FreeBSD/Linux Kernel Cross Reference
sys/vm/vm_object.c
1 /*-
2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3 *
4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
6 *
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
21 *
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 *
34 * from: @(#)vm_object.c 8.5 (Berkeley) 3/22/94
35 *
36 *
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
39 *
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41 *
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
47 *
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51 *
52 * Carnegie Mellon requests users of this software to return to
53 *
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
58 *
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
61 */
62
63 /*
64 * Virtual memory object module.
65 */
66
67 #include <sys/cdefs.h>
68 __FBSDID("$FreeBSD$");
69
70 #include "opt_vm.h"
71
72 #include <sys/param.h>
73 #include <sys/systm.h>
74 #include <sys/cpuset.h>
75 #include <sys/lock.h>
76 #include <sys/mman.h>
77 #include <sys/mount.h>
78 #include <sys/kernel.h>
79 #include <sys/pctrie.h>
80 #include <sys/sysctl.h>
81 #include <sys/mutex.h>
82 #include <sys/proc.h> /* for curproc, pageproc */
83 #include <sys/socket.h>
84 #include <sys/resourcevar.h>
85 #include <sys/rwlock.h>
86 #include <sys/user.h>
87 #include <sys/vnode.h>
88 #include <sys/vmmeter.h>
89 #include <sys/sx.h>
90
91 #include <vm/vm.h>
92 #include <vm/vm_param.h>
93 #include <vm/pmap.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_object.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_pageout.h>
98 #include <vm/vm_pager.h>
99 #include <vm/vm_phys.h>
100 #include <vm/vm_pagequeue.h>
101 #include <vm/swap_pager.h>
102 #include <vm/vm_kern.h>
103 #include <vm/vm_extern.h>
104 #include <vm/vm_radix.h>
105 #include <vm/vm_reserv.h>
106 #include <vm/uma.h>
107
108 static int old_msync;
109 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
110 "Use old (insecure) msync behavior");
111
112 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
113 int pagerflags, int flags, boolean_t *clearobjflags,
114 boolean_t *eio);
115 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
116 boolean_t *clearobjflags);
117 static void vm_object_qcollapse(vm_object_t object);
118 static void vm_object_vndeallocate(vm_object_t object);
119
120 /*
121 * Virtual memory objects maintain the actual data
122 * associated with allocated virtual memory. A given
123 * page of memory exists within exactly one object.
124 *
125 * An object is only deallocated when all "references"
126 * are given up. Only one "reference" to a given
127 * region of an object should be writeable.
128 *
129 * Associated with each object is a list of all resident
130 * memory pages belonging to that object; this list is
131 * maintained by the "vm_page" module, and locked by the object's
132 * lock.
133 *
134 * Each object also records a "pager" routine which is
135 * used to retrieve (and store) pages to the proper backing
136 * storage. In addition, objects may be backed by other
137 * objects from which they were virtual-copied.
138 *
139 * The only items within the object structure which are
140 * modified after time of creation are:
141 * reference count locked by object's lock
142 * pager routine locked by object's lock
143 *
144 */
145
146 struct object_q vm_object_list;
147 struct mtx vm_object_list_mtx; /* lock for object list and count */
148
149 struct vm_object kernel_object_store;
150
151 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
152 "VM object stats");
153
154 static counter_u64_t object_collapses = EARLY_COUNTER;
155 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
156 &object_collapses,
157 "VM object collapses");
158
159 static counter_u64_t object_bypasses = EARLY_COUNTER;
160 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
161 &object_bypasses,
162 "VM object bypasses");
163
164 static void
165 counter_startup(void)
166 {
167
168 object_collapses = counter_u64_alloc(M_WAITOK);
169 object_bypasses = counter_u64_alloc(M_WAITOK);
170 }
171 SYSINIT(object_counters, SI_SUB_CPU, SI_ORDER_ANY, counter_startup, NULL);
172
173 static uma_zone_t obj_zone;
174
175 static int vm_object_zinit(void *mem, int size, int flags);
176
177 #ifdef INVARIANTS
178 static void vm_object_zdtor(void *mem, int size, void *arg);
179
180 static void
181 vm_object_zdtor(void *mem, int size, void *arg)
182 {
183 vm_object_t object;
184
185 object = (vm_object_t)mem;
186 KASSERT(object->ref_count == 0,
187 ("object %p ref_count = %d", object, object->ref_count));
188 KASSERT(TAILQ_EMPTY(&object->memq),
189 ("object %p has resident pages in its memq", object));
190 KASSERT(vm_radix_is_empty(&object->rtree),
191 ("object %p has resident pages in its trie", object));
192 #if VM_NRESERVLEVEL > 0
193 KASSERT(LIST_EMPTY(&object->rvq),
194 ("object %p has reservations",
195 object));
196 #endif
197 KASSERT(object->paging_in_progress == 0,
198 ("object %p paging_in_progress = %d",
199 object, object->paging_in_progress));
200 KASSERT(object->resident_page_count == 0,
201 ("object %p resident_page_count = %d",
202 object, object->resident_page_count));
203 KASSERT(object->shadow_count == 0,
204 ("object %p shadow_count = %d",
205 object, object->shadow_count));
206 KASSERT(object->type == OBJT_DEAD,
207 ("object %p has non-dead type %d",
208 object, object->type));
209 }
210 #endif
211
212 static int
213 vm_object_zinit(void *mem, int size, int flags)
214 {
215 vm_object_t object;
216
217 object = (vm_object_t)mem;
218 rw_init_flags(&object->lock, "vm object", RW_DUPOK | RW_NEW);
219
220 /* These are true for any object that has been freed */
221 object->type = OBJT_DEAD;
222 object->ref_count = 0;
223 vm_radix_init(&object->rtree);
224 object->paging_in_progress = 0;
225 object->resident_page_count = 0;
226 object->shadow_count = 0;
227 object->flags = OBJ_DEAD;
228
229 mtx_lock(&vm_object_list_mtx);
230 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
231 mtx_unlock(&vm_object_list_mtx);
232 return (0);
233 }
234
235 static void
236 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
237 {
238
239 TAILQ_INIT(&object->memq);
240 LIST_INIT(&object->shadow_head);
241
242 object->type = type;
243 if (type == OBJT_SWAP)
244 pctrie_init(&object->un_pager.swp.swp_blks);
245
246 /*
247 * Ensure that swap_pager_swapoff() iteration over object_list
248 * sees up to date type and pctrie head if it observed
249 * non-dead object.
250 */
251 atomic_thread_fence_rel();
252
253 switch (type) {
254 case OBJT_DEAD:
255 panic("_vm_object_allocate: can't create OBJT_DEAD");
256 case OBJT_DEFAULT:
257 case OBJT_SWAP:
258 object->flags = OBJ_ONEMAPPING;
259 break;
260 case OBJT_DEVICE:
261 case OBJT_SG:
262 object->flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
263 break;
264 case OBJT_MGTDEVICE:
265 object->flags = OBJ_FICTITIOUS;
266 break;
267 case OBJT_PHYS:
268 object->flags = OBJ_UNMANAGED;
269 break;
270 case OBJT_VNODE:
271 object->flags = 0;
272 break;
273 default:
274 panic("_vm_object_allocate: type %d is undefined", type);
275 }
276 object->size = size;
277 object->domain.dr_policy = NULL;
278 object->generation = 1;
279 object->ref_count = 1;
280 object->memattr = VM_MEMATTR_DEFAULT;
281 object->cred = NULL;
282 object->charge = 0;
283 object->handle = NULL;
284 object->backing_object = NULL;
285 object->backing_object_offset = (vm_ooffset_t) 0;
286 #if VM_NRESERVLEVEL > 0
287 LIST_INIT(&object->rvq);
288 #endif
289 umtx_shm_object_init(object);
290 }
291
292 /*
293 * vm_object_init:
294 *
295 * Initialize the VM objects module.
296 */
297 void
298 vm_object_init(void)
299 {
300 TAILQ_INIT(&vm_object_list);
301 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
302
303 rw_init(&kernel_object->lock, "kernel vm object");
304 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
305 VM_MIN_KERNEL_ADDRESS), kernel_object);
306 #if VM_NRESERVLEVEL > 0
307 kernel_object->flags |= OBJ_COLORED;
308 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
309 #endif
310 kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
311
312 /*
313 * The lock portion of struct vm_object must be type stable due
314 * to vm_pageout_fallback_object_lock locking a vm object
315 * without holding any references to it.
