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: stable/11/sys/vm/vm_object.c 323537 2017-09-13 11:19:04Z kib $");
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/pctrie.h>
77#include <sys/sysctl.h>
78#include <sys/mutex.h>
79#include <sys/proc.h> /* for curproc, pageproc */
80#include <sys/socket.h>
81#include <sys/resourcevar.h>
82#include <sys/rwlock.h>
83#include <sys/user.h>
84#include <sys/vnode.h>
85#include <sys/vmmeter.h>
86#include <sys/sx.h>
87
88#include <vm/vm.h>
89#include <vm/vm_param.h>
90#include <vm/pmap.h>
91#include <vm/vm_map.h>
92#include <vm/vm_object.h>
93#include <vm/vm_page.h>
94#include <vm/vm_pageout.h>
95#include <vm/vm_pager.h>
96#include <vm/swap_pager.h>
97#include <vm/vm_kern.h>
98#include <vm/vm_extern.h>
99#include <vm/vm_radix.h>
100#include <vm/vm_reserv.h>
101#include <vm/uma.h>
102
103static int old_msync;
104SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
105 "Use old (insecure) msync behavior");
106
107static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
108 int pagerflags, int flags, boolean_t *clearobjflags,
109 boolean_t *eio);
110static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
111 boolean_t *clearobjflags);
112static void vm_object_qcollapse(vm_object_t object);
113static void vm_object_vndeallocate(vm_object_t object);
114
115/*
116 * Virtual memory objects maintain the actual data
117 * associated with allocated virtual memory. A given
118 * page of memory exists within exactly one object.
119 *
120 * An object is only deallocated when all "references"
121 * are given up. Only one "reference" to a given
122 * region of an object should be writeable.
123 *
124 * Associated with each object is a list of all resident
125 * memory pages belonging to that object; this list is
126 * maintained by the "vm_page" module, and locked by the object's
127 * lock.
128 *
129 * Each object also records a "pager" routine which is
130 * used to retrieve (and store) pages to the proper backing
131 * storage. In addition, objects may be backed by other
132 * objects from which they were virtual-copied.
133 *
134 * The only items within the object structure which are
135 * modified after time of creation are:
136 * reference count locked by object's lock
137 * pager routine locked by object's lock
138 *
139 */
140
141struct object_q vm_object_list;
142struct mtx vm_object_list_mtx; /* lock for object list and count */
143
144struct vm_object kernel_object_store;
145struct vm_object kmem_object_store;
146
147static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
148 "VM object stats");
149
150static long object_collapses;
151SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
152 &object_collapses, 0, "VM object collapses");
153
154static long object_bypasses;
155SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
156 &object_bypasses, 0, "VM object bypasses");
157
158static uma_zone_t obj_zone;
159
160static int vm_object_zinit(void *mem, int size, int flags);
161
162#ifdef INVARIANTS
163static void vm_object_zdtor(void *mem, int size, void *arg);
164
165static void
166vm_object_zdtor(void *mem, int size, void *arg)
167{
168 vm_object_t object;
169
170 object = (vm_object_t)mem;
171 KASSERT(object->ref_count == 0,
172 ("object %p ref_count = %d", object, object->ref_count));
173 KASSERT(TAILQ_EMPTY(&object->memq),
174 ("object %p has resident pages in its memq", object));
175 KASSERT(vm_radix_is_empty(&object->rtree),
176 ("object %p has resident pages in its trie", object));
177#if VM_NRESERVLEVEL > 0
178 KASSERT(LIST_EMPTY(&object->rvq),
179 ("object %p has reservations",
180 object));
181#endif
182 KASSERT(object->paging_in_progress == 0,
183 ("object %p paging_in_progress = %d",
184 object, object->paging_in_progress));
185 KASSERT(object->resident_page_count == 0,
186 ("object %p resident_page_count = %d",
187 object, object->resident_page_count));
188 KASSERT(object->shadow_count == 0,
189 ("object %p shadow_count = %d",
190 object, object->shadow_count));
191 KASSERT(object->type == OBJT_DEAD,
192 ("object %p has non-dead type %d",
193 object, object->type));
194}
195#endif
196
197static int
198vm_object_zinit(void *mem, int size, int flags)
199{
200 vm_object_t object;
201
202 object = (vm_object_t)mem;
203 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
204
205 /* These are true for any object that has been freed */
206 object->type = OBJT_DEAD;
207 object->ref_count = 0;
208 vm_radix_init(&object->rtree);
209 object->paging_in_progress = 0;
210 object->resident_page_count = 0;
211 object->shadow_count = 0;
212 object->flags = OBJ_DEAD;
213
214 mtx_lock(&vm_object_list_mtx);
215 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
216 mtx_unlock(&vm_object_list_mtx);
217 return (0);
218}
219
220static void
221_vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
222{
223
224 TAILQ_INIT(&object->memq);
225 LIST_INIT(&object->shadow_head);
226
227 object->type = type;
228 if (type == OBJT_SWAP)
229 pctrie_init(&object->un_pager.swp.swp_blks);
230
231 /*
232 * Ensure that swap_pager_swapoff() iteration over object_list
233 * sees up to date type and pctrie head if it observed
234 * non-dead object.
235 */
236 atomic_thread_fence_rel();
237
238 switch (type) {
239 case OBJT_DEAD:
240 panic("_vm_object_allocate: can't create OBJT_DEAD");
241 case OBJT_DEFAULT:
242 case OBJT_SWAP:
243 object->flags = OBJ_ONEMAPPING;
244 break;
245 case OBJT_DEVICE:
246 case OBJT_SG:
247 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
248 break;
249 case OBJT_MGTDEVICE:
250 object->flags = OBJ_FICTITIOUS;
251 break;
252 case OBJT_PHYS:
253 object->flags = OBJ_UNMANAGED;
254 break;
255 case OBJT_VNODE:
256 object->flags = 0;
257 break;
258 default:
259 panic("_vm_object_allocate: type %d is undefined", type);
260 }
261 object->size = size;
262 object->generation = 1;
263 object->ref_count = 1;
264 object->memattr = VM_MEMATTR_DEFAULT;
265 object->cred = NULL;
266 object->charge = 0;
267 object->handle = NULL;
268 object->backing_object = NULL;
269 object->backing_object_offset = (vm_ooffset_t) 0;
270#if VM_NRESERVLEVEL > 0
271 LIST_INIT(&object->rvq);
272#endif
273 umtx_shm_object_init(object);
274}
275
276/*
277 * vm_object_init:
278 *
279 * Initialize the VM objects module.
280 */
281void
282vm_object_init(void)
283{
284 TAILQ_INIT(&vm_object_list);
285 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
286
287 rw_init(&kernel_object->lock, "kernel vm object");
288 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
289 VM_MIN_KERNEL_ADDRESS), kernel_object);
290#if VM_NRESERVLEVEL > 0
291 kernel_object->flags |= OBJ_COLORED;
292 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
293#endif
294
295 rw_init(&kmem_object->lock, "kmem vm object");
296 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
297 VM_MIN_KERNEL_ADDRESS), kmem_object);
298#if VM_NRESERVLEVEL > 0
299 kmem_object->flags |= OBJ_COLORED;
300 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
301#endif
302
303 /*
304 * The lock portion of struct vm_object must be type stable due
305 * to vm_pageout_fallback_object_lock locking a vm object
306 * without holding any references to it.
307 */
308 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
309#ifdef INVARIANTS
310 vm_object_zdtor,
311#else
312 NULL,
313#endif
314 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
315
316 vm_radix_zinit();
317}
318
319void
320vm_object_clear_flag(vm_object_t object, u_short bits)
321{
322
323 VM_OBJECT_ASSERT_WLOCKED(object);
324 object->flags &= ~bits;
325}
326
327/*
328 * Sets the default memory attribute for the specified object. Pages
329 * that are allocated to this object are by default assigned this memory
330 * attribute.
331 *
332 * Presently, this function must be called before any pages are allocated
333 * to the object. In the future, this requirement may be relaxed for
334 * "default" and "swap" objects.
335 */
336int
337vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
338{
339
340 VM_OBJECT_ASSERT_WLOCKED(object);
341 switch (object->type) {
342 case OBJT_DEFAULT:
343 case OBJT_DEVICE:
344 case OBJT_MGTDEVICE:
345 case OBJT_PHYS:
346 case OBJT_SG:
347 case OBJT_SWAP:
348 case OBJT_VNODE:
349 if (!TAILQ_EMPTY(&object->memq))
350 return (KERN_FAILURE);
351 break;
352 case OBJT_DEAD:
353 return (KERN_INVALID_ARGUMENT);
354 default:
355 panic("vm_object_set_memattr: object %p is of undefined type",
356 object);
357 }
358 object->memattr = memattr;
359 return (KERN_SUCCESS);
360}
361
362void
363vm_object_pip_add(vm_object_t object, short i)
364{
365
366 VM_OBJECT_ASSERT_WLOCKED(object);
367 object->paging_in_progress += i;
368}
369
370void
371vm_object_pip_subtract(vm_object_t object, short i)
372{
373
374 VM_OBJECT_ASSERT_WLOCKED(object);
375 object->paging_in_progress -= i;
376}
377
378void
379vm_object_pip_wakeup(vm_object_t object)
380{
381
382 VM_OBJECT_ASSERT_WLOCKED(object);
383 object->paging_in_progress--;
384 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
385 vm_object_clear_flag(object, OBJ_PIPWNT);
386 wakeup(object);
387 }
388}
389
390void
391vm_object_pip_wakeupn(vm_object_t object, short i)
392{
393
394 VM_OBJECT_ASSERT_WLOCKED(object);
395 if (i)
396 object->paging_in_progress -= i;
397 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
398 vm_object_clear_flag(object, OBJ_PIPWNT);
399 wakeup(object);
400 }
401}
402
403void
404vm_object_pip_wait(vm_object_t object, char *waitid)
405{
406
407 VM_OBJECT_ASSERT_WLOCKED(object);
408 while (object->paging_in_progress) {
409 object->flags |= OBJ_PIPWNT;
410 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
411 }
412}
413
414/*
415 * vm_object_allocate:
416 *
417 * Returns a new object with the given size.
