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(object->resident_page_count == 0,
168 ("object %p resident_page_count = %d",
169 object, object->resident_page_count));
170 KASSERT(TAILQ_EMPTY(&object->memq),
171 ("object %p has resident pages",
173 #if VM_NRESERVLEVEL > 0
174 KASSERT(LIST_EMPTY(&object->rvq),
175 ("object %p has reservations",
178 KASSERT(object->cached_page_count == 0,
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->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 mtx_init(&object->mtx, "vm object", NULL, MTX_DEF | MTX_DUPOK);
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);
213 object->rtree.rt_root = 0;
217 panic("_vm_object_allocate: can't create OBJT_DEAD");
220 object->flags = OBJ_ONEMAPPING;
224 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
227 object->flags = OBJ_FICTITIOUS;
230 object->flags = OBJ_UNMANAGED;
236 panic("_vm_object_allocate: type %d is undefined", type);
239 object->generation = 1;
240 object->ref_count = 1;
241 object->memattr = VM_MEMATTR_DEFAULT;
244 object->pg_color = 0;
245 object->handle = NULL;
246 object->backing_object = NULL;
247 object->backing_object_offset = (vm_ooffset_t) 0;
248 #if VM_NRESERVLEVEL > 0
249 LIST_INIT(&object->rvq);
252 mtx_lock(&vm_object_list_mtx);
253 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
254 mtx_unlock(&vm_object_list_mtx);
260 * Initialize the VM objects module.
265 TAILQ_INIT(&vm_object_list);
266 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
268 mtx_init(&kernel_object->mtx, "vm object", "kernel object", MTX_DEF);
269 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
271 #if VM_NRESERVLEVEL > 0
272 kernel_object->flags |= OBJ_COLORED;
273 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
276 mtx_init(&kmem_object->mtx, "vm object", "kmem object", MTX_DEF);
277 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
279 #if VM_NRESERVLEVEL > 0
280 kmem_object->flags |= OBJ_COLORED;
281 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
285 * The lock portion of struct vm_object must be type stable due
286 * to vm_pageout_fallback_object_lock locking a vm object
287 * without holding any references to it.
289 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
295 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
299 vm_object_clear_flag(vm_object_t object, u_short bits)
302 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
303 object->flags &= ~bits;
307 * Sets the default memory attribute for the specified object. Pages
308 * that are allocated to this object are by default assigned this memory
311 * Presently, this function must be called before any pages are allocated
312 * to the object. In the future, this requirement may be relaxed for
313 * "default" and "swap" objects.
316 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
319 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
320 switch (object->type) {
328 if (object->resident_page_count == 0)
329 return (KERN_FAILURE);
332 return (KERN_INVALID_ARGUMENT);
334 panic("vm_object_set_memattr: object %p is of undefined type",
337 object->memattr = memattr;
338 return (KERN_SUCCESS);
342 vm_object_pip_add(vm_object_t object, short i)
345 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
346 object->paging_in_progress += i;
350 vm_object_pip_subtract(vm_object_t object, short i)
353 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
354 object->paging_in_progress -= i;
358 vm_object_pip_wakeup(vm_object_t object)
361 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
362 object->paging_in_progress--;
363 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
364 vm_object_clear_flag(object, OBJ_PIPWNT);
370 vm_object_pip_wakeupn(vm_object_t object, short i)
373 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
375 object->paging_in_progress -= i;
376 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
377 vm_object_clear_flag(object, OBJ_PIPWNT);
383 vm_object_pip_wait(vm_object_t object, char *waitid)
386 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
387 while (object->paging_in_progress) {
388 object->flags |= OBJ_PIPWNT;
389 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
394 * vm_object_allocate:
396 * Returns a new object with the given size.
399 vm_object_allocate(objtype_t type, vm_pindex_t size)
403 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
404 _vm_object_allocate(type, size, object);
410 * vm_object_reference:
412 * Gets another reference to the given object. Note: OBJ_DEAD
413 * objects can be referenced during final cleaning.
416 vm_object_reference(vm_object_t object)
420 VM_OBJECT_LOCK(object);
421 vm_object_reference_locked(object);
422 VM_OBJECT_UNLOCK(object);
426 * vm_object_reference_locked:
428 * Gets another reference to the given object.
430 * The object must be locked.
433 vm_object_reference_locked(vm_object_t object)
437 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
439 if (object->type == OBJT_VNODE) {
446 * Handle deallocating an object of type OBJT_VNODE.
449 vm_object_vndeallocate(vm_object_t object)
451 struct vnode *vp = (struct vnode *) object->handle;
453 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
454 KASSERT(object->type == OBJT_VNODE,
455 ("vm_object_vndeallocate: not a vnode object"));
456 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
458 if (object->ref_count == 0) {
459 vprint("vm_object_vndeallocate", vp);
460 panic("vm_object_vndeallocate: bad object reference count");
464 if (object->ref_count > 1) {
466 VM_OBJECT_UNLOCK(object);
467 /* vrele may need the vnode lock. */
471 VM_OBJECT_UNLOCK(object);
472 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
474 VM_OBJECT_LOCK(object);
476 if (object->type == OBJT_DEAD) {
477 VM_OBJECT_UNLOCK(object);
480 if (object->ref_count == 0)
482 VM_OBJECT_UNLOCK(object);
489 * vm_object_deallocate:
491 * Release a reference to the specified object,
492 * gained either through a vm_object_allocate
493 * or a vm_object_reference call. When all references
494 * are gone, storage associated with this object
495 * may be relinquished.
497 * No object may be locked.
500 vm_object_deallocate(vm_object_t object)
504 while (object != NULL) {
505 VM_OBJECT_LOCK(object);
506 if (object->type == OBJT_VNODE) {
507 vm_object_vndeallocate(object);
511 KASSERT(object->ref_count != 0,
512 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
515 * If the reference count goes to 0 we start calling
516 * vm_object_terminate() on the object chain.
