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.
64 * $FreeBSD: src/sys/vm/vm_object.c,v 1.171.2.8 2003/05/26 19:17:56 alc Exp $
65 * $DragonFly: src/sys/vm/vm_object.c,v 1.30 2007/03/20 00:55:10 dillon Exp $
69 * Virtual memory object module.
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/proc.h> /* for curproc, pageproc */
75 #include <sys/vnode.h>
76 #include <sys/vmmeter.h>
78 #include <sys/mount.h>
79 #include <sys/kernel.h>
80 #include <sys/sysctl.h>
83 #include <vm/vm_param.h>
85 #include <vm/vm_map.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_pager.h>
90 #include <vm/swap_pager.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_extern.h>
93 #include <vm/vm_zone.h>
95 #define EASY_SCAN_FACTOR 8
97 static void vm_object_qcollapse(vm_object_t object);
98 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
102 * Virtual memory objects maintain the actual data
103 * associated with allocated virtual memory. A given
104 * page of memory exists within exactly one object.
106 * An object is only deallocated when all "references"
107 * are given up. Only one "reference" to a given
108 * region of an object should be writeable.
110 * Associated with each object is a list of all resident
111 * memory pages belonging to that object; this list is
112 * maintained by the "vm_page" module, and locked by the object's
115 * Each object also records a "pager" routine which is
116 * used to retrieve (and store) pages to the proper backing
117 * storage. In addition, objects may be backed by other
118 * objects from which they were virtual-copied.
120 * The only items within the object structure which are
121 * modified after time of creation are:
122 * reference count locked by object's lock
123 * pager routine locked by object's lock
127 struct object_q vm_object_list;
128 struct vm_object kernel_object;
130 static long vm_object_count; /* count of all objects */
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)
146 RB_INIT(&object->rb_memq);
147 LIST_INIT(&object->shadow_head);
151 object->ref_count = 1;
153 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
154 vm_object_set_flag(object, OBJ_ONEMAPPING);
155 object->paging_in_progress = 0;
156 object->resident_page_count = 0;
157 object->shadow_count = 0;
158 object->pg_color = next_index;
159 if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
160 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
163 next_index = (next_index + incr) & PQ_L2_MASK;
164 object->handle = NULL;
165 object->backing_object = NULL;
166 object->backing_object_offset = (vm_ooffset_t) 0;
168 * Try to generate a number that will spread objects out in the
169 * hash table. We 'wipe' new objects across the hash in 128 page
170 * increments plus 1 more to offset it a little more by the time
173 object->hash_rand = object_hash_rand - 129;
175 object->generation++;
178 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
180 object_hash_rand = object->hash_rand;
187 * Initialize the VM objects module.
192 TAILQ_INIT(&vm_object_list);
194 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd),
197 obj_zone = &obj_zone_store;
198 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
199 vm_objects_init, VM_OBJECTS_INIT);
203 vm_object_init2(void)
205 zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1);
209 * vm_object_allocate:
211 * Returns a new object with the given size.
215 vm_object_allocate(objtype_t type, vm_size_t size)
219 result = (vm_object_t) zalloc(obj_zone);
221 _vm_object_allocate(type, size, result);
228 * vm_object_reference:
230 * Gets another reference to the given object.
233 vm_object_reference(vm_object_t object)
239 if (object->type == OBJT_VNODE) {
240 vref(object->handle);
241 /* XXX what if the vnode is being destroyed? */
246 vm_object_vndeallocate(vm_object_t object)
248 struct vnode *vp = (struct vnode *) object->handle;
250 KASSERT(object->type == OBJT_VNODE,
251 ("vm_object_vndeallocate: not a vnode object"));
252 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
254 if (object->ref_count == 0) {
255 vprint("vm_object_vndeallocate", vp);
256 panic("vm_object_vndeallocate: bad object reference count");
261 if (object->ref_count == 0)
262 vp->v_flag &= ~VTEXT;
267 * vm_object_deallocate:
269 * Release a reference to the specified object,
270 * gained either through a vm_object_allocate
271 * or a vm_object_reference call. When all references
272 * are gone, storage associated with this object
273 * may be relinquished.
275 * No object may be locked.