418 */
419vm_object_t
420vm_object_allocate(objtype_t type, vm_pindex_t size)
421{
422 vm_object_t object;
423
424 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
425 _vm_object_allocate(type, size, object);
426 return (object);
427}
428
429
430/*
431 * vm_object_reference:
432 *
433 * Gets another reference to the given object. Note: OBJ_DEAD
434 * objects can be referenced during final cleaning.
435 */
436void
437vm_object_reference(vm_object_t object)
438{
439 if (object == NULL)
440 return;
441 VM_OBJECT_WLOCK(object);
442 vm_object_reference_locked(object);
443 VM_OBJECT_WUNLOCK(object);
444}
445
446/*
447 * vm_object_reference_locked:
448 *
449 * Gets another reference to the given object.
450 *
451 * The object must be locked.
452 */
453void
454vm_object_reference_locked(vm_object_t object)
455{
456 struct vnode *vp;
457
458 VM_OBJECT_ASSERT_WLOCKED(object);
459 object->ref_count++;
460 if (object->type == OBJT_VNODE) {
461 vp = object->handle;
462 vref(vp);
463 }
464}
465
466/*
467 * Handle deallocating an object of type OBJT_VNODE.
468 */
469static void
470vm_object_vndeallocate(vm_object_t object)
471{
472 struct vnode *vp = (struct vnode *) object->handle;
473
474 VM_OBJECT_ASSERT_WLOCKED(object);
475 KASSERT(object->type == OBJT_VNODE,
476 ("vm_object_vndeallocate: not a vnode object"));
477 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
478#ifdef INVARIANTS
479 if (object->ref_count == 0) {
480 vn_printf(vp, "vm_object_vndeallocate ");
481 panic("vm_object_vndeallocate: bad object reference count");
482 }
483#endif
484
485 if (!umtx_shm_vnobj_persistent && object->ref_count == 1)
486 umtx_shm_object_terminated(object);
487
488 /*
489 * The test for text of vp vnode does not need a bypass to
490 * reach right VV_TEXT there, since it is obtained from
491 * object->handle.
492 */
493 if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) {
494 object->ref_count--;
495 VM_OBJECT_WUNLOCK(object);
496 /* vrele may need the vnode lock. */
497 vrele(vp);
498 } else {
499 vhold(vp);
500 VM_OBJECT_WUNLOCK(object);
501 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
502 vdrop(vp);
503 VM_OBJECT_WLOCK(object);
504 object->ref_count--;
505 if (object->type == OBJT_DEAD) {
506 VM_OBJECT_WUNLOCK(object);
507 VOP_UNLOCK(vp, 0);
508 } else {
509 if (object->ref_count == 0)
510 VOP_UNSET_TEXT(vp);
511 VM_OBJECT_WUNLOCK(object);
512 vput(vp);
513 }
514 }
515}
516
517/*
518 * vm_object_deallocate:
519 *
520 * Release a reference to the specified object,
521 * gained either through a vm_object_allocate
522 * or a vm_object_reference call. When all references
523 * are gone, storage associated with this object
524 * may be relinquished.
525 *
526 * No object may be locked.
527 */
528void
529vm_object_deallocate(vm_object_t object)
530{
531 vm_object_t temp;
532 struct vnode *vp;
533
534 while (object != NULL) {
535 VM_OBJECT_WLOCK(object);
536 if (object->type == OBJT_VNODE) {
537 vm_object_vndeallocate(object);
538 return;
539 }
540
541 KASSERT(object->ref_count != 0,
542 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
543
544 /*
545 * If the reference count goes to 0 we start calling
546 * vm_object_terminate() on the object chain.
547 * A ref count of 1 may be a special case depending on the
548 * shadow count being 0 or 1.
549 */
550 object->ref_count--;
551 if (object->ref_count > 1) {
552 VM_OBJECT_WUNLOCK(object);
553 return;
554 } else if (object->ref_count == 1) {
555 if (object->type == OBJT_SWAP &&
556 (object->flags & OBJ_TMPFS) != 0) {
557 vp = object->un_pager.swp.swp_tmpfs;
558 vhold(vp);
559 VM_OBJECT_WUNLOCK(object);
560 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
561 VM_OBJECT_WLOCK(object);
562 if (object->type == OBJT_DEAD ||
563 object->ref_count != 1) {
564 VM_OBJECT_WUNLOCK(object);
565 VOP_UNLOCK(vp, 0);
566 vdrop(vp);
567 return;
568 }
569 if ((object->flags & OBJ_TMPFS) != 0)
570 VOP_UNSET_TEXT(vp);
571 VOP_UNLOCK(vp, 0);
572 vdrop(vp);
573 }
574 if (object->shadow_count == 0 &&
575 object->handle == NULL &&
576 (object->type == OBJT_DEFAULT ||
577 (object->type == OBJT_SWAP &&
578 (object->flags & OBJ_TMPFS_NODE) == 0))) {
579 vm_object_set_flag(object, OBJ_ONEMAPPING);
580 } else if ((object->shadow_count == 1) &&
581 (object->handle == NULL) &&
582 (object->type == OBJT_DEFAULT ||
583 object->type == OBJT_SWAP)) {
584 vm_object_t robject;
585
586 robject = LIST_FIRST(&object->shadow_head);
587 KASSERT(robject != NULL,
588 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
589 object->ref_count,
590 object->shadow_count));
591 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0,
592 ("shadowed tmpfs v_object %p", object));
593 if (!VM_OBJECT_TRYWLOCK(robject)) {
594 /*
595 * Avoid a potential deadlock.
596 */
597 object->ref_count++;
598 VM_OBJECT_WUNLOCK(object);
599 /*
600 * More likely than not the thread
601 * holding robject's lock has lower
602 * priority than the current thread.
603 * Let the lower priority thread run.
604 */
605 pause("vmo_de", 1);
606 continue;
607 }
608 /*
609 * Collapse object into its shadow unless its
610 * shadow is dead. In that case, object will
611 * be deallocated by the thread that is
612 * deallocating its shadow.
613 */
614 if ((robject->flags & OBJ_DEAD) == 0 &&
615 (robject->handle == NULL) &&
616 (robject->type == OBJT_DEFAULT ||
617 robject->type == OBJT_SWAP)) {
618
619 robject->ref_count++;
620retry:
621 if (robject->paging_in_progress) {
622 VM_OBJECT_WUNLOCK(object);
623 vm_object_pip_wait(robject,
624 "objde1");
625 temp = robject->backing_object;
626 if (object == temp) {
627 VM_OBJECT_WLOCK(object);
628 goto retry;
629 }
630 } else if (object->paging_in_progress) {
631 VM_OBJECT_WUNLOCK(robject);
632 object->flags |= OBJ_PIPWNT;
633 VM_OBJECT_SLEEP(object, object,
634 PDROP | PVM, "objde2", 0);
635 VM_OBJECT_WLOCK(robject);
636 temp = robject->backing_object;
637 if (object == temp) {
638 VM_OBJECT_WLOCK(object);
639 goto retry;
640 }
641 } else
642 VM_OBJECT_WUNLOCK(object);
643
644 if (robject->ref_count == 1) {
645 robject->ref_count--;
646 object = robject;
647 goto doterm;
648 }
649 object = robject;
650 vm_object_collapse(object);
651 VM_OBJECT_WUNLOCK(object);
652 continue;
653 }
654 VM_OBJECT_WUNLOCK(robject);
655 }
656 VM_OBJECT_WUNLOCK(object);
657 return;
658 }
659doterm:
660 umtx_shm_object_terminated(object);
661 temp = object->backing_object;
662 if (temp != NULL) {
663 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
664 ("shadowed tmpfs v_object 2 %p", object));
665 VM_OBJECT_WLOCK(temp);
666 LIST_REMOVE(object, shadow_list);
667 temp->shadow_count--;
668 VM_OBJECT_WUNLOCK(temp);
669 object->backing_object = NULL;
670 }
671 /*
672 * Don't double-terminate, we could be in a termination
673 * recursion due to the terminate having to sync data
674 * to disk.
675 */
676 if ((object->flags & OBJ_DEAD) == 0)
677 vm_object_terminate(object);
678 else
679 VM_OBJECT_WUNLOCK(object);
680 object = temp;
681 }
682}
683
684/*
685 * vm_object_destroy removes the object from the global object list
686 * and frees the space for the object.
687 */
688void
689vm_object_destroy(vm_object_t object)
690{
691
692 /*
693 * Release the allocation charge.
694 */
695 if (object->cred != NULL) {
696 swap_release_by_cred(object->charge, object->cred);
697 object->charge = 0;
698 crfree(object->cred);
699 object->cred = NULL;
700 }
701
702 /*
703 * Free the space for the object.
704 */
705 uma_zfree(obj_zone, object);
706}
707
708/*
709 * vm_object_terminate actually destroys the specified object, freeing
710 * up all previously used resources.
711 *
712 * The object must be locked.
713 * This routine may block.
714 */
715void
716vm_object_terminate(vm_object_t object)
717{
718 vm_page_t p, p_next;
719
720 VM_OBJECT_ASSERT_WLOCKED(object);
721
722 /*
723 * Make sure no one uses us.
724 */
725 vm_object_set_flag(object, OBJ_DEAD);
726
727 /*
728 * wait for the pageout daemon to be done with the object
729 */
730 vm_object_pip_wait(object, "objtrm");
731
732 KASSERT(!object->paging_in_progress,
733 ("vm_object_terminate: pageout in progress"));
734
735 /*
736 * Clean and free the pages, as appropriate. All references to the
737 * object are gone, so we don't need to lock it.
738 */
739 if (object->type == OBJT_VNODE) {
740 struct vnode *vp = (struct vnode *)object->handle;
741
742 /*
743 * Clean pages and flush buffers.
744 */
745 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
746 VM_OBJECT_WUNLOCK(object);
747
748 vinvalbuf(vp, V_SAVE, 0, 0);
749
750 BO_LOCK(&vp->v_bufobj);
751 vp->v_bufobj.bo_flag |= BO_DEAD;
752 BO_UNLOCK(&vp->v_bufobj);
753
754 VM_OBJECT_WLOCK(object);
755 }
756
757 KASSERT(object->ref_count == 0,
758 ("vm_object_terminate: object with references, ref_count=%d",
759 object->ref_count));
760
761 /*
762 * Free any remaining pageable pages. This also removes them from the
763 * paging queues. However, don't free wired pages, just remove them
764 * from the object. Rather than incrementally removing each page from
765 * the object, the page and object are reset to any empty state.
