4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. All advertising materials mentioning features or use of this software
19 * must display the following acknowledgement:
20 * This product includes software developed by the University of
21 * California, Berkeley and its contributors.
22 * 4. Neither the name of the University nor the names of its contributors
23 * may be used to endorse or promote products derived from this software
24 * without specific prior written permission.
26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94
41 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42 * All rights reserved.
44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
46 * Permission to use, copy, modify and distribute this software and
47 * its documentation is hereby granted, provided that both the copyright
48 * notice and this permission notice appear in all copies of the
49 * software, derivative works or modified versions, and any portions
50 * thereof, and that both notices appear in supporting documentation.
52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
56 * Carnegie Mellon requests users of this software to return to
58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
59 * School of Computer Science
60 * Carnegie Mellon University
61 * Pittsburgh PA 15213-3890
63 * any improvements or extensions that they make and grant Carnegie the
64 * rights to redistribute these changes.
66 * $FreeBSD: src/sys/vm/vm_object.c,v 1.171.2.8 2003/05/26 19:17:56 alc Exp $
70 * Virtual memory object module.
73 #include <sys/param.h>
74 #include <sys/systm.h>
75 #include <sys/proc.h> /* for curproc, pageproc */
76 #include <sys/thread.h>
77 #include <sys/vnode.h>
78 #include <sys/vmmeter.h>
80 #include <sys/mount.h>
81 #include <sys/kernel.h>
82 #include <sys/sysctl.h>
83 #include <sys/refcount.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_zone.h>
98 #define EASY_SCAN_FACTOR 8
100 static void vm_object_qcollapse(vm_object_t object,
101 vm_object_t backing_object);
102 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
104 static void vm_object_lock_init(vm_object_t);
108 * Virtual memory objects maintain the actual data
109 * associated with allocated virtual memory. A given
110 * page of memory exists within exactly one object.
112 * An object is only deallocated when all "references"
113 * are given up. Only one "reference" to a given
114 * region of an object should be writeable.
116 * Associated with each object is a list of all resident
117 * memory pages belonging to that object; this list is
118 * maintained by the "vm_page" module, and locked by the object's
121 * Each object also records a "pager" routine which is
122 * used to retrieve (and store) pages to the proper backing
123 * storage. In addition, objects may be backed by other
124 * objects from which they were virtual-copied.
126 * The only items within the object structure which are
127 * modified after time of creation are:
128 * reference count locked by object's lock
129 * pager routine locked by object's lock
133 struct object_q vm_object_list; /* locked by vmobj_token */
134 struct vm_object kernel_object;
136 static long vm_object_count; /* locked by vmobj_token */
137 extern int vm_pageout_page_count;
139 static long object_collapses;
140 static long object_bypasses;
141 static int next_index;
142 static vm_zone_t obj_zone;
143 static struct vm_zone obj_zone_store;
144 #define VM_OBJECTS_INIT 256
145 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
148 * Misc low level routines
151 vm_object_lock_init(vm_object_t obj)
153 #if defined(DEBUG_LOCKS)
156 obj->debug_hold_bitmap = 0;
157 obj->debug_hold_ovfl = 0;
158 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) {
159 obj->debug_hold_thrs[i] = NULL;
160 obj->debug_hold_file[i] = NULL;
161 obj->debug_hold_line[i] = 0;
167 vm_object_lock_swap(void)
173 vm_object_lock(vm_object_t obj)
175 lwkt_gettoken(&obj->token);
179 * Returns TRUE on sucesss
182 vm_object_lock_try(vm_object_t obj)
184 return(lwkt_trytoken(&obj->token));
188 vm_object_lock_shared(vm_object_t obj)
190 lwkt_gettoken_shared(&obj->token);
194 vm_object_unlock(vm_object_t obj)
196 lwkt_reltoken(&obj->token);
200 vm_object_assert_held(vm_object_t obj)
202 ASSERT_LWKT_TOKEN_HELD(&obj->token);
207 vm_object_hold(vm_object_t obj)
209 debugvm_object_hold(vm_object_t obj, char *file, int line)
212 KKASSERT(obj != NULL);
215 * Object must be held (object allocation is stable due to callers
216 * context, typically already holding the token on a parent object)
217 * prior to potentially blocking on the lock, otherwise the object
218 * can get ripped away from us.
220 refcount_acquire(&obj->hold_count);
223 #if defined(DEBUG_LOCKS)
228 mask = ~obj->debug_hold_bitmap;
230 if (mask == 0xFFFFFFFFU) {
231 if (obj->debug_hold_ovfl == 0)
232 obj->debug_hold_ovfl = 1;
236 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask,
238 obj->debug_hold_bitmap |= (1 << i);
239 obj->debug_hold_thrs[i] = curthread;
240 obj->debug_hold_file[i] = file;
241 obj->debug_hold_line[i] = line;
250 vm_object_hold_try(vm_object_t obj)
252 debugvm_object_hold_try(vm_object_t obj, char *file, int line)
255 KKASSERT(obj != NULL);
258 * Object must be held (object allocation is stable due to callers
259 * context, typically already holding the token on a parent object)
260 * prior to potentially blocking on the lock, otherwise the object
261 * can get ripped away from us.
263 refcount_acquire(&obj->hold_count);
264 if (vm_object_lock_try(obj) == 0) {
265 if (refcount_release(&obj->hold_count)) {
266 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD))
267 zfree(obj_zone, obj);
272 #if defined(DEBUG_LOCKS)
277 mask = ~obj->debug_hold_bitmap;
279 if (mask == 0xFFFFFFFFU) {
280 if (obj->debug_hold_ovfl == 0)
281 obj->debug_hold_ovfl = 1;
285 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask,
287 obj->debug_hold_bitmap |= (1 << i);
288 obj->debug_hold_thrs[i] = curthread;
289 obj->debug_hold_file[i] = file;
290 obj->debug_hold_line[i] = line;
300 vm_object_hold_shared(vm_object_t obj)
302 debugvm_object_hold_shared(vm_object_t obj, char *file, int line)
305 KKASSERT(obj != NULL);
308 * Object must be held (object allocation is stable due to callers
309 * context, typically already holding the token on a parent object)
310 * prior to potentially blocking on the lock, otherwise the object
311 * can get ripped away from us.
313 refcount_acquire(&obj->hold_count);
314 vm_object_lock_shared(obj);
316 #if defined(DEBUG_LOCKS)
321 mask = ~obj->debug_hold_bitmap;
323 if (mask == 0xFFFFFFFFU) {
324 if (obj->debug_hold_ovfl == 0)
325 obj->debug_hold_ovfl = 1;
329 if (atomic_cmpset_int(&obj->debug_hold_bitmap, ~mask,
331 obj->debug_hold_bitmap |= (1 << i);
332 obj->debug_hold_thrs[i] = curthread;
333 obj->debug_hold_file[i] = file;
334 obj->debug_hold_line[i] = line;
342 * Drop the token and hold_count on the object.
345 vm_object_drop(vm_object_t obj)
350 #if defined(DEBUG_LOCKS)
354 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) {
355 if ((obj->debug_hold_bitmap & (1 << i)) &&
356 (obj->debug_hold_thrs[i] == curthread)) {
357 obj->debug_hold_bitmap &= ~(1 << i);
358 obj->debug_hold_thrs[i] = NULL;
359 obj->debug_hold_file[i] = NULL;
360 obj->debug_hold_line[i] = 0;
366 if (found == 0 && obj->debug_hold_ovfl == 0)
367 panic("vm_object: attempt to drop hold on non-self-held obj");
371 * No new holders should be possible once we drop hold_count 1->0 as
372 * there is no longer any way to reference the object.
374 KKASSERT(obj->hold_count > 0);
375 if (refcount_release(&obj->hold_count)) {
376 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD)) {
377 vm_object_unlock(obj);
378 zfree(obj_zone, obj);
380 vm_object_unlock(obj);
383 vm_object_unlock(obj);
388 * Initialize a freshly allocated object, returning a held object.
