4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. All advertising materials mentioning features or use of this software
19 * must display the following acknowledgement:
20 * This product includes software developed by the University of
21 * California, Berkeley and its contributors.
22 * 4. Neither the name of the University nor the names of its contributors
23 * may be used to endorse or promote products derived from this software
24 * without specific prior written permission.
26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
38 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94
41 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42 * All rights reserved.
44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
46 * Permission to use, copy, modify and distribute this software and
47 * its documentation is hereby granted, provided that both the copyright
48 * notice and this permission notice appear in all copies of the
49 * software, derivative works or modified versions, and any portions
50 * thereof, and that both notices appear in supporting documentation.
52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
56 * Carnegie Mellon requests users of this software to return to
58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
59 * School of Computer Science
60 * Carnegie Mellon University
61 * Pittsburgh PA 15213-3890
63 * any improvements or extensions that they make and grant Carnegie the
64 * rights to redistribute these changes.
66 * $FreeBSD: src/sys/vm/vm_object.c,v 1.171.2.8 2003/05/26 19:17:56 alc Exp $
70 * Virtual memory object module.
73 #include <sys/param.h>
74 #include <sys/systm.h>
75 #include <sys/proc.h> /* for curproc, pageproc */
76 #include <sys/thread.h>
77 #include <sys/vnode.h>
78 #include <sys/vmmeter.h>
80 #include <sys/mount.h>
81 #include <sys/kernel.h>
82 #include <sys/sysctl.h>
83 #include <sys/refcount.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/swap_pager.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_zone.h>
98 #define EASY_SCAN_FACTOR 8
100 static void vm_object_qcollapse(vm_object_t object,
101 vm_object_t backing_object);
102 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
104 static void vm_object_lock_init(vm_object_t);
108 * Virtual memory objects maintain the actual data
109 * associated with allocated virtual memory. A given
110 * page of memory exists within exactly one object.
112 * An object is only deallocated when all "references"
113 * are given up. Only one "reference" to a given
114 * region of an object should be writeable.
116 * Associated with each object is a list of all resident
117 * memory pages belonging to that object; this list is
118 * maintained by the "vm_page" module, and locked by the object's
121 * Each object also records a "pager" routine which is
122 * used to retrieve (and store) pages to the proper backing
123 * storage. In addition, objects may be backed by other
124 * objects from which they were virtual-copied.
126 * The only items within the object structure which are
127 * modified after time of creation are:
128 * reference count locked by object's lock
129 * pager routine locked by object's lock
133 struct object_q vm_object_list; /* locked by vmobj_token */
134 struct vm_object kernel_object;
136 static long vm_object_count; /* locked by vmobj_token */
137 extern int vm_pageout_page_count;
139 static long object_collapses;
140 static long object_bypasses;
141 static int next_index;
142 static vm_zone_t obj_zone;
143 static struct vm_zone obj_zone_store;
144 #define VM_OBJECTS_INIT 256
145 static struct vm_object vm_objects_init[VM_OBJECTS_INIT];
148 * Misc low level routines
151 vm_object_lock_init(vm_object_t obj)
153 #if defined(DEBUG_LOCKS)
156 obj->debug_hold_bitmap = 0;
157 obj->debug_hold_ovfl = 0;
158 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) {
159 obj->debug_hold_thrs[i] = NULL;
160 obj->debug_hold_file[i] = NULL;
161 obj->debug_hold_line[i] = 0;
167 vm_object_lock_swap(void)
173 vm_object_lock(vm_object_t obj)
175 lwkt_getpooltoken(obj);
179 vm_object_unlock(vm_object_t obj)
181 lwkt_relpooltoken(obj);
185 vm_object_assert_held(vm_object_t obj)
187 ASSERT_LWKT_TOKEN_HELD(lwkt_token_pool_lookup(obj));
192 vm_object_hold(vm_object_t obj)
194 debugvm_object_hold(vm_object_t obj, char *file, int line)
197 KKASSERT(obj != NULL);
200 * Object must be held (object allocation is stable due to callers
201 * context, typically already holding the token on a parent object)
202 * prior to potentially blocking on the lock, otherwise the object
203 * can get ripped away from us.
205 refcount_acquire(&obj->hold_count);
208 #if defined(DEBUG_LOCKS)
211 i = ffs(~obj->debug_hold_bitmap) - 1;
213 kprintf("vm_object hold count > VMOBJ_DEBUG_ARRAY_SIZE");
214 obj->debug_hold_ovfl = 1;
217 obj->debug_hold_bitmap |= (1 << i);
218 obj->debug_hold_thrs[i] = curthread;
219 obj->debug_hold_file[i] = file;
220 obj->debug_hold_line[i] = line;
225 * Drop the token and hold_count on the object.
228 vm_object_drop(vm_object_t obj)
233 #if defined(DEBUG_LOCKS)
237 for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) {
238 if ((obj->debug_hold_bitmap & (1 << i)) &&
239 (obj->debug_hold_thrs[i] == curthread)) {
240 obj->debug_hold_bitmap &= ~(1 << i);
241 obj->debug_hold_thrs[i] = NULL;
242 obj->debug_hold_file[i] = NULL;
243 obj->debug_hold_line[i] = 0;
249 if (found == 0 && obj->debug_hold_ovfl == 0)
250 panic("vm_object: attempt to drop hold on non-self-held obj");
254 * The lock is a pool token, no new holders should be possible once
255 * we drop hold_count 1->0 as there is no longer any way to reference
258 if (refcount_release(&obj->hold_count)) {
259 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD))
260 zfree(obj_zone, obj);
262 vm_object_unlock(obj); /* uses pool token, ok to call on freed obj */
266 * Initialize a freshly allocated object
268 * Used only by vm_object_allocate() and zinitna().
273 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
277 RB_INIT(&object->rb_memq);
278 LIST_INIT(&object->shadow_head);
282 object->ref_count = 1;
283 object->hold_count = 0;
285 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
286 vm_object_set_flag(object, OBJ_ONEMAPPING);
287 object->paging_in_progress = 0;
288 object->resident_page_count = 0;
289 object->agg_pv_list_count = 0;
290 object->shadow_count = 0;
292 /* cpu localization twist */
293 object->pg_color = (int)(intptr_t)curthread;
295 object->pg_color = next_index;
297 if ( size > (PQ_L2_SIZE / 3 + PQ_PRIME1))
298 incr = PQ_L2_SIZE / 3 + PQ_PRIME1;
301 next_index = (next_index + incr) & PQ_L2_MASK;
302 object->handle = NULL;
303 object->backing_object = NULL;
304 object->backing_object_offset = (vm_ooffset_t)0;
306 object->generation++;
307 object->swblock_count = 0;
308 RB_INIT(&object->swblock_root);
309 vm_object_lock_init(object);
311 lwkt_gettoken(&vmobj_token);
312 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
314 lwkt_reltoken(&vmobj_token);
318 * Initialize the VM objects module.
320 * Called from the low level boot code only.
325 TAILQ_INIT(&vm_object_list);
327 _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd),
330 obj_zone = &obj_zone_store;
331 zbootinit(obj_zone, "VM OBJECT", sizeof (struct vm_object),
332 vm_objects_init, VM_OBJECTS_INIT);
336 vm_object_init2(void)
338 zinitna(obj_zone, NULL, NULL, 0, 0, ZONE_PANICFAIL, 1);
342 * Allocate and return a new object of the specified type and size.
347 vm_object_allocate(objtype_t type, vm_pindex_t size)
351 result = (vm_object_t) zalloc(obj_zone);
353 _vm_object_allocate(type, size, result);
359 * Add an additional reference to a vm_object. The object must already be
360 * held. The original non-lock version is no longer supported. The object
361 * must NOT be chain locked by anyone at the time the reference is added.
