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 1.13 2004/03/01 06:33:24 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 #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 struct lwkt_token vm_object_list_token;
135 static long vm_object_count; /* count of all objects */
136 vm_object_t kernel_object;
137 vm_object_t kmem_object;
138 static struct vm_object kernel_object_store;
139 static struct vm_object kmem_object_store;
140 extern int vm_pageout_page_count;
142 static long object_collapses;
143 static long object_bypasses;
144 static int next_index;
145 static vm_zone_t obj_zone;
146 static struct vm_zone obj_zone_store;
147 static int object_hash_rand;
148 #define VM_OBJECTS_INIT 256
149 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
152 _vm_object_allocate(type, size, object)
158 TAILQ_INIT(&object->memq);
159 LIST_INIT(&object->shadow_head);
163 object->ref_count = 1;
165 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
166 vm_object_set_flag(object, OBJ_ONEMAPPING);
167 object->paging_in_progress = 0;
168 object->resident_page_count = 0;
169 object->shadow_count = 0;
170 object->pg_color = next_index;
171 if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
172 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
175 next_index = (next_index + incr) & PQ_L2_MASK;
176 object->handle = NULL;
177 object->backing_object = NULL;
178 object->backing_object_offset = (vm_ooffset_t) 0;
180 * Try to generate a number that will spread objects out in the
181 * hash table. We 'wipe' new objects across the hash in 128 page
182 * increments plus 1 more to offset it a little more by the time
185 object->hash_rand = object_hash_rand - 129;
187 object->generation++;
189 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
191 object_hash_rand = object->hash_rand;
197 * Initialize the VM objects module.
202 TAILQ_INIT(&vm_object_list);
203 lwkt_token_init(&vm_object_list_token);
206 kernel_object = &kernel_object_store;
207 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
210 kmem_object = &kmem_object_store;
211 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
214 obj_zone = &obj_zone_store;
215 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
216 vm_objects_init, VM_OBJECTS_INIT);
222 zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1);
226 * vm_object_allocate:
228 * Returns a new object with the given size.
232 vm_object_allocate(type, size)
238 result = (vm_object_t) zalloc(obj_zone);
240 _vm_object_allocate(type, size, result);
247 * vm_object_reference:
249 * Gets another reference to the given object.
252 vm_object_reference(object)
259 /* object can be re-referenced during final cleaning */
260 KASSERT(!(object->flags & OBJ_DEAD),
261 ("vm_object_reference: attempting to reference dead obj"));
265 if (object->type == OBJT_VNODE) {
266 while (vget((struct vnode *) object->handle, NULL,
267 LK_RETRY|LK_NOOBJ, curthread)) {
268 printf("vm_object_reference: delay in getting object\n");
274 vm_object_vndeallocate(object)
277 struct vnode *vp = (struct vnode *) object->handle;
279 KASSERT(object->type == OBJT_VNODE,
280 ("vm_object_vndeallocate: not a vnode object"));
281 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
283 if (object->ref_count == 0) {
284 vprint("vm_object_vndeallocate", vp);
285 panic("vm_object_vndeallocate: bad object reference count");
290 if (object->ref_count == 0) {
291 vp->v_flag &= ~VTEXT;
292 vm_object_clear_flag(object, OBJ_OPT);
298 * vm_object_deallocate:
300 * Release a reference to the specified object,
301 * gained either through a vm_object_allocate
302 * or a vm_object_reference call. When all references
303 * are gone, storage associated with this object
304 * may be relinquished.
306 * No object may be locked.
