2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. All advertising materials mentioning features or use of this software
17 * must display the following acknowledgement:
18 * This product includes software developed by the University of
19 * California, Berkeley and its contributors.
20 * 4. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 * Carnegie Mellon requests users of this software to return to
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
68 * Virtual memory object module.
71 #include <sys/param.h>
72 #include <sys/systm.h>
75 #include <sys/mount.h>
76 #include <sys/mutex.h>
77 #include <sys/proc.h> /* for curproc, pageproc */
78 #include <sys/socket.h>
79 #include <sys/vnode.h>
80 #include <sys/vmmeter.h>
84 #include <vm/vm_param.h>
86 #include <vm/vm_map.h>
87 #include <vm/vm_object.h>
88 #include <vm/vm_page.h>
89 #include <vm/vm_pageout.h>
90 #include <vm/vm_pager.h>
91 #include <vm/vm_zone.h>
92 #include <vm/swap_pager.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
96 static void vm_object_qcollapse __P((vm_object_t object));
99 * Virtual memory objects maintain the actual data
100 * associated with allocated virtual memory. A given
101 * page of memory exists within exactly one object.
103 * An object is only deallocated when all "references"
104 * are given up. Only one "reference" to a given
105 * region of an object should be writeable.
107 * Associated with each object is a list of all resident
108 * memory pages belonging to that object; this list is
109 * maintained by the "vm_page" module, and locked by the object's
112 * Each object also records a "pager" routine which is
113 * used to retrieve (and store) pages to the proper backing
114 * storage. In addition, objects may be backed by other
115 * objects from which they were virtual-copied.
117 * The only items within the object structure which are
118 * modified after time of creation are:
119 * reference count locked by object's lock
120 * pager routine locked by object's lock
124 struct object_q vm_object_list;
125 static struct mtx vm_object_list_mtx; /* lock for object list and count */
126 static long vm_object_count; /* count of all objects */
127 vm_object_t kernel_object;
128 vm_object_t kmem_object;
129 static struct vm_object kernel_object_store;
130 static struct vm_object kmem_object_store;
131 extern int vm_pageout_page_count;
133 static long object_collapses;
134 static long object_bypasses;
135 static int next_index;
136 static vm_zone_t obj_zone;
137 static struct vm_zone obj_zone_store;
138 static int object_hash_rand;
139 #define VM_OBJECTS_INIT 256
140 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
143 _vm_object_allocate(objtype_t type, vm_size_t size, vm_object_t object)
149 TAILQ_INIT(&object->memq);
150 TAILQ_INIT(&object->shadow_head);
154 object->ref_count = 1;
156 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
157 vm_object_set_flag(object, OBJ_ONEMAPPING);
158 object->paging_in_progress = 0;
159 object->resident_page_count = 0;
160 object->shadow_count = 0;
161 object->pg_color = next_index;
162 if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
163 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
166 next_index = (next_index + incr) & PQ_L2_MASK;
167 object->handle = NULL;
168 object->backing_object = NULL;
169 object->backing_object_offset = (vm_ooffset_t) 0;
171 * Try to generate a number that will spread objects out in the
172 * hash table. We 'wipe' new objects across the hash in 128 page
173 * increments plus 1 more to offset it a little more by the time
176 object->hash_rand = object_hash_rand - 129;
178 object->generation++;
180 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
182 object_hash_rand = object->hash_rand;
188 * Initialize the VM objects module.
195 TAILQ_INIT(&vm_object_list);
196 mtx_init(&vm_object_list_mtx, "vm object_list", MTX_DEF);
199 kernel_object = &kernel_object_store;
200 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
203 kmem_object = &kmem_object_store;
204 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
207 obj_zone = &obj_zone_store;
208 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
209 vm_objects_init, VM_OBJECTS_INIT);
213 vm_object_init2(void)
215 zinitna(obj_zone, NULL, NULL, 0, 0, 0, 1);
219 vm_object_set_flag(vm_object_t object, u_short bits)
222 object->flags |= bits;
226 vm_object_clear_flag(vm_object_t object, u_short bits)
229 object->flags &= ~bits;
233 vm_object_pip_add(vm_object_t object, short i)
236 object->paging_in_progress += i;
240 vm_object_pip_subtract(vm_object_t object, short i)
243 object->paging_in_progress -= i;
247 vm_object_pip_wakeup(vm_object_t object)
250 object->paging_in_progress--;
251 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
252 vm_object_clear_flag(object, OBJ_PIPWNT);
258 vm_object_pip_wakeupn(vm_object_t object, short i)
262 object->paging_in_progress -= i;
263 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
264 vm_object_clear_flag(object, OBJ_PIPWNT);
270 vm_object_pip_sleep(vm_object_t object, char *waitid)
273 if (object->paging_in_progress) {
275 if (object->paging_in_progress) {
276 vm_object_set_flag(object, OBJ_PIPWNT);
277 tsleep(object, PVM, waitid, 0);
284 vm_object_pip_wait(vm_object_t object, char *waitid)
287 while (object->paging_in_progress)
288 vm_object_pip_sleep(object, waitid);
292 * vm_object_allocate:
294 * Returns a new object with the given size.
