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.
68 * Virtual memory object module.
71 #include <sys/param.h>
72 #include <sys/systm.h>
75 #include <sys/mount.h>
76 #include <sys/kernel.h>
77 #include <sys/sysctl.h>
78 #include <sys/mutex.h>
79 #include <sys/proc.h> /* for curproc, pageproc */
80 #include <sys/socket.h>
81 #include <sys/vnode.h>
82 #include <sys/vmmeter.h>
86 #include <vm/vm_param.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_pager.h>
93 #include <vm/vm_zone.h>
94 #include <vm/swap_pager.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_extern.h>
98 #define EASY_SCAN_FACTOR 8
100 #define MSYNC_FLUSH_HARDSEQ 0x01
101 #define MSYNC_FLUSH_SOFTSEQ 0x02
104 * msync / VM object flushing optimizations
106 static int msync_flush_flags = MSYNC_FLUSH_HARDSEQ | MSYNC_FLUSH_SOFTSEQ;
107 SYSCTL_INT(_vm, OID_AUTO, msync_flush_flags,
108 CTLFLAG_RW, &msync_flush_flags, 0, "");
110 static void vm_object_qcollapse(vm_object_t object);
111 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
114 * Virtual memory objects maintain the actual data
115 * associated with allocated virtual memory. A given
116 * page of memory exists within exactly one object.
118 * An object is only deallocated when all "references"
119 * are given up. Only one "reference" to a given
120 * region of an object should be writeable.
122 * Associated with each object is a list of all resident
123 * memory pages belonging to that object; this list is
124 * maintained by the "vm_page" module, and locked by the object's
127 * Each object also records a "pager" routine which is
128 * used to retrieve (and store) pages to the proper backing
129 * storage. In addition, objects may be backed by other
130 * objects from which they were virtual-copied.
132 * The only items within the object structure which are
133 * modified after time of creation are:
134 * reference count locked by object's lock
135 * pager routine locked by object's lock
139 struct object_q vm_object_list;
140 static struct mtx vm_object_list_mtx; /* lock for object list and count */
141 static long vm_object_count; /* count of all objects */
142 vm_object_t kernel_object;
143 vm_object_t kmem_object;
144 static struct vm_object kernel_object_store;
145 static struct vm_object kmem_object_store;
146 extern int vm_pageout_page_count;
148 static long object_collapses;
149 static long object_bypasses;
150 static int next_index;
151 static vm_zone_t obj_zone;
152 static struct vm_zone obj_zone_store;
153 static int object_hash_rand;
154 #define VM_OBJECTS_INIT 256
155 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
158 _vm_object_allocate(objtype_t type, vm_size_t size, vm_object_t object)
164 TAILQ_INIT(&object->memq);
165 TAILQ_INIT(&object->shadow_head);
169 object->ref_count = 1;
171 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
172 vm_object_set_flag(object, OBJ_ONEMAPPING);
173 object->paging_in_progress = 0;
174 object->resident_page_count = 0;
175 object->shadow_count = 0;
176 object->pg_color = next_index;
177 if (size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
178 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
181 next_index = (next_index + incr) & PQ_L2_MASK;
182 object->handle = NULL;
183 object->backing_object = NULL;
184 object->backing_object_offset = (vm_ooffset_t) 0;
186 * Try to generate a number that will spread objects out in the
187 * hash table. We 'wipe' new objects across the hash in 128 page
188 * increments plus 1 more to offset it a little more by the time
191 object->hash_rand = object_hash_rand - 129;
193 object->generation++;
195 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
197 object_hash_rand = object->hash_rand;
203 * Initialize the VM objects module.
210 TAILQ_INIT(&vm_object_list);
211 mtx_init(&vm_object_list_mtx, "vm object_list", MTX_DEF);
214 kernel_object = &kernel_object_store;
215 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
218 kmem_object = &kmem_object_store;
219 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
222 obj_zone = &obj_zone_store;
223 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
224 vm_objects_init, VM_OBJECTS_INIT);
228 vm_object_init2(void)
230 zinitna(obj_zone, NULL, NULL, 0, 0, 0, 1);
234 vm_object_set_flag(vm_object_t object, u_short bits)
237 object->flags |= bits;
241 vm_object_clear_flag(vm_object_t object, u_short bits)
244 object->flags &= ~bits;
248 vm_object_pip_add(vm_object_t object, short i)
251 object->paging_in_progress += i;
255 vm_object_pip_subtract(vm_object_t object, short i)
258 object->paging_in_progress -= i;
262 vm_object_pip_wakeup(vm_object_t object)
265 object->paging_in_progress--;
266 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
267 vm_object_clear_flag(object, OBJ_PIPWNT);
273 vm_object_pip_wakeupn(vm_object_t object, short i)
277 object->paging_in_progress -= i;
278 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
279 vm_object_clear_flag(object, OBJ_PIPWNT);
285 vm_object_pip_sleep(vm_object_t object, char *waitid)
288 if (object->paging_in_progress) {
290 if (object->paging_in_progress) {
291 vm_object_set_flag(object, OBJ_PIPWNT);
292 tsleep(object, PVM, waitid, 0);
299 vm_object_pip_wait(vm_object_t object, char *waitid)
302 while (object->paging_in_progress)
303 vm_object_pip_sleep(object, waitid);
307 * vm_object_allocate:
309 * Returns a new object with the given size.