316 */
317 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
318 #ifdef INVARIANTS
319 vm_object_zdtor,
320 #else
321 NULL,
322 #endif
323 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
324
325 vm_radix_zinit();
326 }
327
328 void
329 vm_object_clear_flag(vm_object_t object, u_short bits)
330 {
331
332 VM_OBJECT_ASSERT_WLOCKED(object);
333 object->flags &= ~bits;
334 }
335
336 /*
337 * Sets the default memory attribute for the specified object. Pages
338 * that are allocated to this object are by default assigned this memory
339 * attribute.
340 *
341 * Presently, this function must be called before any pages are allocated
342 * to the object. In the future, this requirement may be relaxed for
343 * "default" and "swap" objects.
344 */
345 int
346 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
347 {
348
349 VM_OBJECT_ASSERT_WLOCKED(object);
350 switch (object->type) {
351 case OBJT_DEFAULT:
352 case OBJT_DEVICE:
353 case OBJT_MGTDEVICE:
354 case OBJT_PHYS:
355 case OBJT_SG:
356 case OBJT_SWAP:
357 case OBJT_VNODE:
358 if (!TAILQ_EMPTY(&object->memq))
359 return (KERN_FAILURE);
360 break;
361 case OBJT_DEAD:
362 return (KERN_INVALID_ARGUMENT);
363 default:
364 panic("vm_object_set_memattr: object %p is of undefined type",
365 object);
366 }
367 object->memattr = memattr;
368 return (KERN_SUCCESS);
369 }
370
371 void
372 vm_object_pip_add(vm_object_t object, short i)
373 {
374
375 VM_OBJECT_ASSERT_WLOCKED(object);
376 object->paging_in_progress += i;
377 }
378
379 void
380 vm_object_pip_subtract(vm_object_t object, short i)
381 {
382
383 VM_OBJECT_ASSERT_WLOCKED(object);
384 object->paging_in_progress -= i;
385 }
386
387 void
388 vm_object_pip_wakeup(vm_object_t object)
389 {
390
391 VM_OBJECT_ASSERT_WLOCKED(object);
392 object->paging_in_progress--;
393 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
394 vm_object_clear_flag(object, OBJ_PIPWNT);
395 wakeup(object);
396 }
397 }
398
399 void
400 vm_object_pip_wakeupn(vm_object_t object, short i)
401 {
402
403 VM_OBJECT_ASSERT_WLOCKED(object);
404 if (i)
405 object->paging_in_progress -= i;
406 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
407 vm_object_clear_flag(object, OBJ_PIPWNT);
408 wakeup(object);
409 }
410 }
411
412 void
413 vm_object_pip_wait(vm_object_t object, char *waitid)
414 {
415
416 VM_OBJECT_ASSERT_WLOCKED(object);
417 while (object->paging_in_progress) {
418 object->flags |= OBJ_PIPWNT;
419 VM_OBJECT_SLEEP(object, object, PVM, waitid, 0);
420 }
421 }
422
423 /*
424 * vm_object_allocate:
425 *
426 * Returns a new object with the given size.
427 */
428 vm_object_t
429 vm_object_allocate(objtype_t type, vm_pindex_t size)
430 {
431 vm_object_t object;
432
433 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
434 _vm_object_allocate(type, size, object);
435 return (object);
436 }
437
438
439 /*
440 * vm_object_reference:
441 *
442 * Gets another reference to the given object. Note: OBJ_DEAD
443 * objects can be referenced during final cleaning.
444 */
445 void
446 vm_object_reference(vm_object_t object)
447 {
448 if (object == NULL)
449 return;
450 VM_OBJECT_WLOCK(object);
451 vm_object_reference_locked(object);
452 VM_OBJECT_WUNLOCK(object);
453 }
454
455 /*
456 * vm_object_reference_locked:
457 *
458 * Gets another reference to the given object.
459 *
460 * The object must be locked.
461 */
462 void
463 vm_object_reference_locked(vm_object_t object)
464 {
465 struct vnode *vp;
466
467 VM_OBJECT_ASSERT_WLOCKED(object);
468 object->ref_count++;
469 if (object->type == OBJT_VNODE) {
470 vp = object->handle;
471 vref(vp);
472 }
473 }
474
475 /*
476 * Handle deallocating an object of type OBJT_VNODE.
477 */
478 static void
479 vm_object_vndeallocate(vm_object_t object)
480 {
481 struct vnode *vp = (struct vnode *) object->handle;
482
483 VM_OBJECT_ASSERT_WLOCKED(object);
484 KASSERT(object->type == OBJT_VNODE,
485 ("vm_object_vndeallocate: not a vnode object"));
486 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
487 #ifdef INVARIANTS
488 if (object->ref_count == 0) {
489 vn_printf(vp, "vm_object_vndeallocate ");
490 panic("vm_object_vndeallocate: bad object reference count");
491 }
492 #endif
493
494 if (!umtx_shm_vnobj_persistent && object->ref_count == 1)
495 umtx_shm_object_terminated(object);
496
497 object->ref_count--;
498
499 /* vrele may need the vnode lock. */
500 VM_OBJECT_WUNLOCK(object);
501 vrele(vp);
502 }
503
504 /*
505 * vm_object_deallocate:
506 *
507 * Release a reference to the specified object,
508 * gained either through a vm_object_allocate
509 * or a vm_object_reference call. When all references
510 * are gone, storage associated with this object
511 * may be relinquished.
512 *
513 * No object may be locked.
514 */
515 void
516 vm_object_deallocate(vm_object_t object)
517 {
518 vm_object_t temp;
519
520 while (object != NULL) {
521 VM_OBJECT_WLOCK(object);
522 if (object->type == OBJT_VNODE) {
523 vm_object_vndeallocate(object);
524 return;
525 }
526
527 KASSERT(object->ref_count != 0,
528 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
529
530 /*
531 * If the reference count goes to 0 we start calling
532 * vm_object_terminate() on the object chain.
533 * A ref count of 1 may be a special case depending on the
534 * shadow count being 0 or 1.
535 */
536 object->ref_count--;
537 if (object->ref_count > 1) {
538 VM_OBJECT_WUNLOCK(object);
539 return;
540 } else if (object->ref_count == 1) {
541 if (object->shadow_count == 0 &&
542 object->handle == NULL &&
543 (object->type == OBJT_DEFAULT ||
544 (object->type == OBJT_SWAP &&
545 (object->flags & OBJ_TMPFS_NODE) == 0))) {
546 vm_object_set_flag(object, OBJ_ONEMAPPING);
547 } else if ((object->shadow_count == 1) &&
548 (object->handle == NULL) &&
549 (object->type == OBJT_DEFAULT ||
550 object->type == OBJT_SWAP)) {
551 vm_object_t robject;
552
553 robject = LIST_FIRST(&object->shadow_head);
554 KASSERT(robject != NULL,
555 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
556 object->ref_count,
557 object->shadow_count));
558 KASSERT((robject->flags & OBJ_TMPFS_NODE) == 0,
559 ("shadowed tmpfs v_object %p", object));
560 if (!VM_OBJECT_TRYWLOCK(robject)) {
561 /*
562 * Avoid a potential deadlock.
563 */
564 object->ref_count++;
565 VM_OBJECT_WUNLOCK(object);
566 /*
567 * More likely than not the thread
568 * holding robject's lock has lower
569 * priority than the current thread.
570 * Let the lower priority thread run.
571 */
572 pause("vmo_de", 1);
573 continue;
574 }
575 /*
576 * Collapse object into its shadow unless its
577 * shadow is dead. In that case, object will
578 * be deallocated by the thread that is
579 * deallocating its shadow.
