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 1.20 2004/10/12 19:21:16 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 #include <sys/thread2.h>
97 #define EASY_SCAN_FACTOR 8
99 #define MSYNC_FLUSH_HARDSEQ 0x01
100 #define MSYNC_FLUSH_SOFTSEQ 0x02
102 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ;
103 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags,
104 CTLFLAG_RW, &msync_flush_flags, 0, "");
106 static void vm_object_qcollapse (vm_object_t object);
107 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
110 * Virtual memory objects maintain the actual data
111 * associated with allocated virtual memory. A given
112 * page of memory exists within exactly one object.
114 * An object is only deallocated when all "references"
115 * are given up. Only one "reference" to a given
116 * region of an object should be writeable.
118 * Associated with each object is a list of all resident
119 * memory pages belonging to that object; this list is
120 * maintained by the "vm_page" module, and locked by the object's
123 * Each object also records a "pager" routine which is
124 * used to retrieve (and store) pages to the proper backing
125 * storage. In addition, objects may be backed by other
126 * objects from which they were virtual-copied.
128 * The only items within the object structure which are
129 * modified after time of creation are:
130 * reference count locked by object's lock
131 * pager routine locked by object's lock
135 struct object_q vm_object_list;
136 static struct lwkt_token vm_object_list_token;
137 static long vm_object_count; /* count of all objects */
138 vm_object_t kernel_object;
139 vm_object_t kmem_object;
140 static struct vm_object kernel_object_store;
141 static struct vm_object kmem_object_store;
142 extern int vm_pageout_page_count;
144 static long object_collapses;
145 static long object_bypasses;
146 static int next_index;
147 static vm_zone_t obj_zone;
148 static struct vm_zone obj_zone_store;
149 static int object_hash_rand;
150 #define VM_OBJECTS_INIT 256
151 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
154 _vm_object_allocate(objtype_t type, vm_size_t size, vm_object_t object)
157 TAILQ_INIT(&object->memq);
158 LIST_INIT(&object->shadow_head);
162 object->ref_count = 1;
164 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
165 vm_object_set_flag(object, OBJ_ONEMAPPING);
166 object->paging_in_progress = 0;
167 object->resident_page_count = 0;
168 object->shadow_count = 0;
169 object->pg_color = next_index;
170 if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
171 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
174 next_index = (next_index + incr) & PQ_L2_MASK;
175 object->handle = NULL;
176 object->backing_object = NULL;
177 object->backing_object_offset = (vm_ooffset_t) 0;
179 * Try to generate a number that will spread objects out in the
180 * hash table. We 'wipe' new objects across the hash in 128 page
181 * increments plus 1 more to offset it a little more by the time
184 object->hash_rand = object_hash_rand - 129;
186 object->generation++;
188 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
190 object_hash_rand = object->hash_rand;
196 * Initialize the VM objects module.
201 TAILQ_INIT(&vm_object_list);
202 lwkt_token_init(&vm_object_list_token);
205 kernel_object = &kernel_object_store;
206 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
209 kmem_object = &kmem_object_store;
210 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
213 obj_zone = &obj_zone_store;
214 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
215 vm_objects_init, VM_OBJECTS_INIT);
219 vm_object_init2(void)
221 zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1);
225 * vm_object_allocate:
227 * Returns a new object with the given size.
231 vm_object_allocate(objtype_t type, vm_size_t size)
235 result = (vm_object_t) zalloc(obj_zone);
237 _vm_object_allocate(type, size, result);
244 * vm_object_reference:
246 * Gets another reference to the given object.
