object->paging_in_progress--;
369 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
370 vm_object_clear_flag(object, OBJ_PIPWNT);
376 vm_object_pip_wakeupn(vm_object_t object, short i)
379 VM_OBJECT_ASSERT_WLOCKED(object);
381 object->paging_in_progress -= i;
382 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
383 vm_object_clear_flag(object, OBJ_PIPWNT);
389 vm_object_pip_wait(vm_object_t object, char *waitid)
392 VM_OBJECT_ASSERT_WLOCKED(object);
393 while (object->paging_in_progress) {
394 object->flags |= OBJ_PIPWNT;
395 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
400 * vm_object_allocate:
402 * Returns a new object with the given size.
405 vm_object_allocate(objtype_t type, vm_pindex_t size)
409 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
410 _vm_object_allocate(type, size, object);
416 * vm_object_reference:
418 * Gets another reference to the given object. Note: OBJ_DEAD
419 * objects can be referenced during final cleaning.
422 vm_object_reference(vm_object_t object)
426 VM_OBJECT_WLOCK(object);
427 vm_object_reference_locked(object);
428 VM_OBJECT_WUNLOCK(object);
432 * vm_object_reference_locked:
434 * Gets another reference to the given object.
436 * The object must be locked.
439 vm_object_reference_locked(vm_object_t object)
443 VM_OBJECT_ASSERT_WLOCKED(object);
445 if (object->type == OBJT_VNODE) {
452 * Handle deallocating an object of type OBJT_VNODE.
455 vm_object_vndeallocate(vm_object_t object)
457 struct vnode *vp = (struct vnode *) object->handle;
459 VM_OBJECT_ASSERT_WLOCKED(object);
460 KASSERT(object->type == OBJT_VNODE,
461 ("vm_object_vndeallocate: not a vnode object"));
462 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
464 if (object->ref_count == 0) {
465 vprint("vm_object_vndeallocate", vp);
466 panic("vm_object_vndeallocate: bad object reference count");
471 * The test for text of vp vnode does not need a bypass to
472 * reach right VV_TEXT there, since it is obtained from
475 if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) {
477 VM_OBJECT_WUNLOCK(object);
478 /* vrele may need the vnode lock. */
482 VM_OBJECT_WUNLOCK(object);
483 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
485 VM_OBJECT_WLOCK(object);
487 if (object->type == OBJT_DEAD) {
488 VM_OBJECT_WUNLOCK(object);
491 if (object->ref_count == 0)
493 VM_OBJECT_WUNLOCK(object);
500 * vm_object_deallocate:
502 * Release a reference to the specified object,
503 * gained either through a vm_object_allocate
504 * or a vm_object_reference call. When all references
505 * are gone, storage associated with this object
506 * may be relinquished.
508 * No object may be locked.
511 vm_object_deallocate(vm_object_t object)
516 while (object != NULL) {
517 VM_OBJECT_WLOCK(object);
518 if (object->type == OBJT_VNODE) {
519 vm_object_vndeallocate(object);
523 KASSERT(object->ref_count != 0,
524 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
527 * If the reference count goes to 0 we start calling
528 * vm_object_terminate() on the object chain.
529 * A ref count of 1 may be a special case depending on the
530 * shadow count being 0 or 1.
533 if (object->ref_count > 1) {
534 VM_OBJECT_WUNLOCK(object);
536 } else if (object->ref_count == 1) {
537 if (object->type == OBJT_SWAP &&
538 (object->flags & OBJ_TMPFS) != 0) {
539 vp = object->un_pager.swp.swp_tmpfs;
541 VM_OBJECT_WUNLOCK(object);
542 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
543 VM_OBJECT_WLOCK(object);
544 if (object->type == OBJT_DEAD ||
545 object->ref_count != 1) {
546 VM_OBJECT_WUNLOCK(object);
551 if ((object->flags & OBJ_TMPFS) != 0)
556 if (object->shadow_count == 0 &&
557 object->handle == NULL &&
558 (object->type == OBJT_DEFAULT ||
559 (object->type == OBJT_SWAP &&
560 (object->flags & OBJ_TMPFS_NODE) == 0))) {
561 vm_object_set_flag(object, OBJ_ONEMAPPING);
562 } else if ((object->shadow_count == 1) &&
563 (object->handle == NULL) &&
564 (object->type == OBJT_DEFAULT ||
565 object->type == OBJT_SWAP)) {
568 robject = LIST_FIRST(&object->shadow_head);
569 KASSERT(robject != NULL,
570 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
572 object->shadow_count));
573 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0,
574 ("shadowed tmpfs v_object %p", object));
575 if (!VM_OBJECT_TRYWLOCK(robject)) {
577 * Avoid a potential deadlock.
580 VM_OBJECT_WUNLOCK(object);
582 * More likely than not the thread
583 * holding robject's lock has lower
584 * priority than the current thread.
585 * Let the lower priority thread run.
591 * Collapse object into its shadow unless its
592 * shadow is dead. In that case, object will
593 * be deallocated by the thread that is
594 * deallocating its shadow.
596 if ((robject->flags & OBJ_DEAD) == 0 &&
597 (robject->handle == NULL) &&
598 (robject->type == OBJT_DEFAULT ||
599 robject->type == OBJT_SWAP)) {
601 robject->ref_count++;
603 if (robject->paging_in_progress) {
604 VM_OBJECT_WUNLOCK(object);
605 vm_object_pip_wait(robject,
607 temp = robject->backing_object;
608 if (object == temp) {
609 VM_OBJECT_WLOCK(object);
612 } else if (object->paging_in_progress) {
613 VM_OBJECT_WUNLOCK(robject);
614 object->flags |= OBJ_PIPWNT;
615 VM_OBJECT_SLEEP(object, object,
616 PDROP | PVM, "objde2", 0);
617 VM_OBJECT_WLOCK(robject);
618 temp = robject->backing_object;
619 if (object == temp) {
620 VM_OBJECT_WLOCK(object);
624 VM_OBJECT_WUNLOCK(object);
626 if (robject->ref_count == 1) {
627 robject->ref_count--;
632 vm_object_collapse(object);
633 VM_OBJECT_WUNLOCK(object);
636 VM_OBJECT_WUNLOCK(robject);
638 VM_OBJECT_WUNLOCK(object);
642 temp = object->backing_object;
644 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
645 ("shadowed tmpfs v_object 2 %p", object));
646 VM_OBJECT_WLOCK(temp);
647 LIST_REMOVE(object, shadow_list);
648 temp->shadow_count--;
649 VM_OBJECT_WUNLOCK(temp);
650 object->backing_object = NULL;
653 * Don't double-terminate, we could be in a termination
654 * recursion due to the terminate having to sync data
657 if ((object->flags & OBJ_DEAD) == 0)
658 vm_object_terminate(object);
660 VM_OBJECT_WUNLOCK(object);
666 * vm_object_destroy removes the object from the global object list
667 * and frees the space for the object.