309 vm_object_deallocate(object)
314 while (object != NULL) {
316 if (object->type == OBJT_VNODE) {
317 vm_object_vndeallocate(object);
321 if (object->ref_count == 0) {
322 panic("vm_object_deallocate: object deallocated too many times: %d", object->type);
323 } else if (object->ref_count > 2) {
329 * Here on ref_count of one or two, which are special cases for
332 if ((object->ref_count == 2) && (object->shadow_count == 0)) {
333 vm_object_set_flag(object, OBJ_ONEMAPPING);
336 } else if ((object->ref_count == 2) && (object->shadow_count == 1)) {
338 if ((object->handle == NULL) &&
339 (object->type == OBJT_DEFAULT ||
340 object->type == OBJT_SWAP)) {
343 robject = LIST_FIRST(&object->shadow_head);
344 KASSERT(robject != NULL,
345 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
347 object->shadow_count));
348 if ((robject->handle == NULL) &&
349 (robject->type == OBJT_DEFAULT ||
350 robject->type == OBJT_SWAP)) {
352 robject->ref_count++;
355 robject->paging_in_progress ||
356 object->paging_in_progress
358 vm_object_pip_sleep(robject, "objde1");
359 vm_object_pip_sleep(object, "objde2");
362 if (robject->ref_count == 1) {
363 robject->ref_count--;
369 vm_object_collapse(object);
378 if (object->ref_count != 0)
384 temp = object->backing_object;
386 LIST_REMOVE(object, shadow_list);
387 temp->shadow_count--;
388 if (temp->ref_count == 0)
389 vm_object_clear_flag(temp, OBJ_OPT);
391 object->backing_object = NULL;
395 * Don't double-terminate, we could be in a termination
396 * recursion due to the terminate having to sync data
399 if ((object->flags & OBJ_DEAD) == 0)
400 vm_object_terminate(object);
406 * vm_object_terminate actually destroys the specified object, freeing
407 * up all previously used resources.
409 * The object must be locked.
410 * This routine may block.
413 vm_object_terminate(object)
421 * Make sure no one uses us.
423 vm_object_set_flag(object, OBJ_DEAD);
426 * wait for the pageout daemon to be done with the object
428 vm_object_pip_wait(object, "objtrm");
430 KASSERT(!object->paging_in_progress,
431 ("vm_object_terminate: pageout in progress"));
434 * Clean and free the pages, as appropriate. All references to the
435 * object are gone, so we don't need to lock it.
437 if (object->type == OBJT_VNODE) {
441 * Freeze optimized copies.
443 vm_freeze_copyopts(object, 0, object->size);
446 * Clean pages and flush buffers.
448 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
450 vp = (struct vnode *) object->handle;
451 vinvalbuf(vp, V_SAVE, NULL, 0, 0);
455 * Wait for any I/O to complete, after which there had better not
456 * be any references left on the object.
458 vm_object_pip_wait(object, "objtrm");
460 if (object->ref_count != 0)
461 panic("vm_object_terminate: object with references, ref_count=%d", object->ref_count);
464 * Now free any remaining pages. For internal objects, this also
465 * removes them from paging queues. Don't free wired pages, just
466 * remove them from the object.
469 while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
470 if (p->busy || (p->flags & PG_BUSY))
471 panic("vm_object_terminate: freeing busy page %p\n", p);
472 if (p->wire_count == 0) {
475 mycpu->gd_cnt.v_pfree++;
484 * Let the pager know object is dead.
486 vm_pager_deallocate(object);
489 * Remove the object from the global object list.
491 lwkt_gettoken(&ilock, &vm_object_list_token);
492 TAILQ_REMOVE(&vm_object_list, object, object_list);
493 lwkt_reltoken(&ilock);
498 * Free the space for the object.
500 zfree(obj_zone, object);
504 * vm_object_page_clean
506 * Clean all dirty pages in the specified range of object. Leaves page
507 * on whatever queue it is currently on. If NOSYNC is set then do not
508 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
509 * leaving the object dirty.
511 * When stuffing pages asynchronously, allow clustering. XXX we need a
512 * synchronous clustering mode implementation.
514 * Odd semantics: if start == end, we clean everything.
516 * The object must be locked.
520 vm_object_page_clean(object, start, end, flags)
527 vm_offset_t tstart, tend;
535 if (object->type != OBJT_VNODE ||
536 (object->flags & OBJ_MIGHTBEDIRTY) == 0)
539 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
540 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
544 vm_object_set_flag(object, OBJ_CLEANING);
547 * Handle 'entire object' case
557 * If the caller is smart and only msync()s a range he knows is
558 * dirty, we may be able to avoid an object scan. This results in
559 * a phenominal improvement in performance. We cannot do this
560 * as a matter of course because the object may be huge - e.g.
561 * the size might be in the gigabytes or terrabytes.
563 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
568 scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
571 pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
573 scanlimit = scanreset;
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, curgeneration, pagerflags);
612 * If everything was dirty and we flushed it successfully,
613 * and the requested range is not the entire object, we
614 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
615 * return immediately.
617 if (tscan >= tend && (tstart || tend < object->size)) {
618 vm_object_clear_flag(object, OBJ_CLEANING);
621 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
625 * Generally set CLEANCHK interlock and make the page read-only so
626 * we can then clear the object flags.
