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 */
1021 * The caller must hold the object
1025 vm_object_page_clean_pass2(struct vm_page *p, void *data)
1027 struct rb_vm_page_scan_info *info = data;
1031 * Do not mess with pages that were inserted after we started
1032 * the cleaning pass.
1034 if ((p->flags & PG_CLEANCHK) == 0)
1037 generation = info->object->generation;
1038 vm_page_busy_wait(p, TRUE, "vpcwai");
1039 if (p->object != info->object ||
1040 info->object->generation != generation) {
1047 * Before wasting time traversing the pmaps, check for trivial
1048 * cases where the page cannot be dirty.
1050 if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) {
1051 KKASSERT((p->dirty & p->valid) == 0);
1057 * Check whether the page is dirty or not. The page has been set
1058 * to be read-only so the check will not race a user dirtying the
1061 vm_page_test_dirty(p);
1062 if ((p->dirty & p->valid) == 0) {
1063 vm_page_flag_clear(p, PG_CLEANCHK);
1069 * If we have been asked to skip nosync pages and this is a
1070 * nosync page, skip it. Note that the object flags were
1071 * not cleared in this case (because pass1 will have returned an
1072 * error), so we do not have to set them.
1074 if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1075 vm_page_flag_clear(p, PG_CLEANCHK);
1081 * Flush as many pages as we can. PG_CLEANCHK will be cleared on
1082 * the pages that get successfully flushed. Set info->error if
1083 * we raced an object modification.
1085 vm_object_page_collect_flush(info->object, p, info->pagerflags);
1092 * Collect the specified page and nearby pages and flush them out.
1093 * The number of pages flushed is returned. The passed page is busied
1094 * by the caller and we are responsible for its disposition.
1096 * The caller must hold the object.
1099 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags)
1108 vm_page_t maf[vm_pageout_page_count];
1109 vm_page_t mab[vm_pageout_page_count];
1110 vm_page_t ma[vm_pageout_page_count];
1112 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1117 for(i = 1; i < vm_pageout_page_count; i++) {
1120 tp = vm_page_lookup_busy_try(object, pi + i, TRUE, &error);
1125 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1126 (tp->flags & PG_CLEANCHK) == 0) {
1130 if ((tp->queue - tp->pc) == PQ_CACHE) {
1131 vm_page_flag_clear(tp, PG_CLEANCHK);
1135 vm_page_test_dirty(tp);
1136 if ((tp->dirty & tp->valid) == 0) {
1137 vm_page_flag_clear(tp, PG_CLEANCHK);
1146 chkb = vm_pageout_page_count - maxf;
1148 * NOTE: chkb can be 0
1150 for(i = 1; chkb && i < chkb; i++) {
1153 tp = vm_page_lookup_busy_try(object, pi - i, TRUE, &error);
1158 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1159 (tp->flags & PG_CLEANCHK) == 0) {
1163 if ((tp->queue - tp->pc) == PQ_CACHE) {
1164 vm_page_flag_clear(tp, PG_CLEANCHK);
1168 vm_page_test_dirty(tp);
1169 if ((tp->dirty & tp->valid) == 0) {
1170 vm_page_flag_clear(tp, PG_CLEANCHK);
1179 * All pages in the maf[] and mab[] array are busied.
1181 for (i = 0; i < maxb; i++) {
1182 int index = (maxb - i) - 1;
1184 vm_page_flag_clear(ma[index], PG_CLEANCHK);
1186 vm_page_flag_clear(p, PG_CLEANCHK);
1188 for(i = 0; i < maxf; i++) {
1189 int index = (maxb + i) + 1;
1191 vm_page_flag_clear(ma[index], PG_CLEANCHK);
1193 runlen = maxb + maxf + 1;
1195 for (i = 0; i < runlen; i++)
1196 vm_page_hold(ma[i]);
1198 vm_pageout_flush(ma, runlen, pagerflags);
1200 for (i = 0; i < runlen; i++) {
1201 if (ma[i]->valid & ma[i]->dirty) {
1202 vm_page_protect(ma[i], VM_PROT_READ);
1203 vm_page_flag_set(ma[i], PG_CLEANCHK);
1206 * maxf will end up being the actual number of pages
1207 * we wrote out contiguously, non-inclusive of the
1208 * first page. We do not count look-behind pages.
1210 if (i >= maxb + 1 && (maxf > i - maxb - 1))
1211 maxf = i - maxb - 1;
1213 vm_page_unhold(ma[i]);
1219 * Same as vm_object_pmap_copy, except range checking really
1220 * works, and is meant for small sections of an object.
1222 * This code protects resident pages by making them read-only
1223 * and is typically called on a fork or split when a page
1224 * is converted to copy-on-write.
1226 * NOTE: If the page is already at VM_PROT_NONE, calling
1227 * vm_page_protect will have no effect.