249 vm_object_reference(vm_object_t object)
255 /* object can be re-referenced during final cleaning */
256 KASSERT(!(object->flags & OBJ_DEAD),
257 ("vm_object_reference: attempting to reference dead obj"));
261 if (object->type == OBJT_VNODE) {
262 vref(object->handle);
263 /* XXX what if the vnode is being destroyed? */
265 while (vget((struct vnode *) object->handle,
266 LK_RETRY|LK_NOOBJ, curthread)) {
267 printf("vm_object_reference: delay in getting object\n");
274 vm_object_vndeallocate(vm_object_t object)
276 struct vnode *vp = (struct vnode *) object->handle;
278 KASSERT(object->type == OBJT_VNODE,
279 ("vm_object_vndeallocate: not a vnode object"));
280 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
282 if (object->ref_count == 0) {
283 vprint("vm_object_vndeallocate", vp);
284 panic("vm_object_vndeallocate: bad object reference count");
289 if (object->ref_count == 0) {
290 vp->v_flag &= ~VTEXT;
291 vm_object_clear_flag(object, OBJ_OPT);
297 * vm_object_deallocate:
299 * Release a reference to the specified object,
300 * gained either through a vm_object_allocate
301 * or a vm_object_reference call. When all references
302 * are gone, storage associated with this object
303 * may be relinquished.
305 * No object may be locked.
308 vm_object_deallocate(vm_object_t object)
312 while (object != NULL) {
314 if (object->type == OBJT_VNODE) {
315 vm_object_vndeallocate(object);
319 if (object->ref_count == 0) {
320 panic("vm_object_deallocate: object deallocated too many times: %d", object->type);
321 } else if (object->ref_count > 2) {
327 * Here on ref_count of one or two, which are special cases for
330 if ((object->ref_count == 2) && (object->shadow_count == 0)) {
331 vm_object_set_flag(object, OBJ_ONEMAPPING);
334 } else if ((object->ref_count == 2) && (object->shadow_count == 1)) {
336 if ((object->handle == NULL) &&
337 (object->type == OBJT_DEFAULT ||
338 object->type == OBJT_SWAP)) {
341 robject = LIST_FIRST(&object->shadow_head);
342 KASSERT(robject != NULL,
343 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
345 object->shadow_count));
346 if ((robject->handle == NULL) &&
347 (robject->type == OBJT_DEFAULT ||
348 robject->type == OBJT_SWAP)) {
350 robject->ref_count++;
353 robject->paging_in_progress ||
354 object->paging_in_progress
356 vm_object_pip_sleep(robject, "objde1");
357 vm_object_pip_sleep(object, "objde2");
360 if (robject->ref_count == 1) {
361 robject->ref_count--;
367 vm_object_collapse(object);
376 if (object->ref_count != 0)
382 temp = object->backing_object;
384 LIST_REMOVE(object, shadow_list);
385 temp->shadow_count--;
386 if (temp->ref_count == 0)
387 vm_object_clear_flag(temp, OBJ_OPT);
389 object->backing_object = NULL;
393 * Don't double-terminate, we could be in a termination
394 * recursion due to the terminate having to sync data
397 if ((object->flags & OBJ_DEAD) == 0)
398 vm_object_terminate(object);
404 * vm_object_terminate actually destroys the specified object, freeing
405 * up all previously used resources.
407 * The object must be locked.
408 * This routine may block.
411 vm_object_terminate(vm_object_t object)
417 * Make sure no one uses us.
419 vm_object_set_flag(object, OBJ_DEAD);
422 * wait for the pageout daemon to be done with the object
424 vm_object_pip_wait(object, "objtrm");
426 KASSERT(!object->paging_in_progress,
427 ("vm_object_terminate: pageout in progress"));
430 * Clean and free the pages, as appropriate. All references to the
431 * object are gone, so we don't need to lock it.
433 if (object->type == OBJT_VNODE) {
437 * Freeze optimized copies.
439 vm_freeze_copyopts(object, 0, object->size);
442 * Clean pages and flush buffers.
444 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
446 vp = (struct vnode *) object->handle;
447 vinvalbuf(vp, V_SAVE, NULL, 0, 0);
451 * Wait for any I/O to complete, after which there had better not
452 * be any references left on the object.
454 vm_object_pip_wait(object, "objtrm");
456 if (object->ref_count != 0)
457 panic("vm_object_terminate: object with references, ref_count=%d", object->ref_count);
460 * Now free any remaining pages. For internal objects, this also
461 * removes them from paging queues. Don't free wired pages, just
462 * remove them from the object.