628 * However, if this is a nosync mmap then the object is likely to
629 * stay dirty so do not mess with the page and do not clear the
635 for(p = TAILQ_FIRST(&object->memq); p; p = TAILQ_NEXT(p, listq)) {
636 vm_page_flag_set(p, PG_CLEANCHK);
637 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
640 vm_page_protect(p, VM_PROT_READ);
643 if (clearobjflags && (tstart == 0) && (tend == object->size)) {
646 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
647 if (object->type == OBJT_VNODE &&
648 (vp = (struct vnode *)object->handle) != NULL) {
649 if (vp->v_flag & VOBJDIRTY) {
650 lwkt_gettoken(&vlock, vp->v_interlock);
651 vp->v_flag &= ~VOBJDIRTY;
652 lwkt_reltoken(&vlock);
658 curgeneration = object->generation;
660 for(p = TAILQ_FIRST(&object->memq); p; p = np) {
663 np = TAILQ_NEXT(p, listq);
667 if (((p->flags & PG_CLEANCHK) == 0) ||
668 (pi < tstart) || (pi >= tend) ||
670 ((p->queue - p->pc) == PQ_CACHE)) {
671 vm_page_flag_clear(p, PG_CLEANCHK);
675 vm_page_test_dirty(p);
676 if ((p->dirty & p->valid) == 0) {
677 vm_page_flag_clear(p, PG_CLEANCHK);
682 * If we have been asked to skip nosync pages and this is a
683 * nosync page, skip it. Note that the object flags were
684 * not cleared in this case so we do not have to set them.
686 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
687 vm_page_flag_clear(p, PG_CLEANCHK);
691 n = vm_object_page_collect_flush(object, p,
692 curgeneration, pagerflags);
695 if (object->generation != curgeneration)
699 * Try to optimize the next page. If we can't we pick up
700 * our (random) scan where we left off.
702 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
703 if ((p = vm_page_lookup(object, pi + n)) != NULL)
709 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
712 vm_object_clear_flag(object, OBJ_CLEANING);
717 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
726 vm_page_t maf[vm_pageout_page_count];
727 vm_page_t mab[vm_pageout_page_count];
728 vm_page_t ma[vm_pageout_page_count];
732 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
733 if (object->generation != curgeneration) {
740 for(i = 1; i < vm_pageout_page_count; i++) {
743 if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
744 if ((tp->flags & PG_BUSY) ||
745 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
746 (tp->flags & PG_CLEANCHK) == 0) ||
749 if((tp->queue - tp->pc) == PQ_CACHE) {
750 vm_page_flag_clear(tp, PG_CLEANCHK);
753 vm_page_test_dirty(tp);
754 if ((tp->dirty & tp->valid) == 0) {
755 vm_page_flag_clear(tp, PG_CLEANCHK);
766 chkb = vm_pageout_page_count - maxf;
768 for(i = 1; i < chkb;i++) {
771 if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
772 if ((tp->flags & PG_BUSY) ||
773 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
774 (tp->flags & PG_CLEANCHK) == 0) ||
777 if((tp->queue - tp->pc) == PQ_CACHE) {
778 vm_page_flag_clear(tp, PG_CLEANCHK);
781 vm_page_test_dirty(tp);
782 if ((tp->dirty & tp->valid) == 0) {
783 vm_page_flag_clear(tp, PG_CLEANCHK);
794 for(i = 0; i < maxb; i++) {
795 int index = (maxb - i) - 1;
797 vm_page_flag_clear(ma[index], PG_CLEANCHK);
799 vm_page_flag_clear(p, PG_CLEANCHK);
801 for(i = 0; i < maxf; i++) {
802 int index = (maxb + i) + 1;
804 vm_page_flag_clear(ma[index], PG_CLEANCHK);
806 runlen = maxb + maxf + 1;
809 vm_pageout_flush(ma, runlen, pagerflags);
810 for (i = 0; i < runlen; i++) {
811 if (ma[i]->valid & ma[i]->dirty) {
812 vm_page_protect(ma[i], VM_PROT_READ);
813 vm_page_flag_set(ma[i], PG_CLEANCHK);
816 * maxf will end up being the actual number of pages
817 * we wrote out contiguously, non-inclusive of the
818 * first page. We do not count look-behind pages.
820 if (i >= maxb + 1 && (maxf > i - maxb - 1))
828 /* XXX I cannot tell if this should be an exported symbol */
830 * vm_object_deactivate_pages
832 * Deactivate all pages in the specified object. (Keep its pages
833 * in memory even though it is no longer referenced.)