670 vm_object_destroy(vm_object_t object)
674 * Remove the object from the global object list.
676 mtx_lock(&vm_object_list_mtx);
677 TAILQ_REMOVE(&vm_object_list, object, object_list);
678 mtx_unlock(&vm_object_list_mtx);
681 * Release the allocation charge.
683 if (object->cred != NULL) {
684 KASSERT(object->type == OBJT_DEFAULT ||
685 object->type == OBJT_SWAP,
686 ("%s: non-swap obj %p has cred", __func__, object));
687 swap_release_by_cred(object->charge, object->cred);
689 crfree(object->cred);
694 * Free the space for the object.
696 uma_zfree(obj_zone, object);
700 * vm_object_terminate actually destroys the specified object, freeing
701 * up all previously used resources.
703 * The object must be locked.
704 * This routine may block.
707 vm_object_terminate(vm_object_t object)
711 VM_OBJECT_ASSERT_WLOCKED(object);
714 * Make sure no one uses us.
716 vm_object_set_flag(object, OBJ_DEAD);
719 * wait for the pageout daemon to be done with the object
721 vm_object_pip_wait(object, "objtrm");
723 KASSERT(!object->paging_in_progress,
724 ("vm_object_terminate: pageout in progress"));
727 * Clean and free the pages, as appropriate. All references to the
728 * object are gone, so we don't need to lock it.
730 if (object->type == OBJT_VNODE) {
731 struct vnode *vp = (struct vnode *)object->handle;
734 * Clean pages and flush buffers.
736 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
737 VM_OBJECT_WUNLOCK(object);
739 vinvalbuf(vp, V_SAVE, 0, 0);
741 VM_OBJECT_WLOCK(object);
744 KASSERT(object->ref_count == 0,
745 ("vm_object_terminate: object with references, ref_count=%d",
749 * Free any remaining pageable pages. This also removes them from the
750 * paging queues. However, don't free wired pages, just remove them
751 * from the object. Rather than incrementally removing each page from
752 * the object, the page and object are reset to any empty state.
754 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
755 vm_page_assert_unbusied(p);
758 * Optimize the page's removal from the object by resetting
759 * its "object" field. Specifically, if the page is not
760 * wired, then the effect of this assignment is that
761 * vm_page_free()'s call to vm_page_remove() will return
762 * immediately without modifying the page or the object.
765 if (p->wire_count == 0) {
767 PCPU_INC(cnt.v_pfree);
772 * If the object contained any pages, then reset it to an empty state.
773 * None of the object's fields, including "resident_page_count", were
774 * modified by the preceding loop.
776 if (object->resident_page_count != 0) {
777 vm_radix_reclaim_allnodes(&object->rtree);
778 TAILQ_INIT(&object->memq);
779 object->resident_page_count = 0;
780 if (object->type == OBJT_VNODE)
781 vdrop(object->handle);
784 #if VM_NRESERVLEVEL > 0
785 if (__predict_false(!LIST_EMPTY(&object->rvq)))
786 vm_reserv_break_all(object);
788 if (__predict_false(!vm_object_cache_is_empty(object)))
789 vm_page_cache_free(object, 0, 0);
792 * Let the pager know object is dead.
794 vm_pager_deallocate(object);
795 VM_OBJECT_WUNLOCK(object);
797 vm_object_destroy(object);
801 * Make the page read-only so that we can clear the object flags. However, if
802 * this is a nosync mmap then the object is likely to stay dirty so do not
803 * mess with the page and do not clear the object flags. Returns TRUE if the
804 * page should be flushed, and FALSE otherwise.
807 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
811 * If we have been asked to skip nosync pages and this is a
812 * nosync page, skip it. Note that the object flags were not
813 * cleared in this case so we do not have to set them.
815 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
816 *clearobjflags = FALSE;
819 pmap_remove_write(p);
820 return (p->dirty != 0);
825 * vm_object_page_clean
827 * Clean all dirty pages in the specified range of object. Leaves page
828 * on whatever queue it is currently on. If NOSYNC is set then do not
829 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
830 * leaving the object dirty.
832 * When stuffing pages asynchronously, allow clustering. XXX we need a
833 * synchronous clustering mode implementation.
835 * Odd semantics: if start == end, we clean everything.
837 * The object must be locked.
839 * Returns FALSE if some page from the range was not written, as
840 * reported by the pager, and TRUE otherwise.
843 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
847 vm_pindex_t pi, tend, tstart;
848 int curgeneration, n, pagerflags;
849 boolean_t clearobjflags, eio, res;
851 VM_OBJECT_ASSERT_WLOCKED(object);
854 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
855 * objects. The check below prevents the function from
856 * operating on non-vnode objects.
858 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
859 object->resident_page_count == 0)
862 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
863 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
864 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
866 tstart = OFF_TO_IDX(start);
867 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
868 clearobjflags = tstart == 0 && tend >= object->size;
872 curgeneration = object->generation;
874 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
878 np = TAILQ_NEXT(p, listq);
881 if (vm_page_sleep_if_busy(p, "vpcwai")) {
882 if (object->generation != curgeneration) {
883 if ((flags & OBJPC_SYNC) != 0)
886 clearobjflags = FALSE;
888 np = vm_page_find_least(object, pi);
891 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
894 n = vm_object_page_collect_flush(object, p, pagerflags,
895 flags, &clearobjflags, &eio);
898 clearobjflags = FALSE;
900 if (object->generation != curgeneration) {
901 if ((flags & OBJPC_SYNC) != 0)
904 clearobjflags = FALSE;
908 * If the VOP_PUTPAGES() did a truncated write, so
909 * that even the first page of the run is not fully
910 * written, vm_pageout_flush() returns 0 as the run
911 * length. Since the condition that caused truncated
912 * write may be permanent, e.g. exhausted free space,
913 * accepting n == 0 would cause an infinite loop.