766 */
767 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
768 vm_page_assert_unbusied(p);
769 vm_page_lock(p);
770 /*
771 * Optimize the page's removal from the object by resetting
772 * its "object" field. Specifically, if the page is not
773 * wired, then the effect of this assignment is that
774 * vm_page_free()'s call to vm_page_remove() will return
775 * immediately without modifying the page or the object.
776 */
777 p->object = NULL;
778 if (p->wire_count == 0) {
779 vm_page_free(p);
780 PCPU_INC(cnt.v_pfree);
781 }
782 vm_page_unlock(p);
783 }
784 /*
785 * If the object contained any pages, then reset it to an empty state.
786 * None of the object's fields, including "resident_page_count", were
787 * modified by the preceding loop.
788 */
789 if (object->resident_page_count != 0) {
790 vm_radix_reclaim_allnodes(&object->rtree);
791 TAILQ_INIT(&object->memq);
792 object->resident_page_count = 0;
793 if (object->type == OBJT_VNODE)
794 vdrop(object->handle);
795 }
796
797#if VM_NRESERVLEVEL > 0
798 if (__predict_false(!LIST_EMPTY(&object->rvq)))
799 vm_reserv_break_all(object);
800#endif
801
802 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
803 object->type == OBJT_SWAP,
804 ("%s: non-swap obj %p has cred", __func__, object));
805
806 /*
807 * Let the pager know object is dead.
808 */
809 vm_pager_deallocate(object);
810 VM_OBJECT_WUNLOCK(object);
811
812 vm_object_destroy(object);
813}
814
815/*
816 * Make the page read-only so that we can clear the object flags. However, if
817 * this is a nosync mmap then the object is likely to stay dirty so do not
818 * mess with the page and do not clear the object flags. Returns TRUE if the
819 * page should be flushed, and FALSE otherwise.
820 */
821static boolean_t
822vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
823{
824
825 /*
826 * If we have been asked to skip nosync pages and this is a
827 * nosync page, skip it. Note that the object flags were not
828 * cleared in this case so we do not have to set them.
829 */
830 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
831 *clearobjflags = FALSE;
832 return (FALSE);
833 } else {
834 pmap_remove_write(p);
835 return (p->dirty != 0);
836 }
837}
838
839/*
840 * vm_object_page_clean
841 *
842 * Clean all dirty pages in the specified range of object. Leaves page
843 * on whatever queue it is currently on. If NOSYNC is set then do not
844 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
845 * leaving the object dirty.
846 *
847 * When stuffing pages asynchronously, allow clustering. XXX we need a
848 * synchronous clustering mode implementation.
849 *
850 * Odd semantics: if start == end, we clean everything.
851 *
852 * The object must be locked.
853 *
854 * Returns FALSE if some page from the range was not written, as
855 * reported by the pager, and TRUE otherwise.
856 */
857boolean_t
858vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
859 int flags)
860{
861 vm_page_t np, p;
862 vm_pindex_t pi, tend, tstart;
863 int curgeneration, n, pagerflags;
864 boolean_t clearobjflags, eio, res;
865
866 VM_OBJECT_ASSERT_WLOCKED(object);
867
868 /*
869 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
870 * objects. The check below prevents the function from
871 * operating on non-vnode objects.
872 */
873 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
874 object->resident_page_count == 0)
875 return (TRUE);
876
877 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
878 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
879 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
880
881 tstart = OFF_TO_IDX(start);
882 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
883 clearobjflags = tstart == 0 && tend >= object->size;
884 res = TRUE;
885
886rescan:
887 curgeneration = object->generation;
888
889 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
890 pi = p->pindex;
891 if (pi >= tend)
892 break;
893 np = TAILQ_NEXT(p, listq);
894 if (p->valid == 0)
895 continue;
896 if (vm_page_sleep_if_busy(p, "vpcwai")) {
897 if (object->generation != curgeneration) {
898 if ((flags & OBJPC_SYNC) != 0)
899 goto rescan;
900 else
901 clearobjflags = FALSE;
902 }
903 np = vm_page_find_least(object, pi);
904 continue;
905 }
906 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
907 continue;
908
909 n = vm_object_page_collect_flush(object, p, pagerflags,
910 flags, &clearobjflags, &eio);
911 if (eio) {
912 res = FALSE;
913 clearobjflags = FALSE;
914 }
915 if (object->generation != curgeneration) {
916 if ((flags & OBJPC_SYNC) != 0)
917 goto rescan;
918 else
919 clearobjflags = FALSE;
920 }
921
922 /*
923 * If the VOP_PUTPAGES() did a truncated write, so
924 * that even the first page of the run is not fully
925 * written, vm_pageout_flush() returns 0 as the run
926 * length. Since the condition that caused truncated
927 * write may be permanent, e.g. exhausted free space,
928 * accepting n == 0 would cause an infinite loop.
929 *
930 * Forwarding the iterator leaves the unwritten page
931 * behind, but there is not much we can do there if
932 * filesystem refuses to write it.
933 */
934 if (n == 0) {
935 n = 1;
936 clearobjflags = FALSE;
937 }
938 np = vm_page_find_least(object, pi + n);
939 }
940#if 0
941 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
942#endif
943
944 if (clearobjflags)
945 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
946 return (res);
947}
948
949static int
950vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
951 int flags, boolean_t *clearobjflags, boolean_t *eio)
952{
953 vm_page_t ma[vm_pageout_page_count], p_first, tp;
954 int count, i, mreq, runlen;
955
956 vm_page_lock_assert(p, MA_NOTOWNED);
957 VM_OBJECT_ASSERT_WLOCKED(object);
958
959 count = 1;
960 mreq = 0;
961
962 for (tp = p; count < vm_pageout_page_count; count++) {
963 tp = vm_page_next(tp);
964 if (tp == NULL || vm_page_busied(tp))
965 break;
966 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
967 break;
968 }
969
970 for (p_first = p; count < vm_pageout_page_count; count++) {
971 tp = vm_page_prev(p_first);
972 if (tp == NULL || vm_page_busied(tp))
973 break;
974 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
975 break;
976 p_first = tp;
977 mreq++;
978 }
979
980 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
981 ma[i] = tp;
982
983 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
984 return (runlen);
985}
986
987/*
988 * Note that there is absolutely no sense in writing out
989 * anonymous objects, so we track down the vnode object
990 * to write out.
991 * We invalidate (remove) all pages from the address space
992 * for semantic correctness.
993 *
994 * If the backing object is a device object with unmanaged pages, then any
995 * mappings to the specified range of pages must be removed before this
996 * function is called.
997 *
998 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
999 * may start out with a NULL object.
1000 */
1001boolean_t
1002vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1003 boolean_t syncio, boolean_t invalidate)
1004{
1005 vm_object_t backing_object;
1006 struct vnode *vp;
1007 struct mount *mp;
1008 int error, flags, fsync_after;
1009 boolean_t res;
1010
1011 if (object == NULL)
1012 return (TRUE);
1013 res = TRUE;
1014 error = 0;
1015 VM_OBJECT_WLOCK(object);
1016 while ((backing_object = object->backing_object) != NULL) {
1017 VM_OBJECT_WLOCK(backing_object);
1018 offset += object->backing_object_offset;
1019 VM_OBJECT_WUNLOCK(object);
1020 object = backing_object;
1021 if (object->size < OFF_TO_IDX(offset + size))
1022 size = IDX_TO_OFF(object->size) - offset;
1023 }
1024 /*
1025 * Flush pages if writing is allowed, invalidate them
1026 * if invalidation requested. Pages undergoing I/O
1027 * will be ignored by vm_object_page_remove().
1028 *
1029 * We cannot lock the vnode and then wait for paging
1030 * to complete without deadlocking against vm_fault.
1031 * Instead we simply call vm_object_page_remove() and
1032 * allow it to block internally on a page-by-page
1033 * basis when it encounters pages undergoing async
1034 * I/O.
1035 */
1036 if (object->type == OBJT_VNODE &&
1037 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1038 vp = object->handle;
1039 VM_OBJECT_WUNLOCK(object);
1040 (void) vn_start_write(vp, &mp, V_WAIT);
1041 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1042 if (syncio && !invalidate && offset == 0 &&
1043 atop(size) == object->size) {
1044 /*
1045 * If syncing the whole mapping of the file,
1046 * it is faster to schedule all the writes in
1047 * async mode, also allowing the clustering,
1048 * and then wait for i/o to complete.
1049 */
1050 flags = 0;
1051 fsync_after = TRUE;
1052 } else {
1053 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1054 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1055 fsync_after = FALSE;
1056 }
1057 VM_OBJECT_WLOCK(object);
1058 res = vm_object_page_clean(object, offset, offset + size,
1059 flags);
1060 VM_OBJECT_WUNLOCK(object);
1061 if (fsync_after)
1062 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1063 VOP_UNLOCK(vp, 0);
1064 vn_finished_write(mp);
1065 if (error != 0)
1066 res = FALSE;
1067 VM_OBJECT_WLOCK(object);
1068 }
1069 if ((object->type == OBJT_VNODE ||
1070 object->type == OBJT_DEVICE) && invalidate) {
1071 if (object->type == OBJT_DEVICE)
1072 /*
1073 * The option OBJPR_NOTMAPPED must be passed here
1074 * because vm_object_page_remove() cannot remove
1075 * unmanaged mappings.
1076 */
1077 flags = OBJPR_NOTMAPPED;
1078 else if (old_msync)
1079 flags = 0;
1080 else
1081 flags = OBJPR_CLEANONLY;
1082 vm_object_page_remove(object, OFF_TO_IDX(offset),
1083 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1084 }
1085 VM_OBJECT_WUNLOCK(object);
1086 return (res);
1087}
1088
1089/*
1090 * Determine whether the given advice can be applied to the object. Advice is
1091 * not applied to unmanaged pages since they never belong to page queues, and
1092 * since MADV_FREE is destructive, it can apply only to anonymous pages that
1093 * have been mapped at most once.