363 * Referencing a chain-locked object can blow up the fairly sensitive
364 * ref_count and shadow_count tests in the deallocator. Most callers
365 * will call vm_object_chain_wait() prior to calling
366 * vm_object_reference_locked() to avoid the case.
368 * The object must be held.
371 vm_object_reference_locked(vm_object_t object)
373 KKASSERT(object != NULL);
374 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
375 KKASSERT((object->flags & OBJ_CHAINLOCK) == 0);
377 if (object->type == OBJT_VNODE) {
378 vref(object->handle);
379 /* XXX what if the vnode is being destroyed? */
384 * Object OBJ_CHAINLOCK lock handling.
386 * The caller can chain-lock backing objects recursively and then
387 * use vm_object_chain_release_all() to undo the whole chain.
389 * Chain locks are used to prevent collapses and are only applicable
390 * to OBJT_DEFAULT and OBJT_SWAP objects. Chain locking operations
391 * on other object types are ignored. This is also important because
392 * it allows e.g. the vnode underlying a memory mapping to take concurrent
395 * The object must usually be held on entry, though intermediate
396 * objects need not be held on release.
399 vm_object_chain_wait(vm_object_t object)
401 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
402 while (object->flags & OBJ_CHAINLOCK) {
403 vm_object_set_flag(object, OBJ_CHAINWANT);
404 tsleep(object, 0, "objchain", 0);
409 vm_object_chain_acquire(vm_object_t object)
411 if (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) {
412 vm_object_chain_wait(object);
413 vm_object_set_flag(object, OBJ_CHAINLOCK);
418 vm_object_chain_release(vm_object_t object)
420 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
421 if (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) {
422 KKASSERT(object->flags & OBJ_CHAINLOCK);
423 if (object->flags & OBJ_CHAINWANT) {
424 vm_object_clear_flag(object,
425 OBJ_CHAINLOCK | OBJ_CHAINWANT);
428 vm_object_clear_flag(object, OBJ_CHAINLOCK);
434 * This releases the entire chain of objects from first_object to and
435 * including stopobj, flowing through object->backing_object.
437 * We release stopobj first as an optimization as this object is most
438 * likely to be shared across multiple processes.
441 vm_object_chain_release_all(vm_object_t first_object, vm_object_t stopobj)
443 vm_object_t backing_object;
446 vm_object_chain_release(stopobj);
447 object = first_object;
449 while (object != stopobj) {
451 if (object != first_object)
452 vm_object_hold(object);
453 backing_object = object->backing_object;
454 vm_object_chain_release(object);
455 if (object != first_object)
456 vm_object_drop(object);
457 object = backing_object;
462 * Dereference an object and its underlying vnode.
464 * The object must be held and will be held on return.
467 vm_object_vndeallocate(vm_object_t object)
469 struct vnode *vp = (struct vnode *) object->handle;
471 KASSERT(object->type == OBJT_VNODE,
472 ("vm_object_vndeallocate: not a vnode object"));
473 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
474 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
476 if (object->ref_count == 0) {
477 vprint("vm_object_vndeallocate", vp);
478 panic("vm_object_vndeallocate: bad object reference count");
482 if (object->ref_count == 0)
483 vclrflags(vp, VTEXT);
488 * Release a reference to the specified object, gained either through a
489 * vm_object_allocate or a vm_object_reference call. When all references
490 * are gone, storage associated with this object may be relinquished.
492 * The caller does not have to hold the object locked but must have control
493 * over the reference in question in order to guarantee that the object
494 * does not get ripped out from under us.
497 vm_object_deallocate(vm_object_t object)
500 vm_object_hold(object);
501 vm_object_deallocate_locked(object);
502 vm_object_drop(object);
507 vm_object_deallocate_locked(vm_object_t object)
509 struct vm_object_dealloc_list *dlist = NULL;
510 struct vm_object_dealloc_list *dtmp;
515 * We may chain deallocate object, but additional objects may
516 * collect on the dlist which also have to be deallocated. We
517 * must avoid a recursion, vm_object chains can get deep.
520 while (object != NULL) {
523 * Don't rip a ref_count out from under an object undergoing
524 * collapse, it will confuse the collapse code.
526 vm_object_chain_wait(object);
528 if (object->type == OBJT_VNODE) {
529 vm_object_vndeallocate(object);
533 if (object->ref_count == 0) {
534 panic("vm_object_deallocate: object deallocated "
535 "too many times: %d", object->type);
537 if (object->ref_count > 2) {
543 * Here on ref_count of one or two, which are special cases for
546 * Nominal ref_count > 1 case if the second ref is not from
549 if (object->ref_count == 2 && object->shadow_count == 0) {
550 vm_object_set_flag(object, OBJ_ONEMAPPING);
556 * If the second ref is from a shadow we chain along it
557 * upwards if object's handle is exhausted.
559 * We have to decrement object->ref_count before potentially
560 * collapsing the first shadow object or the collapse code
561 * will not be able to handle the degenerate case to remove
562 * object. However, if we do it too early the object can
563 * get ripped out from under us.
565 if (object->ref_count == 2 && object->shadow_count == 1 &&
566 object->handle == NULL && (object->type == OBJT_DEFAULT ||
567 object->type == OBJT_SWAP)) {
568 temp = LIST_FIRST(&object->shadow_head);
569 KKASSERT(temp != NULL);
570 vm_object_hold(temp);
573 * Wait for any paging to complete so the collapse
574 * doesn't (or isn't likely to) qcollapse. pip
575 * waiting must occur before we acquire the
579 temp->paging_in_progress ||
580 object->paging_in_progress
582 vm_object_pip_wait(temp, "objde1");
583 vm_object_pip_wait(object, "objde2");
587 * If the parent is locked we have to give up, as
588 * otherwise we would be acquiring locks in the
589 * wrong order and potentially deadlock.
591 if (temp->flags & OBJ_CHAINLOCK) {
592 vm_object_drop(temp);
595 vm_object_chain_acquire(temp);
598 * Recheck/retry after the hold and the paging
599 * wait, both of which can block us.
601 if (object->ref_count != 2 ||
602 object->shadow_count != 1 ||
604 LIST_FIRST(&object->shadow_head) != temp ||
605 (object->type != OBJT_DEFAULT &&
606 object->type != OBJT_SWAP)) {
607 vm_object_chain_release(temp);
608 vm_object_drop(temp);
613 * We can safely drop object's ref_count now.
615 KKASSERT(object->ref_count == 2);
619 * If our single parent is not collapseable just
620 * decrement ref_count (2->1) and stop.
622 if (temp->handle || (temp->type != OBJT_DEFAULT &&
623 temp->type != OBJT_SWAP)) {
624 vm_object_chain_release(temp);
625 vm_object_drop(temp);
630 * At this point we have already dropped object's
631 * ref_count so it is possible for a race to
632 * deallocate obj out from under us. Any collapse
633 * will re-check the situation. We must not block
634 * until we are able to collapse.
636 * Bump temp's ref_count to avoid an unwanted
637 * degenerate recursion (can't call
638 * vm_object_reference_locked() because it asserts
639 * that CHAINLOCK is not set).
642 KKASSERT(temp->ref_count > 1);
645 * Collapse temp, then deallocate the extra ref
648 vm_object_collapse(temp, &dlist);
649 vm_object_chain_release(temp);
651 vm_object_lock_swap();
652 vm_object_drop(object);
660 * Drop the ref and handle termination on the 1->0 transition.
661 * We may have blocked above so we have to recheck.
