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 1.10 2003/10/19 00:23:30 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_inittoken(&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);
221 zinitna(obj_zone, NULL, NULL, 0, 0, 0, 1);
225 * vm_object_allocate:
227 * Returns a new object with the given size.
231 vm_object_allocate(type, size)
237 result = (vm_object_t) zalloc(obj_zone);
239 _vm_object_allocate(type, size, result);
246 * vm_object_reference:
248 * Gets another reference to the given object.
251 vm_object_reference(object)
258 /* object can be re-referenced during final cleaning */
259 KASSERT(!(object->flags & OBJ_DEAD),
260 ("vm_object_reference: attempting to reference dead obj"));
264 if (object->type == OBJT_VNODE) {
265 while (vget((struct vnode *) object->handle, LK_RETRY|LK_NOOBJ, curthread)) {
266 printf("vm_object_reference: delay in getting object\n");
272 vm_object_vndeallocate(object)
275 struct vnode *vp = (struct vnode *) object->handle;
277 KASSERT(object->type == OBJT_VNODE,
278 ("vm_object_vndeallocate: not a vnode object"));
279 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
281 if (object->ref_count == 0) {
282 vprint("vm_object_vndeallocate", vp);
283 panic("vm_object_vndeallocate: bad object reference count");
288 if (object->ref_count == 0) {
289 vp->v_flag &= ~VTEXT;
290 vm_object_clear_flag(object, OBJ_OPT);
296 * vm_object_deallocate:
298 * Release a reference to the specified object,
299 * gained either through a vm_object_allocate
300 * or a vm_object_reference call. When all references
301 * are gone, storage associated with this object
302 * may be relinquished.
304 * No object may be locked.
307 vm_object_deallocate(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(object)
418 * Make sure no one uses us.
420 vm_object_set_flag(object, OBJ_DEAD);
423 * wait for the pageout daemon to be done with the object
425 vm_object_pip_wait(object, "objtrm");
427 KASSERT(!object->paging_in_progress,
428 ("vm_object_terminate: pageout in progress"));
431 * Clean and free the pages, as appropriate. All references to the
432 * object are gone, so we don't need to lock it.
434 if (object->type == OBJT_VNODE) {
438 * Freeze optimized copies.
440 vm_freeze_copyopts(object, 0, object->size);
443 * Clean pages and flush buffers.
445 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
447 vp = (struct vnode *) object->handle;
448 vinvalbuf(vp, V_SAVE, NULL, 0, 0);
452 * Wait for any I/O to complete, after which there had better not
453 * be any references left on the object.
455 vm_object_pip_wait(object, "objtrm");
457 if (object->ref_count != 0)
458 panic("vm_object_terminate: object with references, ref_count=%d", object->ref_count);
461 * Now free any remaining pages. For internal objects, this also
462 * removes them from paging queues. Don't free wired pages, just
463 * remove them from the object.
466 while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
467 if (p->busy || (p->flags & PG_BUSY))
468 panic("vm_object_terminate: freeing busy page %p\n", p);
469 if (p->wire_count == 0) {
472 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(&vm_object_list_token);
489 TAILQ_REMOVE(&vm_object_list, object, object_list);
490 lwkt_reltoken(&vm_object_list_token);
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.
513 * The object must be locked.
517 vm_object_page_clean(object, start, end, flags)
524 vm_offset_t tstart, tend;
531 if (object->type != OBJT_VNODE ||
532 (object->flags & OBJ_MIGHTBEDIRTY) == 0)
535 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
536 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
540 vm_object_set_flag(object, OBJ_CLEANING);
543 * Handle 'entire object' case
553 * If the caller is smart and only msync()s a range he knows is
554 * dirty, we may be able to avoid an object scan. This results in
555 * a phenominal improvement in performance. We cannot do this
556 * as a matter of course because the object may be huge - e.g.
557 * the size might be in the gigabytes or terrabytes.