312 vm_object_allocate(objtype_t type, vm_size_t size)
318 result = (vm_object_t) zalloc(obj_zone);
319 _vm_object_allocate(type, size, result);
326 * vm_object_reference:
328 * Gets another reference to the given object.
331 vm_object_reference(vm_object_t object)
339 /* object can be re-referenced during final cleaning */
340 KASSERT(!(object->flags & OBJ_DEAD),
341 ("vm_object_reference: attempting to reference dead obj"));
345 if (object->type == OBJT_VNODE) {
346 while (vget((struct vnode *) object->handle, LK_RETRY|LK_NOOBJ, curthread)) {
347 printf("vm_object_reference: delay in getting object\n");
353 * handle deallocating a object of type OBJT_VNODE
356 vm_object_vndeallocate(vm_object_t object)
358 struct vnode *vp = (struct vnode *) object->handle;
361 KASSERT(object->type == OBJT_VNODE,
362 ("vm_object_vndeallocate: not a vnode object"));
363 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
365 if (object->ref_count == 0) {
366 vprint("vm_object_vndeallocate", vp);
367 panic("vm_object_vndeallocate: bad object reference count");
372 if (object->ref_count == 0) {
373 vp->v_flag &= ~VTEXT;
374 vm_object_clear_flag(object, OBJ_OPT);
377 * vrele may need a vop lock
383 * vm_object_deallocate:
385 * Release a reference to the specified object,
386 * gained either through a vm_object_allocate
387 * or a vm_object_reference call. When all references
388 * are gone, storage associated with this object
389 * may be relinquished.
391 * No object may be locked.
394 vm_object_deallocate(vm_object_t object)
400 while (object != NULL) {
402 if (object->type == OBJT_VNODE) {
403 vm_object_vndeallocate(object);
407 KASSERT(object->ref_count != 0,
408 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
411 * If the reference count goes to 0 we start calling
412 * vm_object_terminate() on the object chain.
413 * A ref count of 1 may be a special case depending on the
414 * shadow count being 0 or 1.
417 if (object->ref_count > 1) {
419 } else if (object->ref_count == 1) {
420 if (object->shadow_count == 0) {
421 vm_object_set_flag(object, OBJ_ONEMAPPING);
422 } else if ((object->shadow_count == 1) &&
423 (object->handle == NULL) &&
424 (object->type == OBJT_DEFAULT ||
425 object->type == OBJT_SWAP)) {
428 robject = TAILQ_FIRST(&object->shadow_head);
429 KASSERT(robject != NULL,
430 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
432 object->shadow_count));
433 if ((robject->handle == NULL) &&
434 (robject->type == OBJT_DEFAULT ||
435 robject->type == OBJT_SWAP)) {
437 robject->ref_count++;
440 robject->paging_in_progress ||
441 object->paging_in_progress
443 vm_object_pip_sleep(robject, "objde1");
444 vm_object_pip_sleep(object, "objde2");
447 if (robject->ref_count == 1) {
448 robject->ref_count--;
454 vm_object_collapse(object);
465 temp = object->backing_object;
467 TAILQ_REMOVE(&temp->shadow_head, object, shadow_list);
468 temp->shadow_count--;
469 if (temp->ref_count == 0)
470 vm_object_clear_flag(temp, OBJ_OPT);
472 object->backing_object = NULL;
475 * Don't double-terminate, we could be in a termination
476 * recursion due to the terminate having to sync data
479 if ((object->flags & OBJ_DEAD) == 0)
480 vm_object_terminate(object);
486 * vm_object_terminate actually destroys the specified object, freeing
487 * up all previously used resources.
489 * The object must be locked.
490 * This routine may block.
493 vm_object_terminate(vm_object_t object)
501 * Make sure no one uses us.
503 vm_object_set_flag(object, OBJ_DEAD);
506 * wait for the pageout daemon to be done with the object
508 vm_object_pip_wait(object, "objtrm");
510 KASSERT(!object->paging_in_progress,
511 ("vm_object_terminate: pageout in progress"));
514 * Clean and free the pages, as appropriate. All references to the
515 * object are gone, so we don't need to lock it.
517 if (object->type == OBJT_VNODE) {
521 * Freeze optimized copies.
523 vm_freeze_copyopts(object, 0, object->size);
526 * Clean pages and flush buffers.