1230 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1235 if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
1238 vm_object_hold(object);
1239 for (idx = start; idx < end; idx++) {
1240 p = vm_page_lookup(object, idx);
1243 vm_page_protect(p, VM_PROT_READ);
1245 vm_object_drop(object);
1249 * Removes all physical pages in the specified object range from all
1252 * The object must *not* be locked.
1255 static int vm_object_pmap_remove_callback(vm_page_t p, void *data);
1258 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1260 struct rb_vm_page_scan_info info;
1264 info.start_pindex = start;
1265 info.end_pindex = end - 1;
1267 vm_object_hold(object);
1268 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1269 vm_object_pmap_remove_callback, &info);
1270 if (start == 0 && end == object->size)
1271 vm_object_clear_flag(object, OBJ_WRITEABLE);
1272 vm_object_drop(object);
1276 * The caller must hold the object
1279 vm_object_pmap_remove_callback(vm_page_t p, void *data __unused)
1281 vm_page_protect(p, VM_PROT_NONE);
1286 * Implements the madvise function at the object/page level.
1288 * MADV_WILLNEED (any object)
1290 * Activate the specified pages if they are resident.
1292 * MADV_DONTNEED (any object)
1294 * Deactivate the specified pages if they are resident.
1296 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only)
1298 * Deactivate and clean the specified pages if they are
1299 * resident. This permits the process to reuse the pages
1300 * without faulting or the kernel to reclaim the pages
1306 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, int count, int advise)
1308 vm_pindex_t end, tpindex;
1309 vm_object_t tobject;
1317 end = pindex + count;
1319 vm_object_hold(object);
1323 * Locate and adjust resident pages
1325 for (; pindex < end; pindex += 1) {
1327 if (tobject != object)
1328 vm_object_drop(tobject);
1333 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1334 * and those pages must be OBJ_ONEMAPPING.
1336 if (advise == MADV_FREE) {
1337 if ((tobject->type != OBJT_DEFAULT &&
1338 tobject->type != OBJT_SWAP) ||
1339 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1344 m = vm_page_lookup_busy_try(tobject, tpindex, TRUE, &error);
1347 vm_page_sleep_busy(m, TRUE, "madvpo");
1352 * There may be swap even if there is no backing page
1354 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1355 swap_pager_freespace(tobject, tpindex, 1);
1360 while ((xobj = tobject->backing_object) != NULL) {
1361 KKASSERT(xobj != object);
1362 vm_object_hold(xobj);
1363 if (xobj == tobject->backing_object)
1365 vm_object_drop(xobj);
1369 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1370 if (tobject != object) {
1371 vm_object_lock_swap();
1372 vm_object_drop(tobject);
1379 * If the page is not in a normal active state, we skip it.
1380 * If the page is not managed there are no page queues to
1381 * mess with. Things can break if we mess with pages in
1382 * any of the below states.
1385 /*m->hold_count ||*/
1387 (m->flags & PG_UNMANAGED) ||
1388 m->valid != VM_PAGE_BITS_ALL
1395 * Theoretically once a page is known not to be busy, an
1396 * interrupt cannot come along and rip it out from under us.
1399 if (advise == MADV_WILLNEED) {
1400 vm_page_activate(m);
1401 } else if (advise == MADV_DONTNEED) {
1402 vm_page_dontneed(m);
1403 } else if (advise == MADV_FREE) {
1405 * Mark the page clean. This will allow the page
1406 * to be freed up by the system. However, such pages
1407 * are often reused quickly by malloc()/free()
1408 * so we do not do anything that would cause
1409 * a page fault if we can help it.
1411 * Specifically, we do not try to actually free
1412 * the page now nor do we try to put it in the
1413 * cache (which would cause a page fault on reuse).
1415 * But we do make the page is freeable as we
1416 * can without actually taking the step of unmapping
1419 pmap_clear_modify(m);
1422 vm_page_dontneed(m);
1423 if (tobject->type == OBJT_SWAP)
1424 swap_pager_freespace(tobject, tpindex, 1);
1428 if (tobject != object)
1429 vm_object_drop(tobject);
1430 vm_object_drop(object);
1434 * Create a new object which is backed by the specified existing object
1435 * range. Replace the pointer and offset that was pointing at the existing
1436 * object with the pointer/offset for the new object.
1438 * No other requirements.
1441 vm_object_shadow(vm_object_t *objectp, vm_ooffset_t *offset, vm_size_t length,
1450 * Don't create the new object if the old object isn't shared.
1451 * We have to chain wait before adding the reference to avoid
1452 * racing a collapse or deallocation.