559 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
564 scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
567 pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
569 scanlimit = scanreset;
571 while (tscan < tend) {
572 curgeneration = object->generation;
573 p = vm_page_lookup(object, tscan);
574 if (p == NULL || p->valid == 0 ||
575 (p->queue - p->pc) == PQ_CACHE) {
576 if (--scanlimit == 0)
581 vm_page_test_dirty(p);
582 if ((p->dirty & p->valid) == 0) {
583 if (--scanlimit == 0)
589 * If we have been asked to skip nosync pages and
590 * this is a nosync page, we can't continue.
592 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
593 if (--scanlimit == 0)
598 scanlimit = scanreset;
601 * This returns 0 if it was unable to busy the first
602 * page (i.e. had to sleep).
604 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
608 * If everything was dirty and we flushed it successfully,
609 * and the requested range is not the entire object, we
610 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
611 * return immediately.
613 if (tscan >= tend && (tstart || tend < object->size)) {
614 vm_object_clear_flag(object, OBJ_CLEANING);
617 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
621 * Generally set CLEANCHK interlock and make the page read-only so
622 * we can then clear the object flags.
624 * However, if this is a nosync mmap then the object is likely to
625 * stay dirty so do not mess with the page and do not clear the
631 for(p = TAILQ_FIRST(&object->memq); p; p = TAILQ_NEXT(p, listq)) {
632 vm_page_flag_set(p, PG_CLEANCHK);
633 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
636 vm_page_protect(p, VM_PROT_READ);
639 if (clearobjflags && (tstart == 0) && (tend == object->size)) {
642 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
643 if (object->type == OBJT_VNODE &&
644 (vp = (struct vnode *)object->handle) != NULL) {
645 if (vp->v_flag & VOBJDIRTY) {
646 lwkt_gettoken(&vp->v_interlock);
647 vp->v_flag &= ~VOBJDIRTY;
648 lwkt_reltoken(&vp->v_interlock);
654 curgeneration = object->generation;
656 for(p = TAILQ_FIRST(&object->memq); p; p = np) {
659 np = TAILQ_NEXT(p, listq);
663 if (((p->flags & PG_CLEANCHK) == 0) ||
664 (pi < tstart) || (pi >= tend) ||
666 ((p->queue - p->pc) == PQ_CACHE)) {
667 vm_page_flag_clear(p, PG_CLEANCHK);
671 vm_page_test_dirty(p);
672 if ((p->dirty & p->valid) == 0) {
673 vm_page_flag_clear(p, PG_CLEANCHK);
678 * If we have been asked to skip nosync pages and this is a
679 * nosync page, skip it. Note that the object flags were
680 * not cleared in this case so we do not have to set them.
682 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
683 vm_page_flag_clear(p, PG_CLEANCHK);
687 n = vm_object_page_collect_flush(object, p,
688 curgeneration, pagerflags);
691 if (object->generation != curgeneration)
695 * Try to optimize the next page. If we can't we pick up
696 * our (random) scan where we left off.