465 while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
466 if (p->busy || (p->flags & PG_BUSY))
467 panic("vm_object_terminate: freeing busy page %p", p);
468 if (p->wire_count == 0) {
471 mycpu->gd_cnt.v_pfree++;
481 * Let the pager know object is dead.
483 vm_pager_deallocate(object);
486 * Remove the object from the global object list.
488 lwkt_gettoken(&ilock, &vm_object_list_token);
489 TAILQ_REMOVE(&vm_object_list, object, object_list);
490 lwkt_reltoken(&ilock);
495 * Free the space for the object.
497 zfree(obj_zone, object);
501 * vm_object_page_clean
503 * Clean all dirty pages in the specified range of object. Leaves page
504 * on whatever queue it is currently on. If NOSYNC is set then do not
505 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
506 * leaving the object dirty.
508 * When stuffing pages asynchronously, allow clustering. XXX we need a
509 * synchronous clustering mode implementation.
511 * Odd semantics: if start == end, we clean everything.
515 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
519 vm_offset_t tstart, tend;
526 if (object->type != OBJT_VNODE ||
527 (object->flags & OBJ_MIGHTBEDIRTY) == 0)
530 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
531 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
535 vm_object_set_flag(object, OBJ_CLEANING);
538 * Handle 'entire object' case
548 * If the caller is smart and only msync()s a range he knows is
549 * dirty, we may be able to avoid an object scan. This results in
550 * a phenominal improvement in performance. We cannot do this
551 * as a matter of course because the object may be huge - e.g.
552 * the size might be in the gigabytes or terrabytes.
554 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
559 scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
562 pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
564 scanlimit = scanreset;
568 * spl protection is required despite the obj generation
569 * tracking because we cannot safely call vm_page_test_dirty()
570 * or avoid page field tests against an interrupt unbusy/free
571 * race that might occur prior to the busy check in
572 * vm_object_page_collect_flush().
575 while (tscan < tend) {
576 curgeneration = object->generation;
577 p = vm_page_lookup(object, tscan);
578 if (p == NULL || p->valid == 0 ||
579 (p->queue - p->pc) == PQ_CACHE) {
580 if (--scanlimit == 0)
585 vm_page_test_dirty(p);
586 if ((p->dirty & p->valid) == 0) {
587 if (--scanlimit == 0)
593 * If we have been asked to skip nosync pages and
594 * this is a nosync page, we can't continue.
596 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
597 if (--scanlimit == 0)
602 scanlimit = scanreset;
605 * This returns 0 if it was unable to busy the first
606 * page (i.e. had to sleep).
608 tscan += vm_object_page_collect_flush(object, p,
609 curgeneration, pagerflags);
614 * If everything was dirty and we flushed it successfully,
615 * and the requested range is not the entire object, we
616 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
617 * return immediately.
619 if (tscan >= tend && (tstart || tend < object->size)) {
620 vm_object_clear_flag(object, OBJ_CLEANING);
623 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
627 * Generally set CLEANCHK interlock and make the page read-only so
628 * we can then clear the object flags.
630 * However, if this is a nosync mmap then the object is likely to
631 * stay dirty so do not mess with the page and do not clear the
634 * spl protection is required because an interrupt can remove page
640 for (p = TAILQ_FIRST(&object->memq); p; p = TAILQ_NEXT(p, listq)) {
641 vm_page_flag_set(p, PG_CLEANCHK);
642 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
645 vm_page_protect(p, VM_PROT_READ);
649 if (clearobjflags && (tstart == 0) && (tend == object->size)) {
652 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
653 if (object->type == OBJT_VNODE &&
654 (vp = (struct vnode *)object->handle) != NULL) {
655 if (vp->v_flag & VOBJDIRTY)
656 vclrflags(vp, VOBJDIRTY);
661 * spl protection is required both to avoid an interrupt unbusy/free
662 * race against a vm_page_lookup(), and also to ensure that the
663 * memq is consistent. We do not want a busy page to be ripped out
670 curgeneration = object->generation;
672 for (p = TAILQ_FIRST(&object->memq); p; p = np) {
675 np = TAILQ_NEXT(p, listq);
679 if (((p->flags & PG_CLEANCHK) == 0) ||
680 (pi < tstart) || (pi >= tend) ||
682 ((p->queue - p->pc) == PQ_CACHE)) {
683 vm_page_flag_clear(p, PG_CLEANCHK);
687 vm_page_test_dirty(p);
688 if ((p->dirty & p->valid) == 0) {
689 vm_page_flag_clear(p, PG_CLEANCHK);
694 * If we have been asked to skip nosync pages and this is a
695 * nosync page, skip it. Note that the object flags were
696 * not cleared in this case so we do not have to set them.