517 * A ref count of 1 may be a special case depending on the
518 * shadow count being 0 or 1.
521 if (object->ref_count > 1) {
522 VM_OBJECT_UNLOCK(object);
524 } else if (object->ref_count == 1) {
525 if (object->shadow_count == 0 &&
526 object->handle == NULL &&
527 (object->type == OBJT_DEFAULT ||
528 object->type == OBJT_SWAP)) {
529 vm_object_set_flag(object, OBJ_ONEMAPPING);
530 } else if ((object->shadow_count == 1) &&
531 (object->handle == NULL) &&
532 (object->type == OBJT_DEFAULT ||
533 object->type == OBJT_SWAP)) {
536 robject = LIST_FIRST(&object->shadow_head);
537 KASSERT(robject != NULL,
538 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
540 object->shadow_count));
541 if (!VM_OBJECT_TRYLOCK(robject)) {
543 * Avoid a potential deadlock.
546 VM_OBJECT_UNLOCK(object);
548 * More likely than not the thread
549 * holding robject's lock has lower
550 * priority than the current thread.
551 * Let the lower priority thread run.
557 * Collapse object into its shadow unless its
558 * shadow is dead. In that case, object will
559 * be deallocated by the thread that is
560 * deallocating its shadow.
562 if ((robject->flags & OBJ_DEAD) == 0 &&
563 (robject->handle == NULL) &&
564 (robject->type == OBJT_DEFAULT ||
565 robject->type == OBJT_SWAP)) {
567 robject->ref_count++;
569 if (robject->paging_in_progress) {
570 VM_OBJECT_UNLOCK(object);
571 vm_object_pip_wait(robject,
573 temp = robject->backing_object;
574 if (object == temp) {
575 VM_OBJECT_LOCK(object);
578 } else if (object->paging_in_progress) {
579 VM_OBJECT_UNLOCK(robject);
580 object->flags |= OBJ_PIPWNT;
581 VM_OBJECT_SLEEP(object, object,
582 PDROP | PVM, "objde2", 0);
583 VM_OBJECT_LOCK(robject);
584 temp = robject->backing_object;
585 if (object == temp) {
586 VM_OBJECT_LOCK(object);
590 VM_OBJECT_UNLOCK(object);
592 if (robject->ref_count == 1) {
593 robject->ref_count--;
598 vm_object_collapse(object);
599 VM_OBJECT_UNLOCK(object);
602 VM_OBJECT_UNLOCK(robject);
604 VM_OBJECT_UNLOCK(object);
608 temp = object->backing_object;
610 VM_OBJECT_LOCK(temp);
611 LIST_REMOVE(object, shadow_list);
612 temp->shadow_count--;
613 VM_OBJECT_UNLOCK(temp);
614 object->backing_object = NULL;
617 * Don't double-terminate, we could be in a termination
618 * recursion due to the terminate having to sync data
621 if ((object->flags & OBJ_DEAD) == 0)
622 vm_object_terminate(object);
624 VM_OBJECT_UNLOCK(object);
630 * vm_object_destroy removes the object from the global object list
631 * and frees the space for the object.
634 vm_object_destroy(vm_object_t object)
638 * Remove the object from the global object list.
640 mtx_lock(&vm_object_list_mtx);
641 TAILQ_REMOVE(&vm_object_list, object, object_list);
642 mtx_unlock(&vm_object_list_mtx);
645 * Release the allocation charge.
647 if (object->cred != NULL) {
648 KASSERT(object->type == OBJT_DEFAULT ||
649 object->type == OBJT_SWAP,
650 ("vm_object_terminate: non-swap obj %p has cred",
652 swap_release_by_cred(object->charge, object->cred);
654 crfree(object->cred);
659 * Free the space for the object.
661 uma_zfree(obj_zone, object);
665 * vm_object_terminate actually destroys the specified object, freeing
666 * up all previously used resources.
668 * The object must be locked.
669 * This routine may block.
672 vm_object_terminate(vm_object_t object)
674 vm_page_t pa[VM_RADIX_STACK];
680 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
683 * Make sure no one uses us.
685 vm_object_set_flag(object, OBJ_DEAD);
688 * wait for the pageout daemon to be done with the object
690 vm_object_pip_wait(object, "objtrm");
692 KASSERT(!object->paging_in_progress,
693 ("vm_object_terminate: pageout in progress"));
696 * Clean and free the pages, as appropriate. All references to the
697 * object are gone, so we don't need to lock it.
699 if (object->type == OBJT_VNODE) {
700 struct vnode *vp = (struct vnode *)object->handle;
703 * Clean pages and flush buffers.
705 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
706 VM_OBJECT_UNLOCK(object);
708 vinvalbuf(vp, V_SAVE, 0, 0);
710 VM_OBJECT_LOCK(object);
713 KASSERT(object->ref_count == 0,
714 ("vm_object_terminate: object with references, ref_count=%d",
718 * Free any remaining pageable pages. This also removes them from the
719 * paging queues. However, don't free wired pages, just remove them
720 * from the object. Rather than incrementally removing each page from
721 * the object, the page and object are reset to any empty state.
725 while (exhausted == 0 && (n = vm_radix_lookupn(&object->rtree, start,
726 0, VM_RADIX_ANY, (void **)pa, VM_RADIX_STACK, &start,
728 for (i = 0; i < n; i++) {
731 * Another thread may allocate this cached page from
732 * the queue before we acquire the page queue free
735 if (p->flags & PG_CACHED) {
736 mtx_lock(&vm_page_queue_free_mtx);
737 if (p->object == object) {
740 /* Clear PG_CACHED and set PG_FREE. */
741 p->flags ^= PG_CACHED | PG_FREE;
745 mtx_unlock(&vm_page_queue_free_mtx);
747 } else if (p->object != object)
749 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
750 ("vm_object_terminate: freeing busy page %p", p));
753 * Optimize the page's removal from the object by
754 * resetting its "object" field. Specifically, if
755 * the page is not wired, then the effect of this
756 * assignment is that vm_page_free()'s call to
757 * vm_page_remove() will return immediately without
758 * modifying the page or the object.