278 vm_object_deallocate(vm_object_t object)
282 while (object != NULL) {
283 if (object->type == OBJT_VNODE) {
284 vm_object_vndeallocate(object);
288 if (object->ref_count == 0) {
289 panic("vm_object_deallocate: object deallocated too many times: %d", object->type);
290 } else if (object->ref_count > 2) {
296 * Here on ref_count of one or two, which are special cases for
299 if ((object->ref_count == 2) && (object->shadow_count == 0)) {
300 vm_object_set_flag(object, OBJ_ONEMAPPING);
303 } else if ((object->ref_count == 2) && (object->shadow_count == 1)) {
305 if ((object->handle == NULL) &&
306 (object->type == OBJT_DEFAULT ||
307 object->type == OBJT_SWAP)) {
310 robject = LIST_FIRST(&object->shadow_head);
311 KASSERT(robject != NULL,
312 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
314 object->shadow_count));
315 if ((robject->handle == NULL) &&
316 (robject->type == OBJT_DEFAULT ||
317 robject->type == OBJT_SWAP)) {
319 robject->ref_count++;
322 robject->paging_in_progress ||
323 object->paging_in_progress
325 vm_object_pip_sleep(robject, "objde1");
326 vm_object_pip_sleep(object, "objde2");
329 if (robject->ref_count == 1) {
330 robject->ref_count--;
336 vm_object_collapse(object);
345 if (object->ref_count != 0)
351 temp = object->backing_object;
353 LIST_REMOVE(object, shadow_list);
354 temp->shadow_count--;
356 object->backing_object = NULL;
360 * Don't double-terminate, we could be in a termination
361 * recursion due to the terminate having to sync data
364 if ((object->flags & OBJ_DEAD) == 0)
365 vm_object_terminate(object);
371 * vm_object_terminate actually destroys the specified object, freeing
372 * up all previously used resources.
374 * The object must be locked.
375 * This routine may block.
377 static int vm_object_terminate_callback(vm_page_t p, void *data);
380 vm_object_terminate(vm_object_t object)
383 * Make sure no one uses us.
385 vm_object_set_flag(object, OBJ_DEAD);
388 * wait for the pageout daemon to be done with the object
390 vm_object_pip_wait(object, "objtrm");
392 KASSERT(!object->paging_in_progress,
393 ("vm_object_terminate: pageout in progress"));
396 * Clean and free the pages, as appropriate. All references to the
397 * object are gone, so we don't need to lock it.
399 if (object->type == OBJT_VNODE) {
403 * Clean pages and flush buffers.
405 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
407 vp = (struct vnode *) object->handle;
408 vinvalbuf(vp, V_SAVE, 0, 0);
412 * Wait for any I/O to complete, after which there had better not
413 * be any references left on the object.
415 vm_object_pip_wait(object, "objtrm");
417 if (object->ref_count != 0)
418 panic("vm_object_terminate: object with references, ref_count=%d", object->ref_count);
421 * Now free any remaining pages. For internal objects, this also
422 * removes them from paging queues. Don't free wired pages, just
423 * remove them from the object.
426 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
427 vm_object_terminate_callback, NULL);
431 * Let the pager know object is dead.
433 vm_pager_deallocate(object);
436 * Remove the object from the global object list.
439 TAILQ_REMOVE(&vm_object_list, object, object_list);
444 if (object->ref_count != 0)
445 panic("vm_object_terminate2: object with references, ref_count=%d", object->ref_count);
448 * Free the space for the object.
450 zfree(obj_zone, object);
454 vm_object_terminate_callback(vm_page_t p, void *data __unused)
456 if (p->busy || (p->flags & PG_BUSY))
457 panic("vm_object_terminate: freeing busy page %p", p);
458 if (p->wire_count == 0) {
461 mycpu->gd_cnt.v_pfree++;
463 if (p->queue != PQ_NONE)
464 kprintf("vm_object_terminate: Warning: Encountered wired page %p on queue %d\n", p, p->queue);
473 * vm_object_page_clean
475 * Clean all dirty pages in the specified range of object. Leaves page
476 * on whatever queue it is currently on. If NOSYNC is set then do not
477 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
478 * leaving the object dirty.
480 * When stuffing pages asynchronously, allow clustering. XXX we need a
481 * synchronous clustering mode implementation.
483 * Odd semantics: if start == end, we clean everything.
485 static int vm_object_page_clean_pass1(struct vm_page *p, void *data);
486 static int vm_object_page_clean_pass2(struct vm_page *p, void *data);
489 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
492 struct rb_vm_page_scan_info info;
498 if (object->type != OBJT_VNODE ||
499 (object->flags & OBJ_MIGHTBEDIRTY) == 0)
502 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ?
503 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
504 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
509 * Interlock other major object operations. This allows us to
510 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY.
513 vm_object_set_flag(object, OBJ_CLEANING);
516 * Handle 'entire object' case
518 info.start_pindex = start;
520 info.end_pindex = object->size - 1;
522 info.end_pindex = end - 1;
524 wholescan = (start == 0 && info.end_pindex == object->size - 1);
526 info.pagerflags = pagerflags;
527 info.object = object;
530 * If cleaning the entire object do a pass to mark the pages read-only.