298 vm_object_allocate(objtype_t type, vm_size_t size)
304 result = (vm_object_t) zalloc(obj_zone);
305 _vm_object_allocate(type, size, result);
312 * vm_object_reference:
314 * Gets another reference to the given object.
317 vm_object_reference(vm_object_t object)
325 /* object can be re-referenced during final cleaning */
326 KASSERT(!(object->flags & OBJ_DEAD),
327 ("vm_object_reference: attempting to reference dead obj"));
331 if (object->type == OBJT_VNODE) {
332 while (vget((struct vnode *) object->handle, LK_RETRY|LK_NOOBJ, curthread)) {
333 printf("vm_object_reference: delay in getting object\n");
339 * handle deallocating a object of type OBJT_VNODE
342 vm_object_vndeallocate(vm_object_t object)
344 struct vnode *vp = (struct vnode *) object->handle;
347 KASSERT(object->type == OBJT_VNODE,
348 ("vm_object_vndeallocate: not a vnode object"));
349 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
351 if (object->ref_count == 0) {
352 vprint("vm_object_vndeallocate", vp);
353 panic("vm_object_vndeallocate: bad object reference count");
358 if (object->ref_count == 0) {
359 vp->v_flag &= ~VTEXT;
360 vm_object_clear_flag(object, OBJ_OPT);
363 * vrele may need a vop lock
369 * vm_object_deallocate:
371 * Release a reference to the specified object,
372 * gained either through a vm_object_allocate
373 * or a vm_object_reference call. When all references
374 * are gone, storage associated with this object
375 * may be relinquished.
377 * No object may be locked.
380 vm_object_deallocate(vm_object_t object)
386 while (object != NULL) {
388 if (object->type == OBJT_VNODE) {
389 vm_object_vndeallocate(object);
393 KASSERT(object->ref_count != 0,
394 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
397 * If the reference count goes to 0 we start calling
398 * vm_object_terminate() on the object chain.
399 * A ref count of 1 may be a special case depending on the
400 * shadow count being 0 or 1.
403 if (object->ref_count > 1) {
405 } else if (object->ref_count == 1) {
406 if (object->shadow_count == 0) {
407 vm_object_set_flag(object, OBJ_ONEMAPPING);
408 } else if ((object->shadow_count == 1) &&
409 (object->handle == NULL) &&
410 (object->type == OBJT_DEFAULT ||
411 object->type == OBJT_SWAP)) {
414 robject = TAILQ_FIRST(&object->shadow_head);
415 KASSERT(robject != NULL,
416 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
418 object->shadow_count));
419 if ((robject->handle == NULL) &&
420 (robject->type == OBJT_DEFAULT ||
421 robject->type == OBJT_SWAP)) {
423 robject->ref_count++;
426 robject->paging_in_progress ||
427 object->paging_in_progress
429 vm_object_pip_sleep(robject, "objde1");
430 vm_object_pip_sleep(object, "objde2");
433 if (robject->ref_count == 1) {
434 robject->ref_count--;
440 vm_object_collapse(object);
451 temp = object->backing_object;
453 TAILQ_REMOVE(&temp->shadow_head, object, shadow_list);
454 temp->shadow_count--;
455 if (temp->ref_count == 0)
456 vm_object_clear_flag(temp, OBJ_OPT);
458 object->backing_object = NULL;
461 * Don't double-terminate, we could be in a termination
462 * recursion due to the terminate having to sync data
465 if ((object->flags & OBJ_DEAD) == 0)
466 vm_object_terminate(object);
472 * vm_object_terminate actually destroys the specified object, freeing
473 * up all previously used resources.
475 * The object must be locked.
476 * This routine may block.
479 vm_object_terminate(vm_object_t object)
487 * Make sure no one uses us.