1094 */
1095static bool
1096vm_object_advice_applies(vm_object_t object, int advice)
1097{
1098
1099 if ((object->flags & OBJ_UNMANAGED) != 0)
1100 return (false);
1101 if (advice != MADV_FREE)
1102 return (true);
1103 return ((object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) &&
1104 (object->flags & OBJ_ONEMAPPING) != 0);
1105}
1106
1107static void
1108vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1109 vm_size_t size)
1110{
1111
1112 if (advice == MADV_FREE && object->type == OBJT_SWAP)
1113 swap_pager_freespace(object, pindex, size);
1114}
1115
1116/*
1117 * vm_object_madvise:
1118 *
1119 * Implements the madvise function at the object/page level.
1120 *
1121 * MADV_WILLNEED (any object)
1122 *
1123 * Activate the specified pages if they are resident.
1124 *
1125 * MADV_DONTNEED (any object)
1126 *
1127 * Deactivate the specified pages if they are resident.
1128 *
1129 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1130 * OBJ_ONEMAPPING only)
1131 *
1132 * Deactivate and clean the specified pages if they are
1133 * resident. This permits the process to reuse the pages
1134 * without faulting or the kernel to reclaim the pages
1135 * without I/O.
1136 */
1137void
1138vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1139 int advice)
1140{
1141 vm_pindex_t tpindex;
1142 vm_object_t backing_object, tobject;
1143 vm_page_t m, tm;
1144
1145 if (object == NULL)
1146 return;
1147
1148relookup:
1149 VM_OBJECT_WLOCK(object);
1150 if (!vm_object_advice_applies(object, advice)) {
1151 VM_OBJECT_WUNLOCK(object);
1152 return;
1153 }
1154 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1155 tobject = object;
1156
1157 /*
1158 * If the next page isn't resident in the top-level object, we
1159 * need to search the shadow chain. When applying MADV_FREE, we
1160 * take care to release any swap space used to store
1161 * non-resident pages.
1162 */
1163 if (m == NULL || pindex < m->pindex) {
1164 /*
1165 * Optimize a common case: if the top-level object has
1166 * no backing object, we can skip over the non-resident
1167 * range in constant time.
1168 */
1169 if (object->backing_object == NULL) {
1170 tpindex = (m != NULL && m->pindex < end) ?
1171 m->pindex : end;
1172 vm_object_madvise_freespace(object, advice,
1173 pindex, tpindex - pindex);
1174 if ((pindex = tpindex) == end)
1175 break;
1176 goto next_page;
1177 }
1178
1179 tpindex = pindex;
1180 do {
1181 vm_object_madvise_freespace(tobject, advice,
1182 tpindex, 1);
1183 /*
1184 * Prepare to search the next object in the
1185 * chain.
1186 */
1187 backing_object = tobject->backing_object;
1188 if (backing_object == NULL)
1189 goto next_pindex;
1190 VM_OBJECT_WLOCK(backing_object);
1191 tpindex +=
1192 OFF_TO_IDX(tobject->backing_object_offset);
1193 if (tobject != object)
1194 VM_OBJECT_WUNLOCK(tobject);
1195 tobject = backing_object;
1196 if (!vm_object_advice_applies(tobject, advice))
1197 goto next_pindex;
1198 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
1199 NULL);
1200 } else {
1201next_page:
1202 tm = m;
1203 m = TAILQ_NEXT(m, listq);
1204 }
1205
1206 /*
1207 * If the page is not in a normal state, skip it.
1208 */
1209 if (tm->valid != VM_PAGE_BITS_ALL)
1210 goto next_pindex;
1211 vm_page_lock(tm);
1212 if (tm->hold_count != 0 || tm->wire_count != 0) {
1213 vm_page_unlock(tm);
1214 goto next_pindex;
1215 }
1216 KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1217 ("vm_object_madvise: page %p is fictitious", tm));
1218 KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1219 ("vm_object_madvise: page %p is not managed", tm));
1220 if (vm_page_busied(tm)) {
1221 if (object != tobject)
1222 VM_OBJECT_WUNLOCK(tobject);
1223 VM_OBJECT_WUNLOCK(object);
1224 if (advice == MADV_WILLNEED) {
1225 /*
1226 * Reference the page before unlocking and
1227 * sleeping so that the page daemon is less
1228 * likely to reclaim it.
1229 */
1230 vm_page_aflag_set(tm, PGA_REFERENCED);
1231 }
1232 vm_page_busy_sleep(tm, "madvpo", false);
1233 goto relookup;
1234 }
1235 vm_page_advise(tm, advice);
1236 vm_page_unlock(tm);
1237 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1238next_pindex:
1239 if (tobject != object)
1240 VM_OBJECT_WUNLOCK(tobject);
1241 }
1242 VM_OBJECT_WUNLOCK(object);
1243}
1244
1245/*
1246 * vm_object_shadow:
1247 *
1248 * Create a new object which is backed by the
1249 * specified existing object range. The source
1250 * object reference is deallocated.
1251 *
1252 * The new object and offset into that object
1253 * are returned in the source parameters.
1254 */
1255void
1256vm_object_shadow(
1257 vm_object_t *object, /* IN/OUT */
1258 vm_ooffset_t *offset, /* IN/OUT */
1259 vm_size_t length)
1260{
1261 vm_object_t source;
1262 vm_object_t result;
1263
1264 source = *object;
1265
1266 /*
1267 * Don't create the new object if the old object isn't shared.
1268 */
1269 if (source != NULL) {
1270 VM_OBJECT_WLOCK(source);
1271 if (source->ref_count == 1 &&
1272 source->handle == NULL &&
1273 (source->type == OBJT_DEFAULT ||
1274 source->type == OBJT_SWAP)) {
1275 VM_OBJECT_WUNLOCK(source);
1276 return;
1277 }
1278 VM_OBJECT_WUNLOCK(source);
1279 }
1280
1281 /*
1282 * Allocate a new object with the given length.
1283 */
1284 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1285
1286 /*
1287 * The new object shadows the source object, adding a reference to it.
1288 * Our caller changes his reference to point to the new object,
1289 * removing a reference to the source object. Net result: no change
1290 * of reference count.
1291 *
1292 * Try to optimize the result object's page color when shadowing
1293 * in order to maintain page coloring consistency in the combined
1294 * shadowed object.
1295 */
1296 result->backing_object = source;
1297 /*
1298 * Store the offset into the source object, and fix up the offset into
1299 * the new object.
1300 */
1301 result->backing_object_offset = *offset;
1302 if (source != NULL) {
1303 VM_OBJECT_WLOCK(source);
1304 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1305 source->shadow_count++;
1306#if VM_NRESERVLEVEL > 0
1307 result->flags |= source->flags & OBJ_COLORED;
1308 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1309 ((1 << (VM_NFREEORDER - 1)) - 1);
1310#endif
1311 VM_OBJECT_WUNLOCK(source);
1312 }
1313
1314
1315 /*
1316 * Return the new things
1317 */
1318 *offset = 0;
1319 *object = result;
1320}
1321
1322/*
1323 * vm_object_split:
1324 *
1325 * Split the pages in a map entry into a new object. This affords
1326 * easier removal of unused pages, and keeps object inheritance from
1327 * being a negative impact on memory usage.
1328 */
1329void
1330vm_object_split(vm_map_entry_t entry)
1331{
1332 vm_page_t m, m_next;
1333 vm_object_t orig_object, new_object, source;
1334 vm_pindex_t idx, offidxstart;
1335 vm_size_t size;
1336
1337 orig_object = entry->object.vm_object;
1338 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1339 return;
1340 if (orig_object->ref_count <= 1)
1341 return;
1342 VM_OBJECT_WUNLOCK(orig_object);
1343
1344 offidxstart = OFF_TO_IDX(entry->offset);
1345 size = atop(entry->end - entry->start);
1346
1347 /*
1348 * If swap_pager_copy() is later called, it will convert new_object
1349 * into a swap object.
1350 */
1351 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1352
1353 /*
1354 * At this point, the new object is still private, so the order in
1355 * which the original and new objects are locked does not matter.
1356 */
1357 VM_OBJECT_WLOCK(new_object);
1358 VM_OBJECT_WLOCK(orig_object);
1359 source = orig_object->backing_object;
1360 if (source != NULL) {
1361 VM_OBJECT_WLOCK(source);
1362 if ((source->flags & OBJ_DEAD) != 0) {
1363 VM_OBJECT_WUNLOCK(source);
1364 VM_OBJECT_WUNLOCK(orig_object);
1365 VM_OBJECT_WUNLOCK(new_object);
1366 vm_object_deallocate(new_object);
1367 VM_OBJECT_WLOCK(orig_object);
1368 return;
1369 }
1370 LIST_INSERT_HEAD(&source->shadow_head,
1371 new_object, shadow_list);
1372 source->shadow_count++;
1373 vm_object_reference_locked(source); /* for new_object */
1374 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1375 VM_OBJECT_WUNLOCK(source);
1376 new_object->backing_object_offset =
1377 orig_object->backing_object_offset + entry->offset;
1378 new_object->backing_object = source;
1379 }
1380 if (orig_object->cred != NULL) {
1381 new_object->cred = orig_object->cred;
1382 crhold(orig_object->cred);
1383 new_object->charge = ptoa(size);
1384 KASSERT(orig_object->charge >= ptoa(size),
1385 ("orig_object->charge < 0"));
1386 orig_object->charge -= ptoa(size);
1387 }
1388retry:
1389 m = vm_page_find_least(orig_object, offidxstart);
1390 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1391 m = m_next) {
1392 m_next = TAILQ_NEXT(m, listq);
1393
1394 /*
1395 * We must wait for pending I/O to complete before we can
1396 * rename the page.
1397 *
1398 * We do not have to VM_PROT_NONE the page as mappings should
1399 * not be changed by this operation.