664 KKASSERT(object->ref_count != 0);
665 if (object->ref_count >= 2) {
669 KKASSERT(object->ref_count == 1);
672 * 1->0 transition. Chain through the backing_object.
673 * Maintain the ref until we've located the backing object,
676 while ((temp = object->backing_object) != NULL) {
677 vm_object_hold(temp);
678 if (temp == object->backing_object)
680 vm_object_drop(temp);
684 * 1->0 transition verified, retry if ref_count is no longer
685 * 1. Otherwise disconnect the backing_object (temp) and
688 if (object->ref_count != 1) {
689 vm_object_drop(temp);
694 * It shouldn't be possible for the object to be chain locked
695 * if we're removing the last ref on it.
697 KKASSERT((object->flags & OBJ_CHAINLOCK) == 0);
700 LIST_REMOVE(object, shadow_list);
701 temp->shadow_count--;
703 object->backing_object = NULL;
707 if ((object->flags & OBJ_DEAD) == 0)
708 vm_object_terminate(object);
709 if (must_drop && temp)
710 vm_object_lock_swap();
712 vm_object_drop(object);
716 if (must_drop && object)
717 vm_object_drop(object);
720 * Additional tail recursion on dlist. Avoid a recursion. Objects
721 * on the dlist have a hold count but are not locked.
723 if ((dtmp = dlist) != NULL) {
725 object = dtmp->object;
728 vm_object_lock(object); /* already held, add lock */
729 must_drop = 1; /* and we're responsible for it */
735 * Destroy the specified object, freeing up related resources.
737 * The object must have zero references.
739 * The object must held. The caller is responsible for dropping the object
740 * after terminate returns. Terminate does NOT drop the object.
742 static int vm_object_terminate_callback(vm_page_t p, void *data);
745 vm_object_terminate(vm_object_t object)
748 * Make sure no one uses us. Once we set OBJ_DEAD we should be
749 * able to safely block.
751 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
752 KKASSERT((object->flags & OBJ_DEAD) == 0);
753 vm_object_set_flag(object, OBJ_DEAD);
756 * Wait for the pageout daemon to be done with the object
758 vm_object_pip_wait(object, "objtrm1");
760 KASSERT(!object->paging_in_progress,
761 ("vm_object_terminate: pageout in progress"));
764 * Clean and free the pages, as appropriate. All references to the
765 * object are gone, so we don't need to lock it.
767 if (object->type == OBJT_VNODE) {
771 * Clean pages and flush buffers.
773 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
775 vp = (struct vnode *) object->handle;
776 vinvalbuf(vp, V_SAVE, 0, 0);
780 * Wait for any I/O to complete, after which there had better not
781 * be any references left on the object.
783 vm_object_pip_wait(object, "objtrm2");
785 if (object->ref_count != 0) {
786 panic("vm_object_terminate: object with references, "
787 "ref_count=%d", object->ref_count);
791 * Now free any remaining pages. For internal objects, this also
792 * removes them from paging queues. Don't free wired pages, just
793 * remove them from the object.
795 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
796 vm_object_terminate_callback, NULL);
799 * Let the pager know object is dead.
801 vm_pager_deallocate(object);
804 * Wait for the object hold count to hit 1, clean out pages as
805 * we go. vmobj_token interlocks any race conditions that might
806 * pick the object up from the vm_object_list after we have cleared
810 if (RB_ROOT(&object->rb_memq) == NULL)
812 kprintf("vm_object_terminate: Warning, object %p "
813 "still has %d pages\n",
814 object, object->resident_page_count);
815 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
816 vm_object_terminate_callback, NULL);
820 * There had better not be any pages left
822 KKASSERT(object->resident_page_count == 0);
825 * Remove the object from the global object list.
827 lwkt_gettoken(&vmobj_token);
828 TAILQ_REMOVE(&vm_object_list, object, object_list);
830 lwkt_reltoken(&vmobj_token);
831 vm_object_dead_wakeup(object);
833 if (object->ref_count != 0) {
834 panic("vm_object_terminate2: object with references, "
835 "ref_count=%d", object->ref_count);
839 * NOTE: The object hold_count is at least 1, so we cannot zfree()
840 * the object here. See vm_object_drop().
845 * The caller must hold the object.
848 vm_object_terminate_callback(vm_page_t p, void *data __unused)
853 vm_page_busy_wait(p, FALSE, "vmpgtrm");
854 if (object != p->object) {
855 kprintf("vm_object_terminate: Warning: Encountered "
856 "busied page %p on queue %d\n", p, p->queue);
858 } else if (p->wire_count == 0) {
860 mycpu->gd_cnt.v_pfree++;
862 if (p->queue != PQ_NONE)
863 kprintf("vm_object_terminate: Warning: Encountered "
864 "wired page %p on queue %d\n", p, p->queue);
873 * The object is dead but still has an object<->pager association. Sleep
874 * and return. The caller typically retests the association in a loop.
876 * The caller must hold the object.
879 vm_object_dead_sleep(vm_object_t object, const char *wmesg)
881 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
882 if (object->handle) {
883 vm_object_set_flag(object, OBJ_DEADWNT);
884 tsleep(object, 0, wmesg, 0);
885 /* object may be invalid after this point */
890 * Wakeup anyone waiting for the object<->pager disassociation on
893 * The caller must hold the object.
896 vm_object_dead_wakeup(vm_object_t object)
898 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
899 if (object->flags & OBJ_DEADWNT) {
900 vm_object_clear_flag(object, OBJ_DEADWNT);
906 * Clean all dirty pages in the specified range of object. Leaves page
907 * on whatever queue it is currently on. If NOSYNC is set then do not
908 * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
909 * leaving the object dirty.
911 * When stuffing pages asynchronously, allow clustering. XXX we need a
912 * synchronous clustering mode implementation.
914 * Odd semantics: if start == end, we clean everything.
916 * The object must be locked? XXX
918 static int vm_object_page_clean_pass1(struct vm_page *p, void *data);
919 static int vm_object_page_clean_pass2(struct vm_page *p, void *data);
922 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
925 struct rb_vm_page_scan_info info;
931 vm_object_hold(object);
932 if (object->type != OBJT_VNODE ||
933 (object->flags & OBJ_MIGHTBEDIRTY) == 0) {
934 vm_object_drop(object);
938 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ?
939 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
940 pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
945 * Interlock other major object operations. This allows us to
946 * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY.
948 vm_object_set_flag(object, OBJ_CLEANING);
951 * Handle 'entire object' case
953 info.start_pindex = start;
955 info.end_pindex = object->size - 1;
957 info.end_pindex = end - 1;
959 wholescan = (start == 0 && info.end_pindex == object->size - 1);
961 info.pagerflags = pagerflags;
962 info.object = object;
965 * If cleaning the entire object do a pass to mark the pages read-only.
966 * If everything worked out ok, clear OBJ_WRITEABLE and
971 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
972 vm_object_page_clean_pass1, &info);
973 if (info.error == 0) {
974 vm_object_clear_flag(object,
975 OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
976 if (object->type == OBJT_VNODE &&
977 (vp = (struct vnode *)object->handle) != NULL) {
978 if (vp->v_flag & VOBJDIRTY)
979 vclrflags(vp, VOBJDIRTY);
985 * Do a pass to clean all the dirty pages we find.
989 generation = object->generation;
990 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
991 vm_object_page_clean_pass2, &info);
992 } while (info.error || generation != object->generation);
994 vm_object_clear_flag(object, OBJ_CLEANING);
995 vm_object_drop(object);
999 * The caller must hold the object.