835 * The object must be locked.
838 vm_object_deactivate_pages(object)
843 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
844 next = TAILQ_NEXT(p, listq);
845 vm_page_deactivate(p);
851 * Same as vm_object_pmap_copy, except range checking really
852 * works, and is meant for small sections of an object.
854 * This code protects resident pages by making them read-only
855 * and is typically called on a fork or split when a page
856 * is converted to copy-on-write.
858 * NOTE: If the page is already at VM_PROT_NONE, calling
859 * vm_page_protect will have no effect.
863 vm_object_pmap_copy_1(object, start, end)
871 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
874 for (idx = start; idx < end; idx++) {
875 p = vm_page_lookup(object, idx);
878 vm_page_protect(p, VM_PROT_READ);
883 * vm_object_pmap_remove:
885 * Removes all physical pages in the specified
886 * object range from all physical maps.
888 * The object must *not* be locked.
891 vm_object_pmap_remove(object, start, end)
900 for (p = TAILQ_FIRST(&object->memq);
902 p = TAILQ_NEXT(p, listq)) {
903 if (p->pindex >= start && p->pindex < end)
904 vm_page_protect(p, VM_PROT_NONE);
906 if ((start == 0) && (object->size == end))
907 vm_object_clear_flag(object, OBJ_WRITEABLE);
913 * Implements the madvise function at the object/page level.
915 * MADV_WILLNEED (any object)
917 * Activate the specified pages if they are resident.
919 * MADV_DONTNEED (any object)
921 * Deactivate the specified pages if they are resident.
923 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
924 * OBJ_ONEMAPPING only)
926 * Deactivate and clean the specified pages if they are
927 * resident. This permits the process to reuse the pages
928 * without faulting or the kernel to reclaim the pages
932 vm_object_madvise(object, pindex, count, advise)
938 vm_pindex_t end, tpindex;
945 end = pindex + count;
948 * Locate and adjust resident pages
951 for (; pindex < end; pindex += 1) {
957 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
958 * and those pages must be OBJ_ONEMAPPING.
960 if (advise == MADV_FREE) {
961 if ((tobject->type != OBJT_DEFAULT &&
962 tobject->type != OBJT_SWAP) ||
963 (tobject->flags & OBJ_ONEMAPPING) == 0) {
968 m = vm_page_lookup(tobject, tpindex);
972 * There may be swap even if there is no backing page
974 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
975 swap_pager_freespace(tobject, tpindex, 1);
980 tobject = tobject->backing_object;
983 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
988 * If the page is busy or not in a normal active state,
989 * we skip it. If the page is not managed there are no
990 * page queues to mess with. Things can break if we mess
991 * with pages in any of the below states.
996 (m->flags & PG_UNMANAGED) ||
997 m->valid != VM_PAGE_BITS_ALL
1002 if (vm_page_sleep_busy(m, TRUE, "madvpo"))
1005 if (advise == MADV_WILLNEED) {
1006 vm_page_activate(m);
1007 } else if (advise == MADV_DONTNEED) {
1008 vm_page_dontneed(m);
1009 } else if (advise == MADV_FREE) {
1011 * Mark the page clean. This will allow the page
1012 * to be freed up by the system. However, such pages
1013 * are often reused quickly by malloc()/free()
1014 * so we do not do anything that would cause
1015 * a page fault if we can help it.
1017 * Specifically, we do not try to actually free
1018 * the page now nor do we try to put it in the
1019 * cache (which would cause a page fault on reuse).
1021 * But we do make the page is freeable as we
1022 * can without actually taking the step of unmapping
1025 pmap_clear_modify(m);
1028 vm_page_dontneed(m);
1029 if (tobject->type == OBJT_SWAP)
1030 swap_pager_freespace(tobject, tpindex, 1);
1038 * Create a new object which is backed by the
1039 * specified existing object range. The source
1040 * object reference is deallocated.
1042 * The new object and offset into that object
1043 * are returned in the source parameters.
1047 vm_object_shadow(object, offset, length)
1048 vm_object_t *object; /* IN/OUT */
1049 vm_ooffset_t *offset; /* IN/OUT */
1058 * Don't create the new object if the old object isn't shared.
