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 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94
35 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36 * All rights reserved.
38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
40 * Permission to use, copy, modify and distribute this software and
41 * its documentation is hereby granted, provided that both the copyright
42 * notice and this permission notice appear in all copies of the
43 * software, derivative works or modified versions, and any portions
44 * thereof, and that both notices appear in supporting documentation.
46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
50 * Carnegie Mellon requests users of this software to return to
52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
53 * School of Computer Science
54 * Carnegie Mellon University
55 * Pittsburgh PA 15213-3890
57 * any improvements or extensions that they make and grant Carnegie the
58 * rights to redistribute these changes.
62 * Virtual memory object module.
65 #include <sys/cdefs.h>
66 __FBSDID("$FreeBSD$");
70 #include <sys/param.h>
71 #include <sys/systm.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/mutex.h>
78 #include <sys/proc.h> /* for curproc, pageproc */
79 #include <sys/socket.h>
80 #include <sys/vnode.h>
81 #include <sys/vmmeter.h>
85 #include <vm/vm_param.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_pager.h>
92 #include <vm/swap_pager.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_extern.h>
95 #include <vm/vm_reserv.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, CTLFLAG_RW, &msync_flush_flags, 0,
108 "Enable sequential iteration optimization");
110 static int old_msync;
111 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
112 "Use old (insecure) msync behavior");
114 static void vm_object_qcollapse(vm_object_t object);
115 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags);
116 static void vm_object_vndeallocate(vm_object_t object);
119 * Virtual memory objects maintain the actual data
120 * associated with allocated virtual memory. A given
121 * page of memory exists within exactly one object.
123 * An object is only deallocated when all "references"
124 * are given up. Only one "reference" to a given
125 * region of an object should be writeable.
127 * Associated with each object is a list of all resident
128 * memory pages belonging to that object; this list is
129 * maintained by the "vm_page" module, and locked by the object's
132 * Each object also records a "pager" routine which is
133 * used to retrieve (and store) pages to the proper backing
134 * storage. In addition, objects may be backed by other
135 * objects from which they were virtual-copied.
137 * The only items within the object structure which are
138 * modified after time of creation are:
139 * reference count locked by object's lock
140 * pager routine locked by object's lock
144 struct object_q vm_object_list;
145 struct mtx vm_object_list_mtx; /* lock for object list and count */
147 struct vm_object kernel_object_store;
148 struct vm_object kmem_object_store;
150 SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0, "VM object stats");
152 static long object_collapses;
153 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
154 &object_collapses, 0, "VM object collapses");
156 static long object_bypasses;
157 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
158 &object_bypasses, 0, "VM object bypasses");
160 static uma_zone_t obj_zone;
162 static int vm_object_zinit(void *mem, int size, int flags);
165 static void vm_object_zdtor(void *mem, int size, void *arg);
168 vm_object_zdtor(void *mem, int size, void *arg)
172 object = (vm_object_t)mem;
173 KASSERT(TAILQ_EMPTY(&object->memq),
174 ("object %p has resident pages",
176 #if VM_NRESERVLEVEL > 0
177 KASSERT(LIST_EMPTY(&object->rvq),
178 ("object %p has reservations",
181 KASSERT(object->cache == NULL,
182 ("object %p has cached pages",
184 KASSERT(object->paging_in_progress == 0,
185 ("object %p paging_in_progress = %d",
186 object, object->paging_in_progress));
187 KASSERT(object->resident_page_count == 0,
188 ("object %p resident_page_count = %d",
189 object, object->resident_page_count));
190 KASSERT(object->shadow_count == 0,
191 ("object %p shadow_count = %d",
192 object, object->shadow_count));
197 vm_object_zinit(void *mem, int size, int flags)
201 object = (vm_object_t)mem;
202 bzero(&object->mtx, sizeof(object->mtx));
203 VM_OBJECT_LOCK_INIT(object, "standard object");
205 /* These are true for any object that has been freed */
206 object->paging_in_progress = 0;
207 object->resident_page_count = 0;
208 object->shadow_count = 0;
213 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
216 TAILQ_INIT(&object->memq);
217 LIST_INIT(&object->shadow_head);
222 object->generation = 1;
223 object->ref_count = 1;
225 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
226 object->flags = OBJ_ONEMAPPING;
227 object->pg_color = 0;
228 object->handle = NULL;
229 object->backing_object = NULL;
230 object->backing_object_offset = (vm_ooffset_t) 0;
231 #if VM_NRESERVLEVEL > 0
232 LIST_INIT(&object->rvq);
234 object->cache = NULL;
236 mtx_lock(&vm_object_list_mtx);
237 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
238 mtx_unlock(&vm_object_list_mtx);
244 * Initialize the VM objects module.
249 TAILQ_INIT(&vm_object_list);
250 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
252 VM_OBJECT_LOCK_INIT(&kernel_object_store, "kernel object");
253 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
255 #if VM_NRESERVLEVEL > 0
256 kernel_object->flags |= OBJ_COLORED;
257 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
260 VM_OBJECT_LOCK_INIT(&kmem_object_store, "kmem object");
261 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
263 #if VM_NRESERVLEVEL > 0
264 kmem_object->flags |= OBJ_COLORED;
265 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
269 * The lock portion of struct vm_object must be type stable due
270 * to vm_pageout_fallback_object_lock locking a vm object
271 * without holding any references to it.
273 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
279 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
283 vm_object_clear_flag(vm_object_t object, u_short bits)
286 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
287 object->flags &= ~bits;
291 vm_object_pip_add(vm_object_t object, short i)
294 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
295 object->paging_in_progress += i;
299 vm_object_pip_subtract(vm_object_t object, short i)
302 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
303 object->paging_in_progress -= i;
307 vm_object_pip_wakeup(vm_object_t object)
310 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
311 object->paging_in_progress--;
312 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
313 vm_object_clear_flag(object, OBJ_PIPWNT);
319 vm_object_pip_wakeupn(vm_object_t object, short i)
322 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
324 object->paging_in_progress -= i;
325 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
326 vm_object_clear_flag(object, OBJ_PIPWNT);
332 vm_object_pip_wait(vm_object_t object, char *waitid)
335 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
336 while (object->paging_in_progress) {
337 object->flags |= OBJ_PIPWNT;
338 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
343 * vm_object_allocate:
345 * Returns a new object with the given size.