1003 vm_object_page_clean_pass1(struct vm_page *p, void *data)
1005 struct rb_vm_page_scan_info *info = data;
1007 vm_page_flag_set(p, PG_CLEANCHK);
1008 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1010 } else if (vm_page_busy_try(p, FALSE) == 0) {
1011 vm_page_protect(p, VM_PROT_READ); /* must not block */
1020 * The caller must hold the object
1024 vm_object_page_clean_pass2(struct vm_page *p, void *data)
1026 struct rb_vm_page_scan_info *info = data;
1030 * Do not mess with pages that were inserted after we started
1031 * the cleaning pass.
1033 if ((p->flags & PG_CLEANCHK) == 0)
1036 generation = info->object->generation;
1037 vm_page_busy_wait(p, TRUE, "vpcwai");
1038 if (p->object != info->object ||
1039 info->object->generation != generation) {
1046 * Before wasting time traversing the pmaps, check for trivial
1047 * cases where the page cannot be dirty.
1049 if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) {
1050 KKASSERT((p->dirty & p->valid) == 0);
1056 * Check whether the page is dirty or not. The page has been set
1057 * to be read-only so the check will not race a user dirtying the
1060 vm_page_test_dirty(p);
1061 if ((p->dirty & p->valid) == 0) {
1062 vm_page_flag_clear(p, PG_CLEANCHK);
1068 * If we have been asked to skip nosync pages and this is a
1069 * nosync page, skip it. Note that the object flags were
1070 * not cleared in this case (because pass1 will have returned an
1071 * error), so we do not have to set them.
1073 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1074 vm_page_flag_clear(p, PG_CLEANCHK);
1080 * Flush as many pages as we can. PG_CLEANCHK will be cleared on
1081 * the pages that get successfully flushed. Set info->error if
1082 * we raced an object modification.
1084 vm_object_page_collect_flush(info->object, p, info->pagerflags);
1089 * Collect the specified page and nearby pages and flush them out.
1090 * The number of pages flushed is returned. The passed page is busied
1091 * by the caller and we are responsible for its disposition.
1093 * The caller must hold the object.
1096 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags)
1105 vm_page_t maf[vm_pageout_page_count];
1106 vm_page_t mab[vm_pageout_page_count];
1107 vm_page_t ma[vm_pageout_page_count];
1109 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1114 for(i = 1; i < vm_pageout_page_count; i++) {
1117 tp = vm_page_lookup_busy_try(object, pi + i, TRUE, &error);
1122 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1123 (tp->flags & PG_CLEANCHK) == 0) {
1127 if ((tp->queue - tp->pc) == PQ_CACHE) {
1128 vm_page_flag_clear(tp, PG_CLEANCHK);
1132 vm_page_test_dirty(tp);
1133 if ((tp->dirty & tp->valid) == 0) {
1134 vm_page_flag_clear(tp, PG_CLEANCHK);
1143 chkb = vm_pageout_page_count - maxf;
1145 * NOTE: chkb can be 0
1147 for(i = 1; chkb && i < chkb; i++) {
1150 tp = vm_page_lookup_busy_try(object, pi - i, TRUE, &error);
1155 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1156 (tp->flags & PG_CLEANCHK) == 0) {
1160 if ((tp->queue - tp->pc) == PQ_CACHE) {
1161 vm_page_flag_clear(tp, PG_CLEANCHK);
1165 vm_page_test_dirty(tp);
1166 if ((tp->dirty & tp->valid) == 0) {
1167 vm_page_flag_clear(tp, PG_CLEANCHK);
1176 * All pages in the maf[] and mab[] array are busied.
1178 for (i = 0; i < maxb; i++) {
1179 int index = (maxb - i) - 1;
1181 vm_page_flag_clear(ma[index], PG_CLEANCHK);
1183 vm_page_flag_clear(p, PG_CLEANCHK);
1185 for(i = 0; i < maxf; i++) {
1186 int index = (maxb + i) + 1;
1188 vm_page_flag_clear(ma[index], PG_CLEANCHK);
1190 runlen = maxb + maxf + 1;
1192 for (i = 0; i < runlen; i++)
1193 vm_page_hold(ma[i]);
1195 vm_pageout_flush(ma, runlen, pagerflags);
1197 for (i = 0; i < runlen; i++) {
1198 if (ma[i]->valid & ma[i]->dirty) {
1199 vm_page_protect(ma[i], VM_PROT_READ);
1200 vm_page_flag_set(ma[i], PG_CLEANCHK);
1203 * maxf will end up being the actual number of pages
1204 * we wrote out contiguously, non-inclusive of the
1205 * first page. We do not count look-behind pages.
1207 if (i >= maxb + 1 && (maxf > i - maxb - 1))
1208 maxf = i - maxb - 1;
1210 vm_page_unhold(ma[i]);
1216 * Same as vm_object_pmap_copy, except range checking really
1217 * works, and is meant for small sections of an object.
1219 * This code protects resident pages by making them read-only
1220 * and is typically called on a fork or split when a page
1221 * is converted to copy-on-write.
1223 * NOTE: If the page is already at VM_PROT_NONE, calling
1224 * vm_page_protect will have no effect.
1227 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1232 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
1235 vm_object_hold(object);
1236 for (idx = start; idx < end; idx++) {
1237 p = vm_page_lookup(object, idx);
1240 vm_page_protect(p, VM_PROT_READ);
1242 vm_object_drop(object);
1246 * Removes all physical pages in the specified object range from all
1249 * The object must *not* be locked.
1252 static int vm_object_pmap_remove_callback(vm_page_t p, void *data);
1255 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1257 struct rb_vm_page_scan_info info;
1261 info.start_pindex = start;
1262 info.end_pindex = end - 1;
1264 vm_object_hold(object);
1265 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1266 vm_object_pmap_remove_callback, &info);
1267 if (start == 0 && end == object->size)
1268 vm_object_clear_flag(object, OBJ_WRITEABLE);
1269 vm_object_drop(object);
1273 * The caller must hold the object
1276 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused)
1278 vm_page_protect(p, VM_PROT_NONE);
1283 * Implements the madvise function at the object/page level.
1285 * MADV_WILLNEED (any object)
1287 * Activate the specified pages if they are resident.
1289 * MADV_DONTNEED (any object)
1291 * Deactivate the specified pages if they are resident.
1293 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only)
1295 * Deactivate and clean the specified pages if they are
1296 * resident. This permits the process to reuse the pages
1297 * without faulting or the kernel to reclaim the pages
1303 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1305 vm_pindex_t end, tpindex;
1306 vm_object_t tobject;
1314 end = pindex + count;
1316 vm_object_hold(object);
1320 * Locate and adjust resident pages
1322 for (; pindex < end; pindex += 1) {
1324 if (tobject != object)
1325 vm_object_drop(tobject);
1330 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1331 * and those pages must be OBJ_ONEMAPPING.
1333 if (advise == MADV_FREE) {
1334 if ((tobject->type != OBJT_DEFAULT &&
1335 tobject->type != OBJT_SWAP) ||
1336 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1341 m = vm_page_lookup_busy_try(tobject, tpindex, TRUE, &error);
1344 vm_page_sleep_busy(m, TRUE, "madvpo");
1349 * There may be swap even if there is no backing page
1351 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1352 swap_pager_freespace(tobject, tpindex, 1);
1357 while ((xobj = tobject->backing_object) != NULL) {
1358 KKASSERT(xobj != object);
1359 vm_object_hold(xobj);
1360 if (xobj == tobject->backing_object)
1362 vm_object_drop(xobj);
1366 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1367 if (tobject != object) {
1368 vm_object_lock_swap();
1369 vm_object_drop(tobject);
1376 * If the page is not in a normal active state, we skip it.
