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 1.22 2005/06/02 20:57:21 swildner 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 #define MSYNC_FLUSH_HARDSEQ 0x01
98 #define MSYNC_FLUSH_SOFTSEQ 0x02
100 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ;
101 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags,
102 CTLFLAG_RW, &msync_flush_flags, 0, "");
104 static void vm_object_qcollapse (vm_object_t object);
105 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
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;
134 static long vm_object_count; /* count of all objects */
135 vm_object_t kernel_object;
136 vm_object_t kmem_object;
137 static struct vm_object kernel_object_store;
138 static struct vm_object kmem_object_store;
139 extern int vm_pageout_page_count;
141 static long object_collapses;
142 static long object_bypasses;
143 static int next_index;
144 static vm_zone_t obj_zone;
145 static struct vm_zone obj_zone_store;
146 static int object_hash_rand;
147 #define VM_OBJECTS_INIT 256
148 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
151 _vm_object_allocate(objtype_t type, vm_size_t size, vm_object_t object)
154 TAILQ_INIT(&object->memq);
155 LIST_INIT(&object->shadow_head);
159 object->ref_count = 1;
161 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
162 vm_object_set_flag(object, OBJ_ONEMAPPING);
163 object->paging_in_progress = 0;
164 object->resident_page_count = 0;
165 object->shadow_count = 0;
166 object->pg_color = next_index;
167 if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
168 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
171 next_index = (next_index + incr) & PQ_L2_MASK;
172 object->handle = NULL;
173 object->backing_object = NULL;
174 object->backing_object_offset = (vm_ooffset_t) 0;
176 * Try to generate a number that will spread objects out in the
177 * hash table. We 'wipe' new objects across the hash in 128 page
178 * increments plus 1 more to offset it a little more by the time
181 object->hash_rand = object_hash_rand - 129;
183 object->generation++;
186 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
188 object_hash_rand = object->hash_rand;
195 * Initialize the VM objects module.
200 TAILQ_INIT(&vm_object_list);
202 kernel_object = &kernel_object_store;
203 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
206 kmem_object = &kmem_object_store;
207 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
210 obj_zone = &obj_zone_store;
211 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
212 vm_objects_init, VM_OBJECTS_INIT);
216 vm_object_init2(void)
218 zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1);
222 * vm_object_allocate:
224 * Returns a new object with the given size.
228 vm_object_allocate(objtype_t type, vm_size_t size)
232 result = (vm_object_t) zalloc(obj_zone);
234 _vm_object_allocate(type, size, result);
241 * vm_object_reference:
243 * Gets another reference to the given object.
246 vm_object_reference(vm_object_t object)
252 /* object can be re-referenced during final cleaning */
253 KASSERT(!(object->flags & OBJ_DEAD),
254 ("vm_object_reference: attempting to reference dead obj"));
258 if (object->type == OBJT_VNODE) {
259 vref(object->handle);
260 /* XXX what if the vnode is being destroyed? */
262 while (vget((struct vnode *) object->handle,
263 LK_RETRY|LK_NOOBJ, curthread)) {
264 printf("vm_object_reference: delay in getting object\n");
271 vm_object_vndeallocate(vm_object_t object)
273 struct vnode *vp = (struct vnode *) object->handle;
275 KASSERT(object->type == OBJT_VNODE,
276 ("vm_object_vndeallocate: not a vnode object"));
277 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
279 if (object->ref_count == 0) {
280 vprint("vm_object_vndeallocate", vp);
281 panic("vm_object_vndeallocate: bad object reference count");
286 if (object->ref_count == 0) {
287 vp->v_flag &= ~VTEXT;
288 vm_object_clear_flag(object, OBJ_OPT);
294 * vm_object_deallocate:
296 * Release a reference to the specified object,
297 * gained either through a vm_object_allocate
298 * or a vm_object_reference call. When all references
299 * are gone, storage associated with this object
300 * may be relinquished.
302 * No object may be locked.
