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 void 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);
433 pmap_object_init(object);
435 vm_object_hold(object);
436 lwkt_gettoken(&vmobj_token);
437 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
439 lwkt_reltoken(&vmobj_token);
443 * Initialize the VM objects module.
445 * Called from the low level boot code only.
450 TAILQ_INIT(&vm_object_list);
452 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd),
454 vm_object_drop(&kernel_object);
456 obj_zone = &obj_zone_store;
457 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
458 vm_objects_init, VM_OBJECTS_INIT);
462 vm_object_init2(void)
464 zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1);
468 * Allocate and return a new object of the specified type and size.
473 vm_object_allocate(objtype_t type, vm_pindex_t size)
477 result = (vm_object_t) zalloc(obj_zone);
479 _vm_object_allocate(type, size, result);
480 vm_object_drop(result);
486 * This version returns a held object, allowing further atomic initialization
490 vm_object_allocate_hold(objtype_t type, vm_pindex_t size)
494 result = (vm_object_t) zalloc(obj_zone);
496 _vm_object_allocate(type, size, result);
502 * Add an additional reference to a vm_object. The object must already be
503 * held. The original non-lock version is no longer supported. The object
504 * must NOT be chain locked by anyone at the time the reference is added.
506 * Referencing a chain-locked object can blow up the fairly sensitive
507 * ref_count and shadow_count tests in the deallocator. Most callers
508 * will call vm_object_chain_wait() prior to calling
509 * vm_object_reference_locked() to avoid the case.
511 * The object must be held.
514 vm_object_reference_locked(vm_object_t object)
516 KKASSERT(object != NULL);
517 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
518 KKASSERT((object->flags & OBJ_CHAINLOCK) == 0);
520 if (object->type == OBJT_VNODE) {
521 vref(object->handle);
522 /* XXX what if the vnode is being destroyed? */
527 * Object OBJ_CHAINLOCK lock handling.
529 * The caller can chain-lock backing objects recursively and then
530 * use vm_object_chain_release_all() to undo the whole chain.
532 * Chain locks are used to prevent collapses and are only applicable
533 * to OBJT_DEFAULT and OBJT_SWAP objects. Chain locking operations
534 * on other object types are ignored. This is also important because
535 * it allows e.g. the vnode underlying a memory mapping to take concurrent
538 * The object must usually be held on entry, though intermediate
539 * objects need not be held on release.
542 vm_object_chain_wait(vm_object_t object)
544 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
545 while (object->flags & OBJ_CHAINLOCK) {
546 vm_object_set_flag(object, OBJ_CHAINWANT);
547 tsleep(object, 0, "objchain", 0);
552 vm_object_chain_acquire(vm_object_t object)
554 if (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) {
555 vm_object_chain_wait(object);
556 vm_object_set_flag(object, OBJ_CHAINLOCK);
561 vm_object_chain_release(vm_object_t object)
563 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
564 if (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) {
565 KKASSERT(object->flags & OBJ_CHAINLOCK);
566 if (object->flags & OBJ_CHAINWANT) {
567 vm_object_clear_flag(object,
568 OBJ_CHAINLOCK | OBJ_CHAINWANT);
571 vm_object_clear_flag(object, OBJ_CHAINLOCK);
577 * This releases the entire chain of objects from first_object to and
578 * including stopobj, flowing through object->backing_object.
580 * We release stopobj first as an optimization as this object is most
581 * likely to be shared across multiple processes.
584 vm_object_chain_release_all(vm_object_t first_object, vm_object_t stopobj)
586 vm_object_t backing_object;
589 vm_object_chain_release(stopobj);
590 object = first_object;
592 while (object != stopobj) {
594 if (object != first_object)
595 vm_object_hold(object);
596 backing_object = object->backing_object;
597 vm_object_chain_release(object);
598 if (object != first_object)
599 vm_object_drop(object);
600 object = backing_object;
605 * Dereference an object and its underlying vnode.
607 * The object must be held and will be held on return.
610 vm_object_vndeallocate(vm_object_t object)
612 struct vnode *vp = (struct vnode *) object->handle;
614 KASSERT(object->type == OBJT_VNODE,
615 ("vm_object_vndeallocate: not a vnode object"));
616 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
617 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
619 if (object->ref_count == 0) {
620 vprint("vm_object_vndeallocate", vp);
621 panic("vm_object_vndeallocate: bad object reference count");
625 if (object->ref_count == 0)
626 vclrflags(vp, VTEXT);
631 * Release a reference to the specified object, gained either through a
632 * vm_object_allocate or a vm_object_reference call. When all references
633 * are gone, storage associated with this object may be relinquished.
635 * The caller does not have to hold the object locked but must have control
636 * over the reference in question in order to guarantee that the object
637 * does not get ripped out from under us.
640 vm_object_deallocate(vm_object_t object)
643 vm_object_hold(object);
644 vm_object_deallocate_locked(object);
645 vm_object_drop(object);
650 vm_object_deallocate_locked(vm_object_t object)
652 struct vm_object_dealloc_list *dlist = NULL;
653 struct vm_object_dealloc_list *dtmp;
658 * We may chain deallocate object, but additional objects may
659 * collect on the dlist which also have to be deallocated. We
660 * must avoid a recursion, vm_object chains can get deep.
663 while (object != NULL) {
666 * Don't rip a ref_count out from under an object undergoing
667 * collapse, it will confuse the collapse code.
669 vm_object_chain_wait(object);
671 if (object->type == OBJT_VNODE) {
672 vm_object_vndeallocate(object);
676 if (object->ref_count == 0) {
677 panic("vm_object_deallocate: object deallocated "
678 "too many times: %d", object->type);
680 if (object->ref_count > 2) {
686 * Here on ref_count of one or two, which are special cases for
689 * Nominal ref_count > 1 case if the second ref is not from
692 if (object->ref_count == 2 && object->shadow_count == 0) {
693 vm_object_set_flag(object, OBJ_ONEMAPPING);
699 * If the second ref is from a shadow we chain along it
700 * upwards if object's handle is exhausted.
702 * We have to decrement object->ref_count before potentially
703 * collapsing the first shadow object or the collapse code
704 * will not be able to handle the degenerate case to remove
705 * object. However, if we do it too early the object can
706 * get ripped out from under us.
708 if (object->ref_count == 2 && object->shadow_count == 1 &&
709 object->handle == NULL && (object->type == OBJT_DEFAULT ||
710 object->type == OBJT_SWAP)) {
711 temp = LIST_FIRST(&object->shadow_head);
712 KKASSERT(temp != NULL);
713 vm_object_hold(temp);
716 * Wait for any paging to complete so the collapse
717 * doesn't (or isn't likely to) qcollapse. pip
718 * waiting must occur before we acquire the
722 temp->paging_in_progress ||
723 object->paging_in_progress
725 vm_object_pip_wait(temp, "objde1");
726 vm_object_pip_wait(object, "objde2");
730 * If the parent is locked we have to give up, as
731 * otherwise we would be acquiring locks in the
732 * wrong order and potentially deadlock.