580 */
581 if ((robject->flags & OBJ_DEAD) == 0 &&
582 (robject->handle == NULL) &&
583 (robject->type == OBJT_DEFAULT ||
584 robject->type == OBJT_SWAP)) {
585
586 robject->ref_count++;
587 retry:
588 if (robject->paging_in_progress) {
589 VM_OBJECT_WUNLOCK(object);
590 vm_object_pip_wait(robject,
591 "objde1");
592 temp = robject->backing_object;
593 if (object == temp) {
594 VM_OBJECT_WLOCK(object);
595 goto retry;
596 }
597 } else if (object->paging_in_progress) {
598 VM_OBJECT_WUNLOCK(robject);
599 object->flags |= OBJ_PIPWNT;
600 VM_OBJECT_SLEEP(object, object,
601 PDROP | PVM, "objde2", 0);
602 VM_OBJECT_WLOCK(robject);
603 temp = robject->backing_object;
604 if (object == temp) {
605 VM_OBJECT_WLOCK(object);
606 goto retry;
607 }
608 } else
609 VM_OBJECT_WUNLOCK(object);
610
611 if (robject->ref_count == 1) {
612 robject->ref_count--;
613 object = robject;
614 goto doterm;
615 }
616 object = robject;
617 vm_object_collapse(object);
618 VM_OBJECT_WUNLOCK(object);
619 continue;
620 }
621 VM_OBJECT_WUNLOCK(robject);
622 }
623 VM_OBJECT_WUNLOCK(object);
624 return;
625 }
626 doterm:
627 umtx_shm_object_terminated(object);
628 temp = object->backing_object;
629 if (temp != NULL) {
630 KASSERT((object->flags & OBJ_TMPFS_NODE) == 0,
631 ("shadowed tmpfs v_object 2 %p", object));
632 VM_OBJECT_WLOCK(temp);
633 LIST_REMOVE(object, shadow_list);
634 temp->shadow_count--;
635 VM_OBJECT_WUNLOCK(temp);
636 object->backing_object = NULL;
637 }
638 /*
639 * Don't double-terminate, we could be in a termination
640 * recursion due to the terminate having to sync data
641 * to disk.
642 */
643 if ((object->flags & OBJ_DEAD) == 0)
644 vm_object_terminate(object);
645 else
646 VM_OBJECT_WUNLOCK(object);
647 object = temp;
648 }
649 }
650
651 /*
652 * vm_object_destroy removes the object from the global object list
653 * and frees the space for the object.
654 */
655 void
656 vm_object_destroy(vm_object_t object)
657 {
658
659 /*
660 * Release the allocation charge.
661 */
662 if (object->cred != NULL) {
663 swap_release_by_cred(object->charge, object->cred);
664 object->charge = 0;
665 crfree(object->cred);
666 object->cred = NULL;
667 }
668
669 /*
670 * Free the space for the object.
671 */
672 uma_zfree(obj_zone, object);
673 }
674
675 /*
676 * vm_object_terminate_pages removes any remaining pageable pages
677 * from the object and resets the object to an empty state.
678 */
679 static void
680 vm_object_terminate_pages(vm_object_t object)
681 {
682 vm_page_t p, p_next;
683 struct mtx *mtx;
684
685 VM_OBJECT_ASSERT_WLOCKED(object);
686
687 mtx = NULL;
688
689 /*
690 * Free any remaining pageable pages. This also removes them from the
691 * paging queues. However, don't free wired pages, just remove them
692 * from the object. Rather than incrementally removing each page from
693 * the object, the page and object are reset to any empty state.
694 */
695 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
696 vm_page_assert_unbusied(p);
697 if ((object->flags & OBJ_UNMANAGED) == 0)
698 /*
699 * vm_page_free_prep() only needs the page
700 * lock for managed pages.
701 */
702 vm_page_change_lock(p, &mtx);
703 p->object = NULL;
704 if (vm_page_wired(p))
705 continue;
706 VM_CNT_INC(v_pfree);
707 vm_page_free(p);
708 }
709 if (mtx != NULL)
710 mtx_unlock(mtx);
711
712 /*
713 * If the object contained any pages, then reset it to an empty state.
714 * None of the object's fields, including "resident_page_count", were
715 * modified by the preceding loop.
716 */
717 if (object->resident_page_count != 0) {
718 vm_radix_reclaim_allnodes(&object->rtree);
719 TAILQ_INIT(&object->memq);
720 object->resident_page_count = 0;
721 if (object->type == OBJT_VNODE)
722 vdrop(object->handle);
723 }
724 }
725
726 /*
727 * vm_object_terminate actually destroys the specified object, freeing
728 * up all previously used resources.
729 *
730 * The object must be locked.
731 * This routine may block.
732 */
733 void
734 vm_object_terminate(vm_object_t object)
735 {
736
737 VM_OBJECT_ASSERT_WLOCKED(object);
738
739 /*
740 * Make sure no one uses us.
741 */
742 vm_object_set_flag(object, OBJ_DEAD);
743
744 /*
745 * wait for the pageout daemon to be done with the object
746 */
747 vm_object_pip_wait(object, "objtrm");
748
749 KASSERT(!object->paging_in_progress,
750 ("vm_object_terminate: pageout in progress"));
751
752 /*
753 * Clean and free the pages, as appropriate. All references to the
754 * object are gone, so we don't need to lock it.
755 */
756 if (object->type == OBJT_VNODE) {
757 struct vnode *vp = (struct vnode *)object->handle;
758
759 /*
760 * Clean pages and flush buffers.
761 */
762 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
763 VM_OBJECT_WUNLOCK(object);
764
765 vinvalbuf(vp, V_SAVE, 0, 0);
766
767 BO_LOCK(&vp->v_bufobj);
768 vp->v_bufobj.bo_flag |= BO_DEAD;
769 BO_UNLOCK(&vp->v_bufobj);
770
771 VM_OBJECT_WLOCK(object);
772 }
773
774 KASSERT(object->ref_count == 0,
775 ("vm_object_terminate: object with references, ref_count=%d",
776 object->ref_count));
777
778 if ((object->flags & OBJ_PG_DTOR) == 0)
779 vm_object_terminate_pages(object);
780
781 #if VM_NRESERVLEVEL > 0
782 if (__predict_false(!LIST_EMPTY(&object->rvq)))
783 vm_reserv_break_all(object);
784 #endif
785
786 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
787 object->type == OBJT_SWAP,
788 ("%s: non-swap obj %p has cred", __func__, object));
789
790 /*
791 * Let the pager know object is dead.
792 */
793 vm_pager_deallocate(object);
794 VM_OBJECT_WUNLOCK(object);
795
796 vm_object_destroy(object);
797 }
798
799 /*
800 * Make the page read-only so that we can clear the object flags. However, if
801 * this is a nosync mmap then the object is likely to stay dirty so do not
802 * mess with the page and do not clear the object flags. Returns TRUE if the
803 * page should be flushed, and FALSE otherwise.
804 */
805 static boolean_t
806 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
807 {
808
809 /*
810 * If we have been asked to skip nosync pages and this is a
811 * nosync page, skip it. Note that the object flags were not
812 * cleared in this case so we do not have to set them.
813 */
814 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
815 *clearobjflags = FALSE;
816 return (FALSE);
817 } else {
818 pmap_remove_write(p);
819 return (p->dirty != 0);
820 }
821 }
822
823 /*
824 * vm_object_page_clean
825 *
826 * Clean all dirty pages in the specified range of object. Leaves page
827 * on whatever queue it is currently on. If NOSYNC is set then do not
828 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
829 * leaving the object dirty.
830 *
831 * When stuffing pages asynchronously, allow clustering. XXX we need a
832 * synchronous clustering mode implementation.
833 *
834 * Odd semantics: if start == end, we clean everything.
835 *
836 * The object must be locked.
837 *
838 * Returns FALSE if some page from the range was not written, as
839 * reported by the pager, and TRUE otherwise.
840 */
841 boolean_t
842 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
843 int flags)
844 {
845 vm_page_t np, p;
846 vm_pindex_t pi, tend, tstart;
847 int curgeneration, n, pagerflags;
848 boolean_t clearobjflags, eio, res;
849
850 VM_OBJECT_ASSERT_WLOCKED(object);
851
852 /*
853 * The OBJ_MIGHTBEDIRTY flag is only set for OBJT_VNODE
854 * objects. The check below prevents the function from
855 * operating on non-vnode objects.