1454 * Add the additional ref to source here to avoid racing a later
1455 * collapse or deallocation. Clear the ONEMAPPING flag whether
1456 * addref is TRUE or not in this case because the original object
1460 vm_object_hold(source);
1461 vm_object_chain_wait(source);
1462 if (source->ref_count == 1 &&
1463 source->handle == NULL &&
1464 (source->type == OBJT_DEFAULT ||
1465 source->type == OBJT_SWAP)) {
1466 vm_object_drop(source);
1468 vm_object_reference_locked(source);
1469 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1473 vm_object_reference_locked(source);
1474 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1478 * Allocate a new object with the given length. The new object
1479 * is returned referenced but we may have to add another one.
1480 * If we are adding a second reference we must clear OBJ_ONEMAPPING.
1481 * (typically because the caller is about to clone a vm_map_entry).
1483 * The source object currently has an extra reference to prevent
1484 * collapses into it while we mess with its shadow list, which
1485 * we will remove later in this routine.
1487 if ((result = vm_object_allocate(OBJT_DEFAULT, length)) == NULL)
1488 panic("vm_object_shadow: no object for shadowing");
1489 vm_object_hold(result);
1491 vm_object_reference_locked(result);
1492 vm_object_clear_flag(result, OBJ_ONEMAPPING);
1496 * The new object shadows the source object. Chain wait before
1497 * adjusting shadow_count or the shadow list to avoid races.
1499 * Try to optimize the result object's page color when shadowing
1500 * in order to maintain page coloring consistency in the combined
1503 KKASSERT(result->backing_object == NULL);
1504 result->backing_object = source;
1506 vm_object_chain_wait(source);
1507 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1508 source->shadow_count++;
1509 source->generation++;
1511 /* cpu localization twist */
1512 result->pg_color = (int)(intptr_t)curthread;
1514 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1520 * Adjust the return storage. Drop the ref on source before
1523 result->backing_object_offset = *offset;
1524 vm_object_drop(result);
1527 vm_object_deallocate_locked(source);
1528 vm_object_drop(source);
1532 * Return the new things
1537 #define OBSC_TEST_ALL_SHADOWED 0x0001
1538 #define OBSC_COLLAPSE_NOWAIT 0x0002
1539 #define OBSC_COLLAPSE_WAIT 0x0004
1541 static int vm_object_backing_scan_callback(vm_page_t p, void *data);
1544 * The caller must hold the object.
1547 vm_object_backing_scan(vm_object_t object, vm_object_t backing_object, int op)
1549 struct rb_vm_page_scan_info info;
1551 vm_object_assert_held(object);
1552 vm_object_assert_held(backing_object);
1554 KKASSERT(backing_object == object->backing_object);
1555 info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1558 * Initial conditions
1560 if (op & OBSC_TEST_ALL_SHADOWED) {
1562 * We do not want to have to test for the existence of
1563 * swap pages in the backing object. XXX but with the
1564 * new swapper this would be pretty easy to do.
1566 * XXX what about anonymous MAP_SHARED memory that hasn't
1567 * been ZFOD faulted yet? If we do not test for this, the
1568 * shadow test may succeed! XXX
1570 if (backing_object->type != OBJT_DEFAULT)
1573 if (op & OBSC_COLLAPSE_WAIT) {
1574 KKASSERT((backing_object->flags & OBJ_DEAD) == 0);
1575 vm_object_set_flag(backing_object, OBJ_DEAD);
1576 lwkt_gettoken(&vmobj_token);
1577 TAILQ_REMOVE(&vm_object_list, backing_object, object_list);
1579 lwkt_reltoken(&vmobj_token);
1580 vm_object_dead_wakeup(backing_object);
1584 * Our scan. We have to retry if a negative error code is returned,
1585 * otherwise 0 or 1 will be returned in info.error. 0 Indicates that
1586 * the scan had to be stopped because the parent does not completely
1589 info.object = object;
1590 info.backing_object = backing_object;
1594 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL,
1595 vm_object_backing_scan_callback,
1597 } while (info.error < 0);
1603 * The caller must hold the object.
1606 vm_object_backing_scan_callback(vm_page_t p, void *data)
1608 struct rb_vm_page_scan_info *info = data;
1609 vm_object_t backing_object;
1611 vm_pindex_t new_pindex;
1612 vm_pindex_t backing_offset_index;
1615 new_pindex = p->pindex - info->backing_offset_index;
1617 object = info->object;
1618 backing_object = info->backing_object;
1619 backing_offset_index = info->backing_offset_index;
1621 if (op & OBSC_TEST_ALL_SHADOWED) {
1625 * Ignore pages outside the parent object's range
1626 * and outside the parent object's mapping of the
1629 * note that we do not busy the backing object's
1633 p->pindex < backing_offset_index ||
1634 new_pindex >= object->size
1640 * See if the parent has the page or if the parent's
1641 * object pager has the page. If the parent has the
1642 * page but the page is not valid, the parent's
1643 * object pager must have the page.
1645 * If this fails, the parent does not completely shadow
1646 * the object and we might as well give up now.
