1/*-
2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 *
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * SUCH DAMAGE.
31 *
32 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94
33 *
34 *
35 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36 * All rights reserved.
37 *
38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39 *
40 * Permission to use, copy, modify and distribute this software and
41 * its documentation is hereby granted, provided that both the copyright
42 * notice and this permission notice appear in all copies of the
43 * software, derivative works or modified versions, and any portions
44 * thereof, and that both notices appear in supporting documentation.
45 *
46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49 *
50 * Carnegie Mellon requests users of this software to return to
51 *
52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
53 * School of Computer Science
54 * Carnegie Mellon University
55 * Pittsburgh PA 15213-3890
56 *
57 * any improvements or extensions that they make and grant Carnegie the
58 * rights to redistribute these changes.
59 */
60
61/*
62 * Virtual memory object module.
63 */
64
65#include <sys/cdefs.h>
66__FBSDID("$FreeBSD: head/sys/vm/vm_object.c 254138 2013-08-09 11:11:11Z attilio $");
67
68#include "opt_vm.h"
69
70#include <sys/param.h>
71#include <sys/systm.h>
72#include <sys/lock.h>
73#include <sys/mman.h>
74#include <sys/mount.h>
75#include <sys/kernel.h>
76#include <sys/sysctl.h>
77#include <sys/mutex.h>
78#include <sys/proc.h> /* for curproc, pageproc */
79#include <sys/socket.h>
80#include <sys/resourcevar.h>
81#include <sys/rwlock.h>
82#include <sys/vnode.h>
83#include <sys/vmmeter.h>
84#include <sys/sx.h>
85
86#include <vm/vm.h>
87#include <vm/vm_param.h>
88#include <vm/pmap.h>
89#include <vm/vm_map.h>
90#include <vm/vm_object.h>
91#include <vm/vm_page.h>
92#include <vm/vm_pageout.h>
93#include <vm/vm_pager.h>
94#include <vm/swap_pager.h>
95#include <vm/vm_kern.h>
96#include <vm/vm_extern.h>
97#include <vm/vm_radix.h>
98#include <vm/vm_reserv.h>
99#include <vm/uma.h>
100
101static int old_msync;
102SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
103 "Use old (insecure) msync behavior");
104
105static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
106 int pagerflags, int flags, boolean_t *clearobjflags,
107 boolean_t *eio);
108static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
109 boolean_t *clearobjflags);
110static void vm_object_qcollapse(vm_object_t object);
111static void vm_object_vndeallocate(vm_object_t object);
112
113/*
114 * Virtual memory objects maintain the actual data
115 * associated with allocated virtual memory. A given
116 * page of memory exists within exactly one object.
117 *
118 * An object is only deallocated when all "references"
119 * are given up. Only one "reference" to a given
120 * region of an object should be writeable.
121 *
122 * Associated with each object is a list of all resident
123 * memory pages belonging to that object; this list is
124 * maintained by the "vm_page" module, and locked by the object's
125 * lock.
126 *
127 * Each object also records a "pager" routine which is
128 * used to retrieve (and store) pages to the proper backing
129 * storage. In addition, objects may be backed by other
130 * objects from which they were virtual-copied.
131 *
132 * The only items within the object structure which are
133 * modified after time of creation are:
134 * reference count locked by object's lock
135 * pager routine locked by object's lock
136 *
137 */
138
139struct object_q vm_object_list;
140struct mtx vm_object_list_mtx; /* lock for object list and count */
141
142struct vm_object kernel_object_store;
143struct vm_object kmem_object_store;
144
145static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
146 "VM object stats");
147
148static long object_collapses;
149SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
150 &object_collapses, 0, "VM object collapses");
151
152static long object_bypasses;
153SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
154 &object_bypasses, 0, "VM object bypasses");
155
156static uma_zone_t obj_zone;
157
158static int vm_object_zinit(void *mem, int size, int flags);
159
160#ifdef INVARIANTS
161static void vm_object_zdtor(void *mem, int size, void *arg);
162
163static void
164vm_object_zdtor(void *mem, int size, void *arg)
165{
166 vm_object_t object;
167
168 object = (vm_object_t)mem;
169 KASSERT(TAILQ_EMPTY(&object->memq),
170 ("object %p has resident pages in its memq", object));
171 KASSERT(vm_radix_is_empty(&object->rtree),
172 ("object %p has resident pages in its trie", object));
173#if VM_NRESERVLEVEL > 0
174 KASSERT(LIST_EMPTY(&object->rvq),
175 ("object %p has reservations",
176 object));
177#endif
178 KASSERT(vm_object_cache_is_empty(object),
179 ("object %p has cached pages",
180 object));
181 KASSERT(object->paging_in_progress == 0,
182 ("object %p paging_in_progress = %d",
183 object, object->paging_in_progress));
184 KASSERT(object->resident_page_count == 0,
185 ("object %p resident_page_count = %d",
186 object, object->resident_page_count));
187 KASSERT(object->shadow_count == 0,
188 ("object %p shadow_count = %d",
189 object, object->shadow_count));
190}
191#endif
192
193static int
194vm_object_zinit(void *mem, int size, int flags)
195{
196 vm_object_t object;
197
198 object = (vm_object_t)mem;
199 bzero(&object->lock, sizeof(object->lock));
200 rw_init_flags(&object->lock, "vm object", RW_DUPOK);
201
202 /* These are true for any object that has been freed */
203 object->rtree.rt_root = 0;
204 object->paging_in_progress = 0;
205 object->resident_page_count = 0;
206 object->shadow_count = 0;
207 object->cache.rt_root = 0;
208 return (0);
209}
210
211static void
212_vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
213{
214
215 TAILQ_INIT(&object->memq);
216 LIST_INIT(&object->shadow_head);
217
218 object->type = type;
219 switch (type) {
220 case OBJT_DEAD:
221 panic("_vm_object_allocate: can't create OBJT_DEAD");
222 case OBJT_DEFAULT:
223 case OBJT_SWAP:
224 object->flags = OBJ_ONEMAPPING;
225 break;
226 case OBJT_DEVICE:
227 case OBJT_SG:
228 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
229 break;
230 case OBJT_MGTDEVICE:
231 object->flags = OBJ_FICTITIOUS;
232 break;
233 case OBJT_PHYS:
234 object->flags = OBJ_UNMANAGED;
235 break;
236 case OBJT_VNODE:
237 object->flags = 0;
238 break;
239 default:
240 panic("_vm_object_allocate: type %d is undefined", type);
241 }
242 object->size = size;
243 object->generation = 1;
244 object->ref_count = 1;
245 object->memattr = VM_MEMATTR_DEFAULT;
246 object->cred = NULL;
247 object->charge = 0;
248 object->handle = NULL;
249 object->backing_object = NULL;
250 object->backing_object_offset = (vm_ooffset_t) 0;
251#if VM_NRESERVLEVEL > 0
252 LIST_INIT(&object->rvq);
253#endif
254
255 mtx_lock(&vm_object_list_mtx);
256 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
257 mtx_unlock(&vm_object_list_mtx);
258}
259
260/*
261 * vm_object_init:
262 *
263 * Initialize the VM objects module.
264 */
265void
266vm_object_init(void)
267{
268 TAILQ_INIT(&vm_object_list);
269 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
270
271 rw_init(&kernel_object->lock, "kernel vm object");
272 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
273 kernel_object);
274#if VM_NRESERVLEVEL > 0
275 kernel_object->flags |= OBJ_COLORED;
276 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
277#endif
278
279 rw_init(&kmem_object->lock, "kmem vm object");
280 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
281 kmem_object);
282#if VM_NRESERVLEVEL > 0
283 kmem_object->flags |= OBJ_COLORED;
284 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
285#endif
286
287 /*
288 * The lock portion of struct vm_object must be type stable due
289 * to vm_pageout_fallback_object_lock locking a vm object
290 * without holding any references to it.
291 */
292 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
293#ifdef INVARIANTS
294 vm_object_zdtor,
295#else
296 NULL,
297#endif
298 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
299
300 vm_radix_init();
301}
302
303void
304vm_object_clear_flag(vm_object_t object, u_short bits)
305{
306
307 VM_OBJECT_ASSERT_WLOCKED(object);
308 object->flags &= ~bits;
309}
310
311/*
312 * Sets the default memory attribute for the specified object. Pages
313 * that are allocated to this object are by default assigned this memory
314 * attribute.
315 *
316 * Presently, this function must be called before any pages are allocated
317 * to the object. In the future, this requirement may be relaxed for
318 * "default" and "swap" objects.
319 */
320int
321vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
322{
323
324 VM_OBJECT_ASSERT_WLOCKED(object);
325 switch (object->type) {
326 case OBJT_DEFAULT:
327 case OBJT_DEVICE:
328 case OBJT_MGTDEVICE:
329 case OBJT_PHYS:
330 case OBJT_SG:
331 case OBJT_SWAP:
332 case OBJT_VNODE:
333 if (!TAILQ_EMPTY(&object->memq))
334 return (KERN_FAILURE);
335 break;
336 case OBJT_DEAD:
337 return (KERN_INVALID_ARGUMENT);
338 default:
339 panic("vm_object_set_memattr: object %p is of undefined type",
340 object);
341 }
342 object->memattr = memattr;
343 return (KERN_SUCCESS);
344}
345
346void
347vm_object_pip_add(vm_object_t object, short i)
348{
349
350 VM_OBJECT_ASSERT_WLOCKED(object);
351 object->paging_in_progress += i;
352}
353
354void
355vm_object_pip_subtract(vm_object_t object, short i)
356{
357
358 VM_OBJECT_ASSERT_WLOCKED(object);
359 object->paging_in_progress -= i;
360}
361
362void
363vm_object_pip_wakeup(vm_object_t object)
364{
365
366 VM_OBJECT_ASSERT_WLOCKED(object);
367 object->paging_in_progress--;
368 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
369 vm_object_clear_flag(object, OBJ_PIPWNT);
370 wakeup(object);
371 }
372}
373
374void
375vm_object_pip_wakeupn(vm_object_t object, short i)
376{
377
378 VM_OBJECT_ASSERT_WLOCKED(object);
379 if (i)
380 object->paging_in_progress -= i;
381 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
382 vm_object_clear_flag(object, OBJ_PIPWNT);
383 wakeup(object);
384 }
385}
386
387void
388vm_object_pip_wait(vm_object_t object, char *waitid)
389{
390
391 VM_OBJECT_ASSERT_WLOCKED(object);
392 while (object->paging_in_progress) {
393 object->flags |= OBJ_PIPWNT;
394 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
395 }
396}
397
398/*
399 * vm_object_allocate:
400 *
401 * Returns a new object with the given size.
402 */
403vm_object_t
404vm_object_allocate(objtype_t type, vm_pindex_t size)
405{
406 vm_object_t object;
407
408 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
409 _vm_object_allocate(type, size, object);
410 return (object);
411}
412
413
414/*
415 * vm_object_reference:
416 *
417 * Gets another reference to the given object. Note: OBJ_DEAD
418 * objects can be referenced during final cleaning.
419 */
420void
421vm_object_reference(vm_object_t object)
422{
423 if (object == NULL)
424 return;
425 VM_OBJECT_WLOCK(object);
426 vm_object_reference_locked(object);
427 VM_OBJECT_WUNLOCK(object);
428}
429
430/*
431 * vm_object_reference_locked:
432 *
433 * Gets another reference to the given object.
434 *
435 * The object must be locked.
436 */
437void
438vm_object_reference_locked(vm_object_t object)
439{
440 struct vnode *vp;
441
442 VM_OBJECT_ASSERT_WLOCKED(object);
443 object->ref_count++;
444 if (object->type == OBJT_VNODE) {
445 vp = object->handle;
446 vref(vp);
447 }
448}
449
450/*
451 * Handle deallocating an object of type OBJT_VNODE.
452 */
453static void
454vm_object_vndeallocate(vm_object_t object)
455{
456 struct vnode *vp = (struct vnode *) object->handle;
457
458 VM_OBJECT_ASSERT_WLOCKED(object);
459 KASSERT(object->type == OBJT_VNODE,
460 ("vm_object_vndeallocate: not a vnode object"));
461 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
462#ifdef INVARIANTS
463 if (object->ref_count == 0) {
464 vprint("vm_object_vndeallocate", vp);
465 panic("vm_object_vndeallocate: bad object reference count");
466 }
467#endif
468
469 if (object->ref_count > 1) {
470 object->ref_count--;
471 VM_OBJECT_WUNLOCK(object);
472 /* vrele may need the vnode lock. */
473 vrele(vp);
474 } else {
475 vhold(vp);
476 VM_OBJECT_WUNLOCK(object);
477 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
478 vdrop(vp);
479 VM_OBJECT_WLOCK(object);
480 object->ref_count--;
481 if (object->type == OBJT_DEAD) {
482 VM_OBJECT_WUNLOCK(object);
483 VOP_UNLOCK(vp, 0);
484 } else {
485 if (object->ref_count == 0)
486 VOP_UNSET_TEXT(vp);
487 VM_OBJECT_WUNLOCK(object);
488 vput(vp);
489 }
490 }
491}
492
493/*
494 * vm_object_deallocate:
495 *
496 * Release a reference to the specified object,
497 * gained either through a vm_object_allocate
498 * or a vm_object_reference call. When all references
499 * are gone, storage associated with this object
500 * may be relinquished.
501 *
502 * No object may be locked.
503 */
504void
505vm_object_deallocate(vm_object_t object)
506{
507 vm_object_t temp;
508 struct vnode *vp;
509
510 while (object != NULL) {
511 VM_OBJECT_WLOCK(object);
512 if (object->type == OBJT_VNODE) {
513 vm_object_vndeallocate(object);
514 return;
515 }
516
517 KASSERT(object->ref_count != 0,
518 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
519
520 /*
521 * If the reference count goes to 0 we start calling
522 * vm_object_terminate() on the object chain.
523 * A ref count of 1 may be a special case depending on the
524 * shadow count being 0 or 1.
525 */
526 object->ref_count--;
527 if (object->ref_count > 1) {
528 VM_OBJECT_WUNLOCK(object);
529 return;
530 } else if (object->ref_count == 1) {
531 if (object->type == OBJT_SWAP &&
532 (object->flags & OBJ_TMPFS) != 0) {
533 vp = object->un_pager.swp.swp_tmpfs;
534 vhold(vp);
535 VM_OBJECT_WUNLOCK(object);
536 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
537 vdrop(vp);
538 VM_OBJECT_WLOCK(object);
539 if (object->type == OBJT_DEAD ||
540 object->ref_count != 1) {
541 VM_OBJECT_WUNLOCK(object);
542 VOP_UNLOCK(vp, 0);
543 return;
544 }
545 if ((object->flags & OBJ_TMPFS) != 0)
546 VOP_UNSET_TEXT(vp);
547 VOP_UNLOCK(vp, 0);
548 }
549 if (object->shadow_count == 0 &&
550 object->handle == NULL &&
551 (object->type == OBJT_DEFAULT ||
552 (object->type == OBJT_SWAP &&
553 (object->flags & OBJ_TMPFS) == 0))) {
554 vm_object_set_flag(object, OBJ_ONEMAPPING);
555 } else if ((object->shadow_count == 1) &&
556 (object->handle == NULL) &&
557 (object->type == OBJT_DEFAULT ||
558 object->type == OBJT_SWAP)) {
559 KASSERT((object->flags & OBJ_TMPFS) == 0,
560 ("shadowed tmpfs v_object %p", object));
561 vm_object_t robject;
562
563 robject = LIST_FIRST(&object->shadow_head);
564 KASSERT(robject != NULL,
565 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
566 object->ref_count,
567 object->shadow_count));
568 if (!VM_OBJECT_TRYWLOCK(robject)) {
569 /*
570 * Avoid a potential deadlock.
571 */
572 object->ref_count++;
573 VM_OBJECT_WUNLOCK(object);
574 /*
575 * More likely than not the thread
576 * holding robject's lock has lower
577 * priority than the current thread.
578 * Let the lower priority thread run.
579 */
580 pause("vmo_de", 1);
581 continue;
582 }
583 /*
584 * Collapse object into its shadow unless its
585 * shadow is dead. In that case, object will
586 * be deallocated by the thread that is
587 * deallocating its shadow.
588 */
589 if ((robject->flags & OBJ_DEAD) == 0 &&
590 (robject->handle == NULL) &&
591 (robject->type == OBJT_DEFAULT ||
592 robject->type == OBJT_SWAP)) {
593
594 robject->ref_count++;
595retry:
596 if (robject->paging_in_progress) {
597 VM_OBJECT_WUNLOCK(object);
598 vm_object_pip_wait(robject,
599 "objde1");
600 temp = robject->backing_object;
601 if (object == temp) {
602 VM_OBJECT_WLOCK(object);
603 goto retry;
604 }
605 } else if (object->paging_in_progress) {
606 VM_OBJECT_WUNLOCK(robject);
607 object->flags |= OBJ_PIPWNT;
608 VM_OBJECT_SLEEP(object, object,
609 PDROP | PVM, "objde2", 0);
610 VM_OBJECT_WLOCK(robject);
611 temp = robject->backing_object;
612 if (object == temp) {
613 VM_OBJECT_WLOCK(object);
614 goto retry;
615 }
616 } else
617 VM_OBJECT_WUNLOCK(object);
618
619 if (robject->ref_count == 1) {
620 robject->ref_count--;
621 object = robject;
622 goto doterm;
623 }
624 object = robject;
625 vm_object_collapse(object);
626 VM_OBJECT_WUNLOCK(object);
627 continue;
628 }
629 VM_OBJECT_WUNLOCK(robject);
630 }
631 VM_OBJECT_WUNLOCK(object);
632 return;
633 }
634doterm:
635 temp = object->backing_object;
636 if (temp != NULL) {
637 VM_OBJECT_WLOCK(temp);
638 LIST_REMOVE(object, shadow_list);
639 temp->shadow_count--;
640 VM_OBJECT_WUNLOCK(temp);
641 object->backing_object = NULL;
642 }
643 /*
644 * Don't double-terminate, we could be in a termination
645 * recursion due to the terminate having to sync data
646 * to disk.