305 vm_object_deallocate(vm_object_t object)
309 while (object != NULL) {
311 if (object->type == OBJT_VNODE) {
312 vm_object_vndeallocate(object);
316 if (object->ref_count == 0) {
317 panic("vm_object_deallocate: object deallocated too many times: %d", object->type);
318 } else if (object->ref_count > 2) {
324 * Here on ref_count of one or two, which are special cases for
327 if ((object->ref_count == 2) && (object->shadow_count == 0)) {
328 vm_object_set_flag(object, OBJ_ONEMAPPING);
331 } else if ((object->ref_count == 2) && (object->shadow_count == 1)) {
333 if ((object->handle == NULL) &&
334 (object->type == OBJT_DEFAULT ||
335 object->type == OBJT_SWAP)) {
338 robject = LIST_FIRST(&object->shadow_head);
339 KASSERT(robject != NULL,
340 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
342 object->shadow_count));
343 if ((robject->handle == NULL) &&
344 (robject->type == OBJT_DEFAULT ||
345 robject->type == OBJT_SWAP)) {
347 robject->ref_count++;
350 robject->paging_in_progress ||
351 object->paging_in_progress
353 vm_object_pip_sleep(robject, "objde1");
354 vm_object_pip_sleep(object, "objde2");
357 if (robject->ref_count == 1) {
358 robject->ref_count--;
364 vm_object_collapse(object);
373 if (object->ref_count != 0)
379 temp = object->backing_object;
381 LIST_REMOVE(object, shadow_list);
382 temp->shadow_count--;
383 if (temp->ref_count == 0)
384 vm_object_clear_flag(temp, OBJ_OPT);
386 object->backing_object = NULL;
390 * Don't double-terminate, we could be in a termination
391 * recursion due to the terminate having to sync data
394 if ((object->flags & OBJ_DEAD) == 0)
395 vm_object_terminate(object);
401 * vm_object_terminate actually destroys the specified object, freeing
402 * up all previously used resources.
404 * The object must be locked.
405 * This routine may block.
408 vm_object_terminate(vm_object_t object)
413 * Make sure no one uses us.
415 vm_object_set_flag(object, OBJ_DEAD);
418 * wait for the pageout daemon to be done with the object
420 vm_object_pip_wait(object, "objtrm");
422 KASSERT(!object->paging_in_progress,
423 ("vm_object_terminate: pageout in progress"));
426 * Clean and free the pages, as appropriate. All references to the
427 * object are gone, so we don't need to lock it.
429 if (object->type == OBJT_VNODE) {
433 * Freeze optimized copies.
435 vm_freeze_copyopts(object, 0, object->size);
438 * Clean pages and flush buffers.
440 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
442 vp = (struct vnode *) object->handle;
443 vinvalbuf(vp, V_SAVE, NULL, 0, 0);
447 * Wait for any I/O to complete, after which there had better not
448 * be any references left on the object.
450 vm_object_pip_wait(object, "objtrm");
452 if (object->ref_count != 0)
453 panic("vm_object_terminate: object with references, ref_count=%d", object->ref_count);
456 * Now free any remaining pages. For internal objects, this also
457 * removes them from paging queues. Don't free wired pages, just
458 * remove them from the object.
461 while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
462 if (p->busy || (p->flags & PG_BUSY))
463 panic("vm_object_terminate: freeing busy page %p", p);
464 if (p->wire_count == 0) {
467 mycpu->gd_cnt.v_pfree++;
477 * Let the pager know object is dead.
479 vm_pager_deallocate(object);
482 * Remove the object from the global object list.
485 TAILQ_REMOVE(&vm_object_list, object, object_list);
490 if (object->ref_count != 0)
491 panic("vm_object_terminate2: object with references, ref_count=%d", object->ref_count);
494 * Free the space for the object.
496 zfree(obj_zone, object);
500 * vm_object_page_clean
502 * Clean all dirty pages in the specified range of object. Leaves page
503 * on whatever queue it is currently on. If NOSYNC is set then do not
504 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
505 * leaving the object dirty.
507 * When stuffing pages asynchronously, allow clustering. XXX we need a
508 * synchronous clustering mode implementation.
510 * Odd semantics: if start == end, we clean everything.
514 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
518 vm_offset_t tstart, tend;
525 if (object->type != OBJT_VNODE ||
526 (object->flags & OBJ_MIGHTBEDIRTY) == 0)
529 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
530 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
534 vm_object_set_flag(object, OBJ_CLEANING);
537 * Handle 'entire object' case
547 * If the caller is smart and only msync()s a range he knows is
548 * dirty, we may be able to avoid an object scan. This results in
549 * a phenominal improvement in performance. We cannot do this
550 * as a matter of course because the object may be huge - e.g.
551 * the size might be in the gigabytes or terrabytes.
553 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
558 scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
561 pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
563 scanlimit = scanreset;
567 * spl protection is required despite the obj generation
568 * tracking because we cannot safely call vm_page_test_dirty()
569 * or avoid page field tests against an interrupt unbusy/free
570 * race that might occur prior to the busy check in
571 * vm_object_page_collect_flush().
574 while (tscan < tend) {
575 curgeneration = object->generation;
576 p = vm_page_lookup(object, tscan);
577 if (p == NULL || p->valid == 0 ||
578 (p->queue - p->pc) == PQ_CACHE) {
579 if (--scanlimit == 0)
584 vm_page_test_dirty(p);
585 if ((p->dirty & p->valid) == 0) {
586 if (--scanlimit == 0)
592 * If we have been asked to skip nosync pages and
593 * this is a nosync page, we can't continue.