1649 pp = vm_page_lookup(object, new_pindex);
1650 if ((pp == NULL || pp->valid == 0) &&
1651 !vm_pager_has_page(object, new_pindex)
1653 info->error = 0; /* problemo */
1654 return(-1); /* stop the scan */
1659 * Check for busy page
1661 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1664 if (vm_page_busy_try(p, TRUE)) {
1665 if (op & OBSC_COLLAPSE_NOWAIT) {
1669 * If we slept, anything could have
1670 * happened. Ask that the scan be restarted.
1672 * Since the object is marked dead, the
1673 * backing offset should not have changed.
1675 vm_page_sleep_busy(p, TRUE, "vmocol");
1680 if (op & OBSC_COLLAPSE_NOWAIT) {
1681 if (p->valid == 0 /*|| p->hold_count*/ ||
1687 /* XXX what if p->valid == 0 , hold_count, etc? */
1691 p->object == backing_object,
1692 ("vm_object_qcollapse(): object mismatch")
1696 * Destroy any associated swap
1698 if (backing_object->type == OBJT_SWAP)
1699 swap_pager_freespace(backing_object, p->pindex, 1);
1702 p->pindex < backing_offset_index ||
1703 new_pindex >= object->size
1706 * Page is out of the parent object's range, we
1707 * can simply destroy it.
1709 vm_page_protect(p, VM_PROT_NONE);
1714 pp = vm_page_lookup(object, new_pindex);
1715 if (pp != NULL || vm_pager_has_page(object, new_pindex)) {
1717 * page already exists in parent OR swap exists
1718 * for this location in the parent. Destroy
1719 * the original page from the backing object.
1721 * Leave the parent's page alone
1723 vm_page_protect(p, VM_PROT_NONE);
1729 * Page does not exist in parent, rename the
1730 * page from the backing object to the main object.
1732 * If the page was mapped to a process, it can remain
1733 * mapped through the rename.
1735 if ((p->queue - p->pc) == PQ_CACHE)
1736 vm_page_deactivate(p);
1738 vm_page_rename(p, object, new_pindex);
1740 /* page automatically made dirty by rename */
1746 * This version of collapse allows the operation to occur earlier and
1747 * when paging_in_progress is true for an object... This is not a complete
1748 * operation, but should plug 99.9% of the rest of the leaks.
1750 * The caller must hold the object and backing_object and both must be
1753 * (only called from vm_object_collapse)
1756 vm_object_qcollapse(vm_object_t object, vm_object_t backing_object)
1758 if (backing_object->ref_count == 1) {
1759 backing_object->ref_count += 2;
1760 vm_object_backing_scan(object, backing_object,
1761 OBSC_COLLAPSE_NOWAIT);
1762 backing_object->ref_count -= 2;
1767 * Collapse an object with the object backing it. Pages in the backing
1768 * object are moved into the parent, and the backing object is deallocated.
1769 * Any conflict is resolved in favor of the parent's existing pages.
1771 * object must be held and chain-locked on call.
1773 * The caller must have an extra ref on object to prevent a race from
1774 * destroying it during the collapse.
1777 vm_object_collapse(vm_object_t object, struct vm_object_dealloc_list **dlistp)
1779 struct vm_object_dealloc_list *dlist = NULL;
1780 vm_object_t backing_object;
1783 * Only one thread is attempting a collapse at any given moment.
1784 * There are few restrictions for (object) that callers of this
1785 * function check so reentrancy is likely.
1787 KKASSERT(object != NULL);
1788 vm_object_assert_held(object);
1789 KKASSERT(object->flags & OBJ_CHAINLOCK);
1796 * We have to hold the backing object, check races.
1798 while ((backing_object = object->backing_object) != NULL) {
1799 vm_object_hold(backing_object);
1800 if (backing_object == object->backing_object)
1802 vm_object_drop(backing_object);
1806 * No backing object? Nothing to collapse then.
1808 if (backing_object == NULL)
1812 * You can't collapse with a non-default/non-swap object.
1814 if (backing_object->type != OBJT_DEFAULT &&
1815 backing_object->type != OBJT_SWAP) {
1816 vm_object_drop(backing_object);
1817 backing_object = NULL;
1822 * Chain-lock the backing object too because if we
1823 * successfully merge its pages into the top object we
1824 * will collapse backing_object->backing_object as the
1825 * new backing_object. Re-check that it is still our
1828 vm_object_chain_acquire(backing_object);
1829 if (backing_object != object->backing_object) {
1830 vm_object_chain_release(backing_object);
1831 vm_object_drop(backing_object);
1836 * we check the backing object first, because it is most likely
1839 if (backing_object->handle != NULL ||
1840 (backing_object->type != OBJT_DEFAULT &&
1841 backing_object->type != OBJT_SWAP) ||
1842 (backing_object->flags & OBJ_DEAD) ||
1843 object->handle != NULL ||
1844 (object->type != OBJT_DEFAULT &&
1845 object->type != OBJT_SWAP) ||
1846 (object->flags & OBJ_DEAD)) {
1851 * If paging is in progress we can't do a normal collapse.