647 */
648 if ((object->flags & OBJ_DEAD) == 0)
649 vm_object_terminate(object);
650 else
651 VM_OBJECT_WUNLOCK(object);
652 object = temp;
653 }
654}
655
656/*
657 * vm_object_destroy removes the object from the global object list
658 * and frees the space for the object.
659 */
660void
661vm_object_destroy(vm_object_t object)
662{
663
664 /*
665 * Remove the object from the global object list.
666 */
667 mtx_lock(&vm_object_list_mtx);
668 TAILQ_REMOVE(&vm_object_list, object, object_list);
669 mtx_unlock(&vm_object_list_mtx);
670
671 /*
672 * Release the allocation charge.
673 */
674 if (object->cred != NULL) {
675 KASSERT(object->type == OBJT_DEFAULT ||
676 object->type == OBJT_SWAP,
677 ("vm_object_terminate: non-swap obj %p has cred",
678 object));
679 swap_release_by_cred(object->charge, object->cred);
680 object->charge = 0;
681 crfree(object->cred);
682 object->cred = NULL;
683 }
684
685 /*
686 * Free the space for the object.
687 */
688 uma_zfree(obj_zone, object);
689}
690
691/*
692 * vm_object_terminate actually destroys the specified object, freeing
693 * up all previously used resources.
694 *
695 * The object must be locked.
696 * This routine may block.
697 */
698void
699vm_object_terminate(vm_object_t object)
700{
701 vm_page_t p, p_next;
702
703 VM_OBJECT_ASSERT_WLOCKED(object);
704
705 /*
706 * Make sure no one uses us.
707 */
708 vm_object_set_flag(object, OBJ_DEAD);
709
710 /*
711 * wait for the pageout daemon to be done with the object
712 */
713 vm_object_pip_wait(object, "objtrm");
714
715 KASSERT(!object->paging_in_progress,
716 ("vm_object_terminate: pageout in progress"));
717
718 /*
719 * Clean and free the pages, as appropriate. All references to the
720 * object are gone, so we don't need to lock it.
721 */
722 if (object->type == OBJT_VNODE) {
723 struct vnode *vp = (struct vnode *)object->handle;
724
725 /*
726 * Clean pages and flush buffers.
727 */
728 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
729 VM_OBJECT_WUNLOCK(object);
730
731 vinvalbuf(vp, V_SAVE, 0, 0);
732
733 VM_OBJECT_WLOCK(object);
734 }
735
736 KASSERT(object->ref_count == 0,
737 ("vm_object_terminate: object with references, ref_count=%d",
738 object->ref_count));
739
740 /*
741 * Free any remaining pageable pages. This also removes them from the
742 * paging queues. However, don't free wired pages, just remove them
743 * from the object. Rather than incrementally removing each page from
744 * the object, the page and object are reset to any empty state.
745 */
746 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
747 vm_page_assert_unbusied(p);
748 vm_page_lock(p);
749 /*
750 * Optimize the page's removal from the object by resetting
751 * its "object" field. Specifically, if the page is not
752 * wired, then the effect of this assignment is that
753 * vm_page_free()'s call to vm_page_remove() will return
754 * immediately without modifying the page or the object.
755 */
756 p->object = NULL;
757 if (p->wire_count == 0) {
758 vm_page_free(p);
759 PCPU_INC(cnt.v_pfree);
760 }
761 vm_page_unlock(p);
762 }
763 /*
764 * If the object contained any pages, then reset it to an empty state.
765 * None of the object's fields, including "resident_page_count", were
766 * modified by the preceding loop.
767 */
768 if (object->resident_page_count != 0) {
769 vm_radix_reclaim_allnodes(&object->rtree);
770 TAILQ_INIT(&object->memq);
771 object->resident_page_count = 0;
772 if (object->type == OBJT_VNODE)
773 vdrop(object->handle);
774 }
775
776#if VM_NRESERVLEVEL > 0
777 if (__predict_false(!LIST_EMPTY(&object->rvq)))
778 vm_reserv_break_all(object);
779#endif
780 if (__predict_false(!vm_object_cache_is_empty(object)))
781 vm_page_cache_free(object, 0, 0);
782
783 /*
784 * Let the pager know object is dead.
785 */
786 vm_pager_deallocate(object);
787 VM_OBJECT_WUNLOCK(object);
788
789 vm_object_destroy(object);
790}
791
792/*
793 * Make the page read-only so that we can clear the object flags. However, if
794 * this is a nosync mmap then the object is likely to stay dirty so do not
795 * mess with the page and do not clear the object flags. Returns TRUE if the
796 * page should be flushed, and FALSE otherwise.
797 */
798static boolean_t
799vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
800{
801
802 /*
803 * If we have been asked to skip nosync pages and this is a
804 * nosync page, skip it. Note that the object flags were not
805 * cleared in this case so we do not have to set them.
806 */
807 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
808 *clearobjflags = FALSE;
809 return (FALSE);
810 } else {
811 pmap_remove_write(p);
812 return (p->dirty != 0);
813 }
814}
815
816/*
817 * vm_object_page_clean
818 *
819 * Clean all dirty pages in the specified range of object. Leaves page
820 * on whatever queue it is currently on. If NOSYNC is set then do not
821 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
822 * leaving the object dirty.
823 *
824 * When stuffing pages asynchronously, allow clustering. XXX we need a
825 * synchronous clustering mode implementation.
826 *
827 * Odd semantics: if start == end, we clean everything.
828 *
829 * The object must be locked.
830 *
831 * Returns FALSE if some page from the range was not written, as
832 * reported by the pager, and TRUE otherwise.
833 */
834boolean_t
835vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
836 int flags)
837{
838 vm_page_t np, p;
839 vm_pindex_t pi, tend, tstart;
840 int curgeneration, n, pagerflags;
841 boolean_t clearobjflags, eio, res;
842
843 VM_OBJECT_ASSERT_WLOCKED(object);
844
845 /*
846 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
847 * objects. The check below prevents the function from
848 * operating on non-vnode objects.
849 */
850 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
851 object->resident_page_count == 0)
852 return (TRUE);
853
854 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
855 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
856 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
857
858 tstart = OFF_TO_IDX(start);
859 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
860 clearobjflags = tstart == 0 && tend >= object->size;
861 res = TRUE;
862
863rescan:
864 curgeneration = object->generation;
865
866 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
867 pi = p->pindex;
868 if (pi >= tend)
869 break;
870 np = TAILQ_NEXT(p, listq);
871 if (p->valid == 0)
872 continue;
873 if (vm_page_sleep_if_busy(p, "vpcwai")) {
874 if (object->generation != curgeneration) {
875 if ((flags & OBJPC_SYNC) != 0)
876 goto rescan;
877 else
878 clearobjflags = FALSE;
879 }
880 np = vm_page_find_least(object, pi);
881 continue;
882 }
883 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
884 continue;
885
886 n = vm_object_page_collect_flush(object, p, pagerflags,
887 flags, &clearobjflags, &eio);
888 if (eio) {
889 res = FALSE;
890 clearobjflags = FALSE;
891 }
892 if (object->generation != curgeneration) {
893 if ((flags & OBJPC_SYNC) != 0)
894 goto rescan;
895 else
896 clearobjflags = FALSE;
897 }
898
899 /*
900 * If the VOP_PUTPAGES() did a truncated write, so
901 * that even the first page of the run is not fully
902 * written, vm_pageout_flush() returns 0 as the run
903 * length. Since the condition that caused truncated
904 * write may be permanent, e.g. exhausted free space,
905 * accepting n == 0 would cause an infinite loop.
906 *
907 * Forwarding the iterator leaves the unwritten page
908 * behind, but there is not much we can do there if
909 * filesystem refuses to write it.
910 */
911 if (n == 0) {
912 n = 1;
913 clearobjflags = FALSE;
914 }
915 np = vm_page_find_least(object, pi + n);
916 }
917#if 0
918 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
919#endif
920
921 if (clearobjflags)
922 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
923 return (res);
924}
925
926static int
927vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
928 int flags, boolean_t *clearobjflags, boolean_t *eio)
929{
930 vm_page_t ma[vm_pageout_page_count], p_first, tp;
931 int count, i, mreq, runlen;
932
933 vm_page_lock_assert(p, MA_NOTOWNED);
934 VM_OBJECT_ASSERT_WLOCKED(object);
935
936 count = 1;
937 mreq = 0;
938
939 for (tp = p; count < vm_pageout_page_count; count++) {
940 tp = vm_page_next(tp);
941 if (tp == NULL || vm_page_busied(tp))
942 break;
943 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
944 break;
945 }
946
947 for (p_first = p; count < vm_pageout_page_count; count++) {
948 tp = vm_page_prev(p_first);
949 if (tp == NULL || vm_page_busied(tp))
950 break;
951 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
952 break;
953 p_first = tp;
954 mreq++;
955 }
956
957 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
958 ma[i] = tp;
959
960 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
961 return (runlen);
962}
963
964/*
965 * Note that there is absolutely no sense in writing out
966 * anonymous objects, so we track down the vnode object
967 * to write out.
968 * We invalidate (remove) all pages from the address space
969 * for semantic correctness.
970 *
971 * If the backing object is a device object with unmanaged pages, then any
972 * mappings to the specified range of pages must be removed before this
973 * function is called.
974 *
975 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
976 * may start out with a NULL object.
977 */
978boolean_t
979vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
980 boolean_t syncio, boolean_t invalidate)
981{
982 vm_object_t backing_object;
983 struct vnode *vp;
984 struct mount *mp;
985 int error, flags, fsync_after;
986 boolean_t res;
987
988 if (object == NULL)
989 return (TRUE);
990 res = TRUE;
991 error = 0;
992 VM_OBJECT_WLOCK(object);
993 while ((backing_object = object->backing_object) != NULL) {
994 VM_OBJECT_WLOCK(backing_object);
995 offset += object->backing_object_offset;
996 VM_OBJECT_WUNLOCK(object);
997 object = backing_object;
998 if (object->size < OFF_TO_IDX(offset + size))
999 size = IDX_TO_OFF(object->size) - offset;
1000 }
1001 /*
1002 * Flush pages if writing is allowed, invalidate them
1003 * if invalidation requested. Pages undergoing I/O
1004 * will be ignored by vm_object_page_remove().
1005 *
1006 * We cannot lock the vnode and then wait for paging
1007 * to complete without deadlocking against vm_fault.
1008 * Instead we simply call vm_object_page_remove() and
1009 * allow it to block internally on a page-by-page
1010 * basis when it encounters pages undergoing async
1011 * I/O.
1012 */
1013 if (object->type == OBJT_VNODE &&
1014 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1015 vp = object->handle;
1016 VM_OBJECT_WUNLOCK(object);
1017 (void) vn_start_write(vp, &mp, V_WAIT);
1018 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1019 if (syncio && !invalidate && offset == 0 &&
1020 OFF_TO_IDX(size) == object->size) {
1021 /*
1022 * If syncing the whole mapping of the file,
1023 * it is faster to schedule all the writes in
1024 * async mode, also allowing the clustering,
1025 * and then wait for i/o to complete.
1026 */
1027 flags = 0;
1028 fsync_after = TRUE;
1029 } else {
1030 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1031 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1032 fsync_after = FALSE;
1033 }
1034 VM_OBJECT_WLOCK(object);
1035 res = vm_object_page_clean(object, offset, offset + size,
1036 flags);
1037 VM_OBJECT_WUNLOCK(object);
1038 if (fsync_after)
1039 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1040 VOP_UNLOCK(vp, 0);
1041 vn_finished_write(mp);
1042 if (error != 0)
1043 res = FALSE;
1044 VM_OBJECT_WLOCK(object);
1045 }
1046 if ((object->type == OBJT_VNODE ||
1047 object->type == OBJT_DEVICE) && invalidate) {
1048 if (object->type == OBJT_DEVICE)
1049 /*
1050 * The option OBJPR_NOTMAPPED must be passed here
1051 * because vm_object_page_remove() cannot remove
1052 * unmanaged mappings.
1053 */
1054 flags = OBJPR_NOTMAPPED;
1055 else if (old_msync)
1056 flags = OBJPR_NOTWIRED;
1057 else
1058 flags = OBJPR_CLEANONLY | OBJPR_NOTWIRED;
1059 vm_object_page_remove(object, OFF_TO_IDX(offset),
1060 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1061 }
1062 VM_OBJECT_WUNLOCK(object);
1063 return (res);
1064}
1065
1066/*
1067 * vm_object_madvise:
1068 *
1069 * Implements the madvise function at the object/page level.
1070 *
1071 * MADV_WILLNEED (any object)
1072 *
1073 * Activate the specified pages if they are resident.
1074 *
1075 * MADV_DONTNEED (any object)
1076 *
1077 * Deactivate the specified pages if they are resident.
1078 *
1079 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1080 * OBJ_ONEMAPPING only)
1081 *
1082 * Deactivate and clean the specified pages if they are
1083 * resident. This permits the process to reuse the pages
1084 * without faulting or the kernel to reclaim the pages
1085 * without I/O.
1086 */
1087void
1088vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1089 int advise)
1090{
1091 vm_pindex_t tpindex;
1092 vm_object_t backing_object, tobject;
1093 vm_page_t m;
1094
1095 if (object == NULL)
1096 return;
1097 VM_OBJECT_WLOCK(object);
1098 /*
1099 * Locate and adjust resident pages
1100 */
1101 for (; pindex < end; pindex += 1) {
1102relookup:
1103 tobject = object;
1104 tpindex = pindex;
1105shadowlookup:
1106 /*
1107 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1108 * and those pages must be OBJ_ONEMAPPING.
1109 */
1110 if (advise == MADV_FREE) {
1111 if ((tobject->type != OBJT_DEFAULT &&
1112 tobject->type != OBJT_SWAP) ||
1113 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1114 goto unlock_tobject;
1115 }
1116 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1117 goto unlock_tobject;
1118 m = vm_page_lookup(tobject, tpindex);
1119 if (m == NULL && advise == MADV_WILLNEED) {
1120 /*
1121 * If the page is cached, reactivate it.
1122 */
1123 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1124 VM_ALLOC_NOBUSY);
1125 }
1126 if (m == NULL) {
1127 /*
1128 * There may be swap even if there is no backing page
1129 */
1130 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1131 swap_pager_freespace(tobject, tpindex, 1);
1132 /*
1133 * next object
1134 */
1135 backing_object = tobject->backing_object;
1136 if (backing_object == NULL)
1137 goto unlock_tobject;
1138 VM_OBJECT_WLOCK(backing_object);
1139 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1140 if (tobject != object)
1141 VM_OBJECT_WUNLOCK(tobject);
1142 tobject = backing_object;
1143 goto shadowlookup;
1144 } else if (m->valid != VM_PAGE_BITS_ALL)
1145 goto unlock_tobject;
1146 /*
1147 * If the page is not in a normal state, skip it.
1148 */
1149 vm_page_lock(m);
1150 if (m->hold_count != 0 || m->wire_count != 0) {
1151 vm_page_unlock(m);
1152 goto unlock_tobject;
1153 }
1154 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1155 ("vm_object_madvise: page %p is fictitious", m));
1156 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1157 ("vm_object_madvise: page %p is not managed", m));
1158 if (vm_page_busied(m)) {
1159 if (advise == MADV_WILLNEED) {
1160 /*
1161 * Reference the page before unlocking and
1162 * sleeping so that the page daemon is less
1163 * likely to reclaim it.
1164 */
1165 vm_page_aflag_set(m, PGA_REFERENCED);
1166 }
1167 if (object != tobject)
1168 VM_OBJECT_WUNLOCK(object);
1169 VM_OBJECT_WUNLOCK(tobject);
1170 vm_page_busy_sleep(m, "madvpo");
1171 VM_OBJECT_WLOCK(object);
1172 goto relookup;
1173 }
1174 if (advise == MADV_WILLNEED) {
1175 vm_page_activate(m);
1176 } else {
1177 vm_page_advise(m, advise);
1178 }
1179 vm_page_unlock(m);
1180 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1181 swap_pager_freespace(tobject, tpindex, 1);
1182unlock_tobject:
1183 if (tobject != object)
1184 VM_OBJECT_WUNLOCK(tobject);
1185 }
1186 VM_OBJECT_WUNLOCK(object);
1187}
1188
1189/*
1190 * vm_object_shadow:
1191 *
1192 * Create a new object which is backed by the
1193 * specified existing object range. The source
1194 * object reference is deallocated.
1195 *
1196 * The new object and offset into that object
1197 * are returned in the source parameters.
1198 */
1199void
1200vm_object_shadow(
1201 vm_object_t *object, /* IN/OUT */
1202 vm_ooffset_t *offset, /* IN/OUT */
1203 vm_size_t length)
1204{
1205 vm_object_t source;
1206 vm_object_t result;
1207
1208 source = *object;
1209
1210 /*
1211 * Don't create the new object if the old object isn't shared.
1212 */
1213 if (source != NULL) {
1214 VM_OBJECT_WLOCK(source);
1215 if (source->ref_count == 1 &&
1216 source->handle == NULL &&
1217 (source->type == OBJT_DEFAULT ||
1218 source->type == OBJT_SWAP)) {
1219 VM_OBJECT_WUNLOCK(source);
1220 return;
1221 }
1222 VM_OBJECT_WUNLOCK(source);
1223 }
1224
1225 /*
1226 * Allocate a new object with the given length.
1227 */
1228 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1229
1230 /*
1231 * The new object shadows the source object, adding a reference to it.
1232 * Our caller changes his reference to point to the new object,
1233 * removing a reference to the source object. Net result: no change
1234 * of reference count.
1235 *
1236 * Try to optimize the result object's page color when shadowing
1237 * in order to maintain page coloring consistency in the combined
1238 * shadowed object.
1239 */
1240 result->backing_object = source;
1241 /*
1242 * Store the offset into the source object, and fix up the offset into
1243 * the new object.