734 if (temp->flags & OBJ_CHAINLOCK) {
735 vm_object_drop(temp);
738 vm_object_chain_acquire(temp);
741 * Recheck/retry after the hold and the paging
742 * wait, both of which can block us.
744 if (object->ref_count != 2 ||
745 object->shadow_count != 1 ||
747 LIST_FIRST(&object->shadow_head) != temp ||
748 (object->type != OBJT_DEFAULT &&
749 object->type != OBJT_SWAP)) {
750 vm_object_chain_release(temp);
751 vm_object_drop(temp);
756 * We can safely drop object's ref_count now.
758 KKASSERT(object->ref_count == 2);
762 * If our single parent is not collapseable just
763 * decrement ref_count (2->1) and stop.
765 if (temp->handle || (temp->type != OBJT_DEFAULT &&
766 temp->type != OBJT_SWAP)) {
767 vm_object_chain_release(temp);
768 vm_object_drop(temp);
773 * At this point we have already dropped object's
774 * ref_count so it is possible for a race to
775 * deallocate obj out from under us. Any collapse
776 * will re-check the situation. We must not block
777 * until we are able to collapse.
779 * Bump temp's ref_count to avoid an unwanted
780 * degenerate recursion (can't call
781 * vm_object_reference_locked() because it asserts
782 * that CHAINLOCK is not set).
785 KKASSERT(temp->ref_count > 1);
788 * Collapse temp, then deallocate the extra ref
791 vm_object_collapse(temp, &dlist);
792 vm_object_chain_release(temp);
794 vm_object_lock_swap();
795 vm_object_drop(object);
803 * Drop the ref and handle termination on the 1->0 transition.
804 * We may have blocked above so we have to recheck.
807 KKASSERT(object->ref_count != 0);
808 if (object->ref_count >= 2) {
812 KKASSERT(object->ref_count == 1);
815 * 1->0 transition. Chain through the backing_object.
816 * Maintain the ref until we've located the backing object,
819 while ((temp = object->backing_object) != NULL) {
820 vm_object_hold(temp);
821 if (temp == object->backing_object)
823 vm_object_drop(temp);
827 * 1->0 transition verified, retry if ref_count is no longer
828 * 1. Otherwise disconnect the backing_object (temp) and
831 if (object->ref_count != 1) {
832 vm_object_drop(temp);
837 * It shouldn't be possible for the object to be chain locked
838 * if we're removing the last ref on it.
840 KKASSERT((object->flags & OBJ_CHAINLOCK) == 0);
843 LIST_REMOVE(object, shadow_list);
844 temp->shadow_count--;
846 object->backing_object = NULL;
850 if ((object->flags & OBJ_DEAD) == 0)
851 vm_object_terminate(object);
852 if (must_drop && temp)
853 vm_object_lock_swap();
855 vm_object_drop(object);
859 if (must_drop && object)
860 vm_object_drop(object);
863 * Additional tail recursion on dlist. Avoid a recursion. Objects
864 * on the dlist have a hold count but are not locked.
866 if ((dtmp = dlist) != NULL) {
868 object = dtmp->object;
871 vm_object_lock(object); /* already held, add lock */
872 must_drop = 1; /* and we're responsible for it */
878 * Destroy the specified object, freeing up related resources.
880 * The object must have zero references.
882 * The object must held. The caller is responsible for dropping the object
883 * after terminate returns. Terminate does NOT drop the object.
885 static int vm_object_terminate_callback(vm_page_t p, void *data);
888 vm_object_terminate(vm_object_t object)
891 * Make sure no one uses us. Once we set OBJ_DEAD we should be
892 * able to safely block.
894 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
895 KKASSERT((object->flags & OBJ_DEAD) == 0);
896 vm_object_set_flag(object, OBJ_DEAD);
899 * Wait for the pageout daemon to be done with the object
901 vm_object_pip_wait(object, "objtrm1");
903 KASSERT(!object->paging_in_progress,
904 ("vm_object_terminate: pageout in progress"));
907 * Clean and free the pages, as appropriate. All references to the
908 * object are gone, so we don't need to lock it.
910 if (object->type == OBJT_VNODE) {
914 * Clean pages and flush buffers.
916 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
918 vp = (struct vnode *) object->handle;
919 vinvalbuf(vp, V_SAVE, 0, 0);
923 * Wait for any I/O to complete, after which there had better not
924 * be any references left on the object.
926 vm_object_pip_wait(object, "objtrm2");
928 if (object->ref_count != 0) {
929 panic("vm_object_terminate: object with references, "
930 "ref_count=%d", object->ref_count);
934 * Cleanup any shared pmaps associated with this object.
936 pmap_object_free(object);
939 * Now free any remaining pages. For internal objects, this also
940 * removes them from paging queues. Don't free wired pages, just
941 * remove them from the object.
943 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
944 vm_object_terminate_callback, NULL);
947 * Let the pager know object is dead.
949 vm_pager_deallocate(object);
952 * Wait for the object hold count to hit 1, clean out pages as
953 * we go. vmobj_token interlocks any race conditions that might
954 * pick the object up from the vm_object_list after we have cleared
958 if (RB_ROOT(&object->rb_memq) == NULL)
960 kprintf("vm_object_terminate: Warning, object %p "
961 "still has %d pages\n",
962 object, object->resident_page_count);
963 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
964 vm_object_terminate_callback, NULL);
968 * There had better not be any pages left
970 KKASSERT(object->resident_page_count == 0);
973 * Remove the object from the global object list.
975 lwkt_gettoken(&vmobj_token);
976 TAILQ_REMOVE(&vm_object_list, object, object_list);
978 lwkt_reltoken(&vmobj_token);
979 vm_object_dead_wakeup(object);
981 if (object->ref_count != 0) {
982 panic("vm_object_terminate2: object with references, "
983 "ref_count=%d", object->ref_count);
987 * NOTE: The object hold_count is at least 1, so we cannot zfree()
988 * the object here. See vm_object_drop().
993 * The caller must hold the object.
996 vm_object_terminate_callback(vm_page_t p, void *data __unused)
1001 vm_page_busy_wait(p, TRUE, "vmpgtrm");
1002 if (object != p->object) {
1003 kprintf("vm_object_terminate: Warning: Encountered "
1004 "busied page %p on queue %d\n", p, p->queue);
1006 } else if (p->wire_count == 0) {
1008 * NOTE: PG_NEED_COMMIT is ignored.
1011 mycpu->gd_cnt.v_pfree++;
1013 if (p->queue != PQ_NONE)
1014 kprintf("vm_object_terminate: Warning: Encountered "
1015 "wired page %p on queue %d\n", p, p->queue);
1024 * The object is dead but still has an object<->pager association. Sleep
1025 * and return. The caller typically retests the association in a loop.
1027 * The caller must hold the object.