698 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
699 vm_page_flag_clear(p, PG_CLEANCHK);
703 n = vm_object_page_collect_flush(object, p,
704 curgeneration, pagerflags);
707 if (object->generation != curgeneration)
711 * Try to optimize the next page. If we can't we pick up
712 * our (random) scan where we left off.
714 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
715 if ((p = vm_page_lookup(object, pi + n)) != NULL)
722 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
725 vm_object_clear_flag(object, OBJ_CLEANING);
730 * This routine must be called within a critical section to properly avoid
731 * an interrupt unbusy/free race that can occur prior to the busy check.
733 * Using the object generation number here to detect page ripout is not
734 * the best idea in the world. XXX
736 * NOTE: we operate under the assumption that a page found to not be busy
737 * will not be ripped out from under us by an interrupt. XXX we should
738 * recode this to explicitly busy the pages.
741 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
749 vm_page_t maf[vm_pageout_page_count];
750 vm_page_t mab[vm_pageout_page_count];
751 vm_page_t ma[vm_pageout_page_count];
754 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
755 if (object->generation != curgeneration) {
761 for(i = 1; i < vm_pageout_page_count; i++) {
764 if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
765 if ((tp->flags & PG_BUSY) ||
766 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
767 (tp->flags & PG_CLEANCHK) == 0) ||
770 if((tp->queue - tp->pc) == PQ_CACHE) {
771 vm_page_flag_clear(tp, PG_CLEANCHK);
774 vm_page_test_dirty(tp);
775 if ((tp->dirty & tp->valid) == 0) {
776 vm_page_flag_clear(tp, PG_CLEANCHK);
787 chkb = vm_pageout_page_count - maxf;
789 for(i = 1; i < chkb;i++) {
792 if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
793 if ((tp->flags & PG_BUSY) ||
794 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
795 (tp->flags & PG_CLEANCHK) == 0) ||
798 if((tp->queue - tp->pc) == PQ_CACHE) {
799 vm_page_flag_clear(tp, PG_CLEANCHK);
802 vm_page_test_dirty(tp);
803 if ((tp->dirty & tp->valid) == 0) {
804 vm_page_flag_clear(tp, PG_CLEANCHK);
815 for(i = 0; i < maxb; i++) {
816 int index = (maxb - i) - 1;
818 vm_page_flag_clear(ma[index], PG_CLEANCHK);
820 vm_page_flag_clear(p, PG_CLEANCHK);
822 for(i = 0; i < maxf; i++) {
823 int index = (maxb + i) + 1;
825 vm_page_flag_clear(ma[index], PG_CLEANCHK);
827 runlen = maxb + maxf + 1;
829 vm_pageout_flush(ma, runlen, pagerflags);
830 for (i = 0; i < runlen; i++) {
831 if (ma[i]->valid & ma[i]->dirty) {
832 vm_page_protect(ma[i], VM_PROT_READ);
833 vm_page_flag_set(ma[i], PG_CLEANCHK);
836 * maxf will end up being the actual number of pages
837 * we wrote out contiguously, non-inclusive of the
838 * first page. We do not count look-behind pages.
840 if (i >= maxb + 1 && (maxf > i - maxb - 1))
848 /* XXX I cannot tell if this should be an exported symbol */
850 * vm_object_deactivate_pages
852 * Deactivate all pages in the specified object. (Keep its pages
853 * in memory even though it is no longer referenced.)
855 * The object must be locked.
858 vm_object_deactivate_pages(vm_object_t object)
864 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
865 next = TAILQ_NEXT(p, listq);
866 vm_page_deactivate(p);
873 * Same as vm_object_pmap_copy, except range checking really
874 * works, and is meant for small sections of an object.