856 */
857 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
858 object->resident_page_count == 0)
859 return (TRUE);
860
861 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
862 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
863 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
864
865 tstart = OFF_TO_IDX(start);
866 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
867 clearobjflags = tstart == 0 && tend >= object->size;
868 res = TRUE;
869
870 rescan:
871 curgeneration = object->generation;
872
873 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
874 pi = p->pindex;
875 if (pi >= tend)
876 break;
877 np = TAILQ_NEXT(p, listq);
878 if (p->valid == 0)
879 continue;
880 if (vm_page_sleep_if_busy(p, "vpcwai")) {
881 if (object->generation != curgeneration) {
882 if ((flags & OBJPC_SYNC) != 0)
883 goto rescan;
884 else
885 clearobjflags = FALSE;
886 }
887 np = vm_page_find_least(object, pi);
888 continue;
889 }
890 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
891 continue;
892
893 n = vm_object_page_collect_flush(object, p, pagerflags,
894 flags, &clearobjflags, &eio);
895 if (eio) {
896 res = FALSE;
897 clearobjflags = FALSE;
898 }
899 if (object->generation != curgeneration) {
900 if ((flags & OBJPC_SYNC) != 0)
901 goto rescan;
902 else
903 clearobjflags = FALSE;
904 }
905
906 /*
907 * If the VOP_PUTPAGES() did a truncated write, so
908 * that even the first page of the run is not fully
909 * written, vm_pageout_flush() returns 0 as the run
910 * length. Since the condition that caused truncated
911 * write may be permanent, e.g. exhausted free space,
912 * accepting n == 0 would cause an infinite loop.
913 *
914 * Forwarding the iterator leaves the unwritten page
915 * behind, but there is not much we can do there if
916 * filesystem refuses to write it.
917 */
918 if (n == 0) {
919 n = 1;
920 clearobjflags = FALSE;
921 }
922 np = vm_page_find_least(object, pi + n);
923 }
924 #if 0
925 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
926 #endif
927
928 if (clearobjflags)
929 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
930 return (res);
931 }
932
933 static int
934 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
935 int flags, boolean_t *clearobjflags, boolean_t *eio)
936 {
937 vm_page_t ma[vm_pageout_page_count], p_first, tp;
938 int count, i, mreq, runlen;
939
940 vm_page_lock_assert(p, MA_NOTOWNED);
941 VM_OBJECT_ASSERT_WLOCKED(object);
942
943 count = 1;
944 mreq = 0;
945
946 for (tp = p; count < vm_pageout_page_count; count++) {
947 tp = vm_page_next(tp);
948 if (tp == NULL || vm_page_busied(tp))
949 break;
950 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
951 break;
952 }
953
954 for (p_first = p; count < vm_pageout_page_count; count++) {
955 tp = vm_page_prev(p_first);
956 if (tp == NULL || vm_page_busied(tp))
957 break;
958 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
959 break;
960 p_first = tp;
961 mreq++;
962 }
963
964 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
965 ma[i] = tp;
966
967 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
968 return (runlen);
969 }
970
971 /*
972 * Note that there is absolutely no sense in writing out
973 * anonymous objects, so we track down the vnode object
974 * to write out.
975 * We invalidate (remove) all pages from the address space
976 * for semantic correctness.
977 *
978 * If the backing object is a device object with unmanaged pages, then any
979 * mappings to the specified range of pages must be removed before this
980 * function is called.
981 *
982 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
983 * may start out with a NULL object.
984 */
985 boolean_t
986 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
987 boolean_t syncio, boolean_t invalidate)
988 {
989 vm_object_t backing_object;
990 struct vnode *vp;
991 struct mount *mp;
992 int error, flags, fsync_after;
993 boolean_t res;
994
995 if (object == NULL)
996 return (TRUE);
997 res = TRUE;
998 error = 0;
999 VM_OBJECT_WLOCK(object);
1000 while ((backing_object = object->backing_object) != NULL) {
1001 VM_OBJECT_WLOCK(backing_object);
1002 offset += object->backing_object_offset;
1003 VM_OBJECT_WUNLOCK(object);
1004 object = backing_object;
1005 if (object->size < OFF_TO_IDX(offset + size))
1006 size = IDX_TO_OFF(object->size) - offset;
1007 }
1008 /*
1009 * Flush pages if writing is allowed, invalidate them
1010 * if invalidation requested. Pages undergoing I/O
1011 * will be ignored by vm_object_page_remove().
1012 *
1013 * We cannot lock the vnode and then wait for paging
1014 * to complete without deadlocking against vm_fault.
1015 * Instead we simply call vm_object_page_remove() and
1016 * allow it to block internally on a page-by-page
1017 * basis when it encounters pages undergoing async
1018 * I/O.
1019 */
1020 if (object->type == OBJT_VNODE &&
1021 (object->flags & OBJ_MIGHTBEDIRTY) != 0 &&
1022 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1023 VM_OBJECT_WUNLOCK(object);
1024 (void) vn_start_write(vp, &mp, V_WAIT);
1025 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1026 if (syncio && !invalidate && offset == 0 &&
1027 atop(size) == object->size) {
1028 /*
1029 * If syncing the whole mapping of the file,
1030 * it is faster to schedule all the writes in
1031 * async mode, also allowing the clustering,
1032 * and then wait for i/o to complete.
1033 */
1034 flags = 0;
1035 fsync_after = TRUE;
1036 } else {
1037 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1038 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1039 fsync_after = FALSE;
1040 }
1041 VM_OBJECT_WLOCK(object);
1042 res = vm_object_page_clean(object, offset, offset + size,
1043 flags);
1044 VM_OBJECT_WUNLOCK(object);
1045 if (fsync_after)
1046 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1047 VOP_UNLOCK(vp, 0);
1048 vn_finished_write(mp);
1049 if (error != 0)
1050 res = FALSE;
1051 VM_OBJECT_WLOCK(object);
1052 }
1053 if ((object->type == OBJT_VNODE ||
1054 object->type == OBJT_DEVICE) && invalidate) {
1055 if (object->type == OBJT_DEVICE)
1056 /*
1057 * The option OBJPR_NOTMAPPED must be passed here
1058 * because vm_object_page_remove() cannot remove
1059 * unmanaged mappings.
1060 */
1061 flags = OBJPR_NOTMAPPED;
1062 else if (old_msync)
1063 flags = 0;
1064 else
1065 flags = OBJPR_CLEANONLY;
1066 vm_object_page_remove(object, OFF_TO_IDX(offset),
1067 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1068 }
1069 VM_OBJECT_WUNLOCK(object);
1070 return (res);
1071 }
1072
1073 /*
1074 * Determine whether the given advice can be applied to the object. Advice is
1075 * not applied to unmanaged pages since they never belong to page queues, and
1076 * since MADV_FREE is destructive, it can apply only to anonymous pages that
1077 * have been mapped at most once.
1078 */
1079 static bool
1080 vm_object_advice_applies(vm_object_t object, int advice)
1081 {
1082
1083 if ((object->flags & OBJ_UNMANAGED) != 0)
1084 return (false);
1085 if (advice != MADV_FREE)
1086 return (true);
1087 return ((object->type == OBJT_DEFAULT || object->type == OBJT_SWAP) &&
1088 (object->flags & OBJ_ONEMAPPING) != 0);
1089 }
1090
1091 static void
1092 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1093 vm_size_t size)
1094 {
1095
1096 if (advice == MADV_FREE && object->type == OBJT_SWAP)
1097 swap_pager_freespace(object, pindex, size);
1098 }
1099
1100 /*
1101 * vm_object_madvise:
1102 *
1103 * Implements the madvise function at the object/page level.
1104 *
1105 * MADV_WILLNEED (any object)
1106 *
1107 * Activate the specified pages if they are resident.
1108 *
1109 * MADV_DONTNEED (any object)
1110 *
1111 * Deactivate the specified pages if they are resident.
1112 *
1113 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1114 * OBJ_ONEMAPPING only)
1115 *
1116 * Deactivate and clean the specified pages if they are
1117 * resident. This permits the process to reuse the pages
1118 * without faulting or the kernel to reclaim the pages
1119 * without I/O.
1120 */
1121 void
1122 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1123 int advice)
1124 {
1125 vm_pindex_t tpindex;
1126 vm_object_t backing_object, tobject;
1127 vm_page_t m, tm;
1128
1129 if (object == NULL)
1130 return;
1131
1132 relookup:
1133 VM_OBJECT_WLOCK(object);
1134 if (!vm_object_advice_applies(object, advice)) {
1135 VM_OBJECT_WUNLOCK(object);
1136 return;
1137 }
1138 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1139 tobject = object;
1140
1141 /*
1142 * If the next page isn't resident in the top-level object, we
1143 * need to search the shadow chain. When applying MADV_FREE, we
1144 * take care to release any swap space used to store
1145 * non-resident pages.