759 * Anyway, the radix tree cannot be accessed anymore
760 * from within the object, thus all the nodes need
761 * to be reclaimed later on.
764 if (p->wire_count == 0) {
766 PCPU_INC(cnt.v_pfree);
770 if (n < VM_RADIX_STACK)
773 vm_radix_reclaim_allnodes(&object->rtree);
775 * If the object contained any pages, then reset it to an empty state.
776 * None of the object's fields, including "resident_page_count", were
777 * modified by the preceding loop.
779 if (object->resident_page_count != 0) {
780 TAILQ_INIT(&object->memq);
781 object->resident_page_count = 0;
782 if (object->type == OBJT_VNODE)
783 vdrop(object->handle);
785 if (object->cached_page_count != 0) {
786 object->cached_page_count = 0;
787 if (object->type == OBJT_VNODE)
788 vdrop(object->handle);
791 #if VM_NRESERVLEVEL > 0
792 if (__predict_false(!LIST_EMPTY(&object->rvq)))
793 vm_reserv_break_all(object);
797 * Let the pager know object is dead.
799 vm_pager_deallocate(object);
800 VM_OBJECT_UNLOCK(object);
802 vm_object_destroy(object);
806 * Make the page read-only so that we can clear the object flags. However, if
807 * this is a nosync mmap then the object is likely to stay dirty so do not
808 * mess with the page and do not clear the object flags. Returns TRUE if the
809 * page should be flushed, and FALSE otherwise.
812 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
816 * If we have been asked to skip nosync pages and this is a
817 * nosync page, skip it. Note that the object flags were not
818 * cleared in this case so we do not have to set them.
820 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
821 *clearobjflags = FALSE;
824 pmap_remove_write(p);
825 return (p->dirty != 0);
830 * vm_object_page_clean
832 * Clean all dirty pages in the specified range of object. Leaves page
833 * on whatever queue it is currently on. If NOSYNC is set then do not
834 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
835 * leaving the object dirty.
837 * When stuffing pages asynchronously, allow clustering. XXX we need a
838 * synchronous clustering mode implementation.
840 * Odd semantics: if start == end, we clean everything.
842 * The object must be locked.
844 * Returns FALSE if some page from the range was not written, as
845 * reported by the pager, and TRUE otherwise.
848 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
852 vm_pindex_t pi, tend, tstart;
853 int curgeneration, n, pagerflags;
854 boolean_t clearobjflags, eio, res;
856 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
857 KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
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, TRUE, "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_LOCK_ASSERT(object, MA_OWNED);
947 for (tp = p; count < vm_pageout_page_count; count++) {
948 tp = vm_page_next(tp);
949 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
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 || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
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_LOCK(object);
1001 while ((backing_object = object->backing_object) != NULL) {
1002 VM_OBJECT_LOCK(backing_object);
1003 offset += object->backing_object_offset;
1004 VM_OBJECT_UNLOCK(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_UNLOCK(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_LOCK(object);
1043 res = vm_object_page_clean(object, offset, offset + size,
1045 VM_OBJECT_UNLOCK(object);
1047 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1049 vn_finished_write(mp);
1052 VM_OBJECT_LOCK(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;
1066 flags = OBJPR_CLEANONLY;
1067 vm_object_page_remove(object, OFF_TO_IDX(offset),
1068 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1070 VM_OBJECT_UNLOCK(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_LOCK(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_LOCK(backing_object);
1147 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1148 if (tobject != object)
1149 VM_OBJECT_UNLOCK(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 ((m->oflags & VPO_BUSY) || m->busy) {
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);
1176 if (object != tobject)
1177 VM_OBJECT_UNLOCK(object);
1178 m->oflags |= VPO_WANTED;
1179 VM_OBJECT_SLEEP(tobject, m, PDROP | PVM, "madvpo", 0);
1180 VM_OBJECT_LOCK(object);
1183 if (advise == MADV_WILLNEED) {
1184 vm_page_activate(m);
1185 } else if (advise == MADV_DONTNEED) {
1186 vm_page_dontneed(m);
1187 } else if (advise == MADV_FREE) {
1189 * Mark the page clean. This will allow the page
1190 * to be freed up by the system. However, such pages
1191 * are often reused quickly by malloc()/free()
1192 * so we do not do anything that would cause
1193 * a page fault if we can help it.
1195 * Specifically, we do not try to actually free
1196 * the page now nor do we try to put it in the
1197 * cache (which would cause a page fault on reuse).
1199 * But we do make the page is freeable as we
1200 * can without actually taking the step of unmapping
1203 pmap_clear_modify(m);
1206 vm_page_dontneed(m);
1209 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1210 swap_pager_freespace(tobject, tpindex, 1);
1212 if (tobject != object)
1213 VM_OBJECT_UNLOCK(tobject);
1215 VM_OBJECT_UNLOCK(object);
1221 * Create a new object which is backed by the
1222 * specified existing object range. The source
1223 * object reference is deallocated.
1225 * The new object and offset into that object
1226 * are returned in the source parameters.
1230 vm_object_t *object, /* IN/OUT */
1231 vm_ooffset_t *offset, /* IN/OUT */
1240 * Don't create the new object if the old object isn't shared.