531 * If everything worked out ok, clear OBJ_WRITEABLE and
536 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
537 vm_object_page_clean_pass1, &info);
538 if (info.error == 0) {
539 vm_object_clear_flag(object,
540 OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
541 if (object->type == OBJT_VNODE &&
542 (vp = (struct vnode *)object->handle) != NULL) {
543 if (vp->v_flag & VOBJDIRTY)
544 vclrflags(vp, VOBJDIRTY);
550 * Do a pass to clean all the dirty pages we find.
554 curgeneration = object->generation;
555 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
556 vm_object_page_clean_pass2, &info);
557 } while (info.error || curgeneration != object->generation);
559 vm_object_clear_flag(object, OBJ_CLEANING);
565 vm_object_page_clean_pass1(struct vm_page *p, void *data)
567 struct rb_vm_page_scan_info *info = data;
569 vm_page_flag_set(p, PG_CLEANCHK);
570 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
573 vm_page_protect(p, VM_PROT_READ);
579 vm_object_page_clean_pass2(struct vm_page *p, void *data)
581 struct rb_vm_page_scan_info *info = data;
585 * Do not mess with pages that were inserted after we started
588 if ((p->flags & PG_CLEANCHK) == 0)
592 * Before wasting time traversing the pmaps, check for trivial
593 * cases where the page cannot be dirty.
595 if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) {
596 KKASSERT((p->dirty & p->valid) == 0);
601 * Check whether the page is dirty or not. The page has been set
602 * to be read-only so the check will not race a user dirtying the
605 vm_page_test_dirty(p);
606 if ((p->dirty & p->valid) == 0) {
607 vm_page_flag_clear(p, PG_CLEANCHK);
612 * If we have been asked to skip nosync pages and this is a
613 * nosync page, skip it. Note that the object flags were
614 * not cleared in this case (because pass1 will have returned an
615 * error), so we do not have to set them.
617 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
618 vm_page_flag_clear(p, PG_CLEANCHK);
623 * Flush as many pages as we can. PG_CLEANCHK will be cleared on
624 * the pages that get successfully flushed. Set info->error if
625 * we raced an object modification.
627 n = vm_object_page_collect_flush(info->object, p, info->pagerflags);
634 * This routine must be called within a critical section to properly avoid
635 * an interrupt unbusy/free race that can occur prior to the busy check.
637 * Using the object generation number here to detect page ripout is not
638 * the best idea in the world. XXX
640 * NOTE: we operate under the assumption that a page found to not be busy
641 * will not be ripped out from under us by an interrupt. XXX we should
642 * recode this to explicitly busy the pages.
645 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags)
654 vm_page_t maf[vm_pageout_page_count];
655 vm_page_t mab[vm_pageout_page_count];
656 vm_page_t ma[vm_pageout_page_count];
658 curgeneration = object->generation;
661 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
662 if (object->generation != curgeneration) {
666 KKASSERT(p->object == object && p->pindex == pi);
669 for(i = 1; i < vm_pageout_page_count; i++) {
672 if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
673 if ((tp->flags & PG_BUSY) ||
674 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
675 (tp->flags & PG_CLEANCHK) == 0) ||
678 if((tp->queue - tp->pc) == PQ_CACHE) {
679 vm_page_flag_clear(tp, PG_CLEANCHK);
682 vm_page_test_dirty(tp);
683 if ((tp->dirty & tp->valid) == 0) {
684 vm_page_flag_clear(tp, PG_CLEANCHK);
695 chkb = vm_pageout_page_count - maxf;
697 for(i = 1; i < chkb;i++) {
700 if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
701 if ((tp->flags & PG_BUSY) ||
702 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
703 (tp->flags & PG_CLEANCHK) == 0) ||
706 if((tp->queue - tp->pc) == PQ_CACHE) {
707 vm_page_flag_clear(tp, PG_CLEANCHK);
710 vm_page_test_dirty(tp);
711 if ((tp->dirty & tp->valid) == 0) {
712 vm_page_flag_clear(tp, PG_CLEANCHK);
723 for(i = 0; i < maxb; i++) {
724 int index = (maxb - i) - 1;
726 vm_page_flag_clear(ma[index], PG_CLEANCHK);
728 vm_page_flag_clear(p, PG_CLEANCHK);
730 for(i = 0; i < maxf; i++) {
731 int index = (maxb + i) + 1;
733 vm_page_flag_clear(ma[index], PG_CLEANCHK);
735 runlen = maxb + maxf + 1;
737 vm_pageout_flush(ma, runlen, pagerflags);
738 for (i = 0; i < runlen; i++) {
739 if (ma[i]->valid & ma[i]->dirty) {
740 vm_page_protect(ma[i], VM_PROT_READ);
741 vm_page_flag_set(ma[i], PG_CLEANCHK);
744 * maxf will end up being the actual number of pages
745 * we wrote out contiguously, non-inclusive of the
746 * first page. We do not count look-behind pages.
