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/11.0/sys/vm/vm_object.c 303290 2016-07-25 13:28:59Z 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/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(vm_object_cache_is_empty(object),
182 ("object %p has cached pages",
183 object));
184 KASSERT(object->paging_in_progress == 0,
185 ("object %p paging_in_progress = %d",
186 object, object->paging_in_progress));
187 KASSERT(object->resident_page_count == 0,
188 ("object %p resident_page_count = %d",
189 object, object->resident_page_count));
190 KASSERT(object->shadow_count == 0,
191 ("object %p shadow_count = %d",
192 object, object->shadow_count));
193 KASSERT(object->type == OBJT_DEAD,
194 ("object %p has non-dead type %d",
195 object, object->type));
196 }
197 #endif
198
199 static int
200 vm_object_zinit(void *mem, int size, int flags)
201 {
202 vm_object_t object;
203
204 object = (vm_object_t)mem;
205 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
206
207 /* These are true for any object that has been freed */
208 object->type = OBJT_DEAD;
209 object->ref_count = 0;
210 object->rtree.rt_root = 0;
211 object->rtree.rt_flags = 0;
212 object->paging_in_progress = 0;
213 object->resident_page_count = 0;
214 object->shadow_count = 0;
215 object->cache.rt_root = 0;
216 object->cache.rt_flags = 0;
217
218 mtx_lock(&vm_object_list_mtx);
219 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
220 mtx_unlock(&vm_object_list_mtx);
221 return (0);
222 }
223
224 static void
225 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
226 {
227
228 TAILQ_INIT(&object->memq);
229 LIST_INIT(&object->shadow_head);
230
231 object->type = type;
232 switch (type) {
233 case OBJT_DEAD:
234 panic("_vm_object_allocate: can't create OBJT_DEAD");
235 case OBJT_DEFAULT:
236 case OBJT_SWAP:
237 object->flags = OBJ_ONEMAPPING;
238 break;
239 case OBJT_DEVICE:
240 case OBJT_SG:
241 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
242 break;
243 case OBJT_MGTDEVICE:
244 object->flags = OBJ_FICTITIOUS;
245 break;
246 case OBJT_PHYS:
247 object->flags = OBJ_UNMANAGED;
248 break;
249 case OBJT_VNODE:
250 object->flags = 0;
251 break;
252 default:
253 panic("_vm_object_allocate: type %d is undefined", type);
254 }
255 object->size = size;
256 object->generation = 1;
257 object->ref_count = 1;
258 object->memattr = VM_MEMATTR_DEFAULT;
259 object->cred = NULL;
260 object->charge = 0;
261 object->handle = NULL;
262 object->backing_object = NULL;
263 object->backing_object_offset = (vm_ooffset_t) 0;
264 #if VM_NRESERVLEVEL > 0
265 LIST_INIT(&object->rvq);
266 #endif
267 umtx_shm_object_init(object);
268 }
269
270 /*
271 * vm_object_init:
272 *
273 * Initialize the VM objects module.
274 */
275 void
276 vm_object_init(void)
277 {
278 TAILQ_INIT(&vm_object_list);
279 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
280
281 rw_init(&kernel_object->lock, "kernel vm object");
282 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
283 kernel_object);
284 #if VM_NRESERVLEVEL > 0
285 kernel_object->flags |= OBJ_COLORED;
286 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
287 #endif
288
289 rw_init(&kmem_object->lock, "kmem vm object");
290 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
291 kmem_object);
292 #if VM_NRESERVLEVEL > 0
293 kmem_object->flags |= OBJ_COLORED;
294 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
295 #endif
296
297 /*
298 * The lock portion of struct vm_object must be type stable due
299 * to vm_pageout_fallback_object_lock locking a vm object
300 * without holding any references to it.
301 */
302 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
303 #ifdef INVARIANTS
304 vm_object_zdtor,
305 #else
306 NULL,
307 #endif
308 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
309
310 vm_radix_init();
311 }
312
313 void
314 vm_object_clear_flag(vm_object_t object, u_short bits)
315 {
316
317 VM_OBJECT_ASSERT_WLOCKED(object);
318 object->flags &= ~bits;
319 }
320
321 /*
322 * Sets the default memory attribute for the specified object. Pages
323 * that are allocated to this object are by default assigned this memory
324 * attribute.
325 *
326 * Presently, this function must be called before any pages are allocated
327 * to the object. In the future, this requirement may be relaxed for
328 * "default" and "swap" objects.
329 */
330 int
331 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
332 {
333
334 VM_OBJECT_ASSERT_WLOCKED(object);
335 switch (object->type) {
336 case OBJT_DEFAULT:
337 case OBJT_DEVICE:
338 case OBJT_MGTDEVICE:
339 case OBJT_PHYS:
340 case OBJT_SG:
341 case OBJT_SWAP:
342 case OBJT_VNODE:
343 if (!TAILQ_EMPTY(&object->memq))
344 return (KERN_FAILURE);
345 break;
346 case OBJT_DEAD:
347 return (KERN_INVALID_ARGUMENT);
348 default:
349 panic("vm_object_set_memattr: object %p is of undefined type",
350 object);
351 }
352 object->memattr = memattr;
353 return (KERN_SUCCESS);
354 }
355
356 void
357 vm_object_pip_add(vm_object_t object, short i)
358 {
359
360 VM_OBJECT_ASSERT_WLOCKED(object);
361 object->paging_in_progress += i;
362 }
363
364 void
365 vm_object_pip_subtract(vm_object_t object, short i)
366 {
367
368 VM_OBJECT_ASSERT_WLOCKED(object);
369 object->paging_in_progress -= i;
370 }
371
372 void
373 vm_object_pip_wakeup(vm_object_t object)
374 {
375
376 VM_OBJECT_ASSERT_WLOCKED(object);
377 object->paging_in_progress--;
378 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
379 vm_object_clear_flag(object, OBJ_PIPWNT);
380 wakeup(object);
381 }
382 }
383
384 void
385 vm_object_pip_wakeupn(vm_object_t object, short i)
386 {
387
388 VM_OBJECT_ASSERT_WLOCKED(object);
389 if (i)
390 object->paging_in_progress -= i;
391 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
392 vm_object_clear_flag(object, OBJ_PIPWNT);
393 wakeup(object);
394 }
395 }
396
397 void
398 vm_object_pip_wait(vm_object_t object, char *waitid)
399 {
400
401 VM_OBJECT_ASSERT_WLOCKED(object);
402 while (object->paging_in_progress) {
403 object->flags |= OBJ_PIPWNT;
404 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
405 }
406 }
407
408 /*
409 * vm_object_allocate:
410 *
411 * Returns a new object with the given size.
412 */
413 vm_object_t
414 vm_object_allocate(objtype_t type, vm_pindex_t size)
415 {
416 vm_object_t object;
417
418 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
419 _vm_object_allocate(type, size, object);
420 return (object);
421 }
422
423
424 /*
425 * vm_object_reference:
426 *
427 * Gets another reference to the given object. Note: OBJ_DEAD
428 * objects can be referenced during final cleaning.
429 */
430 void
431 vm_object_reference(vm_object_t object)
432 {
433 if (object == NULL)
434 return;
435 VM_OBJECT_WLOCK(object);
436 vm_object_reference_locked(object);
437 VM_OBJECT_WUNLOCK(object);
438 }
439
440 /*
441 * vm_object_reference_locked:
442 *
443 * Gets another reference to the given object.
444 *
445 * The object must be locked.
446 */
447 void
448 vm_object_reference_locked(vm_object_t object)
449 {
450 struct vnode *vp;
451
452 VM_OBJECT_ASSERT_WLOCKED(object);
453 object->ref_count++;
454 if (object->type == OBJT_VNODE) {
455 vp = object->handle;
456 vref(vp);
457 }
458 }
459
460 /*
461 * Handle deallocating an object of type OBJT_VNODE.
462 */
463 static void
464 vm_object_vndeallocate(vm_object_t object)
465 {
466 struct vnode *vp = (struct vnode *) object->handle;
467
468 VM_OBJECT_ASSERT_WLOCKED(object);
469 KASSERT(object->type == OBJT_VNODE,
470 ("vm_object_vndeallocate: not a vnode object"));
471 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
472 #ifdef INVARIANTS
473 if (object->ref_count == 0) {
474 vprint("vm_object_vndeallocate", vp);
475 panic("vm_object_vndeallocate: bad object reference count");
476 }
477 #endif
478
479 if (!umtx_shm_vnobj_persistent && object->ref_count == 1)
480 umtx_shm_object_terminated(object);
481
482 /*
483 * The test for text of vp vnode does not need a bypass to
484 * reach right VV_TEXT there, since it is obtained from
485 * object->handle.
486 */
487 if (object->ref_count > 1 || (vp->v_vflag & VV_TEXT) == 0) {
488 object->ref_count--;
489 VM_OBJECT_WUNLOCK(object);
490 /* vrele may need the vnode lock. */
491 vrele(vp);
492 } else {
493 vhold(vp);
494 VM_OBJECT_WUNLOCK(object);
495 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
496 vdrop(vp);
497 VM_OBJECT_WLOCK(object);
498 object->ref_count--;
499 if (object->type == OBJT_DEAD) {
500 VM_OBJECT_WUNLOCK(object);
501 VOP_UNLOCK(vp, 0);
502 } else {
503 if (object->ref_count == 0)
504 VOP_UNSET_TEXT(vp);
505 VM_OBJECT_WUNLOCK(object);
506 vput(vp);
507 }
508 }
509 }
510
511 /*
512 * vm_object_deallocate:
513 *
514 * Release a reference to the specified object,
515 * gained either through a vm_object_allocate
516 * or a vm_object_reference call. When all references
517 * are gone, storage associated with this object
518 * may be relinquished.
519 *
520 * No object may be locked.