348 vm_object_allocate(objtype_t type, vm_pindex_t size)
352 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
353 _vm_object_allocate(type, size, object);
359 * vm_object_reference:
361 * Gets another reference to the given object. Note: OBJ_DEAD
362 * objects can be referenced during final cleaning.
365 vm_object_reference(vm_object_t object)
369 VM_OBJECT_LOCK(object);
370 vm_object_reference_locked(object);
371 VM_OBJECT_UNLOCK(object);
375 * vm_object_reference_locked:
377 * Gets another reference to the given object.
379 * The object must be locked.
382 vm_object_reference_locked(vm_object_t object)
386 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
388 if (object->type == OBJT_VNODE) {
395 * Handle deallocating an object of type OBJT_VNODE.
398 vm_object_vndeallocate(vm_object_t object)
400 struct vnode *vp = (struct vnode *) object->handle;
402 VFS_ASSERT_GIANT(vp->v_mount);
403 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
404 KASSERT(object->type == OBJT_VNODE,
405 ("vm_object_vndeallocate: not a vnode object"));
406 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
408 if (object->ref_count == 0) {
409 vprint("vm_object_vndeallocate", vp);
410 panic("vm_object_vndeallocate: bad object reference count");
415 if (object->ref_count == 0) {
416 mp_fixme("Unlocked vflag access.");
417 vp->v_vflag &= ~VV_TEXT;
419 VM_OBJECT_UNLOCK(object);
421 * vrele may need a vop lock
427 * vm_object_deallocate:
429 * Release a reference to the specified object,
430 * gained either through a vm_object_allocate
431 * or a vm_object_reference call. When all references
432 * are gone, storage associated with this object
433 * may be relinquished.
435 * No object may be locked.
438 vm_object_deallocate(vm_object_t object)
442 while (object != NULL) {
447 VM_OBJECT_LOCK(object);
448 if (object->type == OBJT_VNODE) {
449 struct vnode *vp = (struct vnode *) object->handle;
452 * Conditionally acquire Giant for a vnode-backed
453 * object. We have to be careful since the type of
454 * a vnode object can change while the object is
457 if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
459 if (!mtx_trylock(&Giant)) {
460 VM_OBJECT_UNLOCK(object);
465 vm_object_vndeallocate(object);
466 VFS_UNLOCK_GIANT(vfslocked);
470 * This is to handle the case that the object
471 * changed type while we dropped its lock to
474 VFS_UNLOCK_GIANT(vfslocked);
476 KASSERT(object->ref_count != 0,
477 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
480 * If the reference count goes to 0 we start calling
481 * vm_object_terminate() on the object chain.
482 * A ref count of 1 may be a special case depending on the
483 * shadow count being 0 or 1.
486 if (object->ref_count > 1) {
487 VM_OBJECT_UNLOCK(object);
489 } else if (object->ref_count == 1) {
490 if (object->shadow_count == 0 &&
491 object->handle == NULL &&
492 (object->type == OBJT_DEFAULT ||
493 object->type == OBJT_SWAP)) {
494 vm_object_set_flag(object, OBJ_ONEMAPPING);
495 } else if ((object->shadow_count == 1) &&
496 (object->handle == NULL) &&
497 (object->type == OBJT_DEFAULT ||
498 object->type == OBJT_SWAP)) {
501 robject = LIST_FIRST(&object->shadow_head);
502 KASSERT(robject != NULL,
503 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
505 object->shadow_count));
506 if (!VM_OBJECT_TRYLOCK(robject)) {
508 * Avoid a potential deadlock.
511 VM_OBJECT_UNLOCK(object);
513 * More likely than not the thread
514 * holding robject's lock has lower
515 * priority than the current thread.
516 * Let the lower priority thread run.
522 * Collapse object into its shadow unless its
523 * shadow is dead. In that case, object will
524 * be deallocated by the thread that is
525 * deallocating its shadow.
527 if ((robject->flags & OBJ_DEAD) == 0 &&
528 (robject->handle == NULL) &&
529 (robject->type == OBJT_DEFAULT ||
530 robject->type == OBJT_SWAP)) {
532 robject->ref_count++;
534 if (robject->paging_in_progress) {
535 VM_OBJECT_UNLOCK(object);
536 vm_object_pip_wait(robject,
538 temp = robject->backing_object;
539 if (object == temp) {
540 VM_OBJECT_LOCK(object);
543 } else if (object->paging_in_progress) {
544 VM_OBJECT_UNLOCK(robject);
545 object->flags |= OBJ_PIPWNT;
547 VM_OBJECT_MTX(object),
548 PDROP | PVM, "objde2", 0);
549 VM_OBJECT_LOCK(robject);
550 temp = robject->backing_object;
551 if (object == temp) {
552 VM_OBJECT_LOCK(object);
556 VM_OBJECT_UNLOCK(object);
558 if (robject->ref_count == 1) {
559 robject->ref_count--;
564 vm_object_collapse(object);
565 VM_OBJECT_UNLOCK(object);
568 VM_OBJECT_UNLOCK(robject);
570 VM_OBJECT_UNLOCK(object);
574 temp = object->backing_object;
576 VM_OBJECT_LOCK(temp);
577 LIST_REMOVE(object, shadow_list);
578 temp->shadow_count--;
580 VM_OBJECT_UNLOCK(temp);
581 object->backing_object = NULL;
584 * Don't double-terminate, we could be in a termination
585 * recursion due to the terminate having to sync data
588 if ((object->flags & OBJ_DEAD) == 0)
589 vm_object_terminate(object);
591 VM_OBJECT_UNLOCK(object);
597 * vm_object_destroy removes the object from the global object list
598 * and frees the space for the object.
601 vm_object_destroy(vm_object_t object)
605 * Remove the object from the global object list.
607 mtx_lock(&vm_object_list_mtx);
608 TAILQ_REMOVE(&vm_object_list, object, object_list);
609 mtx_unlock(&vm_object_list_mtx);
612 * Free the space for the object.
614 uma_zfree(obj_zone, object);
618 * vm_object_terminate actually destroys the specified object, freeing
619 * up all previously used resources.
621 * The object must be locked.
622 * This routine may block.
625 vm_object_terminate(vm_object_t object)
629 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
632 * Make sure no one uses us.
634 vm_object_set_flag(object, OBJ_DEAD);
637 * wait for the pageout daemon to be done with the object
639 vm_object_pip_wait(object, "objtrm");
641 KASSERT(!object->paging_in_progress,
642 ("vm_object_terminate: pageout in progress"));
645 * Clean and free the pages, as appropriate. All references to the
646 * object are gone, so we don't need to lock it.