1146 */
1147 if (m == NULL || pindex < m->pindex) {
1148 /*
1149 * Optimize a common case: if the top-level object has
1150 * no backing object, we can skip over the non-resident
1151 * range in constant time.
1152 */
1153 if (object->backing_object == NULL) {
1154 tpindex = (m != NULL && m->pindex < end) ?
1155 m->pindex : end;
1156 vm_object_madvise_freespace(object, advice,
1157 pindex, tpindex - pindex);
1158 if ((pindex = tpindex) == end)
1159 break;
1160 goto next_page;
1161 }
1162
1163 tpindex = pindex;
1164 do {
1165 vm_object_madvise_freespace(tobject, advice,
1166 tpindex, 1);
1167 /*
1168 * Prepare to search the next object in the
1169 * chain.
1170 */
1171 backing_object = tobject->backing_object;
1172 if (backing_object == NULL)
1173 goto next_pindex;
1174 VM_OBJECT_WLOCK(backing_object);
1175 tpindex +=
1176 OFF_TO_IDX(tobject->backing_object_offset);
1177 if (tobject != object)
1178 VM_OBJECT_WUNLOCK(tobject);
1179 tobject = backing_object;
1180 if (!vm_object_advice_applies(tobject, advice))
1181 goto next_pindex;
1182 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
1183 NULL);
1184 } else {
1185 next_page:
1186 tm = m;
1187 m = TAILQ_NEXT(m, listq);
1188 }
1189
1190 /*
1191 * If the page is not in a normal state, skip it.
1192 */
1193 if (tm->valid != VM_PAGE_BITS_ALL)
1194 goto next_pindex;
1195 vm_page_lock(tm);
1196 if (vm_page_held(tm)) {
1197 vm_page_unlock(tm);
1198 goto next_pindex;
1199 }
1200 KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1201 ("vm_object_madvise: page %p is fictitious", tm));
1202 KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1203 ("vm_object_madvise: page %p is not managed", tm));
1204 if (vm_page_busied(tm)) {
1205 if (object != tobject)
1206 VM_OBJECT_WUNLOCK(tobject);
1207 VM_OBJECT_WUNLOCK(object);
1208 if (advice == MADV_WILLNEED) {
1209 /*
1210 * Reference the page before unlocking and
1211 * sleeping so that the page daemon is less
1212 * likely to reclaim it.
1213 */
1214 vm_page_aflag_set(tm, PGA_REFERENCED);
1215 }
1216 vm_page_busy_sleep(tm, "madvpo", false);
1217 goto relookup;
1218 }
1219 vm_page_advise(tm, advice);
1220 vm_page_unlock(tm);
1221 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1222 next_pindex:
1223 if (tobject != object)
1224 VM_OBJECT_WUNLOCK(tobject);
1225 }
1226 VM_OBJECT_WUNLOCK(object);
1227 }
1228
1229 /*
1230 * vm_object_shadow:
1231 *
1232 * Create a new object which is backed by the
1233 * specified existing object range. The source
1234 * object reference is deallocated.
1235 *
1236 * The new object and offset into that object
1237 * are returned in the source parameters.
1238 */
1239 void
1240 vm_object_shadow(
1241 vm_object_t *object, /* IN/OUT */
1242 vm_ooffset_t *offset, /* IN/OUT */
1243 vm_size_t length)
1244 {
1245 vm_object_t source;
1246 vm_object_t result;
1247
1248 source = *object;
1249
1250 /*
1251 * Don't create the new object if the old object isn't shared.
1252 */
1253 if (source != NULL) {
1254 VM_OBJECT_WLOCK(source);
1255 if (source->ref_count == 1 &&
1256 source->handle == NULL &&
1257 (source->type == OBJT_DEFAULT ||
1258 source->type == OBJT_SWAP)) {
1259 VM_OBJECT_WUNLOCK(source);
1260 return;
1261 }
1262 VM_OBJECT_WUNLOCK(source);
1263 }
1264
1265 /*
1266 * Allocate a new object with the given length.
1267 */
1268 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1269
1270 /*
1271 * The new object shadows the source object, adding a reference to it.
1272 * Our caller changes his reference to point to the new object,
1273 * removing a reference to the source object. Net result: no change
1274 * of reference count.
1275 *
1276 * Try to optimize the result object's page color when shadowing
1277 * in order to maintain page coloring consistency in the combined
1278 * shadowed object.
1279 */
1280 result->backing_object = source;
1281 /*
1282 * Store the offset into the source object, and fix up the offset into
1283 * the new object.
1284 */
1285 result->backing_object_offset = *offset;
1286 if (source != NULL) {
1287 VM_OBJECT_WLOCK(source);
1288 result->domain = source->domain;
1289 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1290 source->shadow_count++;
1291 #if VM_NRESERVLEVEL > 0
1292 result->flags |= source->flags & OBJ_COLORED;
1293 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1294 ((1 << (VM_NFREEORDER - 1)) - 1);
1295 #endif
1296 VM_OBJECT_WUNLOCK(source);
1297 }
1298
1299
1300 /*
1301 * Return the new things
1302 */
1303 *offset = 0;
1304 *object = result;
1305 }
1306
1307 /*
1308 * vm_object_split:
1309 *
1310 * Split the pages in a map entry into a new object. This affords
1311 * easier removal of unused pages, and keeps object inheritance from
1312 * being a negative impact on memory usage.
1313 */
1314 void
1315 vm_object_split(vm_map_entry_t entry)
1316 {
1317 vm_page_t m, m_busy, m_next;
1318 vm_object_t orig_object, new_object, source;
1319 vm_pindex_t idx, offidxstart;
1320 vm_size_t size;
1321
1322 orig_object = entry->object.vm_object;
1323 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1324 return;
1325 if (orig_object->ref_count <= 1)
1326 return;
1327 VM_OBJECT_WUNLOCK(orig_object);
1328
1329 offidxstart = OFF_TO_IDX(entry->offset);
1330 size = atop(entry->end - entry->start);
1331
1332 /*
1333 * If swap_pager_copy() is later called, it will convert new_object
1334 * into a swap object.
1335 */
1336 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1337
1338 /*
1339 * At this point, the new object is still private, so the order in
1340 * which the original and new objects are locked does not matter.
1341 */
1342 VM_OBJECT_WLOCK(new_object);
1343 VM_OBJECT_WLOCK(orig_object);
1344 new_object->domain = orig_object->domain;
1345 source = orig_object->backing_object;
1346 if (source != NULL) {
1347 VM_OBJECT_WLOCK(source);
1348 if ((source->flags & OBJ_DEAD) != 0) {
1349 VM_OBJECT_WUNLOCK(source);
1350 VM_OBJECT_WUNLOCK(orig_object);
1351 VM_OBJECT_WUNLOCK(new_object);
1352 vm_object_deallocate(new_object);
1353 VM_OBJECT_WLOCK(orig_object);
1354 return;
1355 }
1356 LIST_INSERT_HEAD(&source->shadow_head,
1357 new_object, shadow_list);
1358 source->shadow_count++;
1359 vm_object_reference_locked(source); /* for new_object */
1360 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1361 VM_OBJECT_WUNLOCK(source);
1362 new_object->backing_object_offset =
1363 orig_object->backing_object_offset + entry->offset;
1364 new_object->backing_object = source;
1365 }
1366 if (orig_object->cred != NULL) {
1367 new_object->cred = orig_object->cred;
1368 crhold(orig_object->cred);
1369 new_object->charge = ptoa(size);
1370 KASSERT(orig_object->charge >= ptoa(size),
1371 ("orig_object->charge < 0"));
1372 orig_object->charge -= ptoa(size);
1373 }
1374 m_busy = NULL;
1375 #ifdef INVARIANTS
1376 idx = 0;
1377 #endif
1378 retry:
1379 m = vm_page_find_least(orig_object, offidxstart);
1380 KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1381 ("%s: object %p was repopulated", __func__, orig_object));
1382 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1383 m = m_next) {
1384 m_next = TAILQ_NEXT(m, listq);
1385
1386 /*
1387 * We must wait for pending I/O to complete before we can
1388 * rename the page.