390 * Used only by vm_object_allocate() and zinitna().
395 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
399 RB_INIT(&object->rb_memq);
400 LIST_INIT(&object->shadow_head);
401 lwkt_token_init(&object->token, "vmobj");
405 object->ref_count = 1;
406 object->hold_count = 0;
408 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
409 vm_object_set_flag(object, OBJ_ONEMAPPING);
410 object->paging_in_progress = 0;
411 object->resident_page_count = 0;
412 object->agg_pv_list_count = 0;
413 object->shadow_count = 0;
415 /* cpu localization twist */
416 object->pg_color = (int)(intptr_t)curthread;
418 object->pg_color = next_index;
420 if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
421 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
424 next_index = (next_index + incr) & PQ_L2_MASK;
425 object->handle = NULL;
426 object->backing_object = NULL;
427 object->backing_object_offset = (vm_ooffset_t)0;
429 object->generation++;
430 object->swblock_count = 0;
431 RB_INIT(&object->swblock_root);
432 vm_object_lock_init(object);
434 vm_object_hold(object);
435 lwkt_gettoken(&vmobj_token);
436 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
438 lwkt_reltoken(&vmobj_token);
442 * Initialize the VM objects module.
444 * Called from the low level boot code only.
449 TAILQ_INIT(&vm_object_list);
451 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd),
453 vm_object_drop(&kernel_object);
455 obj_zone = &obj_zone_store;
456 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
457 vm_objects_init, VM_OBJECTS_INIT);
461 vm_object_init2(void)
463 zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1);
467 * Allocate and return a new object of the specified type and size.
472 vm_object_allocate(objtype_t type, vm_pindex_t size)
476 result = (vm_object_t) zalloc(obj_zone);
478 _vm_object_allocate(type, size, result);
479 vm_object_drop(result);
485 * This version returns a held object, allowing further atomic initialization
489 vm_object_allocate_hold(objtype_t type, vm_pindex_t size)
493 result = (vm_object_t) zalloc(obj_zone);
495 _vm_object_allocate(type, size, result);
501 * Add an additional reference to a vm_object. The object must already be
502 * held. The original non-lock version is no longer supported. The object
503 * must NOT be chain locked by anyone at the time the reference is added.
505 * Referencing a chain-locked object can blow up the fairly sensitive
506 * ref_count and shadow_count tests in the deallocator. Most callers
507 * will call vm_object_chain_wait() prior to calling
508 * vm_object_reference_locked() to avoid the case.
510 * The object must be held.
513 vm_object_reference_locked(vm_object_t object)
515 KKASSERT(object != NULL);
516 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
517 KKASSERT((object->flags & OBJ_CHAINLOCK) == 0);
519 if (object->type == OBJT_VNODE) {
520 vref(object->handle);
521 /* XXX what if the vnode is being destroyed? */
526 * Object OBJ_CHAINLOCK lock handling.
528 * The caller can chain-lock backing objects recursively and then
529 * use vm_object_chain_release_all() to undo the whole chain.
531 * Chain locks are used to prevent collapses and are only applicable
532 * to OBJT_DEFAULT and OBJT_SWAP objects. Chain locking operations
533 * on other object types are ignored. This is also important because
534 * it allows e.g. the vnode underlying a memory mapping to take concurrent
537 * The object must usually be held on entry, though intermediate
538 * objects need not be held on release.
541 vm_object_chain_wait(vm_object_t object)
543 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
544 while (object->flags & OBJ_CHAINLOCK) {
545 vm_object_set_flag(object, OBJ_CHAINWANT);
546 tsleep(object, 0, "objchain", 0);
551 vm_object_chain_acquire(vm_object_t object)
553 if (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) {
554 vm_object_chain_wait(object);
555 vm_object_set_flag(object, OBJ_CHAINLOCK);
560 vm_object_chain_release(vm_object_t object)
562 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
563 if (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) {
564 KKASSERT(object->flags & OBJ_CHAINLOCK);
565 if (object->flags & OBJ_CHAINWANT) {
566 vm_object_clear_flag(object,
567 OBJ_CHAINLOCK | OBJ_CHAINWANT);
570 vm_object_clear_flag(object, OBJ_CHAINLOCK);
576 * This releases the entire chain of objects from first_object to and
577 * including stopobj, flowing through object->backing_object.
579 * We release stopobj first as an optimization as this object is most
580 * likely to be shared across multiple processes.
583 vm_object_chain_release_all(vm_object_t first_object, vm_object_t stopobj)
585 vm_object_t backing_object;
588 vm_object_chain_release(stopobj);
589 object = first_object;
591 while (object != stopobj) {
593 if (object != first_object)
594 vm_object_hold(object);
595 backing_object = object->backing_object;
596 vm_object_chain_release(object);
597 if (object != first_object)
598 vm_object_drop(object);
599 object = backing_object;
604 * Dereference an object and its underlying vnode.
606 * The object must be held and will be held on return.
609 vm_object_vndeallocate(vm_object_t object)
611 struct vnode *vp = (struct vnode *) object->handle;
613 KASSERT(object->type == OBJT_VNODE,
614 ("vm_object_vndeallocate: not a vnode object"));
615 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
616 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
618 if (object->ref_count == 0) {
619 vprint("vm_object_vndeallocate", vp);
620 panic("vm_object_vndeallocate: bad object reference count");
624 if (object->ref_count == 0)
625 vclrflags(vp, VTEXT);
630 * Release a reference to the specified object, gained either through a
631 * vm_object_allocate or a vm_object_reference call. When all references
632 * are gone, storage associated with this object may be relinquished.
634 * The caller does not have to hold the object locked but must have control
635 * over the reference in question in order to guarantee that the object
636 * does not get ripped out from under us.
639 vm_object_deallocate(vm_object_t object)
642 vm_object_hold(object);
643 vm_object_deallocate_locked(object);
644 vm_object_drop(object);
649 vm_object_deallocate_locked(vm_object_t object)
651 struct vm_object_dealloc_list *dlist = NULL;
652 struct vm_object_dealloc_list *dtmp;
657 * We may chain deallocate object, but additional objects may
658 * collect on the dlist which also have to be deallocated. We
659 * must avoid a recursion, vm_object chains can get deep.
662 while (object != NULL) {
665 * Don't rip a ref_count out from under an object undergoing
666 * collapse, it will confuse the collapse code.
668 vm_object_chain_wait(object);
670 if (object->type == OBJT_VNODE) {
671 vm_object_vndeallocate(object);
675 if (object->ref_count == 0) {
676 panic("vm_object_deallocate: object deallocated "
677 "too many times: %d", object->type);
679 if (object->ref_count > 2) {
685 * Here on ref_count of one or two, which are special cases for
688 * Nominal ref_count > 1 case if the second ref is not from
691 if (object->ref_count == 2 && object->shadow_count == 0) {
692 vm_object_set_flag(object, OBJ_ONEMAPPING);
698 * If the second ref is from a shadow we chain along it
699 * upwards if object's handle is exhausted.
701 * We have to decrement object->ref_count before potentially
702 * collapsing the first shadow object or the collapse code
703 * will not be able to handle the degenerate case to remove
704 * object. However, if we do it too early the object can
705 * get ripped out from under us.
707 if (object->ref_count == 2 && object->shadow_count == 1 &&
708 object->handle == NULL && (object->type == OBJT_DEFAULT ||
709 object->type == OBJT_SWAP)) {
710 temp = LIST_FIRST(&object->shadow_head);
711 KKASSERT(temp != NULL);
712 vm_object_hold(temp);
715 * Wait for any paging to complete so the collapse
716 * doesn't (or isn't likely to) qcollapse. pip
717 * waiting must occur before we acquire the
721 temp->paging_in_progress ||
722 object->paging_in_progress
724 vm_object_pip_wait(temp, "objde1");
725 vm_object_pip_wait(object, "objde2");
729 * If the parent is locked we have to give up, as
730 * otherwise we would be acquiring locks in the
731 * wrong order and potentially deadlock.