748 if (i >= maxb + 1 && (maxf > i - maxb - 1))
756 /* XXX I cannot tell if this should be an exported symbol */
758 * vm_object_deactivate_pages
760 * Deactivate all pages in the specified object. (Keep its pages
761 * in memory even though it is no longer referenced.)
763 * The object must be locked.
765 static int vm_object_deactivate_pages_callback(vm_page_t p, void *data);
768 vm_object_deactivate_pages(vm_object_t object)
771 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
772 vm_object_deactivate_pages_callback, NULL);
777 vm_object_deactivate_pages_callback(vm_page_t p, void *data __unused)
779 vm_page_deactivate(p);
786 * Same as vm_object_pmap_copy, except range checking really
787 * works, and is meant for small sections of an object.
789 * This code protects resident pages by making them read-only
790 * and is typically called on a fork or split when a page
791 * is converted to copy-on-write.
793 * NOTE: If the page is already at VM_PROT_NONE, calling
794 * vm_page_protect will have no effect.
797 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
802 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
806 * spl protection needed to prevent races between the lookup,
807 * an interrupt unbusy/free, and our protect call.
810 for (idx = start; idx < end; idx++) {
811 p = vm_page_lookup(object, idx);
814 vm_page_protect(p, VM_PROT_READ);
820 * vm_object_pmap_remove:
822 * Removes all physical pages in the specified
823 * object range from all physical maps.
825 * The object must *not* be locked.
828 static int vm_object_pmap_remove_callback(vm_page_t p, void *data);
831 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
833 struct rb_vm_page_scan_info info;
837 info.start_pindex = start;
838 info.end_pindex = end - 1;
840 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
841 vm_object_pmap_remove_callback, &info);
842 if (start == 0 && end == object->size)
843 vm_object_clear_flag(object, OBJ_WRITEABLE);
848 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused)
850 vm_page_protect(p, VM_PROT_NONE);
857 * Implements the madvise function at the object/page level.
859 * MADV_WILLNEED (any object)
861 * Activate the specified pages if they are resident.
863 * MADV_DONTNEED (any object)
865 * Deactivate the specified pages if they are resident.
867 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
868 * OBJ_ONEMAPPING only)
870 * Deactivate and clean the specified pages if they are
871 * resident. This permits the process to reuse the pages
872 * without faulting or the kernel to reclaim the pages
876 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
878 vm_pindex_t end, tpindex;
885 end = pindex + count;
888 * Locate and adjust resident pages
891 for (; pindex < end; pindex += 1) {
897 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
898 * and those pages must be OBJ_ONEMAPPING.
900 if (advise == MADV_FREE) {
901 if ((tobject->type != OBJT_DEFAULT &&
902 tobject->type != OBJT_SWAP) ||
903 (tobject->flags & OBJ_ONEMAPPING) == 0) {
909 * spl protection is required to avoid a race between the
910 * lookup, an interrupt unbusy/free, and our busy check.
914 m = vm_page_lookup(tobject, tpindex);
918 * There may be swap even if there is no backing page
920 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
921 swap_pager_freespace(tobject, tpindex, 1);
927 if (tobject->backing_object == NULL)
929 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
930 tobject = tobject->backing_object;
935 * If the page is busy or not in a normal active state,
936 * we skip it. If the page is not managed there are no
937 * page queues to mess with. Things can break if we mess
938 * with pages in any of the below states.
943 (m->flags & PG_UNMANAGED) ||
944 m->valid != VM_PAGE_BITS_ALL
950 if (vm_page_sleep_busy(m, TRUE, "madvpo")) {
957 * Theoretically once a page is known not to be busy, an
958 * interrupt cannot come along and rip it out from under us.
961 if (advise == MADV_WILLNEED) {
963 } else if (advise == MADV_DONTNEED) {
965 } else if (advise == MADV_FREE) {
967 * Mark the page clean. This will allow the page
968 * to be freed up by the system. However, such pages
969 * are often reused quickly by malloc()/free()
970 * so we do not do anything that would cause
971 * a page fault if we can help it.
973 * Specifically, we do not try to actually free
974 * the page now nor do we try to put it in the
975 * cache (which would cause a page fault on reuse).
977 * But we do make the page is freeable as we
978 * can without actually taking the step of unmapping
981 pmap_clear_modify(m);
985 if (tobject->type == OBJT_SWAP)
986 swap_pager_freespace(tobject, tpindex, 1);
994 * Create a new object which is backed by the
995 * specified existing object range. The source
996 * object reference is deallocated.
998 * The new object and offset into that object
999 * are returned in the source parameters.
1003 vm_object_shadow(vm_object_t *object, /* IN/OUT */
1004 vm_ooffset_t *offset, /* IN/OUT */
1013 * Don't create the new object if the old object isn't shared.