1244 */
1245 result->backing_object_offset = *offset;
1246 if (source != NULL) {
1247 VM_OBJECT_WLOCK(source);
1248 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1249 source->shadow_count++;
1250#if VM_NRESERVLEVEL > 0
1251 result->flags |= source->flags & OBJ_COLORED;
1252 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1253 ((1 << (VM_NFREEORDER - 1)) - 1);
1254#endif
1255 VM_OBJECT_WUNLOCK(source);
1256 }
1257
1258
1259 /*
1260 * Return the new things
1261 */
1262 *offset = 0;
1263 *object = result;
1264}
1265
1266/*
1267 * vm_object_split:
1268 *
1269 * Split the pages in a map entry into a new object. This affords
1270 * easier removal of unused pages, and keeps object inheritance from
1271 * being a negative impact on memory usage.
1272 */
1273void
1274vm_object_split(vm_map_entry_t entry)
1275{
1276 vm_page_t m, m_next;
1277 vm_object_t orig_object, new_object, source;
1278 vm_pindex_t idx, offidxstart;
1279 vm_size_t size;
1280
1281 orig_object = entry->object.vm_object;
1282 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1283 return;
1284 if (orig_object->ref_count <= 1)
1285 return;
1286 VM_OBJECT_WUNLOCK(orig_object);
1287
1288 offidxstart = OFF_TO_IDX(entry->offset);
1289 size = atop(entry->end - entry->start);
1290
1291 /*
1292 * If swap_pager_copy() is later called, it will convert new_object
1293 * into a swap object.
1294 */
1295 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1296
1297 /*
1298 * At this point, the new object is still private, so the order in
1299 * which the original and new objects are locked does not matter.
1300 */
1301 VM_OBJECT_WLOCK(new_object);
1302 VM_OBJECT_WLOCK(orig_object);
1303 source = orig_object->backing_object;
1304 if (source != NULL) {
1305 VM_OBJECT_WLOCK(source);
1306 if ((source->flags & OBJ_DEAD) != 0) {
1307 VM_OBJECT_WUNLOCK(source);
1308 VM_OBJECT_WUNLOCK(orig_object);
1309 VM_OBJECT_WUNLOCK(new_object);
1310 vm_object_deallocate(new_object);
1311 VM_OBJECT_WLOCK(orig_object);
1312 return;
1313 }
1314 LIST_INSERT_HEAD(&source->shadow_head,
1315 new_object, shadow_list);
1316 source->shadow_count++;
1317 vm_object_reference_locked(source); /* for new_object */
1318 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1319 VM_OBJECT_WUNLOCK(source);
1320 new_object->backing_object_offset =
1321 orig_object->backing_object_offset + entry->offset;
1322 new_object->backing_object = source;
1323 }
1324 if (orig_object->cred != NULL) {
1325 new_object->cred = orig_object->cred;
1326 crhold(orig_object->cred);
1327 new_object->charge = ptoa(size);
1328 KASSERT(orig_object->charge >= ptoa(size),
1329 ("orig_object->charge < 0"));
1330 orig_object->charge -= ptoa(size);
1331 }
1332retry:
1333 m = vm_page_find_least(orig_object, offidxstart);
1334 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1335 m = m_next) {
1336 m_next = TAILQ_NEXT(m, listq);
1337
1338 /*
1339 * We must wait for pending I/O to complete before we can
1340 * rename the page.
1341 *
1342 * We do not have to VM_PROT_NONE the page as mappings should
1343 * not be changed by this operation.
1344 */
1345 if (vm_page_busied(m)) {
1346 VM_OBJECT_WUNLOCK(new_object);
1347 vm_page_lock(m);
1348 VM_OBJECT_WUNLOCK(orig_object);
1349 vm_page_busy_sleep(m, "spltwt");
1350 VM_OBJECT_WLOCK(orig_object);
1351 VM_OBJECT_WLOCK(new_object);
1352 goto retry;
1353 }
1354#if VM_NRESERVLEVEL > 0
1355 /*
1356 * If some of the reservation's allocated pages remain with
1357 * the original object, then transferring the reservation to
1358 * the new object is neither particularly beneficial nor
1359 * particularly harmful as compared to leaving the reservation
1360 * with the original object. If, however, all of the
1361 * reservation's allocated pages are transferred to the new
1362 * object, then transferring the reservation is typically
1363 * beneficial. Determining which of these two cases applies
1364 * would be more costly than unconditionally renaming the
1365 * reservation.
1366 */
1367 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1368#endif
1369 vm_page_lock(m);
1370 vm_page_rename(m, new_object, idx);
1371 vm_page_unlock(m);
1372 /* page automatically made dirty by rename and cache handled */
1373 if (orig_object->type == OBJT_SWAP)
1374 vm_page_xbusy(m);
1375 }
1376 if (orig_object->type == OBJT_SWAP) {
1377 /*
1378 * swap_pager_copy() can sleep, in which case the orig_object's
1379 * and new_object's locks are released and reacquired.
1380 */
1381 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1382 TAILQ_FOREACH(m, &new_object->memq, listq)
1383 vm_page_xunbusy(m);
1384
1385 /*
1386 * Transfer any cached pages from orig_object to new_object.
1387 * If swap_pager_copy() found swapped out pages within the
1388 * specified range of orig_object, then it changed
1389 * new_object's type to OBJT_SWAP when it transferred those
1390 * pages to new_object. Otherwise, new_object's type
1391 * should still be OBJT_DEFAULT and orig_object should not
1392 * contain any cached pages within the specified range.
1393 */
1394 if (__predict_false(!vm_object_cache_is_empty(orig_object)))
1395 vm_page_cache_transfer(orig_object, offidxstart,
1396 new_object);
1397 }
1398 VM_OBJECT_WUNLOCK(orig_object);
1399 VM_OBJECT_WUNLOCK(new_object);
1400 entry->object.vm_object = new_object;
1401 entry->offset = 0LL;
1402 vm_object_deallocate(orig_object);
1403 VM_OBJECT_WLOCK(new_object);
1404}
1405
1406#define OBSC_TEST_ALL_SHADOWED 0x0001
1407#define OBSC_COLLAPSE_NOWAIT 0x0002
1408#define OBSC_COLLAPSE_WAIT 0x0004
1409
1410static int
1411vm_object_backing_scan(vm_object_t object, int op)
1412{
1413 int r = 1;
1414 vm_page_t p;
1415 vm_object_t backing_object;
1416 vm_pindex_t backing_offset_index;
1417
1418 VM_OBJECT_ASSERT_WLOCKED(object);
1419 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1420
1421 backing_object = object->backing_object;
1422 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1423
1424 /*
1425 * Initial conditions
1426 */
1427 if (op & OBSC_TEST_ALL_SHADOWED) {
1428 /*
1429 * We do not want to have to test for the existence of cache
1430 * or swap pages in the backing object. XXX but with the
1431 * new swapper this would be pretty easy to do.
1432 *
1433 * XXX what about anonymous MAP_SHARED memory that hasn't
1434 * been ZFOD faulted yet? If we do not test for this, the
1435 * shadow test may succeed! XXX
1436 */
1437 if (backing_object->type != OBJT_DEFAULT) {
1438 return (0);
1439 }
1440 }
1441 if (op & OBSC_COLLAPSE_WAIT) {
1442 vm_object_set_flag(backing_object, OBJ_DEAD);
1443 }
1444
1445 /*
1446 * Our scan
1447 */
1448 p = TAILQ_FIRST(&backing_object->memq);
1449 while (p) {
1450 vm_page_t next = TAILQ_NEXT(p, listq);
1451 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1452
1453 if (op & OBSC_TEST_ALL_SHADOWED) {
1454 vm_page_t pp;
1455
1456 /*
1457 * Ignore pages outside the parent object's range
1458 * and outside the parent object's mapping of the
1459 * backing object.
1460 *
1461 * note that we do not busy the backing object's
1462 * page.
1463 */
1464 if (
1465 p->pindex < backing_offset_index ||
1466 new_pindex >= object->size
1467 ) {
1468 p = next;
1469 continue;
1470 }
1471
1472 /*
1473 * See if the parent has the page or if the parent's
1474 * object pager has the page. If the parent has the
1475 * page but the page is not valid, the parent's
1476 * object pager must have the page.
1477 *
1478 * If this fails, the parent does not completely shadow
1479 * the object and we might as well give up now.
1480 */
1481
1482 pp = vm_page_lookup(object, new_pindex);
1483 if (
1484 (pp == NULL || pp->valid == 0) &&
1485 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1486 ) {
1487 r = 0;
1488 break;
1489 }
1490 }
1491
1492 /*
1493 * Check for busy page
1494 */
1495 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1496 vm_page_t pp;
1497
1498 if (op & OBSC_COLLAPSE_NOWAIT) {
1499 if (!p->valid || vm_page_busied(p)) {
1500 p = next;
1501 continue;
1502 }
1503 } else if (op & OBSC_COLLAPSE_WAIT) {
1504 if (vm_page_busied(p)) {
1505 VM_OBJECT_WUNLOCK(object);
1506 vm_page_lock(p);
1507 VM_OBJECT_WUNLOCK(backing_object);
1508 vm_page_busy_sleep(p, "vmocol");
1509 VM_OBJECT_WLOCK(object);
1510 VM_OBJECT_WLOCK(backing_object);
1511 /*
1512 * If we slept, anything could have
1513 * happened. Since the object is
1514 * marked dead, the backing offset
1515 * should not have changed so we
1516 * just restart our scan.
1517 */
1518 p = TAILQ_FIRST(&backing_object->memq);
1519 continue;
1520 }
1521 }
1522
1523 KASSERT(
1524 p->object == backing_object,
1525 ("vm_object_backing_scan: object mismatch")
1526 );
1527
1528 /*
1529 * Destroy any associated swap
1530 */
1531 if (backing_object->type == OBJT_SWAP) {
1532 swap_pager_freespace(
1533 backing_object,
1534 p->pindex,
1535 1
1536 );
1537 }
1538
1539 if (
1540 p->pindex < backing_offset_index ||
1541 new_pindex >= object->size
1542 ) {
1543 /*
1544 * Page is out of the parent object's range, we
1545 * can simply destroy it.
1546 */
1547 vm_page_lock(p);
1548 KASSERT(!pmap_page_is_mapped(p),
1549 ("freeing mapped page %p", p));
1550 if (p->wire_count == 0)
1551 vm_page_free(p);
1552 else
1553 vm_page_remove(p);
1554 vm_page_unlock(p);
1555 p = next;
1556 continue;
1557 }
1558
1559 pp = vm_page_lookup(object, new_pindex);
1560 if (
1561 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1562 (pp != NULL && pp->valid == 0)
1563 ) {
1564 /*
1565 * The page in the parent is not (yet) valid.
1566 * We don't know anything about the state of
1567 * the original page. It might be mapped,
1568 * so we must avoid the next if here.
1569 *
1570 * This is due to a race in vm_fault() where
1571 * we must unbusy the original (backing_obj)
1572 * page before we can (re)lock the parent.
1573 * Hence we can get here.
1574 */
1575 p = next;
1576 continue;
1577 }
1578 if (
1579 pp != NULL ||
1580 vm_pager_has_page(object, new_pindex, NULL, NULL)
1581 ) {
1582 /*
1583 * page already exists in parent OR swap exists
1584 * for this location in the parent. Destroy
1585 * the original page from the backing object.
1586 *
1587 * Leave the parent's page alone
1588 */
1589 vm_page_lock(p);
1590 KASSERT(!pmap_page_is_mapped(p),
1591 ("freeing mapped page %p", p));
1592 if (p->wire_count == 0)
1593 vm_page_free(p);
1594 else
1595 vm_page_remove(p);
1596 vm_page_unlock(p);
1597 p = next;
1598 continue;
1599 }
1600
1601#if VM_NRESERVLEVEL > 0
1602 /*
1603 * Rename the reservation.
1604 */
1605 vm_reserv_rename(p, object, backing_object,
1606 backing_offset_index);
1607#endif
1608
1609 /*
1610 * Page does not exist in parent, rename the
1611 * page from the backing object to the main object.
1612 *
1613 * If the page was mapped to a process, it can remain
1614 * mapped through the rename.
1615 */
1616 vm_page_lock(p);
1617 vm_page_rename(p, object, new_pindex);
1618 vm_page_unlock(p);
1619 /* page automatically made dirty by rename */
1620 }
1621 p = next;
1622 }
1623 return (r);
1624}
1625
1626
1627/*
1628 * this version of collapse allows the operation to occur earlier and
1629 * when paging_in_progress is true for an object... This is not a complete
1630 * operation, but should plug 99.9% of the rest of the leaks.
1631 */
1632static void
1633vm_object_qcollapse(vm_object_t object)
1634{
1635 vm_object_t backing_object = object->backing_object;
1636
1637 VM_OBJECT_ASSERT_WLOCKED(object);
1638 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1639
1640 if (backing_object->ref_count != 1)
1641 return;
1642
1643 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1644}
1645
1646/*
1647 * vm_object_collapse:
1648 *
1649 * Collapse an object with the object backing it.
1650 * Pages in the backing object are moved into the
1651 * parent, and the backing object is deallocated.
1652 */
1653void
1654vm_object_collapse(vm_object_t object)
1655{
1656 VM_OBJECT_ASSERT_WLOCKED(object);
1657
1658 while (TRUE) {
1659 vm_object_t backing_object;
1660
1661 /*
1662 * Verify that the conditions are right for collapse:
1663 *
1664 * The object exists and the backing object exists.
1665 */
1666 if ((backing_object = object->backing_object) == NULL)
1667 break;
1668
1669 /*
1670 * we check the backing object first, because it is most likely
1671 * not collapsable.
1672 */
1673 VM_OBJECT_WLOCK(backing_object);
1674 if (backing_object->handle != NULL ||
1675 (backing_object->type != OBJT_DEFAULT &&
1676 backing_object->type != OBJT_SWAP) ||
1677 (backing_object->flags & OBJ_DEAD) ||
1678 object->handle != NULL ||
1679 (object->type != OBJT_DEFAULT &&
1680 object->type != OBJT_SWAP) ||
1681 (object->flags & OBJ_DEAD)) {
1682 VM_OBJECT_WUNLOCK(backing_object);
1683 break;
1684 }
1685
1686 if (
1687 object->paging_in_progress != 0 ||
1688 backing_object->paging_in_progress != 0
1689 ) {
1690 vm_object_qcollapse(object);
1691 VM_OBJECT_WUNLOCK(backing_object);
1692 break;
1693 }
1694 /*
1695 * We know that we can either collapse the backing object (if
1696 * the parent is the only reference to it) or (perhaps) have
1697 * the parent bypass the object if the parent happens to shadow
1698 * all the resident pages in the entire backing object.
1699 *
1700 * This is ignoring pager-backed pages such as swap pages.
1701 * vm_object_backing_scan fails the shadowing test in this
1702 * case.
1703 */
1704 if (backing_object->ref_count == 1) {
1705 /*
1706 * If there is exactly one reference to the backing
1707 * object, we can collapse it into the parent.
1708 */
1709 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1710
1711#if VM_NRESERVLEVEL > 0
1712 /*
1713 * Break any reservations from backing_object.
1714 */
1715 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1716 vm_reserv_break_all(backing_object);
1717#endif
1718
1719 /*
1720 * Move the pager from backing_object to object.
1721 */
1722 if (backing_object->type == OBJT_SWAP) {
1723 /*
1724 * swap_pager_copy() can sleep, in which case
1725 * the backing_object's and object's locks are
1726 * released and reacquired.
1727 * Since swap_pager_copy() is being asked to
1728 * destroy the source, it will change the
1729 * backing_object's type to OBJT_DEFAULT.
1730 */
1731 swap_pager_copy(
1732 backing_object,
1733 object,
1734 OFF_TO_IDX(object->backing_object_offset), TRUE);
1735
1736 /*
1737 * Free any cached pages from backing_object.
1738 */
1739 if (__predict_false(
1740 !vm_object_cache_is_empty(backing_object)))
1741 vm_page_cache_free(backing_object, 0, 0);
1742 }
1743 /*
1744 * Object now shadows whatever backing_object did.
1745 * Note that the reference to
1746 * backing_object->backing_object moves from within
1747 * backing_object to within object.
1748 */
1749 LIST_REMOVE(object, shadow_list);
1750 backing_object->shadow_count--;
1751 if (backing_object->backing_object) {
1752 VM_OBJECT_WLOCK(backing_object->backing_object);
1753 LIST_REMOVE(backing_object, shadow_list);
1754 LIST_INSERT_HEAD(
1755 &backing_object->backing_object->shadow_head,
1756 object, shadow_list);
1757 /*
1758 * The shadow_count has not changed.
1759 */
1760 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1761 }
1762 object->backing_object = backing_object->backing_object;
1763 object->backing_object_offset +=
1764 backing_object->backing_object_offset;
1765
1766 /*
1767 * Discard backing_object.
1768 *
1769 * Since the backing object has no pages, no pager left,
1770 * and no object references within it, all that is
1771 * necessary is to dispose of it.
1772 */
1773 KASSERT(backing_object->ref_count == 1, (
1774"backing_object %p was somehow re-referenced during collapse!",
1775 backing_object));
1776 VM_OBJECT_WUNLOCK(backing_object);
1777 vm_object_destroy(backing_object);
1778
1779 object_collapses++;
1780 } else {
1781 vm_object_t new_backing_object;
1782
1783 /*
1784 * If we do not entirely shadow the backing object,
1785 * there is nothing we can do so we give up.
1786 */
1787 if (object->resident_page_count != object->size &&
1788 vm_object_backing_scan(object,
1789 OBSC_TEST_ALL_SHADOWED) == 0) {
1790 VM_OBJECT_WUNLOCK(backing_object);
1791 break;
1792 }
1793
1794 /*
1795 * Make the parent shadow the next object in the
1796 * chain. Deallocating backing_object will not remove
1797 * it, since its reference count is at least 2.