521 */
522 void
523 vm_object_deallocate(vm_object_t object)
524 {
525 vm_object_t temp;
526 struct vnode *vp;
527
528 while (object != NULL) {
529 VM_OBJECT_WLOCK(object);
530 if (object->type == OBJT_VNODE) {
531 vm_object_vndeallocate(object);
532 return;
533 }
534
535 KASSERT(object->ref_count != 0,
536 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
537
538 /*
539 * If the reference count goes to 0 we start calling
540 * vm_object_terminate() on the object chain.
541 * A ref count of 1 may be a special case depending on the
542 * shadow count being 0 or 1.
543 */
544 object->ref_count--;
545 if (object->ref_count > 1) {
546 VM_OBJECT_WUNLOCK(object);
547 return;
548 } else if (object->ref_count == 1) {
549 if (object->type == OBJT_SWAP &&
550 (object->flags & OBJ_TMPFS) != 0) {
551 vp = object->un_pager.swp.swp_tmpfs;
552 vhold(vp);
553 VM_OBJECT_WUNLOCK(object);
554 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
555 VM_OBJECT_WLOCK(object);
556 if (object->type == OBJT_DEAD ||
557 object->ref_count != 1) {
558 VM_OBJECT_WUNLOCK(object);
559 VOP_UNLOCK(vp, 0);
560 vdrop(vp);
561 return;
562 }
563 if ((object->flags & OBJ_TMPFS) != 0)
564 VOP_UNSET_TEXT(vp);
565 VOP_UNLOCK(vp, 0);
566 vdrop(vp);
567 }
568 if (object->shadow_count == 0 &&
569 object->handle == NULL &&
570 (object->type == OBJT_DEFAULT ||
571 (object->type == OBJT_SWAP &&
572 (object->flags & OBJ_TMPFS_NODE) == 0))) {
573 vm_object_set_flag(object, OBJ_ONEMAPPING);
574 } else if ((object->shadow_count == 1) &&
575 (object->handle == NULL) &&
576 (object->type == OBJT_DEFAULT ||
577 object->type == OBJT_SWAP)) {
578 vm_object_t robject;
579
580 robject = LIST_FIRST(&object->shadow_head);
581 KASSERT(robject != NULL,
582 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
583 object->ref_count,
584 object->shadow_count));
585 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0,
586 ("shadowed tmpfs v_object %p", object));
587 if (!VM_OBJECT_TRYWLOCK(robject)) {
588 /*
589 * Avoid a potential deadlock.
590 */
591 object->ref_count++;
592 VM_OBJECT_WUNLOCK(object);
593 /*
594 * More likely than not the thread
595 * holding robject's lock has lower
596 * priority than the current thread.
597 * Let the lower priority thread run.
598 */
599 pause("vmo_de", 1);
600 continue;
601 }
602 /*
603 * Collapse object into its shadow unless its
604 * shadow is dead. In that case, object will
605 * be deallocated by the thread that is
606 * deallocating its shadow.
607 */
608 if ((robject->flags & OBJ_DEAD) == 0 &&
609 (robject->handle == NULL) &&
610 (robject->type == OBJT_DEFAULT ||
611 robject->type == OBJT_SWAP)) {
612
613 robject->ref_count++;
614 retry:
615 if (robject->paging_in_progress) {
616 VM_OBJECT_WUNLOCK(object);
617 vm_object_pip_wait(robject,
618 "objde1");
619 temp = robject->backing_object;
620 if (object == temp) {
621 VM_OBJECT_WLOCK(object);
622 goto retry;
623 }
624 } else if (object->paging_in_progress) {
625 VM_OBJECT_WUNLOCK(robject);
626 object->flags |= OBJ_PIPWNT;
627 VM_OBJECT_SLEEP(object, object,
628 PDROP | PVM, "objde2", 0);
629 VM_OBJECT_WLOCK(robject);
630 temp = robject->backing_object;
631 if (object == temp) {
632 VM_OBJECT_WLOCK(object);
633 goto retry;
634 }
635 } else
636 VM_OBJECT_WUNLOCK(object);
637
638 if (robject->ref_count == 1) {
639 robject->ref_count--;
640 object = robject;
641 goto doterm;
642 }
643 object = robject;
644 vm_object_collapse(object);
645 VM_OBJECT_WUNLOCK(object);
646 continue;
647 }
648 VM_OBJECT_WUNLOCK(robject);
649 }
650 VM_OBJECT_WUNLOCK(object);
651 return;
652 }
653 doterm:
654 umtx_shm_object_terminated(object);
655 temp = object->backing_object;
656 if (temp != NULL) {
657 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
658 ("shadowed tmpfs v_object 2 %p", object));
659 VM_OBJECT_WLOCK(temp);
660 LIST_REMOVE(object, shadow_list);
661 temp->shadow_count--;
662 VM_OBJECT_WUNLOCK(temp);
663 object->backing_object = NULL;
664 }
665 /*
666 * Don't double-terminate, we could be in a termination
667 * recursion due to the terminate having to sync data
668 * to disk.
669 */
670 if ((object->flags & OBJ_DEAD) == 0)
671 vm_object_terminate(object);
672 else
673 VM_OBJECT_WUNLOCK(object);
674 object = temp;
675 }
676 }
677
678 /*
679 * vm_object_destroy removes the object from the global object list
680 * and frees the space for the object.
681 */
682 void
683 vm_object_destroy(vm_object_t object)
684 {
685
686 /*
687 * Release the allocation charge.
688 */
689 if (object->cred != NULL) {
690 swap_release_by_cred(object->charge, object->cred);
691 object->charge = 0;
692 crfree(object->cred);
693 object->cred = NULL;
694 }
695
696 /*
697 * Free the space for the object.
698 */
699 uma_zfree(obj_zone, object);
700 }
701
702 /*
703 * vm_object_terminate actually destroys the specified object, freeing
704 * up all previously used resources.
705 *
706 * The object must be locked.
707 * This routine may block.
708 */
709 void
710 vm_object_terminate(vm_object_t object)
711 {
712 vm_page_t p, p_next;
713
714 VM_OBJECT_ASSERT_WLOCKED(object);
715
716 /*
717 * Make sure no one uses us.
718 */
719 vm_object_set_flag(object, OBJ_DEAD);
720
721 /*
722 * wait for the pageout daemon to be done with the object
723 */
724 vm_object_pip_wait(object, "objtrm");
725
726 KASSERT(!object->paging_in_progress,
727 ("vm_object_terminate: pageout in progress"));
728
729 /*
730 * Clean and free the pages, as appropriate. All references to the
731 * object are gone, so we don't need to lock it.
732 */
733 if (object->type == OBJT_VNODE) {
734 struct vnode *vp = (struct vnode *)object->handle;
735
736 /*
737 * Clean pages and flush buffers.
738 */
739 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
740 VM_OBJECT_WUNLOCK(object);
741
742 vinvalbuf(vp, V_SAVE, 0, 0);
743
744 BO_LOCK(&vp->v_bufobj);
745 vp->v_bufobj.bo_flag |= BO_DEAD;
746 BO_UNLOCK(&vp->v_bufobj);
747
748 VM_OBJECT_WLOCK(object);
749 }
750
751 KASSERT(object->ref_count == 0,
752 ("vm_object_terminate: object with references, ref_count=%d",
753 object->ref_count));
754
755 /*
756 * Free any remaining pageable pages. This also removes them from the
757 * paging queues. However, don't free wired pages, just remove them
758 * from the object. Rather than incrementally removing each page from
759 * the object, the page and object are reset to any empty state.
760 */
761 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
762 vm_page_assert_unbusied(p);
763 vm_page_lock(p);
764 /*
765 * Optimize the page's removal from the object by resetting
766 * its "object" field. Specifically, if the page is not
767 * wired, then the effect of this assignment is that
768 * vm_page_free()'s call to vm_page_remove() will return
769 * immediately without modifying the page or the object.
770 */
771 p->object = NULL;
772 if (p->wire_count == 0) {
773 vm_page_free(p);
774 PCPU_INC(cnt.v_pfree);
775 }
776 vm_page_unlock(p);
777 }
778 /*
779 * If the object contained any pages, then reset it to an empty state.
780 * None of the object's fields, including "resident_page_count", were
781 * modified by the preceding loop.
782 */
783 if (object->resident_page_count != 0) {
784 vm_radix_reclaim_allnodes(&object->rtree);
785 TAILQ_INIT(&object->memq);
786 object->resident_page_count = 0;
787 if (object->type == OBJT_VNODE)
788 vdrop(object->handle);
789 }
790
791 #if VM_NRESERVLEVEL > 0
792 if (__predict_false(!LIST_EMPTY(&object->rvq)))
793 vm_reserv_break_all(object);
794 #endif
795 if (__predict_false(!vm_object_cache_is_empty(object)))
796 vm_page_cache_free(object, 0, 0);
797
798 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
799 object->type == OBJT_SWAP,
800 ("%s: non-swap obj %p has cred", __func__, object));
801
802 /*
803 * Let the pager know object is dead.
804 */
805 vm_pager_deallocate(object);
806 VM_OBJECT_WUNLOCK(object);
807
808 vm_object_destroy(object);
809 }
810
811 /*
812 * Make the page read-only so that we can clear the object flags. However, if
813 * this is a nosync mmap then the object is likely to stay dirty so do not
814 * mess with the page and do not clear the object flags. Returns TRUE if the
815 * page should be flushed, and FALSE otherwise.
816 */
817 static boolean_t
818 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
819 {
820
821 /*
822 * If we have been asked to skip nosync pages and this is a
823 * nosync page, skip it. Note that the object flags were not
824 * cleared in this case so we do not have to set them.