733 if (temp->flags & OBJ_CHAINLOCK) {
734 vm_object_drop(temp);
737 vm_object_chain_acquire(temp);
740 * Recheck/retry after the hold and the paging
741 * wait, both of which can block us.
743 if (object->ref_count != 2 ||
744 object->shadow_count != 1 ||
746 LIST_FIRST(&object->shadow_head) != temp ||
747 (object->type != OBJT_DEFAULT &&
748 object->type != OBJT_SWAP)) {
749 vm_object_chain_release(temp);
750 vm_object_drop(temp);
755 * We can safely drop object's ref_count now.
757 KKASSERT(object->ref_count == 2);
761 * If our single parent is not collapseable just
762 * decrement ref_count (2->1) and stop.
764 if (temp->handle || (temp->type != OBJT_DEFAULT &&
765 temp->type != OBJT_SWAP)) {
766 vm_object_chain_release(temp);
767 vm_object_drop(temp);
772 * At this point we have already dropped object's
773 * ref_count so it is possible for a race to
774 * deallocate obj out from under us. Any collapse
775 * will re-check the situation. We must not block
776 * until we are able to collapse.
778 * Bump temp's ref_count to avoid an unwanted
779 * degenerate recursion (can't call
780 * vm_object_reference_locked() because it asserts
781 * that CHAINLOCK is not set).
784 KKASSERT(temp->ref_count > 1);
787 * Collapse temp, then deallocate the extra ref
790 vm_object_collapse(temp, &dlist);
791 vm_object_chain_release(temp);
793 vm_object_lock_swap();
794 vm_object_drop(object);
802 * Drop the ref and handle termination on the 1->0 transition.
803 * We may have blocked above so we have to recheck.
806 KKASSERT(object->ref_count != 0);
807 if (object->ref_count >= 2) {
811 KKASSERT(object->ref_count == 1);
814 * 1->0 transition. Chain through the backing_object.
815 * Maintain the ref until we've located the backing object,
818 while ((temp = object->backing_object) != NULL) {
819 vm_object_hold(temp);
820 if (temp == object->backing_object)
822 vm_object_drop(temp);
826 * 1->0 transition verified, retry if ref_count is no longer
827 * 1. Otherwise disconnect the backing_object (temp) and
830 if (object->ref_count != 1) {
831 vm_object_drop(temp);
836 * It shouldn't be possible for the object to be chain locked
837 * if we're removing the last ref on it.
839 KKASSERT((object->flags & OBJ_CHAINLOCK) == 0);
842 LIST_REMOVE(object, shadow_list);
843 temp->shadow_count--;
845 object->backing_object = NULL;
849 if ((object->flags & OBJ_DEAD) == 0)
850 vm_object_terminate(object);
851 if (must_drop && temp)
852 vm_object_lock_swap();
854 vm_object_drop(object);
858 if (must_drop && object)
859 vm_object_drop(object);
862 * Additional tail recursion on dlist. Avoid a recursion. Objects
863 * on the dlist have a hold count but are not locked.
865 if ((dtmp = dlist) != NULL) {
867 object = dtmp->object;
870 vm_object_lock(object); /* already held, add lock */
871 must_drop = 1; /* and we're responsible for it */
877 * Destroy the specified object, freeing up related resources.
879 * The object must have zero references.
881 * The object must held. The caller is responsible for dropping the object
882 * after terminate returns. Terminate does NOT drop the object.
884 static int vm_object_terminate_callback(vm_page_t p, void *data);
887 vm_object_terminate(vm_object_t object)
890 * Make sure no one uses us. Once we set OBJ_DEAD we should be
891 * able to safely block.
893 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
894 KKASSERT((object->flags & OBJ_DEAD) == 0);
895 vm_object_set_flag(object, OBJ_DEAD);
898 * Wait for the pageout daemon to be done with the object
900 vm_object_pip_wait(object, "objtrm1");
902 KASSERT(!object->paging_in_progress,
903 ("vm_object_terminate: pageout in progress"));
906 * Clean and free the pages, as appropriate. All references to the
907 * object are gone, so we don't need to lock it.
909 if (object->type == OBJT_VNODE) {
913 * Clean pages and flush buffers.
915 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
917 vp = (struct vnode *) object->handle;
918 vinvalbuf(vp, V_SAVE, 0, 0);
922 * Wait for any I/O to complete, after which there had better not
923 * be any references left on the object.
925 vm_object_pip_wait(object, "objtrm2");
927 if (object->ref_count != 0) {
928 panic("vm_object_terminate: object with references, "
929 "ref_count=%d", object->ref_count);
933 * Now free any remaining pages. For internal objects, this also
934 * removes them from paging queues. Don't free wired pages, just
935 * remove them from the object.
937 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
938 vm_object_terminate_callback, NULL);
941 * Let the pager know object is dead.
943 vm_pager_deallocate(object);
946 * Wait for the object hold count to hit 1, clean out pages as
947 * we go. vmobj_token interlocks any race conditions that might
948 * pick the object up from the vm_object_list after we have cleared
952 if (RB_ROOT(&object->rb_memq) == NULL)
954 kprintf("vm_object_terminate: Warning, object %p "
955 "still has %d pages\n",
956 object, object->resident_page_count);
957 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
958 vm_object_terminate_callback, NULL);
962 * There had better not be any pages left
964 KKASSERT(object->resident_page_count == 0);
967 * Remove the object from the global object list.
969 lwkt_gettoken(&vmobj_token);
970 TAILQ_REMOVE(&vm_object_list, object, object_list);
972 lwkt_reltoken(&vmobj_token);
973 vm_object_dead_wakeup(object);
975 if (object->ref_count != 0) {
976 panic("vm_object_terminate2: object with references, "
977 "ref_count=%d", object->ref_count);
981 * NOTE: The object hold_count is at least 1, so we cannot zfree()
982 * the object here. See vm_object_drop().
987 * The caller must hold the object.
990 vm_object_terminate_callback(vm_page_t p, void *data __unused)
995 vm_page_busy_wait(p, TRUE, "vmpgtrm");
996 if (object != p->object) {
997 kprintf("vm_object_terminate: Warning: Encountered "
998 "busied page %p on queue %d\n", p, p->queue);
1000 } else if (p->wire_count == 0) {
1002 mycpu->gd_cnt.v_pfree++;
1004 if (p->queue != PQ_NONE)
1005 kprintf("vm_object_terminate: Warning: Encountered "
1006 "wired page %p on queue %d\n", p, p->queue);
1015 * The object is dead but still has an object<->pager association. Sleep
1016 * and return. The caller typically retests the association in a loop.
1018 * The caller must hold the object.
1021 vm_object_dead_sleep(vm_object_t object, const char *wmesg)
1023 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1024 if (object->handle) {
1025 vm_object_set_flag(object, OBJ_DEADWNT);
1026 tsleep(object, 0, wmesg, 0);
1027 /* object may be invalid after this point */
1032 * Wakeup anyone waiting for the object<->pager disassociation on
1035 * The caller must hold the object.
1038 vm_object_dead_wakeup(vm_object_t object)
1040 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1041 if (object->flags & OBJ_DEADWNT) {
1042 vm_object_clear_flag(object, OBJ_DEADWNT);
1048 * Clean all dirty pages in the specified range of object. Leaves page
1049 * on whatever queue it is currently on. If NOSYNC is set then do not
1050 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
1051 * leaving the object dirty.
1053 * When stuffing pages asynchronously, allow clustering. XXX we need a
1054 * synchronous clustering mode implementation.
1056 * Odd semantics: if start == end, we clean everything.
1058 * The object must be locked? XXX
1060 static int vm_object_page_clean_pass1(struct vm_page *p, void *data);
1061 static int vm_object_page_clean_pass2(struct vm_page *p, void *data);
1064 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1067 struct rb_vm_page_scan_info info;
1073 vm_object_hold(object);
1074 if (object->type != OBJT_VNODE ||
1075 (object->flags & OBJ_MIGHTBEDIRTY) == 0) {
1076 vm_object_drop(object);
1080 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ?