1400 */
1401 if (vm_page_busied(m)) {
1402 VM_OBJECT_WUNLOCK(new_object);
1403 vm_page_lock(m);
1404 VM_OBJECT_WUNLOCK(orig_object);
1405 vm_page_busy_sleep(m, "spltwt", false);
1406 VM_OBJECT_WLOCK(orig_object);
1407 VM_OBJECT_WLOCK(new_object);
1408 goto retry;
1409 }
1410
1411 /* vm_page_rename() will dirty the page. */
1412 if (vm_page_rename(m, new_object, idx)) {
1413 VM_OBJECT_WUNLOCK(new_object);
1414 VM_OBJECT_WUNLOCK(orig_object);
1415 VM_WAIT;
1416 VM_OBJECT_WLOCK(orig_object);
1417 VM_OBJECT_WLOCK(new_object);
1418 goto retry;
1419 }
1420#if VM_NRESERVLEVEL > 0
1421 /*
1422 * If some of the reservation's allocated pages remain with
1423 * the original object, then transferring the reservation to
1424 * the new object is neither particularly beneficial nor
1425 * particularly harmful as compared to leaving the reservation
1426 * with the original object. If, however, all of the
1427 * reservation's allocated pages are transferred to the new
1428 * object, then transferring the reservation is typically
1429 * beneficial. Determining which of these two cases applies
1430 * would be more costly than unconditionally renaming the
1431 * reservation.
1432 */
1433 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1434#endif
1435 if (orig_object->type == OBJT_SWAP)
1436 vm_page_xbusy(m);
1437 }
1438 if (orig_object->type == OBJT_SWAP) {
1439 /*
1440 * swap_pager_copy() can sleep, in which case the orig_object's
1441 * and new_object's locks are released and reacquired.
1442 */
1443 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1444 TAILQ_FOREACH(m, &new_object->memq, listq)
1445 vm_page_xunbusy(m);
1446 }
1447 VM_OBJECT_WUNLOCK(orig_object);
1448 VM_OBJECT_WUNLOCK(new_object);
1449 entry->object.vm_object = new_object;
1450 entry->offset = 0LL;
1451 vm_object_deallocate(orig_object);
1452 VM_OBJECT_WLOCK(new_object);
1453}
1454
1455#define OBSC_COLLAPSE_NOWAIT 0x0002
1456#define OBSC_COLLAPSE_WAIT 0x0004
1457
1458static vm_page_t
1459vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next,
1460 int op)
1461{
1462 vm_object_t backing_object;
1463
1464 VM_OBJECT_ASSERT_WLOCKED(object);
1465 backing_object = object->backing_object;
1466 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1467
1468 KASSERT(p == NULL || vm_page_busied(p), ("unbusy page %p", p));
1469 KASSERT(p == NULL || p->object == object || p->object == backing_object,
1470 ("invalid ownership %p %p %p", p, object, backing_object));
1471 if ((op & OBSC_COLLAPSE_NOWAIT) != 0)
1472 return (next);
1473 if (p != NULL)
1474 vm_page_lock(p);
1475 VM_OBJECT_WUNLOCK(object);
1476 VM_OBJECT_WUNLOCK(backing_object);
1477 if (p == NULL)
1478 VM_WAIT;
1479 else
1480 vm_page_busy_sleep(p, "vmocol", false);
1481 VM_OBJECT_WLOCK(object);
1482 VM_OBJECT_WLOCK(backing_object);
1483 return (TAILQ_FIRST(&backing_object->memq));
1484}
1485
1486static bool
1487vm_object_scan_all_shadowed(vm_object_t object)
1488{
1489 vm_object_t backing_object;
1490 vm_page_t p, pp;
1491 vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1492
1493 VM_OBJECT_ASSERT_WLOCKED(object);
1494 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1495
1496 backing_object = object->backing_object;
1497
1498 /*
1499 * Initial conditions:
1500 *
1501 * We do not want to have to test for the existence of swap
1502 * pages in the backing object. XXX but with the new swapper this
1503 * would be pretty easy to do.
1504 */
1505 if (backing_object->type != OBJT_DEFAULT &&
1506 backing_object->type != OBJT_SWAP)
1507 return (false);
1508
1509 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1510 p = vm_page_find_least(backing_object, pi);
1511 ps = swap_pager_find_least(backing_object, pi);
1512
1513 /*
1514 * Only check pages inside the parent object's range and
1515 * inside the parent object's mapping of the backing object.
1516 */
1517 for (;; pi++) {
1518 if (p != NULL && p->pindex < pi)
1519 p = TAILQ_NEXT(p, listq);
1520 if (ps < pi)
1521 ps = swap_pager_find_least(backing_object, pi);
1522 if (p == NULL && ps >= backing_object->size)
1523 break;
1524 else if (p == NULL)
1525 pi = ps;
1526 else
1527 pi = MIN(p->pindex, ps);
1528
1529 new_pindex = pi - backing_offset_index;
1530 if (new_pindex >= object->size)
1531 break;
1532
1533 /*
1534 * See if the parent has the page or if the parent's object
1535 * pager has the page. If the parent has the page but the page
1536 * is not valid, the parent's object pager must have the page.
1537 *
1538 * If this fails, the parent does not completely shadow the
1539 * object and we might as well give up now.
1540 */
1541 pp = vm_page_lookup(object, new_pindex);
1542 if ((pp == NULL || pp->valid == 0) &&
1543 !vm_pager_has_page(object, new_pindex, NULL, NULL))
1544 return (false);
1545 }
1546 return (true);
1547}
1548
1549static bool
1550vm_object_collapse_scan(vm_object_t object, int op)
1551{
1552 vm_object_t backing_object;
1553 vm_page_t next, p, pp;
1554 vm_pindex_t backing_offset_index, new_pindex;
1555
1556 VM_OBJECT_ASSERT_WLOCKED(object);
1557 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1558
1559 backing_object = object->backing_object;
1560 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1561
1562 /*
1563 * Initial conditions
1564 */
1565 if ((op & OBSC_COLLAPSE_WAIT) != 0)
1566 vm_object_set_flag(backing_object, OBJ_DEAD);
1567
1568 /*
1569 * Our scan
1570 */
1571 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1572 next = TAILQ_NEXT(p, listq);
1573 new_pindex = p->pindex - backing_offset_index;
1574
1575 /*
1576 * Check for busy page
1577 */
1578 if (vm_page_busied(p)) {
1579 next = vm_object_collapse_scan_wait(object, p, next, op);
1580 continue;
1581 }
1582
1583 KASSERT(p->object == backing_object,
1584 ("vm_object_collapse_scan: object mismatch"));
1585
1586 if (p->pindex < backing_offset_index ||
1587 new_pindex >= object->size) {
1588 if (backing_object->type == OBJT_SWAP)
1589 swap_pager_freespace(backing_object, p->pindex,
1590 1);
1591
1592 /*
1593 * Page is out of the parent object's range, we can
1594 * simply destroy it.
1595 */
1596 vm_page_lock(p);
1597 KASSERT(!pmap_page_is_mapped(p),
1598 ("freeing mapped page %p", p));
1599 if (p->wire_count == 0)
1600 vm_page_free(p);
1601 else
1602 vm_page_remove(p);
1603 vm_page_unlock(p);
1604 continue;
1605 }
1606
1607 pp = vm_page_lookup(object, new_pindex);
1608 if (pp != NULL && vm_page_busied(pp)) {
1609 /*
1610 * The page in the parent is busy and possibly not
1611 * (yet) valid. Until its state is finalized by the
1612 * busy bit owner, we can't tell whether it shadows the
1613 * original page. Therefore, we must either skip it
1614 * and the original (backing_object) page or wait for
1615 * its state to be finalized.
1616 *
1617 * This is due to a race with vm_fault() where we must
1618 * unbusy the original (backing_obj) page before we can
1619 * (re)lock the parent. Hence we can get here.
1620 */
1621 next = vm_object_collapse_scan_wait(object, pp, next,
1622 op);
1623 continue;
1624 }
1625
1626 KASSERT(pp == NULL || pp->valid != 0,
1627 ("unbusy invalid page %p", pp));
1628
1629 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1630 NULL)) {
1631 /*
1632 * The page already exists in the parent OR swap exists
1633 * for this location in the parent. Leave the parent's
1634 * page alone. Destroy the original page from the
1635 * backing object.
1636 */
1637 if (backing_object->type == OBJT_SWAP)
1638 swap_pager_freespace(backing_object, p->pindex,
1639 1);
1640 vm_page_lock(p);
1641 KASSERT(!pmap_page_is_mapped(p),
1642 ("freeing mapped page %p", p));
1643 if (p->wire_count == 0)
1644 vm_page_free(p);
1645 else
1646 vm_page_remove(p);
1647 vm_page_unlock(p);
1648 continue;
1649 }
1650
1651 /*
1652 * Page does not exist in parent, rename the page from the
1653 * backing object to the main object.
1654 *
1655 * If the page was mapped to a process, it can remain mapped
1656 * through the rename. vm_page_rename() will dirty the page.
1657 */
1658 if (vm_page_rename(p, object, new_pindex)) {
1659 next = vm_object_collapse_scan_wait(object, NULL, next,
1660 op);
1661 continue;
1662 }
1663
1664 /* Use the old pindex to free the right page. */
1665 if (backing_object->type == OBJT_SWAP)
1666 swap_pager_freespace(backing_object,
1667 new_pindex + backing_offset_index, 1);
1668
1669#if VM_NRESERVLEVEL > 0
1670 /*
1671 * Rename the reservation.
1672 */
1673 vm_reserv_rename(p, object, backing_object,
1674 backing_offset_index);
1675#endif
1676 }
1677 return (true);
1678}
1679
1680
1681/*
1682 * this version of collapse allows the operation to occur earlier and
1683 * when paging_in_progress is true for an object... This is not a complete
1684 * operation, but should plug 99.9% of the rest of the leaks.
1685 */
1686static void
1687vm_object_qcollapse(vm_object_t object)
1688{
1689 vm_object_t backing_object = object->backing_object;
1690
1691 VM_OBJECT_ASSERT_WLOCKED(object);
1692 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1693
1694 if (backing_object->ref_count != 1)
1695 return;
1696
1697 vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT);
1698}
1699
1700/*
1701 * vm_object_collapse:
1702 *
1703 * Collapse an object with the object backing it.
1704 * Pages in the backing object are moved into the
1705 * parent, and the backing object is deallocated.