1798 */
1799 LIST_REMOVE(object, shadow_list);
1800 backing_object->shadow_count--;
1801
1802 new_backing_object = backing_object->backing_object;
1803 if ((object->backing_object = new_backing_object) != NULL) {
1804 VM_OBJECT_WLOCK(new_backing_object);
1805 LIST_INSERT_HEAD(
1806 &new_backing_object->shadow_head,
1807 object,
1808 shadow_list
1809 );
1810 new_backing_object->shadow_count++;
1811 vm_object_reference_locked(new_backing_object);
1812 VM_OBJECT_WUNLOCK(new_backing_object);
1813 object->backing_object_offset +=
1814 backing_object->backing_object_offset;
1815 }
1816
1817 /*
1818 * Drop the reference count on backing_object. Since
1819 * its ref_count was at least 2, it will not vanish.
1820 */
1821 backing_object->ref_count--;
1822 VM_OBJECT_WUNLOCK(backing_object);
1823 object_bypasses++;
1824 }
1825
1826 /*
1827 * Try again with this object's new backing object.
1828 */
1829 }
1830}
1831
1832/*
1833 * vm_object_page_remove:
1834 *
1835 * For the given object, either frees or invalidates each of the
1836 * specified pages. In general, a page is freed. However, if a page is
1837 * wired for any reason other than the existence of a managed, wired
1838 * mapping, then it may be invalidated but not removed from the object.
1839 * Pages are specified by the given range ["start", "end") and the option
1840 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1841 * extends from "start" to the end of the object. If the option
1842 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1843 * specified range are affected. If the option OBJPR_NOTMAPPED is
1844 * specified, then the pages within the specified range must have no
1845 * mappings. Otherwise, if this option is not specified, any mappings to
1846 * the specified pages are removed before the pages are freed or
1847 * invalidated.
1848 *
1849 * In general, this operation should only be performed on objects that
1850 * contain managed pages. There are, however, two exceptions. First, it
1851 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1852 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1853 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1854 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1855 *
1856 * The object must be locked.
1857 */
1858void
1859vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1860 int options)
1861{
1862 vm_page_t p, next;
1863 int wirings;
1864
1865 VM_OBJECT_ASSERT_WLOCKED(object);
1866 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1867 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1868 ("vm_object_page_remove: illegal options for object %p", object));
1869 if (object->resident_page_count == 0)
1870 goto skipmemq;
1871 vm_object_pip_add(object, 1);
1872again:
1873 p = vm_page_find_least(object, start);
1874
1875 /*
1876 * Here, the variable "p" is either (1) the page with the least pindex
1877 * greater than or equal to the parameter "start" or (2) NULL.
1878 */
1879 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1880 next = TAILQ_NEXT(p, listq);
1881
1882 /*
1883 * If the page is wired for any reason besides the existence
1884 * of managed, wired mappings, then it cannot be freed. For
1885 * example, fictitious pages, which represent device memory,
1886 * are inherently wired and cannot be freed. They can,
1887 * however, be invalidated if the option OBJPR_CLEANONLY is
1888 * not specified.
1889 */
1890 vm_page_lock(p);
1891 if ((wirings = p->wire_count) != 0 &&
1892 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1893 if ((options & (OBJPR_NOTWIRED | OBJPR_NOTMAPPED)) ==
1894 0) {
1895 pmap_remove_all(p);
1896 /* Account for removal of wired mappings. */
1897 if (wirings != 0)
1898 p->wire_count -= wirings;
1899 }
1900 if ((options & OBJPR_CLEANONLY) == 0) {
1901 p->valid = 0;
1902 vm_page_undirty(p);
1903 }
1904 goto next;
1905 }
1906 if (vm_page_busied(p)) {
1907 VM_OBJECT_WUNLOCK(object);
1908 vm_page_busy_sleep(p, "vmopar");
1909 VM_OBJECT_WLOCK(object);
1910 goto again;
1911 }
1912 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1913 ("vm_object_page_remove: page %p is fictitious", p));
1914 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1915 if ((options & OBJPR_NOTMAPPED) == 0)
1916 pmap_remove_write(p);
1917 if (p->dirty)
1918 goto next;
1919 }
1920 if ((options & OBJPR_NOTMAPPED) == 0) {
1921 if ((options & OBJPR_NOTWIRED) != 0 && wirings != 0)
1922 goto next;
1923 pmap_remove_all(p);
1924 /* Account for removal of wired mappings. */
1925 if (wirings != 0) {
1926 KASSERT(p->wire_count == wirings,
1927 ("inconsistent wire count %d %d %p",
1928 p->wire_count, wirings, p));
1929 p->wire_count = 0;
1930 atomic_subtract_int(&cnt.v_wire_count, 1);
1931 }
1932 }
1933 vm_page_free(p);
1934next:
1935 vm_page_unlock(p);
1936 }
1937 vm_object_pip_wakeup(object);
1938skipmemq:
1939 if (__predict_false(!vm_object_cache_is_empty(object)))
1940 vm_page_cache_free(object, start, end);
1941}
1942
1943/*
1944 * vm_object_page_cache:
1945 *
1946 * For the given object, attempt to move the specified clean
1947 * pages to the cache queue. If a page is wired for any reason,
1948 * then it will not be changed. Pages are specified by the given
1949 * range ["start", "end"). As a special case, if "end" is zero,
1950 * then the range extends from "start" to the end of the object.
1951 * Any mappings to the specified pages are removed before the
1952 * pages are moved to the cache queue.
1953 *
1954 * This operation should only be performed on objects that
1955 * contain non-fictitious, managed pages.
1956 *
1957 * The object must be locked.
1958 */
1959void
1960vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1961{
1962 struct mtx *mtx, *new_mtx;
1963 vm_page_t p, next;
1964
1965 VM_OBJECT_ASSERT_WLOCKED(object);
1966 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1967 ("vm_object_page_cache: illegal object %p", object));
1968 if (object->resident_page_count == 0)
1969 return;
1970 p = vm_page_find_least(object, start);
1971
1972 /*
1973 * Here, the variable "p" is either (1) the page with the least pindex
1974 * greater than or equal to the parameter "start" or (2) NULL.
1975 */
1976 mtx = NULL;
1977 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1978 next = TAILQ_NEXT(p, listq);
1979
1980 /*
1981 * Avoid releasing and reacquiring the same page lock.
1982 */
1983 new_mtx = vm_page_lockptr(p);
1984 if (mtx != new_mtx) {
1985 if (mtx != NULL)
1986 mtx_unlock(mtx);
1987 mtx = new_mtx;
1988 mtx_lock(mtx);
1989 }
1990 vm_page_try_to_cache(p);
1991 }
1992 if (mtx != NULL)
1993 mtx_unlock(mtx);
1994}
1995
1996/*
1997 * Populate the specified range of the object with valid pages. Returns
1998 * TRUE if the range is successfully populated and FALSE otherwise.
1999 *
2000 * Note: This function should be optimized to pass a larger array of
2001 * pages to vm_pager_get_pages() before it is applied to a non-
2002 * OBJT_DEVICE object.
2003 *
2004 * The object must be locked.
2005 */
2006boolean_t
2007vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2008{
2009 vm_page_t m, ma[1];
2010 vm_pindex_t pindex;
2011 int rv;
2012
2013 VM_OBJECT_ASSERT_WLOCKED(object);
2014 for (pindex = start; pindex < end; pindex++) {
2015 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
2016 VM_ALLOC_RETRY);
2017 if (m->valid != VM_PAGE_BITS_ALL) {
2018 ma[0] = m;
2019 rv = vm_pager_get_pages(object, ma, 1, 0);
2020 m = vm_page_lookup(object, pindex);
2021 if (m == NULL)
2022 break;
2023 if (rv != VM_PAGER_OK) {
2024 vm_page_lock(m);
2025 vm_page_free(m);
2026 vm_page_unlock(m);
2027 break;
2028 }
2029 }
2030 /*
2031 * Keep "m" busy because a subsequent iteration may unlock
2032 * the object.
2033 */
2034 }
2035 if (pindex > start) {
2036 m = vm_page_lookup(object, start);
2037 while (m != NULL && m->pindex < pindex) {
2038 vm_page_xunbusy(m);
2039 m = TAILQ_NEXT(m, listq);
2040 }
2041 }
2042 return (pindex == end);
2043}
2044
2045/*
2046 * Routine: vm_object_coalesce
2047 * Function: Coalesces two objects backing up adjoining
2048 * regions of memory into a single object.
2049 *
2050 * returns TRUE if objects were combined.
2051 *
2052 * NOTE: Only works at the moment if the second object is NULL -
2053 * if it's not, which object do we lock first?
2054 *
2055 * Parameters:
2056 * prev_object First object to coalesce
2057 * prev_offset Offset into prev_object
2058 * prev_size Size of reference to prev_object
2059 * next_size Size of reference to the second object
2060 * reserved Indicator that extension region has
2061 * swap accounted for
2062 *
2063 * Conditions:
2064 * The object must *not* be locked.
2065 */
2066boolean_t
2067vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2068 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2069{
2070 vm_pindex_t next_pindex;
2071
2072 if (prev_object == NULL)
2073 return (TRUE);
2074 VM_OBJECT_WLOCK(prev_object);
2075 if (prev_object->type != OBJT_DEFAULT &&
2076 prev_object->type != OBJT_SWAP) {
2077 VM_OBJECT_WUNLOCK(prev_object);
2078 return (FALSE);
2079 }
2080
2081 /*
2082 * Try to collapse the object first
2083 */
2084 vm_object_collapse(prev_object);
2085
2086 /*
2087 * Can't coalesce if: . more than one reference . paged out . shadows
2088 * another object . has a copy elsewhere (any of which mean that the
2089 * pages not mapped to prev_entry may be in use anyway)
2090 */
2091 if (prev_object->backing_object != NULL) {
2092 VM_OBJECT_WUNLOCK(prev_object);
2093 return (FALSE);
2094 }
2095
2096 prev_size >>= PAGE_SHIFT;
2097 next_size >>= PAGE_SHIFT;
2098 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2099
2100 if ((prev_object->ref_count > 1) &&
2101 (prev_object->size != next_pindex)) {
2102 VM_OBJECT_WUNLOCK(prev_object);
2103 return (FALSE);
2104 }
2105
2106 /*
2107 * Account for the charge.
2108 */
2109 if (prev_object->cred != NULL) {
2110
2111 /*
2112 * If prev_object was charged, then this mapping,
2113 * althought not charged now, may become writable
2114 * later. Non-NULL cred in the object would prevent
2115 * swap reservation during enabling of the write
2116 * access, so reserve swap now. Failed reservation
2117 * cause allocation of the separate object for the map
2118 * entry, and swap reservation for this entry is
2119 * managed in appropriate time.
2120 */
2121 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2122 prev_object->cred)) {
2123 return (FALSE);
2124 }
2125 prev_object->charge += ptoa(next_size);
2126 }
2127
2128 /*
2129 * Remove any pages that may still be in the object from a previous
2130 * deallocation.
2131 */
2132 if (next_pindex < prev_object->size) {
2133 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2134 next_size, 0);
2135 if (prev_object->type == OBJT_SWAP)
2136 swap_pager_freespace(prev_object,
2137 next_pindex, next_size);
2138#if 0
2139 if (prev_object->cred != NULL) {
2140 KASSERT(prev_object->charge >=
2141 ptoa(prev_object->size - next_pindex),
2142 ("object %p overcharged 1 %jx %jx", prev_object,
2143 (uintmax_t)next_pindex, (uintmax_t)next_size));
2144 prev_object->charge -= ptoa(prev_object->size -
2145 next_pindex);
2146 }
2147#endif
2148 }
2149
2150 /*
2151 * Extend the object if necessary.
2152 */
2153 if (next_pindex + next_size > prev_object->size)
2154 prev_object->size = next_pindex + next_size;
2155
2156 VM_OBJECT_WUNLOCK(prev_object);
2157 return (TRUE);
2158}
2159
2160void
2161vm_object_set_writeable_dirty(vm_object_t object)
2162{
2163
2164 VM_OBJECT_ASSERT_WLOCKED(object);
2165 if (object->type != OBJT_VNODE)
2166 return;
2167 object->generation++;
2168 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2169 return;
2170 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2171}
2172
2173#include "opt_ddb.h"
2174#ifdef DDB
2175#include <sys/kernel.h>
2176
2177#include <sys/cons.h>
2178
2179#include <ddb/ddb.h>
2180
2181static int
2182_vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2183{
2184 vm_map_t tmpm;
2185 vm_map_entry_t tmpe;
2186 vm_object_t obj;
2187 int entcount;
2188
2189 if (map == 0)
2190 return 0;
2191
2192 if (entry == 0) {
2193 tmpe = map->header.next;
2194 entcount = map->nentries;
2195 while (entcount-- && (tmpe != &map->header)) {
2196 if (_vm_object_in_map(map, object, tmpe)) {
2197 return 1;
2198 }
2199 tmpe = tmpe->next;
2200 }
2201 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2202 tmpm = entry->object.sub_map;
2203 tmpe = tmpm->header.next;
2204 entcount = tmpm->nentries;
2205 while (entcount-- && tmpe != &tmpm->header) {
2206 if (_vm_object_in_map(tmpm, object, tmpe)) {
2207 return 1;
2208 }
2209 tmpe = tmpe->next;
2210 }
2211 } else if ((obj = entry->object.vm_object) != NULL) {
2212 for (; obj; obj = obj->backing_object)
2213 if (obj == object) {
2214 return 1;
2215 }
2216 }
2217 return 0;
2218}
2219
2220static int
2221vm_object_in_map(vm_object_t object)
2222{
2223 struct proc *p;
2224
2225 /* sx_slock(&allproc_lock); */
2226 FOREACH_PROC_IN_SYSTEM(p) {
2227 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2228 continue;
2229 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2230 /* sx_sunlock(&allproc_lock); */
2231 return 1;
2232 }
2233 }
2234 /* sx_sunlock(&allproc_lock); */
2235 if (_vm_object_in_map(kernel_map, object, 0))
2236 return 1;
2237 return 0;
2238}
2239
2240DB_SHOW_COMMAND(vmochk, vm_object_check)
2241{
2242 vm_object_t object;
2243
2244 /*
2245 * make sure that internal objs are in a map somewhere
2246 * and none have zero ref counts.
2247 */
2248 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2249 if (object->handle == NULL &&
2250 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2251 if (object->ref_count == 0) {
2252 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2253 (long)object->size);
2254 }
2255 if (!vm_object_in_map(object)) {
2256 db_printf(
2257 "vmochk: internal obj is not in a map: "
2258 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2259 object->ref_count, (u_long)object->size,
2260 (u_long)object->size,
2261 (void *)object->backing_object);
2262 }
2263 }
2264 }
2265}
2266
2267/*
2268 * vm_object_print: [ debug ]
2269 */
2270DB_SHOW_COMMAND(object, vm_object_print_static)
2271{
2272 /* XXX convert args. */
2273 vm_object_t object = (vm_object_t)addr;
2274 boolean_t full = have_addr;
2275
2276 vm_page_t p;
2277
2278 /* XXX count is an (unused) arg. Avoid shadowing it. */
2279#define count was_count
2280
2281 int count;
2282
2283 if (object == NULL)
2284 return;
2285
2286 db_iprintf(
2287 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2288 object, (int)object->type, (uintmax_t)object->size,
2289 object->resident_page_count, object->ref_count, object->flags,
2290 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2291 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2292 object->shadow_count,
2293 object->backing_object ? object->backing_object->ref_count : 0,
2294 object->backing_object, (uintmax_t)object->backing_object_offset);
2295
2296 if (!full)
2297 return;
2298
2299 db_indent += 2;
2300 count = 0;
2301 TAILQ_FOREACH(p, &object->memq, listq) {
2302 if (count == 0)
2303 db_iprintf("memory:=");
2304 else if (count == 6) {
2305 db_printf("\n");
2306 db_iprintf(" ...");
2307 count = 0;
2308 } else
2309 db_printf(",");
2310 count++;
2311
2312 db_printf("(off=0x%jx,page=0x%jx)",
2313 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2314 }
2315 if (count != 0)
2316 db_printf("\n");
2317 db_indent -= 2;
2318}
2319
2320/* XXX. */
2321#undef count
2322
2323/* XXX need this non-static entry for calling from vm_map_print. */
2324void
2325vm_object_print(
2326 /* db_expr_t */ long addr,
2327 boolean_t have_addr,
2328 /* db_expr_t */ long count,
2329 char *modif)
2330{
2331 vm_object_print_static(addr, have_addr, count, modif);
2332}
2333
2334DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2335{
2336 vm_object_t object;
2337 vm_pindex_t fidx;
2338 vm_paddr_t pa;
2339 vm_page_t m, prev_m;
2340 int rcount, nl, c;
2341
2342 nl = 0;
2343 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2344 db_printf("new object: %p\n", (void *)object);
2345 if (nl > 18) {
2346 c = cngetc();
2347 if (c != ' ')
2348 return;
2349 nl = 0;
2350 }
2351 nl++;
2352 rcount = 0;
2353 fidx = 0;
2354 pa = -1;
2355 TAILQ_FOREACH(m, &object->memq, listq) {
2356 if (m->pindex > 128)
2357 break;
2358 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2359 prev_m->pindex + 1 != m->pindex) {
2360 if (rcount) {
2361 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2362 (long)fidx, rcount, (long)pa);
2363 if (nl > 18) {
2364 c = cngetc();
2365 if (c != ' ')
2366 return;
2367 nl = 0;
2368 }
2369 nl++;
2370 rcount = 0;
2371 }
2372 }
2373 if (rcount &&
2374 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2375 ++rcount;
2376 continue;
2377 }
2378 if (rcount) {
2379 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2380 (long)fidx, rcount, (long)pa);
2381 if (nl > 18) {
2382 c = cngetc();
2383 if (c != ' ')
2384 return;
2385 nl = 0;
2386 }
2387 nl++;
2388 }
2389 fidx = m->pindex;
2390 pa = VM_PAGE_TO_PHYS(m);
2391 rcount = 1;
2392 }
2393 if (rcount) {
2394 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2395 (long)fidx, rcount, (long)pa);
2396 if (nl > 18) {
2397 c = cngetc();
2398 if (c != ' ')
2399 return;
2400 nl = 0;
2401 }
2402 nl++;
2403 }
2404 }
2405}
2406#endif /* DDB */
2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 *
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 4. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * SUCH DAMAGE.