698 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
699 if ((p = vm_page_lookup(object, pi + n)) != NULL)
705 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
708 vm_object_clear_flag(object, OBJ_CLEANING);
713 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
722 vm_page_t maf[vm_pageout_page_count];
723 vm_page_t mab[vm_pageout_page_count];
724 vm_page_t ma[vm_pageout_page_count];
728 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
729 if (object->generation != curgeneration) {
736 for(i = 1; i < vm_pageout_page_count; i++) {
739 if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
740 if ((tp->flags & PG_BUSY) ||
741 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
742 (tp->flags & PG_CLEANCHK) == 0) ||
745 if((tp->queue - tp->pc) == PQ_CACHE) {
746 vm_page_flag_clear(tp, PG_CLEANCHK);
749 vm_page_test_dirty(tp);
750 if ((tp->dirty & tp->valid) == 0) {
751 vm_page_flag_clear(tp, PG_CLEANCHK);
762 chkb = vm_pageout_page_count - maxf;
764 for(i = 1; i < chkb;i++) {
767 if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
768 if ((tp->flags & PG_BUSY) ||
769 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
770 (tp->flags & PG_CLEANCHK) == 0) ||
773 if((tp->queue - tp->pc) == PQ_CACHE) {
774 vm_page_flag_clear(tp, PG_CLEANCHK);
777 vm_page_test_dirty(tp);
778 if ((tp->dirty & tp->valid) == 0) {
779 vm_page_flag_clear(tp, PG_CLEANCHK);
790 for(i = 0; i < maxb; i++) {
791 int index = (maxb - i) - 1;
793 vm_page_flag_clear(ma[index], PG_CLEANCHK);
795 vm_page_flag_clear(p, PG_CLEANCHK);
797 for(i = 0; i < maxf; i++) {
798 int index = (maxb + i) + 1;
800 vm_page_flag_clear(ma[index], PG_CLEANCHK);
802 runlen = maxb + maxf + 1;
805 vm_pageout_flush(ma, runlen, pagerflags);
806 for (i = 0; i < runlen; i++) {
807 if (ma[i]->valid & ma[i]->dirty) {
808 vm_page_protect(ma[i], VM_PROT_READ);
809 vm_page_flag_set(ma[i], PG_CLEANCHK);
812 * maxf will end up being the actual number of pages
813 * we wrote out contiguously, non-inclusive of the
814 * first page. We do not count look-behind pages.
816 if (i >= maxb + 1 && (maxf > i - maxb - 1))
824 /* XXX I cannot tell if this should be an exported symbol */
826 * vm_object_deactivate_pages
828 * Deactivate all pages in the specified object. (Keep its pages
829 * in memory even though it is no longer referenced.)
831 * The object must be locked.
834 vm_object_deactivate_pages(object)
839 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
840 next = TAILQ_NEXT(p, listq);
841 vm_page_deactivate(p);
847 * Same as vm_object_pmap_copy, except range checking really
848 * works, and is meant for small sections of an object.
850 * This code protects resident pages by making them read-only
851 * and is typically called on a fork or split when a page
852 * is converted to copy-on-write.
854 * NOTE: If the page is already at VM_PROT_NONE, calling
855 * vm_page_protect will have no effect.
859 vm_object_pmap_copy_1(object, start, end)
867 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
870 for (idx = start; idx < end; idx++) {
871 p = vm_page_lookup(object, idx);
874 vm_page_protect(p, VM_PROT_READ);
879 * vm_object_pmap_remove:
881 * Removes all physical pages in the specified
882 * object range from all physical maps.
884 * The object must *not* be locked.
887 vm_object_pmap_remove(object, start, end)
896 for (p = TAILQ_FIRST(&object->memq);
898 p = TAILQ_NEXT(p, listq)) {
899 if (p->pindex >= start && p->pindex < end)
900 vm_page_protect(p, VM_PROT_NONE);
902 if ((start == 0) && (object->size == end))
903 vm_object_clear_flag(object, OBJ_WRITEABLE);
909 * Implements the madvise function at the object/page level.
911 * MADV_WILLNEED (any object)
913 * Activate the specified pages if they are resident.
915 * MADV_DONTNEED (any object)
917 * Deactivate the specified pages if they are resident.
919 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
920 * OBJ_ONEMAPPING only)
922 * Deactivate and clean the specified pages if they are
923 * resident. This permits the process to reuse the pages
924 * without faulting or the kernel to reclaim the pages
928 vm_object_madvise(object, pindex, count, advise)
934 vm_pindex_t end, tpindex;
941 end = pindex + count;
944 * Locate and adjust resident pages
947 for (; pindex < end; pindex += 1) {
953 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
954 * and those pages must be OBJ_ONEMAPPING.