1061 if (source != NULL &&
1062 source->ref_count == 1 &&
1063 source->handle == NULL &&
1064 (source->type == OBJT_DEFAULT ||
1065 source->type == OBJT_SWAP))
1069 * Allocate a new object with the given length
1072 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1073 panic("vm_object_shadow: no object for shadowing");
1076 * The new object shadows the source object, adding a reference to it.
1077 * Our caller changes his reference to point to the new object,
1078 * removing a reference to the source object. Net result: no change
1079 * of reference count.
1081 * Try to optimize the result object's page color when shadowing
1082 * in order to maintain page coloring consistency in the combined
1085 result->backing_object = source;
1087 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1088 source->shadow_count++;
1089 source->generation++;
1090 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1094 * Store the offset into the source object, and fix up the offset into
1098 result->backing_object_offset = *offset;
1101 * Return the new things
1108 #define OBSC_TEST_ALL_SHADOWED 0x0001
1109 #define OBSC_COLLAPSE_NOWAIT 0x0002
1110 #define OBSC_COLLAPSE_WAIT 0x0004
1113 vm_object_backing_scan(vm_object_t object, int op)
1118 vm_object_t backing_object;
1119 vm_pindex_t backing_offset_index;
1123 backing_object = object->backing_object;
1124 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1127 * Initial conditions
1130 if (op & OBSC_TEST_ALL_SHADOWED) {
1132 * We do not want to have to test for the existence of
1133 * swap pages in the backing object. XXX but with the
1134 * new swapper this would be pretty easy to do.
1136 * XXX what about anonymous MAP_SHARED memory that hasn't
1137 * been ZFOD faulted yet? If we do not test for this, the
1138 * shadow test may succeed! XXX
1140 if (backing_object->type != OBJT_DEFAULT) {
1145 if (op & OBSC_COLLAPSE_WAIT) {
1146 vm_object_set_flag(backing_object, OBJ_DEAD);
1153 p = TAILQ_FIRST(&backing_object->memq);
1155 vm_page_t next = TAILQ_NEXT(p, listq);
1156 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1158 if (op & OBSC_TEST_ALL_SHADOWED) {
1162 * Ignore pages outside the parent object's range
1163 * and outside the parent object's mapping of the
1166 * note that we do not busy the backing object's
1171 p->pindex < backing_offset_index ||
1172 new_pindex >= object->size
1179 * See if the parent has the page or if the parent's
1180 * object pager has the page. If the parent has the
1181 * page but the page is not valid, the parent's
1182 * object pager must have the page.
1184 * If this fails, the parent does not completely shadow
1185 * the object and we might as well give up now.
1188 pp = vm_page_lookup(object, new_pindex);
1190 (pp == NULL || pp->valid == 0) &&
1191 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1199 * Check for busy page
1202 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1205 if (op & OBSC_COLLAPSE_NOWAIT) {
1207 (p->flags & PG_BUSY) ||
1216 } else if (op & OBSC_COLLAPSE_WAIT) {
1217 if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1219 * If we slept, anything could have
1220 * happened. Since the object is
1221 * marked dead, the backing offset
1222 * should not have changed so we
1223 * just restart our scan.
1225 p = TAILQ_FIRST(&backing_object->memq);
1236 p->object == backing_object,
1237 ("vm_object_qcollapse(): object mismatch")
1241 * Destroy any associated swap
1243 if (backing_object->type == OBJT_SWAP) {
1244 swap_pager_freespace(
1252 p->pindex < backing_offset_index ||
1253 new_pindex >= object->size
1256 * Page is out of the parent object's range, we
1257 * can simply destroy it.
1259 vm_page_protect(p, VM_PROT_NONE);
1265 pp = vm_page_lookup(object, new_pindex);
1268 vm_pager_has_page(object, new_pindex, NULL, NULL)
1271 * page already exists in parent OR swap exists
1272 * for this location in the parent. Destroy
1273 * the original page from the backing object.
1275 * Leave the parent's page alone
1277 vm_page_protect(p, VM_PROT_NONE);
1284 * Page does not exist in parent, rename the
1285 * page from the backing object to the main object.
1287 * If the page was mapped to a process, it can remain
1288 * mapped through the rename.
1290 if ((p->queue - p->pc) == PQ_CACHE)
1291 vm_page_deactivate(p);
1293 vm_page_rename(p, object, new_pindex);
1294 /* page automatically made dirty by rename */
1304 * this version of collapse allows the operation to occur earlier and
1305 * when paging_in_progress is true for an object... This is not a complete
1306 * operation, but should plug 99.9% of the rest of the leaks.