595 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
596 if (--scanlimit == 0)
601 scanlimit = scanreset;
604 * This returns 0 if it was unable to busy the first
605 * page (i.e. had to sleep).
607 tscan += vm_object_page_collect_flush(object, p,
608 curgeneration, pagerflags);
613 * If everything was dirty and we flushed it successfully,
614 * and the requested range is not the entire object, we
615 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
616 * return immediately.
618 if (tscan >= tend && (tstart || tend < object->size)) {
619 vm_object_clear_flag(object, OBJ_CLEANING);
622 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
626 * Generally set CLEANCHK interlock and make the page read-only so
627 * we can then clear the object flags.
629 * However, if this is a nosync mmap then the object is likely to
630 * stay dirty so do not mess with the page and do not clear the
633 * spl protection is required because an interrupt can remove page
639 for (p = TAILQ_FIRST(&object->memq); p; p = TAILQ_NEXT(p, listq)) {
640 vm_page_flag_set(p, PG_CLEANCHK);
641 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
644 vm_page_protect(p, VM_PROT_READ);
648 if (clearobjflags && (tstart == 0) && (tend == object->size)) {
651 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
652 if (object->type == OBJT_VNODE &&
653 (vp = (struct vnode *)object->handle) != NULL) {
654 if (vp->v_flag & VOBJDIRTY)
655 vclrflags(vp, VOBJDIRTY);
660 * spl protection is required both to avoid an interrupt unbusy/free
661 * race against a vm_page_lookup(), and also to ensure that the
662 * memq is consistent. We do not want a busy page to be ripped out
669 curgeneration = object->generation;
671 for (p = TAILQ_FIRST(&object->memq); p; p = np) {
674 np = TAILQ_NEXT(p, listq);
678 if (((p->flags & PG_CLEANCHK) == 0) ||
679 (pi < tstart) || (pi >= tend) ||
681 ((p->queue - p->pc) == PQ_CACHE)) {
682 vm_page_flag_clear(p, PG_CLEANCHK);
686 vm_page_test_dirty(p);
687 if ((p->dirty & p->valid) == 0) {
688 vm_page_flag_clear(p, PG_CLEANCHK);
693 * If we have been asked to skip nosync pages and this is a
694 * nosync page, skip it. Note that the object flags were
695 * not cleared in this case so we do not have to set them.
697 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
698 vm_page_flag_clear(p, PG_CLEANCHK);
702 n = vm_object_page_collect_flush(object, p,
703 curgeneration, pagerflags);
706 if (object->generation != curgeneration)
710 * Try to optimize the next page. If we can't we pick up
711 * our (random) scan where we left off.
713 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
714 if ((p = vm_page_lookup(object, pi + n)) != NULL)
721 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
724 vm_object_clear_flag(object, OBJ_CLEANING);
729 * This routine must be called within a critical section to properly avoid
730 * an interrupt unbusy/free race that can occur prior to the busy check.
732 * Using the object generation number here to detect page ripout is not
733 * the best idea in the world. XXX
735 * NOTE: we operate under the assumption that a page found to not be busy
736 * will not be ripped out from under us by an interrupt. XXX we should
737 * recode this to explicitly busy the pages.
740 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
748 vm_page_t maf[vm_pageout_page_count];
749 vm_page_t mab[vm_pageout_page_count];
750 vm_page_t ma[vm_pageout_page_count];
753 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
754 if (object->generation != curgeneration) {
760 for(i = 1; i < vm_pageout_page_count; i++) {
763 if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
764 if ((tp->flags & PG_BUSY) ||
765 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
766 (tp->flags & PG_CLEANCHK) == 0) ||
769 if((tp->queue - tp->pc) == PQ_CACHE) {
770 vm_page_flag_clear(tp, PG_CLEANCHK);
773 vm_page_test_dirty(tp);
774 if ((tp->dirty & tp->valid) == 0) {
775 vm_page_flag_clear(tp, PG_CLEANCHK);
786 chkb = vm_pageout_page_count - maxf;
788 for(i = 1; i < chkb;i++) {
791 if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
792 if ((tp->flags & PG_BUSY) ||
793 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
794 (tp->flags & PG_CLEANCHK) == 0) ||
797 if((tp->queue - tp->pc) == PQ_CACHE) {
798 vm_page_flag_clear(tp, PG_CLEANCHK);
801 vm_page_test_dirty(tp);
802 if ((tp->dirty & tp->valid) == 0) {
803 vm_page_flag_clear(tp, PG_CLEANCHK);
814 for(i = 0; i < maxb; i++) {
815 int index = (maxb - i) - 1;
817 vm_page_flag_clear(ma[index], PG_CLEANCHK);
819 vm_page_flag_clear(p, PG_CLEANCHK);
821 for(i = 0; i < maxf; i++) {
822 int index = (maxb + i) + 1;
824 vm_page_flag_clear(ma[index], PG_CLEANCHK);
826 runlen = maxb + maxf + 1;
828 vm_pageout_flush(ma, runlen, pagerflags);
829 for (i = 0; i < runlen; i++) {
830 if (ma[i]->valid & ma[i]->dirty) {
831 vm_page_protect(ma[i], VM_PROT_READ);
832 vm_page_flag_set(ma[i], PG_CLEANCHK);
835 * maxf will end up being the actual number of pages
836 * we wrote out contiguously, non-inclusive of the
837 * first page. We do not count look-behind pages.