1030 vm_object_dead_sleep(vm_object_t object, const char *wmesg)
1032 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1033 if (object->handle) {
1034 vm_object_set_flag(object, OBJ_DEADWNT);
1035 tsleep(object, 0, wmesg, 0);
1036 /* object may be invalid after this point */
1041 * Wakeup anyone waiting for the object<->pager disassociation on
1044 * The caller must hold the object.
1047 vm_object_dead_wakeup(vm_object_t object)
1049 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1050 if (object->flags & OBJ_DEADWNT) {
1051 vm_object_clear_flag(object, OBJ_DEADWNT);
1057 * Clean all dirty pages in the specified range of object. Leaves page
1058 * on whatever queue it is currently on. If NOSYNC is set then do not
1059 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
1060 * leaving the object dirty.
1062 * When stuffing pages asynchronously, allow clustering. XXX we need a
1063 * synchronous clustering mode implementation.
1065 * Odd semantics: if start == end, we clean everything.
1067 * The object must be locked? XXX
1069 static int vm_object_page_clean_pass1(struct vm_page *p, void *data);
1070 static int vm_object_page_clean_pass2(struct vm_page *p, void *data);
1073 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1076 struct rb_vm_page_scan_info info;
1082 vm_object_hold(object);
1083 if (object->type != OBJT_VNODE ||
1084 (object->flags & OBJ_MIGHTBEDIRTY) == 0) {
1085 vm_object_drop(object);
1089 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ?
1090 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1091 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
1093 vp = object->handle;
1096 * Interlock other major object operations. This allows us to
1097 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY.
1099 vm_object_set_flag(object, OBJ_CLEANING);
1102 * Handle 'entire object' case
1104 info.start_pindex = start;
1106 info.end_pindex = object->size - 1;
1108 info.end_pindex = end - 1;
1110 wholescan = (start == 0 && info.end_pindex == object->size - 1);
1112 info.pagerflags = pagerflags;
1113 info.object = object;
1116 * If cleaning the entire object do a pass to mark the pages read-only.
1117 * If everything worked out ok, clear OBJ_WRITEABLE and
1122 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1123 vm_object_page_clean_pass1, &info);
1124 if (info.error == 0) {
1125 vm_object_clear_flag(object,
1126 OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1127 if (object->type == OBJT_VNODE &&
1128 (vp = (struct vnode *)object->handle) != NULL) {
1129 if (vp->v_flag & VOBJDIRTY)
1130 vclrflags(vp, VOBJDIRTY);
1136 * Do a pass to clean all the dirty pages we find.
1140 generation = object->generation;
1141 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1142 vm_object_page_clean_pass2, &info);
1143 } while (info.error || generation != object->generation);
1145 vm_object_clear_flag(object, OBJ_CLEANING);
1146 vm_object_drop(object);
1150 * The caller must hold the object.
1154 vm_object_page_clean_pass1(struct vm_page *p, void *data)
1156 struct rb_vm_page_scan_info *info = data;
1158 vm_page_flag_set(p, PG_CLEANCHK);
1159 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1161 } else if (vm_page_busy_try(p, FALSE) == 0) {
1162 vm_page_protect(p, VM_PROT_READ); /* must not block */
1172 * The caller must hold the object
1176 vm_object_page_clean_pass2(struct vm_page *p, void *data)
1178 struct rb_vm_page_scan_info *info = data;
1182 * Do not mess with pages that were inserted after we started
1183 * the cleaning pass.
1185 if ((p->flags & PG_CLEANCHK) == 0)
1188 generation = info->object->generation;
1189 vm_page_busy_wait(p, TRUE, "vpcwai");
1190 if (p->object != info->object ||
1191 info->object->generation != generation) {
1198 * Before wasting time traversing the pmaps, check for trivial
1199 * cases where the page cannot be dirty.
1201 if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) {
1202 KKASSERT((p->dirty & p->valid) == 0);
1208 * Check whether the page is dirty or not. The page has been set
1209 * to be read-only so the check will not race a user dirtying the
1212 vm_page_test_dirty(p);
1213 if ((p->dirty & p->valid) == 0) {
1214 vm_page_flag_clear(p, PG_CLEANCHK);
1220 * If we have been asked to skip nosync pages and this is a
1221 * nosync page, skip it. Note that the object flags were
1222 * not cleared in this case (because pass1 will have returned an
1223 * error), so we do not have to set them.
1225 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1226 vm_page_flag_clear(p, PG_CLEANCHK);
1232 * Flush as many pages as we can. PG_CLEANCHK will be cleared on
1233 * the pages that get successfully flushed. Set info->error if
1234 * we raced an object modification.
1236 vm_object_page_collect_flush(info->object, p, info->pagerflags);
1244 * Collect the specified page and nearby pages and flush them out.
1245 * The number of pages flushed is returned. The passed page is busied
1246 * by the caller and we are responsible for its disposition.
1248 * The caller must hold the object.
1251 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags)
1260 vm_page_t maf[vm_pageout_page_count];
1261 vm_page_t mab[vm_pageout_page_count];
1262 vm_page_t ma[vm_pageout_page_count];
1264 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1269 for(i = 1; i < vm_pageout_page_count; i++) {
1272 tp = vm_page_lookup_busy_try(object, pi + i, TRUE, &error);
1277 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1278 (tp->flags & PG_CLEANCHK) == 0) {
1282 if ((tp->queue - tp->pc) == PQ_CACHE) {
1283 vm_page_flag_clear(tp, PG_CLEANCHK);
1287 vm_page_test_dirty(tp);
1288 if ((tp->dirty & tp->valid) == 0) {
1289 vm_page_flag_clear(tp, PG_CLEANCHK);
1298 chkb = vm_pageout_page_count - maxf;
1300 * NOTE: chkb can be 0
1302 for(i = 1; chkb && i < chkb; i++) {
1305 tp = vm_page_lookup_busy_try(object, pi - i, TRUE, &error);
1310 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1311 (tp->flags & PG_CLEANCHK) == 0) {
1315 if ((tp->queue - tp->pc) == PQ_CACHE) {
1316 vm_page_flag_clear(tp, PG_CLEANCHK);
1320 vm_page_test_dirty(tp);
1321 if ((tp->dirty & tp->valid) == 0) {
1322 vm_page_flag_clear(tp, PG_CLEANCHK);
1331 * All pages in the maf[] and mab[] array are busied.
1333 for (i = 0; i < maxb; i++) {
1334 int index = (maxb - i) - 1;
1336 vm_page_flag_clear(ma[index], PG_CLEANCHK);
1338 vm_page_flag_clear(p, PG_CLEANCHK);
1340 for(i = 0; i < maxf; i++) {
1341 int index = (maxb + i) + 1;
1343 vm_page_flag_clear(ma[index], PG_CLEANCHK);
1345 runlen = maxb + maxf + 1;
1347 for (i = 0; i < runlen; i++)
1348 vm_page_hold(ma[i]);
1350 vm_pageout_flush(ma, runlen, pagerflags);
1353 * WARNING: Related pages are still held but the BUSY was inherited
1354 * by the pageout I/O, so the pages might not be busy any
1355 * more. We cannot re-protect the page without waiting
1356 * for the I/O to complete and then busying it again.