956 if (advise == MADV_FREE) {
957 if ((tobject->type != OBJT_DEFAULT &&
958 tobject->type != OBJT_SWAP) ||
959 (tobject->flags & OBJ_ONEMAPPING) == 0) {
964 m = vm_page_lookup(tobject, tpindex);
968 * There may be swap even if there is no backing page
970 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
971 swap_pager_freespace(tobject, tpindex, 1);
976 tobject = tobject->backing_object;
979 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
984 * If the page is busy or not in a normal active state,
985 * we skip it. If the page is not managed there are no
986 * page queues to mess with. Things can break if we mess
987 * with pages in any of the below states.
992 (m->flags & PG_UNMANAGED) ||
993 m->valid != VM_PAGE_BITS_ALL
998 if (vm_page_sleep_busy(m, TRUE, "madvpo"))
1001 if (advise == MADV_WILLNEED) {
1002 vm_page_activate(m);
1003 } else if (advise == MADV_DONTNEED) {
1004 vm_page_dontneed(m);
1005 } else if (advise == MADV_FREE) {
1007 * Mark the page clean. This will allow the page
1008 * to be freed up by the system. However, such pages
1009 * are often reused quickly by malloc()/free()
1010 * so we do not do anything that would cause
1011 * a page fault if we can help it.
1013 * Specifically, we do not try to actually free
1014 * the page now nor do we try to put it in the
1015 * cache (which would cause a page fault on reuse).
1017 * But we do make the page is freeable as we
1018 * can without actually taking the step of unmapping
1021 pmap_clear_modify(m);
1024 vm_page_dontneed(m);
1025 if (tobject->type == OBJT_SWAP)
1026 swap_pager_freespace(tobject, tpindex, 1);
1034 * Create a new object which is backed by the
1035 * specified existing object range. The source
1036 * object reference is deallocated.
1038 * The new object and offset into that object
1039 * are returned in the source parameters.
1043 vm_object_shadow(object, offset, length)
1044 vm_object_t *object; /* IN/OUT */
1045 vm_ooffset_t *offset; /* IN/OUT */
1054 * Don't create the new object if the old object isn't shared.
1057 if (source != NULL &&
1058 source->ref_count == 1 &&
1059 source->handle == NULL &&
1060 (source->type == OBJT_DEFAULT ||
1061 source->type == OBJT_SWAP))
1065 * Allocate a new object with the given length
1068 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1069 panic("vm_object_shadow: no object for shadowing");
1072 * The new object shadows the source object, adding a reference to it.
1073 * Our caller changes his reference to point to the new object,
1074 * removing a reference to the source object. Net result: no change
1075 * of reference count.
1077 * Try to optimize the result object's page color when shadowing
1078 * in order to maintain page coloring consistency in the combined
1081 result->backing_object = source;
1083 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1084 source->shadow_count++;
1085 source->generation++;
1086 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1090 * Store the offset into the source object, and fix up the offset into
1094 result->backing_object_offset = *offset;
1097 * Return the new things
1104 #define OBSC_TEST_ALL_SHADOWED 0x0001
1105 #define OBSC_COLLAPSE_NOWAIT 0x0002
1106 #define OBSC_COLLAPSE_WAIT 0x0004
1109 vm_object_backing_scan(vm_object_t object, int op)
1114 vm_object_t backing_object;
1115 vm_pindex_t backing_offset_index;
1119 backing_object = object->backing_object;
1120 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1123 * Initial conditions
1126 if (op & OBSC_TEST_ALL_SHADOWED) {
1128 * We do not want to have to test for the existence of
1129 * swap pages in the backing object. XXX but with the
1130 * new swapper this would be pretty easy to do.
1132 * XXX what about anonymous MAP_SHARED memory that hasn't
1133 * been ZFOD faulted yet? If we do not test for this, the
1134 * shadow test may succeed! XXX
1136 if (backing_object->type != OBJT_DEFAULT) {
1141 if (op & OBSC_COLLAPSE_WAIT) {
1142 vm_object_set_flag(backing_object, OBJ_DEAD);
1149 p = TAILQ_FIRST(&backing_object->memq);
1151 vm_page_t next = TAILQ_NEXT(p, listq);
1152 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1154 if (op & OBSC_TEST_ALL_SHADOWED) {
1158 * Ignore pages outside the parent object's range
1159 * and outside the parent object's mapping of the
1162 * note that we do not busy the backing object's
1167 p->pindex < backing_offset_index ||
1168 new_pindex >= object->size
1175 * See if the parent has the page or if the parent's
1176 * object pager has the page. If the parent has the
1177 * page but the page is not valid, the parent's
1178 * object pager must have the page.