876 * This code protects resident pages by making them read-only
877 * and is typically called on a fork or split when a page
878 * is converted to copy-on-write.
880 * NOTE: If the page is already at VM_PROT_NONE, calling
881 * vm_page_protect will have no effect.
884 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
889 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
893 * spl protection needed to prevent races between the lookup,
894 * an interrupt unbusy/free, and our protect call.
897 for (idx = start; idx < end; idx++) {
898 p = vm_page_lookup(object, idx);
901 vm_page_protect(p, VM_PROT_READ);
907 * vm_object_pmap_remove:
909 * Removes all physical pages in the specified
910 * object range from all physical maps.
912 * The object must *not* be locked.
915 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
923 * spl protection is required because an interrupt can unbusy/free
927 for (p = TAILQ_FIRST(&object->memq);
929 p = TAILQ_NEXT(p, listq)
931 if (p->pindex >= start && p->pindex < end)
932 vm_page_protect(p, VM_PROT_NONE);
935 if ((start == 0) && (object->size == end))
936 vm_object_clear_flag(object, OBJ_WRITEABLE);
942 * Implements the madvise function at the object/page level.
944 * MADV_WILLNEED (any object)
946 * Activate the specified pages if they are resident.
948 * MADV_DONTNEED (any object)
950 * Deactivate the specified pages if they are resident.
952 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
953 * OBJ_ONEMAPPING only)
955 * Deactivate and clean the specified pages if they are
956 * resident. This permits the process to reuse the pages
957 * without faulting or the kernel to reclaim the pages
961 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
963 vm_pindex_t end, tpindex;
970 end = pindex + count;
973 * Locate and adjust resident pages
976 for (; pindex < end; pindex += 1) {
982 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
983 * and those pages must be OBJ_ONEMAPPING.
985 if (advise == MADV_FREE) {
986 if ((tobject->type != OBJT_DEFAULT &&
987 tobject->type != OBJT_SWAP) ||
988 (tobject->flags & OBJ_ONEMAPPING) == 0) {
994 * spl protection is required to avoid a race between the
995 * lookup, an interrupt unbusy/free, and our busy check.
999 m = vm_page_lookup(tobject, tpindex);
1003 * There may be swap even if there is no backing page
1005 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1006 swap_pager_freespace(tobject, tpindex, 1);
1012 if (tobject->backing_object == NULL)
1014 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1015 tobject = tobject->backing_object;
1020 * If the page is busy or not in a normal active state,
1021 * we skip it. If the page is not managed there are no
1022 * page queues to mess with. Things can break if we mess
1023 * with pages in any of the below states.
1028 (m->flags & PG_UNMANAGED) ||
1029 m->valid != VM_PAGE_BITS_ALL
1035 if (vm_page_sleep_busy(m, TRUE, "madvpo")) {
1042 * Theoretically once a page is known not to be busy, an
1043 * interrupt cannot come along and rip it out from under us.
1046 if (advise == MADV_WILLNEED) {
1047 vm_page_activate(m);
1048 } else if (advise == MADV_DONTNEED) {
1049 vm_page_dontneed(m);
1050 } else if (advise == MADV_FREE) {
1052 * Mark the page clean. This will allow the page
1053 * to be freed up by the system. However, such pages
1054 * are often reused quickly by malloc()/free()
1055 * so we do not do anything that would cause
1056 * a page fault if we can help it.
1058 * Specifically, we do not try to actually free
1059 * the page now nor do we try to put it in the
1060 * cache (which would cause a page fault on reuse).
1062 * But we do make the page is freeable as we
1063 * can without actually taking the step of unmapping
1066 pmap_clear_modify(m);
1069 vm_page_dontneed(m);
1070 if (tobject->type == OBJT_SWAP)
1071 swap_pager_freespace(tobject, tpindex, 1);
1079 * Create a new object which is backed by the
1080 * specified existing object range. The source
1081 * object reference is deallocated.
1083 * The new object and offset into that object
1084 * are returned in the source parameters.
1088 vm_object_shadow(vm_object_t *object, /* IN/OUT */
1089 vm_ooffset_t *offset, /* IN/OUT */
1098 * Don't create the new object if the old object isn't shared.