1389 *
1390 * We do not have to VM_PROT_NONE the page as mappings should
1391 * not be changed by this operation.
1392 */
1393 if (vm_page_busied(m)) {
1394 VM_OBJECT_WUNLOCK(new_object);
1395 vm_page_lock(m);
1396 VM_OBJECT_WUNLOCK(orig_object);
1397 vm_page_busy_sleep(m, "spltwt", false);
1398 VM_OBJECT_WLOCK(orig_object);
1399 VM_OBJECT_WLOCK(new_object);
1400 goto retry;
1401 }
1402
1403 /* vm_page_rename() will dirty the page. */
1404 if (vm_page_rename(m, new_object, idx)) {
1405 VM_OBJECT_WUNLOCK(new_object);
1406 VM_OBJECT_WUNLOCK(orig_object);
1407 vm_radix_wait();
1408 VM_OBJECT_WLOCK(orig_object);
1409 VM_OBJECT_WLOCK(new_object);
1410 goto retry;
1411 }
1412 #if VM_NRESERVLEVEL > 0
1413 /*
1414 * If some of the reservation's allocated pages remain with
1415 * the original object, then transferring the reservation to
1416 * the new object is neither particularly beneficial nor
1417 * particularly harmful as compared to leaving the reservation
1418 * with the original object. If, however, all of the
1419 * reservation's allocated pages are transferred to the new
1420 * object, then transferring the reservation is typically
1421 * beneficial. Determining which of these two cases applies
1422 * would be more costly than unconditionally renaming the
1423 * reservation.
1424 */
1425 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1426 #endif
1427
1428 /*
1429 * orig_object's type may change while sleeping, so keep track
1430 * of the beginning of the busied range.
1431 */
1432 if (orig_object->type == OBJT_SWAP) {
1433 vm_page_xbusy(m);
1434 if (m_busy == NULL)
1435 m_busy = m;
1436 }
1437 }
1438 if (orig_object->type == OBJT_SWAP) {
1439 /*
1440 * swap_pager_copy() can sleep, in which case the orig_object's
1441 * and new_object's locks are released and reacquired.
1442 */
1443 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1444 if (m_busy != NULL)
1445 TAILQ_FOREACH_FROM(m_busy, &new_object->memq, listq)
1446 vm_page_xunbusy(m_busy);
1447 }
1448 VM_OBJECT_WUNLOCK(orig_object);
1449 VM_OBJECT_WUNLOCK(new_object);
1450 entry->object.vm_object = new_object;
1451 entry->offset = 0LL;
1452 vm_object_deallocate(orig_object);
1453 VM_OBJECT_WLOCK(new_object);
1454 }
1455
1456 #define OBSC_COLLAPSE_NOWAIT 0x0002
1457 #define OBSC_COLLAPSE_WAIT 0x0004
1458
1459 static vm_page_t
1460 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p, vm_page_t next,
1461 int op)
1462 {
1463 vm_object_t backing_object;
1464
1465 VM_OBJECT_ASSERT_WLOCKED(object);
1466 backing_object = object->backing_object;
1467 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1468
1469 KASSERT(p == NULL || vm_page_busied(p), ("unbusy page %p", p));
1470 KASSERT(p == NULL || p->object == object || p->object == backing_object,
1471 ("invalid ownership %p %p %p", p, object, backing_object));
1472 if ((op & OBSC_COLLAPSE_NOWAIT) != 0)
1473 return (next);
1474 if (p != NULL)
1475 vm_page_lock(p);
1476 VM_OBJECT_WUNLOCK(object);
1477 VM_OBJECT_WUNLOCK(backing_object);
1478 /* The page is only NULL when rename fails. */
1479 if (p == NULL)
1480 vm_radix_wait();
1481 else
1482 vm_page_busy_sleep(p, "vmocol", false);
1483 VM_OBJECT_WLOCK(object);
1484 VM_OBJECT_WLOCK(backing_object);
1485 return (TAILQ_FIRST(&backing_object->memq));
1486 }
1487
1488 static bool
1489 vm_object_scan_all_shadowed(vm_object_t object)
1490 {
1491 vm_object_t backing_object;
1492 vm_page_t p, pp;
1493 vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1494
1495 VM_OBJECT_ASSERT_WLOCKED(object);
1496 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1497
1498 backing_object = object->backing_object;
1499
1500 if (backing_object->type != OBJT_DEFAULT &&
1501 backing_object->type != OBJT_SWAP)
1502 return (false);
1503
1504 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1505 p = vm_page_find_least(backing_object, pi);
1506 ps = swap_pager_find_least(backing_object, pi);
1507
1508 /*
1509 * Only check pages inside the parent object's range and
1510 * inside the parent object's mapping of the backing object.
1511 */
1512 for (;; pi++) {
1513 if (p != NULL && p->pindex < pi)
1514 p = TAILQ_NEXT(p, listq);
1515 if (ps < pi)
1516 ps = swap_pager_find_least(backing_object, pi);
1517 if (p == NULL && ps >= backing_object->size)
1518 break;
1519 else if (p == NULL)
1520 pi = ps;
1521 else
1522 pi = MIN(p->pindex, ps);
1523
1524 new_pindex = pi - backing_offset_index;
1525 if (new_pindex >= object->size)
1526 break;
1527
1528 /*
1529 * See if the parent has the page or if the parent's object
1530 * pager has the page. If the parent has the page but the page
1531 * is not valid, the parent's object pager must have the page.
1532 *
1533 * If this fails, the parent does not completely shadow the
1534 * object and we might as well give up now.
1535 */
1536 pp = vm_page_lookup(object, new_pindex);
1537 if ((pp == NULL || pp->valid == 0) &&
1538 !vm_pager_has_page(object, new_pindex, NULL, NULL))
1539 return (false);
1540 }
1541 return (true);
1542 }
1543
1544 static bool
1545 vm_object_collapse_scan(vm_object_t object, int op)
1546 {
1547 vm_object_t backing_object;
1548 vm_page_t next, p, pp;
1549 vm_pindex_t backing_offset_index, new_pindex;
1550
1551 VM_OBJECT_ASSERT_WLOCKED(object);
1552 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1553
1554 backing_object = object->backing_object;
1555 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1556
1557 /*
1558 * Initial conditions
1559 */
1560 if ((op & OBSC_COLLAPSE_WAIT) != 0)
1561 vm_object_set_flag(backing_object, OBJ_DEAD);
1562
1563 /*
1564 * Our scan
1565 */
1566 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1567 next = TAILQ_NEXT(p, listq);
1568 new_pindex = p->pindex - backing_offset_index;
1569
1570 /*
1571 * Check for busy page
1572 */
1573 if (vm_page_busied(p)) {
1574 next = vm_object_collapse_scan_wait(object, p, next, op);
1575 continue;
1576 }
1577
1578 KASSERT(p->object == backing_object,
1579 ("vm_object_collapse_scan: object mismatch"));
1580
1581 if (p->pindex < backing_offset_index ||
1582 new_pindex >= object->size) {
1583 if (backing_object->type == OBJT_SWAP)
1584 swap_pager_freespace(backing_object, p->pindex,
1585 1);
1586
1587 /*
1588 * Page is out of the parent object's range, we can
1589 * simply destroy it.
1590 */
1591 vm_page_lock(p);
1592 KASSERT(!pmap_page_is_mapped(p),
1593 ("freeing mapped page %p", p));
1594 if (vm_page_remove(p))
1595 vm_page_free(p);
1596 vm_page_unlock(p);
1597 continue;
1598 }
1599
1600 pp = vm_page_lookup(object, new_pindex);
1601 if (pp != NULL && vm_page_busied(pp)) {
1602 /*
1603 * The page in the parent is busy and possibly not
1604 * (yet) valid. Until its state is finalized by the
1605 * busy bit owner, we can't tell whether it shadows the
1606 * original page. Therefore, we must either skip it
1607 * and the original (backing_object) page or wait for
1608 * its state to be finalized.
1609 *
1610 * This is due to a race with vm_fault() where we must
1611 * unbusy the original (backing_obj) page before we can
1612 * (re)lock the parent. Hence we can get here.
