FreeBSD/Linux Kernel Cross Reference
sys/vm/vm_object.c
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: releng/11.1/sys/vm/vm_object.c 318716 2017-05-23 07:27:30Z markj $");
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/user.h>
83 #include <sys/vnode.h>
84 #include <sys/vmmeter.h>
85 #include <sys/sx.h>
86
87 #include <vm/vm.h>
88 #include <vm/vm_param.h>
89 #include <vm/pmap.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_pageout.h>
94 #include <vm/vm_pager.h>
95 #include <vm/swap_pager.h>
96 #include <vm/vm_kern.h>
97 #include <vm/vm_extern.h>
98 #include <vm/vm_radix.h>
99 #include <vm/vm_reserv.h>
100 #include <vm/uma.h>
101
102 static int old_msync;
103 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
104 "Use old (insecure) msync behavior");
105
106 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
107 int pagerflags, int flags, boolean_t *clearobjflags,
108 boolean_t *eio);
109 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
110 boolean_t *clearobjflags);
111 static void vm_object_qcollapse(vm_object_t object);
112 static void vm_object_vndeallocate(vm_object_t object);
113
114 /*
115 * Virtual memory objects maintain the actual data
116 * associated with allocated virtual memory. A given
117 * page of memory exists within exactly one object.
118 *
119 * An object is only deallocated when all "references"
120 * are given up. Only one "reference" to a given
121 * region of an object should be writeable.
122 *
123 * Associated with each object is a list of all resident
124 * memory pages belonging to that object; this list is
125 * maintained by the "vm_page" module, and locked by the object's
126 * lock.
127 *
128 * Each object also records a "pager" routine which is
129 * used to retrieve (and store) pages to the proper backing
130 * storage. In addition, objects may be backed by other
131 * objects from which they were virtual-copied.
132 *
133 * The only items within the object structure which are
134 * modified after time of creation are:
135 * reference count locked by object's lock
136 * pager routine locked by object's lock
137 *
138 */
139
140 struct object_q vm_object_list;
141 struct mtx vm_object_list_mtx; /* lock for object list and count */
142
143 struct vm_object kernel_object_store;
144 struct vm_object kmem_object_store;
145
146 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
147 "VM object stats");
148
149 static long object_collapses;
150 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
151 &object_collapses, 0, "VM object collapses");
152
153 static long object_bypasses;
154 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
155 &object_bypasses, 0, "VM object bypasses");
156
157 static uma_zone_t obj_zone;
158
159 static int vm_object_zinit(void *mem, int size, int flags);
160
161 #ifdef INVARIANTS
162 static void vm_object_zdtor(void *mem, int size, void *arg);
163
164 static void
165 vm_object_zdtor(void *mem, int size, void *arg)
166 {
167 vm_object_t object;
168
169 object = (vm_object_t)mem;
170 KASSERT(object->ref_count == 0,
171 ("object %p ref_count = %d", object, object->ref_count));
172 KASSERT(TAILQ_EMPTY(&object->memq),
173 ("object %p has resident pages in its memq", object));
174 KASSERT(vm_radix_is_empty(&object->rtree),
175 ("object %p has resident pages in its trie", object));
176 #if VM_NRESERVLEVEL > 0
177 KASSERT(LIST_EMPTY(&object->rvq),
178 ("object %p has reservations",
179 object));
180 #endif
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 KASSERT(object->type == OBJT_DEAD,
191 ("object %p has non-dead type %d",
192 object, object->type));
193 }
194 #endif
195
196 static int
197 vm_object_zinit(void *mem, int size, int flags)
198 {
199 vm_object_t object;
200
201 object = (vm_object_t)mem;
202 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
203
204 /* These are true for any object that has been freed */
205 object->type = OBJT_DEAD;
206 object->ref_count = 0;
207 object->rtree.rt_root = 0;
208 object->paging_in_progress = 0;
209 object->resident_page_count = 0;
210 object->shadow_count = 0;
211
212 mtx_lock(&vm_object_list_mtx);
213 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
214 mtx_unlock(&vm_object_list_mtx);
215 return (0);
216 }
217
218 static void
219 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
220 {
221
222 TAILQ_INIT(&object->memq);
223 LIST_INIT(&object->shadow_head);
224
225 object->type = type;
226 switch (type) {
227 case OBJT_DEAD:
228 panic("_vm_object_allocate: can't create OBJT_DEAD");
229 case OBJT_DEFAULT:
230 case OBJT_SWAP:
231 object->flags = OBJ_ONEMAPPING;
232 break;
233 case OBJT_DEVICE:
234 case OBJT_SG:
235 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
236 break;
237 case OBJT_MGTDEVICE:
238 object->flags = OBJ_FICTITIOUS;
239 break;
240 case OBJT_PHYS:
241 object->flags = OBJ_UNMANAGED;
242 break;
243 case OBJT_VNODE:
244 object->flags = 0;
245 break;
246 default:
247 panic("_vm_object_allocate: type %d is undefined", type);
248 }
249 object->size = size;
250 object->generation = 1;
251 object->ref_count = 1;
252 object->memattr = VM_MEMATTR_DEFAULT;
253 object->cred = NULL;
254 object->charge = 0;
255 object->handle = NULL;
256 object->backing_object = NULL;
257 object->backing_object_offset = (vm_ooffset_t) 0;
258 #if VM_NRESERVLEVEL > 0
259 LIST_INIT(&object->rvq);
260 #endif
261 umtx_shm_object_init(object);
262 }
263
264 /*
265 * vm_object_init:
266 *
267 * Initialize the VM objects module.
268 */
269 void
270 vm_object_init(void)
271 {
272 TAILQ_INIT(&vm_object_list);
273 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
274
275 rw_init(&kernel_object->lock, "kernel vm object");
276 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
277 VM_MIN_KERNEL_ADDRESS), kernel_object);
278 #if VM_NRESERVLEVEL > 0
279 kernel_object->flags |= OBJ_COLORED;
280 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
281 #endif
282
283 rw_init(&kmem_object->lock, "kmem vm object");
284 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
285 VM_MIN_KERNEL_ADDRESS), kmem_object);
286 #if VM_NRESERVLEVEL > 0
287 kmem_object->flags |= OBJ_COLORED;
288 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
289 #endif
290
291 /*
292 * The lock portion of struct vm_object must be type stable due
293 * to vm_pageout_fallback_object_lock locking a vm object
294 * without holding any references to it.
295 */
296 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
297 #ifdef INVARIANTS
298 vm_object_zdtor,
299 #else
300 NULL,
301 #endif
302 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
303
304 vm_radix_init();
305 }
306
307 void
308 vm_object_clear_flag(vm_object_t object, u_short bits)
309 {
310
311 VM_OBJECT_ASSERT_WLOCKED(object);
312 object->flags &= ~bits;
313 }
314
315 /*
316 * Sets the default memory attribute for the specified object. Pages
317 * that are allocated to this object are by default assigned this memory
318 * attribute.
319 *
320 * Presently, this function must be called before any pages are allocated
321 * to the object. In the future, this requirement may be relaxed for
322 * "default" and "swap" objects.
323 */
324 int
325 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
326 {
327
328 VM_OBJECT_ASSERT_WLOCKED(object);
329 switch (object->type) {
330 case OBJT_DEFAULT:
331 case OBJT_DEVICE:
332 case OBJT_MGTDEVICE:
333 case OBJT_PHYS:
334 case OBJT_SG:
335 case OBJT_SWAP:
336 case OBJT_VNODE:
337 if (!TAILQ_EMPTY(&object->memq))
338 return (KERN_FAILURE);
339 break;
340 case OBJT_DEAD:
341 return (KERN_INVALID_ARGUMENT);
342 default:
343 panic("vm_object_set_memattr: object %p is of undefined type",
344 object);
345 }
346 object->memattr = memattr;
347 return (KERN_SUCCESS);
348 }
349
350 void
351 vm_object_pip_add(vm_object_t object, short i)
352 {
353
354 VM_OBJECT_ASSERT_WLOCKED(object);
355 object->paging_in_progress += i;
356 }
357
358 void
359 vm_object_pip_subtract(vm_object_t object, short i)
360 {
361
362 VM_OBJECT_ASSERT_WLOCKED(object);
363 object->paging_in_progress -= i;
364 }
365
366 void
367 vm_object_pip_wakeup(vm_object_t object)
368 {
369
370 VM_OBJECT_ASSERT_WLOCKED(object);
371 object->paging_in_progress--;
372 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
373 vm_object_clear_flag(object, OBJ_PIPWNT);
374 wakeup(object);
375 }
376 }
377
378 void
379 vm_object_pip_wakeupn(vm_object_t object, short i)
380 {
381
382 VM_OBJECT_ASSERT_WLOCKED(object);
383 if (i)
384 object->paging_in_progress -= i;
385 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
386 vm_object_clear_flag(object, OBJ_PIPWNT);
387 wakeup(object);
388 }
389 }
390
391 void
392 vm_object_pip_wait(vm_object_t object, char *waitid)
393 {
394
395 VM_OBJECT_ASSERT_WLOCKED(object);
396 while (object->paging_in_progress) {
397 object->flags |= OBJ_PIPWNT;
398 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
399 }
400 }
401
402 /*
403 * vm_object_allocate:
404 *
405 * Returns a new object with the given size.
406 */
407 vm_object_t
408 vm_object_allocate(objtype_t type, vm_pindex_t size)
409 {
410 vm_object_t object;
411
412 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
413 _vm_object_allocate(type, size, object);
414 return (object);
415 }
416
417
418 /*
419 * vm_object_reference:
420 *
421 * Gets another reference to the given object. Note: OBJ_DEAD
422 * objects can be referenced during final cleaning.
423 */
424 void
425 vm_object_reference(vm_object_t object)
426 {
427 if (object == NULL)
428 return;
429 VM_OBJECT_WLOCK(object);
430 vm_object_reference_locked(object);
431 VM_OBJECT_WUNLOCK(object);
432 }
433
434 /*
435 * vm_object_reference_locked:
436 *
437 * Gets another reference to the given object.