528 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
530 vp = (struct vnode *) object->handle;
531 vinvalbuf(vp, V_SAVE, NOCRED, NULL, 0, 0);
534 KASSERT(object->ref_count == 0,
535 ("vm_object_terminate: object with references, ref_count=%d",
539 * Now free any remaining pages. For internal objects, this also
540 * removes them from paging queues. Don't free wired pages, just
541 * remove them from the object.
544 while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
545 KASSERT(!p->busy && (p->flags & PG_BUSY) == 0,
546 ("vm_object_terminate: freeing busy page %p "
547 "p->busy = %d, p->flags %x\n", p, p->busy, p->flags));
548 if (p->wire_count == 0) {
560 * Let the pager know object is dead.
562 vm_pager_deallocate(object);
565 * Remove the object from the global object list.
567 mtx_lock(&vm_object_list_mtx);
568 TAILQ_REMOVE(&vm_object_list, object, object_list);
569 mtx_unlock(&vm_object_list_mtx);
574 * Free the space for the object.
576 zfree(obj_zone, object);
580 * vm_object_page_clean
582 * Clean all dirty pages in the specified range of object. Leaves page
583 * on whatever queue it is currently on. If NOSYNC is set then do not
584 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
585 * leaving the object dirty.
587 * Odd semantics: if start == end, we clean everything.
589 * The object must be locked.
592 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags)
595 vm_offset_t tstart, tend;
604 if (object->type != OBJT_VNODE ||
605 (object->flags & OBJ_MIGHTBEDIRTY) == 0)
608 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : 0;
609 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
613 vm_object_set_flag(object, OBJ_CLEANING);
623 * If the caller is smart and only msync()s a range he knows is
624 * dirty, we may be able to avoid an object scan. This results in
625 * a phenominal improvement in performance. We cannot do this
626 * as a matter of course because the object may be huge - e.g.
627 * the size might be in the gigabytes or terrabytes.
629 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
634 scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
638 scanlimit = scanreset;
640 while (tscan < tend) {
641 curgeneration = object->generation;
642 p = vm_page_lookup(object, tscan);
643 if (p == NULL || p->valid == 0 ||
644 (p->queue - p->pc) == PQ_CACHE) {
645 if (--scanlimit == 0)
650 vm_page_test_dirty(p);
651 if ((p->dirty & p->valid) == 0) {
652 if (--scanlimit == 0)
658 * If we have been asked to skip nosync pages and
659 * this is a nosync page, we can't continue.
661 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
662 if (--scanlimit == 0)
667 scanlimit = scanreset;
670 * This returns 0 if it was unable to busy the first
671 * page (i.e. had to sleep).
673 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
677 * If everything was dirty and we flushed it successfully,
678 * and the requested range is not the entire object, we
679 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
680 * return immediately.
682 if (tscan >= tend && (tstart || tend < object->size)) {
683 vm_object_clear_flag(object, OBJ_CLEANING);
689 * Generally set CLEANCHK interlock and make the page read-only so
690 * we can then clear the object flags.
692 * However, if this is a nosync mmap then the object is likely to
693 * stay dirty so do not mess with the page and do not clear the
698 TAILQ_FOREACH(p, &object->memq, listq) {
699 vm_page_flag_set(p, PG_CLEANCHK);
700 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC))
703 vm_page_protect(p, VM_PROT_READ);
706 if (clearobjflags && (tstart == 0) && (tend == object->size)) {
709 vm_object_clear_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
710 if (object->type == OBJT_VNODE &&
711 (vp = (struct vnode *)object->handle) != NULL) {
712 if (vp->v_flag & VOBJDIRTY) {
713 mtx_lock(&vp->v_interlock);
714 vp->v_flag &= ~VOBJDIRTY;
715 mtx_unlock(&vp->v_interlock);
721 curgeneration = object->generation;
723 for (p = TAILQ_FIRST(&object->memq); p; p = np) {
726 np = TAILQ_NEXT(p, listq);
730 if (((p->flags & PG_CLEANCHK) == 0) ||
731 (pi < tstart) || (pi >= tend) ||
733 ((p->queue - p->pc) == PQ_CACHE)) {
734 vm_page_flag_clear(p, PG_CLEANCHK);
738 vm_page_test_dirty(p);
739 if ((p->dirty & p->valid) == 0) {
740 vm_page_flag_clear(p, PG_CLEANCHK);
745 * If we have been asked to skip nosync pages and this is a
746 * nosync page, skip it. Note that the object flags were
747 * not cleared in this case so we do not have to set them.
749 if ((flags & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
750 vm_page_flag_clear(p, PG_CLEANCHK);
754 n = vm_object_page_collect_flush(object, p,
755 curgeneration, pagerflags);
759 if (object->generation != curgeneration)
763 * Try to optimize the next page. If we can't we pick up
764 * our (random) scan where we left off.