1016 if (source != NULL &&
1017 source->ref_count == 1 &&
1018 source->handle == NULL &&
1019 (source->type == OBJT_DEFAULT ||
1020 source->type == OBJT_SWAP))
1024 * Allocate a new object with the given length
1027 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1028 panic("vm_object_shadow: no object for shadowing");
1031 * The new object shadows the source object, adding a reference to it.
1032 * Our caller changes his reference to point to the new object,
1033 * removing a reference to the source object. Net result: no change
1034 * of reference count.
1036 * Try to optimize the result object's page color when shadowing
1037 * in order to maintain page coloring consistency in the combined
1040 result->backing_object = source;
1042 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1043 source->shadow_count++;
1044 source->generation++;
1045 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1049 * Store the offset into the source object, and fix up the offset into
1053 result->backing_object_offset = *offset;
1056 * Return the new things
1063 #define OBSC_TEST_ALL_SHADOWED 0x0001
1064 #define OBSC_COLLAPSE_NOWAIT 0x0002
1065 #define OBSC_COLLAPSE_WAIT 0x0004
1067 static int vm_object_backing_scan_callback(vm_page_t p, void *data);
1070 vm_object_backing_scan(vm_object_t object, int op)
1072 struct rb_vm_page_scan_info info;
1073 vm_object_t backing_object;
1076 * spl protection is required to avoid races between the memq/lookup,
1077 * an interrupt doing an unbusy/free, and our busy check. Amoung
1082 backing_object = object->backing_object;
1083 info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1086 * Initial conditions
1089 if (op & OBSC_TEST_ALL_SHADOWED) {
1091 * We do not want to have to test for the existence of
1092 * swap pages in the backing object. XXX but with the
1093 * new swapper this would be pretty easy to do.
1095 * XXX what about anonymous MAP_SHARED memory that hasn't
1096 * been ZFOD faulted yet? If we do not test for this, the
1097 * shadow test may succeed! XXX
1099 if (backing_object->type != OBJT_DEFAULT) {
1104 if (op & OBSC_COLLAPSE_WAIT) {
1105 KKASSERT((backing_object->flags & OBJ_DEAD) == 0);
1106 vm_object_set_flag(backing_object, OBJ_DEAD);
1110 * Our scan. We have to retry if a negative error code is returned,
1111 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that
1112 * the scan had to be stopped because the parent does not completely
1115 info.object = object;
1116 info.backing_object = backing_object;
1120 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL,
1121 vm_object_backing_scan_callback,
1123 } while (info.error < 0);
1129 vm_object_backing_scan_callback(vm_page_t p, void *data)
1131 struct rb_vm_page_scan_info *info = data;
1132 vm_object_t backing_object;
1134 vm_pindex_t new_pindex;
1135 vm_pindex_t backing_offset_index;
1138 new_pindex = p->pindex - info->backing_offset_index;
1140 object = info->object;
1141 backing_object = info->backing_object;
1142 backing_offset_index = info->backing_offset_index;
1144 if (op & OBSC_TEST_ALL_SHADOWED) {
1148 * Ignore pages outside the parent object's range
1149 * and outside the parent object's mapping of the
1152 * note that we do not busy the backing object's
1156 p->pindex < backing_offset_index ||
1157 new_pindex >= object->size
1163 * See if the parent has the page or if the parent's
1164 * object pager has the page. If the parent has the
1165 * page but the page is not valid, the parent's
1166 * object pager must have the page.
1168 * If this fails, the parent does not completely shadow
1169 * the object and we might as well give up now.
1172 pp = vm_page_lookup(object, new_pindex);
1174 (pp == NULL || pp->valid == 0) &&
1175 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1177 info->error = 0; /* problemo */
1178 return(-1); /* stop the scan */
1183 * Check for busy page
1186 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1189 if (op & OBSC_COLLAPSE_NOWAIT) {
1191 (p->flags & PG_BUSY) ||
1199 } else if (op & OBSC_COLLAPSE_WAIT) {
1200 if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1202 * If we slept, anything could have
1203 * happened. Ask that the scan be restarted.
1205 * Since the object is marked dead, the
1206 * backing offset should not have changed.
1219 p->object == backing_object,
1220 ("vm_object_qcollapse(): object mismatch")
1224 * Destroy any associated swap
1226 if (backing_object->type == OBJT_SWAP) {
1227 swap_pager_freespace(
1235 p->pindex < backing_offset_index ||
1236 new_pindex >= object->size
1239 * Page is out of the parent object's range, we
1240 * can simply destroy it.
1242 vm_page_protect(p, VM_PROT_NONE);
1247 pp = vm_page_lookup(object, new_pindex);
1250 vm_pager_has_page(object, new_pindex, NULL, NULL)
1253 * page already exists in parent OR swap exists
1254 * for this location in the parent. Destroy
1255 * the original page from the backing object.