438 *
439 * The object must be locked.
440 */
441 void
442 vm_object_reference_locked(vm_object_t object)
443 {
444 struct vnode *vp;
445
446 VM_OBJECT_ASSERT_WLOCKED(object);
447 object->ref_count++;
448 if (object->type == OBJT_VNODE) {
449 vp = object->handle;
450 vref(vp);
451 }
452 }
453
454 /*
455 * Handle deallocating an object of type OBJT_VNODE.
456 */
457 static void
458 vm_object_vndeallocate(vm_object_t object)
459 {
460 struct vnode *vp = (struct vnode *) object->handle;
461
462 VM_OBJECT_ASSERT_WLOCKED(object);
463 KASSERT(object->type == OBJT_VNODE,
464 ("vm_object_vndeallocate: not a vnode object"));
465 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
466 #ifdef INVARIANTS
467 if (object->ref_count == 0) {
468 vn_printf(vp, "vm_object_vndeallocate ");
469 panic("vm_object_vndeallocate: bad object reference count");
470 }
471 #endif
472
473 if (!umtx_shm_vnobj_persistent && object->ref_count == 1)
474 umtx_shm_object_terminated(object);
475
476 /*
477 * The test for text of vp vnode does not need a bypass to
478 * reach right VV_TEXT there, since it is obtained from
479 * object->handle.
480 */
481 if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) {
482 object->ref_count--;
483 VM_OBJECT_WUNLOCK(object);
484 /* vrele may need the vnode lock. */
485 vrele(vp);
486 } else {
487 vhold(vp);
488 VM_OBJECT_WUNLOCK(object);
489 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
490 vdrop(vp);
491 VM_OBJECT_WLOCK(object);
492 object->ref_count--;
493 if (object->type == OBJT_DEAD) {
494 VM_OBJECT_WUNLOCK(object);
495 VOP_UNLOCK(vp, 0);
496 } else {
497 if (object->ref_count == 0)
498 VOP_UNSET_TEXT(vp);
499 VM_OBJECT_WUNLOCK(object);
500 vput(vp);
501 }
502 }
503 }
504
505 /*
506 * vm_object_deallocate:
507 *
508 * Release a reference to the specified object,
509 * gained either through a vm_object_allocate
510 * or a vm_object_reference call. When all references
511 * are gone, storage associated with this object
512 * may be relinquished.
513 *
514 * No object may be locked.
515 */
516 void
517 vm_object_deallocate(vm_object_t object)
518 {
519 vm_object_t temp;
520 struct vnode *vp;
521
522 while (object != NULL) {
523 VM_OBJECT_WLOCK(object);
524 if (object->type == OBJT_VNODE) {
525 vm_object_vndeallocate(object);
526 return;
527 }
528
529 KASSERT(object->ref_count != 0,
530 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
531
532 /*
533 * If the reference count goes to 0 we start calling
534 * vm_object_terminate() on the object chain.
535 * A ref count of 1 may be a special case depending on the
536 * shadow count being 0 or 1.
537 */
538 object->ref_count--;
539 if (object->ref_count > 1) {
540 VM_OBJECT_WUNLOCK(object);
541 return;
542 } else if (object->ref_count == 1) {
543 if (object->type == OBJT_SWAP &&
544 (object->flags & OBJ_TMPFS) != 0) {
545 vp = object->un_pager.swp.swp_tmpfs;
546 vhold(vp);
547 VM_OBJECT_WUNLOCK(object);
548 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
549 VM_OBJECT_WLOCK(object);
550 if (object->type == OBJT_DEAD ||
551 object->ref_count != 1) {
552 VM_OBJECT_WUNLOCK(object);
553 VOP_UNLOCK(vp, 0);
554 vdrop(vp);
555 return;
556 }
557 if ((object->flags & OBJ_TMPFS) != 0)
558 VOP_UNSET_TEXT(vp);
559 VOP_UNLOCK(vp, 0);
560 vdrop(vp);
561 }
562 if (object->shadow_count == 0 &&
563 object->handle == NULL &&
564 (object->type == OBJT_DEFAULT ||
565 (object->type == OBJT_SWAP &&
566 (object->flags & OBJ_TMPFS_NODE) == 0))) {
567 vm_object_set_flag(object, OBJ_ONEMAPPING);
568 } else if ((object->shadow_count == 1) &&
569 (object->handle == NULL) &&
570 (object->type == OBJT_DEFAULT ||
571 object->type == OBJT_SWAP)) {
572 vm_object_t robject;
573
574 robject = LIST_FIRST(&object->shadow_head);
575 KASSERT(robject != NULL,
576 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
577 object->ref_count,
578 object->shadow_count));
579 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0,
580 ("shadowed tmpfs v_object %p", object));
581 if (!VM_OBJECT_TRYWLOCK(robject)) {
582 /*
583 * Avoid a potential deadlock.
584 */
585 object->ref_count++;
586 VM_OBJECT_WUNLOCK(object);
587 /*
588 * More likely than not the thread
589 * holding robject's lock has lower
590 * priority than the current thread.
591 * Let the lower priority thread run.
592 */
593 pause("vmo_de", 1);
594 continue;
595 }
596 /*
597 * Collapse object into its shadow unless its
598 * shadow is dead. In that case, object will
599 * be deallocated by the thread that is
600 * deallocating its shadow.
601 */
602 if ((robject->flags & OBJ_DEAD) == 0 &&
603 (robject->handle == NULL) &&
604 (robject->type == OBJT_DEFAULT ||
605 robject->type == OBJT_SWAP)) {
606
607 robject->ref_count++;
608 retry:
609 if (robject->paging_in_progress) {
610 VM_OBJECT_WUNLOCK(object);
611 vm_object_pip_wait(robject,
612 "objde1");
613 temp = robject->backing_object;
614 if (object == temp) {
615 VM_OBJECT_WLOCK(object);
616 goto retry;
617 }
618 } else if (object->paging_in_progress) {
619 VM_OBJECT_WUNLOCK(robject);
620 object->flags |= OBJ_PIPWNT;
621 VM_OBJECT_SLEEP(object, object,
622 PDROP | PVM, "objde2", 0);
623 VM_OBJECT_WLOCK(robject);
624 temp = robject->backing_object;
625 if (object == temp) {
626 VM_OBJECT_WLOCK(object);
627 goto retry;
628 }
629 } else
630 VM_OBJECT_WUNLOCK(object);
631
632 if (robject->ref_count == 1) {
633 robject->ref_count--;
634 object = robject;
635 goto doterm;
636 }
637 object = robject;
638 vm_object_collapse(object);
639 VM_OBJECT_WUNLOCK(object);
640 continue;
641 }
642 VM_OBJECT_WUNLOCK(robject);
643 }
644 VM_OBJECT_WUNLOCK(object);
645 return;
646 }
647 doterm:
648 umtx_shm_object_terminated(object);
649 temp = object->backing_object;
650 if (temp != NULL) {
651 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
652 ("shadowed tmpfs v_object 2 %p", object));
653 VM_OBJECT_WLOCK(temp);
654 LIST_REMOVE(object, shadow_list);
655 temp->shadow_count--;
656 VM_OBJECT_WUNLOCK(temp);
657 object->backing_object = NULL;
658 }
659 /*
660 * Don't double-terminate, we could be in a termination
661 * recursion due to the terminate having to sync data
662 * to disk.
663 */
664 if ((object->flags & OBJ_DEAD) == 0)
665 vm_object_terminate(object);
666 else
667 VM_OBJECT_WUNLOCK(object);
668 object = temp;
669 }
670 }
671
672 /*
673 * vm_object_destroy removes the object from the global object list
674 * and frees the space for the object.
675 */
676 void
677 vm_object_destroy(vm_object_t object)
678 {
679
680 /*
681 * Release the allocation charge.
682 */
683 if (object->cred != NULL) {
684 swap_release_by_cred(object->charge, object->cred);
685 object->charge = 0;
686 crfree(object->cred);
687 object->cred = NULL;
688 }
689
690 /*
691 * Free the space for the object.
692 */
693 uma_zfree(obj_zone, object);
694 }
695
696 /*
697 * vm_object_terminate actually destroys the specified object, freeing
698 * up all previously used resources.
699 *
700 * The object must be locked.
701 * This routine may block.
702 */
703 void
704 vm_object_terminate(vm_object_t object)
705 {
706 vm_page_t p, p_next;
707
708 VM_OBJECT_ASSERT_WLOCKED(object);
709
710 /*
711 * Make sure no one uses us.
712 */
713 vm_object_set_flag(object, OBJ_DEAD);
714
715 /*
716 * wait for the pageout daemon to be done with the object
717 */
718 vm_object_pip_wait(object, "objtrm");
719
720 KASSERT(!object->paging_in_progress,
721 ("vm_object_terminate: pageout in progress"));
722
723 /*
724 * Clean and free the pages, as appropriate. All references to the
725 * object are gone, so we don't need to lock it.
726 */
727 if (object->type == OBJT_VNODE) {
728 struct vnode *vp = (struct vnode *)object->handle;
729
730 /*
731 * Clean pages and flush buffers.
732 */
733 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
734 VM_OBJECT_WUNLOCK(object);
735
736 vinvalbuf(vp, V_SAVE, 0, 0);
737
738 BO_LOCK(&vp->v_bufobj);
739 vp->v_bufobj.bo_flag |= BO_DEAD;
740 BO_UNLOCK(&vp->v_bufobj);
741
742 VM_OBJECT_WLOCK(object);
743 }
744
745 KASSERT(object->ref_count == 0,
746 ("vm_object_terminate: object with references, ref_count=%d",
747 object->ref_count));
748
749 /*
750 * Free any remaining pageable pages. This also removes them from the
751 * paging queues. However, don't free wired pages, just remove them
752 * from the object. Rather than incrementally removing each page from
753 * the object, the page and object are reset to any empty state.