1358 for (i = 0; i < runlen; i++) {
1359 if (ma[i]->valid & ma[i]->dirty) {
1360 /*vm_page_protect(ma[i], VM_PROT_READ);*/
1361 vm_page_flag_set(ma[i], PG_CLEANCHK);
1364 * maxf will end up being the actual number of pages
1365 * we wrote out contiguously, non-inclusive of the
1366 * first page. We do not count look-behind pages.
1368 if (i >= maxb + 1 && (maxf > i - maxb - 1))
1369 maxf = i - maxb - 1;
1371 vm_page_unhold(ma[i]);
1373 /*return(maxf + 1);*/
1377 * Same as vm_object_pmap_copy, except range checking really
1378 * works, and is meant for small sections of an object.
1380 * This code protects resident pages by making them read-only
1381 * and is typically called on a fork or split when a page
1382 * is converted to copy-on-write.
1384 * NOTE: If the page is already at VM_PROT_NONE, calling
1385 * vm_page_protect will have no effect.
1388 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1393 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
1396 vm_object_hold(object);
1397 for (idx = start; idx < end; idx++) {
1398 p = vm_page_lookup(object, idx);
1401 vm_page_protect(p, VM_PROT_READ);
1403 vm_object_drop(object);
1407 * Removes all physical pages in the specified object range from all
1410 * The object must *not* be locked.
1413 static int vm_object_pmap_remove_callback(vm_page_t p, void *data);
1416 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1418 struct rb_vm_page_scan_info info;
1422 info.start_pindex = start;
1423 info.end_pindex = end - 1;
1425 vm_object_hold(object);
1426 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1427 vm_object_pmap_remove_callback, &info);
1428 if (start == 0 && end == object->size)
1429 vm_object_clear_flag(object, OBJ_WRITEABLE);
1430 vm_object_drop(object);
1434 * The caller must hold the object
1437 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused)
1439 vm_page_protect(p, VM_PROT_NONE);
1444 * Implements the madvise function at the object/page level.
1446 * MADV_WILLNEED (any object)
1448 * Activate the specified pages if they are resident.
1450 * MADV_DONTNEED (any object)
1452 * Deactivate the specified pages if they are resident.
1454 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only)
1456 * Deactivate and clean the specified pages if they are
1457 * resident. This permits the process to reuse the pages
1458 * without faulting or the kernel to reclaim the pages
1464 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1466 vm_pindex_t end, tpindex;
1467 vm_object_t tobject;
1475 end = pindex + count;
1477 vm_object_hold(object);
1481 * Locate and adjust resident pages
1483 for (; pindex < end; pindex += 1) {
1485 if (tobject != object)
1486 vm_object_drop(tobject);
1491 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1492 * and those pages must be OBJ_ONEMAPPING.
1494 if (advise == MADV_FREE) {
1495 if ((tobject->type != OBJT_DEFAULT &&
1496 tobject->type != OBJT_SWAP) ||
1497 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1502 m = vm_page_lookup_busy_try(tobject, tpindex, TRUE, &error);
1505 vm_page_sleep_busy(m, TRUE, "madvpo");
1510 * There may be swap even if there is no backing page
1512 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1513 swap_pager_freespace(tobject, tpindex, 1);
1518 while ((xobj = tobject->backing_object) != NULL) {
1519 KKASSERT(xobj != object);
1520 vm_object_hold(xobj);
1521 if (xobj == tobject->backing_object)
1523 vm_object_drop(xobj);
1527 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1528 if (tobject != object) {
1529 vm_object_lock_swap();
1530 vm_object_drop(tobject);
1537 * If the page is not in a normal active state, we skip it.
1538 * If the page is not managed there are no page queues to
1539 * mess with. Things can break if we mess with pages in
1540 * any of the below states.
1542 if (m->wire_count ||
1543 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
1544 m->valid != VM_PAGE_BITS_ALL
1551 * Theoretically once a page is known not to be busy, an
1552 * interrupt cannot come along and rip it out from under us.
1555 if (advise == MADV_WILLNEED) {
1556 vm_page_activate(m);
1557 } else if (advise == MADV_DONTNEED) {
1558 vm_page_dontneed(m);
1559 } else if (advise == MADV_FREE) {
1561 * Mark the page clean. This will allow the page
1562 * to be freed up by the system. However, such pages
1563 * are often reused quickly by malloc()/free()
1564 * so we do not do anything that would cause
1565 * a page fault if we can help it.
1567 * Specifically, we do not try to actually free
1568 * the page now nor do we try to put it in the
1569 * cache (which would cause a page fault on reuse).
1571 * But we do make the page is freeable as we
1572 * can without actually taking the step of unmapping
1575 pmap_clear_modify(m);
1578 vm_page_dontneed(m);
1579 if (tobject->type == OBJT_SWAP)
1580 swap_pager_freespace(tobject, tpindex, 1);
1584 if (tobject != object)
1585 vm_object_drop(tobject);
1586 vm_object_drop(object);
1590 * Create a new object which is backed by the specified existing object
1591 * range. Replace the pointer and offset that was pointing at the existing
1592 * object with the pointer/offset for the new object.
1594 * No other requirements.
1597 vm_object_shadow(vm_object_t *objectp, vm_ooffset_t *offset, vm_size_t length,
1606 * Don't create the new object if the old object isn't shared.
1607 * We have to chain wait before adding the reference to avoid
1608 * racing a collapse or deallocation.
1610 * Add the additional ref to source here to avoid racing a later
1611 * collapse or deallocation. Clear the ONEMAPPING flag whether
1612 * addref is TRUE or not in this case because the original object
1616 vm_object_hold(source);
1617 vm_object_chain_wait(source);
1618 if (source->ref_count == 1 &&
1619 source->handle == NULL &&
1620 (source->type == OBJT_DEFAULT ||
1621 source->type == OBJT_SWAP)) {
1622 vm_object_drop(source);
1624 vm_object_reference_locked(source);
1625 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1629 vm_object_reference_locked(source);
1630 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1634 * Allocate a new object with the given length. The new object
1635 * is returned referenced but we may have to add another one.
1636 * If we are adding a second reference we must clear OBJ_ONEMAPPING.
1637 * (typically because the caller is about to clone a vm_map_entry).
1639 * The source object currently has an extra reference to prevent
1640 * collapses into it while we mess with its shadow list, which
1641 * we will remove later in this routine.
1643 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1644 panic("vm_object_shadow: no object for shadowing");
1645 vm_object_hold(result);
1647 vm_object_reference_locked(result);
1648 vm_object_clear_flag(result, OBJ_ONEMAPPING);
1652 * The new object shadows the source object. Chain wait before
1653 * adjusting shadow_count or the shadow list to avoid races.