1309 vm_object_qcollapse(object)
1312 vm_object_t backing_object = object->backing_object;
1314 if (backing_object->ref_count != 1)
1317 backing_object->ref_count += 2;
1319 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1321 backing_object->ref_count -= 2;
1325 * vm_object_collapse:
1327 * Collapse an object with the object backing it.
1328 * Pages in the backing object are moved into the
1329 * parent, and the backing object is deallocated.
1332 vm_object_collapse(object)
1336 vm_object_t backing_object;
1339 * Verify that the conditions are right for collapse:
1341 * The object exists and the backing object exists.
1346 if ((backing_object = object->backing_object) == NULL)
1350 * we check the backing object first, because it is most likely
1353 if (backing_object->handle != NULL ||
1354 (backing_object->type != OBJT_DEFAULT &&
1355 backing_object->type != OBJT_SWAP) ||
1356 (backing_object->flags & OBJ_DEAD) ||
1357 object->handle != NULL ||
1358 (object->type != OBJT_DEFAULT &&
1359 object->type != OBJT_SWAP) ||
1360 (object->flags & OBJ_DEAD)) {
1365 object->paging_in_progress != 0 ||
1366 backing_object->paging_in_progress != 0
1368 vm_object_qcollapse(object);
1373 * We know that we can either collapse the backing object (if
1374 * the parent is the only reference to it) or (perhaps) have
1375 * the parent bypass the object if the parent happens to shadow
1376 * all the resident pages in the entire backing object.
1378 * This is ignoring pager-backed pages such as swap pages.
1379 * vm_object_backing_scan fails the shadowing test in this
1383 if (backing_object->ref_count == 1) {
1385 * If there is exactly one reference to the backing
1386 * object, we can collapse it into the parent.
1389 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1392 * Move the pager from backing_object to object.
1395 if (backing_object->type == OBJT_SWAP) {
1396 vm_object_pip_add(backing_object, 1);
1399 * scrap the paging_offset junk and do a
1400 * discrete copy. This also removes major
1401 * assumptions about how the swap-pager
1402 * works from where it doesn't belong. The
1403 * new swapper is able to optimize the
1404 * destroy-source case.
1407 vm_object_pip_add(object, 1);
1411 OFF_TO_IDX(object->backing_object_offset), TRUE);
1412 vm_object_pip_wakeup(object);
1414 vm_object_pip_wakeup(backing_object);
1417 * Object now shadows whatever backing_object did.
1418 * Note that the reference to
1419 * backing_object->backing_object moves from within
1420 * backing_object to within object.
1423 LIST_REMOVE(object, shadow_list);
1424 object->backing_object->shadow_count--;
1425 object->backing_object->generation++;
1426 if (backing_object->backing_object) {
1427 LIST_REMOVE(backing_object, shadow_list);
1428 backing_object->backing_object->shadow_count--;
1429 backing_object->backing_object->generation++;
1431 object->backing_object = backing_object->backing_object;
1432 if (object->backing_object) {
1434 &object->backing_object->shadow_head,
1438 object->backing_object->shadow_count++;
1439 object->backing_object->generation++;
1442 object->backing_object_offset +=
1443 backing_object->backing_object_offset;
1446 * Discard backing_object.
1448 * Since the backing object has no pages, no pager left,
1449 * and no object references within it, all that is
1450 * necessary is to dispose of it.
1453 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1454 KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object));
1462 zfree(obj_zone, backing_object);
1466 vm_object_t new_backing_object;
1469 * If we do not entirely shadow the backing object,
1470 * there is nothing we can do so we give up.
1473 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1478 * Make the parent shadow the next object in the
1479 * chain. Deallocating backing_object will not remove
1480 * it, since its reference count is at least 2.
1483 LIST_REMOVE(object, shadow_list);
1484 backing_object->shadow_count--;
1485 backing_object->generation++;
1487 new_backing_object = backing_object->backing_object;
1488 if ((object->backing_object = new_backing_object) != NULL) {
1489 vm_object_reference(new_backing_object);
1491 &new_backing_object->shadow_head,
1495 new_backing_object->shadow_count++;
1496 new_backing_object->generation++;
1497 object->backing_object_offset +=
1498 backing_object->backing_object_offset;
1502 * Drop the reference count on backing_object. Since
1503 * its ref_count was at least 2, it will not vanish;
1504 * so we don't need to call vm_object_deallocate, but
1507 vm_object_deallocate(backing_object);
1512 * Try again with this object's new backing object.