1242 if (source != NULL) {
1243 VM_OBJECT_LOCK(source);
1244 if (source->ref_count == 1 &&
1245 source->handle == NULL &&
1246 (source->type == OBJT_DEFAULT ||
1247 source->type == OBJT_SWAP)) {
1248 VM_OBJECT_UNLOCK(source);
1251 VM_OBJECT_UNLOCK(source);
1255 * Allocate a new object with the given length.
1257 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1260 * The new object shadows the source object, adding a reference to it.
1261 * Our caller changes his reference to point to the new object,
1262 * removing a reference to the source object. Net result: no change
1263 * of reference count.
1265 * Try to optimize the result object's page color when shadowing
1266 * in order to maintain page coloring consistency in the combined
1269 result->backing_object = source;
1271 * Store the offset into the source object, and fix up the offset into
1274 result->backing_object_offset = *offset;
1275 if (source != NULL) {
1276 VM_OBJECT_LOCK(source);
1277 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1278 source->shadow_count++;
1279 #if VM_NRESERVLEVEL > 0
1280 result->flags |= source->flags & OBJ_COLORED;
1281 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1282 ((1 << (VM_NFREEORDER - 1)) - 1);
1284 VM_OBJECT_UNLOCK(source);
1289 * Return the new things
1298 * Split the pages in a map entry into a new object. This affords
1299 * easier removal of unused pages, and keeps object inheritance from
1300 * being a negative impact on memory usage.
1303 vm_object_split(vm_map_entry_t entry)
1305 vm_page_t ma[VM_RADIX_STACK];
1307 vm_object_t orig_object, new_object, source;
1308 vm_pindex_t idx, offidxstart, start;
1313 orig_object = entry->object.vm_object;
1314 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1316 if (orig_object->ref_count <= 1)
1318 VM_OBJECT_UNLOCK(orig_object);
1320 offidxstart = OFF_TO_IDX(entry->offset);
1321 size = atop(entry->end - entry->start);
1324 * If swap_pager_copy() is later called, it will convert new_object
1325 * into a swap object.
1327 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1330 * At this point, the new object is still private, so the order in
1331 * which the original and new objects are locked does not matter.
1333 VM_OBJECT_LOCK(new_object);
1334 VM_OBJECT_LOCK(orig_object);
1335 source = orig_object->backing_object;
1336 if (source != NULL) {
1337 VM_OBJECT_LOCK(source);
1338 if ((source->flags & OBJ_DEAD) != 0) {
1339 VM_OBJECT_UNLOCK(source);
1340 VM_OBJECT_UNLOCK(orig_object);
1341 VM_OBJECT_UNLOCK(new_object);
1342 vm_object_deallocate(new_object);
1343 VM_OBJECT_LOCK(orig_object);
1346 LIST_INSERT_HEAD(&source->shadow_head,
1347 new_object, shadow_list);
1348 source->shadow_count++;
1349 vm_object_reference_locked(source); /* for new_object */
1350 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1351 VM_OBJECT_UNLOCK(source);
1352 new_object->backing_object_offset =
1353 orig_object->backing_object_offset + entry->offset;
1354 new_object->backing_object = source;
1356 if (orig_object->cred != NULL) {
1357 new_object->cred = orig_object->cred;
1358 crhold(orig_object->cred);
1359 new_object->charge = ptoa(size);
1360 KASSERT(orig_object->charge >= ptoa(size),
1361 ("orig_object->charge < 0"));
1362 orig_object->charge -= ptoa(size);
1364 start = offidxstart;
1367 while (exhausted == 0 && (n = vm_radix_lookupn(&orig_object->rtree,
1368 start, offidxstart + size, VM_RADIX_ANY, (void **)ma,
1369 VM_RADIX_STACK, &start, &exhausted)) != 0) {
1370 for (i = 0; i < n; i++) {
1372 idx = m->pindex - offidxstart;
1373 if (m->flags & PG_CACHED) {
1374 mtx_lock(&vm_page_queue_free_mtx);
1375 if (m->object == orig_object)
1376 vm_page_cache_rename(m, new_object,
1378 mtx_unlock(&vm_page_queue_free_mtx);
1380 } else if (m->object != orig_object)
1383 * We must wait for pending I/O to complete before
1384 * we can rename the page.
1386 * We do not have to VM_PROT_NONE the page as mappings
1387 * should not be changed by this operation.
1389 if ((m->oflags & VPO_BUSY) || m->busy) {
1391 VM_OBJECT_UNLOCK(new_object);
1392 m->oflags |= VPO_WANTED;
1393 VM_OBJECT_SLEEP(orig_object, m, PVM, "spltwt", 0);
1394 VM_OBJECT_LOCK(new_object);
1397 #if VM_NRESERVLEVEL > 0
1399 * If some of the reservation's allocated pages remain
1400 * with the original object, then transferring the
1401 * reservation to the new object is neither
1402 * particularly beneficial nor particularly harmful as
1403 * compared to leaving the reservation with the
1404 * original object. If, however, all of the
1405 * reservation's allocated pages are transferred to
1406 * the new object, then transferring the reservation
1407 * is typically beneficial. Determining which of
1408 * these two cases applies would be more costly than
1409 * unconditionally renaming the reservation.
1411 vm_reserv_rename(m, new_object, orig_object,
1414 vm_page_rename(m, new_object, idx);
1416 * page automatically made dirty by rename and
1421 if (n < VM_RADIX_STACK)
1424 if (orig_object->type == OBJT_SWAP) {
1426 * swap_pager_copy() can sleep, in which case the orig_object's
1427 * and new_object's locks are released and reacquired.