1377 * If the page is not managed there are no page queues to
1378 * mess with. Things can break if we mess with pages in
1379 * any of the below states.
1382 /*m->hold_count ||*/
1384 (m->flags & PG_UNMANAGED) ||
1385 m->valid != VM_PAGE_BITS_ALL
1392 * Theoretically once a page is known not to be busy, an
1393 * interrupt cannot come along and rip it out from under us.
1396 if (advise == MADV_WILLNEED) {
1397 vm_page_activate(m);
1398 } else if (advise == MADV_DONTNEED) {
1399 vm_page_dontneed(m);
1400 } else if (advise == MADV_FREE) {
1402 * Mark the page clean. This will allow the page
1403 * to be freed up by the system. However, such pages
1404 * are often reused quickly by malloc()/free()
1405 * so we do not do anything that would cause
1406 * a page fault if we can help it.
1408 * Specifically, we do not try to actually free
1409 * the page now nor do we try to put it in the
1410 * cache (which would cause a page fault on reuse).
1412 * But we do make the page is freeable as we
1413 * can without actually taking the step of unmapping
1416 pmap_clear_modify(m);
1419 vm_page_dontneed(m);
1420 if (tobject->type == OBJT_SWAP)
1421 swap_pager_freespace(tobject, tpindex, 1);
1425 if (tobject != object)
1426 vm_object_drop(tobject);
1427 vm_object_drop(object);
1431 * Create a new object which is backed by the specified existing object
1432 * range. Replace the pointer and offset that was pointing at the existing
1433 * object with the pointer/offset for the new object.
1435 * No other requirements.
1438 vm_object_shadow(vm_object_t *objectp, vm_ooffset_t *offset, vm_size_t length,
1447 * Don't create the new object if the old object isn't shared.
1448 * We have to chain wait before adding the reference to avoid
1449 * racing a collapse or deallocation.
1451 * Add the additional ref to source here to avoid racing a later
1452 * collapse or deallocation. Clear the ONEMAPPING flag whether
1453 * addref is TRUE or not in this case because the original object
1457 vm_object_hold(source);
1458 vm_object_chain_wait(source);
1459 if (source->ref_count == 1 &&
1460 source->handle == NULL &&
1461 (source->type == OBJT_DEFAULT ||
1462 source->type == OBJT_SWAP)) {
1463 vm_object_drop(source);
1465 vm_object_reference_locked(source);
1466 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1470 vm_object_reference_locked(source);
1471 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1475 * Allocate a new object with the given length. The new object
1476 * is returned referenced but we may have to add another one.
1477 * If we are adding a second reference we must clear OBJ_ONEMAPPING.
1478 * (typically because the caller is about to clone a vm_map_entry).
1480 * The source object currently has an extra reference to prevent
1481 * collapses into it while we mess with its shadow list, which
1482 * we will remove later in this routine.
1484 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1485 panic("vm_object_shadow: no object for shadowing");
1486 vm_object_hold(result);
1488 vm_object_reference_locked(result);
1489 vm_object_clear_flag(result, OBJ_ONEMAPPING);
1493 * The new object shadows the source object. Chain wait before
1494 * adjusting shadow_count or the shadow list to avoid races.
1496 * Try to optimize the result object's page color when shadowing
1497 * in order to maintain page coloring consistency in the combined
1500 KKASSERT(result->backing_object == NULL);
1501 result->backing_object = source;
1503 vm_object_chain_wait(source);
1504 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1505 source->shadow_count++;
1506 source->generation++;
1508 /* cpu localization twist */
1509 result->pg_color = (int)(intptr_t)curthread;
1511 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1517 * Adjust the return storage. Drop the ref on source before
1520 result->backing_object_offset = *offset;
1521 vm_object_drop(result);
1524 vm_object_deallocate_locked(source);
1525 vm_object_drop(source);
1529 * Return the new things
1534 #define OBSC_TEST_ALL_SHADOWED 0x0001
1535 #define OBSC_COLLAPSE_NOWAIT 0x0002
1536 #define OBSC_COLLAPSE_WAIT 0x0004
1538 static int vm_object_backing_scan_callback(vm_page_t p, void *data);
1541 * The caller must hold the object.
1544 vm_object_backing_scan(vm_object_t object, vm_object_t backing_object, int op)
1546 struct rb_vm_page_scan_info info;
1548 vm_object_assert_held(object);
1549 vm_object_assert_held(backing_object);
1551 KKASSERT(backing_object == object->backing_object);
1552 info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1555 * Initial conditions
1557 if (op & OBSC_TEST_ALL_SHADOWED) {
1559 * We do not want to have to test for the existence of
1560 * swap pages in the backing object. XXX but with the
1561 * new swapper this would be pretty easy to do.
1563 * XXX what about anonymous MAP_SHARED memory that hasn't
1564 * been ZFOD faulted yet? If we do not test for this, the
1565 * shadow test may succeed! XXX
1567 if (backing_object->type != OBJT_DEFAULT)
1570 if (op & OBSC_COLLAPSE_WAIT) {
1571 KKASSERT((backing_object->flags & OBJ_DEAD) == 0);
1572 vm_object_set_flag(backing_object, OBJ_DEAD);
1573 lwkt_gettoken(&vmobj_token);
1574 TAILQ_REMOVE(&vm_object_list, backing_object, object_list);
1576 lwkt_reltoken(&vmobj_token);
1577 vm_object_dead_wakeup(backing_object);
1581 * Our scan. We have to retry if a negative error code is returned,
1582 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that
1583 * the scan had to be stopped because the parent does not completely
1586 info.object = object;
1587 info.backing_object = backing_object;
1591 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL,
1592 vm_object_backing_scan_callback,
1594 } while (info.error < 0);
1600 * The caller must hold the object.
1603 vm_object_backing_scan_callback(vm_page_t p, void *data)
1605 struct rb_vm_page_scan_info *info = data;
1606 vm_object_t backing_object;
1608 vm_pindex_t new_pindex;
1609 vm_pindex_t backing_offset_index;
1612 new_pindex = p->pindex - info->backing_offset_index;
1614 object = info->object;
1615 backing_object = info->backing_object;
1616 backing_offset_index = info->backing_offset_index;
1618 if (op & OBSC_TEST_ALL_SHADOWED) {
1622 * Ignore pages outside the parent object's range
1623 * and outside the parent object's mapping of the
1626 * note that we do not busy the backing object's
1630 p->pindex < backing_offset_index ||
1631 new_pindex >= object->size
1637 * See if the parent has the page or if the parent's
1638 * object pager has the page. If the parent has the
1639 * page but the page is not valid, the parent's
1640 * object pager must have the page.
1642 * If this fails, the parent does not completely shadow
1643 * the object and we might as well give up now.
1646 pp = vm_page_lookup(object, new_pindex);
1647 if ((pp == NULL || pp->valid == 0) &&
1648 !vm_pager_has_page(object, new_pindex)
1650 info->error = 0; /* problemo */
1651 return(-1); /* stop the scan */
1656 * Check for busy page
1658 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1661 if (vm_page_busy_try(p, TRUE)) {
1662 if (op & OBSC_COLLAPSE_NOWAIT) {
1666 * If we slept, anything could have
1667 * happened. Ask that the scan be restarted.
1669 * Since the object is marked dead, the
1670 * backing offset should not have changed.