766 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
767 if ((p = vm_page_lookup(object, pi + n)) != NULL)
773 VOP_FSYNC(vp, NULL, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
776 vm_object_clear_flag(object, OBJ_CLEANING);
781 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
790 vm_page_t maf[vm_pageout_page_count];
791 vm_page_t mab[vm_pageout_page_count];
792 vm_page_t ma[vm_pageout_page_count];
796 while (vm_page_sleep_busy(p, TRUE, "vpcwai")) {
797 if (object->generation != curgeneration) {
804 for(i = 1; i < vm_pageout_page_count; i++) {
807 if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
808 if ((tp->flags & PG_BUSY) ||
809 (tp->flags & PG_CLEANCHK) == 0 ||
812 if((tp->queue - tp->pc) == PQ_CACHE) {
813 vm_page_flag_clear(tp, PG_CLEANCHK);
816 vm_page_test_dirty(tp);
817 if ((tp->dirty & tp->valid) == 0) {
818 vm_page_flag_clear(tp, PG_CLEANCHK);
829 chkb = vm_pageout_page_count - maxf;
831 for(i = 1; i < chkb;i++) {
834 if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
835 if ((tp->flags & PG_BUSY) ||
836 (tp->flags & PG_CLEANCHK) == 0 ||
839 if ((tp->queue - tp->pc) == PQ_CACHE) {
840 vm_page_flag_clear(tp, PG_CLEANCHK);
843 vm_page_test_dirty(tp);
844 if ((tp->dirty & tp->valid) == 0) {
845 vm_page_flag_clear(tp, PG_CLEANCHK);
856 for(i = 0; i < maxb; i++) {
857 int index = (maxb - i) - 1;
859 vm_page_flag_clear(ma[index], PG_CLEANCHK);
861 vm_page_flag_clear(p, PG_CLEANCHK);
863 for(i = 0; i < maxf; i++) {
864 int index = (maxb + i) + 1;
866 vm_page_flag_clear(ma[index], PG_CLEANCHK);
868 runlen = maxb + maxf + 1;
871 vm_pageout_flush(ma, runlen, pagerflags);
872 for (i = 0; i < runlen; i++) {
873 if (ma[i]->valid & ma[i]->dirty) {
874 vm_page_protect(ma[i], VM_PROT_READ);
875 vm_page_flag_set(ma[i], PG_CLEANCHK);
878 * maxf will end up being the actual number of pages
879 * we wrote out contiguously, non-inclusive of the
880 * first page. We do not count look-behind pages.
882 if (i >= maxb + 1 && (maxf > i - maxb - 1))
890 * Same as vm_object_pmap_copy, except range checking really
891 * works, and is meant for small sections of an object.
893 * This code protects resident pages by making them read-only
894 * and is typically called on a fork or split when a page
895 * is converted to copy-on-write.
897 * NOTE: If the page is already at VM_PROT_NONE, calling
898 * vm_page_protect will have no effect.
901 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
908 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
911 for (idx = start; idx < end; idx++) {
912 p = vm_page_lookup(object, idx);
915 vm_page_protect(p, VM_PROT_READ);
920 * vm_object_pmap_remove:
922 * Removes all physical pages in the specified
923 * object range from all physical maps.
925 * The object must *not* be locked.
928 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
935 TAILQ_FOREACH(p, &object->memq, listq) {
936 if (p->pindex >= start && p->pindex < end)
937 vm_page_protect(p, VM_PROT_NONE);
939 if ((start == 0) && (object->size == end))
940 vm_object_clear_flag(object, OBJ_WRITEABLE);
946 * Implements the madvise function at the object/page level.
948 * MADV_WILLNEED (any object)
950 * Activate the specified pages if they are resident.
952 * MADV_DONTNEED (any object)
954 * Deactivate the specified pages if they are resident.
956 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
957 * OBJ_ONEMAPPING only)
959 * Deactivate and clean the specified pages if they are
960 * resident. This permits the process to reuse the pages
961 * without faulting or the kernel to reclaim the pages
965 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
967 vm_pindex_t end, tpindex;
975 end = pindex + count;
978 * Locate and adjust resident pages
980 for (; pindex < end; pindex += 1) {
986 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
987 * and those pages must be OBJ_ONEMAPPING.
989 if (advise == MADV_FREE) {
990 if ((tobject->type != OBJT_DEFAULT &&
991 tobject->type != OBJT_SWAP) ||
992 (tobject->flags & OBJ_ONEMAPPING) == 0) {
997 m = vm_page_lookup(tobject, tpindex);
1001 * There may be swap even if there is no backing page
1003 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1004 swap_pager_freespace(tobject, tpindex, 1);
1009 tobject = tobject->backing_object;
1010 if (tobject == NULL)
1012 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1017 * If the page is busy or not in a normal active state,
1018 * we skip it. If the page is not managed there are no
1019 * page queues to mess with. Things can break if we mess
1020 * with pages in any of the below states.