1854 object->paging_in_progress != 0 ||
1855 backing_object->paging_in_progress != 0
1857 vm_object_qcollapse(object, backing_object);
1862 * We know that we can either collapse the backing object (if
1863 * the parent is the only reference to it) or (perhaps) have
1864 * the parent bypass the object if the parent happens to shadow
1865 * all the resident pages in the entire backing object.
1867 * This is ignoring pager-backed pages such as swap pages.
1868 * vm_object_backing_scan fails the shadowing test in this
1871 if (backing_object->ref_count == 1) {
1873 * If there is exactly one reference to the backing
1874 * object, we can collapse it into the parent.
1876 KKASSERT(object->backing_object == backing_object);
1877 vm_object_backing_scan(object, backing_object,
1878 OBSC_COLLAPSE_WAIT);
1881 * Move the pager from backing_object to object.
1883 if (backing_object->type == OBJT_SWAP) {
1884 vm_object_pip_add(backing_object, 1);
1887 * scrap the paging_offset junk and do a
1888 * discrete copy. This also removes major
1889 * assumptions about how the swap-pager
1890 * works from where it doesn't belong. The
1891 * new swapper is able to optimize the
1892 * destroy-source case.
1894 vm_object_pip_add(object, 1);
1895 swap_pager_copy(backing_object, object,
1896 OFF_TO_IDX(object->backing_object_offset),
1898 vm_object_pip_wakeup(object);
1899 vm_object_pip_wakeup(backing_object);
1903 * Object now shadows whatever backing_object did.
1904 * Remove object from backing_object's shadow_list.
1906 LIST_REMOVE(object, shadow_list);
1907 KKASSERT(object->backing_object == backing_object);
1908 backing_object->shadow_count--;
1909 backing_object->generation++;
1912 * backing_object->backing_object moves from within
1913 * backing_object to within object.
1915 while ((bbobj = backing_object->backing_object) != NULL) {
1916 vm_object_hold(bbobj);
1917 if (bbobj == backing_object->backing_object)
1919 vm_object_drop(bbobj);
1922 LIST_REMOVE(backing_object, shadow_list);
1923 bbobj->shadow_count--;
1924 bbobj->generation++;
1925 backing_object->backing_object = NULL;
1927 object->backing_object = bbobj;
1929 LIST_INSERT_HEAD(&bbobj->shadow_head,
1930 object, shadow_list);
1931 bbobj->shadow_count++;
1932 bbobj->generation++;
1935 object->backing_object_offset +=
1936 backing_object->backing_object_offset;
1938 vm_object_drop(bbobj);
1941 * Discard the old backing_object. Nothing should be
1942 * able to ref it, other than a vm_map_split(),
1943 * and vm_map_split() will stall on our chain lock.
1944 * And we control the parent so it shouldn't be
1945 * possible for it to go away either.
1947 * Since the backing object has no pages, no pager
1948 * left, and no object references within it, all
1949 * that is necessary is to dispose of it.
1951 KASSERT(backing_object->ref_count == 1,
1952 ("backing_object %p was somehow "
1953 "re-referenced during collapse!",
1955 KASSERT(RB_EMPTY(&backing_object->rb_memq),
1956 ("backing_object %p somehow has left "
1957 "over pages during collapse!",
1961 * The object can be destroyed.
1963 * XXX just fall through and dodealloc instead
1964 * of forcing destruction?
1966 --backing_object->ref_count;
1967 if ((backing_object->flags & OBJ_DEAD) == 0)
1968 vm_object_terminate(backing_object);
1973 * If we do not entirely shadow the backing object,
1974 * there is nothing we can do so we give up.
1976 if (vm_object_backing_scan(object, backing_object,
1977 OBSC_TEST_ALL_SHADOWED) == 0) {
1982 * bbobj is backing_object->backing_object. Since
1983 * object completely shadows backing_object we can
1984 * bypass it and become backed by bbobj instead.
1986 while ((bbobj = backing_object->backing_object) != NULL) {
1987 vm_object_hold(bbobj);
1988 if (bbobj == backing_object->backing_object)
1990 vm_object_drop(bbobj);
1994 * Make object shadow bbobj instead of backing_object.
1995 * Remove object from backing_object's shadow list.
1997 * Deallocating backing_object will not remove
1998 * it, since its reference count is at least 2.
2000 KKASSERT(object->backing_object == backing_object);
2001 LIST_REMOVE(object, shadow_list);
2002 backing_object->shadow_count--;
2003 backing_object->generation++;
2006 * Add a ref to bbobj, bbobj now shadows object.