839 if (i >= maxb + 1 && (maxf > i - maxb - 1))
847 /* XXX I cannot tell if this should be an exported symbol */
849 * vm_object_deactivate_pages
851 * Deactivate all pages in the specified object. (Keep its pages
852 * in memory even though it is no longer referenced.)
854 * The object must be locked.
857 vm_object_deactivate_pages(vm_object_t object)
862 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
863 next = TAILQ_NEXT(p, listq);
864 vm_page_deactivate(p);
871 * Same as vm_object_pmap_copy, except range checking really
872 * works, and is meant for small sections of an object.
874 * This code protects resident pages by making them read-only
875 * and is typically called on a fork or split when a page
876 * is converted to copy-on-write.
878 * NOTE: If the page is already at VM_PROT_NONE, calling
879 * vm_page_protect will have no effect.
882 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
887 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
891 * spl protection needed to prevent races between the lookup,
892 * an interrupt unbusy/free, and our protect call.
895 for (idx = start; idx < end; idx++) {
896 p = vm_page_lookup(object, idx);
899 vm_page_protect(p, VM_PROT_READ);
905 * vm_object_pmap_remove:
907 * Removes all physical pages in the specified
908 * object range from all physical maps.
910 * The object must *not* be locked.
913 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
921 * spl protection is required because an interrupt can unbusy/free
925 for (p = TAILQ_FIRST(&object->memq);
927 p = TAILQ_NEXT(p, listq)
929 if (p->pindex >= start && p->pindex < end)
930 vm_page_protect(p, VM_PROT_NONE);
933 if ((start == 0) && (object->size == end))
934 vm_object_clear_flag(object, OBJ_WRITEABLE);
940 * Implements the madvise function at the object/page level.
942 * MADV_WILLNEED (any object)
944 * Activate the specified pages if they are resident.
946 * MADV_DONTNEED (any object)
948 * Deactivate the specified pages if they are resident.
950 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
951 * OBJ_ONEMAPPING only)
953 * Deactivate and clean the specified pages if they are
954 * resident. This permits the process to reuse the pages
955 * without faulting or the kernel to reclaim the pages
959 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
961 vm_pindex_t end, tpindex;
968 end = pindex + count;
971 * Locate and adjust resident pages
974 for (; pindex < end; pindex += 1) {
980 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
981 * and those pages must be OBJ_ONEMAPPING.
983 if (advise == MADV_FREE) {
984 if ((tobject->type != OBJT_DEFAULT &&
985 tobject->type != OBJT_SWAP) ||
986 (tobject->flags & OBJ_ONEMAPPING) == 0) {
992 * spl protection is required to avoid a race between the
993 * lookup, an interrupt unbusy/free, and our busy check.
997 m = vm_page_lookup(tobject, tpindex);
1001 * There may be swap even if there is no backing page
1003 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1004 swap_pager_freespace(tobject, tpindex, 1);
1010 if (tobject->backing_object == NULL)
1012 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1013 tobject = tobject->backing_object;
1018 * If the page is busy or not in a normal active state,
1019 * we skip it. If the page is not managed there are no
1020 * page queues to mess with. Things can break if we mess
1021 * with pages in any of the below states.
1026 (m->flags & PG_UNMANAGED) ||
1027 m->valid != VM_PAGE_BITS_ALL
1033 if (vm_page_sleep_busy(m, TRUE, "madvpo")) {
1040 * Theoretically once a page is known not to be busy, an
1041 * interrupt cannot come along and rip it out from under us.
1044 if (advise == MADV_WILLNEED) {
1045 vm_page_activate(m);
1046 } else if (advise == MADV_DONTNEED) {
1047 vm_page_dontneed(m);
1048 } else if (advise == MADV_FREE) {
1050 * Mark the page clean. This will allow the page
1051 * to be freed up by the system. However, such pages
1052 * are often reused quickly by malloc()/free()
1053 * so we do not do anything that would cause
1054 * a page fault if we can help it.