1655 * Try to optimize the result object's page color when shadowing
1656 * in order to maintain page coloring consistency in the combined
1659 KKASSERT(result->backing_object == NULL);
1660 result->backing_object = source;
1662 vm_object_chain_wait(source);
1663 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1664 source->shadow_count++;
1665 source->generation++;
1667 /* cpu localization twist */
1668 result->pg_color = (int)(intptr_t)curthread;
1670 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1676 * Adjust the return storage. Drop the ref on source before
1679 result->backing_object_offset = *offset;
1680 vm_object_drop(result);
1683 vm_object_deallocate_locked(source);
1684 vm_object_drop(source);
1688 * Return the new things
1693 #define OBSC_TEST_ALL_SHADOWED 0x0001
1694 #define OBSC_COLLAPSE_NOWAIT 0x0002
1695 #define OBSC_COLLAPSE_WAIT 0x0004
1697 static int vm_object_backing_scan_callback(vm_page_t p, void *data);
1700 * The caller must hold the object.
1703 vm_object_backing_scan(vm_object_t object, vm_object_t backing_object, int op)
1705 struct rb_vm_page_scan_info info;
1707 vm_object_assert_held(object);
1708 vm_object_assert_held(backing_object);
1710 KKASSERT(backing_object == object->backing_object);
1711 info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1714 * Initial conditions
1716 if (op & OBSC_TEST_ALL_SHADOWED) {
1718 * We do not want to have to test for the existence of
1719 * swap pages in the backing object. XXX but with the
1720 * new swapper this would be pretty easy to do.
1722 * XXX what about anonymous MAP_SHARED memory that hasn't
1723 * been ZFOD faulted yet? If we do not test for this, the
1724 * shadow test may succeed! XXX
1726 if (backing_object->type != OBJT_DEFAULT)
1729 if (op & OBSC_COLLAPSE_WAIT) {
1730 KKASSERT((backing_object->flags & OBJ_DEAD) == 0);
1731 vm_object_set_flag(backing_object, OBJ_DEAD);
1732 lwkt_gettoken(&vmobj_token);
1733 TAILQ_REMOVE(&vm_object_list, backing_object, object_list);
1735 lwkt_reltoken(&vmobj_token);
1736 vm_object_dead_wakeup(backing_object);
1740 * Our scan. We have to retry if a negative error code is returned,
1741 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that
1742 * the scan had to be stopped because the parent does not completely
1745 info.object = object;
1746 info.backing_object = backing_object;
1750 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL,
1751 vm_object_backing_scan_callback,
1753 } while (info.error < 0);
1759 * The caller must hold the object.
1762 vm_object_backing_scan_callback(vm_page_t p, void *data)
1764 struct rb_vm_page_scan_info *info = data;
1765 vm_object_t backing_object;
1768 vm_pindex_t new_pindex;
1769 vm_pindex_t backing_offset_index;
1773 new_pindex = pindex - info->backing_offset_index;
1775 object = info->object;
1776 backing_object = info->backing_object;
1777 backing_offset_index = info->backing_offset_index;
1779 if (op & OBSC_TEST_ALL_SHADOWED) {
1783 * Ignore pages outside the parent object's range
1784 * and outside the parent object's mapping of the
1787 * note that we do not busy the backing object's
1790 if (pindex < backing_offset_index ||
1791 new_pindex >= object->size
1797 * See if the parent has the page or if the parent's
1798 * object pager has the page. If the parent has the
1799 * page but the page is not valid, the parent's
1800 * object pager must have the page.
1802 * If this fails, the parent does not completely shadow
1803 * the object and we might as well give up now.
1805 pp = vm_page_lookup(object, new_pindex);
1806 if ((pp == NULL || pp->valid == 0) &&
1807 !vm_pager_has_page(object, new_pindex)
1809 info->error = 0; /* problemo */
1810 return(-1); /* stop the scan */
1815 * Check for busy page. Note that we may have lost (p) when we
1816 * possibly blocked above.
1818 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1821 if (vm_page_busy_try(p, TRUE)) {
1822 if (op & OBSC_COLLAPSE_NOWAIT) {
1826 * If we slept, anything could have
1827 * happened. Ask that the scan be restarted.
1829 * Since the object is marked dead, the
1830 * backing offset should not have changed.
1832 vm_page_sleep_busy(p, TRUE, "vmocol");
1839 * If (p) is no longer valid restart the scan.
1841 if (p->object != backing_object || p->pindex != pindex) {
1842 kprintf("vm_object_backing_scan: Warning: page "
1843 "%p ripped out from under us\n", p);
1849 if (op & OBSC_COLLAPSE_NOWAIT) {
1850 if (p->valid == 0 ||
1852 (p->flags & PG_NEED_COMMIT)) {
1857 /* XXX what if p->valid == 0 , hold_count, etc? */
1861 p->object == backing_object,
1862 ("vm_object_qcollapse(): object mismatch")
1866 * Destroy any associated swap
1868 if (backing_object->type == OBJT_SWAP)
1869 swap_pager_freespace(backing_object, p->pindex, 1);
1872 p->pindex < backing_offset_index ||
1873 new_pindex >= object->size
1876 * Page is out of the parent object's range, we
1877 * can simply destroy it.
1879 vm_page_protect(p, VM_PROT_NONE);
1884 pp = vm_page_lookup(object, new_pindex);
1885 if (pp != NULL || vm_pager_has_page(object, new_pindex)) {
1887 * page already exists in parent OR swap exists
1888 * for this location in the parent. Destroy
1889 * the original page from the backing object.
1891 * Leave the parent's page alone
1893 vm_page_protect(p, VM_PROT_NONE);
1899 * Page does not exist in parent, rename the
1900 * page from the backing object to the main object.
1902 * If the page was mapped to a process, it can remain
1903 * mapped through the rename.
1905 if ((p->queue - p->pc) == PQ_CACHE)
1906 vm_page_deactivate(p);
1908 vm_page_rename(p, object, new_pindex);
1910 /* page automatically made dirty by rename */
1916 * This version of collapse allows the operation to occur earlier and
1917 * when paging_in_progress is true for an object... This is not a complete
1918 * operation, but should plug 99.9% of the rest of the leaks.
1920 * The caller must hold the object and backing_object and both must be
1923 * (only called from vm_object_collapse)
1926 vm_object_qcollapse(vm_object_t object, vm_object_t backing_object)
1928 if (backing_object->ref_count == 1) {
1929 backing_object->ref_count += 2;
1930 vm_object_backing_scan(object, backing_object,
1931 OBSC_COLLAPSE_NOWAIT);
1932 backing_object->ref_count -= 2;
1937 * Collapse an object with the object backing it. Pages in the backing
1938 * object are moved into the parent, and the backing object is deallocated.
1939 * Any conflict is resolved in favor of the parent's existing pages.
1941 * object must be held and chain-locked on call.
1943 * The caller must have an extra ref on object to prevent a race from
1944 * destroying it during the collapse.