1672 vm_page_sleep_busy(p, TRUE, "vmocol");
1677 if (op & OBSC_COLLAPSE_NOWAIT) {
1678 if (p->valid == 0 /*|| p->hold_count*/ ||
1684 /* XXX what if p->valid == 0 , hold_count, etc? */
1688 p->object == backing_object,
1689 ("vm_object_qcollapse(): object mismatch")
1693 * Destroy any associated swap
1695 if (backing_object->type == OBJT_SWAP)
1696 swap_pager_freespace(backing_object, p->pindex, 1);
1699 p->pindex < backing_offset_index ||
1700 new_pindex >= object->size
1703 * Page is out of the parent object's range, we
1704 * can simply destroy it.
1706 vm_page_protect(p, VM_PROT_NONE);
1711 pp = vm_page_lookup(object, new_pindex);
1712 if (pp != NULL || vm_pager_has_page(object, new_pindex)) {
1714 * page already exists in parent OR swap exists
1715 * for this location in the parent. Destroy
1716 * the original page from the backing object.
1718 * Leave the parent's page alone
1720 vm_page_protect(p, VM_PROT_NONE);
1726 * Page does not exist in parent, rename the
1727 * page from the backing object to the main object.
1729 * If the page was mapped to a process, it can remain
1730 * mapped through the rename.
1732 if ((p->queue - p->pc) == PQ_CACHE)
1733 vm_page_deactivate(p);
1735 vm_page_rename(p, object, new_pindex);
1737 /* page automatically made dirty by rename */
1743 * This version of collapse allows the operation to occur earlier and
1744 * when paging_in_progress is true for an object... This is not a complete
1745 * operation, but should plug 99.9% of the rest of the leaks.
1747 * The caller must hold the object and backing_object and both must be
1750 * (only called from vm_object_collapse)
1753 vm_object_qcollapse(vm_object_t object, vm_object_t backing_object)
1755 if (backing_object->ref_count == 1) {
1756 backing_object->ref_count += 2;
1757 vm_object_backing_scan(object, backing_object,
1758 OBSC_COLLAPSE_NOWAIT);
1759 backing_object->ref_count -= 2;
1764 * Collapse an object with the object backing it. Pages in the backing
1765 * object are moved into the parent, and the backing object is deallocated.
1766 * Any conflict is resolved in favor of the parent's existing pages.
1768 * object must be held and chain-locked on call.
1770 * The caller must have an extra ref on object to prevent a race from
1771 * destroying it during the collapse.
1774 vm_object_collapse(vm_object_t object, struct vm_object_dealloc_list **dlistp)
1776 struct vm_object_dealloc_list *dlist = NULL;
1777 vm_object_t backing_object;
1780 * Only one thread is attempting a collapse at any given moment.
1781 * There are few restrictions for (object) that callers of this
1782 * function check so reentrancy is likely.
1784 KKASSERT(object != NULL);
1785 vm_object_assert_held(object);
1786 KKASSERT(object->flags & OBJ_CHAINLOCK);
1793 * We have to hold the backing object, check races.
1795 while ((backing_object = object->backing_object) != NULL) {
1796 vm_object_hold(backing_object);
1797 if (backing_object == object->backing_object)
1799 vm_object_drop(backing_object);
1803 * No backing object? Nothing to collapse then.
1805 if (backing_object == NULL)
1809 * You can't collapse with a non-default/non-swap object.
1811 if (backing_object->type != OBJT_DEFAULT &&
1812 backing_object->type != OBJT_SWAP) {
1813 vm_object_drop(backing_object);
1814 backing_object = NULL;
1819 * Chain-lock the backing object too because if we
1820 * successfully merge its pages into the top object we
1821 * will collapse backing_object->backing_object as the
1822 * new backing_object. Re-check that it is still our
1825 vm_object_chain_acquire(backing_object);
1826 if (backing_object != object->backing_object) {
1827 vm_object_chain_release(backing_object);
1828 vm_object_drop(backing_object);
1833 * we check the backing object first, because it is most likely
1836 if (backing_object->handle != NULL ||
1837 (backing_object->type != OBJT_DEFAULT &&
1838 backing_object->type != OBJT_SWAP) ||
1839 (backing_object->flags & OBJ_DEAD) ||
1840 object->handle != NULL ||
1841 (object->type != OBJT_DEFAULT &&
1842 object->type != OBJT_SWAP) ||
1843 (object->flags & OBJ_DEAD)) {
1848 * If paging is in progress we can't do a normal collapse.
1851 object->paging_in_progress != 0 ||
1852 backing_object->paging_in_progress != 0
1854 vm_object_qcollapse(object, backing_object);
1859 * We know that we can either collapse the backing object (if
1860 * the parent is the only reference to it) or (perhaps) have
1861 * the parent bypass the object if the parent happens to shadow
1862 * all the resident pages in the entire backing object.
1864 * This is ignoring pager-backed pages such as swap pages.
1865 * vm_object_backing_scan fails the shadowing test in this
1868 if (backing_object->ref_count == 1) {
1870 * If there is exactly one reference to the backing
1871 * object, we can collapse it into the parent.
1873 KKASSERT(object->backing_object == backing_object);
1874 vm_object_backing_scan(object, backing_object,
1875 OBSC_COLLAPSE_WAIT);
1878 * Move the pager from backing_object to object.
1880 if (backing_object->type == OBJT_SWAP) {
1881 vm_object_pip_add(backing_object, 1);
1884 * scrap the paging_offset junk and do a
1885 * discrete copy. This also removes major
1886 * assumptions about how the swap-pager
1887 * works from where it doesn't belong. The
1888 * new swapper is able to optimize the
1889 * destroy-source case.
1891 vm_object_pip_add(object, 1);
1892 swap_pager_copy(backing_object, object,
1893 OFF_TO_IDX(object->backing_object_offset),
1895 vm_object_pip_wakeup(object);
1896 vm_object_pip_wakeup(backing_object);
1900 * Object now shadows whatever backing_object did.
1901 * Remove object from backing_object's shadow_list.
1903 LIST_REMOVE(object, shadow_list);
1904 KKASSERT(object->backing_object == backing_object);
1905 backing_object->shadow_count--;
1906 backing_object->generation++;
1909 * backing_object->backing_object moves from within
1910 * backing_object to within object.
1912 while ((bbobj = backing_object->backing_object) != NULL) {
1913 vm_object_hold(bbobj);
1914 if (bbobj == backing_object->backing_object)
1916 vm_object_drop(bbobj);
1919 LIST_REMOVE(backing_object, shadow_list);
1920 bbobj->shadow_count--;
1921 bbobj->generation++;
1922 backing_object->backing_object = NULL;
1924 object->backing_object = bbobj;
1926 LIST_INSERT_HEAD(&bbobj->shadow_head,
1927 object, shadow_list);
1928 bbobj->shadow_count++;
1929 bbobj->generation++;
1932 object->backing_object_offset +=
1933 backing_object->backing_object_offset;
1935 vm_object_drop(bbobj);
1938 * Discard the old backing_object. Nothing should be
1939 * able to ref it, other than a vm_map_split(),
1940 * and vm_map_split() will stall on our chain lock.
1941 * And we control the parent so it shouldn't be
1942 * possible for it to go away either.
1944 * Since the backing object has no pages, no pager
1945 * left, and no object references within it, all
1946 * that is necessary is to dispose of it.
1948 KASSERT(backing_object->ref_count == 1,
1949 ("backing_object %p was somehow "
1950 "re-referenced during collapse!",
1952 KASSERT(RB_EMPTY(&backing_object->rb_memq),
1953 ("backing_object %p somehow has left "
1954 "over pages during collapse!",
1958 * The object can be destroyed.
1960 * XXX just fall through and dodealloc instead
1961 * of forcing destruction?