1706 */
1707void
1708vm_object_collapse(vm_object_t object)
1709{
1710 vm_object_t backing_object, new_backing_object;
1711
1712 VM_OBJECT_ASSERT_WLOCKED(object);
1713
1714 while (TRUE) {
1715 /*
1716 * Verify that the conditions are right for collapse:
1717 *
1718 * The object exists and the backing object exists.
1719 */
1720 if ((backing_object = object->backing_object) == NULL)
1721 break;
1722
1723 /*
1724 * we check the backing object first, because it is most likely
1725 * not collapsable.
1726 */
1727 VM_OBJECT_WLOCK(backing_object);
1728 if (backing_object->handle != NULL ||
1729 (backing_object->type != OBJT_DEFAULT &&
1730 backing_object->type != OBJT_SWAP) ||
1731 (backing_object->flags & OBJ_DEAD) ||
1732 object->handle != NULL ||
1733 (object->type != OBJT_DEFAULT &&
1734 object->type != OBJT_SWAP) ||
1735 (object->flags & OBJ_DEAD)) {
1736 VM_OBJECT_WUNLOCK(backing_object);
1737 break;
1738 }
1739
1740 if (object->paging_in_progress != 0 ||
1741 backing_object->paging_in_progress != 0) {
1742 vm_object_qcollapse(object);
1743 VM_OBJECT_WUNLOCK(backing_object);
1744 break;
1745 }
1746
1747 /*
1748 * We know that we can either collapse the backing object (if
1749 * the parent is the only reference to it) or (perhaps) have
1750 * the parent bypass the object if the parent happens to shadow
1751 * all the resident pages in the entire backing object.
1752 *
1753 * This is ignoring pager-backed pages such as swap pages.
1754 * vm_object_collapse_scan fails the shadowing test in this
1755 * case.
1756 */
1757 if (backing_object->ref_count == 1) {
1758 vm_object_pip_add(object, 1);
1759 vm_object_pip_add(backing_object, 1);
1760
1761 /*
1762 * If there is exactly one reference to the backing
1763 * object, we can collapse it into the parent.
1764 */
1765 vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT);
1766
1767#if VM_NRESERVLEVEL > 0
1768 /*
1769 * Break any reservations from backing_object.
1770 */
1771 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1772 vm_reserv_break_all(backing_object);
1773#endif
1774
1775 /*
1776 * Move the pager from backing_object to object.
1777 */
1778 if (backing_object->type == OBJT_SWAP) {
1779 /*
1780 * swap_pager_copy() can sleep, in which case
1781 * the backing_object's and object's locks are
1782 * released and reacquired.
1783 * Since swap_pager_copy() is being asked to
1784 * destroy the source, it will change the
1785 * backing_object's type to OBJT_DEFAULT.
1786 */
1787 swap_pager_copy(
1788 backing_object,
1789 object,
1790 OFF_TO_IDX(object->backing_object_offset), TRUE);
1791 }
1792 /*
1793 * Object now shadows whatever backing_object did.
1794 * Note that the reference to
1795 * backing_object->backing_object moves from within
1796 * backing_object to within object.
1797 */
1798 LIST_REMOVE(object, shadow_list);
1799 backing_object->shadow_count--;
1800 if (backing_object->backing_object) {
1801 VM_OBJECT_WLOCK(backing_object->backing_object);
1802 LIST_REMOVE(backing_object, shadow_list);
1803 LIST_INSERT_HEAD(
1804 &backing_object->backing_object->shadow_head,
1805 object, shadow_list);
1806 /*
1807 * The shadow_count has not changed.
1808 */
1809 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1810 }
1811 object->backing_object = backing_object->backing_object;
1812 object->backing_object_offset +=
1813 backing_object->backing_object_offset;
1814
1815 /*
1816 * Discard backing_object.
1817 *
1818 * Since the backing object has no pages, no pager left,
1819 * and no object references within it, all that is
1820 * necessary is to dispose of it.
1821 */
1822 KASSERT(backing_object->ref_count == 1, (
1823"backing_object %p was somehow re-referenced during collapse!",
1824 backing_object));
1825 vm_object_pip_wakeup(backing_object);
1826 backing_object->type = OBJT_DEAD;
1827 backing_object->ref_count = 0;
1828 VM_OBJECT_WUNLOCK(backing_object);
1829 vm_object_destroy(backing_object);
1830
1831 vm_object_pip_wakeup(object);
1832 object_collapses++;
1833 } else {
1834 /*
1835 * If we do not entirely shadow the backing object,
1836 * there is nothing we can do so we give up.
1837 */
1838 if (object->resident_page_count != object->size &&
1839 !vm_object_scan_all_shadowed(object)) {
1840 VM_OBJECT_WUNLOCK(backing_object);
1841 break;
1842 }
1843
1844 /*
1845 * Make the parent shadow the next object in the
1846 * chain. Deallocating backing_object will not remove
1847 * it, since its reference count is at least 2.
1848 */
1849 LIST_REMOVE(object, shadow_list);
1850 backing_object->shadow_count--;
1851
1852 new_backing_object = backing_object->backing_object;
1853 if ((object->backing_object = new_backing_object) != NULL) {
1854 VM_OBJECT_WLOCK(new_backing_object);
1855 LIST_INSERT_HEAD(
1856 &new_backing_object->shadow_head,
1857 object,
1858 shadow_list
1859 );
1860 new_backing_object->shadow_count++;
1861 vm_object_reference_locked(new_backing_object);
1862 VM_OBJECT_WUNLOCK(new_backing_object);
1863 object->backing_object_offset +=
1864 backing_object->backing_object_offset;
1865 }
1866
1867 /*
1868 * Drop the reference count on backing_object. Since
1869 * its ref_count was at least 2, it will not vanish.
1870 */
1871 backing_object->ref_count--;
1872 VM_OBJECT_WUNLOCK(backing_object);
1873 object_bypasses++;
1874 }
1875
1876 /*
1877 * Try again with this object's new backing object.
1878 */
1879 }
1880}
1881
1882/*
1883 * vm_object_page_remove:
1884 *
1885 * For the given object, either frees or invalidates each of the
1886 * specified pages. In general, a page is freed. However, if a page is
1887 * wired for any reason other than the existence of a managed, wired
1888 * mapping, then it may be invalidated but not removed from the object.
1889 * Pages are specified by the given range ["start", "end") and the option
1890 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1891 * extends from "start" to the end of the object. If the option
1892 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1893 * specified range are affected. If the option OBJPR_NOTMAPPED is
1894 * specified, then the pages within the specified range must have no
1895 * mappings. Otherwise, if this option is not specified, any mappings to
1896 * the specified pages are removed before the pages are freed or
1897 * invalidated.
1898 *
1899 * In general, this operation should only be performed on objects that
1900 * contain managed pages. There are, however, two exceptions. First, it
1901 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1902 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1903 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1904 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1905 *
1906 * The object must be locked.
1907 */
1908void
1909vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1910 int options)
1911{
1912 vm_page_t p, next;
1913
1914 VM_OBJECT_ASSERT_WLOCKED(object);
1915 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1916 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1917 ("vm_object_page_remove: illegal options for object %p", object));
1918 if (object->resident_page_count == 0)
1919 return;
1920 vm_object_pip_add(object, 1);
1921again:
1922 p = vm_page_find_least(object, start);
1923
1924 /*
1925 * Here, the variable "p" is either (1) the page with the least pindex
1926 * greater than or equal to the parameter "start" or (2) NULL.
1927 */
1928 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1929 next = TAILQ_NEXT(p, listq);
1930
1931 /*
1932 * If the page is wired for any reason besides the existence
1933 * of managed, wired mappings, then it cannot be freed. For
1934 * example, fictitious pages, which represent device memory,
1935 * are inherently wired and cannot be freed. They can,
1936 * however, be invalidated if the option OBJPR_CLEANONLY is
1937 * not specified.
1938 */
1939 vm_page_lock(p);
1940 if (vm_page_xbusied(p)) {
1941 VM_OBJECT_WUNLOCK(object);
1942 vm_page_busy_sleep(p, "vmopax", true);
1943 VM_OBJECT_WLOCK(object);
1944 goto again;
1945 }
1946 if (p->wire_count != 0) {
1947 if ((options & OBJPR_NOTMAPPED) == 0)
1948 pmap_remove_all(p);
1949 if ((options & OBJPR_CLEANONLY) == 0) {
1950 p->valid = 0;
1951 vm_page_undirty(p);
1952 }
1953 goto next;
1954 }
1955 if (vm_page_busied(p)) {
1956 VM_OBJECT_WUNLOCK(object);
1957 vm_page_busy_sleep(p, "vmopar", false);
1958 VM_OBJECT_WLOCK(object);
1959 goto again;
1960 }
1961 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1962 ("vm_object_page_remove: page %p is fictitious", p));
1963 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1964 if ((options & OBJPR_NOTMAPPED) == 0)
1965 pmap_remove_write(p);
1966 if (p->dirty)
1967 goto next;
1968 }
1969 if ((options & OBJPR_NOTMAPPED) == 0)
1970 pmap_remove_all(p);
1971 vm_page_free(p);
1972next:
1973 vm_page_unlock(p);
1974 }
1975 vm_object_pip_wakeup(object);
1976}
1977
1978/*
1979 * vm_object_page_noreuse:
1980 *
1981 * For the given object, attempt to move the specified pages to
1982 * the head of the inactive queue. This bypasses regular LRU
1983 * operation and allows the pages to be reused quickly under memory
1984 * pressure. If a page is wired for any reason, then it will not
1985 * be queued. Pages are specified by the range ["start", "end").
1986 * As a special case, if "end" is zero, then the range extends from
1987 * "start" to the end of the object.
1988 *
1989 * This operation should only be performed on objects that
1990 * contain non-fictitious, managed pages.
1991 *
1992 * The object must be locked.
1993 */
1994void
1995vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1996{
1997 struct mtx *mtx, *new_mtx;
1998 vm_page_t p, next;
1999
2000 VM_OBJECT_ASSERT_LOCKED(object);
2001 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2002 ("vm_object_page_noreuse: illegal object %p", object));
2003 if (object->resident_page_count == 0)
2004 return;
2005 p = vm_page_find_least(object, start);
2006
2007 /*
2008 * Here, the variable "p" is either (1) the page with the least pindex
2009 * greater than or equal to the parameter "start" or (2) NULL.