648 if (object->type == OBJT_VNODE) {
649 struct vnode *vp = (struct vnode *)object->handle;
652 * Clean pages and flush buffers.
654 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
655 VM_OBJECT_UNLOCK(object);
657 vinvalbuf(vp, V_SAVE, 0, 0);
659 VM_OBJECT_LOCK(object);
662 KASSERT(object->ref_count == 0,
663 ("vm_object_terminate: object with references, ref_count=%d",
667 * Now free any remaining pages. For internal objects, this also
668 * removes them from paging queues. Don't free wired pages, just
669 * remove them from the object.
671 vm_page_lock_queues();
672 while ((p = TAILQ_FIRST(&object->memq)) != NULL) {
673 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
674 ("vm_object_terminate: freeing busy page %p "
675 "p->busy = %d, p->oflags %x\n", p, p->busy, p->oflags));
676 if (p->wire_count == 0) {
683 vm_page_unlock_queues();
685 #if VM_NRESERVLEVEL > 0
686 if (__predict_false(!LIST_EMPTY(&object->rvq)))
687 vm_reserv_break_all(object);
689 if (__predict_false(object->cache != NULL))
690 vm_page_cache_free(object, 0, 0);
693 * Let the pager know object is dead.
695 vm_pager_deallocate(object);
696 VM_OBJECT_UNLOCK(object);
698 vm_object_destroy(object);
702 * vm_object_page_clean
704 * Clean all dirty pages in the specified range of object. Leaves page
705 * on whatever queue it is currently on. If NOSYNC is set then do not
706 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
707 * leaving the object dirty.
709 * When stuffing pages asynchronously, allow clustering. XXX we need a
710 * synchronous clustering mode implementation.
712 * Odd semantics: if start == end, we clean everything.
714 * The object must be locked.
717 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end, int flags)
720 vm_pindex_t tstart, tend;
726 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
727 if (object->type != OBJT_VNODE ||
728 (object->flags & OBJ_MIGHTBEDIRTY) == 0)
731 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
732 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
734 vm_object_set_flag(object, OBJ_CLEANING);
743 vm_page_lock_queues();
745 * If the caller is smart and only msync()s a range he knows is
746 * dirty, we may be able to avoid an object scan. This results in
747 * a phenominal improvement in performance. We cannot do this
748 * as a matter of course because the object may be huge - e.g.
749 * the size might be in the gigabytes or terrabytes.
751 if (msync_flush_flags & MSYNC_FLUSH_HARDSEQ) {
756 scanreset = object->resident_page_count / EASY_SCAN_FACTOR;
759 pagerflags |= VM_PAGER_IGNORE_CLEANCHK;
761 scanlimit = scanreset;
763 while (tscan < tend) {
764 curgeneration = object->generation;
765 p = vm_page_lookup(object, tscan);
766 if (p == NULL || p->valid == 0) {
767 if (--scanlimit == 0)
772 vm_page_test_dirty(p);
773 if ((p->dirty & p->valid) == 0) {
774 if (--scanlimit == 0)
780 * If we have been asked to skip nosync pages and
781 * this is a nosync page, we can't continue.
783 if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
784 if (--scanlimit == 0)
789 scanlimit = scanreset;
792 * This returns 0 if it was unable to busy the first
793 * page (i.e. had to sleep).
795 tscan += vm_object_page_collect_flush(object, p, curgeneration, pagerflags);
799 * If everything was dirty and we flushed it successfully,
800 * and the requested range is not the entire object, we
801 * don't have to mess with CLEANCHK or MIGHTBEDIRTY and can
802 * return immediately.
804 if (tscan >= tend && (tstart || tend < object->size)) {
805 vm_page_unlock_queues();
806 vm_object_clear_flag(object, OBJ_CLEANING);
809 pagerflags &= ~VM_PAGER_IGNORE_CLEANCHK;
813 * Generally set CLEANCHK interlock and make the page read-only so
814 * we can then clear the object flags.
816 * However, if this is a nosync mmap then the object is likely to
817 * stay dirty so do not mess with the page and do not clear the
821 TAILQ_FOREACH(p, &object->memq, listq) {
822 p->oflags |= VPO_CLEANCHK;
823 if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC))
826 pmap_remove_write(p);
829 if (clearobjflags && (tstart == 0) && (tend == object->size)) {
832 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
833 if (object->type == OBJT_VNODE &&
834 (vp = (struct vnode *)object->handle) != NULL) {
836 if (vp->v_iflag & VI_OBJDIRTY)
837 vp->v_iflag &= ~VI_OBJDIRTY;
843 curgeneration = object->generation;
845 for (p = TAILQ_FIRST(&object->memq); p; p = np) {
848 np = TAILQ_NEXT(p, listq);
852 if ((p->oflags & VPO_CLEANCHK) == 0 ||
853 (pi < tstart) || (pi >= tend) ||
855 p->oflags &= ~VPO_CLEANCHK;
859 vm_page_test_dirty(p);
860 if ((p->dirty & p->valid) == 0) {
861 p->oflags &= ~VPO_CLEANCHK;
866 * If we have been asked to skip nosync pages and this is a
867 * nosync page, skip it. Note that the object flags were
868 * not cleared in this case so we do not have to set them.
870 if ((flags & OBJPC_NOSYNC) && (p->oflags & VPO_NOSYNC)) {
871 p->oflags &= ~VPO_CLEANCHK;
875 n = vm_object_page_collect_flush(object, p,
876 curgeneration, pagerflags);
880 if (object->generation != curgeneration)
884 * Try to optimize the next page. If we can't we pick up
885 * our (random) scan where we left off.