825 */
826 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
827 *clearobjflags = FALSE;
828 return (FALSE);
829 } else {
830 pmap_remove_write(p);
831 return (p->dirty != 0);
832 }
833 }
834
835 /*
836 * vm_object_page_clean
837 *
838 * Clean all dirty pages in the specified range of object. Leaves page
839 * on whatever queue it is currently on. If NOSYNC is set then do not
840 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
841 * leaving the object dirty.
842 *
843 * When stuffing pages asynchronously, allow clustering. XXX we need a
844 * synchronous clustering mode implementation.
845 *
846 * Odd semantics: if start == end, we clean everything.
847 *
848 * The object must be locked.
849 *
850 * Returns FALSE if some page from the range was not written, as
851 * reported by the pager, and TRUE otherwise.
852 */
853 boolean_t
854 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
855 int flags)
856 {
857 vm_page_t np, p;
858 vm_pindex_t pi, tend, tstart;
859 int curgeneration, n, pagerflags;
860 boolean_t clearobjflags, eio, res;
861
862 VM_OBJECT_ASSERT_WLOCKED(object);
863
864 /*
865 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
866 * objects. The check below prevents the function from
867 * operating on non-vnode objects.
868 */
869 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
870 object->resident_page_count == 0)
871 return (TRUE);
872
873 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
874 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
875 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
876
877 tstart = OFF_TO_IDX(start);
878 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
879 clearobjflags = tstart == 0 && tend >= object->size;
880 res = TRUE;
881
882 rescan:
883 curgeneration = object->generation;
884
885 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
886 pi = p->pindex;
887 if (pi >= tend)
888 break;
889 np = TAILQ_NEXT(p, listq);
890 if (p->valid == 0)
891 continue;
892 if (vm_page_sleep_if_busy(p, "vpcwai")) {
893 if (object->generation != curgeneration) {
894 if ((flags & OBJPC_SYNC) != 0)
895 goto rescan;
896 else
897 clearobjflags = FALSE;
898 }
899 np = vm_page_find_least(object, pi);
900 continue;
901 }
902 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
903 continue;
904
905 n = vm_object_page_collect_flush(object, p, pagerflags,
906 flags, &clearobjflags, &eio);
907 if (eio) {
908 res = FALSE;
909 clearobjflags = FALSE;
910 }
911 if (object->generation != curgeneration) {
912 if ((flags & OBJPC_SYNC) != 0)
913 goto rescan;
914 else
915 clearobjflags = FALSE;
916 }
917
918 /*
919 * If the VOP_PUTPAGES() did a truncated write, so
920 * that even the first page of the run is not fully
921 * written, vm_pageout_flush() returns 0 as the run
922 * length. Since the condition that caused truncated
923 * write may be permanent, e.g. exhausted free space,
924 * accepting n == 0 would cause an infinite loop.
925 *
926 * Forwarding the iterator leaves the unwritten page
927 * behind, but there is not much we can do there if
928 * filesystem refuses to write it.
929 */
930 if (n == 0) {
931 n = 1;
932 clearobjflags = FALSE;
933 }
934 np = vm_page_find_least(object, pi + n);
935 }
936 #if 0
937 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
938 #endif
939
940 if (clearobjflags)
941 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
942 return (res);
943 }
944
945 static int
946 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
947 int flags, boolean_t *clearobjflags, boolean_t *eio)
948 {
949 vm_page_t ma[vm_pageout_page_count], p_first, tp;
950 int count, i, mreq, runlen;
951
952 vm_page_lock_assert(p, MA_NOTOWNED);
953 VM_OBJECT_ASSERT_WLOCKED(object);
954
955 count = 1;
956 mreq = 0;
957
958 for (tp = p; count < vm_pageout_page_count; count++) {
959 tp = vm_page_next(tp);
960 if (tp == NULL || vm_page_busied(tp))
961 break;
962 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
963 break;
964 }
965
966 for (p_first = p; count < vm_pageout_page_count; count++) {
967 tp = vm_page_prev(p_first);
968 if (tp == NULL || vm_page_busied(tp))
969 break;
970 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
971 break;
972 p_first = tp;
973 mreq++;
974 }
975
976 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
977 ma[i] = tp;
978
979 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
980 return (runlen);
981 }
982
983 /*
984 * Note that there is absolutely no sense in writing out
985 * anonymous objects, so we track down the vnode object
986 * to write out.
987 * We invalidate (remove) all pages from the address space
988 * for semantic correctness.
989 *
990 * If the backing object is a device object with unmanaged pages, then any
991 * mappings to the specified range of pages must be removed before this
992 * function is called.
993 *
994 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
995 * may start out with a NULL object.
996 */
997 boolean_t
998 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
999 boolean_t syncio, boolean_t invalidate)
1000 {
1001 vm_object_t backing_object;
1002 struct vnode *vp;
1003 struct mount *mp;
1004 int error, flags, fsync_after;
1005 boolean_t res;
1006
1007 if (object == NULL)
1008 return (TRUE);
1009 res = TRUE;
1010 error = 0;
1011 VM_OBJECT_WLOCK(object);
1012 while ((backing_object = object->backing_object) != NULL) {
1013 VM_OBJECT_WLOCK(backing_object);
1014 offset += object->backing_object_offset;
1015 VM_OBJECT_WUNLOCK(object);
1016 object = backing_object;
1017 if (object->size < OFF_TO_IDX(offset + size))
1018 size = IDX_TO_OFF(object->size) - offset;
1019 }
1020 /*
1021 * Flush pages if writing is allowed, invalidate them
1022 * if invalidation requested. Pages undergoing I/O
1023 * will be ignored by vm_object_page_remove().
1024 *
1025 * We cannot lock the vnode and then wait for paging
1026 * to complete without deadlocking against vm_fault.
1027 * Instead we simply call vm_object_page_remove() and
1028 * allow it to block internally on a page-by-page
1029 * basis when it encounters pages undergoing async
1030 * I/O.
1031 */
1032 if (object->type == OBJT_VNODE &&
1033 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
1034 vp = object->handle;
1035 VM_OBJECT_WUNLOCK(object);
1036 (void) vn_start_write(vp, &mp, V_WAIT);
1037 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1038 if (syncio && !invalidate && offset == 0 &&
1039 OFF_TO_IDX(size) == object->size) {
1040 /*
1041 * If syncing the whole mapping of the file,
1042 * it is faster to schedule all the writes in
1043 * async mode, also allowing the clustering,
1044 * and then wait for i/o to complete.
1045 */
1046 flags = 0;
1047 fsync_after = TRUE;
1048 } else {
1049 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1050 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1051 fsync_after = FALSE;
1052 }
1053 VM_OBJECT_WLOCK(object);
1054 res = vm_object_page_clean(object, offset, offset + size,
1055 flags);
1056 VM_OBJECT_WUNLOCK(object);
1057 if (fsync_after)
1058 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1059 VOP_UNLOCK(vp, 0);
1060 vn_finished_write(mp);
1061 if (error != 0)
1062 res = FALSE;
1063 VM_OBJECT_WLOCK(object);
1064 }
1065 if ((object->type == OBJT_VNODE ||
1066 object->type == OBJT_DEVICE) && invalidate) {
1067 if (object->type == OBJT_DEVICE)
1068 /*
1069 * The option OBJPR_NOTMAPPED must be passed here
1070 * because vm_object_page_remove() cannot remove
1071 * unmanaged mappings.
1072 */
1073 flags = OBJPR_NOTMAPPED;
1074 else if (old_msync)
1075 flags = 0;
1076 else
1077 flags = OBJPR_CLEANONLY;
1078 vm_object_page_remove(object, OFF_TO_IDX(offset),
1079 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1080 }
1081 VM_OBJECT_WUNLOCK(object);
1082 return (res);
1083 }
1084
1085 /*
1086 * vm_object_madvise:
1087 *
1088 * Implements the madvise function at the object/page level.
1089 *
1090 * MADV_WILLNEED (any object)
1091 *
1092 * Activate the specified pages if they are resident.
1093 *
1094 * MADV_DONTNEED (any object)
1095 *
1096 * Deactivate the specified pages if they are resident.
1097 *
1098 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1099 * OBJ_ONEMAPPING only)
1100 *
1101 * Deactivate and clean the specified pages if they are
1102 * resident. This permits the process to reuse the pages
1103 * without faulting or the kernel to reclaim the pages
1104 * without I/O.
1105 */
1106 void
1107 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1108 int advise)
1109 {
1110 vm_pindex_t tpindex;
1111 vm_object_t backing_object, tobject;
1112 vm_page_t m;
1113
1114 if (object == NULL)
1115 return;
1116 VM_OBJECT_WLOCK(object);
1117 /*
1118 * Locate and adjust resident pages
1119 */
1120 for (; pindex < end; pindex += 1) {
1121 relookup:
1122 tobject = object;
1123 tpindex = pindex;
1124 shadowlookup:
1125 /*
1126 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1127 * and those pages must be OBJ_ONEMAPPING.
1128 */
1129 if (advise == MADV_FREE) {
1130 if ((tobject->type != OBJT_DEFAULT &&
1131 tobject->type != OBJT_SWAP) ||
1132 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1133 goto unlock_tobject;
1134 }
1135 } else if ((tobject->flags & OBJ_UNMANAGED) != 0)
1136 goto unlock_tobject;
1137 m = vm_page_lookup(tobject, tpindex);
1138 if (m == NULL && advise == MADV_WILLNEED) {
1139 /*
1140 * If the page is cached, reactivate it.
1141 */
1142 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1143 VM_ALLOC_NOBUSY);
1144 }
1145 if (m == NULL) {
1146 /*
1147 * There may be swap even if there is no backing page
1148 */
1149 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1150 swap_pager_freespace(tobject, tpindex, 1);
1151 /*
1152 * next object
1153 */
1154 backing_object = tobject->backing_object;
1155 if (backing_object == NULL)
1156 goto unlock_tobject;
1157 VM_OBJECT_WLOCK(backing_object);
1158 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1159 if (tobject != object)
1160 VM_OBJECT_WUNLOCK(tobject);
1161 tobject = backing_object;
1162 goto shadowlookup;
1163 } else if (m->valid != VM_PAGE_BITS_ALL)
1164 goto unlock_tobject;
1165 /*
1166 * If the page is not in a normal state, skip it.