1180 * If this fails, the parent does not completely shadow
1181 * the object and we might as well give up now.
1184 pp = vm_page_lookup(object, new_pindex);
1186 (pp == NULL || pp->valid == 0) &&
1187 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1195 * Check for busy page
1198 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1201 if (op & OBSC_COLLAPSE_NOWAIT) {
1203 (p->flags & PG_BUSY) ||
1212 } else if (op & OBSC_COLLAPSE_WAIT) {
1213 if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1215 * If we slept, anything could have
1216 * happened. Since the object is
1217 * marked dead, the backing offset
1218 * should not have changed so we
1219 * just restart our scan.
1221 p = TAILQ_FIRST(&backing_object->memq);
1232 p->object == backing_object,
1233 ("vm_object_qcollapse(): object mismatch")
1237 * Destroy any associated swap
1239 if (backing_object->type == OBJT_SWAP) {
1240 swap_pager_freespace(
1248 p->pindex < backing_offset_index ||
1249 new_pindex >= object->size
1252 * Page is out of the parent object's range, we
1253 * can simply destroy it.
1255 vm_page_protect(p, VM_PROT_NONE);
1261 pp = vm_page_lookup(object, new_pindex);
1264 vm_pager_has_page(object, new_pindex, NULL, NULL)
1267 * page already exists in parent OR swap exists
1268 * for this location in the parent. Destroy
1269 * the original page from the backing object.
1271 * Leave the parent's page alone
1273 vm_page_protect(p, VM_PROT_NONE);
1280 * Page does not exist in parent, rename the
1281 * page from the backing object to the main object.
1283 * If the page was mapped to a process, it can remain
1284 * mapped through the rename.
1286 if ((p->queue - p->pc) == PQ_CACHE)
1287 vm_page_deactivate(p);
1289 vm_page_rename(p, object, new_pindex);
1290 /* page automatically made dirty by rename */
1300 * this version of collapse allows the operation to occur earlier and
1301 * when paging_in_progress is true for an object... This is not a complete
1302 * operation, but should plug 99.9% of the rest of the leaks.
1305 vm_object_qcollapse(object)
1308 vm_object_t backing_object = object->backing_object;
1310 if (backing_object->ref_count != 1)
1313 backing_object->ref_count += 2;
1315 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1317 backing_object->ref_count -= 2;
1321 * vm_object_collapse:
1323 * Collapse an object with the object backing it.
1324 * Pages in the backing object are moved into the
1325 * parent, and the backing object is deallocated.
1328 vm_object_collapse(object)
1332 vm_object_t backing_object;
1335 * Verify that the conditions are right for collapse:
1337 * The object exists and the backing object exists.
1342 if ((backing_object = object->backing_object) == NULL)
1346 * we check the backing object first, because it is most likely
1349 if (backing_object->handle != NULL ||
1350 (backing_object->type != OBJT_DEFAULT &&
1351 backing_object->type != OBJT_SWAP) ||
1352 (backing_object->flags & OBJ_DEAD) ||
1353 object->handle != NULL ||
1354 (object->type != OBJT_DEFAULT &&
1355 object->type != OBJT_SWAP) ||
1356 (object->flags & OBJ_DEAD)) {
1361 object->paging_in_progress != 0 ||
1362 backing_object->paging_in_progress != 0
1364 vm_object_qcollapse(object);
1369 * We know that we can either collapse the backing object (if
1370 * the parent is the only reference to it) or (perhaps) have
1371 * the parent bypass the object if the parent happens to shadow
1372 * all the resident pages in the entire backing object.