887 if (msync_flush_flags & MSYNC_FLUSH_SOFTSEQ) {
888 if ((p = vm_page_lookup(object, pi + n)) != NULL)
892 vm_page_unlock_queues();
894 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC)?MNT_WAIT:0, curproc);
897 vm_object_clear_flag(object, OBJ_CLEANING);
902 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int curgeneration, int pagerflags)
910 vm_page_t maf[vm_pageout_page_count];
911 vm_page_t mab[vm_pageout_page_count];
912 vm_page_t ma[vm_pageout_page_count];
914 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
916 while (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
917 vm_page_lock_queues();
918 if (object->generation != curgeneration) {
923 for(i = 1; i < vm_pageout_page_count; i++) {
926 if ((tp = vm_page_lookup(object, pi + i)) != NULL) {
927 if ((tp->oflags & VPO_BUSY) ||
928 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
929 (tp->oflags & VPO_CLEANCHK) == 0) ||
932 vm_page_test_dirty(tp);
933 if ((tp->dirty & tp->valid) == 0) {
934 tp->oflags &= ~VPO_CLEANCHK;
945 chkb = vm_pageout_page_count - maxf;
947 for(i = 1; i < chkb;i++) {
950 if ((tp = vm_page_lookup(object, pi - i)) != NULL) {
951 if ((tp->oflags & VPO_BUSY) ||
952 ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
953 (tp->oflags & VPO_CLEANCHK) == 0) ||
956 vm_page_test_dirty(tp);
957 if ((tp->dirty & tp->valid) == 0) {
958 tp->oflags &= ~VPO_CLEANCHK;
969 for(i = 0; i < maxb; i++) {
970 int index = (maxb - i) - 1;
972 ma[index]->oflags &= ~VPO_CLEANCHK;
974 p->oflags &= ~VPO_CLEANCHK;
976 for(i = 0; i < maxf; i++) {
977 int index = (maxb + i) + 1;
979 ma[index]->oflags &= ~VPO_CLEANCHK;
981 runlen = maxb + maxf + 1;
983 vm_pageout_flush(ma, runlen, pagerflags);
984 for (i = 0; i < runlen; i++) {
985 if (ma[i]->valid & ma[i]->dirty) {
986 pmap_remove_write(ma[i]);
987 ma[i]->oflags |= VPO_CLEANCHK;
990 * maxf will end up being the actual number of pages
991 * we wrote out contiguously, non-inclusive of the
992 * first page. We do not count look-behind pages.
994 if (i >= maxb + 1 && (maxf > i - maxb - 1))
1002 * Note that there is absolutely no sense in writing out
1003 * anonymous objects, so we track down the vnode object
1005 * We invalidate (remove) all pages from the address space
1006 * for semantic correctness.
1008 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1009 * may start out with a NULL object.
1012 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1013 boolean_t syncio, boolean_t invalidate)
1015 vm_object_t backing_object;
1022 VM_OBJECT_LOCK(object);
1023 while ((backing_object = object->backing_object) != NULL) {
1024 VM_OBJECT_LOCK(backing_object);
1025 offset += object->backing_object_offset;
1026 VM_OBJECT_UNLOCK(object);
1027 object = backing_object;
1028 if (object->size < OFF_TO_IDX(offset + size))
1029 size = IDX_TO_OFF(object->size) - offset;
1032 * Flush pages if writing is allowed, invalidate them
1033 * if invalidation requested. Pages undergoing I/O
1034 * will be ignored by vm_object_page_remove().
1036 * We cannot lock the vnode and then wait for paging
1037 * to complete without deadlocking against vm_fault.
1038 * Instead we simply call vm_object_page_remove() and
1039 * allow it to block internally on a page-by-page
1040 * basis when it encounters pages undergoing async
1043 if (object->type == OBJT_VNODE &&
1044 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1046 vp = object->handle;
1047 VM_OBJECT_UNLOCK(object);
1048 (void) vn_start_write(vp, &mp, V_WAIT);
1049 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
1050 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1051 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1052 flags |= invalidate ? OBJPC_INVAL : 0;
1053 VM_OBJECT_LOCK(object);
1054 vm_object_page_clean(object,
1056 OFF_TO_IDX(offset + size + PAGE_MASK),
1058 VM_OBJECT_UNLOCK(object);
1060 VFS_UNLOCK_GIANT(vfslocked);
1061 vn_finished_write(mp);
1062 VM_OBJECT_LOCK(object);
1064 if ((object->type == OBJT_VNODE ||
1065 object->type == OBJT_DEVICE) && invalidate) {
1067 purge = old_msync || (object->type == OBJT_DEVICE);
1068 vm_object_page_remove(object,
1070 OFF_TO_IDX(offset + size + PAGE_MASK),
1071 purge ? FALSE : TRUE);
1073 VM_OBJECT_UNLOCK(object);
1077 * vm_object_madvise:
1079 * Implements the madvise function at the object/page level.
1081 * MADV_WILLNEED (any object)
1083 * Activate the specified pages if they are resident.
1085 * MADV_DONTNEED (any object)
1087 * Deactivate the specified pages if they are resident.
1089 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1090 * OBJ_ONEMAPPING only)
1092 * Deactivate and clean the specified pages if they are
1093 * resident. This permits the process to reuse the pages
1094 * without faulting or the kernel to reclaim the pages
1098 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1100 vm_pindex_t end, tpindex;
1101 vm_object_t backing_object, tobject;
1106 VM_OBJECT_LOCK(object);
1107 end = pindex + count;
1109 * Locate and adjust resident pages
1111 for (; pindex < end; pindex += 1) {
1117 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1118 * and those pages must be OBJ_ONEMAPPING.
1120 if (advise == MADV_FREE) {
1121 if ((tobject->type != OBJT_DEFAULT &&
1122 tobject->type != OBJT_SWAP) ||
1123 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1124 goto unlock_tobject;
1127 m = vm_page_lookup(tobject, tpindex);
1128 if (m == NULL && advise == MADV_WILLNEED) {
1130 * If the page is cached, reactivate it.
1132 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1137 * There may be swap even if there is no backing page
1139 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1140 swap_pager_freespace(tobject, tpindex, 1);
1144 backing_object = tobject->backing_object;
1145 if (backing_object == NULL)
1146 goto unlock_tobject;
1147 VM_OBJECT_LOCK(backing_object);
1148 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1149 if (tobject != object)
1150 VM_OBJECT_UNLOCK(tobject);
1151 tobject = backing_object;
1155 * If the page is busy or not in a normal active state,
1156 * we skip it. If the page is not managed there are no
1157 * page queues to mess with. Things can break if we mess
1158 * with pages in any of the below states.