1167 */
1168 vm_page_lock(m);
1169 if (m->hold_count != 0 || m->wire_count != 0) {
1170 vm_page_unlock(m);
1171 goto unlock_tobject;
1172 }
1173 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1174 ("vm_object_madvise: page %p is fictitious", m));
1175 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1176 ("vm_object_madvise: page %p is not managed", m));
1177 if (vm_page_busied(m)) {
1178 if (advise == MADV_WILLNEED) {
1179 /*
1180 * Reference the page before unlocking and
1181 * sleeping so that the page daemon is less
1182 * likely to reclaim it.
1183 */
1184 vm_page_aflag_set(m, PGA_REFERENCED);
1185 }
1186 if (object != tobject)
1187 VM_OBJECT_WUNLOCK(object);
1188 VM_OBJECT_WUNLOCK(tobject);
1189 vm_page_busy_sleep(m, "madvpo");
1190 VM_OBJECT_WLOCK(object);
1191 goto relookup;
1192 }
1193 if (advise == MADV_WILLNEED) {
1194 vm_page_activate(m);
1195 } else {
1196 vm_page_advise(m, advise);
1197 }
1198 vm_page_unlock(m);
1199 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1200 swap_pager_freespace(tobject, tpindex, 1);
1201 unlock_tobject:
1202 if (tobject != object)
1203 VM_OBJECT_WUNLOCK(tobject);
1204 }
1205 VM_OBJECT_WUNLOCK(object);
1206 }
1207
1208 /*
1209 * vm_object_shadow:
1210 *
1211 * Create a new object which is backed by the
1212 * specified existing object range. The source
1213 * object reference is deallocated.
1214 *
1215 * The new object and offset into that object
1216 * are returned in the source parameters.
1217 */
1218 void
1219 vm_object_shadow(
1220 vm_object_t *object, /* IN/OUT */
1221 vm_ooffset_t *offset, /* IN/OUT */
1222 vm_size_t length)
1223 {
1224 vm_object_t source;
1225 vm_object_t result;
1226
1227 source = *object;
1228
1229 /*
1230 * Don't create the new object if the old object isn't shared.
1231 */
1232 if (source != NULL) {
1233 VM_OBJECT_WLOCK(source);
1234 if (source->ref_count == 1 &&
1235 source->handle == NULL &&
1236 (source->type == OBJT_DEFAULT ||
1237 source->type == OBJT_SWAP)) {
1238 VM_OBJECT_WUNLOCK(source);
1239 return;
1240 }
1241 VM_OBJECT_WUNLOCK(source);
1242 }
1243
1244 /*
1245 * Allocate a new object with the given length.
1246 */
1247 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1248
1249 /*
1250 * The new object shadows the source object, adding a reference to it.
1251 * Our caller changes his reference to point to the new object,
1252 * removing a reference to the source object. Net result: no change
1253 * of reference count.
1254 *
1255 * Try to optimize the result object's page color when shadowing
1256 * in order to maintain page coloring consistency in the combined
1257 * shadowed object.
1258 */
1259 result->backing_object = source;
1260 /*
1261 * Store the offset into the source object, and fix up the offset into
1262 * the new object.
1263 */
1264 result->backing_object_offset = *offset;
1265 if (source != NULL) {
1266 VM_OBJECT_WLOCK(source);
1267 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1268 source->shadow_count++;
1269 #if VM_NRESERVLEVEL > 0
1270 result->flags |= source->flags & OBJ_COLORED;
1271 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1272 ((1 << (VM_NFREEORDER - 1)) - 1);
1273 #endif
1274 VM_OBJECT_WUNLOCK(source);
1275 }
1276
1277
1278 /*
1279 * Return the new things
1280 */
1281 *offset = 0;
1282 *object = result;
1283 }
1284
1285 /*
1286 * vm_object_split:
1287 *
1288 * Split the pages in a map entry into a new object. This affords
1289 * easier removal of unused pages, and keeps object inheritance from
1290 * being a negative impact on memory usage.
1291 */
1292 void
1293 vm_object_split(vm_map_entry_t entry)
1294 {
1295 vm_page_t m, m_next;
1296 vm_object_t orig_object, new_object, source;
1297 vm_pindex_t idx, offidxstart;
1298 vm_size_t size;
1299
1300 orig_object = entry->object.vm_object;
1301 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1302 return;
1303 if (orig_object->ref_count <= 1)
1304 return;
1305 VM_OBJECT_WUNLOCK(orig_object);
1306
1307 offidxstart = OFF_TO_IDX(entry->offset);
1308 size = atop(entry->end - entry->start);
1309
1310 /*
1311 * If swap_pager_copy() is later called, it will convert new_object
1312 * into a swap object.
1313 */
1314 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1315
1316 /*
1317 * At this point, the new object is still private, so the order in
1318 * which the original and new objects are locked does not matter.
1319 */
1320 VM_OBJECT_WLOCK(new_object);
1321 VM_OBJECT_WLOCK(orig_object);
1322 source = orig_object->backing_object;
1323 if (source != NULL) {
1324 VM_OBJECT_WLOCK(source);
1325 if ((source->flags & OBJ_DEAD) != 0) {
1326 VM_OBJECT_WUNLOCK(source);
1327 VM_OBJECT_WUNLOCK(orig_object);
1328 VM_OBJECT_WUNLOCK(new_object);
1329 vm_object_deallocate(new_object);
1330 VM_OBJECT_WLOCK(orig_object);
1331 return;
1332 }
1333 LIST_INSERT_HEAD(&source->shadow_head,
1334 new_object, shadow_list);
1335 source->shadow_count++;
1336 vm_object_reference_locked(source); /* for new_object */
1337 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1338 VM_OBJECT_WUNLOCK(source);
1339 new_object->backing_object_offset =
1340 orig_object->backing_object_offset + entry->offset;
1341 new_object->backing_object = source;
1342 }
1343 if (orig_object->cred != NULL) {
1344 new_object->cred = orig_object->cred;
1345 crhold(orig_object->cred);
1346 new_object->charge = ptoa(size);
1347 KASSERT(orig_object->charge >= ptoa(size),
1348 ("orig_object->charge < 0"));
1349 orig_object->charge -= ptoa(size);
1350 }
1351 retry:
1352 m = vm_page_find_least(orig_object, offidxstart);
1353 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1354 m = m_next) {
1355 m_next = TAILQ_NEXT(m, listq);
1356
1357 /*
1358 * We must wait for pending I/O to complete before we can
1359 * rename the page.
1360 *
1361 * We do not have to VM_PROT_NONE the page as mappings should
1362 * not be changed by this operation.
1363 */
1364 if (vm_page_busied(m)) {
1365 VM_OBJECT_WUNLOCK(new_object);
1366 vm_page_lock(m);
1367 VM_OBJECT_WUNLOCK(orig_object);
1368 vm_page_busy_sleep(m, "spltwt");
1369 VM_OBJECT_WLOCK(orig_object);
1370 VM_OBJECT_WLOCK(new_object);
1371 goto retry;
1372 }
1373
1374 /* vm_page_rename() will handle dirty and cache. */
1375 if (vm_page_rename(m, new_object, idx)) {
1376 VM_OBJECT_WUNLOCK(new_object);
1377 VM_OBJECT_WUNLOCK(orig_object);
1378 VM_WAIT;
1379 VM_OBJECT_WLOCK(orig_object);
1380 VM_OBJECT_WLOCK(new_object);
1381 goto retry;
1382 }
1383 #if VM_NRESERVLEVEL > 0
1384 /*
1385 * If some of the reservation's allocated pages remain with
1386 * the original object, then transferring the reservation to
1387 * the new object is neither particularly beneficial nor
1388 * particularly harmful as compared to leaving the reservation
1389 * with the original object. If, however, all of the
1390 * reservation's allocated pages are transferred to the new
1391 * object, then transferring the reservation is typically
1392 * beneficial. Determining which of these two cases applies
1393 * would be more costly than unconditionally renaming the
1394 * reservation.
1395 */
1396 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1397 #endif
1398 if (orig_object->type == OBJT_SWAP)
1399 vm_page_xbusy(m);
1400 }
1401 if (orig_object->type == OBJT_SWAP) {
1402 /*
1403 * swap_pager_copy() can sleep, in which case the orig_object's
1404 * and new_object's locks are released and reacquired.
1405 */
1406 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1407 TAILQ_FOREACH(m, &new_object->memq, listq)
1408 vm_page_xunbusy(m);
1409
1410 /*
1411 * Transfer any cached pages from orig_object to new_object.
1412 * If swap_pager_copy() found swapped out pages within the
1413 * specified range of orig_object, then it changed
1414 * new_object's type to OBJT_SWAP when it transferred those
1415 * pages to new_object. Otherwise, new_object's type
1416 * should still be OBJT_DEFAULT and orig_object should not
1417 * contain any cached pages within the specified range.