1429 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1431 VM_OBJECT_UNLOCK(orig_object);
1432 TAILQ_FOREACH(m, &new_object->memq, listq)
1434 VM_OBJECT_UNLOCK(new_object);
1435 entry->object.vm_object = new_object;
1436 entry->offset = 0LL;
1437 vm_object_deallocate(orig_object);
1438 VM_OBJECT_LOCK(new_object);
1441 #define OBSC_TEST_ALL_SHADOWED 0x0001
1442 #define OBSC_COLLAPSE_NOWAIT 0x0002
1443 #define OBSC_COLLAPSE_WAIT 0x0004
1446 vm_object_backing_scan(vm_object_t object, int op)
1448 vm_page_t pa[VM_RADIX_STACK];
1450 vm_object_t backing_object;
1451 vm_pindex_t backing_offset_index, new_pindex;
1457 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1458 VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1460 backing_object = object->backing_object;
1461 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1464 * Initial conditions
1466 if (op & OBSC_TEST_ALL_SHADOWED) {
1468 * We do not want to have to test for the existence of cache
1469 * or swap pages in the backing object. XXX but with the
1470 * new swapper this would be pretty easy to do.
1472 * XXX what about anonymous MAP_SHARED memory that hasn't
1473 * been ZFOD faulted yet? If we do not test for this, the
1474 * shadow test may succeed! XXX
1476 if (backing_object->type != OBJT_DEFAULT) {
1480 if (op & OBSC_COLLAPSE_WAIT) {
1481 vm_object_set_flag(backing_object, OBJ_DEAD);
1483 color = VM_RADIX_BLACK;
1484 if (op & OBSC_COLLAPSE_WAIT)
1485 color |= VM_RADIX_RED;
1491 i = n = VM_RADIX_STACK;
1495 if (n < VM_RADIX_STACK)
1497 if (exhausted != 0 ||
1498 (n = vm_radix_lookupn(&backing_object->rtree,
1499 start, 0, color, (void **)pa, VM_RADIX_STACK,
1500 &start, &exhausted)) == 0)
1506 * Free cached pages. XXX Why? Emulating old behavior here.
1508 if (p->flags & PG_CACHED) {
1509 mtx_lock(&vm_page_queue_free_mtx);
1510 if (p->object == backing_object)
1511 vm_page_cache_free(p);
1512 mtx_unlock(&vm_page_queue_free_mtx);
1514 } else if (p->object != backing_object)
1517 new_pindex = p->pindex - backing_offset_index;
1518 if (op & OBSC_TEST_ALL_SHADOWED) {
1522 * Ignore pages outside the parent object's range
1523 * and outside the parent object's mapping of the
1526 * note that we do not busy the backing object's
1529 if (p->pindex < backing_offset_index ||
1530 new_pindex >= object->size)
1534 * See if the parent has the page or if the parent's
1535 * object pager has the page. If the parent has the
1536 * page but the page is not valid, the parent's
1537 * object pager must have the page.
1539 * If this fails, the parent does not completely shadow
1540 * the object and we might as well give up now.
1543 pp = vm_page_lookup(object, new_pindex);
1545 (pp == NULL || pp->valid == 0) &&
1546 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1554 * Check for busy page
1556 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1559 if (op & OBSC_COLLAPSE_NOWAIT) {
1560 if ((p->oflags & VPO_BUSY) || !p->valid ||
1563 } else if (op & OBSC_COLLAPSE_WAIT) {
1564 if ((p->oflags & VPO_BUSY) || p->busy) {
1565 VM_OBJECT_UNLOCK(object);
1566 p->oflags |= VPO_WANTED;
1567 VM_OBJECT_SLEEP(backing_object, p,
1568 PDROP | PVM, "vmocol", 0);
1569 VM_OBJECT_LOCK(object);
1570 VM_OBJECT_LOCK(backing_object);
1572 * If we slept, anything could have
1573 * happened. Since the object is
1574 * marked dead, the backing offset
1575 * should not have changed so we
1576 * just restart our scan.
1583 p->object == backing_object,
1584 ("vm_object_backing_scan: object mismatch")
1588 * Destroy any associated swap
1590 if (backing_object->type == OBJT_SWAP) {
1591 swap_pager_freespace(
1599 p->pindex < backing_offset_index ||
1600 new_pindex >= object->size
1603 * Page is out of the parent object's range, we
1604 * can simply destroy it.
1607 KASSERT(!pmap_page_is_mapped(p),
1608 ("freeing mapped page %p", p));
1609 if (p->wire_count == 0)
1617 pp = vm_page_lookup(object, new_pindex);
1619 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1620 (pp != NULL && pp->valid == 0)
1623 * The page in the parent is not (yet) valid.
1624 * We don't know anything about the state of
1625 * the original page. It might be mapped,
1626 * so we must avoid the next if here.
1628 * This is due to a race in vm_fault() where
1629 * we must unbusy the original (backing_obj)
1630 * page before we can (re)lock the parent.
1631 * Hence we can get here.
1637 vm_pager_has_page(object, new_pindex, NULL, NULL)
1640 * page already exists in parent OR swap exists
1641 * for this location in the parent. Destroy
1642 * the original page from the backing object.
1644 * Leave the parent's page alone
1647 KASSERT(!pmap_page_is_mapped(p),
1648 ("freeing mapped page %p", p));
1649 if (p->wire_count == 0)
1657 #if VM_NRESERVLEVEL > 0
1659 * Rename the reservation.
1661 vm_reserv_rename(p, object, backing_object,
1662 backing_offset_index);
1666 * Page does not exist in parent, rename the
1667 * page from the backing object to the main object.
1669 * If the page was mapped to a process, it can remain
1670 * mapped through the rename.
1672 vm_page_rename(p, object, new_pindex);
1673 /* page automatically made dirty by rename */
1681 * this version of collapse allows the operation to occur earlier and
1682 * when paging_in_progress is true for an object... This is not a complete
1683 * operation, but should plug 99.9% of the rest of the leaks.