489 vm_object_set_flag(object, OBJ_DEAD);
492 * wait for the pageout daemon to be done with the object
494 vm_object_pip_wait(object, "objtrm");
496 KASSERT(!object->paging_in_progress,
497 ("vm_object_terminate: pageout in progress"));
500 * Clean and free the pages, as appropriate. All references to the
501 * object are gone, so we don't need to lock it.
503 if (object->type == OBJT_VNODE) {
507 * Freeze optimized copies.
509 vm_freeze_copyopts(object, 0, object->size);
512 * Clean pages and flush buffers.
514 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
516 vp = (struct vnode *) object->handle;
517 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0);
520 KASSERT(object->ref_count == 0,
521 ("vm_object_terminate: object with references, ref_count=%d",
525 * Now free any remaining pages. For internal objects, this also
526 * removes them from paging queues. Don't free wired pages, just
527 * remove them from the object.
530 while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
531 KASSERT(!p->busy && (p->flags & PG_BUSY) == 0,
532 ("vm_object_terminate: freeing busy page %p "
533 "p->busy = %d, p->flags %x\n", p, p->busy, p->flags));
534 if (p->wire_count == 0) {
546 * Let the pager know object is dead.
548 vm_pager_deallocate(object);
551 * Remove the object from the global object list.
553 mtx_lock(&vm_object_list_mtx);
554 TAILQ_REMOVE(&vm_object_list, object, object_list);
555 mtx_unlock(&vm_object_list_mtx);
560 * Free the space for the object.
562 zfree(obj_zone, object);
566 * vm_object_page_clean
568 * Clean all dirty pages in the specified range of object. Leaves page
569 * on whatever queue it is currently on. If NOSYNC is set then do not
570 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
571 * leaving the object dirty.
573 * Odd semantics: if start == end, we clean everything.
575 * The object must be locked.
579 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags)
582 vm_offset_t tstart, tend;
593 vm_page_t maf[vm_pageout_page_count];
594 vm_page_t mab[vm_pageout_page_count];
595 vm_page_t ma[vm_pageout_page_count];
600 if (object->type != OBJT_VNODE ||
601 (object->flags & OBJ_MIGHTBEDIRTY) == 0)
604 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : 0;
605 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
609 vm_object_set_flag(object, OBJ_CLEANING);
619 * Generally set CLEANCHK interlock and make the page read-only so
620 * we can then clear the object flags.
622 * However, if this is a nosync mmap then the object is likely to
623 * stay dirty so do not mess with the page and do not clear the
629 TAILQ_FOREACH(p, &object->memq, listq) {
630 vm_page_flag_set(p, PG_CLEANCHK);
631 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
634 vm_page_protect(p, VM_PROT_READ);
637 if (clearobjflags && (tstart == 0) && (tend == object->size)) {
640 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
641 if (object->type == OBJT_VNODE &&
642 (vp = (struct vnode *)object->handle) != NULL) {
643 if (vp->v_flag & VOBJDIRTY) {
644 mtx_lock(&vp->v_interlock);
645 vp->v_flag &= ~VOBJDIRTY;
646 mtx_unlock(&vp->v_interlock);
652 curgeneration = object->generation;
654 for (p = TAILQ_FIRST(&object->memq); p; p = np) {
655 np = TAILQ_NEXT(p, listq);
658 if (((p->flags & PG_CLEANCHK) == 0) ||
659 (pi < tstart) || (pi >= tend) ||
661 ((p->queue - p->pc) == PQ_CACHE)) {
662 vm_page_flag_clear(p, PG_CLEANCHK);
666 vm_page_test_dirty(p);
667 if ((p->dirty & p->valid) == 0) {
668 vm_page_flag_clear(p, PG_CLEANCHK);
673 * If we have been asked to skip nosync pages and this is a
674 * nosync page, skip it. Note that the object flags were
675 * not cleared in this case so we do not have to set them.