1947 vm_object_collapse(vm_object_t object, struct vm_object_dealloc_list **dlistp)
1949 struct vm_object_dealloc_list *dlist = NULL;
1950 vm_object_t backing_object;
1953 * Only one thread is attempting a collapse at any given moment.
1954 * There are few restrictions for (object) that callers of this
1955 * function check so reentrancy is likely.
1957 KKASSERT(object != NULL);
1958 vm_object_assert_held(object);
1959 KKASSERT(object->flags & OBJ_CHAINLOCK);
1966 * We have to hold the backing object, check races.
1968 while ((backing_object = object->backing_object) != NULL) {
1969 vm_object_hold(backing_object);
1970 if (backing_object == object->backing_object)
1972 vm_object_drop(backing_object);
1976 * No backing object? Nothing to collapse then.
1978 if (backing_object == NULL)
1982 * You can't collapse with a non-default/non-swap object.
1984 if (backing_object->type != OBJT_DEFAULT &&
1985 backing_object->type != OBJT_SWAP) {
1986 vm_object_drop(backing_object);
1987 backing_object = NULL;
1992 * Chain-lock the backing object too because if we
1993 * successfully merge its pages into the top object we
1994 * will collapse backing_object->backing_object as the
1995 * new backing_object. Re-check that it is still our
1998 vm_object_chain_acquire(backing_object);
1999 if (backing_object != object->backing_object) {
2000 vm_object_chain_release(backing_object);
2001 vm_object_drop(backing_object);
2006 * we check the backing object first, because it is most likely
2009 if (backing_object->handle != NULL ||
2010 (backing_object->type != OBJT_DEFAULT &&
2011 backing_object->type != OBJT_SWAP) ||
2012 (backing_object->flags & OBJ_DEAD) ||
2013 object->handle != NULL ||
2014 (object->type != OBJT_DEFAULT &&
2015 object->type != OBJT_SWAP) ||
2016 (object->flags & OBJ_DEAD)) {
2021 * If paging is in progress we can't do a normal collapse.
2024 object->paging_in_progress != 0 ||
2025 backing_object->paging_in_progress != 0
2027 vm_object_qcollapse(object, backing_object);
2032 * We know that we can either collapse the backing object (if
2033 * the parent is the only reference to it) or (perhaps) have
2034 * the parent bypass the object if the parent happens to shadow
2035 * all the resident pages in the entire backing object.
2037 * This is ignoring pager-backed pages such as swap pages.
2038 * vm_object_backing_scan fails the shadowing test in this
2041 if (backing_object->ref_count == 1) {
2043 * If there is exactly one reference to the backing
2044 * object, we can collapse it into the parent.
2046 KKASSERT(object->backing_object == backing_object);
2047 vm_object_backing_scan(object, backing_object,
2048 OBSC_COLLAPSE_WAIT);
2051 * Move the pager from backing_object to object.
2053 if (backing_object->type == OBJT_SWAP) {
2054 vm_object_pip_add(backing_object, 1);
2057 * scrap the paging_offset junk and do a
2058 * discrete copy. This also removes major
2059 * assumptions about how the swap-pager
2060 * works from where it doesn't belong. The
2061 * new swapper is able to optimize the
2062 * destroy-source case.
2064 vm_object_pip_add(object, 1);
2065 swap_pager_copy(backing_object, object,
2066 OFF_TO_IDX(object->backing_object_offset),
2068 vm_object_pip_wakeup(object);
2069 vm_object_pip_wakeup(backing_object);
2073 * Object now shadows whatever backing_object did.
2074 * Remove object from backing_object's shadow_list.
2076 LIST_REMOVE(object, shadow_list);
2077 KKASSERT(object->backing_object == backing_object);
2078 backing_object->shadow_count--;
2079 backing_object->generation++;
2082 * backing_object->backing_object moves from within
2083 * backing_object to within object.
2085 while ((bbobj = backing_object->backing_object) != NULL) {
2086 vm_object_hold(bbobj);
2087 if (bbobj == backing_object->backing_object)
2089 vm_object_drop(bbobj);
2092 LIST_REMOVE(backing_object, shadow_list);
2093 bbobj->shadow_count--;
2094 bbobj->generation++;
2095 backing_object->backing_object = NULL;
2097 object->backing_object = bbobj;
2099 LIST_INSERT_HEAD(&bbobj->shadow_head,
2100 object, shadow_list);
2101 bbobj->shadow_count++;
2102 bbobj->generation++;
2105 object->backing_object_offset +=
2106 backing_object->backing_object_offset;
2108 vm_object_drop(bbobj);
2111 * Discard the old backing_object. Nothing should be
2112 * able to ref it, other than a vm_map_split(),
2113 * and vm_map_split() will stall on our chain lock.
2114 * And we control the parent so it shouldn't be
2115 * possible for it to go away either.
2117 * Since the backing object has no pages, no pager
2118 * left, and no object references within it, all
2119 * that is necessary is to dispose of it.
2121 KASSERT(backing_object->ref_count == 1,
2122 ("backing_object %p was somehow "
2123 "re-referenced during collapse!",
2125 KASSERT(RB_EMPTY(&backing_object->rb_memq),
2126 ("backing_object %p somehow has left "
2127 "over pages during collapse!",
2131 * The object can be destroyed.
2133 * XXX just fall through and dodealloc instead
2134 * of forcing destruction?
2136 --backing_object->ref_count;
2137 if ((backing_object->flags & OBJ_DEAD) == 0)
2138 vm_object_terminate(backing_object);
2143 * If we do not entirely shadow the backing object,
2144 * there is nothing we can do so we give up.
2146 if (vm_object_backing_scan(object, backing_object,
2147 OBSC_TEST_ALL_SHADOWED) == 0) {
2152 * bbobj is backing_object->backing_object. Since
2153 * object completely shadows backing_object we can
2154 * bypass it and become backed by bbobj instead.
2156 while ((bbobj = backing_object->backing_object) != NULL) {
2157 vm_object_hold(bbobj);
2158 if (bbobj == backing_object->backing_object)
2160 vm_object_drop(bbobj);
2164 * Make object shadow bbobj instead of backing_object.
2165 * Remove object from backing_object's shadow list.
2167 * Deallocating backing_object will not remove
2168 * it, since its reference count is at least 2.
2170 KKASSERT(object->backing_object == backing_object);
2171 LIST_REMOVE(object, shadow_list);
2172 backing_object->shadow_count--;
2173 backing_object->generation++;
2176 * Add a ref to bbobj, bbobj now shadows object.
2178 * NOTE: backing_object->backing_object still points
2179 * to bbobj. That relationship remains intact
2180 * because backing_object has > 1 ref, so
2181 * someone else is pointing to it (hence why
2182 * we can't collapse it into object and can
2183 * only handle the all-shadowed bypass case).