1101 if (source != NULL &&
1102 source->ref_count == 1 &&
1103 source->handle == NULL &&
1104 (source->type == OBJT_DEFAULT ||
1105 source->type == OBJT_SWAP))
1109 * Allocate a new object with the given length
1112 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1113 panic("vm_object_shadow: no object for shadowing");
1116 * The new object shadows the source object, adding a reference to it.
1117 * Our caller changes his reference to point to the new object,
1118 * removing a reference to the source object. Net result: no change
1119 * of reference count.
1121 * Try to optimize the result object's page color when shadowing
1122 * in order to maintain page coloring consistency in the combined
1125 result->backing_object = source;
1127 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1128 source->shadow_count++;
1129 source->generation++;
1130 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1134 * Store the offset into the source object, and fix up the offset into
1138 result->backing_object_offset = *offset;
1141 * Return the new things
1148 #define OBSC_TEST_ALL_SHADOWED 0x0001
1149 #define OBSC_COLLAPSE_NOWAIT 0x0002
1150 #define OBSC_COLLAPSE_WAIT 0x0004
1153 vm_object_backing_scan(vm_object_t object, int op)
1157 vm_object_t backing_object;
1158 vm_pindex_t backing_offset_index;
1161 * spl protection is required to avoid races between the memq/lookup,
1162 * an interrupt doing an unbusy/free, and our busy check. Amoung
1167 backing_object = object->backing_object;
1168 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1171 * Initial conditions
1174 if (op & OBSC_TEST_ALL_SHADOWED) {
1176 * We do not want to have to test for the existence of
1177 * swap pages in the backing object. XXX but with the
1178 * new swapper this would be pretty easy to do.
1180 * XXX what about anonymous MAP_SHARED memory that hasn't
1181 * been ZFOD faulted yet? If we do not test for this, the
1182 * shadow test may succeed! XXX
1184 if (backing_object->type != OBJT_DEFAULT) {
1189 if (op & OBSC_COLLAPSE_WAIT) {
1190 vm_object_set_flag(backing_object, OBJ_DEAD);
1197 p = TAILQ_FIRST(&backing_object->memq);
1199 vm_page_t next = TAILQ_NEXT(p, listq);
1200 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1202 if (op & OBSC_TEST_ALL_SHADOWED) {
1206 * Ignore pages outside the parent object's range
1207 * and outside the parent object's mapping of the
1210 * note that we do not busy the backing object's
1215 p->pindex < backing_offset_index ||
1216 new_pindex >= object->size
1223 * See if the parent has the page or if the parent's
1224 * object pager has the page. If the parent has the
1225 * page but the page is not valid, the parent's
1226 * object pager must have the page.
1228 * If this fails, the parent does not completely shadow
1229 * the object and we might as well give up now.
1232 pp = vm_page_lookup(object, new_pindex);
1234 (pp == NULL || pp->valid == 0) &&
1235 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1243 * Check for busy page
1246 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1249 if (op & OBSC_COLLAPSE_NOWAIT) {
1251 (p->flags & PG_BUSY) ||
1260 } else if (op & OBSC_COLLAPSE_WAIT) {
1261 if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1263 * If we slept, anything could have
1264 * happened. Since the object is
1265 * marked dead, the backing offset
1266 * should not have changed so we
1267 * just restart our scan.
1269 p = TAILQ_FIRST(&backing_object->memq);
1280 p->object == backing_object,
1281 ("vm_object_qcollapse(): object mismatch")
1285 * Destroy any associated swap
1287 if (backing_object->type == OBJT_SWAP) {
1288 swap_pager_freespace(
1296 p->pindex < backing_offset_index ||
1297 new_pindex >= object->size
1300 * Page is out of the parent object's range, we
1301 * can simply destroy it.
1303 vm_page_protect(p, VM_PROT_NONE);
1309 pp = vm_page_lookup(object, new_pindex);
1312 vm_pager_has_page(object, new_pindex, NULL, NULL)
1315 * page already exists in parent OR swap exists
1316 * for this location in the parent. Destroy
1317 * the original page from the backing object.