1613 */
1614 next = vm_object_collapse_scan_wait(object, pp, next,
1615 op);
1616 continue;
1617 }
1618
1619 KASSERT(pp == NULL || pp->valid != 0,
1620 ("unbusy invalid page %p", pp));
1621
1622 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1623 NULL)) {
1624 /*
1625 * The page already exists in the parent OR swap exists
1626 * for this location in the parent. Leave the parent's
1627 * page alone. Destroy the original page from the
1628 * backing object.
1629 */
1630 if (backing_object->type == OBJT_SWAP)
1631 swap_pager_freespace(backing_object, p->pindex,
1632 1);
1633 vm_page_lock(p);
1634 KASSERT(!pmap_page_is_mapped(p),
1635 ("freeing mapped page %p", p));
1636 if (vm_page_remove(p))
1637 vm_page_free(p);
1638 vm_page_unlock(p);
1639 continue;
1640 }
1641
1642 /*
1643 * Page does not exist in parent, rename the page from the
1644 * backing object to the main object.
1645 *
1646 * If the page was mapped to a process, it can remain mapped
1647 * through the rename. vm_page_rename() will dirty the page.
1648 */
1649 if (vm_page_rename(p, object, new_pindex)) {
1650 next = vm_object_collapse_scan_wait(object, NULL, next,
1651 op);
1652 continue;
1653 }
1654
1655 /* Use the old pindex to free the right page. */
1656 if (backing_object->type == OBJT_SWAP)
1657 swap_pager_freespace(backing_object,
1658 new_pindex + backing_offset_index, 1);
1659
1660 #if VM_NRESERVLEVEL > 0
1661 /*
1662 * Rename the reservation.
1663 */
1664 vm_reserv_rename(p, object, backing_object,
1665 backing_offset_index);
1666 #endif
1667 }
1668 return (true);
1669 }
1670
1671
1672 /*
1673 * this version of collapse allows the operation to occur earlier and
1674 * when paging_in_progress is true for an object... This is not a complete
1675 * operation, but should plug 99.9% of the rest of the leaks.
1676 */
1677 static void
1678 vm_object_qcollapse(vm_object_t object)
1679 {
1680 vm_object_t backing_object = object->backing_object;
1681
1682 VM_OBJECT_ASSERT_WLOCKED(object);
1683 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1684
1685 if (backing_object->ref_count != 1)
1686 return;
1687
1688 vm_object_collapse_scan(object, OBSC_COLLAPSE_NOWAIT);
1689 }
1690
1691 /*
1692 * vm_object_collapse:
1693 *
1694 * Collapse an object with the object backing it.
1695 * Pages in the backing object are moved into the
1696 * parent, and the backing object is deallocated.
1697 */
1698 void
1699 vm_object_collapse(vm_object_t object)
1700 {
1701 vm_object_t backing_object, new_backing_object;
1702
1703 VM_OBJECT_ASSERT_WLOCKED(object);
1704
1705 while (TRUE) {
1706 /*
1707 * Verify that the conditions are right for collapse:
1708 *
1709 * The object exists and the backing object exists.
1710 */
1711 if ((backing_object = object->backing_object) == NULL)
1712 break;
1713
1714 /*
1715 * we check the backing object first, because it is most likely
1716 * not collapsable.
1717 */
1718 VM_OBJECT_WLOCK(backing_object);
1719 if (backing_object->handle != NULL ||
1720 (backing_object->type != OBJT_DEFAULT &&
1721 backing_object->type != OBJT_SWAP) ||
1722 (backing_object->flags & (OBJ_DEAD | OBJ_NOSPLIT)) != 0 ||
1723 object->handle != NULL ||
1724 (object->type != OBJT_DEFAULT &&
1725 object->type != OBJT_SWAP) ||
1726 (object->flags & OBJ_DEAD)) {
1727 VM_OBJECT_WUNLOCK(backing_object);
1728 break;
1729 }
1730
1731 if (object->paging_in_progress != 0 ||
1732 backing_object->paging_in_progress != 0) {
1733 vm_object_qcollapse(object);
1734 VM_OBJECT_WUNLOCK(backing_object);
1735 break;
1736 }
1737
1738 /*
1739 * We know that we can either collapse the backing object (if
1740 * the parent is the only reference to it) or (perhaps) have
1741 * the parent bypass the object if the parent happens to shadow
1742 * all the resident pages in the entire backing object.
1743 *
1744 * This is ignoring pager-backed pages such as swap pages.
1745 * vm_object_collapse_scan fails the shadowing test in this
1746 * case.
1747 */
1748 if (backing_object->ref_count == 1) {
1749 vm_object_pip_add(object, 1);
1750 vm_object_pip_add(backing_object, 1);
1751
1752 /*
1753 * If there is exactly one reference to the backing
1754 * object, we can collapse it into the parent.
1755 */
1756 vm_object_collapse_scan(object, OBSC_COLLAPSE_WAIT);
1757
1758 #if VM_NRESERVLEVEL > 0
1759 /*
1760 * Break any reservations from backing_object.
1761 */
1762 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1763 vm_reserv_break_all(backing_object);
1764 #endif
1765
1766 /*
1767 * Move the pager from backing_object to object.
1768 */
1769 if (backing_object->type == OBJT_SWAP) {
1770 /*
1771 * swap_pager_copy() can sleep, in which case
1772 * the backing_object's and object's locks are
1773 * released and reacquired.
1774 * Since swap_pager_copy() is being asked to
1775 * destroy the source, it will change the
1776 * backing_object's type to OBJT_DEFAULT.
1777 */
1778 swap_pager_copy(
1779 backing_object,
1780 object,
1781 OFF_TO_IDX(object->backing_object_offset), TRUE);
1782 }
1783 /*
1784 * Object now shadows whatever backing_object did.
1785 * Note that the reference to
1786 * backing_object->backing_object moves from within
1787 * backing_object to within object.
1788 */
1789 LIST_REMOVE(object, shadow_list);
1790 backing_object->shadow_count--;
1791 if (backing_object->backing_object) {
1792 VM_OBJECT_WLOCK(backing_object->backing_object);
1793 LIST_REMOVE(backing_object, shadow_list);
1794 LIST_INSERT_HEAD(
1795 &backing_object->backing_object->shadow_head,
1796 object, shadow_list);
1797 /*
1798 * The shadow_count has not changed.
1799 */
1800 VM_OBJECT_WUNLOCK(backing_object->backing_object);
1801 }
1802 object->backing_object = backing_object->backing_object;
1803 object->backing_object_offset +=
1804 backing_object->backing_object_offset;
1805
1806 /*
1807 * Discard backing_object.
1808 *
1809 * Since the backing object has no pages, no pager left,
1810 * and no object references within it, all that is
1811 * necessary is to dispose of it.
1812 */
1813 KASSERT(backing_object->ref_count == 1, (
1814 "backing_object %p was somehow re-referenced during collapse!",
1815 backing_object));
1816 vm_object_pip_wakeup(backing_object);
1817 backing_object->type = OBJT_DEAD;
1818 backing_object->ref_count = 0;
1819 VM_OBJECT_WUNLOCK(backing_object);
1820 vm_object_destroy(backing_object);
1821
1822 vm_object_pip_wakeup(object);
1823 counter_u64_add(object_collapses, 1);
1824 } else {
1825 /*
1826 * If we do not entirely shadow the backing object,
1827 * there is nothing we can do so we give up.
1828 */
1829 if (object->resident_page_count != object->size &&
1830 !vm_object_scan_all_shadowed(object)) {
1831 VM_OBJECT_WUNLOCK(backing_object);
1832 break;
1833 }
1834
1835 /*
1836 * Make the parent shadow the next object in the
1837 * chain. Deallocating backing_object will not remove
1838 * it, since its reference count is at least 2.
1839 */
1840 LIST_REMOVE(object, shadow_list);
1841 backing_object->shadow_count--;
1842
1843 new_backing_object = backing_object->backing_object;
1844 if ((object->backing_object = new_backing_object) != NULL) {
1845 VM_OBJECT_WLOCK(new_backing_object);
1846 LIST_INSERT_HEAD(
1847 &new_backing_object->shadow_head,
1848 object,
1849 shadow_list
1850 );
1851 new_backing_object->shadow_count++;
1852 vm_object_reference_locked(new_backing_object);
1853 VM_OBJECT_WUNLOCK(new_backing_object);
1854 object->backing_object_offset +=
1855 backing_object->backing_object_offset;
1856 }
1857
1858 /*
1859 * Drop the reference count on backing_object. Since
1860 * its ref_count was at least 2, it will not vanish.