1056 * Specifically, we do not try to actually free
1057 * the page now nor do we try to put it in the
1058 * cache (which would cause a page fault on reuse).
1060 * But we do make the page is freeable as we
1061 * can without actually taking the step of unmapping
1064 pmap_clear_modify(m);
1067 vm_page_dontneed(m);
1068 if (tobject->type == OBJT_SWAP)
1069 swap_pager_freespace(tobject, tpindex, 1);
1077 * Create a new object which is backed by the
1078 * specified existing object range. The source
1079 * object reference is deallocated.
1081 * The new object and offset into that object
1082 * are returned in the source parameters.
1086 vm_object_shadow(vm_object_t *object, /* IN/OUT */
1087 vm_ooffset_t *offset, /* IN/OUT */
1096 * Don't create the new object if the old object isn't shared.
1099 if (source != NULL &&
1100 source->ref_count == 1 &&
1101 source->handle == NULL &&
1102 (source->type == OBJT_DEFAULT ||
1103 source->type == OBJT_SWAP))
1107 * Allocate a new object with the given length
1110 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1111 panic("vm_object_shadow: no object for shadowing");
1114 * The new object shadows the source object, adding a reference to it.
1115 * Our caller changes his reference to point to the new object,
1116 * removing a reference to the source object. Net result: no change
1117 * of reference count.
1119 * Try to optimize the result object's page color when shadowing
1120 * in order to maintain page coloring consistency in the combined
1123 result->backing_object = source;
1125 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1126 source->shadow_count++;
1127 source->generation++;
1128 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1132 * Store the offset into the source object, and fix up the offset into
1136 result->backing_object_offset = *offset;
1139 * Return the new things
1146 #define OBSC_TEST_ALL_SHADOWED 0x0001
1147 #define OBSC_COLLAPSE_NOWAIT 0x0002
1148 #define OBSC_COLLAPSE_WAIT 0x0004
1151 vm_object_backing_scan(vm_object_t object, int op)
1155 vm_object_t backing_object;
1156 vm_pindex_t backing_offset_index;
1159 * spl protection is required to avoid races between the memq/lookup,
1160 * an interrupt doing an unbusy/free, and our busy check. Amoung
1165 backing_object = object->backing_object;
1166 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1169 * Initial conditions
1172 if (op & OBSC_TEST_ALL_SHADOWED) {
1174 * We do not want to have to test for the existence of
1175 * swap pages in the backing object. XXX but with the
1176 * new swapper this would be pretty easy to do.
1178 * XXX what about anonymous MAP_SHARED memory that hasn't
1179 * been ZFOD faulted yet? If we do not test for this, the
1180 * shadow test may succeed! XXX
1182 if (backing_object->type != OBJT_DEFAULT) {
1187 if (op & OBSC_COLLAPSE_WAIT) {
1188 KKASSERT((backing_object->flags & OBJ_DEAD) == 0);
1189 vm_object_set_flag(backing_object, OBJ_DEAD);
1196 p = TAILQ_FIRST(&backing_object->memq);
1198 vm_page_t next = TAILQ_NEXT(p, listq);
1199 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1201 if (op & OBSC_TEST_ALL_SHADOWED) {
1205 * Ignore pages outside the parent object's range
1206 * and outside the parent object's mapping of the
1209 * note that we do not busy the backing object's
1214 p->pindex < backing_offset_index ||
1215 new_pindex >= object->size
1222 * See if the parent has the page or if the parent's
1223 * object pager has the page. If the parent has the
1224 * page but the page is not valid, the parent's
1225 * object pager must have the page.
1227 * If this fails, the parent does not completely shadow
1228 * the object and we might as well give up now.
1231 pp = vm_page_lookup(object, new_pindex);
1233 (pp == NULL || pp->valid == 0) &&
1234 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1242 * Check for busy page
1245 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1248 if (op & OBSC_COLLAPSE_NOWAIT) {
1250 (p->flags & PG_BUSY) ||
1259 } else if (op & OBSC_COLLAPSE_WAIT) {
1260 if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1262 * If we slept, anything could have
1263 * happened. Since the object is
1264 * marked dead, the backing offset
1265 * should not have changed so we
1266 * just restart our scan.
1268 p = TAILQ_FIRST(&backing_object->memq);
1279 p->object == backing_object,
1280 ("vm_object_qcollapse(): object mismatch")
1284 * Destroy any associated swap
1286 if (backing_object->type == OBJT_SWAP) {
1287 swap_pager_freespace(
1295 p->pindex < backing_offset_index ||
1296 new_pindex >= object->size
1299 * Page is out of the parent object's range, we
1300 * can simply destroy it.