2008 * NOTE: backing_object->backing_object still points
2009 * to bbobj. That relationship remains intact
2010 * because backing_object has > 1 ref, so
2011 * someone else is pointing to it (hence why
2012 * we can't collapse it into object and can
2013 * only handle the all-shadowed bypass case).
2016 vm_object_chain_wait(bbobj);
2017 vm_object_reference_locked(bbobj);
2018 LIST_INSERT_HEAD(&bbobj->shadow_head,
2019 object, shadow_list);
2020 bbobj->shadow_count++;
2021 bbobj->generation++;
2022 object->backing_object_offset +=
2023 backing_object->backing_object_offset;
2024 object->backing_object = bbobj;
2025 vm_object_drop(bbobj);
2027 object->backing_object = NULL;
2031 * Drop the reference count on backing_object. To
2032 * handle ref_count races properly we can't assume
2033 * that the ref_count is still at least 2 so we
2034 * have to actually call vm_object_deallocate()
2035 * (after clearing the chainlock).
2042 * Ok, we want to loop on the new object->bbobj association,
2043 * possibly collapsing it further. However if dodealloc is
2044 * non-zero we have to deallocate the backing_object which
2045 * itself can potentially undergo a collapse, creating a
2046 * recursion depth issue with the LWKT token subsystem.
2048 * In the case where we must deallocate the backing_object
2049 * it is possible now that the backing_object has a single
2050 * shadow count on some other object (not represented here
2051 * as yet), since it no longer shadows us. Thus when we
2052 * call vm_object_deallocate() it may attempt to collapse
2053 * itself into its remaining parent.
2056 struct vm_object_dealloc_list *dtmp;
2058 vm_object_chain_release(backing_object);
2059 vm_object_unlock(backing_object);
2060 /* backing_object remains held */
2063 * Auto-deallocation list for caller convenience.
2068 dtmp = kmalloc(sizeof(*dtmp), M_TEMP, M_WAITOK);
2069 dtmp->object = backing_object;
2070 dtmp->next = *dlistp;
2073 vm_object_chain_release(backing_object);
2074 vm_object_drop(backing_object);
2076 /* backing_object = NULL; not needed */
2081 * Clean up any left over backing_object
2083 if (backing_object) {
2084 vm_object_chain_release(backing_object);
2085 vm_object_drop(backing_object);
2089 * Clean up any auto-deallocation list. This is a convenience
2090 * for top-level callers so they don't have to pass &dlist.
2091 * Do not clean up any caller-passed dlistp, the caller will
2095 vm_object_deallocate_list(&dlist);
2100 * vm_object_collapse() may collect additional objects in need of
2101 * deallocation. This routine deallocates these objects. The
2102 * deallocation itself can trigger additional collapses (which the
2103 * deallocate function takes care of). This procedure is used to
2104 * reduce procedural recursion since these vm_object shadow chains
2105 * can become quite long.
2108 vm_object_deallocate_list(struct vm_object_dealloc_list **dlistp)
2110 struct vm_object_dealloc_list *dlist;
2112 while ((dlist = *dlistp) != NULL) {
2113 *dlistp = dlist->next;
2114 vm_object_lock(dlist->object);
2115 vm_object_deallocate_locked(dlist->object);
2116 vm_object_drop(dlist->object);
2117 kfree(dlist, M_TEMP);
2122 * Removes all physical pages in the specified object range from the
2123 * object's list of pages.
2127 static int vm_object_page_remove_callback(vm_page_t p, void *data);
2130 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2131 boolean_t clean_only)
2133 struct rb_vm_page_scan_info info;
2137 * Degenerate cases and assertions
2139 vm_object_hold(object);
2140 if (object == NULL ||
2141 (object->resident_page_count == 0 && object->swblock_count == 0)) {
2142 vm_object_drop(object);
2145 KASSERT(object->type != OBJT_PHYS,
2146 ("attempt to remove pages from a physical object"));
2149 * Indicate that paging is occuring on the object
2151 vm_object_pip_add(object, 1);
2154 * Figure out the actual removal range and whether we are removing
2155 * the entire contents of the object or not. If removing the entire
2156 * contents, be sure to get all pages, even those that might be
2157 * beyond the end of the object.
2159 info.start_pindex = start;
2161 info.end_pindex = (vm_pindex_t)-1;
2163 info.end_pindex = end - 1;
2164 info.limit = clean_only;
2165 all = (start == 0 && info.end_pindex >= object->size - 1);
2168 * Loop until we are sure we have gotten them all.
2172 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2173 vm_object_page_remove_callback, &info);
2174 } while (info.error);
2177 * Remove any related swap if throwing away pages, or for
2178 * non-swap objects (the swap is a clean copy in that case).