31 *
32 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94
33 *
34 *
35 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
36 * All rights reserved.
37 *
38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
39 *
40 * Permission to use, copy, modify and distribute this software and
41 * its documentation is hereby granted, provided that both the copyright
42 * notice and this permission notice appear in all copies of the
43 * software, derivative works or modified versions, and any portions
44 * thereof, and that both notices appear in supporting documentation.
45 *
46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
49 *
50 * Carnegie Mellon requests users of this software to return to
51 *
52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
53 * School of Computer Science
54 * Carnegie Mellon University
55 * Pittsburgh PA 15213-3890
56 *
57 * any improvements or extensions that they make and grant Carnegie the
58 * rights to redistribute these changes.
59 */
60
61/*
62 * Virtual memory object module.
63 */
64
65#include <sys/cdefs.h>
66__FBSDID("$FreeBSD: head/sys/vm/vm_object.c 254138 2013-08-09 11:11:11Z attilio $");
67
68#include "opt_vm.h"
69
70#include <sys/param.h>
71#include <sys/systm.h>
72#include <sys/lock.h>
73#include <sys/mman.h>
74#include <sys/mount.h>
75#include <sys/kernel.h>
76#include <sys/sysctl.h>
77#include <sys/mutex.h>
78#include <sys/proc.h> /* for curproc, pageproc */
79#include <sys/socket.h>
80#include <sys/resourcevar.h>
81#include <sys/rwlock.h>
82#include <sys/vnode.h>
83#include <sys/vmmeter.h>
84#include <sys/sx.h>
85
86#include <vm/vm.h>
87#include <vm/vm_param.h>
88#include <vm/pmap.h>
89#include <vm/vm_map.h>
90#include <vm/vm_object.h>
91#include <vm/vm_page.h>
92#include <vm/vm_pageout.h>
93#include <vm/vm_pager.h>
94#include <vm/swap_pager.h>
95#include <vm/vm_kern.h>
96#include <vm/vm_extern.h>
97#include <vm/vm_radix.h>
98#include <vm/vm_reserv.h>
99#include <vm/uma.h>
100
101static int old_msync;
102SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
103 "Use old (insecure) msync behavior");
104
105static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
106 int pagerflags, int flags, boolean_t *clearobjflags,
107 boolean_t *eio);
108static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
109 boolean_t *clearobjflags);
110static void vm_object_qcollapse(vm_object_t object);
111static void vm_object_vndeallocate(vm_object_t object);
112
113/*
114 * Virtual memory objects maintain the actual data
115 * associated with allocated virtual memory. A given
116 * page of memory exists within exactly one object.
117 *
118 * An object is only deallocated when all "references"
119 * are given up. Only one "reference" to a given
120 * region of an object should be writeable.
121 *
122 * Associated with each object is a list of all resident
123 * memory pages belonging to that object; this list is
124 * maintained by the "vm_page" module, and locked by the object's
125 * lock.
126 *
127 * Each object also records a "pager" routine which is
128 * used to retrieve (and store) pages to the proper backing
129 * storage. In addition, objects may be backed by other
130 * objects from which they were virtual-copied.
131 *
132 * The only items within the object structure which are
133 * modified after time of creation are:
134 * reference count locked by object's lock
135 * pager routine locked by object's lock
136 *
137 */
138
139struct object_q vm_object_list;
140struct mtx vm_object_list_mtx; /* lock for object list and count */
141
142struct vm_object kernel_object_store;
143struct vm_object kmem_object_store;
144
145static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
146 "VM object stats");
147
148static long object_collapses;
149SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
150 &object_collapses, 0, "VM object collapses");
151
152static long object_bypasses;
153SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
154 &object_bypasses, 0, "VM object bypasses");
155
156static uma_zone_t obj_zone;
157
158static int vm_object_zinit(void *mem, int size, int flags);
159
160#ifdef INVARIANTS
161static void vm_object_zdtor(void *mem, int size, void *arg);
162
163static void
164vm_object_zdtor(void *mem, int size, void *arg)
165{
166 vm_object_t object;
167
168 object = (vm_object_t)mem;
169 KASSERT(TAILQ_EMPTY(&object->memq),
170 ("object %p has resident pages in its memq", object));
171 KASSERT(vm_radix_is_empty(&object->rtree),
172 ("object %p has resident pages in its trie", object));
173#if VM_NRESERVLEVEL > 0
174 KASSERT(LIST_EMPTY(&object->rvq),
175 ("object %p has reservations",
176 object));
177#endif
178 KASSERT(vm_object_cache_is_empty(object),
179 ("object %p has cached pages",
180 object));
181 KASSERT(object->paging_in_progress == 0,
182 ("object %p paging_in_progress = %d",
183 object, object->paging_in_progress));
184 KASSERT(object->resident_page_count == 0,
185 ("object %p resident_page_count = %d",
186 object, object->resident_page_count));
187 KASSERT(object->shadow_count == 0,
188 ("object %p shadow_count = %d",
189 object, object->shadow_count));
190}
191#endif
192
193static int
194vm_object_zinit(void *mem, int size, int flags)
195{
196 vm_object_t object;
197
198 object = (vm_object_t)mem;
199 bzero(&object->lock, sizeof(object->lock));
200 rw_init_flags(&object->lock, "vm object", RW_DUPOK);
201
202 /* These are true for any object that has been freed */
203 object->rtree.rt_root = 0;
204 object->paging_in_progress = 0;
205 object->resident_page_count = 0;
206 object->shadow_count = 0;
207 object->cache.rt_root = 0;
208 return (0);
209}
210
211static void
212_vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
213{
214
215 TAILQ_INIT(&object->memq);
216 LIST_INIT(&object->shadow_head);
217
218 object->type = type;
219 switch (type) {
220 case OBJT_DEAD:
221 panic("_vm_object_allocate: can't create OBJT_DEAD");
222 case OBJT_DEFAULT:
223 case OBJT_SWAP:
224 object->flags = OBJ_ONEMAPPING;
225 break;
226 case OBJT_DEVICE:
227 case OBJT_SG:
228 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
229 break;
230 case OBJT_MGTDEVICE:
231 object->flags = OBJ_FICTITIOUS;
232 break;
233 case OBJT_PHYS:
234 object->flags = OBJ_UNMANAGED;
235 break;
236 case OBJT_VNODE:
237 object->flags = 0;
238 break;
239 default:
240 panic("_vm_object_allocate: type %d is undefined", type);
241 }
242 object->size = size;
243 object->generation = 1;
244 object->ref_count = 1;
245 object->memattr = VM_MEMATTR_DEFAULT;
246 object->cred = NULL;
247 object->charge = 0;
248 object->handle = NULL;
249 object->backing_object = NULL;
250 object->backing_object_offset = (vm_ooffset_t) 0;
251#if VM_NRESERVLEVEL > 0
252 LIST_INIT(&object->rvq);
253#endif
254
255 mtx_lock(&vm_object_list_mtx);
256 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
257 mtx_unlock(&vm_object_list_mtx);
258}
259
260/*
261 * vm_object_init:
262 *
263 * Initialize the VM objects module.
264 */
265void
266vm_object_init(void)
267{
268 TAILQ_INIT(&vm_object_list);
269 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
270
271 rw_init(&kernel_object->lock, "kernel vm object");
272 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
273 kernel_object);
274#if VM_NRESERVLEVEL > 0
275 kernel_object->flags |= OBJ_COLORED;
276 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
277#endif
278
279 rw_init(&kmem_object->lock, "kmem vm object");
280 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
281 kmem_object);
282#if VM_NRESERVLEVEL > 0
283 kmem_object->flags |= OBJ_COLORED;
284 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
285#endif
286
287 /*
288 * The lock portion of struct vm_object must be type stable due
289 * to vm_pageout_fallback_object_lock locking a vm object
290 * without holding any references to it.
291 */
292 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
293#ifdef INVARIANTS
294 vm_object_zdtor,
295#else
296 NULL,
297#endif
298 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
299
300 vm_radix_init();
301}
302
303void
304vm_object_clear_flag(vm_object_t object, u_short bits)
305{
306
307 VM_OBJECT_ASSERT_WLOCKED(object);
308 object->flags &= ~bits;
309}
310
311/*
312 * Sets the default memory attribute for the specified object. Pages
313 * that are allocated to this object are by default assigned this memory
314 * attribute.
315 *
316 * Presently, this function must be called before any pages are allocated
317 * to the object. In the future, this requirement may be relaxed for
318 * "default" and "swap" objects.
319 */
320int
321vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
322{
323
324 VM_OBJECT_ASSERT_WLOCKED(object);
325 switch (object->type) {
326 case OBJT_DEFAULT:
327 case OBJT_DEVICE:
328 case OBJT_MGTDEVICE:
329 case OBJT_PHYS:
330 case OBJT_SG:
331 case OBJT_SWAP:
332 case OBJT_VNODE:
333 if (!TAILQ_EMPTY(&object->memq))
334 return (KERN_FAILURE);
335 break;
336 case OBJT_DEAD:
337 return (KERN_INVALID_ARGUMENT);
338 default:
339 panic("vm_object_set_memattr: object %p is of undefined type",
340 object);
341 }
342 object->memattr = memattr;
343 return (KERN_SUCCESS);
344}
345
346void
347vm_object_pip_add(vm_object_t object, short i)
348{
349
350 VM_OBJECT_ASSERT_WLOCKED(object);
351 object->paging_in_progress += i;
352}
353
354void
355vm_object_pip_subtract(vm_object_t object, short i)
356{
357
358 VM_OBJECT_ASSERT_WLOCKED(object);
359 object->paging_in_progress -= i;
360}
361
362void
363vm_object_pip_wakeup(vm_object_t object)
364{
365
366 VM_OBJECT_ASSERT_WLOCKED(object);
367 object->paging_in_progress--;
368 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
369 vm_object_clear_flag(object, OBJ_PIPWNT);
370 wakeup(object);
371 }
372}
373
374void
375vm_object_pip_wakeupn(vm_object_t object, short i)
376{
377
378 VM_OBJECT_ASSERT_WLOCKED(object);
379 if (i)
380 object->paging_in_progress -= i;
381 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
382 vm_object_clear_flag(object, OBJ_PIPWNT);
383 wakeup(object);
384 }
385}
386
387void
388vm_object_pip_wait(vm_object_t object, char *waitid)
389{
390
391 VM_OBJECT_ASSERT_WLOCKED(object);
392 while (object->paging_in_progress) {
393 object->flags |= OBJ_PIPWNT;
394 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
395 }
396}
397
398/*
399 * vm_object_allocate:
400 *
401 * Returns a new object with the given size.
402 */
403vm_object_t
404vm_object_allocate(objtype_t type, vm_pindex_t size)
405{
406 vm_object_t object;
407
408 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
409 _vm_object_allocate(type, size, object);
410 return (object);
411}
412
413
414/*
415 * vm_object_reference:
416 *
417 * Gets another reference to the given object. Note: OBJ_DEAD
418 * objects can be referenced during final cleaning.
419 */
420void
421vm_object_reference(vm_object_t object)
422{
423 if (object == NULL)
424 return;
425 VM_OBJECT_WLOCK(object);
426 vm_object_reference_locked(object);
427 VM_OBJECT_WUNLOCK(object);
428}
429
430/*
431 * vm_object_reference_locked:
432 *
433 * Gets another reference to the given object.
434 *
435 * The object must be locked.
436 */
437void
438vm_object_reference_locked(vm_object_t object)
439{
440 struct vnode *vp;
441
442 VM_OBJECT_ASSERT_WLOCKED(object);
443 object->ref_count++;
444 if (object->type == OBJT_VNODE) {
445 vp = object->handle;
446 vref(vp);
447 }
448}
449
450/*
451 * Handle deallocating an object of type OBJT_VNODE.
452 */
453static void
454vm_object_vndeallocate(vm_object_t object)
455{
456 struct vnode *vp = (struct vnode *) object->handle;
457
458 VM_OBJECT_ASSERT_WLOCKED(object);
459 KASSERT(object->type == OBJT_VNODE,
460 ("vm_object_vndeallocate: not a vnode object"));
461 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
462#ifdef INVARIANTS
463 if (object->ref_count == 0) {
464 vprint("vm_object_vndeallocate", vp);
465 panic("vm_object_vndeallocate: bad object reference count");
466 }
467#endif
468
469 if (object->ref_count > 1) {
470 object->ref_count--;
471 VM_OBJECT_WUNLOCK(object);
472 /* vrele may need the vnode lock. */
473 vrele(vp);
474 } else {
475 vhold(vp);
476 VM_OBJECT_WUNLOCK(object);
477 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
478 vdrop(vp);
479 VM_OBJECT_WLOCK(object);
480 object->ref_count--;
481 if (object->type == OBJT_DEAD) {
482 VM_OBJECT_WUNLOCK(object);
483 VOP_UNLOCK(vp, 0);
484 } else {
485 if (object->ref_count == 0)
486 VOP_UNSET_TEXT(vp);
487 VM_OBJECT_WUNLOCK(object);
488 vput(vp);
489 }
490 }
491}
492
493/*
494 * vm_object_deallocate:
495 *
496 * Release a reference to the specified object,
497 * gained either through a vm_object_allocate
498 * or a vm_object_reference call. When all references
499 * are gone, storage associated with this object
500 * may be relinquished.
501 *
502 * No object may be locked.
503 */
504void
505vm_object_deallocate(vm_object_t object)
506{
507 vm_object_t temp;
508 struct vnode *vp;
509
510 while (object != NULL) {
511 VM_OBJECT_WLOCK(object);
512 if (object->type == OBJT_VNODE) {
513 vm_object_vndeallocate(object);
514 return;
515 }
516
517 KASSERT(object->ref_count != 0,
518 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
519
520 /*
521 * If the reference count goes to 0 we start calling
522 * vm_object_terminate() on the object chain.
523 * A ref count of 1 may be a special case depending on the
524 * shadow count being 0 or 1.
525 */
526 object->ref_count--;
527 if (object->ref_count > 1) {
528 VM_OBJECT_WUNLOCK(object);
529 return;
530 } else if (object->ref_count == 1) {
531 if (object->type == OBJT_SWAP &&
532 (object->flags & OBJ_TMPFS) != 0) {
533 vp = object->un_pager.swp.swp_tmpfs;
534 vhold(vp);
535 VM_OBJECT_WUNLOCK(object);
536 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
537 vdrop(vp);
538 VM_OBJECT_WLOCK(object);
539 if (object->type == OBJT_DEAD ||
540 object->ref_count != 1) {
541 VM_OBJECT_WUNLOCK(object);
542 VOP_UNLOCK(vp, 0);
543 return;
544 }
545 if ((object->flags & OBJ_TMPFS) != 0)
546 VOP_UNSET_TEXT(vp);
547 VOP_UNLOCK(vp, 0);
548 }
549 if (object->shadow_count == 0 &&
550 object->handle == NULL &&
551 (object->type == OBJT_DEFAULT ||
552 (object->type == OBJT_SWAP &&
553 (object->flags & OBJ_TMPFS) == 0))) {
554 vm_object_set_flag(object, OBJ_ONEMAPPING);
555 } else if ((object->shadow_count == 1) &&
556 (object->handle == NULL) &&
557 (object->type == OBJT_DEFAULT ||
558 object->type == OBJT_SWAP)) {
559 KASSERT((object->flags & OBJ_TMPFS) == 0,
560 ("shadowed tmpfs v_object %p", object));
561 vm_object_t robject;
562
563 robject = LIST_FIRST(&object->shadow_head);
564 KASSERT(robject != NULL,
565 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
566 object->ref_count,
567 object->shadow_count));
568 if (!VM_OBJECT_TRYWLOCK(robject)) {
569 /*
570 * Avoid a potential deadlock.
571 */
572 object->ref_count++;
573 VM_OBJECT_WUNLOCK(object);
574 /*
575 * More likely than not the thread
576 * holding robject's lock has lower
577 * priority than the current thread.
578 * Let the lower priority thread run.
579 */
580 pause("vmo_de", 1);
581 continue;
582 }
583 /*
584 * Collapse object into its shadow unless its
585 * shadow is dead. In that case, object will
586 * be deallocated by the thread that is
587 * deallocating its shadow.
588 */
589 if ((robject->flags & OBJ_DEAD) == 0 &&
590 (robject->handle == NULL) &&
591 (robject->type == OBJT_DEFAULT ||
592 robject->type == OBJT_SWAP)) {
593
594 robject->ref_count++;
595retry:
596 if (robject->paging_in_progress) {
597 VM_OBJECT_WUNLOCK(object);
598 vm_object_pip_wait(robject,
599 "objde1");
600 temp = robject->backing_object;
601 if (object == temp) {
602 VM_OBJECT_WLOCK(object);
603 goto retry;
604 }
605 } else if (object->paging_in_progress) {
606 VM_OBJECT_WUNLOCK(robject);
607 object->flags |= OBJ_PIPWNT;
608 VM_OBJECT_SLEEP(object, object,
609 PDROP | PVM, "objde2", 0);
610 VM_OBJECT_WLOCK(robject);
611 temp = robject->backing_object;
612 if (object == temp) {
613 VM_OBJECT_WLOCK(object);
614 goto retry;
615 }
616 } else
617 VM_OBJECT_WUNLOCK(object);
618
619 if (robject->ref_count == 1) {
620 robject->ref_count--;
621 object = robject;
622 goto doterm;
623 }
624 object = robject;
625 vm_object_collapse(object);
626 VM_OBJECT_WUNLOCK(object);
627 continue;
628 }
629 VM_OBJECT_WUNLOCK(robject);
630 }
631 VM_OBJECT_WUNLOCK(object);
632 return;
633 }
634doterm:
635 temp = object->backing_object;
636 if (temp != NULL) {
637 VM_OBJECT_WLOCK(temp);
638 LIST_REMOVE(object, shadow_list);
639 temp->shadow_count--;
640 VM_OBJECT_WUNLOCK(temp);
641 object->backing_object = NULL;
642 }
643 /*
644 * Don't double-terminate, we could be in a termination
645 * recursion due to the terminate having to sync data
646 * to disk.