1319 * Leave the parent's page alone
1321 vm_page_protect(p, VM_PROT_NONE);
1328 * Page does not exist in parent, rename the
1329 * page from the backing object to the main object.
1331 * If the page was mapped to a process, it can remain
1332 * mapped through the rename.
1334 if ((p->queue - p->pc) == PQ_CACHE)
1335 vm_page_deactivate(p);
1337 vm_page_rename(p, object, new_pindex);
1338 /* page automatically made dirty by rename */
1348 * this version of collapse allows the operation to occur earlier and
1349 * when paging_in_progress is true for an object... This is not a complete
1350 * operation, but should plug 99.9% of the rest of the leaks.
1353 vm_object_qcollapse(vm_object_t object)
1355 vm_object_t backing_object = object->backing_object;
1357 if (backing_object->ref_count != 1)
1360 backing_object->ref_count += 2;
1362 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1364 backing_object->ref_count -= 2;
1368 * vm_object_collapse:
1370 * Collapse an object with the object backing it.
1371 * Pages in the backing object are moved into the
1372 * parent, and the backing object is deallocated.
1375 vm_object_collapse(vm_object_t object)
1378 vm_object_t backing_object;
1381 * Verify that the conditions are right for collapse:
1383 * The object exists and the backing object exists.
1388 if ((backing_object = object->backing_object) == NULL)
1392 * we check the backing object first, because it is most likely
1395 if (backing_object->handle != NULL ||
1396 (backing_object->type != OBJT_DEFAULT &&
1397 backing_object->type != OBJT_SWAP) ||
1398 (backing_object->flags & OBJ_DEAD) ||
1399 object->handle != NULL ||
1400 (object->type != OBJT_DEFAULT &&
1401 object->type != OBJT_SWAP) ||
1402 (object->flags & OBJ_DEAD)) {
1407 object->paging_in_progress != 0 ||
1408 backing_object->paging_in_progress != 0
1410 vm_object_qcollapse(object);
1415 * We know that we can either collapse the backing object (if
1416 * the parent is the only reference to it) or (perhaps) have
1417 * the parent bypass the object if the parent happens to shadow
1418 * all the resident pages in the entire backing object.
1420 * This is ignoring pager-backed pages such as swap pages.
1421 * vm_object_backing_scan fails the shadowing test in this
1425 if (backing_object->ref_count == 1) {
1427 * If there is exactly one reference to the backing
1428 * object, we can collapse it into the parent.
1431 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1434 * Move the pager from backing_object to object.
1437 if (backing_object->type == OBJT_SWAP) {
1438 vm_object_pip_add(backing_object, 1);
1441 * scrap the paging_offset junk and do a
1442 * discrete copy. This also removes major
1443 * assumptions about how the swap-pager
1444 * works from where it doesn't belong. The
1445 * new swapper is able to optimize the
1446 * destroy-source case.
1449 vm_object_pip_add(object, 1);
1453 OFF_TO_IDX(object->backing_object_offset), TRUE);
1454 vm_object_pip_wakeup(object);
1456 vm_object_pip_wakeup(backing_object);
1459 * Object now shadows whatever backing_object did.
1460 * Note that the reference to
1461 * backing_object->backing_object moves from within
1462 * backing_object to within object.
1465 LIST_REMOVE(object, shadow_list);
1466 object->backing_object->shadow_count--;
1467 object->backing_object->generation++;
1468 if (backing_object->backing_object) {
1469 LIST_REMOVE(backing_object, shadow_list);
1470 backing_object->backing_object->shadow_count--;
1471 backing_object->backing_object->generation++;
1473 object->backing_object = backing_object->backing_object;
1474 if (object->backing_object) {
1476 &object->backing_object->shadow_head,
1480 object->backing_object->shadow_count++;
1481 object->backing_object->generation++;
1484 object->backing_object_offset +=
1485 backing_object->backing_object_offset;
1488 * Discard backing_object.
1490 * Since the backing object has no pages, no pager left,
1491 * and no object references within it, all that is
1492 * necessary is to dispose of it.
1495 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1496 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);