1963 --backing_object->ref_count;
1964 if ((backing_object->flags & OBJ_DEAD) == 0)
1965 vm_object_terminate(backing_object);
1970 * If we do not entirely shadow the backing object,
1971 * there is nothing we can do so we give up.
1973 if (vm_object_backing_scan(object, backing_object,
1974 OBSC_TEST_ALL_SHADOWED) == 0) {
1979 * bbobj is backing_object->backing_object. Since
1980 * object completely shadows backing_object we can
1981 * bypass it and become backed by bbobj instead.
1983 while ((bbobj = backing_object->backing_object) != NULL) {
1984 vm_object_hold(bbobj);
1985 if (bbobj == backing_object->backing_object)
1987 vm_object_drop(bbobj);
1991 * Make object shadow bbobj instead of backing_object.
1992 * Remove object from backing_object's shadow list.
1994 * Deallocating backing_object will not remove
1995 * it, since its reference count is at least 2.
1997 KKASSERT(object->backing_object == backing_object);
1998 LIST_REMOVE(object, shadow_list);
1999 backing_object->shadow_count--;
2000 backing_object->generation++;
2003 * Add a ref to bbobj, bbobj now shadows object.
2005 * NOTE: backing_object->backing_object still points
2006 * to bbobj. That relationship remains intact
2007 * because backing_object has > 1 ref, so
2008 * someone else is pointing to it (hence why
2009 * we can't collapse it into object and can
2010 * only handle the all-shadowed bypass case).
2013 vm_object_chain_wait(bbobj);
2014 vm_object_reference_locked(bbobj);
2015 LIST_INSERT_HEAD(&bbobj->shadow_head,
2016 object, shadow_list);
2017 bbobj->shadow_count++;
2018 bbobj->generation++;
2019 object->backing_object_offset +=
2020 backing_object->backing_object_offset;
2021 object->backing_object = bbobj;
2022 vm_object_drop(bbobj);
2024 object->backing_object = NULL;
2028 * Drop the reference count on backing_object. To
2029 * handle ref_count races properly we can't assume
2030 * that the ref_count is still at least 2 so we
2031 * have to actually call vm_object_deallocate()
2032 * (after clearing the chainlock).
2039 * Ok, we want to loop on the new object->bbobj association,
2040 * possibly collapsing it further. However if dodealloc is
2041 * non-zero we have to deallocate the backing_object which
2042 * itself can potentially undergo a collapse, creating a
2043 * recursion depth issue with the LWKT token subsystem.
2045 * In the case where we must deallocate the backing_object
2046 * it is possible now that the backing_object has a single
2047 * shadow count on some other object (not represented here
2048 * as yet), since it no longer shadows us. Thus when we
2049 * call vm_object_deallocate() it may attempt to collapse
2050 * itself into its remaining parent.
2053 struct vm_object_dealloc_list *dtmp;
2055 vm_object_chain_release(backing_object);
2056 vm_object_unlock(backing_object);
2057 /* backing_object remains held */
2060 * Auto-deallocation list for caller convenience.
2065 dtmp = kmalloc(sizeof(*dtmp), M_TEMP, M_WAITOK);
2066 dtmp->object = backing_object;
2067 dtmp->next = *dlistp;
2070 vm_object_chain_release(backing_object);
2071 vm_object_drop(backing_object);
2073 /* backing_object = NULL; not needed */
2078 * Clean up any left over backing_object
2080 if (backing_object) {
2081 vm_object_chain_release(backing_object);
2082 vm_object_drop(backing_object);
2086 * Clean up any auto-deallocation list. This is a convenience
2087 * for top-level callers so they don't have to pass &dlist.
2088 * Do not clean up any caller-passed dlistp, the caller will
2092 vm_object_deallocate_list(&dlist);
2097 * vm_object_collapse() may collect additional objects in need of
2098 * deallocation. This routine deallocates these objects. The
2099 * deallocation itself can trigger additional collapses (which the
2100 * deallocate function takes care of). This procedure is used to
2101 * reduce procedural recursion since these vm_object shadow chains
2102 * can become quite long.
2105 vm_object_deallocate_list(struct vm_object_dealloc_list **dlistp)
2107 struct vm_object_dealloc_list *dlist;
2109 while ((dlist = *dlistp) != NULL) {
2110 *dlistp = dlist->next;
2111 vm_object_lock(dlist->object);
2112 vm_object_deallocate_locked(dlist->object);
2113 vm_object_drop(dlist->object);
2114 kfree(dlist, M_TEMP);
2119 * Removes all physical pages in the specified object range from the
2120 * object's list of pages.
2124 static int vm_object_page_remove_callback(vm_page_t p, void *data);
2127 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2128 boolean_t clean_only)
2130 struct rb_vm_page_scan_info info;
2134 * Degenerate cases and assertions
2136 vm_object_hold(object);
2137 if (object == NULL ||
2138 (object->resident_page_count == 0 && object->swblock_count == 0)) {
2139 vm_object_drop(object);
2142 KASSERT(object->type != OBJT_PHYS,
2143 ("attempt to remove pages from a physical object"));
2146 * Indicate that paging is occuring on the object
2148 vm_object_pip_add(object, 1);
2151 * Figure out the actual removal range and whether we are removing
2152 * the entire contents of the object or not. If removing the entire
2153 * contents, be sure to get all pages, even those that might be
2154 * beyond the end of the object.
2156 info.start_pindex = start;
2158 info.end_pindex = (vm_pindex_t)-1;
2160 info.end_pindex = end - 1;
2161 info.limit = clean_only;
2162 all = (start == 0 && info.end_pindex >= object->size - 1);
2165 * Loop until we are sure we have gotten them all.
2169 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2170 vm_object_page_remove_callback, &info);
2171 } while (info.error);
2174 * Remove any related swap if throwing away pages, or for
2175 * non-swap objects (the swap is a clean copy in that case).
2177 if (object->type != OBJT_SWAP || clean_only == FALSE) {
2179 swap_pager_freespace_all(object);
2181 swap_pager_freespace(object, info.start_pindex,
2182 info.end_pindex - info.start_pindex + 1);
2188 vm_object_pip_wakeup(object);
2189 vm_object_drop(object);
2193 * The caller must hold the object
2196 vm_object_page_remove_callback(vm_page_t p, void *data)
2198 struct rb_vm_page_scan_info *info = data;
2200 if (vm_page_busy_try(p, TRUE)) {
2201 vm_page_sleep_busy(p, TRUE, "vmopar");
2207 * Wired pages cannot be destroyed, but they can be invalidated
2208 * and we do so if clean_only (limit) is not set.
2210 * WARNING! The page may be wired due to being part of a buffer
2211 * cache buffer, and the buffer might be marked B_CACHE.
2212 * This is fine as part of a truncation but VFSs must be
2213 * sure to fix the buffer up when re-extending the file.
2215 if (p->wire_count != 0) {
2216 vm_page_protect(p, VM_PROT_NONE);
2217 if (info->limit == 0)
2224 * limit is our clean_only flag. If set and the page is dirty, do
2225 * not free it. If set and the page is being held by someone, do
2228 if (info->limit && p->valid) {
2229 vm_page_test_dirty(p);
2230 if (p->valid & p->dirty) {
2235 if (p->hold_count) {
2245 vm_page_protect(p, VM_PROT_NONE);
2251 * Coalesces two objects backing up adjoining regions of memory into a
2254 * returns TRUE if objects were combined.
2256 * NOTE: Only works at the moment if the second object is NULL -
2257 * if it's not, which object do we lock first?