1025 (m->flags & PG_UNMANAGED) ||
1026 m->valid != VM_PAGE_BITS_ALL
1031 if (vm_page_sleep_busy(m, TRUE, "madvpo"))
1034 if (advise == MADV_WILLNEED) {
1035 vm_page_activate(m);
1036 } else if (advise == MADV_DONTNEED) {
1037 vm_page_dontneed(m);
1038 } else if (advise == MADV_FREE) {
1040 * Mark the page clean. This will allow the page
1041 * to be freed up by the system. However, such pages
1042 * are often reused quickly by malloc()/free()
1043 * so we do not do anything that would cause
1044 * a page fault if we can help it.
1046 * Specifically, we do not try to actually free
1047 * the page now nor do we try to put it in the
1048 * cache (which would cause a page fault on reuse).
1050 * But we do make the page is freeable as we
1051 * can without actually taking the step of unmapping
1054 pmap_clear_modify(m);
1057 vm_page_dontneed(m);
1058 if (tobject->type == OBJT_SWAP)
1059 swap_pager_freespace(tobject, tpindex, 1);
1067 * Create a new object which is backed by the
1068 * specified existing object range. The source
1069 * object reference is deallocated.
1071 * The new object and offset into that object
1072 * are returned in the source parameters.
1076 vm_object_t *object, /* IN/OUT */
1077 vm_ooffset_t *offset, /* IN/OUT */
1087 * Don't create the new object if the old object isn't shared.
1089 if (source != NULL &&
1090 source->ref_count == 1 &&
1091 source->handle == NULL &&
1092 (source->type == OBJT_DEFAULT ||
1093 source->type == OBJT_SWAP))
1097 * Allocate a new object with the given length
1099 result = vm_object_allocate(OBJT_DEFAULT, length);
1100 KASSERT(result != NULL, ("vm_object_shadow: no object for shadowing"));
1103 * The new object shadows the source object, adding a reference to it.
1104 * Our caller changes his reference to point to the new object,
1105 * removing a reference to the source object. Net result: no change
1106 * of reference count.
1108 * Try to optimize the result object's page color when shadowing
1109 * in order to maintain page coloring consistency in the combined
1112 result->backing_object = source;
1114 TAILQ_INSERT_TAIL(&source->shadow_head, result, shadow_list);
1115 source->shadow_count++;
1116 source->generation++;
1117 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) & PQ_L2_MASK;
1121 * Store the offset into the source object, and fix up the offset into
1124 result->backing_object_offset = *offset;
1127 * Return the new things
1133 #define OBSC_TEST_ALL_SHADOWED 0x0001
1134 #define OBSC_COLLAPSE_NOWAIT 0x0002
1135 #define OBSC_COLLAPSE_WAIT 0x0004
1138 vm_object_backing_scan(vm_object_t object, int op)
1143 vm_object_t backing_object;
1144 vm_pindex_t backing_offset_index;
1149 backing_object = object->backing_object;
1150 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1153 * Initial conditions
1155 if (op & OBSC_TEST_ALL_SHADOWED) {
1157 * We do not want to have to test for the existence of
1158 * swap pages in the backing object. XXX but with the
1159 * new swapper this would be pretty easy to do.
1161 * XXX what about anonymous MAP_SHARED memory that hasn't
1162 * been ZFOD faulted yet? If we do not test for this, the
1163 * shadow test may succeed! XXX
1165 if (backing_object->type != OBJT_DEFAULT) {
1170 if (op & OBSC_COLLAPSE_WAIT) {
1171 vm_object_set_flag(backing_object, OBJ_DEAD);
1177 p = TAILQ_FIRST(&backing_object->memq);
1179 vm_page_t next = TAILQ_NEXT(p, listq);
1180 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1182 if (op & OBSC_TEST_ALL_SHADOWED) {
1186 * Ignore pages outside the parent object's range
1187 * and outside the parent object's mapping of the
1190 * note that we do not busy the backing object's
1194 p->pindex < backing_offset_index ||
1195 new_pindex >= object->size
1202 * See if the parent has the page or if the parent's
1203 * object pager has the page. If the parent has the
1204 * page but the page is not valid, the parent's
1205 * object pager must have the page.
1207 * If this fails, the parent does not completely shadow
1208 * the object and we might as well give up now.
1211 pp = vm_page_lookup(object, new_pindex);
1213 (pp == NULL || pp->valid == 0) &&
1214 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1222 * Check for busy page
1224 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1227 if (op & OBSC_COLLAPSE_NOWAIT) {
1229 (p->flags & PG_BUSY) ||
1238 } else if (op & OBSC_COLLAPSE_WAIT) {
1239 if (vm_page_sleep_busy(p, TRUE, "vmocol")) {
1241 * If we slept, anything could have
1242 * happened. Since the object is
1243 * marked dead, the backing offset
1244 * should not have changed so we
1245 * just restart our scan.