1081 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1082 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
1084 vp = object->handle;
1087 * Interlock other major object operations. This allows us to
1088 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY.
1090 vm_object_set_flag(object, OBJ_CLEANING);
1093 * Handle 'entire object' case
1095 info.start_pindex = start;
1097 info.end_pindex = object->size - 1;
1099 info.end_pindex = end - 1;
1101 wholescan = (start == 0 && info.end_pindex == object->size - 1);
1103 info.pagerflags = pagerflags;
1104 info.object = object;
1107 * If cleaning the entire object do a pass to mark the pages read-only.
1108 * If everything worked out ok, clear OBJ_WRITEABLE and
1113 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1114 vm_object_page_clean_pass1, &info);
1115 if (info.error == 0) {
1116 vm_object_clear_flag(object,
1117 OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1118 if (object->type == OBJT_VNODE &&
1119 (vp = (struct vnode *)object->handle) != NULL) {
1120 if (vp->v_flag & VOBJDIRTY)
1121 vclrflags(vp, VOBJDIRTY);
1127 * Do a pass to clean all the dirty pages we find.
1131 generation = object->generation;
1132 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1133 vm_object_page_clean_pass2, &info);
1134 } while (info.error || generation != object->generation);
1136 vm_object_clear_flag(object, OBJ_CLEANING);
1137 vm_object_drop(object);
1141 * The caller must hold the object.
1145 vm_object_page_clean_pass1(struct vm_page *p, void *data)
1147 struct rb_vm_page_scan_info *info = data;
1149 vm_page_flag_set(p, PG_CLEANCHK);
1150 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1152 } else if (vm_page_busy_try(p, FALSE) == 0) {
1153 vm_page_protect(p, VM_PROT_READ); /* must not block */
1163 * The caller must hold the object
1167 vm_object_page_clean_pass2(struct vm_page *p, void *data)
1169 struct rb_vm_page_scan_info *info = data;
1173 * Do not mess with pages that were inserted after we started
1174 * the cleaning pass.
1176 if ((p->flags & PG_CLEANCHK) == 0)
1179 generation = info->object->generation;
1180 vm_page_busy_wait(p, TRUE, "vpcwai");
1181 if (p->object != info->object ||
1182 info->object->generation != generation) {
1189 * Before wasting time traversing the pmaps, check for trivial
1190 * cases where the page cannot be dirty.
1192 if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) {
1193 KKASSERT((p->dirty & p->valid) == 0);
1199 * Check whether the page is dirty or not. The page has been set
1200 * to be read-only so the check will not race a user dirtying the
1203 vm_page_test_dirty(p);
1204 if ((p->dirty & p->valid) == 0) {
1205 vm_page_flag_clear(p, PG_CLEANCHK);
1211 * If we have been asked to skip nosync pages and this is a
1212 * nosync page, skip it. Note that the object flags were
1213 * not cleared in this case (because pass1 will have returned an
1214 * error), so we do not have to set them.
1216 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1217 vm_page_flag_clear(p, PG_CLEANCHK);
1223 * Flush as many pages as we can. PG_CLEANCHK will be cleared on
1224 * the pages that get successfully flushed. Set info->error if
1225 * we raced an object modification.
1227 vm_object_page_collect_flush(info->object, p, info->pagerflags);
1234 * Collect the specified page and nearby pages and flush them out.
1235 * The number of pages flushed is returned. The passed page is busied
1236 * by the caller and we are responsible for its disposition.
1238 * The caller must hold the object.
1241 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags)
1250 vm_page_t maf[vm_pageout_page_count];
1251 vm_page_t mab[vm_pageout_page_count];
1252 vm_page_t ma[vm_pageout_page_count];
1254 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1259 for(i = 1; i < vm_pageout_page_count; i++) {
1262 tp = vm_page_lookup_busy_try(object, pi + i, TRUE, &error);
1267 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1268 (tp->flags & PG_CLEANCHK) == 0) {
1272 if ((tp->queue - tp->pc) == PQ_CACHE) {
1273 vm_page_flag_clear(tp, PG_CLEANCHK);
1277 vm_page_test_dirty(tp);
1278 if ((tp->dirty & tp->valid) == 0) {
1279 vm_page_flag_clear(tp, PG_CLEANCHK);
1288 chkb = vm_pageout_page_count - maxf;
1290 * NOTE: chkb can be 0
1292 for(i = 1; chkb && i < chkb; i++) {
1295 tp = vm_page_lookup_busy_try(object, pi - i, TRUE, &error);
1300 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1301 (tp->flags & PG_CLEANCHK) == 0) {
1305 if ((tp->queue - tp->pc) == PQ_CACHE) {
1306 vm_page_flag_clear(tp, PG_CLEANCHK);
1310 vm_page_test_dirty(tp);
1311 if ((tp->dirty & tp->valid) == 0) {
1312 vm_page_flag_clear(tp, PG_CLEANCHK);
1321 * All pages in the maf[] and mab[] array are busied.
1323 for (i = 0; i < maxb; i++) {
1324 int index = (maxb - i) - 1;
1326 vm_page_flag_clear(ma[index], PG_CLEANCHK);
1328 vm_page_flag_clear(p, PG_CLEANCHK);
1330 for(i = 0; i < maxf; i++) {
1331 int index = (maxb + i) + 1;
1333 vm_page_flag_clear(ma[index], PG_CLEANCHK);
1335 runlen = maxb + maxf + 1;
1337 for (i = 0; i < runlen; i++)
1338 vm_page_hold(ma[i]);
1340 vm_pageout_flush(ma, runlen, pagerflags);
1342 for (i = 0; i < runlen; i++) {
1343 if (ma[i]->valid & ma[i]->dirty) {
1344 vm_page_protect(ma[i], VM_PROT_READ);
1345 vm_page_flag_set(ma[i], PG_CLEANCHK);
1348 * maxf will end up being the actual number of pages
1349 * we wrote out contiguously, non-inclusive of the
1350 * first page. We do not count look-behind pages.
1352 if (i >= maxb + 1 && (maxf > i - maxb - 1))
1353 maxf = i - maxb - 1;
1355 vm_page_unhold(ma[i]);
1361 * Same as vm_object_pmap_copy, except range checking really
1362 * works, and is meant for small sections of an object.
1364 * This code protects resident pages by making them read-only
1365 * and is typically called on a fork or split when a page
1366 * is converted to copy-on-write.
1368 * NOTE: If the page is already at VM_PROT_NONE, calling
1369 * vm_page_protect will have no effect.
1372 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1377 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
1380 vm_object_hold(object);
1381 for (idx = start; idx < end; idx++) {
1382 p = vm_page_lookup(object, idx);
1385 vm_page_protect(p, VM_PROT_READ);
1387 vm_object_drop(object);
1391 * Removes all physical pages in the specified object range from all
1394 * The object must *not* be locked.
1397 static int vm_object_pmap_remove_callback(vm_page_t p, void *data);
1400 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1402 struct rb_vm_page_scan_info info;
1406 info.start_pindex = start;
1407 info.end_pindex = end - 1;
1409 vm_object_hold(object);
1410 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1411 vm_object_pmap_remove_callback, &info);
1412 if (start == 0 && end == object->size)
1413 vm_object_clear_flag(object, OBJ_WRITEABLE);
1414 vm_object_drop(object);
1418 * The caller must hold the object
1421 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused)
1423 vm_page_protect(p, VM_PROT_NONE);
1428 * Implements the madvise function at the object/page level.
1430 * MADV_WILLNEED (any object)
1432 * Activate the specified pages if they are resident.
1434 * MADV_DONTNEED (any object)
1436 * Deactivate the specified pages if they are resident.
1438 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only)
1440 * Deactivate and clean the specified pages if they are
1441 * resident. This permits the process to reuse the pages
1442 * without faulting or the kernel to reclaim the pages
1448 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1450 vm_pindex_t end, tpindex;
1451 vm_object_t tobject;
1459 end = pindex + count;
1461 vm_object_hold(object);
1465 * Locate and adjust resident pages
1467 for (; pindex < end; pindex += 1) {
1469 if (tobject != object)
1470 vm_object_drop(tobject);
1475 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1476 * and those pages must be OBJ_ONEMAPPING.