1257 * Leave the parent's page alone
1259 vm_page_protect(p, VM_PROT_NONE);
1265 * Page does not exist in parent, rename the
1266 * page from the backing object to the main object.
1268 * If the page was mapped to a process, it can remain
1269 * mapped through the rename.
1271 if ((p->queue - p->pc) == PQ_CACHE)
1272 vm_page_deactivate(p);
1274 vm_page_rename(p, object, new_pindex);
1275 /* page automatically made dirty by rename */
1281 * this version of collapse allows the operation to occur earlier and
1282 * when paging_in_progress is true for an object... This is not a complete
1283 * operation, but should plug 99.9% of the rest of the leaks.
1286 vm_object_qcollapse(vm_object_t object)
1288 vm_object_t backing_object = object->backing_object;
1290 if (backing_object->ref_count != 1)
1293 backing_object->ref_count += 2;
1295 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1297 backing_object->ref_count -= 2;
1301 * vm_object_collapse:
1303 * Collapse an object with the object backing it.
1304 * Pages in the backing object are moved into the
1305 * parent, and the backing object is deallocated.
1308 vm_object_collapse(vm_object_t object)
1311 vm_object_t backing_object;
1314 * Verify that the conditions are right for collapse:
1316 * The object exists and the backing object exists.
1321 if ((backing_object = object->backing_object) == NULL)
1325 * we check the backing object first, because it is most likely
1328 if (backing_object->handle != NULL ||
1329 (backing_object->type != OBJT_DEFAULT &&
1330 backing_object->type != OBJT_SWAP) ||
1331 (backing_object->flags & OBJ_DEAD) ||
1332 object->handle != NULL ||
1333 (object->type != OBJT_DEFAULT &&
1334 object->type != OBJT_SWAP) ||
1335 (object->flags & OBJ_DEAD)) {
1340 object->paging_in_progress != 0 ||
1341 backing_object->paging_in_progress != 0
1343 vm_object_qcollapse(object);
1348 * We know that we can either collapse the backing object (if
1349 * the parent is the only reference to it) or (perhaps) have
1350 * the parent bypass the object if the parent happens to shadow
1351 * all the resident pages in the entire backing object.
1353 * This is ignoring pager-backed pages such as swap pages.
1354 * vm_object_backing_scan fails the shadowing test in this
1358 if (backing_object->ref_count == 1) {
1360 * If there is exactly one reference to the backing
1361 * object, we can collapse it into the parent.
1363 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1366 * Move the pager from backing_object to object.
1369 if (backing_object->type == OBJT_SWAP) {
1370 vm_object_pip_add(backing_object, 1);
1373 * scrap the paging_offset junk and do a
1374 * discrete copy. This also removes major
1375 * assumptions about how the swap-pager
1376 * works from where it doesn't belong. The
1377 * new swapper is able to optimize the
1378 * destroy-source case.
1381 vm_object_pip_add(object, 1);
1385 OFF_TO_IDX(object->backing_object_offset), TRUE);
1386 vm_object_pip_wakeup(object);
1388 vm_object_pip_wakeup(backing_object);
1391 * Object now shadows whatever backing_object did.
1392 * Note that the reference to
1393 * backing_object->backing_object moves from within
1394 * backing_object to within object.
1397 LIST_REMOVE(object, shadow_list);
1398 object->backing_object->shadow_count--;
1399 object->backing_object->generation++;
1400 if (backing_object->backing_object) {
1401 LIST_REMOVE(backing_object, shadow_list);
1402 backing_object->backing_object->shadow_count--;
1403 backing_object->backing_object->generation++;
1405 object->backing_object = backing_object->backing_object;
1406 if (object->backing_object) {
1408 &object->backing_object->shadow_head,
1412 object->backing_object->shadow_count++;
1413 object->backing_object->generation++;
1416 object->backing_object_offset +=
1417 backing_object->backing_object_offset;
1420 * Discard backing_object.
1422 * Since the backing object has no pages, no pager left,
1423 * and no object references within it, all that is
1424 * necessary is to dispose of it.
1427 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1428 KASSERT(RB_EMPTY(&backing_object->rb_memq), ("backing_object %p somehow has left over pages during collapse!", backing_object));
1438 zfree(obj_zone, backing_object);
1442 vm_object_t new_backing_object;
1445 * If we do not entirely shadow the backing object,
1446 * there is nothing we can do so we give up.
1449 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1454 * Make the parent shadow the next object in the
1455 * chain. Deallocating backing_object will not remove
1456 * it, since its reference count is at least 2.