1686 vm_object_qcollapse(vm_object_t object)
1688 vm_object_t backing_object = object->backing_object;
1690 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1691 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1693 if (backing_object->ref_count != 1)
1696 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1700 * vm_object_collapse:
1702 * Collapse an object with the object backing it.
1703 * Pages in the backing object are moved into the
1704 * parent, and the backing object is deallocated.
1707 vm_object_collapse(vm_object_t object)
1709 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1712 vm_object_t backing_object;
1715 * Verify that the conditions are right for collapse:
1717 * The object exists and the backing object exists.
1719 if ((backing_object = object->backing_object) == NULL)
1723 * we check the backing object first, because it is most likely
1726 VM_OBJECT_LOCK(backing_object);
1727 if (backing_object->handle != NULL ||
1728 (backing_object->type != OBJT_DEFAULT &&
1729 backing_object->type != OBJT_SWAP) ||
1730 (backing_object->flags & OBJ_DEAD) ||
1731 object->handle != NULL ||
1732 (object->type != OBJT_DEFAULT &&
1733 object->type != OBJT_SWAP) ||
1734 (object->flags & OBJ_DEAD)) {
1735 VM_OBJECT_UNLOCK(backing_object);
1740 object->paging_in_progress != 0 ||
1741 backing_object->paging_in_progress != 0
1743 vm_object_qcollapse(object);
1744 VM_OBJECT_UNLOCK(backing_object);
1748 * We know that we can either collapse the backing object (if
1749 * the parent is the only reference to it) or (perhaps) have
1750 * the parent bypass the object if the parent happens to shadow
1751 * all the resident pages in the entire backing object.
1753 * This is ignoring pager-backed pages such as swap pages.
1754 * vm_object_backing_scan fails the shadowing test in this
1757 if (backing_object->ref_count == 1) {
1759 * If there is exactly one reference to the backing
1760 * object, we can collapse it into the parent.
1762 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1764 #if VM_NRESERVLEVEL > 0
1766 * Break any reservations from backing_object.
1768 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1769 vm_reserv_break_all(backing_object);
1773 * Move the pager from backing_object to object.
1775 if (backing_object->type == OBJT_SWAP) {
1777 * swap_pager_copy() can sleep, in which case
1778 * the backing_object's and object's locks are
1779 * released and reacquired.
1780 * Since swap_pager_copy() is being asked to
1781 * destroy the source, it will change the
1782 * backing_object's type to OBJT_DEFAULT.
1787 OFF_TO_IDX(object->backing_object_offset), TRUE);
1790 * Object now shadows whatever backing_object did.
1791 * Note that the reference to
1792 * backing_object->backing_object moves from within
1793 * backing_object to within object.
1795 LIST_REMOVE(object, shadow_list);
1796 backing_object->shadow_count--;
1797 if (backing_object->backing_object) {
1798 VM_OBJECT_LOCK(backing_object->backing_object);
1799 LIST_REMOVE(backing_object, shadow_list);
1801 &backing_object->backing_object->shadow_head,
1802 object, shadow_list);
1804 * The shadow_count has not changed.
1806 VM_OBJECT_UNLOCK(backing_object->backing_object);
1808 object->backing_object = backing_object->backing_object;
1809 object->backing_object_offset +=
1810 backing_object->backing_object_offset;
1813 * Discard backing_object.
1815 * Since the backing object has no pages, no pager left,
1816 * and no object references within it, all that is
1817 * necessary is to dispose of it.
1819 KASSERT(backing_object->ref_count == 1, (
1820 "backing_object %p was somehow re-referenced during collapse!",
1822 VM_OBJECT_UNLOCK(backing_object);
1823 vm_object_destroy(backing_object);
1827 vm_object_t new_backing_object;
1830 * If we do not entirely shadow the backing object,
1831 * there is nothing we can do so we give up.
1833 if (object->resident_page_count != object->size &&
1834 vm_object_backing_scan(object,
1835 OBSC_TEST_ALL_SHADOWED) == 0) {
1836 VM_OBJECT_UNLOCK(backing_object);
1841 * Make the parent shadow the next object in the
1842 * chain. Deallocating backing_object will not remove
1843 * it, since its reference count is at least 2.
1845 LIST_REMOVE(object, shadow_list);
1846 backing_object->shadow_count--;
1848 new_backing_object = backing_object->backing_object;
1849 if ((object->backing_object = new_backing_object) != NULL) {
1850 VM_OBJECT_LOCK(new_backing_object);
1852 &new_backing_object->shadow_head,
1856 new_backing_object->shadow_count++;
1857 vm_object_reference_locked(new_backing_object);
1858 VM_OBJECT_UNLOCK(new_backing_object);
1859 object->backing_object_offset +=
1860 backing_object->backing_object_offset;
1864 * Drop the reference count on backing_object. Since
1865 * its ref_count was at least 2, it will not vanish.
1867 backing_object->ref_count--;
1868 VM_OBJECT_UNLOCK(backing_object);
1873 * Try again with this object's new backing object.
1879 * vm_object_page_remove:
1881 * For the given object, either frees or invalidates each of the
1882 * specified pages. In general, a page is freed. However, if a page is
1883 * wired for any reason other than the existence of a managed, wired
1884 * mapping, then it may be invalidated but not removed from the object.
1885 * Pages are specified by the given range ["start", "end") and the option
1886 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1887 * extends from "start" to the end of the object. If the option
1888 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1889 * specified range are affected. If the option OBJPR_NOTMAPPED is
1890 * specified, then the pages within the specified range must have no
1891 * mappings. Otherwise, if this option is not specified, any mappings to
1892 * the specified pages are removed before the pages are freed or
1895 * In general, this operation should only be performed on objects that
1896 * contain managed pages. There are, however, two exceptions. First, it
1897 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1898 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1899 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1900 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1902 * The object must be locked.