1247 p = TAILQ_FIRST(&backing_object->memq);
1258 p->object == backing_object,
1259 ("vm_object_qcollapse(): object mismatch")
1263 * Destroy any associated swap
1265 if (backing_object->type == OBJT_SWAP) {
1266 swap_pager_freespace(
1274 p->pindex < backing_offset_index ||
1275 new_pindex >= object->size
1278 * Page is out of the parent object's range, we
1279 * can simply destroy it.
1281 vm_page_protect(p, VM_PROT_NONE);
1287 pp = vm_page_lookup(object, new_pindex);
1290 vm_pager_has_page(object, new_pindex, NULL, NULL)
1293 * page already exists in parent OR swap exists
1294 * for this location in the parent. Destroy
1295 * the original page from the backing object.
1297 * Leave the parent's page alone
1299 vm_page_protect(p, VM_PROT_NONE);
1306 * Page does not exist in parent, rename the
1307 * page from the backing object to the main object.
1309 * If the page was mapped to a process, it can remain
1310 * mapped through the rename.
1312 if ((p->queue - p->pc) == PQ_CACHE)
1313 vm_page_deactivate(p);
1315 vm_page_rename(p, object, new_pindex);
1316 /* page automatically made dirty by rename */
1326 * this version of collapse allows the operation to occur earlier and
1327 * when paging_in_progress is true for an object... This is not a complete
1328 * operation, but should plug 99.9% of the rest of the leaks.
1331 vm_object_qcollapse(vm_object_t object)
1333 vm_object_t backing_object = object->backing_object;
1337 if (backing_object->ref_count != 1)
1340 backing_object->ref_count += 2;
1342 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1344 backing_object->ref_count -= 2;
1348 * vm_object_collapse:
1350 * Collapse an object with the object backing it.
1351 * Pages in the backing object are moved into the
1352 * parent, and the backing object is deallocated.
1355 vm_object_collapse(vm_object_t object)
1360 vm_object_t backing_object;
1363 * Verify that the conditions are right for collapse:
1365 * The object exists and the backing object exists.
1370 if ((backing_object = object->backing_object) == NULL)
1374 * we check the backing object first, because it is most likely
1377 if (backing_object->handle != NULL ||
1378 (backing_object->type != OBJT_DEFAULT &&
1379 backing_object->type != OBJT_SWAP) ||
1380 (backing_object->flags & OBJ_DEAD) ||
1381 object->handle != NULL ||
1382 (object->type != OBJT_DEFAULT &&
1383 object->type != OBJT_SWAP) ||
1384 (object->flags & OBJ_DEAD)) {
1389 object->paging_in_progress != 0 ||
1390 backing_object->paging_in_progress != 0
1392 vm_object_qcollapse(object);
1397 * We know that we can either collapse the backing object (if
1398 * the parent is the only reference to it) or (perhaps) have
1399 * the parent bypass the object if the parent happens to shadow
1400 * all the resident pages in the entire backing object.
1402 * This is ignoring pager-backed pages such as swap pages.
1403 * vm_object_backing_scan fails the shadowing test in this
1406 if (backing_object->ref_count == 1) {
1408 * If there is exactly one reference to the backing
1409 * object, we can collapse it into the parent.
1411 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1414 * Move the pager from backing_object to object.
1416 if (backing_object->type == OBJT_SWAP) {
1417 vm_object_pip_add(backing_object, 1);
1420 * scrap the paging_offset junk and do a
1421 * discrete copy. This also removes major
1422 * assumptions about how the swap-pager
1423 * works from where it doesn't belong. The
1424 * new swapper is able to optimize the
1425 * destroy-source case.
1427 vm_object_pip_add(object, 1);
1431 OFF_TO_IDX(object->backing_object_offset), TRUE);
1432 vm_object_pip_wakeup(object);
1434 vm_object_pip_wakeup(backing_object);
1437 * Object now shadows whatever backing_object did.
1438 * Note that the reference to
1439 * backing_object->backing_object moves from within
1440 * backing_object to within object.
1443 &object->backing_object->shadow_head,
1447 object->backing_object->shadow_count--;
1448 object->backing_object->generation++;
1449 if (backing_object->backing_object) {
1451 &backing_object->backing_object->shadow_head,
1455 backing_object->backing_object->shadow_count--;
1456 backing_object->backing_object->generation++;
1458 object->backing_object = backing_object->backing_object;
1459 if (object->backing_object) {
1461 &object->backing_object->shadow_head,
1465 object->backing_object->shadow_count++;
1466 object->backing_object->generation++;
1469 object->backing_object_offset +=
1470 backing_object->backing_object_offset;
1473 * Discard backing_object.
1475 * Since the backing object has no pages, no pager left,
1476 * and no object references within it, all that is
1477 * necessary is to dispose of it.