1459 LIST_REMOVE(object, shadow_list);
1460 backing_object->shadow_count--;
1461 backing_object->generation++;
1463 new_backing_object = backing_object->backing_object;
1464 if ((object->backing_object = new_backing_object) != NULL) {
1465 vm_object_reference(new_backing_object);
1467 &new_backing_object->shadow_head,
1471 new_backing_object->shadow_count++;
1472 new_backing_object->generation++;
1473 object->backing_object_offset +=
1474 backing_object->backing_object_offset;
1478 * Drop the reference count on backing_object. Since
1479 * its ref_count was at least 2, it will not vanish;
1480 * so we don't need to call vm_object_deallocate, but
1483 vm_object_deallocate(backing_object);
1488 * Try again with this object's new backing object.
1494 * vm_object_page_remove: [internal]
1496 * Removes all physical pages in the specified
1497 * object range from the object's list of pages.
1499 static int vm_object_page_remove_callback(vm_page_t p, void *data);
1502 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1503 boolean_t clean_only)
1505 struct rb_vm_page_scan_info info;
1509 * Degenerate cases and assertions
1511 if (object == NULL || object->resident_page_count == 0)
1513 KASSERT(object->type != OBJT_PHYS,
1514 ("attempt to remove pages from a physical object"));
1517 * Indicate that paging is occuring on the object
1520 vm_object_pip_add(object, 1);
1523 * Figure out the actual removal range and whether we are removing
1524 * the entire contents of the object or not. If removing the entire
1525 * contents, be sure to get all pages, even those that might be
1526 * beyond the end of the object.
1528 info.start_pindex = start;
1530 info.end_pindex = (vm_pindex_t)-1;
1532 info.end_pindex = end - 1;
1533 info.limit = clean_only;
1534 all = (start == 0 && info.end_pindex >= object->size - 1);
1537 * Loop until we are sure we have gotten them all.
1541 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1542 vm_object_page_remove_callback, &info);
1543 } while (info.error);
1548 vm_object_pip_wakeup(object);
1553 vm_object_page_remove_callback(vm_page_t p, void *data)
1555 struct rb_vm_page_scan_info *info = data;
1558 * Wired pages cannot be destroyed, but they can be invalidated
1559 * and we do so if clean_only (limit) is not set.
1561 if (p->wire_count != 0) {
1562 vm_page_protect(p, VM_PROT_NONE);
1563 if (info->limit == 0)
1569 * The busy flags are only cleared at
1570 * interrupt -- minimize the spl transitions
1573 if (vm_page_sleep_busy(p, TRUE, "vmopar")) {
1579 * limit is our clean_only flag. If set and the page is dirty, do
1582 if (info->limit && p->valid) {
1583 vm_page_test_dirty(p);
1584 if (p->valid & p->dirty)
1592 vm_page_protect(p, VM_PROT_NONE);
1598 * Routine: vm_object_coalesce
1599 * Function: Coalesces two objects backing up adjoining
1600 * regions of memory into a single object.
1602 * returns TRUE if objects were combined.
1604 * NOTE: Only works at the moment if the second object is NULL -
1605 * if it's not, which object do we lock first?
1608 * prev_object First object to coalesce
1609 * prev_offset Offset into prev_object
1610 * next_object Second object into coalesce
1611 * next_offset Offset into next_object
1613 * prev_size Size of reference to prev_object
1614 * next_size Size of reference to next_object
1617 * The object must *not* be locked.
1620 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
1621 vm_size_t prev_size, vm_size_t next_size)
1623 vm_pindex_t next_pindex;
1625 if (prev_object == NULL) {
1629 if (prev_object->type != OBJT_DEFAULT &&
1630 prev_object->type != OBJT_SWAP) {
1635 * Try to collapse the object first
1637 vm_object_collapse(prev_object);
1640 * Can't coalesce if: . more than one reference . paged out . shadows
1641 * another object . has a copy elsewhere (any of which mean that the
1642 * pages not mapped to prev_entry may be in use anyway)
1645 if (prev_object->backing_object != NULL) {
1649 prev_size >>= PAGE_SHIFT;
1650 next_size >>= PAGE_SHIFT;
1651 next_pindex = prev_pindex + prev_size;
1653 if ((prev_object->ref_count > 1) &&
1654 (prev_object->size != next_pindex)) {
1659 * Remove any pages that may still be in the object from a previous
1662 if (next_pindex < prev_object->size) {
1663 vm_object_page_remove(prev_object,
1665 next_pindex + next_size, FALSE);
1666 if (prev_object->type == OBJT_SWAP)
1667 swap_pager_freespace(prev_object,
1668 next_pindex, next_size);
1672 * Extend the object if necessary.