1478 if (advise == MADV_FREE) {
1479 if ((tobject->type != OBJT_DEFAULT &&
1480 tobject->type != OBJT_SWAP) ||
1481 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1486 m = vm_page_lookup_busy_try(tobject, tpindex, TRUE, &error);
1489 vm_page_sleep_busy(m, TRUE, "madvpo");
1494 * There may be swap even if there is no backing page
1496 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1497 swap_pager_freespace(tobject, tpindex, 1);
1502 while ((xobj = tobject->backing_object) != NULL) {
1503 KKASSERT(xobj != object);
1504 vm_object_hold(xobj);
1505 if (xobj == tobject->backing_object)
1507 vm_object_drop(xobj);
1511 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1512 if (tobject != object) {
1513 vm_object_lock_swap();
1514 vm_object_drop(tobject);
1521 * If the page is not in a normal active state, we skip it.
1522 * If the page is not managed there are no page queues to
1523 * mess with. Things can break if we mess with pages in
1524 * any of the below states.
1527 /*m->hold_count ||*/
1529 (m->flags & PG_UNMANAGED) ||
1530 m->valid != VM_PAGE_BITS_ALL
1537 * Theoretically once a page is known not to be busy, an
1538 * interrupt cannot come along and rip it out from under us.
1541 if (advise == MADV_WILLNEED) {
1542 vm_page_activate(m);
1543 } else if (advise == MADV_DONTNEED) {
1544 vm_page_dontneed(m);
1545 } else if (advise == MADV_FREE) {
1547 * Mark the page clean. This will allow the page
1548 * to be freed up by the system. However, such pages
1549 * are often reused quickly by malloc()/free()
1550 * so we do not do anything that would cause
1551 * a page fault if we can help it.
1553 * Specifically, we do not try to actually free
1554 * the page now nor do we try to put it in the
1555 * cache (which would cause a page fault on reuse).
1557 * But we do make the page is freeable as we
1558 * can without actually taking the step of unmapping
1561 pmap_clear_modify(m);
1564 vm_page_dontneed(m);
1565 if (tobject->type == OBJT_SWAP)
1566 swap_pager_freespace(tobject, tpindex, 1);
1570 if (tobject != object)
1571 vm_object_drop(tobject);
1572 vm_object_drop(object);
1576 * Create a new object which is backed by the specified existing object
1577 * range. Replace the pointer and offset that was pointing at the existing
1578 * object with the pointer/offset for the new object.
1580 * No other requirements.
1583 vm_object_shadow(vm_object_t *objectp, vm_ooffset_t *offset, vm_size_t length,
1592 * Don't create the new object if the old object isn't shared.
1593 * We have to chain wait before adding the reference to avoid
1594 * racing a collapse or deallocation.
1596 * Add the additional ref to source here to avoid racing a later
1597 * collapse or deallocation. Clear the ONEMAPPING flag whether
1598 * addref is TRUE or not in this case because the original object
1602 vm_object_hold(source);
1603 vm_object_chain_wait(source);
1604 if (source->ref_count == 1 &&
1605 source->handle == NULL &&
1606 (source->type == OBJT_DEFAULT ||
1607 source->type == OBJT_SWAP)) {
1608 vm_object_drop(source);
1610 vm_object_reference_locked(source);
1611 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1615 vm_object_reference_locked(source);
1616 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1620 * Allocate a new object with the given length. The new object
1621 * is returned referenced but we may have to add another one.
1622 * If we are adding a second reference we must clear OBJ_ONEMAPPING.
1623 * (typically because the caller is about to clone a vm_map_entry).
1625 * The source object currently has an extra reference to prevent
1626 * collapses into it while we mess with its shadow list, which
1627 * we will remove later in this routine.
1629 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1630 panic("vm_object_shadow: no object for shadowing");
1631 vm_object_hold(result);
1633 vm_object_reference_locked(result);
1634 vm_object_clear_flag(result, OBJ_ONEMAPPING);
1638 * The new object shadows the source object. Chain wait before
1639 * adjusting shadow_count or the shadow list to avoid races.
1641 * Try to optimize the result object's page color when shadowing
1642 * in order to maintain page coloring consistency in the combined
1645 KKASSERT(result->backing_object == NULL);
1646 result->backing_object = source;
1648 vm_object_chain_wait(source);
1649 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1650 source->shadow_count++;
1651 source->generation++;
1653 /* cpu localization twist */
1654 result->pg_color = (int)(intptr_t)curthread;
1656 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1662 * Adjust the return storage. Drop the ref on source before
1665 result->backing_object_offset = *offset;
1666 vm_object_drop(result);
1669 vm_object_deallocate_locked(source);
1670 vm_object_drop(source);
1674 * Return the new things
1679 #define OBSC_TEST_ALL_SHADOWED 0x0001
1680 #define OBSC_COLLAPSE_NOWAIT 0x0002
1681 #define OBSC_COLLAPSE_WAIT 0x0004
1683 static int vm_object_backing_scan_callback(vm_page_t p, void *data);
1686 * The caller must hold the object.
1689 vm_object_backing_scan(vm_object_t object, vm_object_t backing_object, int op)
1691 struct rb_vm_page_scan_info info;
1693 vm_object_assert_held(object);
1694 vm_object_assert_held(backing_object);
1696 KKASSERT(backing_object == object->backing_object);
1697 info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1700 * Initial conditions
1702 if (op & OBSC_TEST_ALL_SHADOWED) {
1704 * We do not want to have to test for the existence of
1705 * swap pages in the backing object. XXX but with the
1706 * new swapper this would be pretty easy to do.
1708 * XXX what about anonymous MAP_SHARED memory that hasn't
1709 * been ZFOD faulted yet? If we do not test for this, the
1710 * shadow test may succeed! XXX
1712 if (backing_object->type != OBJT_DEFAULT)
1715 if (op & OBSC_COLLAPSE_WAIT) {
1716 KKASSERT((backing_object->flags & OBJ_DEAD) == 0);
1717 vm_object_set_flag(backing_object, OBJ_DEAD);
1718 lwkt_gettoken(&vmobj_token);
1719 TAILQ_REMOVE(&vm_object_list, backing_object, object_list);
1721 lwkt_reltoken(&vmobj_token);
1722 vm_object_dead_wakeup(backing_object);
1726 * Our scan. We have to retry if a negative error code is returned,
1727 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that
1728 * the scan had to be stopped because the parent does not completely
1731 info.object = object;
1732 info.backing_object = backing_object;
1736 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL,
1737 vm_object_backing_scan_callback,
1739 } while (info.error < 0);
1745 * The caller must hold the object.
1748 vm_object_backing_scan_callback(vm_page_t p, void *data)
1750 struct rb_vm_page_scan_info *info = data;
1751 vm_object_t backing_object;
1754 vm_pindex_t new_pindex;
1755 vm_pindex_t backing_offset_index;
1759 new_pindex = pindex - info->backing_offset_index;
1761 object = info->object;
1762 backing_object = info->backing_object;
1763 backing_offset_index = info->backing_offset_index;
1765 if (op & OBSC_TEST_ALL_SHADOWED) {
1769 * Ignore pages outside the parent object's range
1770 * and outside the parent object's mapping of the
1773 * note that we do not busy the backing object's
1776 if (pindex < backing_offset_index ||
1777 new_pindex >= object->size
1783 * See if the parent has the page or if the parent's
1784 * object pager has the page. If the parent has the
1785 * page but the page is not valid, the parent's
1786 * object pager must have the page.
1788 * If this fails, the parent does not completely shadow
1789 * the object and we might as well give up now.