754 */
755 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
756 vm_page_assert_unbusied(p);
757 vm_page_lock(p);
758 /*
759 * Optimize the page's removal from the object by resetting
760 * its "object" field. Specifically, if the page is not
761 * wired, then the effect of this assignment is that
762 * vm_page_free()'s call to vm_page_remove() will return
763 * immediately without modifying the page or the object.
764 */
765 p->object = NULL;
766 if (p->wire_count == 0) {
767 vm_page_free(p);
768 PCPU_INC(cnt.v_pfree);
769 }
770 vm_page_unlock(p);
771 }
772 /*
773 * If the object contained any pages, then reset it to an empty state.
774 * None of the object's fields, including "resident_page_count", were
775 * modified by the preceding loop.
776 */
777 if (object->resident_page_count != 0) {
778 vm_radix_reclaim_allnodes(&object->rtree);
779 TAILQ_INIT(&object->memq);
780 object->resident_page_count = 0;
781 if (object->type == OBJT_VNODE)
782 vdrop(object->handle);
783 }
784
785 #if VM_NRESERVLEVEL > 0
786 if (__predict_false(!LIST_EMPTY(&object->rvq)))
787 vm_reserv_break_all(object);
788 #endif
789
790 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
791 object->type == OBJT_SWAP,
792 ("%s: non-swap obj %p has cred", __func__, object));
793
794 /*
795 * Let the pager know object is dead.
796 */
797 vm_pager_deallocate(object);
798 VM_OBJECT_WUNLOCK(object);
799
800 vm_object_destroy(object);
801 }
802
803 /*
804 * Make the page read-only so that we can clear the object flags. However, if
805 * this is a nosync mmap then the object is likely to stay dirty so do not
806 * mess with the page and do not clear the object flags. Returns TRUE if the
807 * page should be flushed, and FALSE otherwise.
808 */
809 static boolean_t
810 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
811 {
812
813 /*
814 * If we have been asked to skip nosync pages and this is a
815 * nosync page, skip it. Note that the object flags were not
816 * cleared in this case so we do not have to set them.
817 */
818 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
819 *clearobjflags = FALSE;
820 return (FALSE);
821 } else {
822 pmap_remove_write(p);
823 return (p->dirty != 0);
824 }
825 }
826
827 /*
828 * vm_object_page_clean
829 *
830 * Clean all dirty pages in the specified range of object. Leaves page
831 * on whatever queue it is currently on. If NOSYNC is set then do not
832 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
833 * leaving the object dirty.
834 *
835 * When stuffing pages asynchronously, allow clustering. XXX we need a
836 * synchronous clustering mode implementation.
837 *
838 * Odd semantics: if start == end, we clean everything.
839 *
840 * The object must be locked.
841 *
842 * Returns FALSE if some page from the range was not written, as
843 * reported by the pager, and TRUE otherwise.
844 */
845 boolean_t
846 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
847 int flags)
848 {
849 vm_page_t np, p;
850 vm_pindex_t pi, tend, tstart;
851 int curgeneration, n, pagerflags;
852 boolean_t clearobjflags, eio, res;
853
854 VM_OBJECT_ASSERT_WLOCKED(object);
855
856 /*
857 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
858 * objects. The check below prevents the function from
859 * operating on non-vnode objects.
860 */
861 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
862 object->resident_page_count == 0)
863 return (TRUE);
864
865 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
866 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
867 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
868
869 tstart = OFF_TO_IDX(start);
870 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
871 clearobjflags = tstart == 0 && tend >= object->size;
872 res = TRUE;
873
874 rescan:
875 curgeneration = object->generation;
876
877 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
878 pi = p->pindex;
879 if (pi >= tend)
880 break;
881 np = TAILQ_NEXT(p, listq);
882 if (p->valid == 0)
883 continue;
884 if (vm_page_sleep_if_busy(p, "vpcwai")) {
885 if (object->generation != curgeneration) {
886 if ((flags & OBJPC_SYNC) != 0)
887 goto rescan;
888 else
889 clearobjflags = FALSE;
890 }
891 np = vm_page_find_least(object, pi);
892 continue;
893 }
894 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
895 continue;
896
897 n = vm_object_page_collect_flush(object, p, pagerflags,
898 flags, &clearobjflags, &eio);
899 if (eio) {
900 res = FALSE;
901 clearobjflags = FALSE;
902 }
903 if (object->generation != curgeneration) {
904 if ((flags & OBJPC_SYNC) != 0)
905 goto rescan;
906 else
907 clearobjflags = FALSE;
908 }
909
910 /*
911 * If the VOP_PUTPAGES() did a truncated write, so
912 * that even the first page of the run is not fully
913 * written, vm_pageout_flush() returns 0 as the run
914 * length. Since the condition that caused truncated
915 * write may be permanent, e.g. exhausted free space,
916 * accepting n == 0 would cause an infinite loop.
917 *
918 * Forwarding the iterator leaves the unwritten page
919 * behind, but there is not much we can do there if
920 * filesystem refuses to write it.
921 */
922 if (n == 0) {
923 n = 1;
924 clearobjflags = FALSE;
925 }
926 np = vm_page_find_least(object, pi + n);
927 }
928 #if 0
929 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
930 #endif
931
932 if (clearobjflags)
933 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
934 return (res);
935 }
936
937 static int
938 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
939 int flags, boolean_t *clearobjflags, boolean_t *eio)
940 {
941 vm_page_t ma[vm_pageout_page_count], p_first, tp;
942 int count, i, mreq, runlen;
943
944 vm_page_lock_assert(p, MA_NOTOWNED);
945 VM_OBJECT_ASSERT_WLOCKED(object);
946
947 count = 1;
948 mreq = 0;
949
950 for (tp = p; count < vm_pageout_page_count; count++) {
951 tp = vm_page_next(tp);
952 if (tp == NULL || vm_page_busied(tp))
953 break;
954 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
955 break;
956 }
957
958 for (p_first = p; count < vm_pageout_page_count; count++) {
959 tp = vm_page_prev(p_first);
960 if (tp == NULL || vm_page_busied(tp))
961 break;
962 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
963 break;
964 p_first = tp;
965 mreq++;
966 }
967
968 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
969 ma[i] = tp;
970
971 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
972 return (runlen);
973 }
974
975 /*
976 * Note that there is absolutely no sense in writing out
977 * anonymous objects, so we track down the vnode object
978 * to write out.
979 * We invalidate (remove) all pages from the address space
980 * for semantic correctness.
981 *
982 * If the backing object is a device object with unmanaged pages, then any
983 * mappings to the specified range of pages must be removed before this
984 * function is called.
985 *
986 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
987 * may start out with a NULL object.
988 */
989 boolean_t
990 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
991 boolean_t syncio, boolean_t invalidate)
992 {
993 vm_object_t backing_object;
994 struct vnode *vp;
995 struct mount *mp;
996 int error, flags, fsync_after;
997 boolean_t res;
998
999 if (object == NULL)
1000 return (TRUE);
1001 res = TRUE;
1002 error = 0;
1003 VM_OBJECT_WLOCK(object);
1004 while ((backing_object = object->backing_object) != NULL) {
1005 VM_OBJECT_WLOCK(backing_object);
1006 offset += object->backing_object_offset;
1007 VM_OBJECT_WUNLOCK(object);
1008 object = backing_object;
1009 if (object->size < OFF_TO_IDX(offset + size))
1010 size = IDX_TO_OFF(object->size) - offset;
1011 }
1012 /*
1013 * Flush pages if writing is allowed, invalidate them
1014 * if invalidation requested. Pages undergoing I/O
1015 * will be ignored by vm_object_page_remove().
1016 *
1017 * We cannot lock the vnode and then wait for paging
1018 * to complete without deadlocking against vm_fault.
1019 * Instead we simply call vm_object_page_remove() and
1020 * allow it to block internally on a page-by-page
1021 * basis when it encounters pages undergoing async
1022 * I/O.
1023 */
1024 if (object->type == OBJT_VNODE &&
1025 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1026 vp = object->handle;
1027 VM_OBJECT_WUNLOCK(object);
1028 (void) vn_start_write(vp, &mp, V_WAIT);
1029 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1030 if (syncio && !invalidate && offset == 0 &&
1031 atop(size) == object->size) {
1032 /*
1033 * If syncing the whole mapping of the file,
1034 * it is faster to schedule all the writes in
1035 * async mode, also allowing the clustering,
1036 * and then wait for i/o to complete.
1037 */
1038 flags = 0;
1039 fsync_after = TRUE;
1040 } else {
1041 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1042 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1043 fsync_after = FALSE;
1044 }
1045 VM_OBJECT_WLOCK(object);
1046 res = vm_object_page_clean(object, offset, offset + size,
1047 flags);
1048 VM_OBJECT_WUNLOCK(object);
1049 if (fsync_after)
1050 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1051 VOP_UNLOCK(vp, 0);
1052 vn_finished_write(mp);
1053 if (error != 0)
1054 res = FALSE;
1055 VM_OBJECT_WLOCK(object);
1056 }
1057 if ((object->type == OBJT_VNODE ||
1058 object->type == OBJT_DEVICE) && invalidate) {
1059 if (object->type == OBJT_DEVICE)
1060 /*
1061 * The option OBJPR_NOTMAPPED must be passed here
1062 * because vm_object_page_remove() cannot remove
1063 * unmanaged mappings.
1064 */
1065 flags = OBJPR_NOTMAPPED;
1066 else if (old_msync)
1067 flags = 0;
1068 else
1069 flags = OBJPR_CLEANONLY;
1070 vm_object_page_remove(object, OFF_TO_IDX(offset),
1071 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1072 }
1073 VM_OBJECT_WUNLOCK(object);
1074 return (res);
1075 }
1076
1077 /*
1078 * vm_object_madvise:
1079 *
1080 * Implements the madvise function at the object/page level.
1081 *
1082 * MADV_WILLNEED (any object)
1083 *
1084 * Activate the specified pages if they are resident.
1085 *
1086 * MADV_DONTNEED (any object)
1087 *
1088 * Deactivate the specified pages if they are resident.