1674 if (next_pindex + next_size > prev_object->size)
1675 prev_object->size = next_pindex + next_size;
1681 vm_object_set_writeable_dirty(vm_object_t object)
1685 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1686 if (object->type == OBJT_VNODE &&
1687 (vp = (struct vnode *)object->handle) != NULL) {
1688 if ((vp->v_flag & VOBJDIRTY) == 0) {
1689 vsetflags(vp, VOBJDIRTY);
1696 #include "opt_ddb.h"
1698 #include <sys/kernel.h>
1700 #include <sys/cons.h>
1702 #include <ddb/ddb.h>
1704 static int _vm_object_in_map (vm_map_t map, vm_object_t object,
1705 vm_map_entry_t entry);
1706 static int vm_object_in_map (vm_object_t object);
1709 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
1712 vm_map_entry_t tmpe;
1719 tmpe = map->header.next;
1720 entcount = map->nentries;
1721 while (entcount-- && (tmpe != &map->header)) {
1722 if( _vm_object_in_map(map, object, tmpe)) {
1729 switch(entry->maptype) {
1730 case VM_MAPTYPE_SUBMAP:
1731 tmpm = entry->object.sub_map;
1732 tmpe = tmpm->header.next;
1733 entcount = tmpm->nentries;
1734 while (entcount-- && tmpe != &tmpm->header) {
1735 if( _vm_object_in_map(tmpm, object, tmpe)) {
1741 case VM_MAPTYPE_NORMAL:
1742 case VM_MAPTYPE_VPAGETABLE:
1743 obj = entry->object.vm_object;
1747 obj = obj->backing_object;
1756 static int vm_object_in_map_callback(struct proc *p, void *data);
1758 struct vm_object_in_map_info {
1764 vm_object_in_map(vm_object_t object)
1766 struct vm_object_in_map_info info;
1769 info.object = object;
1771 allproc_scan(vm_object_in_map_callback, &info);
1774 if( _vm_object_in_map(&kernel_map, object, 0))
1776 if( _vm_object_in_map(&pager_map, object, 0))
1778 if( _vm_object_in_map(&buffer_map, object, 0))
1784 vm_object_in_map_callback(struct proc *p, void *data)
1786 struct vm_object_in_map_info *info = data;
1789 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) {
1797 DB_SHOW_COMMAND(vmochk, vm_object_check)
1802 * make sure that internal objs are in a map somewhere
1803 * and none have zero ref counts.
1805 for (object = TAILQ_FIRST(&vm_object_list);
1807 object = TAILQ_NEXT(object, object_list)) {
1808 if (object->handle == NULL &&
1809 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1810 if (object->ref_count == 0) {
1811 db_printf("vmochk: internal obj has zero ref count: %ld\n",
1812 (long)object->size);
1814 if (!vm_object_in_map(object)) {
1816 "vmochk: internal obj is not in a map: "
1817 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1818 object->ref_count, (u_long)object->size,
1819 (u_long)object->size,
1820 (void *)object->backing_object);
1827 * vm_object_print: [ debug ]
1829 DB_SHOW_COMMAND(object, vm_object_print_static)
1831 /* XXX convert args. */
1832 vm_object_t object = (vm_object_t)addr;
1833 boolean_t full = have_addr;
1837 /* XXX count is an (unused) arg. Avoid shadowing it. */
1838 #define count was_count
1846 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1847 object, (int)object->type, (u_long)object->size,
1848 object->resident_page_count, object->ref_count, object->flags);
1850 * XXX no %qd in kernel. Truncate object->backing_object_offset.
1852 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1853 object->shadow_count,
1854 object->backing_object ? object->backing_object->ref_count : 0,
1855 object->backing_object, (long)object->backing_object_offset);
1862 RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) {
1864 db_iprintf("memory:=");
1865 else if (count == 6) {
1873 db_printf("(off=0x%lx,page=0x%lx)",
1874 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1884 /* XXX need this non-static entry for calling from vm_map_print. */
1886 vm_object_print(/* db_expr_t */ long addr,
1887 boolean_t have_addr,
1888 /* db_expr_t */ long count,
1891 vm_object_print_static(addr, have_addr, count, modif);
1894 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1899 for (object = TAILQ_FIRST(&vm_object_list);
1901 object = TAILQ_NEXT(object, object_list)) {
1902 vm_pindex_t idx, fidx;
1904 vm_paddr_t pa = -1, padiff;
1908 db_printf("new object: %p\n", (void *)object);
1918 osize = object->size;
1921 for (idx = 0; idx < osize; idx++) {
1922 m = vm_page_lookup(object, idx);
1925 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1926 (long)fidx, rcount, (long)pa);
1941 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1946 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1947 padiff >>= PAGE_SHIFT;
1948 padiff &= PQ_L2_MASK;
1950 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1954 db_printf(" index(%ld)run(%d)pa(0x%lx)",
1955 (long)fidx, rcount, (long)pa);
1956 db_printf("pd(%ld)\n", (long)padiff);
1966 pa = VM_PAGE_TO_PHYS(m);
1970 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1971 (long)fidx, rcount, (long)pa);