1791 pp = vm_page_lookup(object, new_pindex);
1792 if ((pp == NULL || pp->valid == 0) &&
1793 !vm_pager_has_page(object, new_pindex)
1795 info->error = 0; /* problemo */
1796 return(-1); /* stop the scan */
1801 * Check for busy page. Note that we may have lost (p) when we
1802 * possibly blocked above.
1804 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1807 if (vm_page_busy_try(p, TRUE)) {
1808 if (op & OBSC_COLLAPSE_NOWAIT) {
1812 * If we slept, anything could have
1813 * happened. Ask that the scan be restarted.
1815 * Since the object is marked dead, the
1816 * backing offset should not have changed.
1818 vm_page_sleep_busy(p, TRUE, "vmocol");
1825 * If (p) is no longer valid restart the scan.
1827 if (p->object != backing_object || p->pindex != pindex) {
1828 kprintf("vm_object_backing_scan: Warning: page "
1829 "%p ripped out from under us\n", p);
1835 if (op & OBSC_COLLAPSE_NOWAIT) {
1836 if (p->valid == 0 /*|| p->hold_count*/ ||
1842 /* XXX what if p->valid == 0 , hold_count, etc? */
1846 p->object == backing_object,
1847 ("vm_object_qcollapse(): object mismatch")
1851 * Destroy any associated swap
1853 if (backing_object->type == OBJT_SWAP)
1854 swap_pager_freespace(backing_object, p->pindex, 1);
1857 p->pindex < backing_offset_index ||
1858 new_pindex >= object->size
1861 * Page is out of the parent object's range, we
1862 * can simply destroy it.
1864 vm_page_protect(p, VM_PROT_NONE);
1869 pp = vm_page_lookup(object, new_pindex);
1870 if (pp != NULL || vm_pager_has_page(object, new_pindex)) {
1872 * page already exists in parent OR swap exists
1873 * for this location in the parent. Destroy
1874 * the original page from the backing object.
1876 * Leave the parent's page alone
1878 vm_page_protect(p, VM_PROT_NONE);
1884 * Page does not exist in parent, rename the
1885 * page from the backing object to the main object.
1887 * If the page was mapped to a process, it can remain
1888 * mapped through the rename.
1890 if ((p->queue - p->pc) == PQ_CACHE)
1891 vm_page_deactivate(p);
1893 vm_page_rename(p, object, new_pindex);
1895 /* page automatically made dirty by rename */
1901 * This version of collapse allows the operation to occur earlier and
1902 * when paging_in_progress is true for an object... This is not a complete
1903 * operation, but should plug 99.9% of the rest of the leaks.
1905 * The caller must hold the object and backing_object and both must be
1908 * (only called from vm_object_collapse)
1911 vm_object_qcollapse(vm_object_t object, vm_object_t backing_object)
1913 if (backing_object->ref_count == 1) {
1914 backing_object->ref_count += 2;
1915 vm_object_backing_scan(object, backing_object,
1916 OBSC_COLLAPSE_NOWAIT);
1917 backing_object->ref_count -= 2;
1922 * Collapse an object with the object backing it. Pages in the backing
1923 * object are moved into the parent, and the backing object is deallocated.
1924 * Any conflict is resolved in favor of the parent's existing pages.
1926 * object must be held and chain-locked on call.
1928 * The caller must have an extra ref on object to prevent a race from
1929 * destroying it during the collapse.
1932 vm_object_collapse(vm_object_t object, struct vm_object_dealloc_list **dlistp)
1934 struct vm_object_dealloc_list *dlist = NULL;
1935 vm_object_t backing_object;
1938 * Only one thread is attempting a collapse at any given moment.
1939 * There are few restrictions for (object) that callers of this
1940 * function check so reentrancy is likely.
1942 KKASSERT(object != NULL);
1943 vm_object_assert_held(object);
1944 KKASSERT(object->flags & OBJ_CHAINLOCK);
1951 * We have to hold the backing object, check races.
1953 while ((backing_object = object->backing_object) != NULL) {
1954 vm_object_hold(backing_object);
1955 if (backing_object == object->backing_object)
1957 vm_object_drop(backing_object);
1961 * No backing object? Nothing to collapse then.
1963 if (backing_object == NULL)
1967 * You can't collapse with a non-default/non-swap object.
1969 if (backing_object->type != OBJT_DEFAULT &&
1970 backing_object->type != OBJT_SWAP) {
1971 vm_object_drop(backing_object);
1972 backing_object = NULL;
1977 * Chain-lock the backing object too because if we
1978 * successfully merge its pages into the top object we
1979 * will collapse backing_object->backing_object as the
1980 * new backing_object. Re-check that it is still our
1983 vm_object_chain_acquire(backing_object);
1984 if (backing_object != object->backing_object) {
1985 vm_object_chain_release(backing_object);
1986 vm_object_drop(backing_object);
1991 * we check the backing object first, because it is most likely
1994 if (backing_object->handle != NULL ||
1995 (backing_object->type != OBJT_DEFAULT &&
1996 backing_object->type != OBJT_SWAP) ||
1997 (backing_object->flags & OBJ_DEAD) ||
1998 object->handle != NULL ||
1999 (object->type != OBJT_DEFAULT &&
2000 object->type != OBJT_SWAP) ||
2001 (object->flags & OBJ_DEAD)) {
2006 * If paging is in progress we can't do a normal collapse.
2009 object->paging_in_progress != 0 ||
2010 backing_object->paging_in_progress != 0
2012 vm_object_qcollapse(object, backing_object);
2017 * We know that we can either collapse the backing object (if
2018 * the parent is the only reference to it) or (perhaps) have
2019 * the parent bypass the object if the parent happens to shadow
2020 * all the resident pages in the entire backing object.
2022 * This is ignoring pager-backed pages such as swap pages.
2023 * vm_object_backing_scan fails the shadowing test in this
2026 if (backing_object->ref_count == 1) {
2028 * If there is exactly one reference to the backing
2029 * object, we can collapse it into the parent.
2031 KKASSERT(object->backing_object == backing_object);
2032 vm_object_backing_scan(object, backing_object,
2033 OBSC_COLLAPSE_WAIT);
2036 * Move the pager from backing_object to object.
2038 if (backing_object->type == OBJT_SWAP) {
2039 vm_object_pip_add(backing_object, 1);
2042 * scrap the paging_offset junk and do a
2043 * discrete copy. This also removes major
2044 * assumptions about how the swap-pager
2045 * works from where it doesn't belong. The
2046 * new swapper is able to optimize the
2047 * destroy-source case.
2049 vm_object_pip_add(object, 1);
2050 swap_pager_copy(backing_object, object,
2051 OFF_TO_IDX(object->backing_object_offset),
2053 vm_object_pip_wakeup(object);
2054 vm_object_pip_wakeup(backing_object);
2058 * Object now shadows whatever backing_object did.
2059 * Remove object from backing_object's shadow_list.
2061 LIST_REMOVE(object, shadow_list);
2062 KKASSERT(object->backing_object == backing_object);
2063 backing_object->shadow_count--;
2064 backing_object->generation++;
2067 * backing_object->backing_object moves from within
2068 * backing_object to within object.
2070 while ((bbobj = backing_object->backing_object) != NULL) {
2071 vm_object_hold(bbobj);
2072 if (bbobj == backing_object->backing_object)
2074 vm_object_drop(bbobj);
2077 LIST_REMOVE(backing_object, shadow_list);
2078 bbobj->shadow_count--;
2079 bbobj->generation++;
2080 backing_object->backing_object = NULL;
2082 object->backing_object = bbobj;
2084 LIST_INSERT_HEAD(&bbobj->shadow_head,
2085 object, shadow_list);
2086 bbobj->shadow_count++;
2087 bbobj->generation++;
2090 object->backing_object_offset +=
2091 backing_object->backing_object_offset;
2093 vm_object_drop(bbobj);
2096 * Discard the old backing_object. Nothing should be
2097 * able to ref it, other than a vm_map_split(),
2098 * and vm_map_split() will stall on our chain lock.
2099 * And we control the parent so it shouldn't be
2100 * possible for it to go away either.