915 * Forwarding the iterator leaves the unwritten page
916 * behind, but there is not much we can do there if
917 * filesystem refuses to write it.
921 clearobjflags = FALSE;
923 np = vm_page_find_least(object, pi + n);
926 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
930 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
935 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
936 int flags, boolean_t *clearobjflags, boolean_t *eio)
938 vm_page_t ma[vm_pageout_page_count], p_first, tp;
939 int count, i, mreq, runlen;
941 vm_page_lock_assert(p, MA_NOTOWNED);
942 VM_OBJECT_ASSERT_WLOCKED(object);
947 for (tp = p; count < vm_pageout_page_count; count++) {
948 tp = vm_page_next(tp);
949 if (tp == NULL || vm_page_busied(tp))
951 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
955 for (p_first = p; count < vm_pageout_page_count; count++) {
956 tp = vm_page_prev(p_first);
957 if (tp == NULL || vm_page_busied(tp))
959 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
965 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
968 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
973 * Note that there is absolutely no sense in writing out
974 * anonymous objects, so we track down the vnode object
976 * We invalidate (remove) all pages from the address space
977 * for semantic correctness.
979 * If the backing object is a device object with unmanaged pages, then any
980 * mappings to the specified range of pages must be removed before this
981 * function is called.
983 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
984 * may start out with a NULL object.
987 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
988 boolean_t syncio, boolean_t invalidate)
990 vm_object_t backing_object;
993 int error, flags, fsync_after;
1000 VM_OBJECT_WLOCK(object);
1001 while ((backing_object = object->backing_object) != NULL) {
1002 VM_OBJECT_WLOCK(backing_object);
1003 offset += object->backing_object_offset;
1004 VM_OBJECT_WUNLOCK(object);
1005 object = backing_object;
1006 if (object->size < OFF_TO_IDX(offset + size))
1007 size = IDX_TO_OFF(object->size) - offset;
1010 * Flush pages if writing is allowed, invalidate them
1011 * if invalidation requested. Pages undergoing I/O
1012 * will be ignored by vm_object_page_remove().
1014 * We cannot lock the vnode and then wait for paging
1015 * to complete without deadlocking against vm_fault.
1016 * Instead we simply call vm_object_page_remove() and
1017 * allow it to block internally on a page-by-page
1018 * basis when it encounters pages undergoing async
1021 if (object->type == OBJT_VNODE &&
1022 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1023 vp = object->handle;
1024 VM_OBJECT_WUNLOCK(object);
1025 (void) vn_start_write(vp, &mp, V_WAIT);
1026 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1027 if (syncio && !invalidate && offset == 0 &&
1028 OFF_TO_IDX(size) == object->size) {
1030 * If syncing the whole mapping of the file,
1031 * it is faster to schedule all the writes in
1032 * async mode, also allowing the clustering,
1033 * and then wait for i/o to complete.
1038 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1039 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1040 fsync_after = FALSE;
1042 VM_OBJECT_WLOCK(object);
1043 res = vm_object_page_clean(object, offset, offset + size,
1045 VM_OBJECT_WUNLOCK(object);
1047 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1049 vn_finished_write(mp);
1052 VM_OBJECT_WLOCK(object);
1054 if ((object->type == OBJT_VNODE ||
1055 object->type == OBJT_DEVICE) && invalidate) {
1056 if (object->type == OBJT_DEVICE)
1058 * The option OBJPR_NOTMAPPED must be passed here
1059 * because vm_object_page_remove() cannot remove
1060 * unmanaged mappings.
1062 flags = OBJPR_NOTMAPPED;
1064 flags = OBJPR_NOTWIRED;
1066 flags = OBJPR_CLEANONLY | OBJPR_NOTWIRED;
1067 vm_object_page_remove(object, OFF_TO_IDX(offset),
1068 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1070 VM_OBJECT_WUNLOCK(object);
1075 * vm_object_madvise:
1077 * Implements the madvise function at the object/page level.
1079 * MADV_WILLNEED (any object)
1081 * Activate the specified pages if they are resident.
1083 * MADV_DONTNEED (any object)
1085 * Deactivate the specified pages if they are resident.
1087 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1088 * OBJ_ONEMAPPING only)
1090 * Deactivate and clean the specified pages if they are
1091 * resident. This permits the process to reuse the pages
1092 * without faulting or the kernel to reclaim the pages
1096 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1099 vm_pindex_t tpindex;
1100 vm_object_t backing_object, tobject;
1105 VM_OBJECT_WLOCK(object);
1107 * Locate and adjust resident pages
1109 for (; pindex < end; pindex += 1) {
1115 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1116 * and those pages must be OBJ_ONEMAPPING.
1118 if (advise == MADV_FREE) {
1119 if ((tobject->type != OBJT_DEFAULT &&
1120 tobject->type != OBJT_SWAP) ||
1121 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1122 goto unlock_tobject;
1124 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1125 goto unlock_tobject;
1126 m = vm_page_lookup(tobject, tpindex);
1127 if (m == NULL && advise == MADV_WILLNEED) {
1129 * If the page is cached, reactivate it.
1131 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1136 * There may be swap even if there is no backing page
1138 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1139 swap_pager_freespace(tobject, tpindex, 1);
1143 backing_object = tobject->backing_object;
1144 if (backing_object == NULL)
1145 goto unlock_tobject;
1146 VM_OBJECT_WLOCK(backing_object);
1147 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1148 if (tobject != object)
1149 VM_OBJECT_WUNLOCK(tobject);
1150 tobject = backing_object;
1152 } else if (m->valid != VM_PAGE_BITS_ALL)
1153 goto unlock_tobject;
1155 * If the page is not in a normal state, skip it.
1158 if (m->hold_count != 0 || m->wire_count != 0) {
1160 goto unlock_tobject;
1162 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1163 ("vm_object_madvise: page %p is fictitious", m));
1164 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1165 ("vm_object_madvise: page %p is not managed", m));
1166 if (vm_page_busied(m)) {
1167 if (advise == MADV_WILLNEED) {
1169 * Reference the page before unlocking and
1170 * sleeping so that the page daemon is less
1171 * likely to reclaim it.