2186 vm_object_chain_wait(bbobj);
2187 vm_object_reference_locked(bbobj);
2188 LIST_INSERT_HEAD(&bbobj->shadow_head,
2189 object, shadow_list);
2190 bbobj->shadow_count++;
2191 bbobj->generation++;
2192 object->backing_object_offset +=
2193 backing_object->backing_object_offset;
2194 object->backing_object = bbobj;
2195 vm_object_drop(bbobj);
2197 object->backing_object = NULL;
2201 * Drop the reference count on backing_object. To
2202 * handle ref_count races properly we can't assume
2203 * that the ref_count is still at least 2 so we
2204 * have to actually call vm_object_deallocate()
2205 * (after clearing the chainlock).
2212 * Ok, we want to loop on the new object->bbobj association,
2213 * possibly collapsing it further. However if dodealloc is
2214 * non-zero we have to deallocate the backing_object which
2215 * itself can potentially undergo a collapse, creating a
2216 * recursion depth issue with the LWKT token subsystem.
2218 * In the case where we must deallocate the backing_object
2219 * it is possible now that the backing_object has a single
2220 * shadow count on some other object (not represented here
2221 * as yet), since it no longer shadows us. Thus when we
2222 * call vm_object_deallocate() it may attempt to collapse
2223 * itself into its remaining parent.
2226 struct vm_object_dealloc_list *dtmp;
2228 vm_object_chain_release(backing_object);
2229 vm_object_unlock(backing_object);
2230 /* backing_object remains held */
2233 * Auto-deallocation list for caller convenience.
2238 dtmp = kmalloc(sizeof(*dtmp), M_TEMP, M_WAITOK);
2239 dtmp->object = backing_object;
2240 dtmp->next = *dlistp;
2243 vm_object_chain_release(backing_object);
2244 vm_object_drop(backing_object);
2246 /* backing_object = NULL; not needed */
2251 * Clean up any left over backing_object
2253 if (backing_object) {
2254 vm_object_chain_release(backing_object);
2255 vm_object_drop(backing_object);
2259 * Clean up any auto-deallocation list. This is a convenience
2260 * for top-level callers so they don't have to pass &dlist.
2261 * Do not clean up any caller-passed dlistp, the caller will
2265 vm_object_deallocate_list(&dlist);
2270 * vm_object_collapse() may collect additional objects in need of
2271 * deallocation. This routine deallocates these objects. The
2272 * deallocation itself can trigger additional collapses (which the
2273 * deallocate function takes care of). This procedure is used to
2274 * reduce procedural recursion since these vm_object shadow chains
2275 * can become quite long.
2278 vm_object_deallocate_list(struct vm_object_dealloc_list **dlistp)
2280 struct vm_object_dealloc_list *dlist;
2282 while ((dlist = *dlistp) != NULL) {
2283 *dlistp = dlist->next;
2284 vm_object_lock(dlist->object);
2285 vm_object_deallocate_locked(dlist->object);
2286 vm_object_drop(dlist->object);
2287 kfree(dlist, M_TEMP);
2292 * Removes all physical pages in the specified object range from the
2293 * object's list of pages.
2297 static int vm_object_page_remove_callback(vm_page_t p, void *data);
2300 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2301 boolean_t clean_only)
2303 struct rb_vm_page_scan_info info;
2307 * Degenerate cases and assertions
2309 vm_object_hold(object);
2310 if (object == NULL ||
2311 (object->resident_page_count == 0 && object->swblock_count == 0)) {
2312 vm_object_drop(object);
2315 KASSERT(object->type != OBJT_PHYS,
2316 ("attempt to remove pages from a physical object"));
2319 * Indicate that paging is occuring on the object
2321 vm_object_pip_add(object, 1);
2324 * Figure out the actual removal range and whether we are removing
2325 * the entire contents of the object or not. If removing the entire
2326 * contents, be sure to get all pages, even those that might be
2327 * beyond the end of the object.
2329 info.start_pindex = start;
2331 info.end_pindex = (vm_pindex_t)-1;
2333 info.end_pindex = end - 1;
2334 info.limit = clean_only;
2335 all = (start == 0 && info.end_pindex >= object->size - 1);
2338 * Loop until we are sure we have gotten them all.
2342 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2343 vm_object_page_remove_callback, &info);
2344 } while (info.error);
2347 * Remove any related swap if throwing away pages, or for
2348 * non-swap objects (the swap is a clean copy in that case).
2350 if (object->type != OBJT_SWAP || clean_only == FALSE) {
2352 swap_pager_freespace_all(object);
2354 swap_pager_freespace(object, info.start_pindex,
2355 info.end_pindex - info.start_pindex + 1);
2361 vm_object_pip_wakeup(object);
2362 vm_object_drop(object);
2366 * The caller must hold the object
2369 vm_object_page_remove_callback(vm_page_t p, void *data)
2371 struct rb_vm_page_scan_info *info = data;
2373 if (vm_page_busy_try(p, TRUE)) {
2374 vm_page_sleep_busy(p, TRUE, "vmopar");
2380 * Wired pages cannot be destroyed, but they can be invalidated
2381 * and we do so if clean_only (limit) is not set.
2383 * WARNING! The page may be wired due to being part of a buffer
2384 * cache buffer, and the buffer might be marked B_CACHE.
2385 * This is fine as part of a truncation but VFSs must be
2386 * sure to fix the buffer up when re-extending the file.
2388 * NOTE! PG_NEED_COMMIT is ignored.
2390 if (p->wire_count != 0) {
2391 vm_page_protect(p, VM_PROT_NONE);
2392 if (info->limit == 0)
2399 * limit is our clean_only flag. If set and the page is dirty, do
2400 * not free it. If set and the page is being held by someone, do
2403 if (info->limit && p->valid) {
2404 vm_page_test_dirty(p);
2405 if (p->valid & p->dirty) {
2410 if (p->hold_count) {
2420 vm_page_protect(p, VM_PROT_NONE);
2426 * Coalesces two objects backing up adjoining regions of memory into a
2429 * returns TRUE if objects were combined.
2431 * NOTE: Only works at the moment if the second object is NULL -
2432 * if it's not, which object do we lock first?
2435 * prev_object First object to coalesce
2436 * prev_offset Offset into prev_object
2437 * next_object Second object into coalesce
2438 * next_offset Offset into next_object
2440 * prev_size Size of reference to prev_object
2441 * next_size Size of reference to next_object
2443 * The caller does not need to hold (prev_object) but must have a stable
2444 * pointer to it (typically by holding the vm_map locked).