1374 * This is ignoring pager-backed pages such as swap pages.
1375 * vm_object_backing_scan fails the shadowing test in this
1379 if (backing_object->ref_count == 1) {
1381 * If there is exactly one reference to the backing
1382 * object, we can collapse it into the parent.
1385 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1388 * Move the pager from backing_object to object.
1391 if (backing_object->type == OBJT_SWAP) {
1392 vm_object_pip_add(backing_object, 1);
1395 * scrap the paging_offset junk and do a
1396 * discrete copy. This also removes major
1397 * assumptions about how the swap-pager
1398 * works from where it doesn't belong. The
1399 * new swapper is able to optimize the
1400 * destroy-source case.
1403 vm_object_pip_add(object, 1);
1407 OFF_TO_IDX(object->backing_object_offset), TRUE);
1408 vm_object_pip_wakeup(object);
1410 vm_object_pip_wakeup(backing_object);
1413 * Object now shadows whatever backing_object did.
1414 * Note that the reference to
1415 * backing_object->backing_object moves from within
1416 * backing_object to within object.
1419 LIST_REMOVE(object, shadow_list);
1420 object->backing_object->shadow_count--;
1421 object->backing_object->generation++;
1422 if (backing_object->backing_object) {
1423 LIST_REMOVE(backing_object, shadow_list);
1424 backing_object->backing_object->shadow_count--;
1425 backing_object->backing_object->generation++;
1427 object->backing_object = backing_object->backing_object;
1428 if (object->backing_object) {
1430 &object->backing_object->shadow_head,
1434 object->backing_object->shadow_count++;
1435 object->backing_object->generation++;
1438 object->backing_object_offset +=
1439 backing_object->backing_object_offset;
1442 * Discard backing_object.
1444 * Since the backing object has no pages, no pager left,
1445 * and no object references within it, all that is
1446 * necessary is to dispose of it.
1449 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1450 KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object));
1458 zfree(obj_zone, backing_object);
1462 vm_object_t new_backing_object;
1465 * If we do not entirely shadow the backing object,
1466 * there is nothing we can do so we give up.
1469 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1474 * Make the parent shadow the next object in the
1475 * chain. Deallocating backing_object will not remove
1476 * it, since its reference count is at least 2.
1479 LIST_REMOVE(object, shadow_list);
1480 backing_object->shadow_count--;
1481 backing_object->generation++;
1483 new_backing_object = backing_object->backing_object;
1484 if ((object->backing_object = new_backing_object) != NULL) {
1485 vm_object_reference(new_backing_object);
1487 &new_backing_object->shadow_head,
1491 new_backing_object->shadow_count++;
1492 new_backing_object->generation++;
1493 object->backing_object_offset +=
1494 backing_object->backing_object_offset;
1498 * Drop the reference count on backing_object. Since
1499 * its ref_count was at least 2, it will not vanish;
1500 * so we don't need to call vm_object_deallocate, but
1503 vm_object_deallocate(backing_object);
1508 * Try again with this object's new backing object.
1514 * vm_object_page_remove: [internal]
1516 * Removes all physical pages in the specified
1517 * object range from the object's list of pages.
1519 * The object must be locked.
1522 vm_object_page_remove(object, start, end, clean_only)
1526 boolean_t clean_only;
1532 if (object == NULL ||
1533 object->resident_page_count == 0)
1536 all = ((end == 0) && (start == 0));
1539 * Since physically-backed objects do not use managed pages, we can't
1540 * remove pages from the object (we must instead remove the page
1541 * references, and then destroy the object).