1479 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1480 KASSERT(TAILQ_FIRST(&backing_object->memq) == NULL, ("backing_object %p somehow has left over pages during collapse!", backing_object));
1489 zfree(obj_zone, backing_object);
1493 vm_object_t new_backing_object;
1496 * If we do not entirely shadow the backing object,
1497 * there is nothing we can do so we give up.
1499 if (vm_object_backing_scan(object, OBSC_TEST_ALL_SHADOWED) == 0) {
1504 * Make the parent shadow the next object in the
1505 * chain. Deallocating backing_object will not remove
1506 * it, since its reference count is at least 2.
1509 &backing_object->shadow_head,
1513 backing_object->shadow_count--;
1514 backing_object->generation++;
1516 new_backing_object = backing_object->backing_object;
1517 if ((object->backing_object = new_backing_object) != NULL) {
1518 vm_object_reference(new_backing_object);
1520 &new_backing_object->shadow_head,
1524 new_backing_object->shadow_count++;
1525 new_backing_object->generation++;
1526 object->backing_object_offset +=
1527 backing_object->backing_object_offset;
1531 * Drop the reference count on backing_object. Since
1532 * its ref_count was at least 2, it will not vanish;
1533 * so we don't need to call vm_object_deallocate, but
1536 vm_object_deallocate(backing_object);
1541 * Try again with this object's new backing object.
1547 * vm_object_page_remove: [internal]
1549 * Removes all physical pages in the specified
1550 * object range from the object's list of pages.
1552 * The object must be locked.
1555 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end, boolean_t clean_only)
1563 if (object == NULL ||
1564 object->resident_page_count == 0)
1567 all = ((end == 0) && (start == 0));
1570 * Since physically-backed objects do not use managed pages, we can't
1571 * remove pages from the object (we must instead remove the page
1572 * references, and then destroy the object).
1574 KASSERT(object->type != OBJT_PHYS, ("attempt to remove pages from a physical object"));
1576 vm_object_pip_add(object, 1);
1579 if (all || size > object->resident_page_count / 4) {
1580 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = next) {
1581 next = TAILQ_NEXT(p, listq);
1582 if (all || ((start <= p->pindex) && (p->pindex < end))) {
1583 if (p->wire_count != 0) {
1584 vm_page_protect(p, VM_PROT_NONE);
1591 * The busy flags are only cleared at
1592 * interrupt -- minimize the spl transitions
1594 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1597 if (clean_only && p->valid) {
1598 vm_page_test_dirty(p);
1599 if (p->valid & p->dirty)
1604 vm_page_protect(p, VM_PROT_NONE);
1610 if ((p = vm_page_lookup(object, start)) != 0) {
1612 if (p->wire_count != 0) {
1613 vm_page_protect(p, VM_PROT_NONE);
1622 * The busy flags are only cleared at
1623 * interrupt -- minimize the spl transitions
1625 if (vm_page_sleep_busy(p, TRUE, "vmopar"))
1628 if (clean_only && p->valid) {
1629 vm_page_test_dirty(p);
1630 if (p->valid & p->dirty) {
1638 vm_page_protect(p, VM_PROT_NONE);
1645 vm_object_pip_wakeup(object);
1649 * Routine: vm_object_coalesce
1650 * Function: Coalesces two objects backing up adjoining
1651 * regions of memory into a single object.
1653 * returns TRUE if objects were combined.
1655 * NOTE: Only works at the moment if the second object is NULL -
1656 * if it's not, which object do we lock first?
1659 * prev_object First object to coalesce
1660 * prev_offset Offset into prev_object
1661 * next_object Second object into coalesce
1662 * next_offset Offset into next_object
1664 * prev_size Size of reference to prev_object
1665 * next_size Size of reference to next_object
1668 * The object must *not* be locked.
1671 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex, vm_size_t prev_size, vm_size_t next_size)
1673 vm_pindex_t next_pindex;
1677 if (prev_object == NULL) {
1681 if (prev_object->type != OBJT_DEFAULT &&
1682 prev_object->type != OBJT_SWAP) {
1687 * Try to collapse the object first
1689 vm_object_collapse(prev_object);
1692 * Can't coalesce if: . more than one reference . paged out . shadows
1693 * another object . has a copy elsewhere (any of which mean that the
1694 * pages not mapped to prev_entry may be in use anyway)
1696 if (prev_object->backing_object != NULL) {
1700 prev_size >>= PAGE_SHIFT;
1701 next_size >>= PAGE_SHIFT;
1702 next_pindex = prev_pindex + prev_size;
1704 if ((prev_object->ref_count > 1) &&
1705 (prev_object->size != next_pindex)) {
1710 * Remove any pages that may still be in the object from a previous
1713 if (next_pindex < prev_object->size) {
1714 vm_object_page_remove(prev_object,
1716 next_pindex + next_size, FALSE);
1717 if (prev_object->type == OBJT_SWAP)
1718 swap_pager_freespace(prev_object,
1719 next_pindex, next_size);
1723 * Extend the object if necessary.