1173 vm_page_aflag_set(m, PGA_REFERENCED);
1175 if (object != tobject)
1176 VM_OBJECT_WUNLOCK(object);
1177 VM_OBJECT_WUNLOCK(tobject);
1178 vm_page_busy_sleep(m, "madvpo");
1179 VM_OBJECT_WLOCK(object);
1182 if (advise == MADV_WILLNEED) {
1183 vm_page_activate(m);
1185 vm_page_advise(m, advise);
1188 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1189 swap_pager_freespace(tobject, tpindex, 1);
1191 if (tobject != object)
1192 VM_OBJECT_WUNLOCK(tobject);
1194 VM_OBJECT_WUNLOCK(object);
1200 * Create a new object which is backed by the
1201 * specified existing object range. The source
1202 * object reference is deallocated.
1204 * The new object and offset into that object
1205 * are returned in the source parameters.
1209 vm_object_t *object, /* IN/OUT */
1210 vm_ooffset_t *offset, /* IN/OUT */
1219 * Don't create the new object if the old object isn't shared.
1221 if (source != NULL) {
1222 VM_OBJECT_WLOCK(source);
1223 if (source->ref_count == 1 &&
1224 source->handle == NULL &&
1225 (source->type == OBJT_DEFAULT ||
1226 source->type == OBJT_SWAP)) {
1227 VM_OBJECT_WUNLOCK(source);
1230 VM_OBJECT_WUNLOCK(source);
1234 * Allocate a new object with the given length.
1236 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1239 * The new object shadows the source object, adding a reference to it.
1240 * Our caller changes his reference to point to the new object,
1241 * removing a reference to the source object. Net result: no change
1242 * of reference count.
1244 * Try to optimize the result object's page color when shadowing
1245 * in order to maintain page coloring consistency in the combined
1248 result->backing_object = source;
1250 * Store the offset into the source object, and fix up the offset into
1253 result->backing_object_offset = *offset;
1254 if (source != NULL) {
1255 VM_OBJECT_WLOCK(source);
1256 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1257 source->shadow_count++;
1258 #if VM_NRESERVLEVEL > 0
1259 result->flags |= source->flags & OBJ_COLORED;
1260 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1261 ((1 << (VM_NFREEORDER - 1)) - 1);
1263 VM_OBJECT_WUNLOCK(source);
1268 * Return the new things
1277 * Split the pages in a map entry into a new object. This affords
1278 * easier removal of unused pages, and keeps object inheritance from
1279 * being a negative impact on memory usage.
1282 vm_object_split(vm_map_entry_t entry)
1284 vm_page_t m, m_next;
1285 vm_object_t orig_object, new_object, source;
1286 vm_pindex_t idx, offidxstart;
1289 orig_object = entry->object.vm_object;
1290 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1292 if (orig_object->ref_count <= 1)
1294 VM_OBJECT_WUNLOCK(orig_object);
1296 offidxstart = OFF_TO_IDX(entry->offset);
1297 size = atop(entry->end - entry->start);
1300 * If swap_pager_copy() is later called, it will convert new_object
1301 * into a swap object.
1303 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1306 * At this point, the new object is still private, so the order in
1307 * which the original and new objects are locked does not matter.
1309 VM_OBJECT_WLOCK(new_object);
1310 VM_OBJECT_WLOCK(orig_object);
1311 source = orig_object->backing_object;
1312 if (source != NULL) {
1313 VM_OBJECT_WLOCK(source);
1314 if ((source->flags & OBJ_DEAD) != 0) {
1315 VM_OBJECT_WUNLOCK(source);
1316 VM_OBJECT_WUNLOCK(orig_object);
1317 VM_OBJECT_WUNLOCK(new_object);
1318 vm_object_deallocate(new_object);
1319 VM_OBJECT_WLOCK(orig_object);
1322 LIST_INSERT_HEAD(&source->shadow_head,
1323 new_object, shadow_list);
1324 source->shadow_count++;
1325 vm_object_reference_locked(source); /* for new_object */
1326 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1327 VM_OBJECT_WUNLOCK(source);
1328 new_object->backing_object_offset =
1329 orig_object->backing_object_offset + entry->offset;
1330 new_object->backing_object = source;
1332 if (orig_object->cred != NULL) {
1333 new_object->cred = orig_object->cred;
1334 crhold(orig_object->cred);
1335 new_object->charge = ptoa(size);
1336 KASSERT(orig_object->charge >= ptoa(size),
1337 ("orig_object->charge < 0"));
1338 orig_object->charge -= ptoa(size);
1341 m = vm_page_find_least(orig_object, offidxstart);
1342 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1344 m_next = TAILQ_NEXT(m, listq);
1347 * We must wait for pending I/O to complete before we can
1350 * We do not have to VM_PROT_NONE the page as mappings should
1351 * not be changed by this operation.
1353 if (vm_page_busied(m)) {
1354 VM_OBJECT_WUNLOCK(new_object);
1356 VM_OBJECT_WUNLOCK(orig_object);
1357 vm_page_busy_sleep(m, "spltwt");
1358 VM_OBJECT_WLOCK(orig_object);
1359 VM_OBJECT_WLOCK(new_object);
1363 /* vm_page_rename() will handle dirty and cache. */
1364 if (vm_page_rename(m, new_object, idx)) {
1365 VM_OBJECT_WUNLOCK(new_object);
1366 VM_OBJECT_WUNLOCK(orig_object);
1368 VM_OBJECT_WLOCK(orig_object);
1369 VM_OBJECT_WLOCK(new_object);
1372 #if VM_NRESERVLEVEL > 0
1374 * If some of the reservation's allocated pages remain with
1375 * the original object, then transferring the reservation to
1376 * the new object is neither particularly beneficial nor
1377 * particularly harmful as compared to leaving the reservation
1378 * with the original object. If, however, all of the
1379 * reservation's allocated pages are transferred to the new
1380 * object, then transferring the reservation is typically
1381 * beneficial. Determining which of these two cases applies
1382 * would be more costly than unconditionally renaming the
1385 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1387 if (orig_object->type == OBJT_SWAP)
1390 if (orig_object->type == OBJT_SWAP) {
1392 * swap_pager_copy() can sleep, in which case the orig_object's
1393 * and new_object's locks are released and reacquired.
1395 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1396 TAILQ_FOREACH(m, &new_object->memq, listq)
1400 * Transfer any cached pages from orig_object to new_object.
1401 * If swap_pager_copy() found swapped out pages within the
1402 * specified range of orig_object, then it changed
1403 * new_object's type to OBJT_SWAP when it transferred those
1404 * pages to new_object. Otherwise, new_object's type
1405 * should still be OBJT_DEFAULT and orig_object should not
1406 * contain any cached pages within the specified range.