1160 vm_page_lock_queues();
1161 if (m->hold_count ||
1163 (m->flags & PG_UNMANAGED) ||
1164 m->valid != VM_PAGE_BITS_ALL) {
1165 vm_page_unlock_queues();
1166 goto unlock_tobject;
1168 if ((m->oflags & VPO_BUSY) || m->busy) {
1169 vm_page_flag_set(m, PG_REFERENCED);
1170 vm_page_unlock_queues();
1171 if (object != tobject)
1172 VM_OBJECT_UNLOCK(object);
1173 m->oflags |= VPO_WANTED;
1174 msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo", 0);
1175 VM_OBJECT_LOCK(object);
1178 if (advise == MADV_WILLNEED) {
1179 vm_page_activate(m);
1180 } else if (advise == MADV_DONTNEED) {
1181 vm_page_dontneed(m);
1182 } else if (advise == MADV_FREE) {
1184 * Mark the page clean. This will allow the page
1185 * to be freed up by the system. However, such pages
1186 * are often reused quickly by malloc()/free()
1187 * so we do not do anything that would cause
1188 * a page fault if we can help it.
1190 * Specifically, we do not try to actually free
1191 * the page now nor do we try to put it in the
1192 * cache (which would cause a page fault on reuse).
1194 * But we do make the page is freeable as we
1195 * can without actually taking the step of unmapping
1198 pmap_clear_modify(m);
1201 vm_page_dontneed(m);
1203 vm_page_unlock_queues();
1204 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1205 swap_pager_freespace(tobject, tpindex, 1);
1207 if (tobject != object)
1208 VM_OBJECT_UNLOCK(tobject);
1210 VM_OBJECT_UNLOCK(object);
1216 * Create a new object which is backed by the
1217 * specified existing object range. The source
1218 * object reference is deallocated.
1220 * The new object and offset into that object
1221 * are returned in the source parameters.
1225 vm_object_t *object, /* IN/OUT */
1226 vm_ooffset_t *offset, /* IN/OUT */
1235 * Don't create the new object if the old object isn't shared.
1237 if (source != NULL) {
1238 VM_OBJECT_LOCK(source);
1239 if (source->ref_count == 1 &&
1240 source->handle == NULL &&
1241 (source->type == OBJT_DEFAULT ||
1242 source->type == OBJT_SWAP)) {
1243 VM_OBJECT_UNLOCK(source);
1246 VM_OBJECT_UNLOCK(source);
1250 * Allocate a new object with the given length.
1252 result = vm_object_allocate(OBJT_DEFAULT, length);
1255 * The new object shadows the source object, adding a reference to it.
1256 * Our caller changes his reference to point to the new object,
1257 * removing a reference to the source object. Net result: no change
1258 * of reference count.
1260 * Try to optimize the result object's page color when shadowing
1261 * in order to maintain page coloring consistency in the combined
1264 result->backing_object = source;
1266 * Store the offset into the source object, and fix up the offset into
1269 result->backing_object_offset = *offset;
1270 if (source != NULL) {
1271 VM_OBJECT_LOCK(source);
1272 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1273 source->shadow_count++;
1274 source->generation++;
1275 #if VM_NRESERVLEVEL > 0
1276 result->flags |= source->flags & (OBJ_NEEDGIANT | OBJ_COLORED);
1277 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1278 ((1 << (VM_NFREEORDER - 1)) - 1);
1280 result->flags |= source->flags & OBJ_NEEDGIANT;
1282 VM_OBJECT_UNLOCK(source);
1287 * Return the new things
1296 * Split the pages in a map entry into a new object. This affords
1297 * easier removal of unused pages, and keeps object inheritance from
1298 * being a negative impact on memory usage.
1301 vm_object_split(vm_map_entry_t entry)
1303 vm_page_t m, m_next;
1304 vm_object_t orig_object, new_object, source;
1305 vm_pindex_t idx, offidxstart;
1308 orig_object = entry->object.vm_object;
1309 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1311 if (orig_object->ref_count <= 1)
1313 VM_OBJECT_UNLOCK(orig_object);
1315 offidxstart = OFF_TO_IDX(entry->offset);
1316 size = atop(entry->end - entry->start);
1319 * If swap_pager_copy() is later called, it will convert new_object
1320 * into a swap object.
1322 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1325 * At this point, the new object is still private, so the order in
1326 * which the original and new objects are locked does not matter.
1328 VM_OBJECT_LOCK(new_object);
1329 VM_OBJECT_LOCK(orig_object);
1330 source = orig_object->backing_object;
1331 if (source != NULL) {
1332 VM_OBJECT_LOCK(source);
1333 if ((source->flags & OBJ_DEAD) != 0) {
1334 VM_OBJECT_UNLOCK(source);
1335 VM_OBJECT_UNLOCK(orig_object);
1336 VM_OBJECT_UNLOCK(new_object);
1337 vm_object_deallocate(new_object);
1338 VM_OBJECT_LOCK(orig_object);
1341 LIST_INSERT_HEAD(&source->shadow_head,
1342 new_object, shadow_list);
1343 source->shadow_count++;
1344 source->generation++;
1345 vm_object_reference_locked(source); /* for new_object */
1346 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1347 VM_OBJECT_UNLOCK(source);
1348 new_object->backing_object_offset =
1349 orig_object->backing_object_offset + entry->offset;
1350 new_object->backing_object = source;
1352 new_object->flags |= orig_object->flags & OBJ_NEEDGIANT;
1354 if ((m = TAILQ_FIRST(&orig_object->memq)) != NULL) {
1355 if (m->pindex < offidxstart) {
1356 m = vm_page_splay(offidxstart, orig_object->root);
1357 if ((orig_object->root = m)->pindex < offidxstart)
1358 m = TAILQ_NEXT(m, listq);
1361 vm_page_lock_queues();
1362 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1364 m_next = TAILQ_NEXT(m, listq);
1367 * We must wait for pending I/O to complete before we can
1370 * We do not have to VM_PROT_NONE the page as mappings should
1371 * not be changed by this operation.
1373 if ((m->oflags & VPO_BUSY) || m->busy) {
1374 vm_page_flag_set(m, PG_REFERENCED);
1375 vm_page_unlock_queues();
1376 VM_OBJECT_UNLOCK(new_object);
1377 m->oflags |= VPO_WANTED;
1378 msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1379 VM_OBJECT_LOCK(new_object);
1382 vm_page_rename(m, new_object, idx);
1383 /* page automatically made dirty by rename and cache handled */
1386 vm_page_unlock_queues();
1387 if (orig_object->type == OBJT_SWAP) {
1389 * swap_pager_copy() can sleep, in which case the orig_object's
1390 * and new_object's locks are released and reacquired.
1392 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1395 * Transfer any cached pages from orig_object to new_object.