1302 vm_page_protect(p, VM_PROT_NONE);
1308 pp = vm_page_lookup(object, new_pindex);
1311 vm_pager_has_page(object, new_pindex, NULL, NULL)
1314 * page already exists in parent OR swap exists
1315 * for this location in the parent. Destroy
1316 * the original page from the backing object.
1318 * Leave the parent's page alone
1320 vm_page_protect(p, VM_PROT_NONE);
1327 * Page does not exist in parent, rename the
1328 * page from the backing object to the main object.
1330 * If the page was mapped to a process, it can remain
1331 * mapped through the rename.
1333 if ((p->queue - p->pc) == PQ_CACHE)
1334 vm_page_deactivate(p);
1336 vm_page_rename(p, object, new_pindex);
1337 /* page automatically made dirty by rename */
1347 * this version of collapse allows the operation to occur earlier and
1348 * when paging_in_progress is true for an object... This is not a complete
1349 * operation, but should plug 99.9% of the rest of the leaks.
1352 vm_object_qcollapse(vm_object_t object)
1354 vm_object_t backing_object = object->backing_object;
1356 if (backing_object->ref_count != 1)
1359 backing_object->ref_count += 2;
1361 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1363 backing_object->ref_count -= 2;
1367 * vm_object_collapse:
1369 * Collapse an object with the object backing it.
1370 * Pages in the backing object are moved into the
1371 * parent, and the backing object is deallocated.
1374 vm_object_collapse(vm_object_t object)
1377 vm_object_t backing_object;
1380 * Verify that the conditions are right for collapse:
1382 * The object exists and the backing object exists.
1387 if ((backing_object = object->backing_object) == NULL)
1391 * we check the backing object first, because it is most likely
1394 if (backing_object->handle != NULL ||
1395 (backing_object->type != OBJT_DEFAULT &&
1396 backing_object->type != OBJT_SWAP) ||
1397 (backing_object->flags & OBJ_DEAD) ||
1398 object->handle != NULL ||
1399 (object->type != OBJT_DEFAULT &&
1400 object->type != OBJT_SWAP) ||
1401 (object->flags & OBJ_DEAD)) {
1406 object->paging_in_progress != 0 ||
1407 backing_object->paging_in_progress != 0
1409 vm_object_qcollapse(object);
1414 * We know that we can either collapse the backing object (if
1415 * the parent is the only reference to it) or (perhaps) have
1416 * the parent bypass the object if the parent happens to shadow
1417 * all the resident pages in the entire backing object.
1419 * This is ignoring pager-backed pages such as swap pages.
1420 * vm_object_backing_scan fails the shadowing test in this
1424 if (backing_object->ref_count == 1) {
1426 * If there is exactly one reference to the backing
1427 * object, we can collapse it into the parent.
1429 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1432 * Move the pager from backing_object to object.
1435 if (backing_object->type == OBJT_SWAP) {
1436 vm_object_pip_add(backing_object, 1);
1439 * scrap the paging_offset junk and do a
1440 * discrete copy. This also removes major
1441 * assumptions about how the swap-pager
1442 * works from where it doesn't belong. The
1443 * new swapper is able to optimize the
1444 * destroy-source case.
1447 vm_object_pip_add(object, 1);
1451 OFF_TO_IDX(object->backing_object_offset), TRUE);
1452 vm_object_pip_wakeup(object);
1454 vm_object_pip_wakeup(backing_object);
1457 * Object now shadows whatever backing_object did.
1458 * Note that the reference to
1459 * backing_object->backing_object moves from within
1460 * backing_object to within object.
1463 LIST_REMOVE(object, shadow_list);
1464 object->backing_object->shadow_count--;
1465 object->backing_object->generation++;
1466 if (backing_object->backing_object) {
1467 LIST_REMOVE(backing_object, shadow_list);
1468 backing_object->backing_object->shadow_count--;
1469 backing_object->backing_object->generation++;
1471 object->backing_object = backing_object->backing_object;
1472 if (object->backing_object) {
1474 &object->backing_object->shadow_head,
1478 object->backing_object->shadow_count++;
1479 object->backing_object->generation++;
1482 object->backing_object_offset +=
1483 backing_object->backing_object_offset;
1486 * Discard backing_object.
1488 * Since the backing object has no pages, no pager left,
1489 * and no object references within it, all that is
1490 * necessary is to dispose of it.
1493 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1494 KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object));
1504 zfree(obj_zone, backing_object);
1508 vm_object_t new_backing_object;
1511 * If we do not entirely shadow the backing object,
1512 * there is nothing we can do so we give up.