1089 *
1090 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1091 * OBJ_ONEMAPPING only)
1092 *
1093 * Deactivate and clean the specified pages if they are
1094 * resident. This permits the process to reuse the pages
1095 * without faulting or the kernel to reclaim the pages
1096 * without I/O.
1097 */
1098 void
1099 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1100 int advise)
1101 {
1102 vm_pindex_t tpindex;
1103 vm_object_t backing_object, tobject;
1104 vm_page_t m;
1105
1106 if (object == NULL)
1107 return;
1108 VM_OBJECT_WLOCK(object);
1109 /*
1110 * Locate and adjust resident pages
1111 */
1112 for (; pindex < end; pindex += 1) {
1113 relookup:
1114 tobject = object;
1115 tpindex = pindex;
1116 shadowlookup:
1117 /*
1118 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1119 * and those pages must be OBJ_ONEMAPPING.
1120 */
1121 if (advise == MADV_FREE) {
1122 if ((tobject->type != OBJT_DEFAULT &&
1123 tobject->type != OBJT_SWAP) ||
1124 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1125 goto unlock_tobject;
1126 }
1127 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1128 goto unlock_tobject;
1129 m = vm_page_lookup(tobject, tpindex);
1130 if (m == NULL) {
1131 /*
1132 * There may be swap even if there is no backing page
1133 */
1134 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1135 swap_pager_freespace(tobject, tpindex, 1);
1136 /*
1137 * next object
1138 */
1139 backing_object = tobject->backing_object;
1140 if (backing_object == NULL)
1141 goto unlock_tobject;
1142 VM_OBJECT_WLOCK(backing_object);
1143 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1144 if (tobject != object)
1145 VM_OBJECT_WUNLOCK(tobject);
1146 tobject = backing_object;
1147 goto shadowlookup;
1148 } else if (m->valid != VM_PAGE_BITS_ALL)
1149 goto unlock_tobject;
1150 /*
1151 * If the page is not in a normal state, skip it.
1152 */
1153 vm_page_lock(m);
1154 if (m->hold_count != 0 || m->wire_count != 0) {
1155 vm_page_unlock(m);
1156 goto unlock_tobject;
1157 }
1158 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1159 ("vm_object_madvise: page %p is fictitious", m));
1160 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1161 ("vm_object_madvise: page %p is not managed", m));
1162 if (vm_page_busied(m)) {
1163 if (advise == MADV_WILLNEED) {
1164 /*
1165 * Reference the page before unlocking and
1166 * sleeping so that the page daemon is less
1167 * likely to reclaim it.
1168 */
1169 vm_page_aflag_set(m, PGA_REFERENCED);
1170 }
1171 if (object != tobject)
1172 VM_OBJECT_WUNLOCK(object);
1173 VM_OBJECT_WUNLOCK(tobject);
1174 vm_page_busy_sleep(m, "madvpo", false);
1175 VM_OBJECT_WLOCK(object);
1176 goto relookup;
1177 }
1178 if (advise == MADV_WILLNEED) {
1179 vm_page_activate(m);
1180 } else {
1181 vm_page_advise(m, advise);
1182 }
1183 vm_page_unlock(m);
1184 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1185 swap_pager_freespace(tobject, tpindex, 1);
1186 unlock_tobject:
1187 if (tobject != object)
1188 VM_OBJECT_WUNLOCK(tobject);
1189 }
1190 VM_OBJECT_WUNLOCK(object);
1191 }
1192
1193 /*
1194 * vm_object_shadow:
1195 *
1196 * Create a new object which is backed by the
1197 * specified existing object range. The source
1198 * object reference is deallocated.
1199 *
1200 * The new object and offset into that object
1201 * are returned in the source parameters.
1202 */
1203 void
1204 vm_object_shadow(
1205 vm_object_t *object, /* IN/OUT */
1206 vm_ooffset_t *offset, /* IN/OUT */
1207 vm_size_t length)
1208 {
1209 vm_object_t source;
1210 vm_object_t result;
1211
1212 source = *object;
1213
1214 /*
1215 * Don't create the new object if the old object isn't shared.
1216 */
1217 if (source != NULL) {
1218 VM_OBJECT_WLOCK(source);
1219 if (source->ref_count == 1 &&
1220 source->handle == NULL &&
1221 (source->type == OBJT_DEFAULT ||
1222 source->type == OBJT_SWAP)) {
1223 VM_OBJECT_WUNLOCK(source);
1224 return;
1225 }
1226 VM_OBJECT_WUNLOCK(source);
1227 }
1228
1229 /*
1230 * Allocate a new object with the given length.
1231 */
1232 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1233
1234 /*
1235 * The new object shadows the source object, adding a reference to it.
1236 * Our caller changes his reference to point to the new object,
1237 * removing a reference to the source object. Net result: no change
1238 * of reference count.
1239 *
1240 * Try to optimize the result object's page color when shadowing
1241 * in order to maintain page coloring consistency in the combined
1242 * shadowed object.
1243 */
1244 result->backing_object = source;
1245 /*
1246 * Store the offset into the source object, and fix up the offset into
1247 * the new object.
1248 */
1249 result->backing_object_offset = *offset;
1250 if (source != NULL) {
1251 VM_OBJECT_WLOCK(source);
1252 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1253 source->shadow_count++;
1254 #if VM_NRESERVLEVEL > 0
1255 result->flags |= source->flags & OBJ_COLORED;
1256 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1257 ((1 << (VM_NFREEORDER - 1)) - 1);
1258 #endif
1259 VM_OBJECT_WUNLOCK(source);
1260 }
1261
1262
1263 /*
1264 * Return the new things
1265 */
1266 *offset = 0;
1267 *object = result;
1268 }
1269
1270 /*
1271 * vm_object_split:
1272 *
1273 * Split the pages in a map entry into a new object. This affords
1274 * easier removal of unused pages, and keeps object inheritance from
1275 * being a negative impact on memory usage.
1276 */
1277 void
1278 vm_object_split(vm_map_entry_t entry)
1279 {
1280 vm_page_t m, m_next;
1281 vm_object_t orig_object, new_object, source;
1282 vm_pindex_t idx, offidxstart;
1283 vm_size_t size;
1284
1285 orig_object = entry->object.vm_object;
1286 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1287 return;
1288 if (orig_object->ref_count <= 1)
1289 return;
1290 VM_OBJECT_WUNLOCK(orig_object);
1291
1292 offidxstart = OFF_TO_IDX(entry->offset);
1293 size = atop(entry->end - entry->start);
1294
1295 /*
1296 * If swap_pager_copy() is later called, it will convert new_object
1297 * into a swap object.
1298 */
1299 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1300
1301 /*
1302 * At this point, the new object is still private, so the order in
1303 * which the original and new objects are locked does not matter.
1304 */
1305 VM_OBJECT_WLOCK(new_object);
1306 VM_OBJECT_WLOCK(orig_object);
1307 source = orig_object->backing_object;
1308 if (source != NULL) {
1309 VM_OBJECT_WLOCK(source);
1310 if ((source->flags & OBJ_DEAD) != 0) {
1311 VM_OBJECT_WUNLOCK(source);
1312 VM_OBJECT_WUNLOCK(orig_object);
1313 VM_OBJECT_WUNLOCK(new_object);
1314 vm_object_deallocate(new_object);
1315 VM_OBJECT_WLOCK(orig_object);
1316 return;
1317 }
1318 LIST_INSERT_HEAD(&source->shadow_head,
1319 new_object, shadow_list);
1320 source->shadow_count++;
1321 vm_object_reference_locked(source); /* for new_object */
1322 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1323 VM_OBJECT_WUNLOCK(source);
1324 new_object->backing_object_offset =
1325 orig_object->backing_object_offset + entry->offset;
1326 new_object->backing_object = source;
1327 }
1328 if (orig_object->cred != NULL) {
1329 new_object->cred = orig_object->cred;
1330 crhold(orig_object->cred);
1331 new_object->charge = ptoa(size);
1332 KASSERT(orig_object->charge >= ptoa(size),
1333 ("orig_object->charge < 0"));
1334 orig_object->charge -= ptoa(size);
1335 }
1336 retry:
1337 m = vm_page_find_least(orig_object, offidxstart);
1338 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1339 m = m_next) {
1340 m_next = TAILQ_NEXT(m, listq);
1341
1342 /*
1343 * We must wait for pending I/O to complete before we can
1344 * rename the page.
1345 *
1346 * We do not have to VM_PROT_NONE the page as mappings should
1347 * not be changed by this operation.
1348 */
1349 if (vm_page_busied(m)) {
1350 VM_OBJECT_WUNLOCK(new_object);
1351 vm_page_lock(m);
1352 VM_OBJECT_WUNLOCK(orig_object);
1353 vm_page_busy_sleep(m, "spltwt", false);
1354 VM_OBJECT_WLOCK(orig_object);
1355 VM_OBJECT_WLOCK(new_object);
1356 goto retry;
1357 }
1358
1359 /* vm_page_rename() will dirty the page. */
1360 if (vm_page_rename(m, new_object, idx)) {
1361 VM_OBJECT_WUNLOCK(new_object);
1362 VM_OBJECT_WUNLOCK(orig_object);
1363 VM_WAIT;
1364 VM_OBJECT_WLOCK(orig_object);
1365 VM_OBJECT_WLOCK(new_object);
1366 goto retry;
1367 }
1368 #if VM_NRESERVLEVEL > 0
1369 /*
1370 * If some of the reservation's allocated pages remain with
1371 * the original object, then transferring the reservation to
1372 * the new object is neither particularly beneficial nor
1373 * particularly harmful as compared to leaving the reservation
1374 * with the original object. If, however, all of the
1375 * reservation's allocated pages are transferred to the new
1376 * object, then transferring the reservation is typically
1377 * beneficial. Determining which of these two cases applies
1378 * would be more costly than unconditionally renaming the
1379 * reservation.