1418 */
1419 if (__predict_false(!vm_object_cache_is_empty(orig_object)))
1420 vm_page_cache_transfer(orig_object, offidxstart,
1421 new_object);
1422 }
1423 VM_OBJECT_WUNLOCK(orig_object);
1424 VM_OBJECT_WUNLOCK(new_object);
1425 entry->object.vm_object = new_object;
1426 entry->offset = 0LL;
1427 vm_object_deallocate(orig_object);
1428 VM_OBJECT_WLOCK(new_object);
1429 }
1430
1431 #define OBSC_COLLAPSE_NOWAIT 0x0002
1432 #define OBSC_COLLAPSE_WAIT 0x0004
1433
1434 static vm_page_t
1435 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next,
1436 int op)
1437 {
1438 vm_object_t backing_object;
1439
1440 VM_OBJECT_ASSERT_WLOCKED(object);
1441 backing_object = object->backing_object;
1442 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1443
1444 KASSERT(p == NULL || vm_page_busied(p), ("unbusy page %p", p));
1445 KASSERT(p == NULL || p->object == object || p->object == backing_object,
1446 ("invalid ownership %p %p %p", p, object, backing_object));
1447 if ((op & OBSC_COLLAPSE_NOWAIT) != 0)
1448 return (next);
1449 if (p != NULL)
1450 vm_page_lock(p);
1451 VM_OBJECT_WUNLOCK(object);
1452 VM_OBJECT_WUNLOCK(backing_object);
1453 if (p == NULL)
1454 VM_WAIT;
1455 else
1456 vm_page_busy_sleep(p, "vmocol");
1457 VM_OBJECT_WLOCK(object);
1458 VM_OBJECT_WLOCK(backing_object);
1459 return (TAILQ_FIRST(&backing_object->memq));
1460 }
1461
1462 static bool
1463 vm_object_scan_all_shadowed(vm_object_t object)
1464 {
1465 vm_object_t backing_object;
1466 vm_page_t p, pp;
1467 vm_pindex_t backing_offset_index, new_pindex;
1468
1469 VM_OBJECT_ASSERT_WLOCKED(object);
1470 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1471
1472 backing_object = object->backing_object;
1473
1474 /*
1475 * Initial conditions:
1476 *
1477 * We do not want to have to test for the existence of cache or swap
1478 * pages in the backing object. XXX but with the new swapper this
1479 * would be pretty easy to do.
1480 */
1481 if (backing_object->type != OBJT_DEFAULT)
1482 return (false);
1483
1484 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1485
1486 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL;
1487 p = TAILQ_NEXT(p, listq)) {
1488 new_pindex = p->pindex - backing_offset_index;
1489
1490 /*
1491 * Ignore pages outside the parent object's range and outside
1492 * the parent object's mapping of the backing object.
1493 */
1494 if (p->pindex < backing_offset_index ||
1495 new_pindex >= object->size)
1496 continue;
1497
1498 /*
1499 * See if the parent has the page or if the parent's object
1500 * pager has the page. If the parent has the page but the page
1501 * is not valid, the parent's object pager must have the page.
1502 *
1503 * If this fails, the parent does not completely shadow the
1504 * object and we might as well give up now.
1505 */
1506 pp = vm_page_lookup(object, new_pindex);
1507 if ((pp == NULL || pp->valid == 0) &&
1508 !vm_pager_has_page(object, new_pindex, NULL, NULL))
1509 return (false);
1510 }
1511 return (true);
1512 }
1513
1514 static bool
1515 vm_object_collapse_scan(vm_object_t object, int op)
1516 {
1517 vm_object_t backing_object;
1518 vm_page_t next, p, pp;
1519 vm_pindex_t backing_offset_index, new_pindex;
1520
1521 VM_OBJECT_ASSERT_WLOCKED(object);
1522 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1523
1524 backing_object = object->backing_object;
1525 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1526
1527 /*
1528 * Initial conditions
1529 */
1530 if ((op & OBSC_COLLAPSE_WAIT) != 0)
1531 vm_object_set_flag(backing_object, OBJ_DEAD);
1532
1533 /*
1534 * Our scan
1535 */
1536 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1537 next = TAILQ_NEXT(p, listq);
1538 new_pindex = p->pindex - backing_offset_index;
1539
1540 /*
1541 * Check for busy page
1542 */
1543 if (vm_page_busied(p)) {
1544 next = vm_object_collapse_scan_wait(object, p, next, op);
1545 continue;
1546 }
1547
1548 KASSERT(p->object == backing_object,
1549 ("vm_object_collapse_scan: object mismatch"));
1550
1551 if (p->pindex < backing_offset_index ||
1552 new_pindex >= object->size) {
1553 if (backing_object->type == OBJT_SWAP)
1554 swap_pager_freespace(backing_object, p->pindex,
1555 1);
1556
1557 /*
1558 * Page is out of the parent object's range, we can
1559 * simply destroy it.
1560 */
1561 vm_page_lock(p);
1562 KASSERT(!pmap_page_is_mapped(p),
1563 ("freeing mapped page %p", p));
1564 if (p->wire_count == 0)
1565 vm_page_free(p);
1566 else
1567 vm_page_remove(p);
1568 vm_page_unlock(p);
1569 continue;
1570 }
1571
1572 pp = vm_page_lookup(object, new_pindex);
1573 if (pp != NULL && vm_page_busied(pp)) {
1574 /*
1575 * The page in the parent is busy and possibly not
1576 * (yet) valid. Until its state is finalized by the
1577 * busy bit owner, we can't tell whether it shadows the
1578 * original page. Therefore, we must either skip it
1579 * and the original (backing_object) page or wait for
1580 * its state to be finalized.
1581 *
1582 * This is due to a race with vm_fault() where we must
1583 * unbusy the original (backing_obj) page before we can
1584 * (re)lock the parent. Hence we can get here.
1585 */
1586 next = vm_object_collapse_scan_wait(object, pp, next,
1587 op);
1588 continue;
1589 }
1590
1591 KASSERT(pp == NULL || pp->valid != 0,
1592 ("unbusy invalid page %p", pp));
1593
1594 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1595 NULL)) {
1596 /*
1597 * The page already exists in the parent OR swap exists
1598 * for this location in the parent. Leave the parent's
1599 * page alone. Destroy the original page from the
1600 * backing object.
1601 */
1602 if (backing_object->type == OBJT_SWAP)
1603 swap_pager_freespace(backing_object, p->pindex,
1604 1);
1605 vm_page_lock(p);
1606 KASSERT(!pmap_page_is_mapped(p),
1607 ("freeing mapped page %p", p));
1608 if (p->wire_count == 0)
1609 vm_page_free(p);
1610 else
1611 vm_page_remove(p);
1612 vm_page_unlock(p);
1613 continue;
1614 }
1615
1616 /*
1617 * Page does not exist in parent, rename the page from the
1618 * backing object to the main object.
1619 *
1620 * If the page was mapped to a process, it can remain mapped
1621 * through the rename. vm_page_rename() will handle dirty and
1622 * cache.
1623 */
1624 if (vm_page_rename(p, object, new_pindex)) {
1625 next = vm_object_collapse_scan_wait(object, NULL, next,
1626 op);
1627 continue;
1628 }
1629
1630 /* Use the old pindex to free the right page. */
1631 if (backing_object->type == OBJT_SWAP)
1632 swap_pager_freespace(backing_object,
1633 new_pindex + backing_offset_index, 1);
1634
1635 #if VM_NRESERVLEVEL > 0
1636 /*
1637 * Rename the reservation.
1638 */
1639 vm_reserv_rename(p, object, backing_object,
1640 backing_offset_index);
1641 #endif
1642 }
1643 return (true);
1644 }
1645
1646
1647 /*
1648 * this version of collapse allows the operation to occur earlier and
1649 * when paging_in_progress is true for an object... This is not a complete
1650 * operation, but should plug 99.9% of the rest of the leaks.
1651 */
1652 static void
1653 vm_object_qcollapse(vm_object_t object)
1654 {
1655 vm_object_t backing_object = object->backing_object;
1656
1657 VM_OBJECT_ASSERT_WLOCKED(object);
1658 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1659
1660 if (backing_object->ref_count != 1)
1661 return;
1662
1663 vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT);
1664 }
1665
1666 /*
1667 * vm_object_collapse:
1668 *
1669 * Collapse an object with the object backing it.
1670 * Pages in the backing object are moved into the
1671 * parent, and the backing object is deallocated.
1672 */
1673 void
1674 vm_object_collapse(vm_object_t object)
1675 {
1676 vm_object_t backing_object, new_backing_object;
1677
1678 VM_OBJECT_ASSERT_WLOCKED(object);
1679
1680 while (TRUE) {
1681 /*
1682 * Verify that the conditions are right for collapse:
1683 *
1684 * The object exists and the backing object exists.
1685 */
1686 if ((backing_object = object->backing_object) == NULL)
1687 break;
1688
1689 /*
1690 * we check the backing object first, because it is most likely
1691 * not collapsable.
1692 */
1693 VM_OBJECT_WLOCK(backing_object);
1694 if (backing_object->handle != NULL ||
1695 (backing_object->type != OBJT_DEFAULT &&
1696 backing_object->type != OBJT_SWAP) ||
1697 (backing_object->flags & OBJ_DEAD) ||
1698 object->handle != NULL ||
1699 (object->type != OBJT_DEFAULT &&
1700 object->type != OBJT_SWAP) ||
1701 (object->flags & OBJ_DEAD)) {
1702 VM_OBJECT_WUNLOCK(backing_object);
1703 break;
1704 }
1705
1706 if (object->paging_in_progress != 0 ||
1707 backing_object->paging_in_progress != 0) {
1708 vm_object_qcollapse(object);
1709 VM_OBJECT_WUNLOCK(backing_object);
1710 break;
1711 }
1712
1713 /*
1714 * We know that we can either collapse the backing object (if
1715 * the parent is the only reference to it) or (perhaps) have
1716 * the parent bypass the object if the parent happens to shadow
1717 * all the resident pages in the entire backing object.
1718 *
1719 * This is ignoring pager-backed pages such as swap pages.
1720 * vm_object_collapse_scan fails the shadowing test in this
1721 * case.
1722 */
1723 if (backing_object->ref_count == 1) {
1724 vm_object_pip_add(object, 1);
1725 vm_object_pip_add(backing_object, 1);
1726
1727 /*
1728 * If there is exactly one reference to the backing
1729 * object, we can collapse it into the parent.