2180 if (object->type != OBJT_SWAP || clean_only == FALSE) {
2182 swap_pager_freespace_all(object);
2184 swap_pager_freespace(object, info.start_pindex,
2185 info.end_pindex - info.start_pindex + 1);
2191 vm_object_pip_wakeup(object);
2192 vm_object_drop(object);
2196 * The caller must hold the object
2199 vm_object_page_remove_callback(vm_page_t p, void *data)
2201 struct rb_vm_page_scan_info *info = data;
2203 if (vm_page_busy_try(p, TRUE)) {
2204 vm_page_sleep_busy(p, TRUE, "vmopar");
2210 * Wired pages cannot be destroyed, but they can be invalidated
2211 * and we do so if clean_only (limit) is not set.
2213 * WARNING! The page may be wired due to being part of a buffer
2214 * cache buffer, and the buffer might be marked B_CACHE.
2215 * This is fine as part of a truncation but VFSs must be
2216 * sure to fix the buffer up when re-extending the file.
2218 if (p->wire_count != 0) {
2219 vm_page_protect(p, VM_PROT_NONE);
2220 if (info->limit == 0)
2227 * limit is our clean_only flag. If set and the page is dirty, do
2228 * not free it. If set and the page is being held by someone, do
2231 if (info->limit && p->valid) {
2232 vm_page_test_dirty(p);
2233 if (p->valid & p->dirty) {
2238 if (p->hold_count) {
2248 vm_page_protect(p, VM_PROT_NONE);
2254 * Coalesces two objects backing up adjoining regions of memory into a
2257 * returns TRUE if objects were combined.
2259 * NOTE: Only works at the moment if the second object is NULL -
2260 * if it's not, which object do we lock first?
2263 * prev_object First object to coalesce
2264 * prev_offset Offset into prev_object
2265 * next_object Second object into coalesce
2266 * next_offset Offset into next_object
2268 * prev_size Size of reference to prev_object
2269 * next_size Size of reference to next_object
2271 * The caller does not need to hold (prev_object) but must have a stable
2272 * pointer to it (typically by holding the vm_map locked).
2275 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
2276 vm_size_t prev_size, vm_size_t next_size)
2278 vm_pindex_t next_pindex;
2280 if (prev_object == NULL)
2283 vm_object_hold(prev_object);
2285 if (prev_object->type != OBJT_DEFAULT &&
2286 prev_object->type != OBJT_SWAP) {
2287 vm_object_drop(prev_object);
2292 * Try to collapse the object first
2294 vm_object_chain_acquire(prev_object);
2295 vm_object_collapse(prev_object, NULL);
2298 * Can't coalesce if: . more than one reference . paged out . shadows
2299 * another object . has a copy elsewhere (any of which mean that the
2300 * pages not mapped to prev_entry may be in use anyway)
2303 if (prev_object->backing_object != NULL) {
2304 vm_object_chain_release(prev_object);
2305 vm_object_drop(prev_object);
2309 prev_size >>= PAGE_SHIFT;
2310 next_size >>= PAGE_SHIFT;
2311 next_pindex = prev_pindex + prev_size;
2313 if ((prev_object->ref_count > 1) &&
2314 (prev_object->size != next_pindex)) {
2315 vm_object_chain_release(prev_object);
2316 vm_object_drop(prev_object);
2321 * Remove any pages that may still be in the object from a previous
2324 if (next_pindex < prev_object->size) {
2325 vm_object_page_remove(prev_object,
2327 next_pindex + next_size, FALSE);
2328 if (prev_object->type == OBJT_SWAP)
2329 swap_pager_freespace(prev_object,
2330 next_pindex, next_size);
2334 * Extend the object if necessary.
2336 if (next_pindex + next_size > prev_object->size)
2337 prev_object->size = next_pindex + next_size;
2339 vm_object_chain_release(prev_object);
2340 vm_object_drop(prev_object);
2345 * Make the object writable and flag is being possibly dirty.
2347 * The caller must hold the object. XXX called from vm_page_dirty(),
2348 * There is currently no requirement to hold the object.
2351 vm_object_set_writeable_dirty(vm_object_t object)
2355 /*vm_object_assert_held(object);*/
2356 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
2357 if (object->type == OBJT_VNODE &&
2358 (vp = (struct vnode *)object->handle) != NULL) {
2359 if ((vp->v_flag & VOBJDIRTY) == 0) {
2360 vsetflags(vp, VOBJDIRTY);
2365 #include "opt_ddb.h"
2367 #include <sys/kernel.h>
2369 #include <sys/cons.h>
2371 #include <ddb/ddb.h>
2373 static int _vm_object_in_map (vm_map_t map, vm_object_t object,
2374 vm_map_entry_t entry);
2375 static int vm_object_in_map (vm_object_t object);
2378 * The caller must hold the object.