1408 if (__predict_false(!vm_object_cache_is_empty(orig_object)))
1409 vm_page_cache_transfer(orig_object, offidxstart,
1412 VM_OBJECT_WUNLOCK(orig_object);
1413 VM_OBJECT_WUNLOCK(new_object);
1414 entry->object.vm_object = new_object;
1415 entry->offset = 0LL;
1416 vm_object_deallocate(orig_object);
1417 VM_OBJECT_WLOCK(new_object);
1420 #define OBSC_TEST_ALL_SHADOWED 0x0001
1421 #define OBSC_COLLAPSE_NOWAIT 0x0002
1422 #define OBSC_COLLAPSE_WAIT 0x0004
1425 vm_object_backing_scan(vm_object_t object, int op)
1429 vm_object_t backing_object;
1430 vm_pindex_t backing_offset_index;
1432 VM_OBJECT_ASSERT_WLOCKED(object);
1433 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1435 backing_object = object->backing_object;
1436 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1439 * Initial conditions
1441 if (op & OBSC_TEST_ALL_SHADOWED) {
1443 * We do not want to have to test for the existence of cache
1444 * or swap pages in the backing object. XXX but with the
1445 * new swapper this would be pretty easy to do.
1447 * XXX what about anonymous MAP_SHARED memory that hasn't
1448 * been ZFOD faulted yet? If we do not test for this, the
1449 * shadow test may succeed! XXX
1451 if (backing_object->type != OBJT_DEFAULT) {
1455 if (op & OBSC_COLLAPSE_WAIT) {
1456 vm_object_set_flag(backing_object, OBJ_DEAD);
1462 p = TAILQ_FIRST(&backing_object->memq);
1464 vm_page_t next = TAILQ_NEXT(p, listq);
1465 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1467 if (op & OBSC_TEST_ALL_SHADOWED) {
1471 * Ignore pages outside the parent object's range
1472 * and outside the parent object's mapping of the
1475 * note that we do not busy the backing object's
1479 p->pindex < backing_offset_index ||
1480 new_pindex >= object->size
1487 * See if the parent has the page or if the parent's
1488 * object pager has the page. If the parent has the
1489 * page but the page is not valid, the parent's
1490 * object pager must have the page.
1492 * If this fails, the parent does not completely shadow
1493 * the object and we might as well give up now.
1496 pp = vm_page_lookup(object, new_pindex);
1498 (pp == NULL || pp->valid == 0) &&
1499 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1507 * Check for busy page
1509 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1512 if (op & OBSC_COLLAPSE_NOWAIT) {
1513 if (!p->valid || vm_page_busied(p)) {
1517 } else if (op & OBSC_COLLAPSE_WAIT) {
1518 if (vm_page_busied(p)) {
1519 VM_OBJECT_WUNLOCK(object);
1521 VM_OBJECT_WUNLOCK(backing_object);
1522 vm_page_busy_sleep(p, "vmocol");
1523 VM_OBJECT_WLOCK(object);
1524 VM_OBJECT_WLOCK(backing_object);
1526 * If we slept, anything could have
1527 * happened. Since the object is
1528 * marked dead, the backing offset
1529 * should not have changed so we
1530 * just restart our scan.
1532 p = TAILQ_FIRST(&backing_object->memq);
1538 p->object == backing_object,
1539 ("vm_object_backing_scan: object mismatch")
1543 p->pindex < backing_offset_index ||
1544 new_pindex >= object->size
1546 if (backing_object->type == OBJT_SWAP)
1547 swap_pager_freespace(backing_object,
1551 * Page is out of the parent object's range, we
1552 * can simply destroy it.
1555 KASSERT(!pmap_page_is_mapped(p),
1556 ("freeing mapped page %p", p));
1557 if (p->wire_count == 0)
1566 pp = vm_page_lookup(object, new_pindex);
1568 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1569 (pp != NULL && pp->valid == 0)
1571 if (backing_object->type == OBJT_SWAP)
1572 swap_pager_freespace(backing_object,
1576 * The page in the parent is not (yet) valid.
1577 * We don't know anything about the state of
1578 * the original page. It might be mapped,
1579 * so we must avoid the next if here.
1581 * This is due to a race in vm_fault() where
1582 * we must unbusy the original (backing_obj)
1583 * page before we can (re)lock the parent.
1584 * Hence we can get here.
1591 vm_pager_has_page(object, new_pindex, NULL, NULL)
1593 if (backing_object->type == OBJT_SWAP)
1594 swap_pager_freespace(backing_object,
1598 * page already exists in parent OR swap exists
1599 * for this location in the parent. Destroy
1600 * the original page from the backing object.
1602 * Leave the parent's page alone
1605 KASSERT(!pmap_page_is_mapped(p),
1606 ("freeing mapped page %p", p));
1607 if (p->wire_count == 0)
1617 * Page does not exist in parent, rename the
1618 * page from the backing object to the main object.
1620 * If the page was mapped to a process, it can remain
1621 * mapped through the rename.
1622 * vm_page_rename() will handle dirty and cache.
1624 if (vm_page_rename(p, object, new_pindex)) {
1625 if (op & OBSC_COLLAPSE_NOWAIT) {
1629 VM_OBJECT_WLOCK(backing_object);
1630 VM_OBJECT_WUNLOCK(object);
1632 VM_OBJECT_WLOCK(object);
1633 VM_OBJECT_WLOCK(backing_object);
1634 p = TAILQ_FIRST(&backing_object->memq);
1638 /* Use the old pindex to free the right page. */
1639 if (backing_object->type == OBJT_SWAP)
1640 swap_pager_freespace(backing_object,
1641 new_pindex + backing_offset_index, 1);
1643 #if VM_NRESERVLEVEL > 0
1645 * Rename the reservation.
1647 vm_reserv_rename(p, object, backing_object,
1648 backing_offset_index);
1658 * this version of collapse allows the operation to occur earlier and
1659 * when paging_in_progress is true for an object... This is not a complete
1660 * operation, but should plug 99.9% of the rest of the leaks.
1663 vm_object_qcollapse(vm_object_t object)
1665 vm_object_t backing_object = object->backing_object;
1667 VM_OBJECT_ASSERT_WLOCKED(object);
1668 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1670 if (backing_object->ref_count != 1)
1673 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1677 * vm_object_collapse:
1679 * Collapse an object with the object backing it.