1543 KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object"));
1545 vm_object_pip_add(object, 1);
1548 if (all || size > object->resident_page_count / 4) {
1549 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
1550 next = TAILQ_NEXT(p, listq);
1551 if (all || ((start <= p->pindex) && (p->pindex < end))) {
1552 if (p->wire_count != 0) {
1553 vm_page_protect(p, VM_PROT_NONE);
1560 * The busy flags are only cleared at
1561 * interrupt -- minimize the spl transitions
1564 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1567 if (clean_only && p->valid) {
1568 vm_page_test_dirty(p);
1569 if (p->valid & p->dirty)
1574 vm_page_protect(p, VM_PROT_NONE);
1580 if ((p = vm_page_lookup(object, start)) != 0) {
1582 if (p->wire_count != 0) {
1583 vm_page_protect(p, VM_PROT_NONE);
1592 * The busy flags are only cleared at
1593 * interrupt -- minimize the spl transitions
1595 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1598 if (clean_only && p->valid) {
1599 vm_page_test_dirty(p);
1600 if (p->valid & p->dirty) {
1608 vm_page_protect(p, VM_PROT_NONE);
1615 vm_object_pip_wakeup(object);
1619 * Routine: vm_object_coalesce
1620 * Function: Coalesces two objects backing up adjoining
1621 * regions of memory into a single object.
1623 * returns TRUE if objects were combined.
1625 * NOTE: Only works at the moment if the second object is NULL -
1626 * if it's not, which object do we lock first?
1629 * prev_object First object to coalesce
1630 * prev_offset Offset into prev_object
1631 * next_object Second object into coalesce
1632 * next_offset Offset into next_object
1634 * prev_size Size of reference to prev_object
1635 * next_size Size of reference to next_object
1638 * The object must *not* be locked.
1641 vm_object_coalesce(prev_object, prev_pindex, prev_size, next_size)
1642 vm_object_t prev_object;
1643 vm_pindex_t prev_pindex;
1644 vm_size_t prev_size, next_size;
1646 vm_pindex_t next_pindex;
1648 if (prev_object == NULL) {
1652 if (prev_object->type != OBJT_DEFAULT &&
1653 prev_object->type != OBJT_SWAP) {
1658 * Try to collapse the object first
1660 vm_object_collapse(prev_object);
1663 * Can't coalesce if: . more than one reference . paged out . shadows
1664 * another object . has a copy elsewhere (any of which mean that the
1665 * pages not mapped to prev_entry may be in use anyway)
1668 if (prev_object->backing_object != NULL) {
1672 prev_size >>= PAGE_SHIFT;
1673 next_size >>= PAGE_SHIFT;
1674 next_pindex = prev_pindex + prev_size;
1676 if ((prev_object->ref_count > 1) &&
1677 (prev_object->size != next_pindex)) {
1682 * Remove any pages that may still be in the object from a previous
1685 if (next_pindex < prev_object->size) {
1686 vm_object_page_remove(prev_object,
1688 next_pindex + next_size, FALSE);
1689 if (prev_object->type == OBJT_SWAP)
1690 swap_pager_freespace(prev_object,
1691 next_pindex, next_size);
1695 * Extend the object if necessary.