1515 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1520 * Make the parent shadow the next object in the
1521 * chain. Deallocating backing_object will not remove
1522 * it, since its reference count is at least 2.
1525 LIST_REMOVE(object, shadow_list);
1526 backing_object->shadow_count--;
1527 backing_object->generation++;
1529 new_backing_object = backing_object->backing_object;
1530 if ((object->backing_object = new_backing_object) != NULL) {
1531 vm_object_reference(new_backing_object);
1533 &new_backing_object->shadow_head,
1537 new_backing_object->shadow_count++;
1538 new_backing_object->generation++;
1539 object->backing_object_offset +=
1540 backing_object->backing_object_offset;
1544 * Drop the reference count on backing_object. Since
1545 * its ref_count was at least 2, it will not vanish;
1546 * so we don't need to call vm_object_deallocate, but
1549 vm_object_deallocate(backing_object);
1554 * Try again with this object's new backing object.
1560 * vm_object_page_remove: [internal]
1562 * Removes all physical pages in the specified
1563 * object range from the object's list of pages.
1566 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1567 boolean_t clean_only)
1573 if (object == NULL || object->resident_page_count == 0)
1576 all = ((end == 0) && (start == 0));
1579 * Since physically-backed objects do not use managed pages, we can't
1580 * remove pages from the object (we must instead remove the page
1581 * references, and then destroy the object).
1583 KASSERT(object->type != OBJT_PHYS,
1584 ("attempt to remove pages from a physical object"));
1587 * Indicating that the object is undergoing paging.
1589 * spl protection is required to avoid a race between the memq scan,
1590 * an interrupt unbusy/free, and the busy check.
1592 vm_object_pip_add(object, 1);
1596 if (all || size > object->resident_page_count / 4) {
1597 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
1598 next = TAILQ_NEXT(p, listq);
1599 if (all || ((start <= p->pindex) && (p->pindex < end))) {
1600 if (p->wire_count != 0) {
1601 vm_page_protect(p, VM_PROT_NONE);
1608 * The busy flags are only cleared at
1609 * interrupt -- minimize the spl transitions
1612 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1615 if (clean_only && p->valid) {
1616 vm_page_test_dirty(p);
1617 if (p->valid & p->dirty)
1622 vm_page_protect(p, VM_PROT_NONE);
1628 if ((p = vm_page_lookup(object, start)) != 0) {
1629 if (p->wire_count != 0) {
1630 vm_page_protect(p, VM_PROT_NONE);
1639 * The busy flags are only cleared at
1640 * interrupt -- minimize the spl transitions
1642 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1645 if (clean_only && p->valid) {
1646 vm_page_test_dirty(p);
1647 if (p->valid & p->dirty) {
1655 vm_page_protect(p, VM_PROT_NONE);
1663 vm_object_pip_wakeup(object);
1667 * Routine: vm_object_coalesce
1668 * Function: Coalesces two objects backing up adjoining
1669 * regions of memory into a single object.
1671 * returns TRUE if objects were combined.
1673 * NOTE: Only works at the moment if the second object is NULL -
1674 * if it's not, which object do we lock first?
1677 * prev_object First object to coalesce
1678 * prev_offset Offset into prev_object
1679 * next_object Second object into coalesce
1680 * next_offset Offset into next_object
1682 * prev_size Size of reference to prev_object
1683 * next_size Size of reference to next_object
1686 * The object must *not* be locked.
1689 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
1690 vm_size_t prev_size, vm_size_t next_size)
1692 vm_pindex_t next_pindex;
1694 if (prev_object == NULL) {
1698 if (prev_object->type != OBJT_DEFAULT &&
1699 prev_object->type != OBJT_SWAP) {
1704 * Try to collapse the object first
1706 vm_object_collapse(prev_object);
1709 * Can't coalesce if: . more than one reference . paged out . shadows
1710 * another object . has a copy elsewhere (any of which mean that the
1711 * pages not mapped to prev_entry may be in use anyway)
1714 if (prev_object->backing_object != NULL) {
1718 prev_size >>= PAGE_SHIFT;
1719 next_size >>= PAGE_SHIFT;
1720 next_pindex = prev_pindex + prev_size;
1722 if ((prev_object->ref_count > 1) &&
1723 (prev_object->size != next_pindex)) {
1728 * Remove any pages that may still be in the object from a previous
1731 if (next_pindex < prev_object->size) {
1732 vm_object_page_remove(prev_object,
1734 next_pindex + next_size, FALSE);
1735 if (prev_object->type == OBJT_SWAP)
1736 swap_pager_freespace(prev_object,
1737 next_pindex, next_size);
1741 * Extend the object if necessary.