2010 */
2011 mtx = NULL;
2012 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2013 next = TAILQ_NEXT(p, listq);
2014
2015 /*
2016 * Avoid releasing and reacquiring the same page lock.
2017 */
2018 new_mtx = vm_page_lockptr(p);
2019 if (mtx != new_mtx) {
2020 if (mtx != NULL)
2021 mtx_unlock(mtx);
2022 mtx = new_mtx;
2023 mtx_lock(mtx);
2024 }
2025 vm_page_deactivate_noreuse(p);
2026 }
2027 if (mtx != NULL)
2028 mtx_unlock(mtx);
2029}
2030
2031/*
2032 * Populate the specified range of the object with valid pages. Returns
2033 * TRUE if the range is successfully populated and FALSE otherwise.
2034 *
2035 * Note: This function should be optimized to pass a larger array of
2036 * pages to vm_pager_get_pages() before it is applied to a non-
2037 * OBJT_DEVICE object.
2038 *
2039 * The object must be locked.
2040 */
2041boolean_t
2042vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2043{
2044 vm_page_t m;
2045 vm_pindex_t pindex;
2046 int rv;
2047
2048 VM_OBJECT_ASSERT_WLOCKED(object);
2049 for (pindex = start; pindex < end; pindex++) {
2050 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
2051 if (m->valid != VM_PAGE_BITS_ALL) {
2052 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
2053 if (rv != VM_PAGER_OK) {
2054 vm_page_lock(m);
2055 vm_page_free(m);
2056 vm_page_unlock(m);
2057 break;
2058 }
2059 }
2060 /*
2061 * Keep "m" busy because a subsequent iteration may unlock
2062 * the object.
2063 */
2064 }
2065 if (pindex > start) {
2066 m = vm_page_lookup(object, start);
2067 while (m != NULL && m->pindex < pindex) {
2068 vm_page_xunbusy(m);
2069 m = TAILQ_NEXT(m, listq);
2070 }
2071 }
2072 return (pindex == end);
2073}
2074
2075/*
2076 * Routine: vm_object_coalesce
2077 * Function: Coalesces two objects backing up adjoining
2078 * regions of memory into a single object.
2079 *
2080 * returns TRUE if objects were combined.
2081 *
2082 * NOTE: Only works at the moment if the second object is NULL -
2083 * if it's not, which object do we lock first?
2084 *
2085 * Parameters:
2086 * prev_object First object to coalesce
2087 * prev_offset Offset into prev_object
2088 * prev_size Size of reference to prev_object
2089 * next_size Size of reference to the second object
2090 * reserved Indicator that extension region has
2091 * swap accounted for
2092 *
2093 * Conditions:
2094 * The object must *not* be locked.
2095 */
2096boolean_t
2097vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2098 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2099{
2100 vm_pindex_t next_pindex;
2101
2102 if (prev_object == NULL)
2103 return (TRUE);
2104 VM_OBJECT_WLOCK(prev_object);
2105 if ((prev_object->type != OBJT_DEFAULT &&
2106 prev_object->type != OBJT_SWAP) ||
2107 (prev_object->flags & OBJ_TMPFS_NODE) != 0) {
2108 VM_OBJECT_WUNLOCK(prev_object);
2109 return (FALSE);
2110 }
2111
2112 /*
2113 * Try to collapse the object first
2114 */
2115 vm_object_collapse(prev_object);
2116
2117 /*
2118 * Can't coalesce if: . more than one reference . paged out . shadows
2119 * another object . has a copy elsewhere (any of which mean that the
2120 * pages not mapped to prev_entry may be in use anyway)
2121 */
2122 if (prev_object->backing_object != NULL) {
2123 VM_OBJECT_WUNLOCK(prev_object);
2124 return (FALSE);
2125 }
2126
2127 prev_size >>= PAGE_SHIFT;
2128 next_size >>= PAGE_SHIFT;
2129 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2130
2131 if ((prev_object->ref_count > 1) &&
2132 (prev_object->size != next_pindex)) {
2133 VM_OBJECT_WUNLOCK(prev_object);
2134 return (FALSE);
2135 }
2136
2137 /*
2138 * Account for the charge.
2139 */
2140 if (prev_object->cred != NULL) {
2141
2142 /*
2143 * If prev_object was charged, then this mapping,
2144 * although not charged now, may become writable
2145 * later. Non-NULL cred in the object would prevent
2146 * swap reservation during enabling of the write
2147 * access, so reserve swap now. Failed reservation
2148 * cause allocation of the separate object for the map
2149 * entry, and swap reservation for this entry is
2150 * managed in appropriate time.
2151 */
2152 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2153 prev_object->cred)) {
2154 VM_OBJECT_WUNLOCK(prev_object);
2155 return (FALSE);
2156 }
2157 prev_object->charge += ptoa(next_size);
2158 }
2159
2160 /*
2161 * Remove any pages that may still be in the object from a previous
2162 * deallocation.
2163 */
2164 if (next_pindex < prev_object->size) {
2165 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2166 next_size, 0);
2167 if (prev_object->type == OBJT_SWAP)
2168 swap_pager_freespace(prev_object,
2169 next_pindex, next_size);
2170#if 0
2171 if (prev_object->cred != NULL) {
2172 KASSERT(prev_object->charge >=
2173 ptoa(prev_object->size - next_pindex),
2174 ("object %p overcharged 1 %jx %jx", prev_object,
2175 (uintmax_t)next_pindex, (uintmax_t)next_size));
2176 prev_object->charge -= ptoa(prev_object->size -
2177 next_pindex);
2178 }
2179#endif
2180 }
2181
2182 /*
2183 * Extend the object if necessary.
2184 */
2185 if (next_pindex + next_size > prev_object->size)
2186 prev_object->size = next_pindex + next_size;
2187
2188 VM_OBJECT_WUNLOCK(prev_object);
2189 return (TRUE);
2190}
2191
2192void
2193vm_object_set_writeable_dirty(vm_object_t object)
2194{
2195
2196 VM_OBJECT_ASSERT_WLOCKED(object);
2197 if (object->type != OBJT_VNODE) {
2198 if ((object->flags & OBJ_TMPFS_NODE) != 0) {
2199 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs"));
2200 vm_object_set_flag(object, OBJ_TMPFS_DIRTY);
2201 }
2202 return;
2203 }
2204 object->generation++;
2205 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2206 return;
2207 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2208}
2209
2210/*
2211 * vm_object_unwire:
2212 *
2213 * For each page offset within the specified range of the given object,
2214 * find the highest-level page in the shadow chain and unwire it. A page
2215 * must exist at every page offset, and the highest-level page must be
2216 * wired.
2217 */
2218void
2219vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2220 uint8_t queue)
2221{
2222 vm_object_t tobject;
2223 vm_page_t m, tm;
2224 vm_pindex_t end_pindex, pindex, tpindex;
2225 int depth, locked_depth;
2226
2227 KASSERT((offset & PAGE_MASK) == 0,
2228 ("vm_object_unwire: offset is not page aligned"));
2229 KASSERT((length & PAGE_MASK) == 0,
2230 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2231 /* The wired count of a fictitious page never changes. */
2232 if ((object->flags & OBJ_FICTITIOUS) != 0)
2233 return;
2234 pindex = OFF_TO_IDX(offset);
2235 end_pindex = pindex + atop(length);
2236 locked_depth = 1;
2237 VM_OBJECT_RLOCK(object);
2238 m = vm_page_find_least(object, pindex);
2239 while (pindex < end_pindex) {
2240 if (m == NULL || pindex < m->pindex) {
2241 /*
2242 * The first object in the shadow chain doesn't
2243 * contain a page at the current index. Therefore,
2244 * the page must exist in a backing object.
2245 */
2246 tobject = object;
2247 tpindex = pindex;
2248 depth = 0;
2249 do {
2250 tpindex +=
2251 OFF_TO_IDX(tobject->backing_object_offset);
2252 tobject = tobject->backing_object;
2253 KASSERT(tobject != NULL,
2254 ("vm_object_unwire: missing page"));
2255 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2256 goto next_page;
2257 depth++;
2258 if (depth == locked_depth) {
2259 locked_depth++;
2260 VM_OBJECT_RLOCK(tobject);
2261 }
2262 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2263 NULL);
2264 } else {
2265 tm = m;
2266 m = TAILQ_NEXT(m, listq);
2267 }
2268 vm_page_lock(tm);
2269 vm_page_unwire(tm, queue);
2270 vm_page_unlock(tm);
2271next_page:
2272 pindex++;
2273 }
2274 /* Release the accumulated object locks. */
2275 for (depth = 0; depth < locked_depth; depth++) {
2276 tobject = object->backing_object;
2277 VM_OBJECT_RUNLOCK(object);
2278 object = tobject;
2279 }
2280}
2281
2282struct vnode *
2283vm_object_vnode(vm_object_t object)
2284{
2285
2286 VM_OBJECT_ASSERT_LOCKED(object);
2287 if (object->type == OBJT_VNODE)
2288 return (object->handle);
2289 if (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS) != 0)
2290 return (object->un_pager.swp.swp_tmpfs);
2291 return (NULL);
2292}
2293
2294static int
2295sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2296{
2297 struct kinfo_vmobject *kvo;
2298 char *fullpath, *freepath;
2299 struct vnode *vp;
2300 struct vattr va;
2301 vm_object_t obj;
2302 vm_page_t m;
2303 int count, error;
2304
2305 if (req->oldptr == NULL) {
2306 /*
2307 * If an old buffer has not been provided, generate an
2308 * estimate of the space needed for a subsequent call.
2309 */
2310 mtx_lock(&vm_object_list_mtx);
2311 count = 0;
2312 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2313 if (obj->type == OBJT_DEAD)
2314 continue;
2315 count++;
2316 }
2317 mtx_unlock(&vm_object_list_mtx);
2318 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2319 count * 11 / 10));
2320 }
2321
2322 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK);
2323 error = 0;
2324
2325 /*
2326 * VM objects are type stable and are never removed from the
2327 * list once added. This allows us to safely read obj->object_list
2328 * after reacquiring the VM object lock.