677 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
678 vm_page_flag_clear(p, PG_CLEANCHK);
683 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
684 if (object->generation != curgeneration) {
691 for (i = 1; i < vm_pageout_page_count; i++) {
692 if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
693 if ((tp->flags & PG_BUSY) ||
694 (tp->flags & PG_CLEANCHK) == 0 ||
697 if((tp->queue - tp->pc) == PQ_CACHE) {
698 vm_page_flag_clear(tp, PG_CLEANCHK);
701 vm_page_test_dirty(tp);
702 if ((tp->dirty & tp->valid) == 0) {
703 vm_page_flag_clear(tp, PG_CLEANCHK);
714 chkb = vm_pageout_page_count - maxf;
716 for (i = 1; i < chkb; i++) {
717 if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
718 if ((tp->flags & PG_BUSY) ||
719 (tp->flags & PG_CLEANCHK) == 0 ||
722 if((tp->queue - tp->pc) == PQ_CACHE) {
723 vm_page_flag_clear(tp, PG_CLEANCHK);
726 vm_page_test_dirty(tp);
727 if ((tp->dirty & tp->valid) == 0) {
728 vm_page_flag_clear(tp, PG_CLEANCHK);
739 for (i = 0; i < maxb; i++) {
740 int index = (maxb - i) - 1;
742 vm_page_flag_clear(ma[index], PG_CLEANCHK);
744 vm_page_flag_clear(p, PG_CLEANCHK);
746 for (i = 0 ; i < maxf; i++) {
747 int index = (maxb + i) + 1;
749 vm_page_flag_clear(ma[index], PG_CLEANCHK);
751 runlen = maxb + maxf + 1;
754 vm_pageout_flush(ma, runlen, pagerflags);
755 for (i = 0; i < runlen; i++) {
756 if (ma[i]->valid & ma[i]->dirty) {
757 vm_page_protect(ma[i], VM_PROT_READ);
758 vm_page_flag_set(ma[i], PG_CLEANCHK);
761 if (object->generation != curgeneration)
766 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
769 vm_object_clear_flag(object, OBJ_CLEANING);
774 * Same as vm_object_pmap_copy, except range checking really
775 * works, and is meant for small sections of an object.
777 * This code protects resident pages by making them read-only
778 * and is typically called on a fork or split when a page
779 * is converted to copy-on-write.
781 * NOTE: If the page is already at VM_PROT_NONE, calling
782 * vm_page_protect will have no effect.
786 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
793 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
796 for (idx = start; idx < end; idx++) {
797 p = vm_page_lookup(object, idx);
800 vm_page_protect(p, VM_PROT_READ);
805 * vm_object_pmap_remove:
807 * Removes all physical pages in the specified
808 * object range from all physical maps.
810 * The object must *not* be locked.
813 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
820 TAILQ_FOREACH(p, &object->memq, listq) {
821 if (p->pindex >= start && p->pindex < end)
822 vm_page_protect(p, VM_PROT_NONE);
824 if ((start == 0) && (object->size == end))
825 vm_object_clear_flag(object, OBJ_WRITEABLE);
831 * Implements the madvise function at the object/page level.
833 * MADV_WILLNEED (any object)
835 * Activate the specified pages if they are resident.
837 * MADV_DONTNEED (any object)
839 * Deactivate the specified pages if they are resident.
841 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
842 * OBJ_ONEMAPPING only)
844 * Deactivate and clean the specified pages if they are
845 * resident. This permits the process to reuse the pages
846 * without faulting or the kernel to reclaim the pages
850 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
852 vm_pindex_t end, tpindex;
860 end = pindex + count;
863 * Locate and adjust resident pages
866 for (; pindex < end; pindex += 1) {
872 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
873 * and those pages must be OBJ_ONEMAPPING.
875 if (advise == MADV_FREE) {
876 if ((tobject->type != OBJT_DEFAULT &&
877 tobject->type != OBJT_SWAP) ||
878 (tobject->flags & OBJ_ONEMAPPING) == 0) {
883 m = vm_page_lookup(tobject, tpindex);
887 * There may be swap even if there is no backing page
889 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
890 swap_pager_freespace(tobject, tpindex, 1);
895 tobject = tobject->backing_object;
898 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
903 * If the page is busy or not in a normal active state,
904 * we skip it. If the page is not managed there are no
905 * page queues to mess with. Things can break if we mess
906 * with pages in any of the below states.
911 (m->flags & PG_UNMANAGED) ||
912 m->valid != VM_PAGE_BITS_ALL
917 if (vm_page_sleep_busy(m, TRUE, "madvpo"))
920 if (advise == MADV_WILLNEED) {
922 } else if (advise == MADV_DONTNEED) {
924 } else if (advise == MADV_FREE) {
926 * Mark the page clean. This will allow the page
927 * to be freed up by the system. However, such pages
928 * are often reused quickly by malloc()/free()
929 * so we do not do anything that would cause
930 * a page fault if we can help it.
932 * Specifically, we do not try to actually free
933 * the page now nor do we try to put it in the
934 * cache (which would cause a page fault on reuse).