2260 * prev_object First object to coalesce
2261 * prev_offset Offset into prev_object
2262 * next_object Second object into coalesce
2263 * next_offset Offset into next_object
2265 * prev_size Size of reference to prev_object
2266 * next_size Size of reference to next_object
2268 * The caller does not need to hold (prev_object) but must have a stable
2269 * pointer to it (typically by holding the vm_map locked).
2272 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
2273 vm_size_t prev_size, vm_size_t next_size)
2275 vm_pindex_t next_pindex;
2277 if (prev_object == NULL)
2280 vm_object_hold(prev_object);
2282 if (prev_object->type != OBJT_DEFAULT &&
2283 prev_object->type != OBJT_SWAP) {
2284 vm_object_drop(prev_object);
2289 * Try to collapse the object first
2291 vm_object_chain_acquire(prev_object);
2292 vm_object_collapse(prev_object, NULL);
2295 * Can't coalesce if: . more than one reference . paged out . shadows
2296 * another object . has a copy elsewhere (any of which mean that the
2297 * pages not mapped to prev_entry may be in use anyway)
2300 if (prev_object->backing_object != NULL) {
2301 vm_object_chain_release(prev_object);
2302 vm_object_drop(prev_object);
2306 prev_size >>= PAGE_SHIFT;
2307 next_size >>= PAGE_SHIFT;
2308 next_pindex = prev_pindex + prev_size;
2310 if ((prev_object->ref_count > 1) &&
2311 (prev_object->size != next_pindex)) {
2312 vm_object_chain_release(prev_object);
2313 vm_object_drop(prev_object);
2318 * Remove any pages that may still be in the object from a previous
2321 if (next_pindex < prev_object->size) {
2322 vm_object_page_remove(prev_object,
2324 next_pindex + next_size, FALSE);
2325 if (prev_object->type == OBJT_SWAP)
2326 swap_pager_freespace(prev_object,
2327 next_pindex, next_size);
2331 * Extend the object if necessary.
2333 if (next_pindex + next_size > prev_object->size)
2334 prev_object->size = next_pindex + next_size;
2336 vm_object_chain_release(prev_object);
2337 vm_object_drop(prev_object);
2342 * Make the object writable and flag is being possibly dirty.
2344 * The caller must hold the object. XXX called from vm_page_dirty(),
2345 * There is currently no requirement to hold the object.
2348 vm_object_set_writeable_dirty(vm_object_t object)
2352 /*vm_object_assert_held(object);*/
2353 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
2354 if (object->type == OBJT_VNODE &&
2355 (vp = (struct vnode *)object->handle) != NULL) {
2356 if ((vp->v_flag & VOBJDIRTY) == 0) {
2357 vsetflags(vp, VOBJDIRTY);
2362 #include "opt_ddb.h"
2364 #include <sys/kernel.h>
2366 #include <sys/cons.h>
2368 #include <ddb/ddb.h>
2370 static int _vm_object_in_map (vm_map_t map, vm_object_t object,
2371 vm_map_entry_t entry);
2372 static int vm_object_in_map (vm_object_t object);
2375 * The caller must hold the object.
2378 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2381 vm_map_entry_t tmpe;
2382 vm_object_t obj, nobj;
2388 tmpe = map->header.next;
2389 entcount = map->nentries;
2390 while (entcount-- && (tmpe != &map->header)) {
2391 if( _vm_object_in_map(map, object, tmpe)) {
2398 switch(entry->maptype) {
2399 case VM_MAPTYPE_SUBMAP:
2400 tmpm = entry->object.sub_map;
2401 tmpe = tmpm->header.next;
2402 entcount = tmpm->nentries;
2403 while (entcount-- && tmpe != &tmpm->header) {
2404 if( _vm_object_in_map(tmpm, object, tmpe)) {
2410 case VM_MAPTYPE_NORMAL:
2411 case VM_MAPTYPE_VPAGETABLE:
2412 obj = entry->object.vm_object;
2414 if (obj == object) {
2415 if (obj != entry->object.vm_object)
2416 vm_object_drop(obj);
2419 while ((nobj = obj->backing_object) != NULL) {
2420 vm_object_hold(nobj);
2421 if (nobj == obj->backing_object)
2423 vm_object_drop(nobj);
2425 if (obj != entry->object.vm_object) {
2427 vm_object_lock_swap();
2428 vm_object_drop(obj);
2439 static int vm_object_in_map_callback(struct proc *p, void *data);
2441 struct vm_object_in_map_info {
2450 vm_object_in_map(vm_object_t object)
2452 struct vm_object_in_map_info info;
2455 info.object = object;
2457 allproc_scan(vm_object_in_map_callback, &info);
2460 if( _vm_object_in_map(&kernel_map, object, 0))
2462 if( _vm_object_in_map(&pager_map, object, 0))
2464 if( _vm_object_in_map(&buffer_map, object, 0))
2473 vm_object_in_map_callback(struct proc *p, void *data)
2475 struct vm_object_in_map_info *info = data;
2478 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) {
2486 DB_SHOW_COMMAND(vmochk, vm_object_check)
2491 * make sure that internal objs are in a map somewhere
2492 * and none have zero ref counts.
2494 for (object = TAILQ_FIRST(&vm_object_list);
2496 object = TAILQ_NEXT(object, object_list)) {
2497 if (object->type == OBJT_MARKER)
2499 if (object->handle == NULL &&
2500 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2501 if (object->ref_count == 0) {
2502 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2503 (long)object->size);
2505 if (!vm_object_in_map(object)) {
2507 "vmochk: internal obj is not in a map: "
2508 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2509 object->ref_count, (u_long)object->size,
2510 (u_long)object->size,
2511 (void *)object->backing_object);
2520 DB_SHOW_COMMAND(object, vm_object_print_static)
2522 /* XXX convert args. */
2523 vm_object_t object = (vm_object_t)addr;
2524 boolean_t full = have_addr;
2528 /* XXX count is an (unused) arg. Avoid shadowing it. */
2529 #define count was_count
2537 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
2538 object, (int)object->type, (u_long)object->size,
2539 object->resident_page_count, object->ref_count, object->flags);
2541 * XXX no %qd in kernel. Truncate object->backing_object_offset.
2543 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
2544 object->shadow_count,
2545 object->backing_object ? object->backing_object->ref_count : 0,
2546 object->backing_object, (long)object->backing_object_offset);
2553 RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) {
2555 db_iprintf("memory:=");
2556 else if (count == 6) {
2564 db_printf("(off=0x%lx,page=0x%lx)",
2565 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
2576 * XXX need this non-static entry for calling from vm_map_print.
2581 vm_object_print(/* db_expr_t */ long addr,
2582 boolean_t have_addr,
2583 /* db_expr_t */ long count,
2586 vm_object_print_static(addr, have_addr, count, modif);
2592 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2597 for (object = TAILQ_FIRST(&vm_object_list);
2599 object = TAILQ_NEXT(object, object_list)) {
2600 vm_pindex_t idx, fidx;
2602 vm_paddr_t pa = -1, padiff;
2606 if (object->type == OBJT_MARKER)
2608 db_printf("new object: %p\n", (void *)object);
2618 osize = object->size;
2621 for (idx = 0; idx < osize; idx++) {
2622 m = vm_page_lookup(object, idx);
2625 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2626 (long)fidx, rcount, (long)pa);
2641 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2646 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
2647 padiff >>= PAGE_SHIFT;
2648 padiff &= PQ_L2_MASK;
2650 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
2654 db_printf(" index(%ld)run(%d)pa(0x%lx)",
2655 (long)fidx, rcount, (long)pa);
2656 db_printf("pd(%ld)\n", (long)padiff);
2666 pa = VM_PAGE_TO_PHYS(m);
2670 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2671 (long)fidx, rcount, (long)pa);