1905 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1909 vm_page_t pa[VM_RADIX_STACK];
1915 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1916 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1917 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1918 ("vm_object_page_remove: illegal options for object %p", object));
1919 if (object->resident_page_count == 0 && object->cached_page_count == 0)
1922 vm_object_pip_add(object, 1);
1925 while (exhausted == 0 && (n = vm_radix_lookupn(&object->rtree, start,
1926 end, VM_RADIX_ANY, (void **)pa, VM_RADIX_STACK, &start,
1927 &exhausted)) != 0) {
1928 for (i = 0; i < n; i++) {
1931 * Another thread may allocate this cached page from
1932 * the queue before we acquire the page queue free
1935 if (p->flags & PG_CACHED) {
1936 mtx_lock(&vm_page_queue_free_mtx);
1937 if (p->object == object) {
1938 vm_page_cache_free(p);
1939 if (object->type == OBJT_VNODE &&
1940 object->cached_page_count == 0)
1941 vp = object->handle;
1943 mtx_unlock(&vm_page_queue_free_mtx);
1945 } else if (p->object != object)
1948 * If the page is wired for any reason besides
1949 * the existence of managed, wired mappings, then
1950 * it cannot be freed. For example, fictitious
1951 * pages, which represent device memory, are
1952 * inherently wired and cannot be freed. They can,
1953 * however, be invalidated if the option
1954 * OBJPR_CLEANONLY is not specified.
1957 if ((wirings = p->wire_count) != 0 &&
1958 (wirings = pmap_page_wired_mappings(p)) !=
1960 if ((options & OBJPR_NOTMAPPED) == 0) {
1963 * Account for removal of wired
1967 p->wire_count -= wirings;
1969 if ((options & OBJPR_CLEANONLY) == 0) {
1976 if (vm_page_sleep_if_busy(p, TRUE, "vmopar")) {
1980 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1981 ("vm_object_page_remove: page %p is fictitious",
1983 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1984 if ((options & OBJPR_NOTMAPPED) == 0)
1985 pmap_remove_write(p);
1991 if ((options & OBJPR_NOTMAPPED) == 0) {
1993 /* Account for removal of wired mappings. */
1995 KASSERT(p->wire_count == wirings,
1996 ("inconsistent wire count %d %d %p",
1997 p->wire_count, wirings, p));
1999 atomic_subtract_int(&cnt.v_wire_count,
2003 p->wire_count -= wirings;
2008 if (n < VM_RADIX_STACK)
2011 vm_object_pip_wakeup(object);
2017 * vm_object_page_cache:
2019 * For the given object, attempt to move the specified clean
2020 * pages to the cache queue. If a page is wired for any reason,
2021 * then it will not be changed. Pages are specified by the given
2022 * range ["start", "end"). As a special case, if "end" is zero,
2023 * then the range extends from "start" to the end of the object.
2024 * Any mappings to the specified pages are removed before the
2025 * pages are moved to the cache queue.
2027 * This operation should only be performed on objects that
2028 * contain non-fictitious, managed pages.
2030 * The object must be locked.
2033 vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2035 struct mtx *mtx, *new_mtx;
2038 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2039 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2040 ("vm_object_page_cache: illegal object %p", object));
2041 if (object->resident_page_count == 0)
2043 p = vm_page_find_least(object, start);
2046 * Here, the variable "p" is either (1) the page with the least pindex
2047 * greater than or equal to the parameter "start" or (2) NULL.
2050 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2051 next = TAILQ_NEXT(p, listq);
2054 * Avoid releasing and reacquiring the same page lock.
2056 new_mtx = vm_page_lockptr(p);
2057 if (mtx != new_mtx) {
2063 vm_page_try_to_cache(p);
2070 * Populate the specified range of the object with valid pages. Returns
2071 * TRUE if the range is successfully populated and FALSE otherwise.
2073 * Note: This function should be optimized to pass a larger array of
2074 * pages to vm_pager_get_pages() before it is applied to a non-
2075 * OBJT_DEVICE object.
2077 * The object must be locked.
2080 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2086 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2087 for (pindex = start; pindex < end; pindex++) {
2088 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
2090 if (m->valid != VM_PAGE_BITS_ALL) {
2092 rv = vm_pager_get_pages(object, ma, 1, 0);
2093 m = vm_page_lookup(object, pindex);
2096 if (rv != VM_PAGER_OK) {
2104 * Keep "m" busy because a subsequent iteration may unlock
2108 if (pindex > start) {
2109 m = vm_page_lookup(object, start);
2110 while (m != NULL && m->pindex < pindex) {
2112 m = TAILQ_NEXT(m, listq);
2115 return (pindex == end);
2119 * Routine: vm_object_coalesce
2120 * Function: Coalesces two objects backing up adjoining
2121 * regions of memory into a single object.
2123 * returns TRUE if objects were combined.
2125 * NOTE: Only works at the moment if the second object is NULL -
2126 * if it's not, which object do we lock first?
2129 * prev_object First object to coalesce
2130 * prev_offset Offset into prev_object
2131 * prev_size Size of reference to prev_object
2132 * next_size Size of reference to the second object
2133 * reserved Indicator that extension region has
2134 * swap accounted for
2137 * The object must *not* be locked.
2140 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2141 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2143 vm_pindex_t next_pindex;
2145 if (prev_object == NULL)
2147 VM_OBJECT_LOCK(prev_object);
2148 if (prev_object->type != OBJT_DEFAULT &&
2149 prev_object->type != OBJT_SWAP) {
2150 VM_OBJECT_UNLOCK(prev_object);
2155 * Try to collapse the object first
2157 vm_object_collapse(prev_object);
2160 * Can't coalesce if: . more than one reference . paged out . shadows
2161 * another object . has a copy elsewhere (any of which mean that the
2162 * pages not mapped to prev_entry may be in use anyway)
2164 if (prev_object->backing_object != NULL) {
2165 VM_OBJECT_UNLOCK(prev_object);
2169 prev_size >>= PAGE_SHIFT;
2170 next_size >>= PAGE_SHIFT;
2171 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2173 if ((prev_object->ref_count > 1) &&
2174 (prev_object->size != next_pindex)) {
2175 VM_OBJECT_UNLOCK(prev_object);
2180 * Account for the charge.