2102 * Since the backing object has no pages, no pager
2103 * left, and no object references within it, all
2104 * that is necessary is to dispose of it.
2106 KASSERT(backing_object->ref_count == 1,
2107 ("backing_object %p was somehow "
2108 "re-referenced during collapse!",
2110 KASSERT(RB_EMPTY(&backing_object->rb_memq),
2111 ("backing_object %p somehow has left "
2112 "over pages during collapse!",
2116 * The object can be destroyed.
2118 * XXX just fall through and dodealloc instead
2119 * of forcing destruction?
2121 --backing_object->ref_count;
2122 if ((backing_object->flags & OBJ_DEAD) == 0)
2123 vm_object_terminate(backing_object);
2128 * If we do not entirely shadow the backing object,
2129 * there is nothing we can do so we give up.
2131 if (vm_object_backing_scan(object, backing_object,
2132 OBSC_TEST_ALL_SHADOWED) == 0) {
2137 * bbobj is backing_object->backing_object. Since
2138 * object completely shadows backing_object we can
2139 * bypass it and become backed by bbobj instead.
2141 while ((bbobj = backing_object->backing_object) != NULL) {
2142 vm_object_hold(bbobj);
2143 if (bbobj == backing_object->backing_object)
2145 vm_object_drop(bbobj);
2149 * Make object shadow bbobj instead of backing_object.
2150 * Remove object from backing_object's shadow list.
2152 * Deallocating backing_object will not remove
2153 * it, since its reference count is at least 2.
2155 KKASSERT(object->backing_object == backing_object);
2156 LIST_REMOVE(object, shadow_list);
2157 backing_object->shadow_count--;
2158 backing_object->generation++;
2161 * Add a ref to bbobj, bbobj now shadows object.
2163 * NOTE: backing_object->backing_object still points
2164 * to bbobj. That relationship remains intact
2165 * because backing_object has > 1 ref, so
2166 * someone else is pointing to it (hence why
2167 * we can't collapse it into object and can
2168 * only handle the all-shadowed bypass case).
2171 vm_object_chain_wait(bbobj);
2172 vm_object_reference_locked(bbobj);
2173 LIST_INSERT_HEAD(&bbobj->shadow_head,
2174 object, shadow_list);
2175 bbobj->shadow_count++;
2176 bbobj->generation++;
2177 object->backing_object_offset +=
2178 backing_object->backing_object_offset;
2179 object->backing_object = bbobj;
2180 vm_object_drop(bbobj);
2182 object->backing_object = NULL;
2186 * Drop the reference count on backing_object. To
2187 * handle ref_count races properly we can't assume
2188 * that the ref_count is still at least 2 so we
2189 * have to actually call vm_object_deallocate()
2190 * (after clearing the chainlock).
2197 * Ok, we want to loop on the new object->bbobj association,
2198 * possibly collapsing it further. However if dodealloc is
2199 * non-zero we have to deallocate the backing_object which
2200 * itself can potentially undergo a collapse, creating a
2201 * recursion depth issue with the LWKT token subsystem.
2203 * In the case where we must deallocate the backing_object
2204 * it is possible now that the backing_object has a single
2205 * shadow count on some other object (not represented here
2206 * as yet), since it no longer shadows us. Thus when we
2207 * call vm_object_deallocate() it may attempt to collapse
2208 * itself into its remaining parent.
2211 struct vm_object_dealloc_list *dtmp;
2213 vm_object_chain_release(backing_object);
2214 vm_object_unlock(backing_object);
2215 /* backing_object remains held */
2218 * Auto-deallocation list for caller convenience.
2223 dtmp = kmalloc(sizeof(*dtmp), M_TEMP, M_WAITOK);
2224 dtmp->object = backing_object;
2225 dtmp->next = *dlistp;
2228 vm_object_chain_release(backing_object);
2229 vm_object_drop(backing_object);
2231 /* backing_object = NULL; not needed */
2236 * Clean up any left over backing_object
2238 if (backing_object) {
2239 vm_object_chain_release(backing_object);
2240 vm_object_drop(backing_object);
2244 * Clean up any auto-deallocation list. This is a convenience
2245 * for top-level callers so they don't have to pass &dlist.
2246 * Do not clean up any caller-passed dlistp, the caller will
2250 vm_object_deallocate_list(&dlist);
2255 * vm_object_collapse() may collect additional objects in need of
2256 * deallocation. This routine deallocates these objects. The
2257 * deallocation itself can trigger additional collapses (which the
2258 * deallocate function takes care of). This procedure is used to
2259 * reduce procedural recursion since these vm_object shadow chains
2260 * can become quite long.
2263 vm_object_deallocate_list(struct vm_object_dealloc_list **dlistp)
2265 struct vm_object_dealloc_list *dlist;
2267 while ((dlist = *dlistp) != NULL) {
2268 *dlistp = dlist->next;
2269 vm_object_lock(dlist->object);
2270 vm_object_deallocate_locked(dlist->object);
2271 vm_object_drop(dlist->object);
2272 kfree(dlist, M_TEMP);
2277 * Removes all physical pages in the specified object range from the
2278 * object's list of pages.
2282 static int vm_object_page_remove_callback(vm_page_t p, void *data);
2285 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2286 boolean_t clean_only)
2288 struct rb_vm_page_scan_info info;
2292 * Degenerate cases and assertions
2294 vm_object_hold(object);
2295 if (object == NULL ||
2296 (object->resident_page_count == 0 && object->swblock_count == 0)) {
2297 vm_object_drop(object);
2300 KASSERT(object->type != OBJT_PHYS,
2301 ("attempt to remove pages from a physical object"));
2304 * Indicate that paging is occuring on the object
2306 vm_object_pip_add(object, 1);
2309 * Figure out the actual removal range and whether we are removing
2310 * the entire contents of the object or not. If removing the entire
2311 * contents, be sure to get all pages, even those that might be
2312 * beyond the end of the object.
2314 info.start_pindex = start;
2316 info.end_pindex = (vm_pindex_t)-1;
2318 info.end_pindex = end - 1;
2319 info.limit = clean_only;
2320 all = (start == 0 && info.end_pindex >= object->size - 1);
2323 * Loop until we are sure we have gotten them all.
2327 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2328 vm_object_page_remove_callback, &info);
2329 } while (info.error);
2332 * Remove any related swap if throwing away pages, or for
2333 * non-swap objects (the swap is a clean copy in that case).
2335 if (object->type != OBJT_SWAP || clean_only == FALSE) {
2337 swap_pager_freespace_all(object);
2339 swap_pager_freespace(object, info.start_pindex,
2340 info.end_pindex - info.start_pindex + 1);
2346 vm_object_pip_wakeup(object);
2347 vm_object_drop(object);
2351 * The caller must hold the object
2354 vm_object_page_remove_callback(vm_page_t p, void *data)
2356 struct rb_vm_page_scan_info *info = data;
2358 if (vm_page_busy_try(p, TRUE)) {
2359 vm_page_sleep_busy(p, TRUE, "vmopar");
2365 * Wired pages cannot be destroyed, but they can be invalidated
2366 * and we do so if clean_only (limit) is not set.
2368 * WARNING! The page may be wired due to being part of a buffer
2369 * cache buffer, and the buffer might be marked B_CACHE.
2370 * This is fine as part of a truncation but VFSs must be
2371 * sure to fix the buffer up when re-extending the file.
2373 if (p->wire_count != 0) {
2374 vm_page_protect(p, VM_PROT_NONE);
2375 if (info->limit == 0)
2382 * limit is our clean_only flag. If set and the page is dirty, do
2383 * not free it. If set and the page is being held by someone, do
2386 if (info->limit && p->valid) {
2387 vm_page_test_dirty(p);
2388 if (p->valid & p->dirty) {
2393 if (p->hold_count) {
2403 vm_page_protect(p, VM_PROT_NONE);
2409 * Coalesces two objects backing up adjoining regions of memory into a
2412 * returns TRUE if objects were combined.
2414 * NOTE: Only works at the moment if the second object is NULL -
2415 * if it's not, which object do we lock first?