936 * But we do make the page is freeable as we
937 * can without actually taking the step of unmapping
940 pmap_clear_modify(m);
944 if (tobject->type == OBJT_SWAP)
945 swap_pager_freespace(tobject, tpindex, 1);
953 * Create a new object which is backed by the
954 * specified existing object range. The source
955 * object reference is deallocated.
957 * The new object and offset into that object
958 * are returned in the source parameters.
963 vm_object_t *object, /* IN/OUT */
964 vm_ooffset_t *offset, /* IN/OUT */
974 * Don't create the new object if the old object isn't shared.
977 if (source != NULL &&
978 source->ref_count == 1 &&
979 source->handle == NULL &&
980 (source->type == OBJT_DEFAULT ||
981 source->type == OBJT_SWAP))
985 * Allocate a new object with the given length
987 result = vm_object_allocate(OBJT_DEFAULT, length);
988 KASSERT(result != NULL, ("vm_object_shadow: no object for shadowing"));
991 * The new object shadows the source object, adding a reference to it.
992 * Our caller changes his reference to point to the new object,
993 * removing a reference to the source object. Net result: no change
994 * of reference count.
996 * Try to optimize the result object's page color when shadowing
997 * in order to maintain page coloring consistency in the combined
1000 result->backing_object = source;
1002 TAILQ_INSERT_TAIL(&source->shadow_head, result, shadow_list);
1003 source->shadow_count++;
1004 source->generation++;
1005 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1009 * Store the offset into the source object, and fix up the offset into
1013 result->backing_object_offset = *offset;
1016 * Return the new things
1023 #define OBSC_TEST_ALL_SHADOWED 0x0001
1024 #define OBSC_COLLAPSE_NOWAIT 0x0002
1025 #define OBSC_COLLAPSE_WAIT 0x0004
1028 vm_object_backing_scan(vm_object_t object, int op)
1033 vm_object_t backing_object;
1034 vm_pindex_t backing_offset_index;
1039 backing_object = object->backing_object;
1040 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1043 * Initial conditions
1046 if (op & OBSC_TEST_ALL_SHADOWED) {
1048 * We do not want to have to test for the existence of
1049 * swap pages in the backing object. XXX but with the
1050 * new swapper this would be pretty easy to do.
1052 * XXX what about anonymous MAP_SHARED memory that hasn't
1053 * been ZFOD faulted yet? If we do not test for this, the
1054 * shadow test may succeed! XXX
1056 if (backing_object->type != OBJT_DEFAULT) {
1061 if (op & OBSC_COLLAPSE_WAIT) {
1062 vm_object_set_flag(backing_object, OBJ_DEAD);
1069 p = TAILQ_FIRST(&backing_object->memq);
1071 vm_page_t next = TAILQ_NEXT(p, listq);
1072 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1074 if (op & OBSC_TEST_ALL_SHADOWED) {
1078 * Ignore pages outside the parent object's range
1079 * and outside the parent object's mapping of the
1082 * note that we do not busy the backing object's
1087 p->pindex < backing_offset_index ||
1088 new_pindex >= object->size
1095 * See if the parent has the page or if the parent's
1096 * object pager has the page. If the parent has the
1097 * page but the page is not valid, the parent's
1098 * object pager must have the page.
1100 * If this fails, the parent does not completely shadow
1101 * the object and we might as well give up now.
1104 pp = vm_page_lookup(object, new_pindex);
1106 (pp == NULL || pp->valid == 0) &&
1107 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1115 * Check for busy page
1118 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1121 if (op & OBSC_COLLAPSE_NOWAIT) {
1123 (p->flags & PG_BUSY) ||
1132 } else if (op & OBSC_COLLAPSE_WAIT) {
1133 if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1135 * If we slept, anything could have
1136 * happened. Since the object is
1137 * marked dead, the backing offset
1138 * should not have changed so we
1139 * just restart our scan.
1141 p = TAILQ_FIRST(&backing_object->memq);
1152 p->object == backing_object,
1153 ("vm_object_qcollapse(): object mismatch")
1157 * Destroy any associated swap
1159 if (backing_object->type == OBJT_SWAP) {
1160 swap_pager_freespace(
1168 p->pindex < backing_offset_index ||
1169 new_pindex >= object->size
1172 * Page is out of the parent object's range, we
1173 * can simply destroy it.
1175 vm_page_protect(p, VM_PROT_NONE);
1181 pp = vm_page_lookup(object, new_pindex);
1184 vm_pager_has_page(object, new_pindex, NULL, NULL)
1187 * page already exists in parent OR swap exists
1188 * for this location in the parent. Destroy
1189 * the original page from the backing object.