2447 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
2448 vm_size_t prev_size, vm_size_t next_size)
2450 vm_pindex_t next_pindex;
2452 if (prev_object == NULL)
2455 vm_object_hold(prev_object);
2457 if (prev_object->type != OBJT_DEFAULT &&
2458 prev_object->type != OBJT_SWAP) {
2459 vm_object_drop(prev_object);
2464 * Try to collapse the object first
2466 vm_object_chain_acquire(prev_object);
2467 vm_object_collapse(prev_object, NULL);
2470 * Can't coalesce if: . more than one reference . paged out . shadows
2471 * another object . has a copy elsewhere (any of which mean that the
2472 * pages not mapped to prev_entry may be in use anyway)
2475 if (prev_object->backing_object != NULL) {
2476 vm_object_chain_release(prev_object);
2477 vm_object_drop(prev_object);
2481 prev_size >>= PAGE_SHIFT;
2482 next_size >>= PAGE_SHIFT;
2483 next_pindex = prev_pindex + prev_size;
2485 if ((prev_object->ref_count > 1) &&
2486 (prev_object->size != next_pindex)) {
2487 vm_object_chain_release(prev_object);
2488 vm_object_drop(prev_object);
2493 * Remove any pages that may still be in the object from a previous
2496 if (next_pindex < prev_object->size) {
2497 vm_object_page_remove(prev_object,
2499 next_pindex + next_size, FALSE);
2500 if (prev_object->type == OBJT_SWAP)
2501 swap_pager_freespace(prev_object,
2502 next_pindex, next_size);
2506 * Extend the object if necessary.
2508 if (next_pindex + next_size > prev_object->size)
2509 prev_object->size = next_pindex + next_size;
2511 vm_object_chain_release(prev_object);
2512 vm_object_drop(prev_object);
2517 * Make the object writable and flag is being possibly dirty.
2519 * The caller must hold the object. XXX called from vm_page_dirty(),
2520 * There is currently no requirement to hold the object.
2523 vm_object_set_writeable_dirty(vm_object_t object)
2527 /*vm_object_assert_held(object);*/
2529 * Avoid contention in vm fault path by checking the state before
2530 * issuing an atomic op on it.
2532 if ((object->flags & (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) !=
2533 (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) {
2534 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
2536 if (object->type == OBJT_VNODE &&
2537 (vp = (struct vnode *)object->handle) != NULL) {
2538 if ((vp->v_flag & VOBJDIRTY) == 0) {
2539 vsetflags(vp, VOBJDIRTY);
2544 #include "opt_ddb.h"
2546 #include <sys/kernel.h>
2548 #include <sys/cons.h>
2550 #include <ddb/ddb.h>
2552 static int _vm_object_in_map (vm_map_t map, vm_object_t object,
2553 vm_map_entry_t entry);
2554 static int vm_object_in_map (vm_object_t object);
2557 * The caller must hold the object.
2560 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2563 vm_map_entry_t tmpe;
2564 vm_object_t obj, nobj;
2570 tmpe = map->header.next;
2571 entcount = map->nentries;
2572 while (entcount-- && (tmpe != &map->header)) {
2573 if( _vm_object_in_map(map, object, tmpe)) {
2580 switch(entry->maptype) {
2581 case VM_MAPTYPE_SUBMAP:
2582 tmpm = entry->object.sub_map;
2583 tmpe = tmpm->header.next;
2584 entcount = tmpm->nentries;
2585 while (entcount-- && tmpe != &tmpm->header) {
2586 if( _vm_object_in_map(tmpm, object, tmpe)) {
2592 case VM_MAPTYPE_NORMAL:
2593 case VM_MAPTYPE_VPAGETABLE:
2594 obj = entry->object.vm_object;
2596 if (obj == object) {
2597 if (obj != entry->object.vm_object)
2598 vm_object_drop(obj);
2601 while ((nobj = obj->backing_object) != NULL) {
2602 vm_object_hold(nobj);
2603 if (nobj == obj->backing_object)
2605 vm_object_drop(nobj);
2607 if (obj != entry->object.vm_object) {
2609 vm_object_lock_swap();
2610 vm_object_drop(obj);
2621 static int vm_object_in_map_callback(struct proc *p, void *data);
2623 struct vm_object_in_map_info {
2632 vm_object_in_map(vm_object_t object)
2634 struct vm_object_in_map_info info;
2637 info.object = object;
2639 allproc_scan(vm_object_in_map_callback, &info);
2642 if( _vm_object_in_map(&kernel_map, object, 0))
2644 if( _vm_object_in_map(&pager_map, object, 0))
2646 if( _vm_object_in_map(&buffer_map, object, 0))
2655 vm_object_in_map_callback(struct proc *p, void *data)
2657 struct vm_object_in_map_info *info = data;
2660 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) {
2668 DB_SHOW_COMMAND(vmochk, vm_object_check)
2673 * make sure that internal objs are in a map somewhere
2674 * and none have zero ref counts.
2676 for (object = TAILQ_FIRST(&vm_object_list);
2678 object = TAILQ_NEXT(object, object_list)) {
2679 if (object->type == OBJT_MARKER)
2681 if (object->handle == NULL &&
2682 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2683 if (object->ref_count == 0) {
2684 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2685 (long)object->size);
2687 if (!vm_object_in_map(object)) {
2689 "vmochk: internal obj is not in a map: "
2690 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2691 object->ref_count, (u_long)object->size,
2692 (u_long)object->size,
2693 (void *)object->backing_object);
2702 DB_SHOW_COMMAND(object, vm_object_print_static)
2704 /* XXX convert args. */
2705 vm_object_t object = (vm_object_t)addr;
2706 boolean_t full = have_addr;
2710 /* XXX count is an (unused) arg. Avoid shadowing it. */
2711 #define count was_count
2719 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
2720 object, (int)object->type, (u_long)object->size,
2721 object->resident_page_count, object->ref_count, object->flags);
2723 * XXX no %qd in kernel. Truncate object->backing_object_offset.
2725 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
2726 object->shadow_count,
2727 object->backing_object ? object->backing_object->ref_count : 0,
2728 object->backing_object, (long)object->backing_object_offset);
2735 RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) {
2737 db_iprintf("memory:=");
2738 else if (count == 6) {
2746 db_printf("(off=0x%lx,page=0x%lx)",
2747 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
2758 * XXX need this non-static entry for calling from vm_map_print.
2763 vm_object_print(/* db_expr_t */ long addr,
2764 boolean_t have_addr,
2765 /* db_expr_t */ long count,
2768 vm_object_print_static(addr, have_addr, count, modif);
2774 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2779 for (object = TAILQ_FIRST(&vm_object_list);
2781 object = TAILQ_NEXT(object, object_list)) {
2782 vm_pindex_t idx, fidx;
2784 vm_paddr_t pa = -1, padiff;
2788 if (object->type == OBJT_MARKER)
2790 db_printf("new object: %p\n", (void *)object);
2800 osize = object->size;
2803 for (idx = 0; idx < osize; idx++) {
2804 m = vm_page_lookup(object, idx);
2807 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2808 (long)fidx, rcount, (long)pa);
2823 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2828 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
2829 padiff >>= PAGE_SHIFT;
2830 padiff &= PQ_L2_MASK;
2832 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
2836 db_printf(" index(%ld)run(%d)pa(0x%lx)",
2837 (long)fidx, rcount, (long)pa);
2838 db_printf("pd(%ld)\n", (long)padiff);
2848 pa = VM_PAGE_TO_PHYS(m);
2852 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2853 (long)fidx, rcount, (long)pa);