1380 */
1381 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1382 #endif
1383 if (orig_object->type == OBJT_SWAP)
1384 vm_page_xbusy(m);
1385 }
1386 if (orig_object->type == OBJT_SWAP) {
1387 /*
1388 * swap_pager_copy() can sleep, in which case the orig_object's
1389 * and new_object's locks are released and reacquired.
1390 */
1391 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1392 TAILQ_FOREACH(m, &new_object->memq, listq)
1393 vm_page_xunbusy(m);
1394 }
1395 VM_OBJECT_WUNLOCK(orig_object);
1396 VM_OBJECT_WUNLOCK(new_object);
1397 entry->object.vm_object = new_object;
1398 entry->offset = 0LL;
1399 vm_object_deallocate(orig_object);
1400 VM_OBJECT_WLOCK(new_object);
1401 }
1402
1403 #define OBSC_COLLAPSE_NOWAIT 0x0002
1404 #define OBSC_COLLAPSE_WAIT 0x0004
1405
1406 static vm_page_t
1407 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next,
1408 int op)
1409 {
1410 vm_object_t backing_object;
1411
1412 VM_OBJECT_ASSERT_WLOCKED(object);
1413 backing_object = object->backing_object;
1414 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1415
1416 KASSERT(p == NULL || vm_page_busied(p), ("unbusy page %p", p));
1417 KASSERT(p == NULL || p->object == object || p->object == backing_object,
1418 ("invalid ownership %p %p %p", p, object, backing_object));
1419 if ((op & OBSC_COLLAPSE_NOWAIT) != 0)
1420 return (next);
1421 if (p != NULL)
1422 vm_page_lock(p);
1423 VM_OBJECT_WUNLOCK(object);
1424 VM_OBJECT_WUNLOCK(backing_object);
1425 if (p == NULL)
1426 VM_WAIT;
1427 else
1428 vm_page_busy_sleep(p, "vmocol", false);
1429 VM_OBJECT_WLOCK(object);
1430 VM_OBJECT_WLOCK(backing_object);
1431 return (TAILQ_FIRST(&backing_object->memq));
1432 }
1433
1434 static bool
1435 vm_object_scan_all_shadowed(vm_object_t object)
1436 {
1437 vm_object_t backing_object;
1438 vm_page_t p, pp;
1439 vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1440
1441 VM_OBJECT_ASSERT_WLOCKED(object);
1442 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1443
1444 backing_object = object->backing_object;
1445
1446 /*
1447 * Initial conditions:
1448 *
1449 * We do not want to have to test for the existence of swap
1450 * pages in the backing object. XXX but with the new swapper this
1451 * would be pretty easy to do.
1452 */
1453 if (backing_object->type != OBJT_DEFAULT &&
1454 backing_object->type != OBJT_SWAP)
1455 return (false);
1456
1457 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1458 p = vm_page_find_least(backing_object, pi);
1459 ps = swap_pager_find_least(backing_object, pi);
1460
1461 /*
1462 * Only check pages inside the parent object's range and
1463 * inside the parent object's mapping of the backing object.
1464 */
1465 for (;; pi++) {
1466 if (p != NULL && p->pindex < pi)
1467 p = TAILQ_NEXT(p, listq);
1468 if (ps < pi)
1469 ps = swap_pager_find_least(backing_object, pi);
1470 if (p == NULL && ps >= backing_object->size)
1471 break;
1472 else if (p == NULL)
1473 pi = ps;
1474 else
1475 pi = MIN(p->pindex, ps);
1476
1477 new_pindex = pi - backing_offset_index;
1478 if (new_pindex >= object->size)
1479 break;
1480
1481 /*
1482 * See if the parent has the page or if the parent's object
1483 * pager has the page. If the parent has the page but the page
1484 * is not valid, the parent's object pager must have the page.
1485 *
1486 * If this fails, the parent does not completely shadow the
1487 * object and we might as well give up now.
1488 */
1489 pp = vm_page_lookup(object, new_pindex);
1490 if ((pp == NULL || pp->valid == 0) &&
1491 !vm_pager_has_page(object, new_pindex, NULL, NULL))
1492 return (false);
1493 }
1494 return (true);
1495 }
1496
1497 static bool
1498 vm_object_collapse_scan(vm_object_t object, int op)
1499 {
1500 vm_object_t backing_object;
1501 vm_page_t next, p, pp;
1502 vm_pindex_t backing_offset_index, new_pindex;
1503
1504 VM_OBJECT_ASSERT_WLOCKED(object);
1505 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1506
1507 backing_object = object->backing_object;
1508 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1509
1510 /*
1511 * Initial conditions
1512 */
1513 if ((op & OBSC_COLLAPSE_WAIT) != 0)
1514 vm_object_set_flag(backing_object, OBJ_DEAD);
1515
1516 /*
1517 * Our scan
1518 */
1519 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1520 next = TAILQ_NEXT(p, listq);
1521 new_pindex = p->pindex - backing_offset_index;
1522
1523 /*
1524 * Check for busy page
1525 */
1526 if (vm_page_busied(p)) {
1527 next = vm_object_collapse_scan_wait(object, p, next, op);
1528 continue;
1529 }
1530
1531 KASSERT(p->object == backing_object,
1532 ("vm_object_collapse_scan: object mismatch"));
1533
1534 if (p->pindex < backing_offset_index ||
1535 new_pindex >= object->size) {
1536 if (backing_object->type == OBJT_SWAP)
1537 swap_pager_freespace(backing_object, p->pindex,
1538 1);
1539
1540 /*
1541 * Page is out of the parent object's range, we can
1542 * simply destroy it.
1543 */
1544 vm_page_lock(p);
1545 KASSERT(!pmap_page_is_mapped(p),
1546 ("freeing mapped page %p", p));
1547 if (p->wire_count == 0)
1548 vm_page_free(p);
1549 else
1550 vm_page_remove(p);
1551 vm_page_unlock(p);
1552 continue;
1553 }
1554
1555 pp = vm_page_lookup(object, new_pindex);
1556 if (pp != NULL && vm_page_busied(pp)) {
1557 /*
1558 * The page in the parent is busy and possibly not
1559 * (yet) valid. Until its state is finalized by the
1560 * busy bit owner, we can't tell whether it shadows the
1561 * original page. Therefore, we must either skip it
1562 * and the original (backing_object) page or wait for
1563 * its state to be finalized.
1564 *
1565 * This is due to a race with vm_fault() where we must
1566 * unbusy the original (backing_obj) page before we can
1567 * (re)lock the parent. Hence we can get here.
1568 */
1569 next = vm_object_collapse_scan_wait(object, pp, next,
1570 op);
1571 continue;
1572 }
1573
1574 KASSERT(pp == NULL || pp->valid != 0,
1575 ("unbusy invalid page %p", pp));
1576
1577 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1578 NULL)) {
1579 /*
1580 * The page already exists in the parent OR swap exists
1581 * for this location in the parent. Leave the parent's
1582 * page alone. Destroy the original page from the
1583 * backing object.
1584 */
1585 if (backing_object->type == OBJT_SWAP)
1586 swap_pager_freespace(backing_object, p->pindex,
1587 1);
1588 vm_page_lock(p);
1589 KASSERT(!pmap_page_is_mapped(p),
1590 ("freeing mapped page %p", p));
1591 if (p->wire_count == 0)
1592 vm_page_free(p);
1593 else
1594 vm_page_remove(p);
1595 vm_page_unlock(p);
1596 continue;
1597 }
1598
1599 /*
1600 * Page does not exist in parent, rename the page from the
1601 * backing object to the main object.
1602 *
1603 * If the page was mapped to a process, it can remain mapped
1604 * through the rename. vm_page_rename() will dirty the page.
1605 */
1606 if (vm_page_rename(p, object, new_pindex)) {
1607 next = vm_object_collapse_scan_wait(object, NULL, next,
1608 op);
1609 continue;
1610 }
1611
1612 /* Use the old pindex to free the right page. */
1613 if (backing_object->type == OBJT_SWAP)
1614 swap_pager_freespace(backing_object,
1615 new_pindex + backing_offset_index, 1);
1616
1617 #if VM_NRESERVLEVEL > 0
1618 /*
1619 * Rename the reservation.
1620 */
1621 vm_reserv_rename(p, object, backing_object,
1622 backing_offset_index);
1623 #endif
1624 }
1625 return (true);
1626 }
1627
1628
1629 /*
1630 * this version of collapse allows the operation to occur earlier and
1631 * when paging_in_progress is true for an object... This is not a complete
1632 * operation, but should plug 99.9% of the rest of the leaks.
1633 */
1634 static void
1635 vm_object_qcollapse(vm_object_t object)
1636 {
1637 vm_object_t backing_object = object->backing_object;
1638
1639 VM_OBJECT_ASSERT_WLOCKED(object);
1640 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1641
1642 if (backing_object->ref_count != 1)
1643 return;
1644
1645 vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT);
1646 }
1647
1648 /*
1649 * vm_object_collapse:
1650 *
1651 * Collapse an object with the object backing it.
1652 * Pages in the backing object are moved into the
1653 * parent, and the backing object is deallocated.
1654 */
1655 void
1656 vm_object_collapse(vm_object_t object)
1657 {
1658 vm_object_t backing_object, new_backing_object;
1659
1660 VM_OBJECT_ASSERT_WLOCKED(object);
1661
1662 while (TRUE) {
1663 /*
1664 * Verify that the conditions are right for collapse:
1665 *
1666 * The object exists and the backing object exists.
1667 */
1668 if ((backing_object = object->backing_object) == NULL)
1669 break;
1670
1671 /*
1672 * we check the backing object first, because it is most likely
1673 * not collapsable.