1191 * Leave the parent's page alone
1193 vm_page_protect(p, VM_PROT_NONE);
1200 * Page does not exist in parent, rename the
1201 * page from the backing object to the main object.
1203 * If the page was mapped to a process, it can remain
1204 * mapped through the rename.
1206 if ((p->queue - p->pc) == PQ_CACHE)
1207 vm_page_deactivate(p);
1209 vm_page_rename(p, object, new_pindex);
1210 /* page automatically made dirty by rename */
1220 * this version of collapse allows the operation to occur earlier and
1221 * when paging_in_progress is true for an object... This is not a complete
1222 * operation, but should plug 99.9% of the rest of the leaks.
1225 vm_object_qcollapse(vm_object_t object)
1227 vm_object_t backing_object = object->backing_object;
1231 if (backing_object->ref_count != 1)
1234 backing_object->ref_count += 2;
1236 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1238 backing_object->ref_count -= 2;
1242 * vm_object_collapse:
1244 * Collapse an object with the object backing it.
1245 * Pages in the backing object are moved into the
1246 * parent, and the backing object is deallocated.
1249 vm_object_collapse(vm_object_t object)
1254 vm_object_t backing_object;
1257 * Verify that the conditions are right for collapse:
1259 * The object exists and the backing object exists.
1264 if ((backing_object = object->backing_object) == NULL)
1268 * we check the backing object first, because it is most likely
1271 if (backing_object->handle != NULL ||
1272 (backing_object->type != OBJT_DEFAULT &&
1273 backing_object->type != OBJT_SWAP) ||
1274 (backing_object->flags & OBJ_DEAD) ||
1275 object->handle != NULL ||
1276 (object->type != OBJT_DEFAULT &&
1277 object->type != OBJT_SWAP) ||
1278 (object->flags & OBJ_DEAD)) {
1283 object->paging_in_progress != 0 ||
1284 backing_object->paging_in_progress != 0
1286 vm_object_qcollapse(object);
1291 * We know that we can either collapse the backing object (if
1292 * the parent is the only reference to it) or (perhaps) have
1293 * the parent bypass the object if the parent happens to shadow
1294 * all the resident pages in the entire backing object.
1296 * This is ignoring pager-backed pages such as swap pages.
1297 * vm_object_backing_scan fails the shadowing test in this
1301 if (backing_object->ref_count == 1) {
1303 * If there is exactly one reference to the backing
1304 * object, we can collapse it into the parent.
1307 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1310 * Move the pager from backing_object to object.
1313 if (backing_object->type == OBJT_SWAP) {
1314 vm_object_pip_add(backing_object, 1);
1317 * scrap the paging_offset junk and do a
1318 * discrete copy. This also removes major
1319 * assumptions about how the swap-pager
1320 * works from where it doesn't belong. The
1321 * new swapper is able to optimize the
1322 * destroy-source case.
1325 vm_object_pip_add(object, 1);
1329 OFF_TO_IDX(object->backing_object_offset), TRUE);
1330 vm_object_pip_wakeup(object);
1332 vm_object_pip_wakeup(backing_object);
1335 * Object now shadows whatever backing_object did.
1336 * Note that the reference to
1337 * backing_object->backing_object moves from within
1338 * backing_object to within object.
1342 &object->backing_object->shadow_head,
1346 object->backing_object->shadow_count--;
1347 object->backing_object->generation++;
1348 if (backing_object->backing_object) {
1350 &backing_object->backing_object->shadow_head,
1354 backing_object->backing_object->shadow_count--;
1355 backing_object->backing_object->generation++;
1357 object->backing_object = backing_object->backing_object;
1358 if (object->backing_object) {
1360 &object->backing_object->shadow_head,
1364 object->backing_object->shadow_count++;
1365 object->backing_object->generation++;
1368 object->backing_object_offset +=
1369 backing_object->backing_object_offset;
1372 * Discard backing_object.
1374 * Since the backing object has no pages, no pager left,
1375 * and no object references within it, all that is
1376 * necessary is to dispose of it.
1378 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1379 KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object));
1388 zfree(obj_zone, backing_object);
1392 vm_object_t new_backing_object;
1395 * If we do not entirely shadow the backing object,
1396 * there is nothing we can do so we give up.
1399 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1404 * Make the parent shadow the next object in the
1405 * chain. Deallocating backing_object will not remove
1406 * it, since its reference count is at least 2.