1730 */
1731 vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT);
1732
1733 #if VM_NRESERVLEVEL > 0
1734 /*
1735 * Break any reservations from backing_object.
1736 */
1737 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1738 vm_reserv_break_all(backing_object);
1739 #endif
1740
1741 /*
1742 * Move the pager from backing_object to object.
1743 */
1744 if (backing_object->type == OBJT_SWAP) {
1745 /*
1746 * swap_pager_copy() can sleep, in which case
1747 * the backing_object's and object's locks are
1748 * released and reacquired.
1749 * Since swap_pager_copy() is being asked to
1750 * destroy the source, it will change the
1751 * backing_object's type to OBJT_DEFAULT.
1752 */
1753 swap_pager_copy(
1754 backing_object,
1755 object,
1756 OFF_TO_IDX(object->backing_object_offset), TRUE);
1757
1758 /*
1759 * Free any cached pages from backing_object.
1760 */
1761 if (__predict_false(
1762 !vm_object_cache_is_empty(backing_object)))
1763 vm_page_cache_free(backing_object, 0, 0);
1764 }
1765 /*
1766 * Object now shadows whatever backing_object did.
1767 * Note that the reference to
1768 * backing_object->backing_object moves from within
1769 * backing_object to within object.
1770 */
1771 LIST_REMOVE(object, shadow_list);
1772 backing_object->shadow_count--;
1773 if (backing_object->backing_object) {
1774 VM_OBJECT_WLOCK(backing_object->backing_object);
1775 LIST_REMOVE(backing_object, shadow_list);
1776 LIST_INSERT_HEAD(
1777 &backing_object->backing_object->shadow_head,
1778 object, shadow_list);
1779 /*
1780 * The shadow_count has not changed.
1781 */
1782 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1783 }
1784 object->backing_object = backing_object->backing_object;
1785 object->backing_object_offset +=
1786 backing_object->backing_object_offset;
1787
1788 /*
1789 * Discard backing_object.
1790 *
1791 * Since the backing object has no pages, no pager left,
1792 * and no object references within it, all that is
1793 * necessary is to dispose of it.
1794 */
1795 KASSERT(backing_object->ref_count == 1, (
1796 "backing_object %p was somehow re-referenced during collapse!",
1797 backing_object));
1798 vm_object_pip_wakeup(backing_object);
1799 backing_object->type = OBJT_DEAD;
1800 backing_object->ref_count = 0;
1801 VM_OBJECT_WUNLOCK(backing_object);
1802 vm_object_destroy(backing_object);
1803
1804 vm_object_pip_wakeup(object);
1805 object_collapses++;
1806 } else {
1807 /*
1808 * If we do not entirely shadow the backing object,
1809 * there is nothing we can do so we give up.
1810 */
1811 if (object->resident_page_count != object->size &&
1812 !vm_object_scan_all_shadowed(object)) {
1813 VM_OBJECT_WUNLOCK(backing_object);
1814 break;
1815 }
1816
1817 /*
1818 * Make the parent shadow the next object in the
1819 * chain. Deallocating backing_object will not remove
1820 * it, since its reference count is at least 2.
1821 */
1822 LIST_REMOVE(object, shadow_list);
1823 backing_object->shadow_count--;
1824
1825 new_backing_object = backing_object->backing_object;
1826 if ((object->backing_object = new_backing_object) != NULL) {
1827 VM_OBJECT_WLOCK(new_backing_object);
1828 LIST_INSERT_HEAD(
1829 &new_backing_object->shadow_head,
1830 object,
1831 shadow_list
1832 );
1833 new_backing_object->shadow_count++;
1834 vm_object_reference_locked(new_backing_object);
1835 VM_OBJECT_WUNLOCK(new_backing_object);
1836 object->backing_object_offset +=
1837 backing_object->backing_object_offset;
1838 }
1839
1840 /*
1841 * Drop the reference count on backing_object. Since
1842 * its ref_count was at least 2, it will not vanish.
1843 */
1844 backing_object->ref_count--;
1845 VM_OBJECT_WUNLOCK(backing_object);
1846 object_bypasses++;
1847 }
1848
1849 /*
1850 * Try again with this object's new backing object.
1851 */
1852 }
1853 }
1854
1855 /*
1856 * vm_object_page_remove:
1857 *
1858 * For the given object, either frees or invalidates each of the
1859 * specified pages. In general, a page is freed. However, if a page is
1860 * wired for any reason other than the existence of a managed, wired
1861 * mapping, then it may be invalidated but not removed from the object.
1862 * Pages are specified by the given range ["start", "end") and the option
1863 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1864 * extends from "start" to the end of the object. If the option
1865 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1866 * specified range are affected. If the option OBJPR_NOTMAPPED is
1867 * specified, then the pages within the specified range must have no
1868 * mappings. Otherwise, if this option is not specified, any mappings to
1869 * the specified pages are removed before the pages are freed or
1870 * invalidated.
1871 *
1872 * In general, this operation should only be performed on objects that
1873 * contain managed pages. There are, however, two exceptions. First, it
1874 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1875 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1876 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1877 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1878 *
1879 * The object must be locked.
1880 */
1881 void
1882 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1883 int options)
1884 {
1885 vm_page_t p, next;
1886
1887 VM_OBJECT_ASSERT_WLOCKED(object);
1888 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1889 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1890 ("vm_object_page_remove: illegal options for object %p", object));
1891 if (object->resident_page_count == 0)
1892 goto skipmemq;
1893 vm_object_pip_add(object, 1);
1894 again:
1895 p = vm_page_find_least(object, start);
1896
1897 /*
1898 * Here, the variable "p" is either (1) the page with the least pindex
1899 * greater than or equal to the parameter "start" or (2) NULL.
1900 */
1901 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1902 next = TAILQ_NEXT(p, listq);
1903
1904 /*
1905 * If the page is wired for any reason besides the existence
1906 * of managed, wired mappings, then it cannot be freed. For
1907 * example, fictitious pages, which represent device memory,
1908 * are inherently wired and cannot be freed. They can,
1909 * however, be invalidated if the option OBJPR_CLEANONLY is
1910 * not specified.
1911 */
1912 vm_page_lock(p);
1913 if (vm_page_xbusied(p)) {
1914 VM_OBJECT_WUNLOCK(object);
1915 vm_page_busy_sleep(p, "vmopax");
1916 VM_OBJECT_WLOCK(object);
1917 goto again;
1918 }
1919 if (p->wire_count != 0) {
1920 if ((options & OBJPR_NOTMAPPED) == 0)
1921 pmap_remove_all(p);
1922 if ((options & OBJPR_CLEANONLY) == 0) {
1923 p->valid = 0;
1924 vm_page_undirty(p);
1925 }
1926 goto next;
1927 }
1928 if (vm_page_busied(p)) {
1929 VM_OBJECT_WUNLOCK(object);
1930 vm_page_busy_sleep(p, "vmopar");
1931 VM_OBJECT_WLOCK(object);
1932 goto again;
1933 }
1934 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1935 ("vm_object_page_remove: page %p is fictitious", p));
1936 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1937 if ((options & OBJPR_NOTMAPPED) == 0)
1938 pmap_remove_write(p);
1939 if (p->dirty)
1940 goto next;
1941 }
1942 if ((options & OBJPR_NOTMAPPED) == 0)
1943 pmap_remove_all(p);
1944 vm_page_free(p);
1945 next:
1946 vm_page_unlock(p);
1947 }
1948 vm_object_pip_wakeup(object);
1949 skipmemq:
1950 if (__predict_false(!vm_object_cache_is_empty(object)))
1951 vm_page_cache_free(object, start, end);
1952 }
1953
1954 /*
1955 * vm_object_page_noreuse:
1956 *
1957 * For the given object, attempt to move the specified pages to
1958 * the head of the inactive queue. This bypasses regular LRU
1959 * operation and allows the pages to be reused quickly under memory
1960 * pressure. If a page is wired for any reason, then it will not
1961 * be queued. Pages are specified by the range ["start", "end").
1962 * As a special case, if "end" is zero, then the range extends from
1963 * "start" to the end of the object.
1964 *
1965 * This operation should only be performed on objects that
1966 * contain non-fictitious, managed pages.
1967 *
1968 * The object must be locked.
1969 */
1970 void
1971 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1972 {
1973 struct mtx *mtx, *new_mtx;
1974 vm_page_t p, next;
1975
1976 VM_OBJECT_ASSERT_WLOCKED(object);
1977 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1978 ("vm_object_page_noreuse: illegal object %p", object));
1979 if (object->resident_page_count == 0)
1980 return;
1981 p = vm_page_find_least(object, start);
1982
1983 /*
1984 * Here, the variable "p" is either (1) the page with the least pindex
1985 * greater than or equal to the parameter "start" or (2) NULL.
1986 */
1987 mtx = NULL;
1988 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1989 next = TAILQ_NEXT(p, listq);
1990
1991 /*
1992 * Avoid releasing and reacquiring the same page lock.
1993 */
1994 new_mtx = vm_page_lockptr(p);
1995 if (mtx != new_mtx) {
1996 if (mtx != NULL)
1997 mtx_unlock(mtx);
1998 mtx = new_mtx;
1999 mtx_lock(mtx);
2000 }
2001 vm_page_deactivate_noreuse(p);
2002 }
2003 if (mtx != NULL)
2004 mtx_unlock(mtx);
2005 }
2006
2007 /*
2008 * Populate the specified range of the object with valid pages. Returns
2009 * TRUE if the range is successfully populated and FALSE otherwise.
2010 *
2011 * Note: This function should be optimized to pass a larger array of
2012 * pages to vm_pager_get_pages() before it is applied to a non-
2013 * OBJT_DEVICE object.
2014 *
2015 * The object must be locked.