1674 */
1675 VM_OBJECT_WLOCK(backing_object);
1676 if (backing_object->handle != NULL ||
1677 (backing_object->type != OBJT_DEFAULT &&
1678 backing_object->type != OBJT_SWAP) ||
1679 (backing_object->flags & OBJ_DEAD) ||
1680 object->handle != NULL ||
1681 (object->type != OBJT_DEFAULT &&
1682 object->type != OBJT_SWAP) ||
1683 (object->flags & OBJ_DEAD)) {
1684 VM_OBJECT_WUNLOCK(backing_object);
1685 break;
1686 }
1687
1688 if (object->paging_in_progress != 0 ||
1689 backing_object->paging_in_progress != 0) {
1690 vm_object_qcollapse(object);
1691 VM_OBJECT_WUNLOCK(backing_object);
1692 break;
1693 }
1694
1695 /*
1696 * We know that we can either collapse the backing object (if
1697 * the parent is the only reference to it) or (perhaps) have
1698 * the parent bypass the object if the parent happens to shadow
1699 * all the resident pages in the entire backing object.
1700 *
1701 * This is ignoring pager-backed pages such as swap pages.
1702 * vm_object_collapse_scan fails the shadowing test in this
1703 * case.
1704 */
1705 if (backing_object->ref_count == 1) {
1706 vm_object_pip_add(object, 1);
1707 vm_object_pip_add(backing_object, 1);
1708
1709 /*
1710 * If there is exactly one reference to the backing
1711 * object, we can collapse it into the parent.
1712 */
1713 vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT);
1714
1715 #if VM_NRESERVLEVEL > 0
1716 /*
1717 * Break any reservations from backing_object.
1718 */
1719 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1720 vm_reserv_break_all(backing_object);
1721 #endif
1722
1723 /*
1724 * Move the pager from backing_object to object.
1725 */
1726 if (backing_object->type == OBJT_SWAP) {
1727 /*
1728 * swap_pager_copy() can sleep, in which case
1729 * the backing_object's and object's locks are
1730 * released and reacquired.
1731 * Since swap_pager_copy() is being asked to
1732 * destroy the source, it will change the
1733 * backing_object's type to OBJT_DEFAULT.
1734 */
1735 swap_pager_copy(
1736 backing_object,
1737 object,
1738 OFF_TO_IDX(object->backing_object_offset), TRUE);
1739 }
1740 /*
1741 * Object now shadows whatever backing_object did.
1742 * Note that the reference to
1743 * backing_object->backing_object moves from within
1744 * backing_object to within object.
1745 */
1746 LIST_REMOVE(object, shadow_list);
1747 backing_object->shadow_count--;
1748 if (backing_object->backing_object) {
1749 VM_OBJECT_WLOCK(backing_object->backing_object);
1750 LIST_REMOVE(backing_object, shadow_list);
1751 LIST_INSERT_HEAD(
1752 &backing_object->backing_object->shadow_head,
1753 object, shadow_list);
1754 /*
1755 * The shadow_count has not changed.
1756 */
1757 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1758 }
1759 object->backing_object = backing_object->backing_object;
1760 object->backing_object_offset +=
1761 backing_object->backing_object_offset;
1762
1763 /*
1764 * Discard backing_object.
1765 *
1766 * Since the backing object has no pages, no pager left,
1767 * and no object references within it, all that is
1768 * necessary is to dispose of it.
1769 */
1770 KASSERT(backing_object->ref_count == 1, (
1771 "backing_object %p was somehow re-referenced during collapse!",
1772 backing_object));
1773 vm_object_pip_wakeup(backing_object);
1774 backing_object->type = OBJT_DEAD;
1775 backing_object->ref_count = 0;
1776 VM_OBJECT_WUNLOCK(backing_object);
1777 vm_object_destroy(backing_object);
1778
1779 vm_object_pip_wakeup(object);
1780 object_collapses++;
1781 } else {
1782 /*
1783 * If we do not entirely shadow the backing object,
1784 * there is nothing we can do so we give up.
1785 */
1786 if (object->resident_page_count != object->size &&
1787 !vm_object_scan_all_shadowed(object)) {
1788 VM_OBJECT_WUNLOCK(backing_object);
1789 break;
1790 }
1791
1792 /*
1793 * Make the parent shadow the next object in the
1794 * chain. Deallocating backing_object will not remove
1795 * it, since its reference count is at least 2.
1796 */
1797 LIST_REMOVE(object, shadow_list);
1798 backing_object->shadow_count--;
1799
1800 new_backing_object = backing_object->backing_object;
1801 if ((object->backing_object = new_backing_object) != NULL) {
1802 VM_OBJECT_WLOCK(new_backing_object);
1803 LIST_INSERT_HEAD(
1804 &new_backing_object->shadow_head,
1805 object,
1806 shadow_list
1807 );
1808 new_backing_object->shadow_count++;
1809 vm_object_reference_locked(new_backing_object);
1810 VM_OBJECT_WUNLOCK(new_backing_object);
1811 object->backing_object_offset +=
1812 backing_object->backing_object_offset;
1813 }
1814
1815 /*
1816 * Drop the reference count on backing_object. Since
1817 * its ref_count was at least 2, it will not vanish.
1818 */
1819 backing_object->ref_count--;
1820 VM_OBJECT_WUNLOCK(backing_object);
1821 object_bypasses++;
1822 }
1823
1824 /*
1825 * Try again with this object's new backing object.
1826 */
1827 }
1828 }
1829
1830 /*
1831 * vm_object_page_remove:
1832 *
1833 * For the given object, either frees or invalidates each of the
1834 * specified pages. In general, a page is freed. However, if a page is
1835 * wired for any reason other than the existence of a managed, wired
1836 * mapping, then it may be invalidated but not removed from the object.
1837 * Pages are specified by the given range ["start", "end") and the option
1838 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1839 * extends from "start" to the end of the object. If the option
1840 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1841 * specified range are affected. If the option OBJPR_NOTMAPPED is
1842 * specified, then the pages within the specified range must have no
1843 * mappings. Otherwise, if this option is not specified, any mappings to
1844 * the specified pages are removed before the pages are freed or
1845 * invalidated.
1846 *
1847 * In general, this operation should only be performed on objects that
1848 * contain managed pages. There are, however, two exceptions. First, it
1849 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1850 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1851 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1852 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1853 *
1854 * The object must be locked.
1855 */
1856 void
1857 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1858 int options)
1859 {
1860 vm_page_t p, next;
1861
1862 VM_OBJECT_ASSERT_WLOCKED(object);
1863 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1864 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1865 ("vm_object_page_remove: illegal options for object %p", object));
1866 if (object->resident_page_count == 0)
1867 return;
1868 vm_object_pip_add(object, 1);
1869 again:
1870 p = vm_page_find_least(object, start);
1871
1872 /*
1873 * Here, the variable "p" is either (1) the page with the least pindex
1874 * greater than or equal to the parameter "start" or (2) NULL.
1875 */
1876 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1877 next = TAILQ_NEXT(p, listq);
1878
1879 /*
1880 * If the page is wired for any reason besides the existence
1881 * of managed, wired mappings, then it cannot be freed. For
1882 * example, fictitious pages, which represent device memory,
1883 * are inherently wired and cannot be freed. They can,
1884 * however, be invalidated if the option OBJPR_CLEANONLY is
1885 * not specified.
1886 */
1887 vm_page_lock(p);
1888 if (vm_page_xbusied(p)) {
1889 VM_OBJECT_WUNLOCK(object);
1890 vm_page_busy_sleep(p, "vmopax", true);
1891 VM_OBJECT_WLOCK(object);
1892 goto again;
1893 }
1894 if (p->wire_count != 0) {
1895 if ((options & OBJPR_NOTMAPPED) == 0)
1896 pmap_remove_all(p);
1897 if ((options & OBJPR_CLEANONLY) == 0) {
1898 p->valid = 0;
1899 vm_page_undirty(p);
1900 }
1901 goto next;
1902 }
1903 if (vm_page_busied(p)) {
1904 VM_OBJECT_WUNLOCK(object);
1905 vm_page_busy_sleep(p, "vmopar", false);
1906 VM_OBJECT_WLOCK(object);
1907 goto again;
1908 }
1909 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1910 ("vm_object_page_remove: page %p is fictitious", p));
1911 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1912 if ((options & OBJPR_NOTMAPPED) == 0)
1913 pmap_remove_write(p);
1914 if (p->dirty)
1915 goto next;
1916 }
1917 if ((options & OBJPR_NOTMAPPED) == 0)
1918 pmap_remove_all(p);
1919 vm_page_free(p);
1920 next:
1921 vm_page_unlock(p);
1922 }
1923 vm_object_pip_wakeup(object);
1924 }
1925
1926 /*
1927 * vm_object_page_noreuse:
1928 *
1929 * For the given object, attempt to move the specified pages to
1930 * the head of the inactive queue. This bypasses regular LRU
1931 * operation and allows the pages to be reused quickly under memory
1932 * pressure. If a page is wired for any reason, then it will not
1933 * be queued. Pages are specified by the range ["start", "end").
1934 * As a special case, if "end" is zero, then the range extends from
1935 * "start" to the end of the object.
1936 *
1937 * This operation should only be performed on objects that
1938 * contain non-fictitious, managed pages.
1939 *
1940 * The object must be locked.
1941 */
1942 void
1943 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1944 {
1945 struct mtx *mtx, *new_mtx;
1946 vm_page_t p, next;
1947
1948 VM_OBJECT_ASSERT_LOCKED(object);
1949 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1950 ("vm_object_page_noreuse: illegal object %p", object));
1951 if (object->resident_page_count == 0)
1952 return;
1953 p = vm_page_find_least(object, start);
1954
1955 /*
1956 * Here, the variable "p" is either (1) the page with the least pindex
1957 * greater than or equal to the parameter "start" or (2) NULL.
1958 */
1959 mtx = NULL;
1960 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1961 next = TAILQ_NEXT(p, listq);
1962
1963 /*
1964 * Avoid releasing and reacquiring the same page lock.
1965 */
1966 new_mtx = vm_page_lockptr(p);
1967 if (mtx != new_mtx) {
1968 if (mtx != NULL)
1969 mtx_unlock(mtx);
1970 mtx = new_mtx;
1971 mtx_lock(mtx);
1972 }
1973 vm_page_deactivate_noreuse(p);
1974 }
1975 if (mtx != NULL)
1976 mtx_unlock(mtx);
1977 }
1978
1979 /*
1980 * Populate the specified range of the object with valid pages. Returns
1981 * TRUE if the range is successfully populated and FALSE otherwise.