1410 &backing_object->shadow_head,
1414 backing_object->shadow_count--;
1415 backing_object->generation++;
1417 new_backing_object = backing_object->backing_object;
1418 if ((object->backing_object = new_backing_object) != NULL) {
1419 vm_object_reference(new_backing_object);
1421 &new_backing_object->shadow_head,
1425 new_backing_object->shadow_count++;
1426 new_backing_object->generation++;
1427 object->backing_object_offset +=
1428 backing_object->backing_object_offset;
1432 * Drop the reference count on backing_object. Since
1433 * its ref_count was at least 2, it will not vanish;
1434 * so we don't need to call vm_object_deallocate, but
1437 vm_object_deallocate(backing_object);
1442 * Try again with this object's new backing object.
1448 * vm_object_page_remove: [internal]
1450 * Removes all physical pages in the specified
1451 * object range from the object's list of pages.
1453 * The object must be locked.
1456 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, boolean_t clean_only)
1464 if (object == NULL ||
1465 object->resident_page_count == 0)
1468 all = ((end == 0) && (start == 0));
1471 * Since physically-backed objects do not use managed pages, we can't
1472 * remove pages from the object (we must instead remove the page
1473 * references, and then destroy the object).
1475 KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object"));
1477 vm_object_pip_add(object, 1);
1480 if (all || size > object->resident_page_count / 4) {
1481 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
1482 next = TAILQ_NEXT(p, listq);
1483 if (all || ((start <= p->pindex) && (p->pindex < end))) {
1484 if (p->wire_count != 0) {
1485 vm_page_protect(p, VM_PROT_NONE);
1492 * The busy flags are only cleared at
1493 * interrupt -- minimize the spl transitions
1496 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1499 if (clean_only && p->valid) {
1500 vm_page_test_dirty(p);
1501 if (p->valid & p->dirty)
1506 vm_page_protect(p, VM_PROT_NONE);
1512 if ((p = vm_page_lookup(object, start)) != 0) {
1514 if (p->wire_count != 0) {
1515 vm_page_protect(p, VM_PROT_NONE);
1524 * The busy flags are only cleared at
1525 * interrupt -- minimize the spl transitions
1527 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1530 if (clean_only && p->valid) {
1531 vm_page_test_dirty(p);
1532 if (p->valid & p->dirty) {
1540 vm_page_protect(p, VM_PROT_NONE);
1547 vm_object_pip_wakeup(object);
1551 * Routine: vm_object_coalesce
1552 * Function: Coalesces two objects backing up adjoining
1553 * regions of memory into a single object.
1555 * returns TRUE if objects were combined.
1557 * NOTE: Only works at the moment if the second object is NULL -
1558 * if it's not, which object do we lock first?
1561 * prev_object First object to coalesce
1562 * prev_offset Offset into prev_object
1563 * next_object Second object into coalesce
1564 * next_offset Offset into next_object
1566 * prev_size Size of reference to prev_object
1567 * next_size Size of reference to next_object
1570 * The object must *not* be locked.
1573 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, vm_size_t prev_size, vm_size_t next_size)
1575 vm_pindex_t next_pindex;
1579 if (prev_object == NULL) {
1583 if (prev_object->type != OBJT_DEFAULT &&
1584 prev_object->type != OBJT_SWAP) {
1589 * Try to collapse the object first
1591 vm_object_collapse(prev_object);
1594 * Can't coalesce if: . more than one reference . paged out . shadows
1595 * another object . has a copy elsewhere (any of which mean that the
1596 * pages not mapped to prev_entry may be in use anyway)
1599 if (prev_object->backing_object != NULL) {
1603 prev_size >>= PAGE_SHIFT;
1604 next_size >>= PAGE_SHIFT;
1605 next_pindex = prev_pindex + prev_size;
1607 if ((prev_object->ref_count > 1) &&
1608 (prev_object->size != next_pindex)) {
1613 * Remove any pages that may still be in the object from a previous
1616 if (next_pindex < prev_object->size) {
1617 vm_object_page_remove(prev_object,
1619 next_pindex + next_size, FALSE);
1620 if (prev_object->type == OBJT_SWAP)
1621 swap_pager_freespace(prev_object,
1622 next_pindex, next_size);
1626 * Extend the object if necessary.