647 */
648 if ((object->flags & OBJ_DEAD) == 0)
649 vm_object_terminate(object);
650 else
651 VM_OBJECT_WUNLOCK(object);
652 object = temp;
653 }
654}
655
656/*
657 * vm_object_destroy removes the object from the global object list
658 * and frees the space for the object.
659 */
660void
661vm_object_destroy(vm_object_t object)
662{
663
664 /*
665 * Remove the object from the global object list.
666 */
667 mtx_lock(&vm_object_list_mtx);
668 TAILQ_REMOVE(&vm_object_list, object, object_list);
669 mtx_unlock(&vm_object_list_mtx);
670
671 /*
672 * Release the allocation charge.
673 */
674 if (object->cred != NULL) {
675 KASSERT(object->type == OBJT_DEFAULT ||
676 object->type == OBJT_SWAP,
677 ("vm_object_terminate: non-swap obj %p has cred",
678 object));
679 swap_release_by_cred(object->charge, object->cred);
680 object->charge = 0;
681 crfree(object->cred);
682 object->cred = NULL;
683 }
684
685 /*
686 * Free the space for the object.
687 */
688 uma_zfree(obj_zone, object);
689}
690
691/*
692 * vm_object_terminate actually destroys the specified object, freeing
693 * up all previously used resources.
694 *
695 * The object must be locked.
696 * This routine may block.
697 */
698void
699vm_object_terminate(vm_object_t object)
700{
701 vm_page_t p, p_next;
702
703 VM_OBJECT_ASSERT_WLOCKED(object);
704
705 /*
706 * Make sure no one uses us.
707 */
708 vm_object_set_flag(object, OBJ_DEAD);
709
710 /*
711 * wait for the pageout daemon to be done with the object
712 */
713 vm_object_pip_wait(object, "objtrm");
714
715 KASSERT(!object->paging_in_progress,
716 ("vm_object_terminate: pageout in progress"));
717
718 /*
719 * Clean and free the pages, as appropriate. All references to the
720 * object are gone, so we don't need to lock it.
721 */
722 if (object->type == OBJT_VNODE) {
723 struct vnode *vp = (struct vnode *)object->handle;
724
725 /*
726 * Clean pages and flush buffers.
727 */
728 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
729 VM_OBJECT_WUNLOCK(object);
730
731 vinvalbuf(vp, V_SAVE, 0, 0);
732
733 VM_OBJECT_WLOCK(object);
734 }
735
736 KASSERT(object->ref_count == 0,
737 ("vm_object_terminate: object with references, ref_count=%d",
738 object->ref_count));
739
740 /*
741 * Free any remaining pageable pages. This also removes them from the
742 * paging queues. However, don't free wired pages, just remove them
743 * from the object. Rather than incrementally removing each page from
744 * the object, the page and object are reset to any empty state.
745 */
746 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
747 vm_page_assert_unbusied(p);
748 vm_page_lock(p);
749 /*
750 * Optimize the page's removal from the object by resetting
751 * its "object" field. Specifically, if the page is not
752 * wired, then the effect of this assignment is that
753 * vm_page_free()'s call to vm_page_remove() will return
754 * immediately without modifying the page or the object.
755 */
756 p->object = NULL;
757 if (p->wire_count == 0) {
758 vm_page_free(p);
759 PCPU_INC(cnt.v_pfree);
760 }
761 vm_page_unlock(p);
762 }
763 /*
764 * If the object contained any pages, then reset it to an empty state.
765 * None of the object's fields, including "resident_page_count", were
766 * modified by the preceding loop.
767 */
768 if (object->resident_page_count != 0) {
769 vm_radix_reclaim_allnodes(&object->rtree);
770 TAILQ_INIT(&object->memq);
771 object->resident_page_count = 0;
772 if (object->type == OBJT_VNODE)
773 vdrop(object->handle);
774 }
775
776#if VM_NRESERVLEVEL > 0
777 if (__predict_false(!LIST_EMPTY(&object->rvq)))
778 vm_reserv_break_all(object);
779#endif
780 if (__predict_false(!vm_object_cache_is_empty(object)))
781 vm_page_cache_free(object, 0, 0);
782
783 /*
784 * Let the pager know object is dead.
785 */
786 vm_pager_deallocate(object);
787 VM_OBJECT_WUNLOCK(object);
788
789 vm_object_destroy(object);
790}
791
792/*
793 * Make the page read-only so that we can clear the object flags. However, if
794 * this is a nosync mmap then the object is likely to stay dirty so do not
795 * mess with the page and do not clear the object flags. Returns TRUE if the
796 * page should be flushed, and FALSE otherwise.
797 */
798static boolean_t
799vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
800{
801
802 /*
803 * If we have been asked to skip nosync pages and this is a
804 * nosync page, skip it. Note that the object flags were not
805 * cleared in this case so we do not have to set them.
806 */
807 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
808 *clearobjflags = FALSE;
809 return (FALSE);
810 } else {
811 pmap_remove_write(p);
812 return (p->dirty != 0);
813 }
814}
815
816/*
817 * vm_object_page_clean
818 *
819 * Clean all dirty pages in the specified range of object. Leaves page
820 * on whatever queue it is currently on. If NOSYNC is set then do not
821 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
822 * leaving the object dirty.
823 *
824 * When stuffing pages asynchronously, allow clustering. XXX we need a
825 * synchronous clustering mode implementation.
826 *
827 * Odd semantics: if start == end, we clean everything.
828 *
829 * The object must be locked.
830 *
831 * Returns FALSE if some page from the range was not written, as
832 * reported by the pager, and TRUE otherwise.
833 */
834boolean_t
835vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
836 int flags)
837{
838 vm_page_t np, p;
839 vm_pindex_t pi, tend, tstart;
840 int curgeneration, n, pagerflags;
841 boolean_t clearobjflags, eio, res;
842
843 VM_OBJECT_ASSERT_WLOCKED(object);
844
845 /*
846 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
847 * objects. The check below prevents the function from
848 * operating on non-vnode objects.
849 */
850 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
851 object->resident_page_count == 0)
852 return (TRUE);
853
854 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
855 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
856 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
857
858 tstart = OFF_TO_IDX(start);
859 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
860 clearobjflags = tstart == 0 && tend >= object->size;
861 res = TRUE;
862
863rescan:
864 curgeneration = object->generation;
865
866 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
867 pi = p->pindex;
868 if (pi >= tend)
869 break;
870 np = TAILQ_NEXT(p, listq);
871 if (p->valid == 0)
872 continue;
873 if (vm_page_sleep_if_busy(p, "vpcwai")) {
874 if (object->generation != curgeneration) {
875 if ((flags & OBJPC_SYNC) != 0)
876 goto rescan;
877 else
878 clearobjflags = FALSE;
879 }
880 np = vm_page_find_least(object, pi);
881 continue;
882 }
883 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
884 continue;
885
886 n = vm_object_page_collect_flush(object, p, pagerflags,
887 flags, &clearobjflags, &eio);
888 if (eio) {
889 res = FALSE;
890 clearobjflags = FALSE;
891 }
892 if (object->generation != curgeneration) {
893 if ((flags & OBJPC_SYNC) != 0)
894 goto rescan;
895 else
896 clearobjflags = FALSE;
897 }
898
899 /*
900 * If the VOP_PUTPAGES() did a truncated write, so
901 * that even the first page of the run is not fully
902 * written, vm_pageout_flush() returns 0 as the run
903 * length. Since the condition that caused truncated
904 * write may be permanent, e.g. exhausted free space,
905 * accepting n == 0 would cause an infinite loop.
906 *
907 * Forwarding the iterator leaves the unwritten page
908 * behind, but there is not much we can do there if
909 * filesystem refuses to write it.
910 */
911 if (n == 0) {
912 n = 1;
913 clearobjflags = FALSE;
914 }
915 np = vm_page_find_least(object, pi + n);
916 }
917#if 0
918 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
919#endif
920
921 if (clearobjflags)
922 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
923 return (res);
924}
925
926static int
927vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
928 int flags, boolean_t *clearobjflags, boolean_t *eio)
929{
930 vm_page_t ma[vm_pageout_page_count], p_first, tp;
931 int count, i, mreq, runlen;
932
933 vm_page_lock_assert(p, MA_NOTOWNED);
934 VM_OBJECT_ASSERT_WLOCKED(object);
935
936 count = 1;
937 mreq = 0;
938
939 for (tp = p; count < vm_pageout_page_count; count++) {
940 tp = vm_page_next(tp);
941 if (tp == NULL || vm_page_busied(tp))
942 break;
943 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
944 break;
945 }
946
947 for (p_first = p; count < vm_pageout_page_count; count++) {
948 tp = vm_page_prev(p_first);
949 if (tp == NULL || vm_page_busied(tp))
950 break;
951 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
952 break;
953 p_first = tp;
954 mreq++;
955 }
956
957 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
958 ma[i] = tp;
959
960 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
961 return (runlen);
962}
963
964/*
965 * Note that there is absolutely no sense in writing out
966 * anonymous objects, so we track down the vnode object
967 * to write out.
968 * We invalidate (remove) all pages from the address space
969 * for semantic correctness.
970 *
971 * If the backing object is a device object with unmanaged pages, then any
972 * mappings to the specified range of pages must be removed before this
973 * function is called.
974 *
975 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
976 * may start out with a NULL object.
977 */
978boolean_t
979vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
980 boolean_t syncio, boolean_t invalidate)
981{
982 vm_object_t backing_object;
983 struct vnode *vp;
984 struct mount *mp;
985 int error, flags, fsync_after;
986 boolean_t res;
987
988 if (object == NULL)
989 return (TRUE);
990 res = TRUE;
991 error = 0;
992 VM_OBJECT_WLOCK(object);
993 while ((backing_object = object->backing_object) != NULL) {
994 VM_OBJECT_WLOCK(backing_object);
995 offset += object->backing_object_offset;
996 VM_OBJECT_WUNLOCK(object);
997 object = backing_object;
998 if (object->size < OFF_TO_IDX(offset + size))
999 size = IDX_TO_OFF(object->size) - offset;
1000 }
1001 /*
1002 * Flush pages if writing is allowed, invalidate them
1003 * if invalidation requested. Pages undergoing I/O
1004 * will be ignored by vm_object_page_remove().
1005 *
1006 * We cannot lock the vnode and then wait for paging
1007 * to complete without deadlocking against vm_fault.
1008 * Instead we simply call vm_object_page_remove() and
1009 * allow it to block internally on a page-by-page
1010 * basis when it encounters pages undergoing async
1011 * I/O.
1012 */
1013 if (object->type == OBJT_VNODE &&
1014 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1015 vp = object->handle;
1016 VM_OBJECT_WUNLOCK(object);
1017 (void) vn_start_write(vp, &mp, V_WAIT);
1018 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1019 if (syncio && !invalidate && offset == 0 &&
1020 OFF_TO_IDX(size) == object->size) {
1021 /*
1022 * If syncing the whole mapping of the file,
1023 * it is faster to schedule all the writes in
1024 * async mode, also allowing the clustering,
1025 * and then wait for i/o to complete.
1026 */
1027 flags = 0;
1028 fsync_after = TRUE;
1029 } else {
1030 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1031 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1032 fsync_after = FALSE;
1033 }
1034 VM_OBJECT_WLOCK(object);
1035 res = vm_object_page_clean(object, offset, offset + size,
1036 flags);
1037 VM_OBJECT_WUNLOCK(object);
1038 if (fsync_after)
1039 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1040 VOP_UNLOCK(vp, 0);
1041 vn_finished_write(mp);
1042 if (error != 0)
1043 res = FALSE;
1044 VM_OBJECT_WLOCK(object);
1045 }
1046 if ((object->type == OBJT_VNODE ||
1047 object->type == OBJT_DEVICE) && invalidate) {
1048 if (object->type == OBJT_DEVICE)
1049 /*
1050 * The option OBJPR_NOTMAPPED must be passed here
1051 * because vm_object_page_remove() cannot remove
1052 * unmanaged mappings.
1053 */
1054 flags = OBJPR_NOTMAPPED;
1055 else if (old_msync)
1056 flags = OBJPR_NOTWIRED;
1057 else
1058 flags = OBJPR_CLEANONLY | OBJPR_NOTWIRED;
1059 vm_object_page_remove(object, OFF_TO_IDX(offset),
1060 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1061 }
1062 VM_OBJECT_WUNLOCK(object);
1063 return (res);
1064}
1065
1066/*
1067 * vm_object_madvise:
1068 *
1069 * Implements the madvise function at the object/page level.
1070 *
1071 * MADV_WILLNEED (any object)
1072 *
1073 * Activate the specified pages if they are resident.
1074 *
1075 * MADV_DONTNEED (any object)
1076 *
1077 * Deactivate the specified pages if they are resident.
1078 *
1079 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1080 * OBJ_ONEMAPPING only)
1081 *
1082 * Deactivate and clean the specified pages if they are
1083 * resident. This permits the process to reuse the pages
1084 * without faulting or the kernel to reclaim the pages
1085 * without I/O.
1086 */
1087void
1088vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1089 int advise)
1090{
1091 vm_pindex_t tpindex;
1092 vm_object_t backing_object, tobject;
1093 vm_page_t m;
1094
1095 if (object == NULL)
1096 return;
1097 VM_OBJECT_WLOCK(object);
1098 /*
1099 * Locate and adjust resident pages
1100 */
1101 for (; pindex < end; pindex += 1) {
1102relookup:
1103 tobject = object;
1104 tpindex = pindex;
1105shadowlookup:
1106 /*
1107 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1108 * and those pages must be OBJ_ONEMAPPING.
1109 */
1110 if (advise == MADV_FREE) {
1111 if ((tobject->type != OBJT_DEFAULT &&
1112 tobject->type != OBJT_SWAP) ||
1113 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1114 goto unlock_tobject;
1115 }
1116 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1117 goto unlock_tobject;
1118 m = vm_page_lookup(tobject, tpindex);
1119 if (m == NULL && advise == MADV_WILLNEED) {
1120 /*
1121 * If the page is cached, reactivate it.
1122 */
1123 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1124 VM_ALLOC_NOBUSY);
1125 }
1126 if (m == NULL) {
1127 /*
1128 * There may be swap even if there is no backing page
1129 */
1130 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1131 swap_pager_freespace(tobject, tpindex, 1);
1132 /*
1133 * next object
1134 */
1135 backing_object = tobject->backing_object;
1136 if (backing_object == NULL)
1137 goto unlock_tobject;
1138 VM_OBJECT_WLOCK(backing_object);
1139 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1140 if (tobject != object)
1141 VM_OBJECT_WUNLOCK(tobject);
1142 tobject = backing_object;
1143 goto shadowlookup;
1144 } else if (m->valid != VM_PAGE_BITS_ALL)
1145 goto unlock_tobject;
1146 /*
1147 * If the page is not in a normal state, skip it.
1148 */
1149 vm_page_lock(m);
1150 if (m->hold_count != 0 || m->wire_count != 0) {
1151 vm_page_unlock(m);
1152 goto unlock_tobject;
1153 }
1154 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1155 ("vm_object_madvise: page %p is fictitious", m));
1156 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1157 ("vm_object_madvise: page %p is not managed", m));
1158 if (vm_page_busied(m)) {
1159 if (advise == MADV_WILLNEED) {
1160 /*
1161 * Reference the page before unlocking and
1162 * sleeping so that the page daemon is less
1163 * likely to reclaim it.
1164 */
1165 vm_page_aflag_set(m, PGA_REFERENCED);
1166 }
1167 if (object != tobject)
1168 VM_OBJECT_WUNLOCK(object);
1169 VM_OBJECT_WUNLOCK(tobject);
1170 vm_page_busy_sleep(m, "madvpo");
1171 VM_OBJECT_WLOCK(object);
1172 goto relookup;
1173 }
1174 if (advise == MADV_WILLNEED) {
1175 vm_page_activate(m);
1176 } else {
1177 vm_page_advise(m, advise);
1178 }
1179 vm_page_unlock(m);
1180 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1181 swap_pager_freespace(tobject, tpindex, 1);
1182unlock_tobject:
1183 if (tobject != object)
1184 VM_OBJECT_WUNLOCK(tobject);
1185 }
1186 VM_OBJECT_WUNLOCK(object);
1187}
1188
1189/*
1190 * vm_object_shadow:
1191 *
1192 * Create a new object which is backed by the
1193 * specified existing object range. The source
1194 * object reference is deallocated.
1195 *
1196 * The new object and offset into that object
1197 * are returned in the source parameters.
1198 */
1199void
1200vm_object_shadow(
1201 vm_object_t *object, /* IN/OUT */
1202 vm_ooffset_t *offset, /* IN/OUT */
1203 vm_size_t length)
1204{
1205 vm_object_t source;
1206 vm_object_t result;
1207
1208 source = *object;
1209
1210 /*
1211 * Don't create the new object if the old object isn't shared.
1212 */
1213 if (source != NULL) {
1214 VM_OBJECT_WLOCK(source);
1215 if (source->ref_count == 1 &&
1216 source->handle == NULL &&
1217 (source->type == OBJT_DEFAULT ||
1218 source->type == OBJT_SWAP)) {
1219 VM_OBJECT_WUNLOCK(source);
1220 return;
1221 }
1222 VM_OBJECT_WUNLOCK(source);
1223 }
1224
1225 /*
1226 * Allocate a new object with the given length.
1227 */
1228 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1229
1230 /*
1231 * The new object shadows the source object, adding a reference to it.
1232 * Our caller changes his reference to point to the new object,
1233 * removing a reference to the source object. Net result: no change
1234 * of reference count.
1235 *
1236 * Try to optimize the result object's page color when shadowing
1237 * in order to maintain page coloring consistency in the combined
1238 * shadowed object.
1239 */
1240 result->backing_object = source;
1241 /*
1242 * Store the offset into the source object, and fix up the offset into
1243 * the new object.