1397 if (__predict_false(orig_object->cache != NULL))
1398 vm_page_cache_transfer(orig_object, offidxstart,
1401 VM_OBJECT_UNLOCK(orig_object);
1402 TAILQ_FOREACH(m, &new_object->memq, listq)
1404 VM_OBJECT_UNLOCK(new_object);
1405 entry->object.vm_object = new_object;
1406 entry->offset = 0LL;
1407 vm_object_deallocate(orig_object);
1408 VM_OBJECT_LOCK(new_object);
1411 #define OBSC_TEST_ALL_SHADOWED 0x0001
1412 #define OBSC_COLLAPSE_NOWAIT 0x0002
1413 #define OBSC_COLLAPSE_WAIT 0x0004
1416 vm_object_backing_scan(vm_object_t object, int op)
1420 vm_object_t backing_object;
1421 vm_pindex_t backing_offset_index;
1423 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1424 VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1426 backing_object = object->backing_object;
1427 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1430 * Initial conditions
1432 if (op & OBSC_TEST_ALL_SHADOWED) {
1434 * We do not want to have to test for the existence of cache
1435 * or swap pages in the backing object. XXX but with the
1436 * new swapper this would be pretty easy to do.
1438 * XXX what about anonymous MAP_SHARED memory that hasn't
1439 * been ZFOD faulted yet? If we do not test for this, the
1440 * shadow test may succeed! XXX
1442 if (backing_object->type != OBJT_DEFAULT) {
1446 if (op & OBSC_COLLAPSE_WAIT) {
1447 vm_object_set_flag(backing_object, OBJ_DEAD);
1453 p = TAILQ_FIRST(&backing_object->memq);
1455 vm_page_t next = TAILQ_NEXT(p, listq);
1456 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1458 if (op & OBSC_TEST_ALL_SHADOWED) {
1462 * Ignore pages outside the parent object's range
1463 * and outside the parent object's mapping of the
1466 * note that we do not busy the backing object's
1470 p->pindex < backing_offset_index ||
1471 new_pindex >= object->size
1478 * See if the parent has the page or if the parent's
1479 * object pager has the page. If the parent has the
1480 * page but the page is not valid, the parent's
1481 * object pager must have the page.
1483 * If this fails, the parent does not completely shadow
1484 * the object and we might as well give up now.
1487 pp = vm_page_lookup(object, new_pindex);
1489 (pp == NULL || pp->valid == 0) &&
1490 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1498 * Check for busy page
1500 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1503 if (op & OBSC_COLLAPSE_NOWAIT) {
1504 if ((p->oflags & VPO_BUSY) ||
1510 } else if (op & OBSC_COLLAPSE_WAIT) {
1511 if ((p->oflags & VPO_BUSY) || p->busy) {
1512 vm_page_lock_queues();
1513 vm_page_flag_set(p, PG_REFERENCED);
1514 vm_page_unlock_queues();
1515 VM_OBJECT_UNLOCK(object);
1516 p->oflags |= VPO_WANTED;
1517 msleep(p, VM_OBJECT_MTX(backing_object),
1518 PDROP | PVM, "vmocol", 0);
1519 VM_OBJECT_LOCK(object);
1520 VM_OBJECT_LOCK(backing_object);
1522 * If we slept, anything could have
1523 * happened. Since the object is
1524 * marked dead, the backing offset
1525 * should not have changed so we
1526 * just restart our scan.
1528 p = TAILQ_FIRST(&backing_object->memq);
1534 p->object == backing_object,
1535 ("vm_object_backing_scan: object mismatch")
1539 * Destroy any associated swap
1541 if (backing_object->type == OBJT_SWAP) {
1542 swap_pager_freespace(
1550 p->pindex < backing_offset_index ||
1551 new_pindex >= object->size
1554 * Page is out of the parent object's range, we
1555 * can simply destroy it.
1557 vm_page_lock_queues();
1558 KASSERT(!pmap_page_is_mapped(p),
1559 ("freeing mapped page %p", p));
1560 if (p->wire_count == 0)
1564 vm_page_unlock_queues();
1569 pp = vm_page_lookup(object, new_pindex);
1572 vm_pager_has_page(object, new_pindex, NULL, NULL)
1575 * page already exists in parent OR swap exists
1576 * for this location in the parent. Destroy
1577 * the original page from the backing object.
1579 * Leave the parent's page alone
1581 vm_page_lock_queues();
1582 KASSERT(!pmap_page_is_mapped(p),
1583 ("freeing mapped page %p", p));
1584 if (p->wire_count == 0)
1588 vm_page_unlock_queues();
1593 #if VM_NRESERVLEVEL > 0
1595 * Rename the reservation.
1597 vm_reserv_rename(p, object, backing_object,
1598 backing_offset_index);
1602 * Page does not exist in parent, rename the
1603 * page from the backing object to the main object.
1605 * If the page was mapped to a process, it can remain
1606 * mapped through the rename.
1608 vm_page_lock_queues();
1609 vm_page_rename(p, object, new_pindex);
1610 vm_page_unlock_queues();
1611 /* page automatically made dirty by rename */
1620 * this version of collapse allows the operation to occur earlier and
1621 * when paging_in_progress is true for an object... This is not a complete
1622 * operation, but should plug 99.9% of the rest of the leaks.
1625 vm_object_qcollapse(vm_object_t object)
1627 vm_object_t backing_object = object->backing_object;
1629 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1630 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1632 if (backing_object->ref_count != 1)
1635 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1639 * vm_object_collapse:
1641 * Collapse an object with the object backing it.
1642 * Pages in the backing object are moved into the
1643 * parent, and the backing object is deallocated.
1646 vm_object_collapse(vm_object_t object)
1648 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1651 vm_object_t backing_object;
1654 * Verify that the conditions are right for collapse:
1656 * The object exists and the backing object exists.
1658 if ((backing_object = object->backing_object) == NULL)
1662 * we check the backing object first, because it is most likely
1665 VM_OBJECT_LOCK(backing_object);
1666 if (backing_object->handle != NULL ||
1667 (backing_object->type != OBJT_DEFAULT &&
1668 backing_object->type != OBJT_SWAP) ||
1669 (backing_object->flags & OBJ_DEAD) ||
1670 object->handle != NULL ||
1671 (object->type != OBJT_DEFAULT &&
1672 object->type != OBJT_SWAP) ||
1673 (object->flags & OBJ_DEAD)) {
1674 VM_OBJECT_UNLOCK(backing_object);
1679 object->paging_in_progress != 0 ||
1680 backing_object->paging_in_progress != 0
1682 vm_object_qcollapse(object);
1683 VM_OBJECT_UNLOCK(backing_object);
1687 * We know that we can either collapse the backing object (if
1688 * the parent is the only reference to it) or (perhaps) have
1689 * the parent bypass the object if the parent happens to shadow
1690 * all the resident pages in the entire backing object.