2016 */
2017 boolean_t
2018 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2019 {
2020 vm_page_t m;
2021 vm_pindex_t pindex;
2022 int rv;
2023
2024 VM_OBJECT_ASSERT_WLOCKED(object);
2025 for (pindex = start; pindex < end; pindex++) {
2026 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
2027 if (m->valid != VM_PAGE_BITS_ALL) {
2028 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
2029 if (rv != VM_PAGER_OK) {
2030 vm_page_lock(m);
2031 vm_page_free(m);
2032 vm_page_unlock(m);
2033 break;
2034 }
2035 }
2036 /*
2037 * Keep "m" busy because a subsequent iteration may unlock
2038 * the object.
2039 */
2040 }
2041 if (pindex > start) {
2042 m = vm_page_lookup(object, start);
2043 while (m != NULL && m->pindex < pindex) {
2044 vm_page_xunbusy(m);
2045 m = TAILQ_NEXT(m, listq);
2046 }
2047 }
2048 return (pindex == end);
2049 }
2050
2051 /*
2052 * Routine: vm_object_coalesce
2053 * Function: Coalesces two objects backing up adjoining
2054 * regions of memory into a single object.
2055 *
2056 * returns TRUE if objects were combined.
2057 *
2058 * NOTE: Only works at the moment if the second object is NULL -
2059 * if it's not, which object do we lock first?
2060 *
2061 * Parameters:
2062 * prev_object First object to coalesce
2063 * prev_offset Offset into prev_object
2064 * prev_size Size of reference to prev_object
2065 * next_size Size of reference to the second object
2066 * reserved Indicator that extension region has
2067 * swap accounted for
2068 *
2069 * Conditions:
2070 * The object must *not* be locked.
2071 */
2072 boolean_t
2073 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2074 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2075 {
2076 vm_pindex_t next_pindex;
2077
2078 if (prev_object == NULL)
2079 return (TRUE);
2080 VM_OBJECT_WLOCK(prev_object);
2081 if ((prev_object->type != OBJT_DEFAULT &&
2082 prev_object->type != OBJT_SWAP) ||
2083 (prev_object->flags & OBJ_TMPFS_NODE) != 0) {
2084 VM_OBJECT_WUNLOCK(prev_object);
2085 return (FALSE);
2086 }
2087
2088 /*
2089 * Try to collapse the object first
2090 */
2091 vm_object_collapse(prev_object);
2092
2093 /*
2094 * Can't coalesce if: . more than one reference . paged out . shadows
2095 * another object . has a copy elsewhere (any of which mean that the
2096 * pages not mapped to prev_entry may be in use anyway)
2097 */
2098 if (prev_object->backing_object != NULL) {
2099 VM_OBJECT_WUNLOCK(prev_object);
2100 return (FALSE);
2101 }
2102
2103 prev_size >>= PAGE_SHIFT;
2104 next_size >>= PAGE_SHIFT;
2105 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2106
2107 if ((prev_object->ref_count > 1) &&
2108 (prev_object->size != next_pindex)) {
2109 VM_OBJECT_WUNLOCK(prev_object);
2110 return (FALSE);
2111 }
2112
2113 /*
2114 * Account for the charge.
2115 */
2116 if (prev_object->cred != NULL) {
2117
2118 /*
2119 * If prev_object was charged, then this mapping,
2120 * although not charged now, may become writable
2121 * later. Non-NULL cred in the object would prevent
2122 * swap reservation during enabling of the write
2123 * access, so reserve swap now. Failed reservation
2124 * cause allocation of the separate object for the map
2125 * entry, and swap reservation for this entry is
2126 * managed in appropriate time.
2127 */
2128 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2129 prev_object->cred)) {
2130 VM_OBJECT_WUNLOCK(prev_object);
2131 return (FALSE);
2132 }
2133 prev_object->charge += ptoa(next_size);
2134 }
2135
2136 /*
2137 * Remove any pages that may still be in the object from a previous
2138 * deallocation.
2139 */
2140 if (next_pindex < prev_object->size) {
2141 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2142 next_size, 0);
2143 if (prev_object->type == OBJT_SWAP)
2144 swap_pager_freespace(prev_object,
2145 next_pindex, next_size);
2146 #if 0
2147 if (prev_object->cred != NULL) {
2148 KASSERT(prev_object->charge >=
2149 ptoa(prev_object->size - next_pindex),
2150 ("object %p overcharged 1 %jx %jx", prev_object,
2151 (uintmax_t)next_pindex, (uintmax_t)next_size));
2152 prev_object->charge -= ptoa(prev_object->size -
2153 next_pindex);
2154 }
2155 #endif
2156 }
2157
2158 /*
2159 * Extend the object if necessary.
2160 */
2161 if (next_pindex + next_size > prev_object->size)
2162 prev_object->size = next_pindex + next_size;
2163
2164 VM_OBJECT_WUNLOCK(prev_object);
2165 return (TRUE);
2166 }
2167
2168 void
2169 vm_object_set_writeable_dirty(vm_object_t object)
2170 {
2171
2172 VM_OBJECT_ASSERT_WLOCKED(object);
2173 if (object->type != OBJT_VNODE) {
2174 if ((object->flags & OBJ_TMPFS_NODE) != 0) {
2175 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs"));
2176 vm_object_set_flag(object, OBJ_TMPFS_DIRTY);
2177 }
2178 return;
2179 }
2180 object->generation++;
2181 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2182 return;
2183 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2184 }
2185
2186 /*
2187 * vm_object_unwire:
2188 *
2189 * For each page offset within the specified range of the given object,
2190 * find the highest-level page in the shadow chain and unwire it. A page
2191 * must exist at every page offset, and the highest-level page must be
2192 * wired.
2193 */
2194 void
2195 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2196 uint8_t queue)
2197 {
2198 vm_object_t tobject;
2199 vm_page_t m, tm;
2200 vm_pindex_t end_pindex, pindex, tpindex;
2201 int depth, locked_depth;
2202
2203 KASSERT((offset & PAGE_MASK) == 0,
2204 ("vm_object_unwire: offset is not page aligned"));
2205 KASSERT((length & PAGE_MASK) == 0,
2206 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2207 /* The wired count of a fictitious page never changes. */
2208 if ((object->flags & OBJ_FICTITIOUS) != 0)
2209 return;
2210 pindex = OFF_TO_IDX(offset);
2211 end_pindex = pindex + atop(length);
2212 locked_depth = 1;
2213 VM_OBJECT_RLOCK(object);
2214 m = vm_page_find_least(object, pindex);
2215 while (pindex < end_pindex) {
2216 if (m == NULL || pindex < m->pindex) {
2217 /*
2218 * The first object in the shadow chain doesn't
2219 * contain a page at the current index. Therefore,
2220 * the page must exist in a backing object.
2221 */
2222 tobject = object;
2223 tpindex = pindex;
2224 depth = 0;
2225 do {
2226 tpindex +=
2227 OFF_TO_IDX(tobject->backing_object_offset);
2228 tobject = tobject->backing_object;
2229 KASSERT(tobject != NULL,
2230 ("vm_object_unwire: missing page"));
2231 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2232 goto next_page;
2233 depth++;
2234 if (depth == locked_depth) {
2235 locked_depth++;
2236 VM_OBJECT_RLOCK(tobject);
2237 }
2238 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2239 NULL);
2240 } else {
2241 tm = m;
2242 m = TAILQ_NEXT(m, listq);
2243 }
2244 vm_page_lock(tm);
2245 vm_page_unwire(tm, queue);
2246 vm_page_unlock(tm);
2247 next_page:
2248 pindex++;
2249 }
2250 /* Release the accumulated object locks. */
2251 for (depth = 0; depth < locked_depth; depth++) {
2252 tobject = object->backing_object;
2253 VM_OBJECT_RUNLOCK(object);
2254 object = tobject;
2255 }
2256 }
2257
2258 struct vnode *
2259 vm_object_vnode(vm_object_t object)
2260 {
2261
2262 VM_OBJECT_ASSERT_LOCKED(object);
2263 if (object->type == OBJT_VNODE)
2264 return (object->handle);
2265 if (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS) != 0)
2266 return (object->un_pager.swp.swp_tmpfs);
2267 return (NULL);
2268 }
2269
2270 static int
2271 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2272 {
2273 struct kinfo_vmobject kvo;
2274 char *fullpath, *freepath;
2275 struct vnode *vp;
2276 struct vattr va;
2277 vm_object_t obj;
2278 vm_page_t m;
2279 int count, error;
2280
2281 if (req->oldptr == NULL) {
2282 /*
2283 * If an old buffer has not been provided, generate an
2284 * estimate of the space needed for a subsequent call.
2285 */
2286 mtx_lock(&vm_object_list_mtx);
2287 count = 0;
2288 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2289 if (obj->type == OBJT_DEAD)
2290 continue;
2291 count++;
2292 }
2293 mtx_unlock(&vm_object_list_mtx);
2294 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2295 count * 11 / 10));
2296 }
2297
2298 error = 0;
2299
2300 /*
2301 * VM objects are type stable and are never removed from the
2302 * list once added. This allows us to safely read obj->object_list
2303 * after reacquiring the VM object lock.
2304 */
2305 mtx_lock(&vm_object_list_mtx);
2306 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2307 if (obj->type == OBJT_DEAD)
2308 continue;
2309 VM_OBJECT_RLOCK(obj);
2310 if (obj->type == OBJT_DEAD) {
2311 VM_OBJECT_RUNLOCK(obj);
2312 continue;
2313 }
2314 mtx_unlock(&vm_object_list_mtx);
2315 kvo.kvo_size = ptoa(obj->size);
2316 kvo.kvo_resident = obj->resident_page_count;
2317 kvo.kvo_ref_count = obj->ref_count;
2318 kvo.kvo_shadow_count = obj->shadow_count;
2319 kvo.kvo_memattr = obj->memattr;
2320 kvo.kvo_active = 0;
2321 kvo.kvo_inactive = 0;
2322 TAILQ_FOREACH(m, &obj->memq, listq) {
2323 /*
2324 * A page may belong to the object but be
2325 * dequeued and set to PQ_NONE while the
2326 * object lock is not held. This makes the
2327 * reads of m->queue below racy, and we do not
2328 * count pages set to PQ_NONE. However, this
2329 * sysctl is only meant to give an
2330 * approximation of the system anyway.