1680 * Pages in the backing object are moved into the
1681 * parent, and the backing object is deallocated.
1684 vm_object_collapse(vm_object_t object)
1686 VM_OBJECT_ASSERT_WLOCKED(object);
1689 vm_object_t backing_object;
1692 * Verify that the conditions are right for collapse:
1694 * The object exists and the backing object exists.
1696 if ((backing_object = object->backing_object) == NULL)
1700 * we check the backing object first, because it is most likely
1703 VM_OBJECT_WLOCK(backing_object);
1704 if (backing_object->handle != NULL ||
1705 (backing_object->type != OBJT_DEFAULT &&
1706 backing_object->type != OBJT_SWAP) ||
1707 (backing_object->flags & OBJ_DEAD) ||
1708 object->handle != NULL ||
1709 (object->type != OBJT_DEFAULT &&
1710 object->type != OBJT_SWAP) ||
1711 (object->flags & OBJ_DEAD)) {
1712 VM_OBJECT_WUNLOCK(backing_object);
1717 object->paging_in_progress != 0 ||
1718 backing_object->paging_in_progress != 0
1720 vm_object_qcollapse(object);
1721 VM_OBJECT_WUNLOCK(backing_object);
1725 * We know that we can either collapse the backing object (if
1726 * the parent is the only reference to it) or (perhaps) have
1727 * the parent bypass the object if the parent happens to shadow
1728 * all the resident pages in the entire backing object.
1730 * This is ignoring pager-backed pages such as swap pages.
1731 * vm_object_backing_scan fails the shadowing test in this
1734 if (backing_object->ref_count == 1) {
1736 * If there is exactly one reference to the backing
1737 * object, we can collapse it into the parent.
1739 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1741 #if VM_NRESERVLEVEL > 0
1743 * Break any reservations from backing_object.
1745 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1746 vm_reserv_break_all(backing_object);
1750 * Move the pager from backing_object to object.
1752 if (backing_object->type == OBJT_SWAP) {
1754 * swap_pager_copy() can sleep, in which case
1755 * the backing_object's and object's locks are
1756 * released and reacquired.
1757 * Since swap_pager_copy() is being asked to
1758 * destroy the source, it will change the
1759 * backing_object's type to OBJT_DEFAULT.
1764 OFF_TO_IDX(object->backing_object_offset), TRUE);
1767 * Free any cached pages from backing_object.
1769 if (__predict_false(
1770 !vm_object_cache_is_empty(backing_object)))
1771 vm_page_cache_free(backing_object, 0, 0);
1774 * Object now shadows whatever backing_object did.
1775 * Note that the reference to
1776 * backing_object->backing_object moves from within
1777 * backing_object to within object.
1779 LIST_REMOVE(object, shadow_list);
1780 backing_object->shadow_count--;
1781 if (backing_object->backing_object) {
1782 VM_OBJECT_WLOCK(backing_object->backing_object);
1783 LIST_REMOVE(backing_object, shadow_list);
1785 &backing_object->backing_object->shadow_head,
1786 object, shadow_list);
1788 * The shadow_count has not changed.
1790 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1792 object->backing_object = backing_object->backing_object;
1793 object->backing_object_offset +=
1794 backing_object->backing_object_offset;
1797 * Discard backing_object.
1799 * Since the backing object has no pages, no pager left,
1800 * and no object references within it, all that is
1801 * necessary is to dispose of it.
1803 KASSERT(backing_object->ref_count == 1, (
1804 "backing_object %p was somehow re-referenced during collapse!",
1806 VM_OBJECT_WUNLOCK(backing_object);
1807 vm_object_destroy(backing_object);
1811 vm_object_t new_backing_object;
1814 * If we do not entirely shadow the backing object,
1815 * there is nothing we can do so we give up.
1817 if (object->resident_page_count != object->size &&
1818 vm_object_backing_scan(object,
1819 OBSC_TEST_ALL_SHADOWED) == 0) {
1820 VM_OBJECT_WUNLOCK(backing_object);
1825 * Make the parent shadow the next object in the
1826 * chain. Deallocating backing_object will not remove
1827 * it, since its reference count is at least 2.
1829 LIST_REMOVE(object, shadow_list);
1830 backing_object->shadow_count--;
1832 new_backing_object = backing_object->backing_object;
1833 if ((object->backing_object = new_backing_object) != NULL) {
1834 VM_OBJECT_WLOCK(new_backing_object);
1836 &new_backing_object->shadow_head,
1840 new_backing_object->shadow_count++;
1841 vm_object_reference_locked(new_backing_object);
1842 VM_OBJECT_WUNLOCK(new_backing_object);
1843 object->backing_object_offset +=
1844 backing_object->backing_object_offset;
1848 * Drop the reference count on backing_object. Since
1849 * its ref_count was at least 2, it will not vanish.
1851 backing_object->ref_count--;
1852 VM_OBJECT_WUNLOCK(backing_object);
1857 * Try again with this object's new backing object.
1863 * vm_object_page_remove:
1865 * For the given object, either frees or invalidates each of the
1866 * specified pages. In general, a page is freed. However, if a page is
1867 * wired for any reason other than the existence of a managed, wired
1868 * mapping, then it may be invalidated but not removed from the object.
1869 * Pages are specified by the given range ["start", "end") and the option
1870 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1871 * extends from "start" to the end of the object. If the option
1872 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1873 * specified range are affected. If the option OBJPR_NOTMAPPED is
1874 * specified, then the pages within the specified range must have no
1875 * mappings. Otherwise, if this option is not specified, any mappings to
1876 * the specified pages are removed before the pages are freed or
1879 * In general, this operation should only be performed on objects that
1880 * contain managed pages. There are, however, two exceptions. First, it
1881 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1882 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1883 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1884 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1886 * The object must be locked.
1889 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1895 VM_OBJECT_ASSERT_WLOCKED(object);
1896 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1897 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1898 ("vm_object_page_remove: illegal options for object %p", object));
1899 if (object->resident_page_count == 0)
1901 vm_object_pip_add(object, 1);
1903 p = vm_page_find_least(object, start);
1906 * Here, the variable "p" is either (1) the page with the least pindex
1907 * greater than or equal to the parameter "start" or (2) NULL.