2329 */
2330 mtx_lock(&vm_object_list_mtx);
2331 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2332 if (obj->type == OBJT_DEAD)
2333 continue;
2334 VM_OBJECT_RLOCK(obj);
2335 if (obj->type == OBJT_DEAD) {
2336 VM_OBJECT_RUNLOCK(obj);
2337 continue;
2338 }
2339 mtx_unlock(&vm_object_list_mtx);
2340 kvo->kvo_size = ptoa(obj->size);
2341 kvo->kvo_resident = obj->resident_page_count;
2342 kvo->kvo_ref_count = obj->ref_count;
2343 kvo->kvo_shadow_count = obj->shadow_count;
2344 kvo->kvo_memattr = obj->memattr;
2345 kvo->kvo_active = 0;
2346 kvo->kvo_inactive = 0;
2347 TAILQ_FOREACH(m, &obj->memq, listq) {
2348 /*
2349 * A page may belong to the object but be
2350 * dequeued and set to PQ_NONE while the
2351 * object lock is not held. This makes the
2352 * reads of m->queue below racy, and we do not
2353 * count pages set to PQ_NONE. However, this
2354 * sysctl is only meant to give an
2355 * approximation of the system anyway.
2356 */
2357 if (vm_page_active(m))
2358 kvo->kvo_active++;
2359 else if (vm_page_inactive(m))
2360 kvo->kvo_inactive++;
2361 }
2362
2363 kvo->kvo_vn_fileid = 0;
2364 kvo->kvo_vn_fsid = 0;
2365 freepath = NULL;
2366 fullpath = "";
2367 vp = NULL;
2368 switch (obj->type) {
2369 case OBJT_DEFAULT:
2370 kvo->kvo_type = KVME_TYPE_DEFAULT;
2371 break;
2372 case OBJT_VNODE:
2373 kvo->kvo_type = KVME_TYPE_VNODE;
2374 vp = obj->handle;
2375 vref(vp);
2376 break;
2377 case OBJT_SWAP:
2378 kvo->kvo_type = KVME_TYPE_SWAP;
2379 break;
2380 case OBJT_DEVICE:
2381 kvo->kvo_type = KVME_TYPE_DEVICE;
2382 break;
2383 case OBJT_PHYS:
2384 kvo->kvo_type = KVME_TYPE_PHYS;
2385 break;
2386 case OBJT_DEAD:
2387 kvo->kvo_type = KVME_TYPE_DEAD;
2388 break;
2389 case OBJT_SG:
2390 kvo->kvo_type = KVME_TYPE_SG;
2391 break;
2392 case OBJT_MGTDEVICE:
2393 kvo->kvo_type = KVME_TYPE_MGTDEVICE;
2394 break;
2395 default:
2396 kvo->kvo_type = KVME_TYPE_UNKNOWN;
2397 break;
2398 }
2399 VM_OBJECT_RUNLOCK(obj);
2400 if (vp != NULL) {
2401 vn_fullpath(curthread, vp, &fullpath, &freepath);
2402 vn_lock(vp, LK_SHARED | LK_RETRY);
2403 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2404 kvo->kvo_vn_fileid = va.va_fileid;
2405 kvo->kvo_vn_fsid = va.va_fsid;
2406 }
2407 vput(vp);
2408 }
2409
2410 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2411 if (freepath != NULL)
2412 free(freepath, M_TEMP);
2413
2414 /* Pack record size down */
2415 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2416 + strlen(kvo->kvo_path) + 1;
2417 kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2418 sizeof(uint64_t));
2419 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2420 mtx_lock(&vm_object_list_mtx);
2421 if (error)
2422 break;
2423 }
2424 mtx_unlock(&vm_object_list_mtx);
2425 free(kvo, M_TEMP);
2426 return (error);
2427}
2428SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2429 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2430 "List of VM objects");
2431
2432#include "opt_ddb.h"
2433#ifdef DDB
2434#include <sys/kernel.h>
2435
2436#include <sys/cons.h>
2437
2438#include <ddb/ddb.h>
2439
2440static int
2441_vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2442{
2443 vm_map_t tmpm;
2444 vm_map_entry_t tmpe;
2445 vm_object_t obj;
2446 int entcount;
2447
2448 if (map == 0)
2449 return 0;
2450
2451 if (entry == 0) {
2452 tmpe = map->header.next;
2453 entcount = map->nentries;
2454 while (entcount-- && (tmpe != &map->header)) {
2455 if (_vm_object_in_map(map, object, tmpe)) {
2456 return 1;
2457 }
2458 tmpe = tmpe->next;
2459 }
2460 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2461 tmpm = entry->object.sub_map;
2462 tmpe = tmpm->header.next;
2463 entcount = tmpm->nentries;
2464 while (entcount-- && tmpe != &tmpm->header) {
2465 if (_vm_object_in_map(tmpm, object, tmpe)) {
2466 return 1;
2467 }
2468 tmpe = tmpe->next;
2469 }
2470 } else if ((obj = entry->object.vm_object) != NULL) {
2471 for (; obj; obj = obj->backing_object)
2472 if (obj == object) {
2473 return 1;
2474 }
2475 }
2476 return 0;
2477}
2478
2479static int
2480vm_object_in_map(vm_object_t object)
2481{
2482 struct proc *p;
2483
2484 /* sx_slock(&allproc_lock); */
2485 FOREACH_PROC_IN_SYSTEM(p) {
2486 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2487 continue;
2488 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2489 /* sx_sunlock(&allproc_lock); */
2490 return 1;
2491 }
2492 }
2493 /* sx_sunlock(&allproc_lock); */
2494 if (_vm_object_in_map(kernel_map, object, 0))
2495 return 1;
2496 return 0;
2497}
2498
2499DB_SHOW_COMMAND(vmochk, vm_object_check)
2500{
2501 vm_object_t object;
2502
2503 /*
2504 * make sure that internal objs are in a map somewhere
2505 * and none have zero ref counts.
2506 */
2507 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2508 if (object->handle == NULL &&
2509 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2510 if (object->ref_count == 0) {
2511 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2512 (long)object->size);
2513 }
2514 if (!vm_object_in_map(object)) {
2515 db_printf(
2516 "vmochk: internal obj is not in a map: "
2517 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2518 object->ref_count, (u_long)object->size,
2519 (u_long)object->size,
2520 (void *)object->backing_object);
2521 }
2522 }
2523 }
2524}
2525
2526/*
2527 * vm_object_print: [ debug ]
2528 */
2529DB_SHOW_COMMAND(object, vm_object_print_static)
2530{
2531 /* XXX convert args. */
2532 vm_object_t object = (vm_object_t)addr;
2533 boolean_t full = have_addr;
2534
2535 vm_page_t p;
2536
2537 /* XXX count is an (unused) arg. Avoid shadowing it. */
2538#define count was_count
2539
2540 int count;
2541
2542 if (object == NULL)
2543 return;
2544
2545 db_iprintf(
2546 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2547 object, (int)object->type, (uintmax_t)object->size,
2548 object->resident_page_count, object->ref_count, object->flags,
2549 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2550 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2551 object->shadow_count,
2552 object->backing_object ? object->backing_object->ref_count : 0,
2553 object->backing_object, (uintmax_t)object->backing_object_offset);
2554
2555 if (!full)
2556 return;
2557
2558 db_indent += 2;
2559 count = 0;
2560 TAILQ_FOREACH(p, &object->memq, listq) {
2561 if (count == 0)
2562 db_iprintf("memory:=");
2563 else if (count == 6) {
2564 db_printf("\n");
2565 db_iprintf(" ...");
2566 count = 0;
2567 } else
2568 db_printf(",");
2569 count++;
2570
2571 db_printf("(off=0x%jx,page=0x%jx)",
2572 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2573 }
2574 if (count != 0)
2575 db_printf("\n");
2576 db_indent -= 2;
2577}
2578
2579/* XXX. */
2580#undef count
2581
2582/* XXX need this non-static entry for calling from vm_map_print. */
2583void
2584vm_object_print(
2585 /* db_expr_t */ long addr,
2586 boolean_t have_addr,
2587 /* db_expr_t */ long count,
2588 char *modif)
2589{
2590 vm_object_print_static(addr, have_addr, count, modif);
2591}
2592
2593DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2594{
2595 vm_object_t object;
2596 vm_pindex_t fidx;
2597 vm_paddr_t pa;
2598 vm_page_t m, prev_m;
2599 int rcount, nl, c;
2600
2601 nl = 0;
2602 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2603 db_printf("new object: %p\n", (void *)object);
2604 if (nl > 18) {
2605 c = cngetc();
2606 if (c != ' ')
2607 return;
2608 nl = 0;
2609 }
2610 nl++;
2611 rcount = 0;
2612 fidx = 0;
2613 pa = -1;
2614 TAILQ_FOREACH(m, &object->memq, listq) {
2615 if (m->pindex > 128)
2616 break;
2617 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2618 prev_m->pindex + 1 != m->pindex) {
2619 if (rcount) {
2620 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2621 (long)fidx, rcount, (long)pa);
2622 if (nl > 18) {
2623 c = cngetc();
2624 if (c != ' ')
2625 return;
2626 nl = 0;
2627 }
2628 nl++;
2629 rcount = 0;
2630 }
2631 }
2632 if (rcount &&
2633 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2634 ++rcount;
2635 continue;
2636 }
2637 if (rcount) {
2638 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2639 (long)fidx, rcount, (long)pa);
2640 if (nl > 18) {
2641 c = cngetc();
2642 if (c != ' ')
2643 return;
2644 nl = 0;
2645 }
2646 nl++;
2647 }
2648 fidx = m->pindex;
2649 pa = VM_PAGE_TO_PHYS(m);
2650 rcount = 1;
2651 }
2652 if (rcount) {
2653 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2654 (long)fidx, rcount, (long)pa);
2655 if (nl > 18) {
2656 c = cngetc();
2657 if (c != ' ')
2658 return;
2659 nl = 0;
2660 }
2661 nl++;
2662 }
2663 }
2664}
2665#endif /* DDB */