2331 */
2332 if (m->queue == PQ_ACTIVE)
2333 kvo.kvo_active++;
2334 else if (m->queue == PQ_INACTIVE)
2335 kvo.kvo_inactive++;
2336 }
2337
2338 kvo.kvo_vn_fileid = 0;
2339 kvo.kvo_vn_fsid = 0;
2340 freepath = NULL;
2341 fullpath = "";
2342 vp = NULL;
2343 switch (obj->type) {
2344 case OBJT_DEFAULT:
2345 kvo.kvo_type = KVME_TYPE_DEFAULT;
2346 break;
2347 case OBJT_VNODE:
2348 kvo.kvo_type = KVME_TYPE_VNODE;
2349 vp = obj->handle;
2350 vref(vp);
2351 break;
2352 case OBJT_SWAP:
2353 kvo.kvo_type = KVME_TYPE_SWAP;
2354 break;
2355 case OBJT_DEVICE:
2356 kvo.kvo_type = KVME_TYPE_DEVICE;
2357 break;
2358 case OBJT_PHYS:
2359 kvo.kvo_type = KVME_TYPE_PHYS;
2360 break;
2361 case OBJT_DEAD:
2362 kvo.kvo_type = KVME_TYPE_DEAD;
2363 break;
2364 case OBJT_SG:
2365 kvo.kvo_type = KVME_TYPE_SG;
2366 break;
2367 case OBJT_MGTDEVICE:
2368 kvo.kvo_type = KVME_TYPE_MGTDEVICE;
2369 break;
2370 default:
2371 kvo.kvo_type = KVME_TYPE_UNKNOWN;
2372 break;
2373 }
2374 VM_OBJECT_RUNLOCK(obj);
2375 if (vp != NULL) {
2376 vn_fullpath(curthread, vp, &fullpath, &freepath);
2377 vn_lock(vp, LK_SHARED | LK_RETRY);
2378 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2379 kvo.kvo_vn_fileid = va.va_fileid;
2380 kvo.kvo_vn_fsid = va.va_fsid;
2381 }
2382 vput(vp);
2383 }
2384
2385 strlcpy(kvo.kvo_path, fullpath, sizeof(kvo.kvo_path));
2386 if (freepath != NULL)
2387 free(freepath, M_TEMP);
2388
2389 /* Pack record size down */
2390 kvo.kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path) +
2391 strlen(kvo.kvo_path) + 1;
2392 kvo.kvo_structsize = roundup(kvo.kvo_structsize,
2393 sizeof(uint64_t));
2394 error = SYSCTL_OUT(req, &kvo, kvo.kvo_structsize);
2395 mtx_lock(&vm_object_list_mtx);
2396 if (error)
2397 break;
2398 }
2399 mtx_unlock(&vm_object_list_mtx);
2400 return (error);
2401 }
2402 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2403 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2404 "List of VM objects");
2405
2406 #include "opt_ddb.h"
2407 #ifdef DDB
2408 #include <sys/kernel.h>
2409
2410 #include <sys/cons.h>
2411
2412 #include <ddb/ddb.h>
2413
2414 static int
2415 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2416 {
2417 vm_map_t tmpm;
2418 vm_map_entry_t tmpe;
2419 vm_object_t obj;
2420 int entcount;
2421
2422 if (map == 0)
2423 return 0;
2424
2425 if (entry == 0) {
2426 tmpe = map->header.next;
2427 entcount = map->nentries;
2428 while (entcount-- && (tmpe != &map->header)) {
2429 if (_vm_object_in_map(map, object, tmpe)) {
2430 return 1;
2431 }
2432 tmpe = tmpe->next;
2433 }
2434 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2435 tmpm = entry->object.sub_map;
2436 tmpe = tmpm->header.next;
2437 entcount = tmpm->nentries;
2438 while (entcount-- && tmpe != &tmpm->header) {
2439 if (_vm_object_in_map(tmpm, object, tmpe)) {
2440 return 1;
2441 }
2442 tmpe = tmpe->next;
2443 }
2444 } else if ((obj = entry->object.vm_object) != NULL) {
2445 for (; obj; obj = obj->backing_object)
2446 if (obj == object) {
2447 return 1;
2448 }
2449 }
2450 return 0;
2451 }
2452
2453 static int
2454 vm_object_in_map(vm_object_t object)
2455 {
2456 struct proc *p;
2457
2458 /* sx_slock(&allproc_lock); */
2459 FOREACH_PROC_IN_SYSTEM(p) {
2460 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2461 continue;
2462 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2463 /* sx_sunlock(&allproc_lock); */
2464 return 1;
2465 }
2466 }
2467 /* sx_sunlock(&allproc_lock); */
2468 if (_vm_object_in_map(kernel_map, object, 0))
2469 return 1;
2470 return 0;
2471 }
2472
2473 DB_SHOW_COMMAND(vmochk, vm_object_check)
2474 {
2475 vm_object_t object;
2476
2477 /*
2478 * make sure that internal objs are in a map somewhere
2479 * and none have zero ref counts.
2480 */
2481 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2482 if (object->handle == NULL &&
2483 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2484 if (object->ref_count == 0) {
2485 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2486 (long)object->size);
2487 }
2488 if (!vm_object_in_map(object)) {
2489 db_printf(
2490 "vmochk: internal obj is not in a map: "
2491 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2492 object->ref_count, (u_long)object->size,
2493 (u_long)object->size,
2494 (void *)object->backing_object);
2495 }
2496 }
2497 }
2498 }
2499
2500 /*
2501 * vm_object_print: [ debug ]
2502 */
2503 DB_SHOW_COMMAND(object, vm_object_print_static)
2504 {
2505 /* XXX convert args. */
2506 vm_object_t object = (vm_object_t)addr;
2507 boolean_t full = have_addr;
2508
2509 vm_page_t p;
2510
2511 /* XXX count is an (unused) arg. Avoid shadowing it. */
2512 #define count was_count
2513
2514 int count;
2515
2516 if (object == NULL)
2517 return;
2518
2519 db_iprintf(
2520 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2521 object, (int)object->type, (uintmax_t)object->size,
2522 object->resident_page_count, object->ref_count, object->flags,
2523 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2524 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2525 object->shadow_count,
2526 object->backing_object ? object->backing_object->ref_count : 0,
2527 object->backing_object, (uintmax_t)object->backing_object_offset);
2528
2529 if (!full)
2530 return;
2531
2532 db_indent += 2;
2533 count = 0;
2534 TAILQ_FOREACH(p, &object->memq, listq) {
2535 if (count == 0)
2536 db_iprintf("memory:=");
2537 else if (count == 6) {
2538 db_printf("\n");
2539 db_iprintf(" ...");
2540 count = 0;
2541 } else
2542 db_printf(",");
2543 count++;
2544
2545 db_printf("(off=0x%jx,page=0x%jx)",
2546 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2547 }
2548 if (count != 0)
2549 db_printf("\n");
2550 db_indent -= 2;
2551 }
2552
2553 /* XXX. */
2554 #undef count
2555
2556 /* XXX need this non-static entry for calling from vm_map_print. */
2557 void
2558 vm_object_print(
2559 /* db_expr_t */ long addr,
2560 boolean_t have_addr,
2561 /* db_expr_t */ long count,
2562 char *modif)
2563 {
2564 vm_object_print_static(addr, have_addr, count, modif);
2565 }
2566
2567 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2568 {
2569 vm_object_t object;
2570 vm_pindex_t fidx;
2571 vm_paddr_t pa;
2572 vm_page_t m, prev_m;
2573 int rcount, nl, c;
2574
2575 nl = 0;
2576 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2577 db_printf("new object: %p\n", (void *)object);
2578 if (nl > 18) {
2579 c = cngetc();
2580 if (c != ' ')
2581 return;
2582 nl = 0;
2583 }
2584 nl++;
2585 rcount = 0;
2586 fidx = 0;
2587 pa = -1;
2588 TAILQ_FOREACH(m, &object->memq, listq) {
2589 if (m->pindex > 128)
2590 break;
2591 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2592 prev_m->pindex + 1 != m->pindex) {
2593 if (rcount) {
2594 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2595 (long)fidx, rcount, (long)pa);
2596 if (nl > 18) {
2597 c = cngetc();
2598 if (c != ' ')
2599 return;
2600 nl = 0;
2601 }
2602 nl++;
2603 rcount = 0;
2604 }
2605 }
2606 if (rcount &&
2607 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2608 ++rcount;
2609 continue;
2610 }
2611 if (rcount) {
2612 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2613 (long)fidx, rcount, (long)pa);
2614 if (nl > 18) {
2615 c = cngetc();
2616 if (c != ' ')
2617 return;
2618 nl = 0;
2619 }
2620 nl++;
2621 }
2622 fidx = m->pindex;
2623 pa = VM_PAGE_TO_PHYS(m);
2624 rcount = 1;
2625 }
2626 if (rcount) {
2627 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2628 (long)fidx, rcount, (long)pa);
2629 if (nl > 18) {
2630 c = cngetc();
2631 if (c != ' ')
2632 return;
2633 nl = 0;
2634 }
2635 nl++;
2636 }
2637 }
2638 }
2639 #endif /* DDB */
Cache object: eaafbf94233f04cb9052e3d1e49a1005
|