2182 if (prev_object->cred != NULL) {
2185 * If prev_object was charged, then this mapping,
2186 * althought not charged now, may become writable
2187 * later. Non-NULL cred in the object would prevent
2188 * swap reservation during enabling of the write
2189 * access, so reserve swap now. Failed reservation
2190 * cause allocation of the separate object for the map
2191 * entry, and swap reservation for this entry is
2192 * managed in appropriate time.
2194 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2195 prev_object->cred)) {
2198 prev_object->charge += ptoa(next_size);
2202 * Remove any pages that may still be in the object from a previous
2205 if (next_pindex < prev_object->size) {
2206 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2208 if (prev_object->type == OBJT_SWAP)
2209 swap_pager_freespace(prev_object,
2210 next_pindex, next_size);
2212 if (prev_object->cred != NULL) {
2213 KASSERT(prev_object->charge >=
2214 ptoa(prev_object->size - next_pindex),
2215 ("object %p overcharged 1 %jx %jx", prev_object,
2216 (uintmax_t)next_pindex, (uintmax_t)next_size));
2217 prev_object->charge -= ptoa(prev_object->size -
2224 * Extend the object if necessary.
2226 if (next_pindex + next_size > prev_object->size)
2227 prev_object->size = next_pindex + next_size;
2229 VM_OBJECT_UNLOCK(prev_object);
2234 vm_object_set_writeable_dirty(vm_object_t object)
2237 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2238 if (object->type != OBJT_VNODE)
2240 object->generation++;
2241 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2243 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2246 #include "opt_ddb.h"
2248 #include <sys/kernel.h>
2250 #include <sys/cons.h>
2252 #include <ddb/ddb.h>
2255 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2258 vm_map_entry_t tmpe;
2266 tmpe = map->header.next;
2267 entcount = map->nentries;
2268 while (entcount-- && (tmpe != &map->header)) {
2269 if (_vm_object_in_map(map, object, tmpe)) {
2274 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2275 tmpm = entry->object.sub_map;
2276 tmpe = tmpm->header.next;
2277 entcount = tmpm->nentries;
2278 while (entcount-- && tmpe != &tmpm->header) {
2279 if (_vm_object_in_map(tmpm, object, tmpe)) {
2284 } else if ((obj = entry->object.vm_object) != NULL) {
2285 for (; obj; obj = obj->backing_object)
2286 if (obj == object) {
2294 vm_object_in_map(vm_object_t object)
2298 /* sx_slock(&allproc_lock); */
2299 FOREACH_PROC_IN_SYSTEM(p) {
2300 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2302 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2303 /* sx_sunlock(&allproc_lock); */
2307 /* sx_sunlock(&allproc_lock); */
2308 if (_vm_object_in_map(kernel_map, object, 0))
2310 if (_vm_object_in_map(kmem_map, object, 0))
2312 if (_vm_object_in_map(pager_map, object, 0))
2314 if (_vm_object_in_map(buffer_map, object, 0))
2319 DB_SHOW_COMMAND(vmochk, vm_object_check)
2324 * make sure that internal objs are in a map somewhere
2325 * and none have zero ref counts.
2327 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2328 if (object->handle == NULL &&
2329 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2330 if (object->ref_count == 0) {
2331 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2332 (long)object->size);
2334 if (!vm_object_in_map(object)) {
2336 "vmochk: internal obj is not in a map: "
2337 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2338 object->ref_count, (u_long)object->size,
2339 (u_long)object->size,
2340 (void *)object->backing_object);
2347 * vm_object_print: [ debug ]
2349 DB_SHOW_COMMAND(object, vm_object_print_static)
2351 /* XXX convert args. */
2352 vm_object_t object = (vm_object_t)addr;
2353 boolean_t full = have_addr;
2357 /* XXX count is an (unused) arg. Avoid shadowing it. */
2358 #define count was_count
2366 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2367 object, (int)object->type, (uintmax_t)object->size,
2368 object->resident_page_count, object->ref_count, object->flags,
2369 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2370 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2371 object->shadow_count,
2372 object->backing_object ? object->backing_object->ref_count : 0,
2373 object->backing_object, (uintmax_t)object->backing_object_offset);
2380 TAILQ_FOREACH(p, &object->memq, listq) {
2382 db_iprintf("memory:=");
2383 else if (count == 6) {
2391 db_printf("(off=0x%jx,page=0x%jx,obj=%p,flags=0x%X)",
2392 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p),
2393 p->object, p->flags);
2403 /* XXX need this non-static entry for calling from vm_map_print. */
2406 /* db_expr_t */ long addr,
2407 boolean_t have_addr,
2408 /* db_expr_t */ long count,
2411 vm_object_print_static(addr, have_addr, count, modif);
2414 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2419 vm_page_t m, prev_m;
2423 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2424 db_printf("new object: %p\n", (void *)object);
2435 TAILQ_FOREACH(m, &object->memq, listq) {
2436 if (m->pindex > 128)
2438 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2439 prev_m->pindex + 1 != m->pindex) {
2441 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2442 (long)fidx, rcount, (long)pa);
2454 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2459 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2460 (long)fidx, rcount, (long)pa);
2470 pa = VM_PAGE_TO_PHYS(m);
2474 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2475 (long)fidx, rcount, (long)pa);