1982 *
1983 * Note: This function should be optimized to pass a larger array of
1984 * pages to vm_pager_get_pages() before it is applied to a non-
1985 * OBJT_DEVICE object.
1986 *
1987 * The object must be locked.
1988 */
1989 boolean_t
1990 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1991 {
1992 vm_page_t m;
1993 vm_pindex_t pindex;
1994 int rv;
1995
1996 VM_OBJECT_ASSERT_WLOCKED(object);
1997 for (pindex = start; pindex < end; pindex++) {
1998 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
1999 if (m->valid != VM_PAGE_BITS_ALL) {
2000 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
2001 if (rv != VM_PAGER_OK) {
2002 vm_page_lock(m);
2003 vm_page_free(m);
2004 vm_page_unlock(m);
2005 break;
2006 }
2007 }
2008 /*
2009 * Keep "m" busy because a subsequent iteration may unlock
2010 * the object.
2011 */
2012 }
2013 if (pindex > start) {
2014 m = vm_page_lookup(object, start);
2015 while (m != NULL && m->pindex < pindex) {
2016 vm_page_xunbusy(m);
2017 m = TAILQ_NEXT(m, listq);
2018 }
2019 }
2020 return (pindex == end);
2021 }
2022
2023 /*
2024 * Routine: vm_object_coalesce
2025 * Function: Coalesces two objects backing up adjoining
2026 * regions of memory into a single object.
2027 *
2028 * returns TRUE if objects were combined.
2029 *
2030 * NOTE: Only works at the moment if the second object is NULL -
2031 * if it's not, which object do we lock first?
2032 *
2033 * Parameters:
2034 * prev_object First object to coalesce
2035 * prev_offset Offset into prev_object
2036 * prev_size Size of reference to prev_object
2037 * next_size Size of reference to the second object
2038 * reserved Indicator that extension region has
2039 * swap accounted for
2040 *
2041 * Conditions:
2042 * The object must *not* be locked.
2043 */
2044 boolean_t
2045 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2046 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2047 {
2048 vm_pindex_t next_pindex;
2049
2050 if (prev_object == NULL)
2051 return (TRUE);
2052 VM_OBJECT_WLOCK(prev_object);
2053 if ((prev_object->type != OBJT_DEFAULT &&
2054 prev_object->type != OBJT_SWAP) ||
2055 (prev_object->flags & OBJ_TMPFS_NODE) != 0) {
2056 VM_OBJECT_WUNLOCK(prev_object);
2057 return (FALSE);
2058 }
2059
2060 /*
2061 * Try to collapse the object first
2062 */
2063 vm_object_collapse(prev_object);
2064
2065 /*
2066 * Can't coalesce if: . more than one reference . paged out . shadows
2067 * another object . has a copy elsewhere (any of which mean that the
2068 * pages not mapped to prev_entry may be in use anyway)
2069 */
2070 if (prev_object->backing_object != NULL) {
2071 VM_OBJECT_WUNLOCK(prev_object);
2072 return (FALSE);
2073 }
2074
2075 prev_size >>= PAGE_SHIFT;
2076 next_size >>= PAGE_SHIFT;
2077 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2078
2079 if ((prev_object->ref_count > 1) &&
2080 (prev_object->size != next_pindex)) {
2081 VM_OBJECT_WUNLOCK(prev_object);
2082 return (FALSE);
2083 }
2084
2085 /*
2086 * Account for the charge.
2087 */
2088 if (prev_object->cred != NULL) {
2089
2090 /*
2091 * If prev_object was charged, then this mapping,
2092 * although not charged now, may become writable
2093 * later. Non-NULL cred in the object would prevent
2094 * swap reservation during enabling of the write
2095 * access, so reserve swap now. Failed reservation
2096 * cause allocation of the separate object for the map
2097 * entry, and swap reservation for this entry is
2098 * managed in appropriate time.
2099 */
2100 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2101 prev_object->cred)) {
2102 VM_OBJECT_WUNLOCK(prev_object);
2103 return (FALSE);
2104 }
2105 prev_object->charge += ptoa(next_size);
2106 }
2107
2108 /*
2109 * Remove any pages that may still be in the object from a previous
2110 * deallocation.
2111 */
2112 if (next_pindex < prev_object->size) {
2113 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2114 next_size, 0);
2115 if (prev_object->type == OBJT_SWAP)
2116 swap_pager_freespace(prev_object,
2117 next_pindex, next_size);
2118 #if 0
2119 if (prev_object->cred != NULL) {
2120 KASSERT(prev_object->charge >=
2121 ptoa(prev_object->size - next_pindex),
2122 ("object %p overcharged 1 %jx %jx", prev_object,
2123 (uintmax_t)next_pindex, (uintmax_t)next_size));
2124 prev_object->charge -= ptoa(prev_object->size -
2125 next_pindex);
2126 }
2127 #endif
2128 }
2129
2130 /*
2131 * Extend the object if necessary.
2132 */
2133 if (next_pindex + next_size > prev_object->size)
2134 prev_object->size = next_pindex + next_size;
2135
2136 VM_OBJECT_WUNLOCK(prev_object);
2137 return (TRUE);
2138 }
2139
2140 void
2141 vm_object_set_writeable_dirty(vm_object_t object)
2142 {
2143
2144 VM_OBJECT_ASSERT_WLOCKED(object);
2145 if (object->type != OBJT_VNODE) {
2146 if ((object->flags & OBJ_TMPFS_NODE) != 0) {
2147 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs"));
2148 vm_object_set_flag(object, OBJ_TMPFS_DIRTY);
2149 }
2150 return;
2151 }
2152 object->generation++;
2153 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2154 return;
2155 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2156 }
2157
2158 /*
2159 * vm_object_unwire:
2160 *
2161 * For each page offset within the specified range of the given object,
2162 * find the highest-level page in the shadow chain and unwire it. A page
2163 * must exist at every page offset, and the highest-level page must be
2164 * wired.
2165 */
2166 void
2167 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2168 uint8_t queue)
2169 {
2170 vm_object_t tobject;
2171 vm_page_t m, tm;
2172 vm_pindex_t end_pindex, pindex, tpindex;
2173 int depth, locked_depth;
2174
2175 KASSERT((offset & PAGE_MASK) == 0,
2176 ("vm_object_unwire: offset is not page aligned"));
2177 KASSERT((length & PAGE_MASK) == 0,
2178 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2179 /* The wired count of a fictitious page never changes. */
2180 if ((object->flags & OBJ_FICTITIOUS) != 0)
2181 return;
2182 pindex = OFF_TO_IDX(offset);
2183 end_pindex = pindex + atop(length);
2184 locked_depth = 1;
2185 VM_OBJECT_RLOCK(object);
2186 m = vm_page_find_least(object, pindex);
2187 while (pindex < end_pindex) {
2188 if (m == NULL || pindex < m->pindex) {
2189 /*
2190 * The first object in the shadow chain doesn't
2191 * contain a page at the current index. Therefore,
2192 * the page must exist in a backing object.
2193 */
2194 tobject = object;
2195 tpindex = pindex;
2196 depth = 0;
2197 do {
2198 tpindex +=
2199 OFF_TO_IDX(tobject->backing_object_offset);
2200 tobject = tobject->backing_object;
2201 KASSERT(tobject != NULL,
2202 ("vm_object_unwire: missing page"));
2203 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2204 goto next_page;
2205 depth++;
2206 if (depth == locked_depth) {
2207 locked_depth++;
2208 VM_OBJECT_RLOCK(tobject);
2209 }
2210 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2211 NULL);
2212 } else {
2213 tm = m;
2214 m = TAILQ_NEXT(m, listq);
2215 }
2216 vm_page_lock(tm);
2217 vm_page_unwire(tm, queue);
2218 vm_page_unlock(tm);
2219 next_page:
2220 pindex++;
2221 }
2222 /* Release the accumulated object locks. */
2223 for (depth = 0; depth < locked_depth; depth++) {
2224 tobject = object->backing_object;
2225 VM_OBJECT_RUNLOCK(object);
2226 object = tobject;
2227 }
2228 }
2229
2230 struct vnode *
2231 vm_object_vnode(vm_object_t object)
2232 {
2233
2234 VM_OBJECT_ASSERT_LOCKED(object);
2235 if (object->type == OBJT_VNODE)
2236 return (object->handle);
2237 if (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS) != 0)
2238 return (object->un_pager.swp.swp_tmpfs);
2239 return (NULL);
2240 }
2241
2242 static int
2243 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2244 {
2245 struct kinfo_vmobject kvo;
2246 char *fullpath, *freepath;
2247 struct vnode *vp;
2248 struct vattr va;
2249 vm_object_t obj;
2250 vm_page_t m;
2251 int count, error;
2252
2253 if (req->oldptr == NULL) {
2254 /*
2255 * If an old buffer has not been provided, generate an
2256 * estimate of the space needed for a subsequent call.
2257 */
2258 mtx_lock(&vm_object_list_mtx);
2259 count = 0;
2260 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2261 if (obj->type == OBJT_DEAD)
2262 continue;
2263 count++;
2264 }
2265 mtx_unlock(&vm_object_list_mtx);
2266 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2267 count * 11 / 10));
2268 }
2269
2270 error = 0;
2271
2272 /*
2273 * VM objects are type stable and are never removed from the
2274 * list once added. This allows us to safely read obj->object_list
2275 * after reacquiring the VM object lock.