1518 * vm_object_page_remove: [internal]
1520 * Removes all physical pages in the specified
1521 * object range from the object's list of pages.
1523 * The object must be locked.
1526 vm_object_page_remove(object, start, end, clean_only)
1530 boolean_t clean_only;
1536 if (object == NULL ||
1537 object->resident_page_count == 0)
1540 all = ((end == 0) && (start == 0));
1543 * Since physically-backed objects do not use managed pages, we can't
1544 * remove pages from the object (we must instead remove the page
1545 * references, and then destroy the object).
1547 KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object"));
1549 vm_object_pip_add(object, 1);
1552 if (all || size > object->resident_page_count / 4) {
1553 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
1554 next = TAILQ_NEXT(p, listq);
1555 if (all || ((start <= p->pindex) && (p->pindex < end))) {
1556 if (p->wire_count != 0) {
1557 vm_page_protect(p, VM_PROT_NONE);
1564 * The busy flags are only cleared at
1565 * interrupt -- minimize the spl transitions
1568 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1571 if (clean_only && p->valid) {
1572 vm_page_test_dirty(p);
1573 if (p->valid & p->dirty)
1578 vm_page_protect(p, VM_PROT_NONE);
1584 if ((p = vm_page_lookup(object, start)) != 0) {
1586 if (p->wire_count != 0) {
1587 vm_page_protect(p, VM_PROT_NONE);
1596 * The busy flags are only cleared at
1597 * interrupt -- minimize the spl transitions
1599 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1602 if (clean_only && p->valid) {
1603 vm_page_test_dirty(p);
1604 if (p->valid & p->dirty) {
1612 vm_page_protect(p, VM_PROT_NONE);
1619 vm_object_pip_wakeup(object);
1623 * Routine: vm_object_coalesce
1624 * Function: Coalesces two objects backing up adjoining
1625 * regions of memory into a single object.
1627 * returns TRUE if objects were combined.
1629 * NOTE: Only works at the moment if the second object is NULL -
1630 * if it's not, which object do we lock first?
1633 * prev_object First object to coalesce
1634 * prev_offset Offset into prev_object
1635 * next_object Second object into coalesce
1636 * next_offset Offset into next_object
1638 * prev_size Size of reference to prev_object
1639 * next_size Size of reference to next_object
1642 * The object must *not* be locked.
1645 vm_object_coalesce(prev_object, prev_pindex, prev_size, next_size)
1646 vm_object_t prev_object;
1647 vm_pindex_t prev_pindex;
1648 vm_size_t prev_size, next_size;
1650 vm_pindex_t next_pindex;
1652 if (prev_object == NULL) {
1656 if (prev_object->type != OBJT_DEFAULT &&
1657 prev_object->type != OBJT_SWAP) {
1662 * Try to collapse the object first
1664 vm_object_collapse(prev_object);
1667 * Can't coalesce if: . more than one reference . paged out . shadows
1668 * another object . has a copy elsewhere (any of which mean that the
1669 * pages not mapped to prev_entry may be in use anyway)
1672 if (prev_object->backing_object != NULL) {
1676 prev_size >>= PAGE_SHIFT;
1677 next_size >>= PAGE_SHIFT;
1678 next_pindex = prev_pindex + prev_size;
1680 if ((prev_object->ref_count > 1) &&
1681 (prev_object->size != next_pindex)) {
1686 * Remove any pages that may still be in the object from a previous
1689 if (next_pindex < prev_object->size) {
1690 vm_object_page_remove(prev_object,
1692 next_pindex + next_size, FALSE);
1693 if (prev_object->type == OBJT_SWAP)
1694 swap_pager_freespace(prev_object,
1695 next_pindex, next_size);
1699 * Extend the object if necessary.