1244 */
1245 result->backing_object_offset = *offset;
1246 if (source != NULL) {
1247 VM_OBJECT_WLOCK(source);
1248 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1249 source->shadow_count++;
1250#if VM_NRESERVLEVEL > 0
1251 result->flags |= source->flags & OBJ_COLORED;
1252 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1253 ((1 << (VM_NFREEORDER - 1)) - 1);
1254#endif
1255 VM_OBJECT_WUNLOCK(source);
1256 }
1257
1258
1259 /*
1260 * Return the new things
1261 */
1262 *offset = 0;
1263 *object = result;
1264}
1265
1266/*
1267 * vm_object_split:
1268 *
1269 * Split the pages in a map entry into a new object. This affords
1270 * easier removal of unused pages, and keeps object inheritance from
1271 * being a negative impact on memory usage.
1272 */
1273void
1274vm_object_split(vm_map_entry_t entry)
1275{
1276 vm_page_t m, m_next;
1277 vm_object_t orig_object, new_object, source;
1278 vm_pindex_t idx, offidxstart;
1279 vm_size_t size;
1280
1281 orig_object = entry->object.vm_object;
1282 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1283 return;
1284 if (orig_object->ref_count <= 1)
1285 return;
1286 VM_OBJECT_WUNLOCK(orig_object);
1287
1288 offidxstart = OFF_TO_IDX(entry->offset);
1289 size = atop(entry->end - entry->start);
1290
1291 /*
1292 * If swap_pager_copy() is later called, it will convert new_object
1293 * into a swap object.
1294 */
1295 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1296
1297 /*
1298 * At this point, the new object is still private, so the order in
1299 * which the original and new objects are locked does not matter.
1300 */
1301 VM_OBJECT_WLOCK(new_object);
1302 VM_OBJECT_WLOCK(orig_object);
1303 source = orig_object->backing_object;
1304 if (source != NULL) {
1305 VM_OBJECT_WLOCK(source);
1306 if ((source->flags & OBJ_DEAD) != 0) {
1307 VM_OBJECT_WUNLOCK(source);
1308 VM_OBJECT_WUNLOCK(orig_object);
1309 VM_OBJECT_WUNLOCK(new_object);
1310 vm_object_deallocate(new_object);
1311 VM_OBJECT_WLOCK(orig_object);
1312 return;
1313 }
1314 LIST_INSERT_HEAD(&source->shadow_head,
1315 new_object, shadow_list);
1316 source->shadow_count++;
1317 vm_object_reference_locked(source); /* for new_object */
1318 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1319 VM_OBJECT_WUNLOCK(source);
1320 new_object->backing_object_offset =
1321 orig_object->backing_object_offset + entry->offset;
1322 new_object->backing_object = source;
1323 }
1324 if (orig_object->cred != NULL) {
1325 new_object->cred = orig_object->cred;
1326 crhold(orig_object->cred);
1327 new_object->charge = ptoa(size);
1328 KASSERT(orig_object->charge >= ptoa(size),
1329 ("orig_object->charge < 0"));
1330 orig_object->charge -= ptoa(size);
1331 }
1332retry:
1333 m = vm_page_find_least(orig_object, offidxstart);
1334 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1335 m = m_next) {
1336 m_next = TAILQ_NEXT(m, listq);
1337
1338 /*
1339 * We must wait for pending I/O to complete before we can
1340 * rename the page.
1341 *
1342 * We do not have to VM_PROT_NONE the page as mappings should
1343 * not be changed by this operation.
1344 */
1345 if (vm_page_busied(m)) {
1346 VM_OBJECT_WUNLOCK(new_object);
1347 vm_page_lock(m);
1348 VM_OBJECT_WUNLOCK(orig_object);
1349 vm_page_busy_sleep(m, "spltwt");
1350 VM_OBJECT_WLOCK(orig_object);
1351 VM_OBJECT_WLOCK(new_object);
1352 goto retry;
1353 }
1354#if VM_NRESERVLEVEL > 0
1355 /*
1356 * If some of the reservation's allocated pages remain with
1357 * the original object, then transferring the reservation to
1358 * the new object is neither particularly beneficial nor
1359 * particularly harmful as compared to leaving the reservation
1360 * with the original object. If, however, all of the
1361 * reservation's allocated pages are transferred to the new
1362 * object, then transferring the reservation is typically
1363 * beneficial. Determining which of these two cases applies
1364 * would be more costly than unconditionally renaming the
1365 * reservation.
1366 */
1367 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1368#endif
1369 vm_page_lock(m);
1370 vm_page_rename(m, new_object, idx);
1371 vm_page_unlock(m);
1372 /* page automatically made dirty by rename and cache handled */
1373 if (orig_object->type == OBJT_SWAP)
1374 vm_page_xbusy(m);
1375 }
1376 if (orig_object->type == OBJT_SWAP) {
1377 /*
1378 * swap_pager_copy() can sleep, in which case the orig_object's
1379 * and new_object's locks are released and reacquired.
1380 */
1381 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1382 TAILQ_FOREACH(m, &new_object->memq, listq)
1383 vm_page_xunbusy(m);
1384
1385 /*
1386 * Transfer any cached pages from orig_object to new_object.
1387 * If swap_pager_copy() found swapped out pages within the
1388 * specified range of orig_object, then it changed
1389 * new_object's type to OBJT_SWAP when it transferred those
1390 * pages to new_object. Otherwise, new_object's type
1391 * should still be OBJT_DEFAULT and orig_object should not
1392 * contain any cached pages within the specified range.
1393 */
1394 if (__predict_false(!vm_object_cache_is_empty(orig_object)))
1395 vm_page_cache_transfer(orig_object, offidxstart,
1396 new_object);
1397 }
1398 VM_OBJECT_WUNLOCK(orig_object);
1399 VM_OBJECT_WUNLOCK(new_object);
1400 entry->object.vm_object = new_object;
1401 entry->offset = 0LL;
1402 vm_object_deallocate(orig_object);
1403 VM_OBJECT_WLOCK(new_object);
1404}
1405
1406#define OBSC_TEST_ALL_SHADOWED 0x0001
1407#define OBSC_COLLAPSE_NOWAIT 0x0002
1408#define OBSC_COLLAPSE_WAIT 0x0004
1409
1410static int
1411vm_object_backing_scan(vm_object_t object, int op)
1412{
1413 int r = 1;
1414 vm_page_t p;
1415 vm_object_t backing_object;
1416 vm_pindex_t backing_offset_index;
1417
1418 VM_OBJECT_ASSERT_WLOCKED(object);
1419 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1420
1421 backing_object = object->backing_object;
1422 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1423
1424 /*
1425 * Initial conditions
1426 */
1427 if (op & OBSC_TEST_ALL_SHADOWED) {
1428 /*
1429 * We do not want to have to test for the existence of cache
1430 * or swap pages in the backing object. XXX but with the
1431 * new swapper this would be pretty easy to do.
1432 *
1433 * XXX what about anonymous MAP_SHARED memory that hasn't
1434 * been ZFOD faulted yet? If we do not test for this, the
1435 * shadow test may succeed! XXX
1436 */
1437 if (backing_object->type != OBJT_DEFAULT) {
1438 return (0);
1439 }
1440 }
1441 if (op & OBSC_COLLAPSE_WAIT) {
1442 vm_object_set_flag(backing_object, OBJ_DEAD);
1443 }
1444
1445 /*
1446 * Our scan
1447 */
1448 p = TAILQ_FIRST(&backing_object->memq);
1449 while (p) {
1450 vm_page_t next = TAILQ_NEXT(p, listq);
1451 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1452
1453 if (op & OBSC_TEST_ALL_SHADOWED) {
1454 vm_page_t pp;
1455
1456 /*
1457 * Ignore pages outside the parent object's range
1458 * and outside the parent object's mapping of the
1459 * backing object.
1460 *
1461 * note that we do not busy the backing object's
1462 * page.
1463 */
1464 if (
1465 p->pindex < backing_offset_index ||
1466 new_pindex >= object->size
1467 ) {
1468 p = next;
1469 continue;
1470 }
1471
1472 /*
1473 * See if the parent has the page or if the parent's
1474 * object pager has the page. If the parent has the
1475 * page but the page is not valid, the parent's
1476 * object pager must have the page.
1477 *
1478 * If this fails, the parent does not completely shadow
1479 * the object and we might as well give up now.
1480 */
1481
1482 pp = vm_page_lookup(object, new_pindex);
1483 if (
1484 (pp == NULL || pp->valid == 0) &&
1485 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1486 ) {
1487 r = 0;
1488 break;
1489 }
1490 }
1491
1492 /*
1493 * Check for busy page
1494 */
1495 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1496 vm_page_t pp;
1497
1498 if (op & OBSC_COLLAPSE_NOWAIT) {
1499 if (!p->valid || vm_page_busied(p)) {
1500 p = next;
1501 continue;
1502 }
1503 } else if (op & OBSC_COLLAPSE_WAIT) {
1504 if (vm_page_busied(p)) {
1505 VM_OBJECT_WUNLOCK(object);
1506 vm_page_lock(p);
1507 VM_OBJECT_WUNLOCK(backing_object);
1508 vm_page_busy_sleep(p, "vmocol");
1509 VM_OBJECT_WLOCK(object);
1510 VM_OBJECT_WLOCK(backing_object);
1511 /*
1512 * If we slept, anything could have
1513 * happened. Since the object is
1514 * marked dead, the backing offset
1515 * should not have changed so we
1516 * just restart our scan.
1517 */
1518 p = TAILQ_FIRST(&backing_object->memq);
1519 continue;
1520 }
1521 }
1522
1523 KASSERT(
1524 p->object == backing_object,
1525 ("vm_object_backing_scan: object mismatch")
1526 );
1527
1528 /*
1529 * Destroy any associated swap
1530 */
1531 if (backing_object->type == OBJT_SWAP) {
1532 swap_pager_freespace(
1533 backing_object,
1534 p->pindex,
1535 1
1536 );
1537 }
1538
1539 if (
1540 p->pindex < backing_offset_index ||
1541 new_pindex >= object->size
1542 ) {
1543 /*
1544 * Page is out of the parent object's range, we
1545 * can simply destroy it.
1546 */
1547 vm_page_lock(p);
1548 KASSERT(!pmap_page_is_mapped(p),
1549 ("freeing mapped page %p", p));
1550 if (p->wire_count == 0)
1551 vm_page_free(p);
1552 else
1553 vm_page_remove(p);
1554 vm_page_unlock(p);
1555 p = next;
1556 continue;
1557 }
1558
1559 pp = vm_page_lookup(object, new_pindex);
1560 if (
1561 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1562 (pp != NULL && pp->valid == 0)
1563 ) {
1564 /*
1565 * The page in the parent is not (yet) valid.
1566 * We don't know anything about the state of
1567 * the original page. It might be mapped,
1568 * so we must avoid the next if here.
1569 *
1570 * This is due to a race in vm_fault() where
1571 * we must unbusy the original (backing_obj)
1572 * page before we can (re)lock the parent.
1573 * Hence we can get here.
1574 */
1575 p = next;
1576 continue;
1577 }
1578 if (
1579 pp != NULL ||
1580 vm_pager_has_page(object, new_pindex, NULL, NULL)
1581 ) {
1582 /*
1583 * page already exists in parent OR swap exists
1584 * for this location in the parent. Destroy
1585 * the original page from the backing object.
1586 *
1587 * Leave the parent's page alone
1588 */
1589 vm_page_lock(p);
1590 KASSERT(!pmap_page_is_mapped(p),
1591 ("freeing mapped page %p", p));
1592 if (p->wire_count == 0)
1593 vm_page_free(p);
1594 else
1595 vm_page_remove(p);
1596 vm_page_unlock(p);
1597 p = next;
1598 continue;
1599 }
1600
1601#if VM_NRESERVLEVEL > 0
1602 /*
1603 * Rename the reservation.
1604 */
1605 vm_reserv_rename(p, object, backing_object,
1606 backing_offset_index);
1607#endif
1608
1609 /*
1610 * Page does not exist in parent, rename the
1611 * page from the backing object to the main object.
1612 *
1613 * If the page was mapped to a process, it can remain
1614 * mapped through the rename.
1615 */
1616 vm_page_lock(p);
1617 vm_page_rename(p, object, new_pindex);
1618 vm_page_unlock(p);
1619 /* page automatically made dirty by rename */
1620 }
1621 p = next;
1622 }
1623 return (r);
1624}
1625
1626
1627/*
1628 * this version of collapse allows the operation to occur earlier and
1629 * when paging_in_progress is true for an object... This is not a complete
1630 * operation, but should plug 99.9% of the rest of the leaks.
1631 */
1632static void
1633vm_object_qcollapse(vm_object_t object)
1634{
1635 vm_object_t backing_object = object->backing_object;
1636
1637 VM_OBJECT_ASSERT_WLOCKED(object);
1638 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1639
1640 if (backing_object->ref_count != 1)
1641 return;
1642
1643 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1644}
1645
1646/*
1647 * vm_object_collapse:
1648 *
1649 * Collapse an object with the object backing it.
1650 * Pages in the backing object are moved into the
1651 * parent, and the backing object is deallocated.
1652 */
1653void
1654vm_object_collapse(vm_object_t object)
1655{
1656 VM_OBJECT_ASSERT_WLOCKED(object);
1657
1658 while (TRUE) {
1659 vm_object_t backing_object;
1660
1661 /*
1662 * Verify that the conditions are right for collapse:
1663 *
1664 * The object exists and the backing object exists.
1665 */
1666 if ((backing_object = object->backing_object) == NULL)
1667 break;
1668
1669 /*
1670 * we check the backing object first, because it is most likely
1671 * not collapsable.
1672 */
1673 VM_OBJECT_WLOCK(backing_object);
1674 if (backing_object->handle != NULL ||
1675 (backing_object->type != OBJT_DEFAULT &&
1676 backing_object->type != OBJT_SWAP) ||
1677 (backing_object->flags & OBJ_DEAD) ||
1678 object->handle != NULL ||
1679 (object->type != OBJT_DEFAULT &&
1680 object->type != OBJT_SWAP) ||
1681 (object->flags & OBJ_DEAD)) {
1682 VM_OBJECT_WUNLOCK(backing_object);
1683 break;
1684 }
1685
1686 if (
1687 object->paging_in_progress != 0 ||
1688 backing_object->paging_in_progress != 0
1689 ) {
1690 vm_object_qcollapse(object);
1691 VM_OBJECT_WUNLOCK(backing_object);
1692 break;
1693 }
1694 /*
1695 * We know that we can either collapse the backing object (if
1696 * the parent is the only reference to it) or (perhaps) have
1697 * the parent bypass the object if the parent happens to shadow
1698 * all the resident pages in the entire backing object.
1699 *
1700 * This is ignoring pager-backed pages such as swap pages.
1701 * vm_object_backing_scan fails the shadowing test in this
1702 * case.
1703 */
1704 if (backing_object->ref_count == 1) {
1705 /*
1706 * If there is exactly one reference to the backing
1707 * object, we can collapse it into the parent.
1708 */
1709 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1710
1711#if VM_NRESERVLEVEL > 0
1712 /*
1713 * Break any reservations from backing_object.
1714 */
1715 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1716 vm_reserv_break_all(backing_object);
1717#endif
1718
1719 /*
1720 * Move the pager from backing_object to object.
1721 */
1722 if (backing_object->type == OBJT_SWAP) {
1723 /*
1724 * swap_pager_copy() can sleep, in which case
1725 * the backing_object's and object's locks are
1726 * released and reacquired.
1727 * Since swap_pager_copy() is being asked to
1728 * destroy the source, it will change the
1729 * backing_object's type to OBJT_DEFAULT.
1730 */
1731 swap_pager_copy(
1732 backing_object,
1733 object,
1734 OFF_TO_IDX(object->backing_object_offset), TRUE);
1735
1736 /*
1737 * Free any cached pages from backing_object.
1738 */
1739 if (__predict_false(
1740 !vm_object_cache_is_empty(backing_object)))
1741 vm_page_cache_free(backing_object, 0, 0);
1742 }
1743 /*
1744 * Object now shadows whatever backing_object did.
1745 * Note that the reference to
1746 * backing_object->backing_object moves from within
1747 * backing_object to within object.
1748 */
1749 LIST_REMOVE(object, shadow_list);
1750 backing_object->shadow_count--;
1751 if (backing_object->backing_object) {
1752 VM_OBJECT_WLOCK(backing_object->backing_object);
1753 LIST_REMOVE(backing_object, shadow_list);
1754 LIST_INSERT_HEAD(
1755 &backing_object->backing_object->shadow_head,
1756 object, shadow_list);
1757 /*
1758 * The shadow_count has not changed.
1759 */
1760 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1761 }
1762 object->backing_object = backing_object->backing_object;
1763 object->backing_object_offset +=
1764 backing_object->backing_object_offset;
1765
1766 /*
1767 * Discard backing_object.
1768 *
1769 * Since the backing object has no pages, no pager left,
1770 * and no object references within it, all that is
1771 * necessary is to dispose of it.
1772 */
1773 KASSERT(backing_object->ref_count == 1, (
1774"backing_object %p was somehow re-referenced during collapse!",
1775 backing_object));
1776 VM_OBJECT_WUNLOCK(backing_object);
1777 vm_object_destroy(backing_object);
1778
1779 object_collapses++;
1780 } else {
1781 vm_object_t new_backing_object;
1782
1783 /*
1784 * If we do not entirely shadow the backing object,
1785 * there is nothing we can do so we give up.
1786 */
1787 if (object->resident_page_count != object->size &&
1788 vm_object_backing_scan(object,
1789 OBSC_TEST_ALL_SHADOWED) == 0) {
1790 VM_OBJECT_WUNLOCK(backing_object);
1791 break;
1792 }
1793
1794 /*
1795 * Make the parent shadow the next object in the
1796 * chain. Deallocating backing_object will not remove
1797 * it, since its reference count is at least 2.