1692 * This is ignoring pager-backed pages such as swap pages.
1693 * vm_object_backing_scan fails the shadowing test in this
1696 if (backing_object->ref_count == 1) {
1698 * If there is exactly one reference to the backing
1699 * object, we can collapse it into the parent.
1701 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1703 #if VM_NRESERVLEVEL > 0
1705 * Break any reservations from backing_object.
1707 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1708 vm_reserv_break_all(backing_object);
1712 * Move the pager from backing_object to object.
1714 if (backing_object->type == OBJT_SWAP) {
1716 * swap_pager_copy() can sleep, in which case
1717 * the backing_object's and object's locks are
1718 * released and reacquired.
1723 OFF_TO_IDX(object->backing_object_offset), TRUE);
1726 * Free any cached pages from backing_object.
1728 if (__predict_false(backing_object->cache != NULL))
1729 vm_page_cache_free(backing_object, 0, 0);
1732 * Object now shadows whatever backing_object did.
1733 * Note that the reference to
1734 * backing_object->backing_object moves from within
1735 * backing_object to within object.
1737 LIST_REMOVE(object, shadow_list);
1738 backing_object->shadow_count--;
1739 backing_object->generation++;
1740 if (backing_object->backing_object) {
1741 VM_OBJECT_LOCK(backing_object->backing_object);
1742 LIST_REMOVE(backing_object, shadow_list);
1744 &backing_object->backing_object->shadow_head,
1745 object, shadow_list);
1747 * The shadow_count has not changed.
1749 backing_object->backing_object->generation++;
1750 VM_OBJECT_UNLOCK(backing_object->backing_object);
1752 object->backing_object = backing_object->backing_object;
1753 object->backing_object_offset +=
1754 backing_object->backing_object_offset;
1757 * Discard backing_object.
1759 * Since the backing object has no pages, no pager left,
1760 * and no object references within it, all that is
1761 * necessary is to dispose of it.
1763 KASSERT(backing_object->ref_count == 1, ("backing_object %p was somehow re-referenced during collapse!", backing_object));
1764 VM_OBJECT_UNLOCK(backing_object);
1766 mtx_lock(&vm_object_list_mtx);
1772 mtx_unlock(&vm_object_list_mtx);
1774 uma_zfree(obj_zone, backing_object);
1778 vm_object_t new_backing_object;
1781 * If we do not entirely shadow the backing object,
1782 * there is nothing we can do so we give up.
1784 if (object->resident_page_count != object->size &&
1785 vm_object_backing_scan(object,
1786 OBSC_TEST_ALL_SHADOWED) == 0) {
1787 VM_OBJECT_UNLOCK(backing_object);
1792 * Make the parent shadow the next object in the
1793 * chain. Deallocating backing_object will not remove
1794 * it, since its reference count is at least 2.
1796 LIST_REMOVE(object, shadow_list);
1797 backing_object->shadow_count--;
1798 backing_object->generation++;
1800 new_backing_object = backing_object->backing_object;
1801 if ((object->backing_object = new_backing_object) != NULL) {
1802 VM_OBJECT_LOCK(new_backing_object);
1804 &new_backing_object->shadow_head,
1808 new_backing_object->shadow_count++;
1809 new_backing_object->generation++;
1810 vm_object_reference_locked(new_backing_object);
1811 VM_OBJECT_UNLOCK(new_backing_object);
1812 object->backing_object_offset +=
1813 backing_object->backing_object_offset;
1817 * Drop the reference count on backing_object. Since
1818 * its ref_count was at least 2, it will not vanish.
1820 backing_object->ref_count--;
1821 VM_OBJECT_UNLOCK(backing_object);
1826 * Try again with this object's new backing object.
1832 * vm_object_page_remove:
1834 * For the given object, either frees or invalidates each of the
1835 * specified pages. In general, a page is freed. However, if a
1836 * page is wired for any reason other than the existence of a
1837 * managed, wired mapping, then it may be invalidated but not
1838 * removed from the object. Pages are specified by the given
1839 * range ["start", "end") and Boolean "clean_only". As a
1840 * special case, if "end" is zero, then the range extends from
1841 * "start" to the end of the object. If "clean_only" is TRUE,
1842 * then only the non-dirty pages within the specified range are
1845 * In general, this operation should only be performed on objects
1846 * that contain managed pages. There are two exceptions. First,
1847 * it may be performed on the kernel and kmem objects. Second,
1848 * it may be used by msync(..., MS_INVALIDATE) to invalidate
1849 * device-backed pages.
1851 * The object must be locked.
1854 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1855 boolean_t clean_only)
1860 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1861 if (object->resident_page_count == 0)
1865 * Since physically-backed objects do not use managed pages, we can't
1866 * remove pages from the object (we must instead remove the page
1867 * references, and then destroy the object).
1869 KASSERT(object->type != OBJT_PHYS || object == kernel_object ||
1870 object == kmem_object,
1871 ("attempt to remove pages from a physical object"));
1873 vm_object_pip_add(object, 1);
1875 vm_page_lock_queues();
1876 if ((p = TAILQ_FIRST(&object->memq)) != NULL) {
1877 if (p->pindex < start) {
1878 p = vm_page_splay(start, object->root);
1879 if ((object->root = p)->pindex < start)
1880 p = TAILQ_NEXT(p, listq);
1884 * Assert: the variable p is either (1) the page with the
1885 * least pindex greater than or equal to the parameter pindex
1889 p != NULL && (p->pindex < end || end == 0);
1891 next = TAILQ_NEXT(p, listq);
1894 * If the page is wired for any reason besides the
1895 * existence of managed, wired mappings, then it cannot
1896 * be freed. For example, fictitious pages, which
1897 * represent device memory, are inherently wired and
1898 * cannot be freed. They can, however, be invalidated
1899 * if "clean_only" is FALSE.