1909 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1910 next = TAILQ_NEXT(p, listq);
1913 * If the page is wired for any reason besides the existence
1914 * of managed, wired mappings, then it cannot be freed. For
1915 * example, fictitious pages, which represent device memory,
1916 * are inherently wired and cannot be freed. They can,
1917 * however, be invalidated if the option OBJPR_CLEANONLY is
1921 if (vm_page_xbusied(p)) {
1922 VM_OBJECT_WUNLOCK(object);
1923 vm_page_busy_sleep(p, "vmopax");
1924 VM_OBJECT_WLOCK(object);
1927 if ((wirings = p->wire_count) != 0 &&
1928 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1929 if ((options & (OBJPR_NOTWIRED | OBJPR_NOTMAPPED)) ==
1932 /* Account for removal of wired mappings. */
1934 p->wire_count -= wirings;
1936 if ((options & OBJPR_CLEANONLY) == 0) {
1942 if (vm_page_busied(p)) {
1943 VM_OBJECT_WUNLOCK(object);
1944 vm_page_busy_sleep(p, "vmopar");
1945 VM_OBJECT_WLOCK(object);
1948 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1949 ("vm_object_page_remove: page %p is fictitious", p));
1950 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1951 if ((options & OBJPR_NOTMAPPED) == 0)
1952 pmap_remove_write(p);
1956 if ((options & OBJPR_NOTMAPPED) == 0) {
1957 if ((options & OBJPR_NOTWIRED) != 0 && wirings != 0)
1960 /* Account for removal of wired mappings. */
1962 KASSERT(p->wire_count == wirings,
1963 ("inconsistent wire count %d %d %p",
1964 p->wire_count, wirings, p));
1966 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1973 vm_object_pip_wakeup(object);
1975 if (__predict_false(!vm_object_cache_is_empty(object)))
1976 vm_page_cache_free(object, start, end);
1980 * vm_object_page_cache:
1982 * For the given object, attempt to move the specified clean
1983 * pages to the cache queue. If a page is wired for any reason,
1984 * then it will not be changed. Pages are specified by the given
1985 * range ["start", "end"). As a special case, if "end" is zero,
1986 * then the range extends from "start" to the end of the object.
1987 * Any mappings to the specified pages are removed before the
1988 * pages are moved to the cache queue.
1990 * This operation should only be performed on objects that
1991 * contain non-fictitious, managed pages.
1993 * The object must be locked.
1996 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1998 struct mtx *mtx, *new_mtx;
2001 VM_OBJECT_ASSERT_WLOCKED(object);
2002 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2003 ("vm_object_page_cache: illegal object %p", object));
2004 if (object->resident_page_count == 0)
2006 p = vm_page_find_least(object, start);
2009 * Here, the variable "p" is either (1) the page with the least pindex
2010 * greater than or equal to the parameter "start" or (2) NULL.
2013 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2014 next = TAILQ_NEXT(p, listq);
2017 * Avoid releasing and reacquiring the same page lock.
2019 new_mtx = vm_page_lockptr(p);
2020 if (mtx != new_mtx) {
2026 vm_page_try_to_cache(p);
2033 * Populate the specified range of the object with valid pages. Returns
2034 * TRUE if the range is successfully populated and FALSE otherwise.
2036 * Note: This function should be optimized to pass a larger array of
2037 * pages to vm_pager_get_pages() before it is applied to a non-
2038 * OBJT_DEVICE object.
2040 * The object must be locked.
2043 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2049 VM_OBJECT_ASSERT_WLOCKED(object);
2050 for (pindex = start; pindex < end; pindex++) {
2051 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
2052 if (m->valid != VM_PAGE_BITS_ALL) {
2054 rv = vm_pager_get_pages(object, ma, 1, 0);
2055 m = vm_page_lookup(object, pindex);
2058 if (rv != VM_PAGER_OK) {
2066 * Keep "m" busy because a subsequent iteration may unlock
2070 if (pindex > start) {
2071 m = vm_page_lookup(object, start);
2072 while (m != NULL && m->pindex < pindex) {
2074 m = TAILQ_NEXT(m, listq);
2077 return (pindex == end);
2081 * Routine: vm_object_coalesce
2082 * Function: Coalesces two objects backing up adjoining
2083 * regions of memory into a single object.
2085 * returns TRUE if objects were combined.
2087 * NOTE: Only works at the moment if the second object is NULL -
2088 * if it's not, which object do we lock first?
2091 * prev_object First object to coalesce
2092 * prev_offset Offset into prev_object
2093 * prev_size Size of reference to prev_object
2094 * next_size Size of reference to the second object
2095 * reserved Indicator that extension region has
2096 * swap accounted for
2099 * The object must *not* be locked.
2102 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2103 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2105 vm_pindex_t next_pindex;
2107 if (prev_object == NULL)
2109 VM_OBJECT_WLOCK(prev_object);
2110 if ((prev_object->type != OBJT_DEFAULT &&
2111 prev_object->type != OBJT_SWAP) ||
2112 (prev_object->flags & OBJ_TMPFS_NODE) != 0) {
2113 VM_OBJECT_WUNLOCK(prev_object);
2118 * Try to collapse the object first
2120 vm_object_collapse(prev_object);
2123 * Can't coalesce if: . more than one reference . paged out . shadows
2124 * another object . has a copy elsewhere (any of which mean that the
2125 * pages not mapped to prev_entry may be in use anyway)
2127 if (prev_object->backing_object != NULL) {
2128 VM_OBJECT_WUNLOCK(prev_object);
2132 prev_size >>= PAGE_SHIFT;
2133 next_size >>= PAGE_SHIFT;
2134 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2136 if ((prev_object->ref_count > 1) &&
2137 (prev_object->size != next_pindex)) {
2138 VM_OBJECT_WUNLOCK(prev_object);
2143 * Account for the charge.
2145 if (prev_object->cred != NULL) {
2148 * If prev_object was charged, then this mapping,
2149 * althought not charged now, may become writable
2150 * later. Non-NULL cred in the object would prevent
2151 * swap reservation during enabling of the write
2152 * access, so reserve swap now. Failed reservation
2153 * cause allocation of the separate object for the map
2154 * entry, and swap reservation for this entry is
2155 * managed in appropriate time.