1697 if (next_pindex + next_size > prev_object->size)
1698 prev_object->size = next_pindex + next_size;
1704 vm_object_set_writeable_dirty(vm_object_t object)
1708 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1709 if (object->type == OBJT_VNODE &&
1710 (vp = (struct vnode *)object->handle) != NULL) {
1711 if ((vp->v_flag & VOBJDIRTY) == 0) {
1712 lwkt_gettoken(&vp->v_interlock);
1713 vp->v_flag |= VOBJDIRTY;
1714 lwkt_reltoken(&vp->v_interlock);
1721 #include "opt_ddb.h"
1723 #include <sys/kernel.h>
1725 #include <sys/cons.h>
1727 #include <ddb/ddb.h>
1729 static int _vm_object_in_map (vm_map_t map, vm_object_t object,
1730 vm_map_entry_t entry);
1731 static int vm_object_in_map (vm_object_t object);
1734 _vm_object_in_map(map, object, entry)
1737 vm_map_entry_t entry;
1740 vm_map_entry_t tmpe;
1748 tmpe = map->header.next;
1749 entcount = map->nentries;
1750 while (entcount-- && (tmpe != &map->header)) {
1751 if( _vm_object_in_map(map, object, tmpe)) {
1756 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
1757 tmpm = entry->object.sub_map;
1758 tmpe = tmpm->header.next;
1759 entcount = tmpm->nentries;
1760 while (entcount-- && tmpe != &tmpm->header) {
1761 if( _vm_object_in_map(tmpm, object, tmpe)) {
1766 } else if ((obj = entry->object.vm_object) != NULL) {
1767 for(; obj; obj=obj->backing_object)
1768 if( obj == object) {
1776 vm_object_in_map( object)
1780 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
1781 if( !p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1783 if( _vm_object_in_map(&p->p_vmspace->vm_map, object, 0))
1786 if( _vm_object_in_map( kernel_map, object, 0))
1788 if( _vm_object_in_map( pager_map, object, 0))
1790 if( _vm_object_in_map( buffer_map, object, 0))
1792 if( _vm_object_in_map( mb_map, object, 0))
1797 DB_SHOW_COMMAND(vmochk, vm_object_check)
1802 * make sure that internal objs are in a map somewhere
1803 * and none have zero ref counts.
1805 for (object = TAILQ_FIRST(&vm_object_list);
1807 object = TAILQ_NEXT(object, object_list)) {
1808 if (object->handle == NULL &&
1809 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1810 if (object->ref_count == 0) {
1811 db_printf("vmochk: internal obj has zero ref count: %ld\n",
1812 (long)object->size);
1814 if (!vm_object_in_map(object)) {
1816 "vmochk: internal obj is not in a map: "
1817 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1818 object->ref_count, (u_long)object->size,
1819 (u_long)object->size,
1820 (void *)object->backing_object);
1827 * vm_object_print: [ debug ]
1829 DB_SHOW_COMMAND(object, vm_object_print_static)
1831 /* XXX convert args. */
1832 vm_object_t object = (vm_object_t)addr;
1833 boolean_t full = have_addr;
1837 /* XXX count is an (unused) arg. Avoid shadowing it. */
1838 #define count was_count
1846 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1847 object, (int)object->type, (u_long)object->size,
1848 object->resident_page_count, object->ref_count, object->flags);
1850 * XXX no %qd in kernel. Truncate object->backing_object_offset.
1852 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1853 object->shadow_count,
1854 object->backing_object ? object->backing_object->ref_count : 0,
1855 object->backing_object, (long)object->backing_object_offset);
1862 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = TAILQ_NEXT(p, listq)) {
1864 db_iprintf("memory:=");
1865 else if (count == 6) {
1873 db_printf("(off=0x%lx,page=0x%lx)",
1874 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1884 /* XXX need this non-static entry for calling from vm_map_print. */
1886 vm_object_print(addr, have_addr, count, modif)
1887 /* db_expr_t */ long addr;
1888 boolean_t have_addr;
1889 /* db_expr_t */ long count;
1892 vm_object_print_static(addr, have_addr, count, modif);
1895 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1900 for (object = TAILQ_FIRST(&vm_object_list);
1902 object = TAILQ_NEXT(object, object_list)) {
1903 vm_pindex_t idx, fidx;
1905 vm_offset_t pa = -1, padiff;
1909 db_printf("new object: %p\n", (void *)object);
1919 osize = object->size;
1922 for(idx=0;idx<osize;idx++) {
1923 m = vm_page_lookup(object, idx);
1926 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1927 (long)fidx, rcount, (long)pa);
1942 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1947 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1948 padiff >>= PAGE_SHIFT;
1949 padiff &= PQ_L2_MASK;
1951 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1955 db_printf(" index(%ld)run(%d)pa(0x%lx)",
1956 (long)fidx, rcount, (long)pa);
1957 db_printf("pd(%ld)\n", (long)padiff);
1967 pa = VM_PAGE_TO_PHYS(m);
1971 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1972 (long)fidx, rcount, (long)pa);