2276 */
2277 mtx_lock(&vm_object_list_mtx);
2278 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2279 if (obj->type == OBJT_DEAD)
2280 continue;
2281 VM_OBJECT_RLOCK(obj);
2282 if (obj->type == OBJT_DEAD) {
2283 VM_OBJECT_RUNLOCK(obj);
2284 continue;
2285 }
2286 mtx_unlock(&vm_object_list_mtx);
2287 kvo.kvo_size = ptoa(obj->size);
2288 kvo.kvo_resident = obj->resident_page_count;
2289 kvo.kvo_ref_count = obj->ref_count;
2290 kvo.kvo_shadow_count = obj->shadow_count;
2291 kvo.kvo_memattr = obj->memattr;
2292 kvo.kvo_active = 0;
2293 kvo.kvo_inactive = 0;
2294 TAILQ_FOREACH(m, &obj->memq, listq) {
2295 /*
2296 * A page may belong to the object but be
2297 * dequeued and set to PQ_NONE while the
2298 * object lock is not held. This makes the
2299 * reads of m->queue below racy, and we do not
2300 * count pages set to PQ_NONE. However, this
2301 * sysctl is only meant to give an
2302 * approximation of the system anyway.
2303 */
2304 if (vm_page_active(m))
2305 kvo.kvo_active++;
2306 else if (vm_page_inactive(m))
2307 kvo.kvo_inactive++;
2308 }
2309
2310 kvo.kvo_vn_fileid = 0;
2311 kvo.kvo_vn_fsid = 0;
2312 freepath = NULL;
2313 fullpath = "";
2314 vp = NULL;
2315 switch (obj->type) {
2316 case OBJT_DEFAULT:
2317 kvo.kvo_type = KVME_TYPE_DEFAULT;
2318 break;
2319 case OBJT_VNODE:
2320 kvo.kvo_type = KVME_TYPE_VNODE;
2321 vp = obj->handle;
2322 vref(vp);
2323 break;
2324 case OBJT_SWAP:
2325 kvo.kvo_type = KVME_TYPE_SWAP;
2326 break;
2327 case OBJT_DEVICE:
2328 kvo.kvo_type = KVME_TYPE_DEVICE;
2329 break;
2330 case OBJT_PHYS:
2331 kvo.kvo_type = KVME_TYPE_PHYS;
2332 break;
2333 case OBJT_DEAD:
2334 kvo.kvo_type = KVME_TYPE_DEAD;
2335 break;
2336 case OBJT_SG:
2337 kvo.kvo_type = KVME_TYPE_SG;
2338 break;
2339 case OBJT_MGTDEVICE:
2340 kvo.kvo_type = KVME_TYPE_MGTDEVICE;
2341 break;
2342 default:
2343 kvo.kvo_type = KVME_TYPE_UNKNOWN;
2344 break;
2345 }
2346 VM_OBJECT_RUNLOCK(obj);
2347 if (vp != NULL) {
2348 vn_fullpath(curthread, vp, &fullpath, &freepath);
2349 vn_lock(vp, LK_SHARED | LK_RETRY);
2350 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2351 kvo.kvo_vn_fileid = va.va_fileid;
2352 kvo.kvo_vn_fsid = va.va_fsid;
2353 }
2354 vput(vp);
2355 }
2356
2357 strlcpy(kvo.kvo_path, fullpath, sizeof(kvo.kvo_path));
2358 if (freepath != NULL)
2359 free(freepath, M_TEMP);
2360
2361 /* Pack record size down */
2362 kvo.kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) +
2363 strlen(kvo.kvo_path) + 1;
2364 kvo.kvo_structsize = roundup(kvo.kvo_structsize,
2365 sizeof(uint64_t));
2366 error = SYSCTL_OUT(req, &kvo, kvo.kvo_structsize);
2367 mtx_lock(&vm_object_list_mtx);
2368 if (error)
2369 break;
2370 }
2371 mtx_unlock(&vm_object_list_mtx);
2372 return (error);
2373 }
2374 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2375 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2376 "List of VM objects");
2377
2378 #include "opt_ddb.h"
2379 #ifdef DDB
2380 #include <sys/kernel.h>
2381
2382 #include <sys/cons.h>
2383
2384 #include <ddb/ddb.h>
2385
2386 static int
2387 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2388 {
2389 vm_map_t tmpm;
2390 vm_map_entry_t tmpe;
2391 vm_object_t obj;
2392 int entcount;
2393
2394 if (map == 0)
2395 return 0;
2396
2397 if (entry == 0) {
2398 tmpe = map->header.next;
2399 entcount = map->nentries;
2400 while (entcount-- && (tmpe != &map->header)) {
2401 if (_vm_object_in_map(map, object, tmpe)) {
2402 return 1;
2403 }
2404 tmpe = tmpe->next;
2405 }
2406 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2407 tmpm = entry->object.sub_map;
2408 tmpe = tmpm->header.next;
2409 entcount = tmpm->nentries;
2410 while (entcount-- && tmpe != &tmpm->header) {
2411 if (_vm_object_in_map(tmpm, object, tmpe)) {
2412 return 1;
2413 }
2414 tmpe = tmpe->next;
2415 }
2416 } else if ((obj = entry->object.vm_object) != NULL) {
2417 for (; obj; obj = obj->backing_object)
2418 if (obj == object) {
2419 return 1;
2420 }
2421 }
2422 return 0;
2423 }
2424
2425 static int
2426 vm_object_in_map(vm_object_t object)
2427 {
2428 struct proc *p;
2429
2430 /* sx_slock(&allproc_lock); */
2431 FOREACH_PROC_IN_SYSTEM(p) {
2432 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2433 continue;
2434 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2435 /* sx_sunlock(&allproc_lock); */
2436 return 1;
2437 }
2438 }
2439 /* sx_sunlock(&allproc_lock); */
2440 if (_vm_object_in_map(kernel_map, object, 0))
2441 return 1;
2442 return 0;
2443 }
2444
2445 DB_SHOW_COMMAND(vmochk, vm_object_check)
2446 {
2447 vm_object_t object;
2448
2449 /*
2450 * make sure that internal objs are in a map somewhere
2451 * and none have zero ref counts.
2452 */
2453 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2454 if (object->handle == NULL &&
2455 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2456 if (object->ref_count == 0) {
2457 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2458 (long)object->size);
2459 }
2460 if (!vm_object_in_map(object)) {
2461 db_printf(
2462 "vmochk: internal obj is not in a map: "
2463 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2464 object->ref_count, (u_long)object->size,
2465 (u_long)object->size,
2466 (void *)object->backing_object);
2467 }
2468 }
2469 }
2470 }
2471
2472 /*
2473 * vm_object_print: [ debug ]
2474 */
2475 DB_SHOW_COMMAND(object, vm_object_print_static)
2476 {
2477 /* XXX convert args. */
2478 vm_object_t object = (vm_object_t)addr;
2479 boolean_t full = have_addr;
2480
2481 vm_page_t p;
2482
2483 /* XXX count is an (unused) arg. Avoid shadowing it. */
2484 #define count was_count
2485
2486 int count;
2487
2488 if (object == NULL)
2489 return;
2490
2491 db_iprintf(
2492 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2493 object, (int)object->type, (uintmax_t)object->size,
2494 object->resident_page_count, object->ref_count, object->flags,
2495 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2496 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2497 object->shadow_count,
2498 object->backing_object ? object->backing_object->ref_count : 0,
2499 object->backing_object, (uintmax_t)object->backing_object_offset);
2500
2501 if (!full)
2502 return;
2503
2504 db_indent += 2;
2505 count = 0;
2506 TAILQ_FOREACH(p, &object->memq, listq) {
2507 if (count == 0)
2508 db_iprintf("memory:=");
2509 else if (count == 6) {
2510 db_printf("\n");
2511 db_iprintf(" ...");
2512 count = 0;
2513 } else
2514 db_printf(",");
2515 count++;
2516
2517 db_printf("(off=0x%jx,page=0x%jx)",
2518 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2519 }
2520 if (count != 0)
2521 db_printf("\n");
2522 db_indent -= 2;
2523 }
2524
2525 /* XXX. */
2526 #undef count
2527
2528 /* XXX need this non-static entry for calling from vm_map_print. */
2529 void
2530 vm_object_print(
2531 /* db_expr_t */ long addr,
2532 boolean_t have_addr,
2533 /* db_expr_t */ long count,
2534 char *modif)
2535 {
2536 vm_object_print_static(addr, have_addr, count, modif);
2537 }
2538
2539 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2540 {
2541 vm_object_t object;
2542 vm_pindex_t fidx;
2543 vm_paddr_t pa;
2544 vm_page_t m, prev_m;
2545 int rcount, nl, c;
2546
2547 nl = 0;
2548 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2549 db_printf("new object: %p\n", (void *)object);
2550 if (nl > 18) {
2551 c = cngetc();
2552 if (c != ' ')
2553 return;
2554 nl = 0;
2555 }
2556 nl++;
2557 rcount = 0;
2558 fidx = 0;
2559 pa = -1;
2560 TAILQ_FOREACH(m, &object->memq, listq) {
2561 if (m->pindex > 128)
2562 break;
2563 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2564 prev_m->pindex + 1 != m->pindex) {
2565 if (rcount) {
2566 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2567 (long)fidx, rcount, (long)pa);
2568 if (nl > 18) {
2569 c = cngetc();
2570 if (c != ' ')
2571 return;
2572 nl = 0;
2573 }
2574 nl++;
2575 rcount = 0;
2576 }
2577 }
2578 if (rcount &&
2579 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2580 ++rcount;
2581 continue;
2582 }
2583 if (rcount) {
2584 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2585 (long)fidx, rcount, (long)pa);
2586 if (nl > 18) {
2587 c = cngetc();
2588 if (c != ' ')
2589 return;
2590 nl = 0;
2591 }
2592 nl++;
2593 }
2594 fidx = m->pindex;
2595 pa = VM_PAGE_TO_PHYS(m);
2596 rcount = 1;
2597 }
2598 if (rcount) {
2599 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2600 (long)fidx, rcount, (long)pa);
2601 if (nl > 18) {
2602 c = cngetc();
2603 if (c != ' ')
2604 return;
2605 nl = 0;
2606 }
2607 nl++;
2608 }
2609 }
2610 }
2611 #endif /* DDB */
Cache object: cc5606ba3df00e2daf23096bc22600bc
|