2418 * prev_object First object to coalesce
2419 * prev_offset Offset into prev_object
2420 * next_object Second object into coalesce
2421 * next_offset Offset into next_object
2423 * prev_size Size of reference to prev_object
2424 * next_size Size of reference to next_object
2426 * The caller does not need to hold (prev_object) but must have a stable
2427 * pointer to it (typically by holding the vm_map locked).
2430 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
2431 vm_size_t prev_size, vm_size_t next_size)
2433 vm_pindex_t next_pindex;
2435 if (prev_object == NULL)
2438 vm_object_hold(prev_object);
2440 if (prev_object->type != OBJT_DEFAULT &&
2441 prev_object->type != OBJT_SWAP) {
2442 vm_object_drop(prev_object);
2447 * Try to collapse the object first
2449 vm_object_chain_acquire(prev_object);
2450 vm_object_collapse(prev_object, NULL);
2453 * Can't coalesce if: . more than one reference . paged out . shadows
2454 * another object . has a copy elsewhere (any of which mean that the
2455 * pages not mapped to prev_entry may be in use anyway)
2458 if (prev_object->backing_object != NULL) {
2459 vm_object_chain_release(prev_object);
2460 vm_object_drop(prev_object);
2464 prev_size >>= PAGE_SHIFT;
2465 next_size >>= PAGE_SHIFT;
2466 next_pindex = prev_pindex + prev_size;
2468 if ((prev_object->ref_count > 1) &&
2469 (prev_object->size != next_pindex)) {
2470 vm_object_chain_release(prev_object);
2471 vm_object_drop(prev_object);
2476 * Remove any pages that may still be in the object from a previous
2479 if (next_pindex < prev_object->size) {
2480 vm_object_page_remove(prev_object,
2482 next_pindex + next_size, FALSE);
2483 if (prev_object->type == OBJT_SWAP)
2484 swap_pager_freespace(prev_object,
2485 next_pindex, next_size);
2489 * Extend the object if necessary.
2491 if (next_pindex + next_size > prev_object->size)
2492 prev_object->size = next_pindex + next_size;
2494 vm_object_chain_release(prev_object);
2495 vm_object_drop(prev_object);
2500 * Make the object writable and flag is being possibly dirty.
2502 * The caller must hold the object. XXX called from vm_page_dirty(),
2503 * There is currently no requirement to hold the object.
2506 vm_object_set_writeable_dirty(vm_object_t object)
2510 /*vm_object_assert_held(object);*/
2512 * Avoid contention in vm fault path by checking the state before
2513 * issuing an atomic op on it.
2515 if ((object->flags & (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) !=
2516 (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) {
2517 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
2519 if (object->type == OBJT_VNODE &&
2520 (vp = (struct vnode *)object->handle) != NULL) {
2521 if ((vp->v_flag & VOBJDIRTY) == 0) {
2522 vsetflags(vp, VOBJDIRTY);
2527 #include "opt_ddb.h"
2529 #include <sys/kernel.h>
2531 #include <sys/cons.h>
2533 #include <ddb/ddb.h>
2535 static int _vm_object_in_map (vm_map_t map, vm_object_t object,
2536 vm_map_entry_t entry);
2537 static int vm_object_in_map (vm_object_t object);
2540 * The caller must hold the object.
2543 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2546 vm_map_entry_t tmpe;
2547 vm_object_t obj, nobj;
2553 tmpe = map->header.next;
2554 entcount = map->nentries;
2555 while (entcount-- && (tmpe != &map->header)) {
2556 if( _vm_object_in_map(map, object, tmpe)) {
2563 switch(entry->maptype) {
2564 case VM_MAPTYPE_SUBMAP:
2565 tmpm = entry->object.sub_map;
2566 tmpe = tmpm->header.next;
2567 entcount = tmpm->nentries;
2568 while (entcount-- && tmpe != &tmpm->header) {
2569 if( _vm_object_in_map(tmpm, object, tmpe)) {
2575 case VM_MAPTYPE_NORMAL:
2576 case VM_MAPTYPE_VPAGETABLE:
2577 obj = entry->object.vm_object;
2579 if (obj == object) {
2580 if (obj != entry->object.vm_object)
2581 vm_object_drop(obj);
2584 while ((nobj = obj->backing_object) != NULL) {
2585 vm_object_hold(nobj);
2586 if (nobj == obj->backing_object)
2588 vm_object_drop(nobj);
2590 if (obj != entry->object.vm_object) {
2592 vm_object_lock_swap();
2593 vm_object_drop(obj);
2604 static int vm_object_in_map_callback(struct proc *p, void *data);
2606 struct vm_object_in_map_info {
2615 vm_object_in_map(vm_object_t object)
2617 struct vm_object_in_map_info info;
2620 info.object = object;
2622 allproc_scan(vm_object_in_map_callback, &info);
2625 if( _vm_object_in_map(&kernel_map, object, 0))
2627 if( _vm_object_in_map(&pager_map, object, 0))
2629 if( _vm_object_in_map(&buffer_map, object, 0))
2638 vm_object_in_map_callback(struct proc *p, void *data)
2640 struct vm_object_in_map_info *info = data;
2643 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) {
2651 DB_SHOW_COMMAND(vmochk, vm_object_check)
2656 * make sure that internal objs are in a map somewhere
2657 * and none have zero ref counts.
2659 for (object = TAILQ_FIRST(&vm_object_list);
2661 object = TAILQ_NEXT(object, object_list)) {
2662 if (object->type == OBJT_MARKER)
2664 if (object->handle == NULL &&
2665 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2666 if (object->ref_count == 0) {
2667 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2668 (long)object->size);
2670 if (!vm_object_in_map(object)) {
2672 "vmochk: internal obj is not in a map: "
2673 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2674 object->ref_count, (u_long)object->size,
2675 (u_long)object->size,
2676 (void *)object->backing_object);
2685 DB_SHOW_COMMAND(object, vm_object_print_static)
2687 /* XXX convert args. */
2688 vm_object_t object = (vm_object_t)addr;
2689 boolean_t full = have_addr;
2693 /* XXX count is an (unused) arg. Avoid shadowing it. */
2694 #define count was_count
2702 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
2703 object, (int)object->type, (u_long)object->size,
2704 object->resident_page_count, object->ref_count, object->flags);
2706 * XXX no %qd in kernel. Truncate object->backing_object_offset.
2708 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
2709 object->shadow_count,
2710 object->backing_object ? object->backing_object->ref_count : 0,
2711 object->backing_object, (long)object->backing_object_offset);
2718 RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) {
2720 db_iprintf("memory:=");
2721 else if (count == 6) {
2729 db_printf("(off=0x%lx,page=0x%lx)",
2730 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
2741 * XXX need this non-static entry for calling from vm_map_print.
2746 vm_object_print(/* db_expr_t */ long addr,
2747 boolean_t have_addr,
2748 /* db_expr_t */ long count,
2751 vm_object_print_static(addr, have_addr, count, modif);
2757 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2762 for (object = TAILQ_FIRST(&vm_object_list);
2764 object = TAILQ_NEXT(object, object_list)) {
2765 vm_pindex_t idx, fidx;
2767 vm_paddr_t pa = -1, padiff;
2771 if (object->type == OBJT_MARKER)
2773 db_printf("new object: %p\n", (void *)object);
2783 osize = object->size;
2786 for (idx = 0; idx < osize; idx++) {
2787 m = vm_page_lookup(object, idx);
2790 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2791 (long)fidx, rcount, (long)pa);
2806 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2811 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
2812 padiff >>= PAGE_SHIFT;
2813 padiff &= PQ_L2_MASK;
2815 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
2819 db_printf(" index(%ld)run(%d)pa(0x%lx)",
2820 (long)fidx, rcount, (long)pa);
2821 db_printf("pd(%ld)\n", (long)padiff);
2831 pa = VM_PAGE_TO_PHYS(m);
2835 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2836 (long)fidx, rcount, (long)pa);