1628 if (next_pindex + next_size > prev_object->size)
1629 prev_object->size = next_pindex + next_size;
1635 vm_object_set_writeable_dirty(vm_object_t object)
1639 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1640 if (object->type == OBJT_VNODE &&
1641 (vp = (struct vnode *)object->handle) != NULL) {
1642 if ((vp->v_flag & VOBJDIRTY) == 0) {
1643 mtx_lock(&vp->v_interlock);
1644 vp->v_flag |= VOBJDIRTY;
1645 mtx_unlock(&vp->v_interlock);
1650 #include "opt_ddb.h"
1652 #include <sys/kernel.h>
1654 #include <sys/cons.h>
1656 #include <ddb/ddb.h>
1659 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
1662 vm_map_entry_t tmpe;
1670 tmpe = map->header.next;
1671 entcount = map->nentries;
1672 while (entcount-- && (tmpe != &map->header)) {
1673 if( _vm_object_in_map(map, object, tmpe)) {
1678 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
1679 tmpm = entry->object.sub_map;
1680 tmpe = tmpm->header.next;
1681 entcount = tmpm->nentries;
1682 while (entcount-- && tmpe != &tmpm->header) {
1683 if( _vm_object_in_map(tmpm, object, tmpe)) {
1688 } else if ((obj = entry->object.vm_object) != NULL) {
1689 for (; obj; obj = obj->backing_object)
1690 if( obj == object) {
1698 vm_object_in_map(vm_object_t object)
1702 /* sx_slock(&allproc_lock); */
1703 LIST_FOREACH(p, &allproc, p_list) {
1704 if( !p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1706 if( _vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
1707 /* sx_sunlock(&allproc_lock); */
1711 /* sx_sunlock(&allproc_lock); */
1712 if( _vm_object_in_map( kernel_map, object, 0))
1714 if( _vm_object_in_map( kmem_map, object, 0))
1716 if( _vm_object_in_map( pager_map, object, 0))
1718 if( _vm_object_in_map( buffer_map, object, 0))
1723 DB_SHOW_COMMAND(vmochk, vm_object_check)
1728 * make sure that internal objs are in a map somewhere
1729 * and none have zero ref counts.
1731 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1732 if (object->handle == NULL &&
1733 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1734 if (object->ref_count == 0) {
1735 db_printf("vmochk: internal obj has zero ref count: %ld\n",
1736 (long)object->size);
1738 if (!vm_object_in_map(object)) {
1740 "vmochk: internal obj is not in a map: "
1741 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1742 object->ref_count, (u_long)object->size,
1743 (u_long)object->size,
1744 (void *)object->backing_object);
1751 * vm_object_print: [ debug ]
1753 DB_SHOW_COMMAND(object, vm_object_print_static)
1755 /* XXX convert args. */
1756 vm_object_t object = (vm_object_t)addr;
1757 boolean_t full = have_addr;
1761 /* XXX count is an (unused) arg. Avoid shadowing it. */
1762 #define count was_count
1770 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1771 object, (int)object->type, (u_long)object->size,
1772 object->resident_page_count, object->ref_count, object->flags);
1774 * XXX no %qd in kernel. Truncate object->backing_object_offset.
1776 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1777 object->shadow_count,
1778 object->backing_object ? object->backing_object->ref_count : 0,
1779 object->backing_object, (long)object->backing_object_offset);
1786 TAILQ_FOREACH(p, &object->memq, listq) {
1788 db_iprintf("memory:=");
1789 else if (count == 6) {
1797 db_printf("(off=0x%lx,page=0x%lx)",
1798 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1808 /* XXX need this non-static entry for calling from vm_map_print. */
1811 /* db_expr_t */ long addr,
1812 boolean_t have_addr,
1813 /* db_expr_t */ long count,
1816 vm_object_print_static(addr, have_addr, count, modif);
1819 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1825 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1826 vm_pindex_t idx, fidx;
1828 vm_offset_t pa = -1, padiff;
1832 db_printf("new object: %p\n", (void *)object);
1842 osize = object->size;
1845 for (idx = 0; idx < osize; idx++) {
1846 m = vm_page_lookup(object, idx);
1849 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1850 (long)fidx, rcount, (long)pa);
1865 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1870 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1871 padiff >>= PAGE_SHIFT;
1872 padiff &= PQ_L2_MASK;
1874 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1878 db_printf(" index(%ld)run(%d)pa(0x%lx)",
1879 (long)fidx, rcount, (long)pa);
1880 db_printf("pd(%ld)\n", (long)padiff);
1890 pa = VM_PAGE_TO_PHYS(m);
1894 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1895 (long)fidx, rcount, (long)pa);