1743 if (next_pindex + next_size > prev_object->size)
1744 prev_object->size = next_pindex + next_size;
1750 vm_object_set_writeable_dirty(vm_object_t object)
1754 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1755 if (object->type == OBJT_VNODE &&
1756 (vp = (struct vnode *)object->handle) != NULL) {
1757 if ((vp->v_flag & VOBJDIRTY) == 0) {
1758 vsetflags(vp, VOBJDIRTY);
1765 #include "opt_ddb.h"
1767 #include <sys/kernel.h>
1769 #include <sys/cons.h>
1771 #include <ddb/ddb.h>
1773 static int _vm_object_in_map (vm_map_t map, vm_object_t object,
1774 vm_map_entry_t entry);
1775 static int vm_object_in_map (vm_object_t object);
1778 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
1781 vm_map_entry_t tmpe;
1789 tmpe = map->header.next;
1790 entcount = map->nentries;
1791 while (entcount-- && (tmpe != &map->header)) {
1792 if( _vm_object_in_map(map, object, tmpe)) {
1797 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
1798 tmpm = entry->object.sub_map;
1799 tmpe = tmpm->header.next;
1800 entcount = tmpm->nentries;
1801 while (entcount-- && tmpe != &tmpm->header) {
1802 if( _vm_object_in_map(tmpm, object, tmpe)) {
1807 } else if ((obj = entry->object.vm_object) != NULL) {
1808 for(; obj; obj=obj->backing_object)
1809 if( obj == object) {
1817 vm_object_in_map(vm_object_t object)
1820 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1821 if( !p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1823 if( _vm_object_in_map(&p->p_vmspace->vm_map, object, 0))
1826 if( _vm_object_in_map( kernel_map, object, 0))
1828 if( _vm_object_in_map( pager_map, object, 0))
1830 if( _vm_object_in_map( buffer_map, object, 0))
1835 DB_SHOW_COMMAND(vmochk, vm_object_check)
1840 * make sure that internal objs are in a map somewhere
1841 * and none have zero ref counts.
1843 for (object = TAILQ_FIRST(&vm_object_list);
1845 object = TAILQ_NEXT(object, object_list)) {
1846 if (object->handle == NULL &&
1847 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1848 if (object->ref_count == 0) {
1849 db_printf("vmochk: internal obj has zero ref count: %ld\n",
1850 (long)object->size);
1852 if (!vm_object_in_map(object)) {
1854 "vmochk: internal obj is not in a map: "
1855 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1856 object->ref_count, (u_long)object->size,
1857 (u_long)object->size,
1858 (void *)object->backing_object);
1865 * vm_object_print: [ debug ]
1867 DB_SHOW_COMMAND(object, vm_object_print_static)
1869 /* XXX convert args. */
1870 vm_object_t object = (vm_object_t)addr;
1871 boolean_t full = have_addr;
1875 /* XXX count is an (unused) arg. Avoid shadowing it. */
1876 #define count was_count
1884 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1885 object, (int)object->type, (u_long)object->size,
1886 object->resident_page_count, object->ref_count, object->flags);
1888 * XXX no %qd in kernel. Truncate object->backing_object_offset.
1890 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1891 object->shadow_count,
1892 object->backing_object ? object->backing_object->ref_count : 0,
1893 object->backing_object, (long)object->backing_object_offset);
1900 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = TAILQ_NEXT(p, listq)) {
1902 db_iprintf("memory:=");
1903 else if (count == 6) {
1911 db_printf("(off=0x%lx,page=0x%lx)",
1912 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1922 /* XXX need this non-static entry for calling from vm_map_print. */
1924 vm_object_print(/* db_expr_t */ long addr,
1925 boolean_t have_addr,
1926 /* db_expr_t */ long count,
1929 vm_object_print_static(addr, have_addr, count, modif);
1932 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1937 for (object = TAILQ_FIRST(&vm_object_list);
1939 object = TAILQ_NEXT(object, object_list)) {
1940 vm_pindex_t idx, fidx;
1942 vm_paddr_t pa = -1, padiff;
1946 db_printf("new object: %p\n", (void *)object);
1956 osize = object->size;
1959 for (idx = 0; idx < osize; idx++) {
1960 m = vm_page_lookup(object, idx);
1963 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1964 (long)fidx, rcount, (long)pa);
1979 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1984 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1985 padiff >>= PAGE_SHIFT;
1986 padiff &= PQ_L2_MASK;
1988 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1992 db_printf(" index(%ld)run(%d)pa(0x%lx)",
1993 (long)fidx, rcount, (long)pa);
1994 db_printf("pd(%ld)\n", (long)padiff);
2004 pa = VM_PAGE_TO_PHYS(m);
2008 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2009 (long)fidx, rcount, (long)pa);