1901 if ((wirings = p->wire_count) != 0 &&
1902 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1903 /* Fictitious pages do not have managed mappings. */
1904 if ((p->flags & PG_FICTITIOUS) == 0)
1906 /* Account for removal of managed, wired mappings. */
1907 p->wire_count -= wirings;
1912 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1914 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1915 ("vm_object_page_remove: page %p is fictitious", p));
1916 if (clean_only && p->valid) {
1917 pmap_remove_write(p);
1918 if (p->valid & p->dirty)
1922 /* Account for removal of managed, wired mappings. */
1924 p->wire_count -= wirings;
1927 vm_page_unlock_queues();
1928 vm_object_pip_wakeup(object);
1930 if (__predict_false(object->cache != NULL))
1931 vm_page_cache_free(object, start, end);
1935 * Routine: vm_object_coalesce
1936 * Function: Coalesces two objects backing up adjoining
1937 * regions of memory into a single object.
1939 * returns TRUE if objects were combined.
1941 * NOTE: Only works at the moment if the second object is NULL -
1942 * if it's not, which object do we lock first?
1945 * prev_object First object to coalesce
1946 * prev_offset Offset into prev_object
1947 * prev_size Size of reference to prev_object
1948 * next_size Size of reference to the second object
1951 * The object must *not* be locked.
1954 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
1955 vm_size_t prev_size, vm_size_t next_size)
1957 vm_pindex_t next_pindex;
1959 if (prev_object == NULL)
1961 VM_OBJECT_LOCK(prev_object);
1962 if (prev_object->type != OBJT_DEFAULT &&
1963 prev_object->type != OBJT_SWAP) {
1964 VM_OBJECT_UNLOCK(prev_object);
1969 * Try to collapse the object first
1971 vm_object_collapse(prev_object);
1974 * Can't coalesce if: . more than one reference . paged out . shadows
1975 * another object . has a copy elsewhere (any of which mean that the
1976 * pages not mapped to prev_entry may be in use anyway)
1978 if (prev_object->backing_object != NULL) {
1979 VM_OBJECT_UNLOCK(prev_object);
1983 prev_size >>= PAGE_SHIFT;
1984 next_size >>= PAGE_SHIFT;
1985 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
1987 if ((prev_object->ref_count > 1) &&
1988 (prev_object->size != next_pindex)) {
1989 VM_OBJECT_UNLOCK(prev_object);
1994 * Remove any pages that may still be in the object from a previous
1997 if (next_pindex < prev_object->size) {
1998 vm_object_page_remove(prev_object,
2000 next_pindex + next_size, FALSE);
2001 if (prev_object->type == OBJT_SWAP)
2002 swap_pager_freespace(prev_object,
2003 next_pindex, next_size);
2007 * Extend the object if necessary.
2009 if (next_pindex + next_size > prev_object->size)
2010 prev_object->size = next_pindex + next_size;
2012 VM_OBJECT_UNLOCK(prev_object);
2017 vm_object_set_writeable_dirty(vm_object_t object)
2021 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2022 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2024 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2025 if (object->type == OBJT_VNODE &&
2026 (vp = (struct vnode *)object->handle) != NULL) {
2028 vp->v_iflag |= VI_OBJDIRTY;
2033 #include "opt_ddb.h"
2035 #include <sys/kernel.h>
2037 #include <sys/cons.h>
2039 #include <ddb/ddb.h>
2042 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2045 vm_map_entry_t tmpe;
2053 tmpe = map->header.next;
2054 entcount = map->nentries;
2055 while (entcount-- && (tmpe != &map->header)) {
2056 if (_vm_object_in_map(map, object, tmpe)) {
2061 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2062 tmpm = entry->object.sub_map;
2063 tmpe = tmpm->header.next;
2064 entcount = tmpm->nentries;
2065 while (entcount-- && tmpe != &tmpm->header) {
2066 if (_vm_object_in_map(tmpm, object, tmpe)) {
2071 } else if ((obj = entry->object.vm_object) != NULL) {
2072 for (; obj; obj = obj->backing_object)
2073 if (obj == object) {
2081 vm_object_in_map(vm_object_t object)
2085 /* sx_slock(&allproc_lock); */
2086 FOREACH_PROC_IN_SYSTEM(p) {
2087 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2089 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2090 /* sx_sunlock(&allproc_lock); */
2094 /* sx_sunlock(&allproc_lock); */
2095 if (_vm_object_in_map(kernel_map, object, 0))
2097 if (_vm_object_in_map(kmem_map, object, 0))
2099 if (_vm_object_in_map(pager_map, object, 0))
2101 if (_vm_object_in_map(buffer_map, object, 0))
2106 DB_SHOW_COMMAND(vmochk, vm_object_check)
2111 * make sure that internal objs are in a map somewhere
2112 * and none have zero ref counts.
2114 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2115 if (object->handle == NULL &&
2116 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2117 if (object->ref_count == 0) {
2118 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2119 (long)object->size);
2121 if (!vm_object_in_map(object)) {
2123 "vmochk: internal obj is not in a map: "
2124 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2125 object->ref_count, (u_long)object->size,
2126 (u_long)object->size,
2127 (void *)object->backing_object);
2134 * vm_object_print: [ debug ]
2136 DB_SHOW_COMMAND(object, vm_object_print_static)
2138 /* XXX convert args. */
2139 vm_object_t object = (vm_object_t)addr;
2140 boolean_t full = have_addr;
2144 /* XXX count is an (unused) arg. Avoid shadowing it. */
2145 #define count was_count
2153 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x\n",
2154 object, (int)object->type, (uintmax_t)object->size,
2155 object->resident_page_count, object->ref_count, object->flags);
2156 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2157 object->shadow_count,
2158 object->backing_object ? object->backing_object->ref_count : 0,
2159 object->backing_object, (uintmax_t)object->backing_object_offset);
2166 TAILQ_FOREACH(p, &object->memq, listq) {
2168 db_iprintf("memory:=");
2169 else if (count == 6) {
2177 db_printf("(off=0x%jx,page=0x%jx)",
2178 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2188 /* XXX need this non-static entry for calling from vm_map_print. */
2191 /* db_expr_t */ long addr,
2192 boolean_t have_addr,
2193 /* db_expr_t */ long count,
2196 vm_object_print_static(addr, have_addr, count, modif);
2199 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2205 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2206 vm_pindex_t idx, fidx;
2212 db_printf("new object: %p\n", (void *)object);
2222 osize = object->size;
2225 for (idx = 0; idx < osize; idx++) {
2226 m = vm_page_lookup(object, idx);
2229 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2230 (long)fidx, rcount, (long)pa);
2245 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2250 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2251 (long)fidx, rcount, (long)pa);
2261 pa = VM_PAGE_TO_PHYS(m);
2265 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2266 (long)fidx, rcount, (long)pa);