1798 */
1799 LIST_REMOVE(object, shadow_list);
1800 backing_object->shadow_count--;
1801
1802 new_backing_object = backing_object->backing_object;
1803 if ((object->backing_object = new_backing_object) != NULL) {
1804 VM_OBJECT_WLOCK(new_backing_object);
1805 LIST_INSERT_HEAD(
1806 &new_backing_object->shadow_head,
1807 object,
1808 shadow_list
1809 );
1810 new_backing_object->shadow_count++;
1811 vm_object_reference_locked(new_backing_object);
1812 VM_OBJECT_WUNLOCK(new_backing_object);
1813 object->backing_object_offset +=
1814 backing_object->backing_object_offset;
1815 }
1816
1817 /*
1818 * Drop the reference count on backing_object. Since
1819 * its ref_count was at least 2, it will not vanish.
1820 */
1821 backing_object->ref_count--;
1822 VM_OBJECT_WUNLOCK(backing_object);
1823 object_bypasses++;
1824 }
1825
1826 /*
1827 * Try again with this object's new backing object.
1828 */
1829 }
1830}
1831
1832/*
1833 * vm_object_page_remove:
1834 *
1835 * For the given object, either frees or invalidates each of the
1836 * specified pages. In general, a page is freed. However, if a page is
1837 * wired for any reason other than the existence of a managed, wired
1838 * mapping, then it may be invalidated but not removed from the object.
1839 * Pages are specified by the given range ["start", "end") and the option
1840 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1841 * extends from "start" to the end of the object. If the option
1842 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1843 * specified range are affected. If the option OBJPR_NOTMAPPED is
1844 * specified, then the pages within the specified range must have no
1845 * mappings. Otherwise, if this option is not specified, any mappings to
1846 * the specified pages are removed before the pages are freed or
1847 * invalidated.
1848 *
1849 * In general, this operation should only be performed on objects that
1850 * contain managed pages. There are, however, two exceptions. First, it
1851 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1852 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1853 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1854 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1855 *
1856 * The object must be locked.
1857 */
1858void
1859vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1860 int options)
1861{
1862 vm_page_t p, next;
1863 int wirings;
1864
1865 VM_OBJECT_ASSERT_WLOCKED(object);
1866 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1867 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1868 ("vm_object_page_remove: illegal options for object %p", object));
1869 if (object->resident_page_count == 0)
1870 goto skipmemq;
1871 vm_object_pip_add(object, 1);
1872again:
1873 p = vm_page_find_least(object, start);
1874
1875 /*
1876 * Here, the variable "p" is either (1) the page with the least pindex
1877 * greater than or equal to the parameter "start" or (2) NULL.
1878 */
1879 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1880 next = TAILQ_NEXT(p, listq);
1881
1882 /*
1883 * If the page is wired for any reason besides the existence
1884 * of managed, wired mappings, then it cannot be freed. For
1885 * example, fictitious pages, which represent device memory,
1886 * are inherently wired and cannot be freed. They can,
1887 * however, be invalidated if the option OBJPR_CLEANONLY is
1888 * not specified.
1889 */
1890 vm_page_lock(p);
1891 if ((wirings = p->wire_count) != 0 &&
1892 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1893 if ((options & (OBJPR_NOTWIRED | OBJPR_NOTMAPPED)) ==
1894 0) {
1895 pmap_remove_all(p);
1896 /* Account for removal of wired mappings. */
1897 if (wirings != 0)
1898 p->wire_count -= wirings;
1899 }
1900 if ((options & OBJPR_CLEANONLY) == 0) {
1901 p->valid = 0;
1902 vm_page_undirty(p);
1903 }
1904 goto next;
1905 }
1906 if (vm_page_busied(p)) {
1907 VM_OBJECT_WUNLOCK(object);
1908 vm_page_busy_sleep(p, "vmopar");
1909 VM_OBJECT_WLOCK(object);
1910 goto again;
1911 }
1912 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1913 ("vm_object_page_remove: page %p is fictitious", p));
1914 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1915 if ((options & OBJPR_NOTMAPPED) == 0)
1916 pmap_remove_write(p);
1917 if (p->dirty)
1918 goto next;
1919 }
1920 if ((options & OBJPR_NOTMAPPED) == 0) {
1921 if ((options & OBJPR_NOTWIRED) != 0 && wirings != 0)
1922 goto next;
1923 pmap_remove_all(p);
1924 /* Account for removal of wired mappings. */
1925 if (wirings != 0) {
1926 KASSERT(p->wire_count == wirings,
1927 ("inconsistent wire count %d %d %p",
1928 p->wire_count, wirings, p));
1929 p->wire_count = 0;
1930 atomic_subtract_int(&cnt.v_wire_count, 1);
1931 }
1932 }
1933 vm_page_free(p);
1934next:
1935 vm_page_unlock(p);
1936 }
1937 vm_object_pip_wakeup(object);
1938skipmemq:
1939 if (__predict_false(!vm_object_cache_is_empty(object)))
1940 vm_page_cache_free(object, start, end);
1941}
1942
1943/*
1944 * vm_object_page_cache:
1945 *
1946 * For the given object, attempt to move the specified clean
1947 * pages to the cache queue. If a page is wired for any reason,
1948 * then it will not be changed. Pages are specified by the given
1949 * range ["start", "end"). As a special case, if "end" is zero,
1950 * then the range extends from "start" to the end of the object.
1951 * Any mappings to the specified pages are removed before the
1952 * pages are moved to the cache queue.
1953 *
1954 * This operation should only be performed on objects that
1955 * contain non-fictitious, managed pages.
1956 *
1957 * The object must be locked.
1958 */
1959void
1960vm_object_page_cache(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1961{
1962 struct mtx *mtx, *new_mtx;
1963 vm_page_t p, next;
1964
1965 VM_OBJECT_ASSERT_WLOCKED(object);
1966 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1967 ("vm_object_page_cache: illegal object %p", object));
1968 if (object->resident_page_count == 0)
1969 return;
1970 p = vm_page_find_least(object, start);
1971
1972 /*
1973 * Here, the variable "p" is either (1) the page with the least pindex
1974 * greater than or equal to the parameter "start" or (2) NULL.
1975 */
1976 mtx = NULL;
1977 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1978 next = TAILQ_NEXT(p, listq);
1979
1980 /*
1981 * Avoid releasing and reacquiring the same page lock.
1982 */
1983 new_mtx = vm_page_lockptr(p);
1984 if (mtx != new_mtx) {
1985 if (mtx != NULL)
1986 mtx_unlock(mtx);
1987 mtx = new_mtx;
1988 mtx_lock(mtx);
1989 }
1990 vm_page_try_to_cache(p);
1991 }
1992 if (mtx != NULL)
1993 mtx_unlock(mtx);
1994}
1995
1996/*
1997 * Populate the specified range of the object with valid pages. Returns
1998 * TRUE if the range is successfully populated and FALSE otherwise.
1999 *
2000 * Note: This function should be optimized to pass a larger array of
2001 * pages to vm_pager_get_pages() before it is applied to a non-
2002 * OBJT_DEVICE object.
2003 *
2004 * The object must be locked.
2005 */
2006boolean_t
2007vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2008{
2009 vm_page_t m, ma[1];
2010 vm_pindex_t pindex;
2011 int rv;
2012
2013 VM_OBJECT_ASSERT_WLOCKED(object);
2014 for (pindex = start; pindex < end; pindex++) {
2015 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
2016 VM_ALLOC_RETRY);
2017 if (m->valid != VM_PAGE_BITS_ALL) {
2018 ma[0] = m;
2019 rv = vm_pager_get_pages(object, ma, 1, 0);
2020 m = vm_page_lookup(object, pindex);
2021 if (m == NULL)
2022 break;
2023 if (rv != VM_PAGER_OK) {
2024 vm_page_lock(m);
2025 vm_page_free(m);
2026 vm_page_unlock(m);
2027 break;
2028 }
2029 }
2030 /*
2031 * Keep "m" busy because a subsequent iteration may unlock
2032 * the object.
2033 */
2034 }
2035 if (pindex > start) {
2036 m = vm_page_lookup(object, start);
2037 while (m != NULL && m->pindex < pindex) {
2038 vm_page_xunbusy(m);
2039 m = TAILQ_NEXT(m, listq);
2040 }
2041 }
2042 return (pindex == end);
2043}
2044
2045/*
2046 * Routine: vm_object_coalesce
2047 * Function: Coalesces two objects backing up adjoining
2048 * regions of memory into a single object.
2049 *
2050 * returns TRUE if objects were combined.
2051 *
2052 * NOTE: Only works at the moment if the second object is NULL -
2053 * if it's not, which object do we lock first?
2054 *
2055 * Parameters:
2056 * prev_object First object to coalesce
2057 * prev_offset Offset into prev_object
2058 * prev_size Size of reference to prev_object
2059 * next_size Size of reference to the second object
2060 * reserved Indicator that extension region has
2061 * swap accounted for
2062 *
2063 * Conditions:
2064 * The object must *not* be locked.
2065 */
2066boolean_t
2067vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2068 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2069{
2070 vm_pindex_t next_pindex;
2071
2072 if (prev_object == NULL)
2073 return (TRUE);
2074 VM_OBJECT_WLOCK(prev_object);
2075 if (prev_object->type != OBJT_DEFAULT &&
2076 prev_object->type != OBJT_SWAP) {
2077 VM_OBJECT_WUNLOCK(prev_object);
2078 return (FALSE);
2079 }
2080
2081 /*
2082 * Try to collapse the object first
2083 */
2084 vm_object_collapse(prev_object);
2085
2086 /*
2087 * Can't coalesce if: . more than one reference . paged out . shadows
2088 * another object . has a copy elsewhere (any of which mean that the
2089 * pages not mapped to prev_entry may be in use anyway)
2090 */
2091 if (prev_object->backing_object != NULL) {
2092 VM_OBJECT_WUNLOCK(prev_object);
2093 return (FALSE);
2094 }
2095
2096 prev_size >>= PAGE_SHIFT;
2097 next_size >>= PAGE_SHIFT;
2098 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2099
2100 if ((prev_object->ref_count > 1) &&
2101 (prev_object->size != next_pindex)) {
2102 VM_OBJECT_WUNLOCK(prev_object);
2103 return (FALSE);
2104 }
2105
2106 /*
2107 * Account for the charge.
2108 */
2109 if (prev_object->cred != NULL) {
2110
2111 /*
2112 * If prev_object was charged, then this mapping,
2113 * althought not charged now, may become writable
2114 * later. Non-NULL cred in the object would prevent
2115 * swap reservation during enabling of the write
2116 * access, so reserve swap now. Failed reservation
2117 * cause allocation of the separate object for the map
2118 * entry, and swap reservation for this entry is
2119 * managed in appropriate time.
2120 */
2121 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2122 prev_object->cred)) {
2123 return (FALSE);
2124 }
2125 prev_object->charge += ptoa(next_size);
2126 }
2127
2128 /*
2129 * Remove any pages that may still be in the object from a previous
2130 * deallocation.
2131 */
2132 if (next_pindex < prev_object->size) {
2133 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2134 next_size, 0);
2135 if (prev_object->type == OBJT_SWAP)
2136 swap_pager_freespace(prev_object,
2137 next_pindex, next_size);
2138#if 0
2139 if (prev_object->cred != NULL) {
2140 KASSERT(prev_object->charge >=
2141 ptoa(prev_object->size - next_pindex),
2142 ("object %p overcharged 1 %jx %jx", prev_object,
2143 (uintmax_t)next_pindex, (uintmax_t)next_size));
2144 prev_object->charge -= ptoa(prev_object->size -
2145 next_pindex);
2146 }
2147#endif
2148 }
2149
2150 /*
2151 * Extend the object if necessary.
2152 */
2153 if (next_pindex + next_size > prev_object->size)
2154 prev_object->size = next_pindex + next_size;
2155
2156 VM_OBJECT_WUNLOCK(prev_object);
2157 return (TRUE);
2158}
2159
2160void
2161vm_object_set_writeable_dirty(vm_object_t object)
2162{
2163
2164 VM_OBJECT_ASSERT_WLOCKED(object);
2165 if (object->type != OBJT_VNODE)
2166 return;
2167 object->generation++;
2168 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2169 return;
2170 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2171}
2172
2173#include "opt_ddb.h"
2174#ifdef DDB
2175#include <sys/kernel.h>
2176
2177#include <sys/cons.h>
2178
2179#include <ddb/ddb.h>
2180
2181static int
2182_vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2183{
2184 vm_map_t tmpm;
2185 vm_map_entry_t tmpe;
2186 vm_object_t obj;
2187 int entcount;
2188
2189 if (map == 0)
2190 return 0;
2191
2192 if (entry == 0) {
2193 tmpe = map->header.next;
2194 entcount = map->nentries;
2195 while (entcount-- && (tmpe != &map->header)) {
2196 if (_vm_object_in_map(map, object, tmpe)) {
2197 return 1;
2198 }
2199 tmpe = tmpe->next;
2200 }
2201 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2202 tmpm = entry->object.sub_map;
2203 tmpe = tmpm->header.next;
2204 entcount = tmpm->nentries;
2205 while (entcount-- && tmpe != &tmpm->header) {
2206 if (_vm_object_in_map(tmpm, object, tmpe)) {
2207 return 1;
2208 }
2209 tmpe = tmpe->next;
2210 }
2211 } else if ((obj = entry->object.vm_object) != NULL) {
2212 for (; obj; obj = obj->backing_object)
2213 if (obj == object) {
2214 return 1;
2215 }
2216 }
2217 return 0;
2218}
2219
2220static int
2221vm_object_in_map(vm_object_t object)
2222{
2223 struct proc *p;
2224
2225 /* sx_slock(&allproc_lock); */
2226 FOREACH_PROC_IN_SYSTEM(p) {
2227 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2228 continue;
2229 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2230 /* sx_sunlock(&allproc_lock); */
2231 return 1;
2232 }
2233 }
2234 /* sx_sunlock(&allproc_lock); */
2235 if (_vm_object_in_map(kernel_map, object, 0))
2236 return 1;
2237 return 0;
2238}
2239
2240DB_SHOW_COMMAND(vmochk, vm_object_check)
2241{
2242 vm_object_t object;
2243
2244 /*
2245 * make sure that internal objs are in a map somewhere
2246 * and none have zero ref counts.
2247 */
2248 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2249 if (object->handle == NULL &&
2250 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2251 if (object->ref_count == 0) {
2252 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2253 (long)object->size);
2254 }
2255 if (!vm_object_in_map(object)) {
2256 db_printf(
2257 "vmochk: internal obj is not in a map: "
2258 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2259 object->ref_count, (u_long)object->size,
2260 (u_long)object->size,
2261 (void *)object->backing_object);
2262 }
2263 }
2264 }
2265}
2266
2267/*
2268 * vm_object_print: [ debug ]
2269 */
2270DB_SHOW_COMMAND(object, vm_object_print_static)
2271{
2272 /* XXX convert args. */
2273 vm_object_t object = (vm_object_t)addr;
2274 boolean_t full = have_addr;
2275
2276 vm_page_t p;
2277
2278 /* XXX count is an (unused) arg. Avoid shadowing it. */
2279#define count was_count
2280
2281 int count;
2282
2283 if (object == NULL)
2284 return;
2285
2286 db_iprintf(
2287 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2288 object, (int)object->type, (uintmax_t)object->size,
2289 object->resident_page_count, object->ref_count, object->flags,
2290 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2291 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2292 object->shadow_count,
2293 object->backing_object ? object->backing_object->ref_count : 0,
2294 object->backing_object, (uintmax_t)object->backing_object_offset);
2295
2296 if (!full)
2297 return;
2298
2299 db_indent += 2;
2300 count = 0;
2301 TAILQ_FOREACH(p, &object->memq, listq) {
2302 if (count == 0)
2303 db_iprintf("memory:=");
2304 else if (count == 6) {
2305 db_printf("\n");
2306 db_iprintf(" ...");
2307 count = 0;
2308 } else
2309 db_printf(",");
2310 count++;
2311
2312 db_printf("(off=0x%jx,page=0x%jx)",
2313 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2314 }
2315 if (count != 0)
2316 db_printf("\n");
2317 db_indent -= 2;
2318}
2319
2320/* XXX. */
2321#undef count
2322
2323/* XXX need this non-static entry for calling from vm_map_print. */
2324void
2325vm_object_print(
2326 /* db_expr_t */ long addr,
2327 boolean_t have_addr,
2328 /* db_expr_t */ long count,
2329 char *modif)
2330{
2331 vm_object_print_static(addr, have_addr, count, modif);
2332}
2333
2334DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2335{
2336 vm_object_t object;
2337 vm_pindex_t fidx;
2338 vm_paddr_t pa;
2339 vm_page_t m, prev_m;
2340 int rcount, nl, c;
2341
2342 nl = 0;
2343 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2344 db_printf("new object: %p\n", (void *)object);
2345 if (nl > 18) {
2346 c = cngetc();
2347 if (c != ' ')
2348 return;
2349 nl = 0;
2350 }
2351 nl++;
2352 rcount = 0;
2353 fidx = 0;
2354 pa = -1;
2355 TAILQ_FOREACH(m, &object->memq, listq) {
2356 if (m->pindex > 128)
2357 break;
2358 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2359 prev_m->pindex + 1 != m->pindex) {
2360 if (rcount) {
2361 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2362 (long)fidx, rcount, (long)pa);
2363 if (nl > 18) {
2364 c = cngetc();
2365 if (c != ' ')
2366 return;
2367 nl = 0;
2368 }
2369 nl++;
2370 rcount = 0;
2371 }
2372 }
2373 if (rcount &&
2374 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2375 ++rcount;
2376 continue;
2377 }
2378 if (rcount) {
2379 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2380 (long)fidx, rcount, (long)pa);
2381 if (nl > 18) {
2382 c = cngetc();
2383 if (c != ' ')
2384 return;
2385 nl = 0;
2386 }
2387 nl++;
2388 }
2389 fidx = m->pindex;
2390 pa = VM_PAGE_TO_PHYS(m);
2391 rcount = 1;
2392 }
2393 if (rcount) {
2394 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2395 (long)fidx, rcount, (long)pa);
2396 if (nl > 18) {
2397 c = cngetc();
2398 if (c != ' ')
2399 return;
2400 nl = 0;
2401 }
2402 nl++;
2403 }
2404 }
2405}
2406#endif /* DDB */