1701 if (next_pindex + next_size > prev_object->size)
1702 prev_object->size = next_pindex + next_size;
1708 vm_object_set_writeable_dirty(vm_object_t object)
1713 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1714 if (object->type == OBJT_VNODE &&
1715 (vp = (struct vnode *)object->handle) != NULL) {
1716 if ((vp->v_flag & VOBJDIRTY) == 0) {
1717 lwkt_gettoken(&vlock, vp->v_interlock);
1718 vp->v_flag |= VOBJDIRTY;
1719 lwkt_reltoken(&vlock);
1726 #include "opt_ddb.h"
1728 #include <sys/kernel.h>
1730 #include <sys/cons.h>
1732 #include <ddb/ddb.h>
1734 static int _vm_object_in_map (vm_map_t map, vm_object_t object,
1735 vm_map_entry_t entry);
1736 static int vm_object_in_map (vm_object_t object);
1739 _vm_object_in_map(map, object, entry)
1742 vm_map_entry_t entry;
1745 vm_map_entry_t tmpe;
1753 tmpe = map->header.next;
1754 entcount = map->nentries;
1755 while (entcount-- && (tmpe != &map->header)) {
1756 if( _vm_object_in_map(map, object, tmpe)) {
1761 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
1762 tmpm = entry->object.sub_map;
1763 tmpe = tmpm->header.next;
1764 entcount = tmpm->nentries;
1765 while (entcount-- && tmpe != &tmpm->header) {
1766 if( _vm_object_in_map(tmpm, object, tmpe)) {
1771 } else if ((obj = entry->object.vm_object) != NULL) {
1772 for(; obj; obj=obj->backing_object)
1773 if( obj == object) {
1781 vm_object_in_map( object)
1785 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1786 if( !p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1788 if( _vm_object_in_map(&p->p_vmspace->vm_map, object, 0))
1791 if( _vm_object_in_map( kernel_map, object, 0))
1793 if( _vm_object_in_map( pager_map, object, 0))
1795 if( _vm_object_in_map( buffer_map, object, 0))
1797 if( _vm_object_in_map( mb_map, object, 0))
1802 DB_SHOW_COMMAND(vmochk, vm_object_check)
1807 * make sure that internal objs are in a map somewhere
1808 * and none have zero ref counts.
1810 for (object = TAILQ_FIRST(&vm_object_list);
1812 object = TAILQ_NEXT(object, object_list)) {
1813 if (object->handle == NULL &&
1814 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1815 if (object->ref_count == 0) {
1816 db_printf("vmochk: internal obj has zero ref count: %ld\n",
1817 (long)object->size);
1819 if (!vm_object_in_map(object)) {
1821 "vmochk: internal obj is not in a map: "
1822 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1823 object->ref_count, (u_long)object->size,
1824 (u_long)object->size,
1825 (void *)object->backing_object);
1832 * vm_object_print: [ debug ]
1834 DB_SHOW_COMMAND(object, vm_object_print_static)
1836 /* XXX convert args. */
1837 vm_object_t object = (vm_object_t)addr;
1838 boolean_t full = have_addr;
1842 /* XXX count is an (unused) arg. Avoid shadowing it. */
1843 #define count was_count
1851 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1852 object, (int)object->type, (u_long)object->size,
1853 object->resident_page_count, object->ref_count, object->flags);
1855 * XXX no %qd in kernel. Truncate object->backing_object_offset.
1857 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1858 object->shadow_count,
1859 object->backing_object ? object->backing_object->ref_count : 0,
1860 object->backing_object, (long)object->backing_object_offset);
1867 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = TAILQ_NEXT(p, listq)) {
1869 db_iprintf("memory:=");
1870 else if (count == 6) {
1878 db_printf("(off=0x%lx,page=0x%lx)",
1879 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1889 /* XXX need this non-static entry for calling from vm_map_print. */
1891 vm_object_print(addr, have_addr, count, modif)
1892 /* db_expr_t */ long addr;
1893 boolean_t have_addr;
1894 /* db_expr_t */ long count;
1897 vm_object_print_static(addr, have_addr, count, modif);
1900 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1905 for (object = TAILQ_FIRST(&vm_object_list);
1907 object = TAILQ_NEXT(object, object_list)) {
1908 vm_pindex_t idx, fidx;
1910 vm_paddr_t pa = -1, padiff;
1914 db_printf("new object: %p\n", (void *)object);
1924 osize = object->size;
1927 for(idx=0;idx<osize;idx++) {
1928 m = vm_page_lookup(object, idx);
1931 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1932 (long)fidx, rcount, (long)pa);
1947 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1952 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1953 padiff >>= PAGE_SHIFT;
1954 padiff &= PQ_L2_MASK;
1956 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1960 db_printf(" index(%ld)run(%d)pa(0x%lx)",
1961 (long)fidx, rcount, (long)pa);
1962 db_printf("pd(%ld)\n", (long)padiff);
1972 pa = VM_PAGE_TO_PHYS(m);
1976 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1977 (long)fidx, rcount, (long)pa);