2157 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2158 prev_object->cred)) {
2161 prev_object->charge += ptoa(next_size);
2165 * Remove any pages that may still be in the object from a previous
2168 if (next_pindex < prev_object->size) {
2169 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2171 if (prev_object->type == OBJT_SWAP)
2172 swap_pager_freespace(prev_object,
2173 next_pindex, next_size);
2175 if (prev_object->cred != NULL) {
2176 KASSERT(prev_object->charge >=
2177 ptoa(prev_object->size - next_pindex),
2178 ("object %p overcharged 1 %jx %jx", prev_object,
2179 (uintmax_t)next_pindex, (uintmax_t)next_size));
2180 prev_object->charge -= ptoa(prev_object->size -
2187 * Extend the object if necessary.
2189 if (next_pindex + next_size > prev_object->size)
2190 prev_object->size = next_pindex + next_size;
2192 VM_OBJECT_WUNLOCK(prev_object);
2197 vm_object_set_writeable_dirty(vm_object_t object)
2200 VM_OBJECT_ASSERT_WLOCKED(object);
2201 if (object->type != OBJT_VNODE) {
2202 if ((object->flags & OBJ_TMPFS_NODE) != 0) {
2203 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs"));
2204 vm_object_set_flag(object, OBJ_TMPFS_DIRTY);
2208 object->generation++;
2209 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2211 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2217 * For each page offset within the specified range of the given object,
2218 * find the highest-level page in the shadow chain and unwire it. A page
2219 * must exist at every page offset, and the highest-level page must be
2223 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2226 vm_object_t tobject;
2228 vm_pindex_t end_pindex, pindex, tpindex;
2229 int depth, locked_depth;
2231 KASSERT((offset & PAGE_MASK) == 0,
2232 ("vm_object_unwire: offset is not page aligned"));
2233 KASSERT((length & PAGE_MASK) == 0,
2234 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2235 /* The wired count of a fictitious page never changes. */
2236 if ((object->flags & OBJ_FICTITIOUS) != 0)
2238 pindex = OFF_TO_IDX(offset);
2239 end_pindex = pindex + atop(length);
2241 VM_OBJECT_RLOCK(object);
2242 m = vm_page_find_least(object, pindex);
2243 while (pindex < end_pindex) {
2244 if (m == NULL || pindex < m->pindex) {
2246 * The first object in the shadow chain doesn't
2247 * contain a page at the current index. Therefore,
2248 * the page must exist in a backing object.
2255 OFF_TO_IDX(tobject->backing_object_offset);
2256 tobject = tobject->backing_object;
2257 KASSERT(tobject != NULL,
2258 ("vm_object_unwire: missing page"));
2259 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2262 if (depth == locked_depth) {
2264 VM_OBJECT_RLOCK(tobject);
2266 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2270 m = TAILQ_NEXT(m, listq);
2273 vm_page_unwire(tm, queue);
2278 /* Release the accumulated object locks. */
2279 for (depth = 0; depth < locked_depth; depth++) {
2280 tobject = object->backing_object;
2281 VM_OBJECT_RUNLOCK(object);
2286 #include "opt_ddb.h"
2288 #include <sys/kernel.h>
2290 #include <sys/cons.h>
2292 #include <ddb/ddb.h>
2295 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2298 vm_map_entry_t tmpe;
2306 tmpe = map->header.next;
2307 entcount = map->nentries;
2308 while (entcount-- && (tmpe != &map->header)) {
2309 if (_vm_object_in_map(map, object, tmpe)) {
2314 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2315 tmpm = entry->object.sub_map;
2316 tmpe = tmpm->header.next;
2317 entcount = tmpm->nentries;
2318 while (entcount-- && tmpe != &tmpm->header) {
2319 if (_vm_object_in_map(tmpm, object, tmpe)) {
2324 } else if ((obj = entry->object.vm_object) != NULL) {
2325 for (; obj; obj = obj->backing_object)
2326 if (obj == object) {
2334 vm_object_in_map(vm_object_t object)
2338 /* sx_slock(&allproc_lock); */
2339 FOREACH_PROC_IN_SYSTEM(p) {
2340 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2342 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2343 /* sx_sunlock(&allproc_lock); */
2347 /* sx_sunlock(&allproc_lock); */
2348 if (_vm_object_in_map(kernel_map, object, 0))
2353 DB_SHOW_COMMAND(vmochk, vm_object_check)
2358 * make sure that internal objs are in a map somewhere
2359 * and none have zero ref counts.
2361 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2362 if (object->handle == NULL &&
2363 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2364 if (object->ref_count == 0) {
2365 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2366 (long)object->size);
2368 if (!vm_object_in_map(object)) {
2370 "vmochk: internal obj is not in a map: "
2371 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2372 object->ref_count, (u_long)object->size,
2373 (u_long)object->size,
2374 (void *)object->backing_object);
2381 * vm_object_print: [ debug ]
2383 DB_SHOW_COMMAND(object, vm_object_print_static)
2385 /* XXX convert args. */
2386 vm_object_t object = (vm_object_t)addr;
2387 boolean_t full = have_addr;
2391 /* XXX count is an (unused) arg. Avoid shadowing it. */
2392 #define count was_count
2400 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2401 object, (int)object->type, (uintmax_t)object->size,
2402 object->resident_page_count, object->ref_count, object->flags,
2403 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2404 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2405 object->shadow_count,
2406 object->backing_object ? object->backing_object->ref_count : 0,
2407 object->backing_object, (uintmax_t)object->backing_object_offset);
2414 TAILQ_FOREACH(p, &object->memq, listq) {
2416 db_iprintf("memory:=");
2417 else if (count == 6) {
2425 db_printf("(off=0x%jx,page=0x%jx)",
2426 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2436 /* XXX need this non-static entry for calling from vm_map_print. */
2439 /* db_expr_t */ long addr,
2440 boolean_t have_addr,
2441 /* db_expr_t */ long count,
2444 vm_object_print_static(addr, have_addr, count, modif);
2447 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2452 vm_page_t m, prev_m;
2456 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2457 db_printf("new object: %p\n", (void *)object);
2468 TAILQ_FOREACH(m, &object->memq, listq) {
2469 if (m->pindex > 128)
2471 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2472 prev_m->pindex + 1 != m->pindex) {
2474 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2475 (long)fidx, rcount, (long)pa);
2487 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2492 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2493 (long)fidx, rcount, (long)pa);
2503 pa = VM_PAGE_TO_PHYS(m);
2507 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2508 (long)fidx, rcount, (long)pa);