1725 if (next_pindex + next_size > prev_object->size)
1726 prev_object->size = next_pindex + next_size;
1732 vm_object_set_writeable_dirty(vm_object_t object)
1736 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1737 if (object->type == OBJT_VNODE &&
1738 (vp = (struct vnode *)object->handle) != NULL) {
1739 if ((vp->v_flag & VOBJDIRTY) == 0) {
1740 mtx_lock(&vp->v_interlock);
1741 vp->v_flag |= VOBJDIRTY;
1742 mtx_unlock(&vp->v_interlock);
1747 #include "opt_ddb.h"
1749 #include <sys/kernel.h>
1751 #include <sys/cons.h>
1753 #include <ddb/ddb.h>
1756 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
1759 vm_map_entry_t tmpe;
1767 tmpe = map->header.next;
1768 entcount = map->nentries;
1769 while (entcount-- && (tmpe != &map->header)) {
1770 if (_vm_object_in_map(map, object, tmpe)) {
1775 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
1776 tmpm = entry->object.sub_map;
1777 tmpe = tmpm->header.next;
1778 entcount = tmpm->nentries;
1779 while (entcount-- && tmpe != &tmpm->header) {
1780 if (_vm_object_in_map(tmpm, object, tmpe)) {
1785 } else if ((obj = entry->object.vm_object) != NULL) {
1786 for (; obj; obj = obj->backing_object)
1787 if (obj == object) {
1795 vm_object_in_map(vm_object_t object)
1799 /* sx_slock(&allproc_lock); */
1800 LIST_FOREACH(p, &allproc, p_list) {
1801 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
1803 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
1804 /* sx_sunlock(&allproc_lock); */
1808 /* sx_sunlock(&allproc_lock); */
1809 if (_vm_object_in_map(kernel_map, object, 0))
1811 if (_vm_object_in_map(kmem_map, object, 0))
1813 if (_vm_object_in_map(pager_map, object, 0))
1815 if (_vm_object_in_map(buffer_map, object, 0))
1820 DB_SHOW_COMMAND(vmochk, vm_object_check)
1825 * make sure that internal objs are in a map somewhere
1826 * and none have zero ref counts.
1828 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1829 if (object->handle == NULL &&
1830 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
1831 if (object->ref_count == 0) {
1832 db_printf("vmochk: internal obj has zero ref count: %ld\n",
1833 (long)object->size);
1835 if (!vm_object_in_map(object)) {
1837 "vmochk: internal obj is not in a map: "
1838 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
1839 object->ref_count, (u_long)object->size,
1840 (u_long)object->size,
1841 (void *)object->backing_object);
1848 * vm_object_print: [ debug ]
1850 DB_SHOW_COMMAND(object, vm_object_print_static)
1852 /* XXX convert args. */
1853 vm_object_t object = (vm_object_t)addr;
1854 boolean_t full = have_addr;
1858 /* XXX count is an (unused) arg. Avoid shadowing it. */
1859 #define count was_count
1867 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
1868 object, (int)object->type, (u_long)object->size,
1869 object->resident_page_count, object->ref_count, object->flags);
1871 * XXX no %qd in kernel. Truncate object->backing_object_offset.
1873 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
1874 object->shadow_count,
1875 object->backing_object ? object->backing_object->ref_count : 0,
1876 object->backing_object, (long)object->backing_object_offset);
1883 TAILQ_FOREACH(p, &object->memq, listq) {
1885 db_iprintf("memory:=");
1886 else if (count == 6) {
1894 db_printf("(off=0x%lx,page=0x%lx)",
1895 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
1905 /* XXX need this non-static entry for calling from vm_map_print. */
1908 /* db_expr_t */ long addr,
1909 boolean_t have_addr,
1910 /* db_expr_t */ long count,
1913 vm_object_print_static(addr, have_addr, count, modif);
1916 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
1922 TAILQ_FOREACH(object, &vm_object_list, object_list) {
1923 vm_pindex_t idx, fidx;
1925 vm_offset_t pa = -1, padiff;
1929 db_printf("new object: %p\n", (void *)object);
1939 osize = object->size;
1942 for (idx = 0; idx < osize; idx++) {
1943 m = vm_page_lookup(object, idx);
1946 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1947 (long)fidx, rcount, (long)pa);
1962 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
1967 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
1968 padiff >>= PAGE_SHIFT;
1969 padiff &= PQ_L2_MASK;
1971 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
1975 db_printf(" index(%ld)run(%d)pa(0x%lx)",
1976 (long)fidx, rcount, (long)pa);
1977 db_printf("pd(%ld)\n", (long)padiff);
1987 pa = VM_PAGE_TO_PHYS(m);
1991 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
1992 (long)fidx, rcount, (long)pa);