2381 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2384 vm_map_entry_t tmpe;
2385 vm_object_t obj, nobj;
2391 tmpe = map->header.next;
2392 entcount = map->nentries;
2393 while (entcount-- && (tmpe != &map->header)) {
2394 if( _vm_object_in_map(map, object, tmpe)) {
2401 switch(entry->maptype) {
2402 case VM_MAPTYPE_SUBMAP:
2403 tmpm = entry->object.sub_map;
2404 tmpe = tmpm->header.next;
2405 entcount = tmpm->nentries;
2406 while (entcount-- && tmpe != &tmpm->header) {
2407 if( _vm_object_in_map(tmpm, object, tmpe)) {
2413 case VM_MAPTYPE_NORMAL:
2414 case VM_MAPTYPE_VPAGETABLE:
2415 obj = entry->object.vm_object;
2417 if (obj == object) {
2418 if (obj != entry->object.vm_object)
2419 vm_object_drop(obj);
2422 while ((nobj = obj->backing_object) != NULL) {
2423 vm_object_hold(nobj);
2424 if (nobj == obj->backing_object)
2426 vm_object_drop(nobj);
2428 if (obj != entry->object.vm_object) {
2430 vm_object_lock_swap();
2431 vm_object_drop(obj);
2442 static int vm_object_in_map_callback(struct proc *p, void *data);
2444 struct vm_object_in_map_info {
2453 vm_object_in_map(vm_object_t object)
2455 struct vm_object_in_map_info info;
2458 info.object = object;
2460 allproc_scan(vm_object_in_map_callback, &info);
2463 if( _vm_object_in_map(&kernel_map, object, 0))
2465 if( _vm_object_in_map(&pager_map, object, 0))
2467 if( _vm_object_in_map(&buffer_map, object, 0))
2476 vm_object_in_map_callback(struct proc *p, void *data)
2478 struct vm_object_in_map_info *info = data;
2481 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) {
2489 DB_SHOW_COMMAND(vmochk, vm_object_check)
2494 * make sure that internal objs are in a map somewhere
2495 * and none have zero ref counts.
2497 for (object = TAILQ_FIRST(&vm_object_list);
2499 object = TAILQ_NEXT(object, object_list)) {
2500 if (object->type == OBJT_MARKER)
2502 if (object->handle == NULL &&
2503 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2504 if (object->ref_count == 0) {
2505 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2506 (long)object->size);
2508 if (!vm_object_in_map(object)) {
2510 "vmochk: internal obj is not in a map: "
2511 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2512 object->ref_count, (u_long)object->size,
2513 (u_long)object->size,
2514 (void *)object->backing_object);
2523 DB_SHOW_COMMAND(object, vm_object_print_static)
2525 /* XXX convert args. */
2526 vm_object_t object = (vm_object_t)addr;
2527 boolean_t full = have_addr;
2531 /* XXX count is an (unused) arg. Avoid shadowing it. */
2532 #define count was_count
2540 "Object %p: type=%d, size=0x%lx, res=%d, ref=%d, flags=0x%x\n",
2541 object, (int)object->type, (u_long)object->size,
2542 object->resident_page_count, object->ref_count, object->flags);
2544 * XXX no %qd in kernel. Truncate object->backing_object_offset.
2546 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
2547 object->shadow_count,
2548 object->backing_object ? object->backing_object->ref_count : 0,
2549 object->backing_object, (long)object->backing_object_offset);
2556 RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) {
2558 db_iprintf("memory:=");
2559 else if (count == 6) {
2567 db_printf("(off=0x%lx,page=0x%lx)",
2568 (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
2579 * XXX need this non-static entry for calling from vm_map_print.
2584 vm_object_print(/* db_expr_t */ long addr,
2585 boolean_t have_addr,
2586 /* db_expr_t */ long count,
2589 vm_object_print_static(addr, have_addr, count, modif);
2595 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2600 for (object = TAILQ_FIRST(&vm_object_list);
2602 object = TAILQ_NEXT(object, object_list)) {
2603 vm_pindex_t idx, fidx;
2605 vm_paddr_t pa = -1, padiff;
2609 if (object->type == OBJT_MARKER)
2611 db_printf("new object: %p\n", (void *)object);
2621 osize = object->size;
2624 for (idx = 0; idx < osize; idx++) {
2625 m = vm_page_lookup(object, idx);
2628 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2629 (long)fidx, rcount, (long)pa);
2644 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2649 padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
2650 padiff >>= PAGE_SHIFT;
2651 padiff &= PQ_L2_MASK;
2653 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
2657 db_printf(" index(%ld)run(%d)pa(0x%lx)",
2658 (long)fidx, rcount, (long)pa);
2659 db_printf("pd(%ld)\n", (long)padiff);
2669 pa = VM_PAGE_TO_PHYS(m);
2673 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2674 (long)fidx, rcount, (long)pa);