1861 */
1862 backing_object->ref_count--;
1863 VM_OBJECT_WUNLOCK(backing_object);
1864 counter_u64_add(object_bypasses, 1);
1865 }
1866
1867 /*
1868 * Try again with this object's new backing object.
1869 */
1870 }
1871 }
1872
1873 /*
1874 * vm_object_page_remove:
1875 *
1876 * For the given object, either frees or invalidates each of the
1877 * specified pages. In general, a page is freed. However, if a page is
1878 * wired for any reason other than the existence of a managed, wired
1879 * mapping, then it may be invalidated but not removed from the object.
1880 * Pages are specified by the given range ["start", "end") and the option
1881 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1882 * extends from "start" to the end of the object. If the option
1883 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1884 * specified range are affected. If the option OBJPR_NOTMAPPED is
1885 * specified, then the pages within the specified range must have no
1886 * mappings. Otherwise, if this option is not specified, any mappings to
1887 * the specified pages are removed before the pages are freed or
1888 * invalidated.
1889 *
1890 * In general, this operation should only be performed on objects that
1891 * contain managed pages. There are, however, two exceptions. First, it
1892 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1893 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1894 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1895 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1896 *
1897 * The object must be locked.
1898 */
1899 void
1900 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1901 int options)
1902 {
1903 vm_page_t p, next;
1904 struct mtx *mtx;
1905
1906 VM_OBJECT_ASSERT_WLOCKED(object);
1907 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
1908 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1909 ("vm_object_page_remove: illegal options for object %p", object));
1910 if (object->resident_page_count == 0)
1911 return;
1912 vm_object_pip_add(object, 1);
1913 again:
1914 p = vm_page_find_least(object, start);
1915 mtx = NULL;
1916
1917 /*
1918 * Here, the variable "p" is either (1) the page with the least pindex
1919 * greater than or equal to the parameter "start" or (2) NULL.
1920 */
1921 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1922 next = TAILQ_NEXT(p, listq);
1923
1924 /*
1925 * If the page is wired for any reason besides the existence
1926 * of managed, wired mappings, then it cannot be freed. For
1927 * example, fictitious pages, which represent device memory,
1928 * are inherently wired and cannot be freed. They can,
1929 * however, be invalidated if the option OBJPR_CLEANONLY is
1930 * not specified.
1931 */
1932 vm_page_change_lock(p, &mtx);
1933 if (vm_page_xbusied(p)) {
1934 VM_OBJECT_WUNLOCK(object);
1935 vm_page_busy_sleep(p, "vmopax", true);
1936 VM_OBJECT_WLOCK(object);
1937 goto again;
1938 }
1939 if (vm_page_wired(p)) {
1940 if ((options & OBJPR_NOTMAPPED) == 0 &&
1941 object->ref_count != 0)
1942 pmap_remove_all(p);
1943 if ((options & OBJPR_CLEANONLY) == 0) {
1944 p->valid = 0;
1945 vm_page_undirty(p);
1946 }
1947 continue;
1948 }
1949 if (vm_page_busied(p)) {
1950 VM_OBJECT_WUNLOCK(object);
1951 vm_page_busy_sleep(p, "vmopar", false);
1952 VM_OBJECT_WLOCK(object);
1953 goto again;
1954 }
1955 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1956 ("vm_object_page_remove: page %p is fictitious", p));
1957 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1958 if ((options & OBJPR_NOTMAPPED) == 0 &&
1959 object->ref_count != 0)
1960 pmap_remove_write(p);
1961 if (p->dirty != 0)
1962 continue;
1963 }
1964 if ((options & OBJPR_NOTMAPPED) == 0 && object->ref_count != 0)
1965 pmap_remove_all(p);
1966 vm_page_free(p);
1967 }
1968 if (mtx != NULL)
1969 mtx_unlock(mtx);
1970 vm_object_pip_wakeup(object);
1971 }
1972
1973 /*
1974 * vm_object_page_noreuse:
1975 *
1976 * For the given object, attempt to move the specified pages to
1977 * the head of the inactive queue. This bypasses regular LRU
1978 * operation and allows the pages to be reused quickly under memory
1979 * pressure. If a page is wired for any reason, then it will not
1980 * be queued. Pages are specified by the range ["start", "end").
1981 * As a special case, if "end" is zero, then the range extends from
1982 * "start" to the end of the object.
1983 *
1984 * This operation should only be performed on objects that
1985 * contain non-fictitious, managed pages.
1986 *
1987 * The object must be locked.
1988 */
1989 void
1990 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1991 {
1992 struct mtx *mtx;
1993 vm_page_t p, next;
1994
1995 VM_OBJECT_ASSERT_LOCKED(object);
1996 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
1997 ("vm_object_page_noreuse: illegal object %p", object));
1998 if (object->resident_page_count == 0)
1999 return;
2000 p = vm_page_find_least(object, start);
2001
2002 /*
2003 * Here, the variable "p" is either (1) the page with the least pindex
2004 * greater than or equal to the parameter "start" or (2) NULL.
2005 */
2006 mtx = NULL;
2007 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2008 next = TAILQ_NEXT(p, listq);
2009 vm_page_change_lock(p, &mtx);
2010 vm_page_deactivate_noreuse(p);
2011 }
2012 if (mtx != NULL)
2013 mtx_unlock(mtx);
2014 }
2015
2016 /*
2017 * Populate the specified range of the object with valid pages. Returns
2018 * TRUE if the range is successfully populated and FALSE otherwise.
2019 *
2020 * Note: This function should be optimized to pass a larger array of
2021 * pages to vm_pager_get_pages() before it is applied to a non-
2022 * OBJT_DEVICE object.
2023 *
2024 * The object must be locked.
2025 */
2026 boolean_t
2027 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2028 {
2029 vm_page_t m;
2030 vm_pindex_t pindex;
2031 int rv;
2032
2033 VM_OBJECT_ASSERT_WLOCKED(object);
2034 for (pindex = start; pindex < end; pindex++) {
2035 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL);
2036 if (m->valid != VM_PAGE_BITS_ALL) {
2037 rv = vm_pager_get_pages(object, &m, 1, NULL, NULL);
2038 if (rv != VM_PAGER_OK) {
2039 vm_page_lock(m);
2040 vm_page_free(m);
2041 vm_page_unlock(m);
2042 break;
2043 }
2044 }
2045 /*
2046 * Keep "m" busy because a subsequent iteration may unlock
2047 * the object.
2048 */
2049 }
2050 if (pindex > start) {
2051 m = vm_page_lookup(object, start);
2052 while (m != NULL && m->pindex < pindex) {
2053 vm_page_xunbusy(m);
2054 m = TAILQ_NEXT(m, listq);
2055 }
2056 }
2057 return (pindex == end);
2058 }
2059
2060 /*
2061 * Routine: vm_object_coalesce
2062 * Function: Coalesces two objects backing up adjoining
2063 * regions of memory into a single object.
2064 *
2065 * returns TRUE if objects were combined.
2066 *
2067 * NOTE: Only works at the moment if the second object is NULL -
2068 * if it's not, which object do we lock first?
2069 *
2070 * Parameters:
2071 * prev_object First object to coalesce
2072 * prev_offset Offset into prev_object
2073 * prev_size Size of reference to prev_object
2074 * next_size Size of reference to the second object
2075 * reserved Indicator that extension region has
2076 * swap accounted for
2077 *
2078 * Conditions:
2079 * The object must *not* be locked.
2080 */
2081 boolean_t
2082 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2083 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2084 {
2085 vm_pindex_t next_pindex;
2086
2087 if (prev_object == NULL)
2088 return (TRUE);
2089 VM_OBJECT_WLOCK(prev_object);
2090 if ((prev_object->type != OBJT_DEFAULT &&
2091 prev_object->type != OBJT_SWAP) ||
2092 (prev_object->flags & OBJ_NOSPLIT) != 0) {
2093 VM_OBJECT_WUNLOCK(prev_object);
2094 return (FALSE);
2095 }
2096
2097 /*
2098 * Try to collapse the object first
2099 */
2100 vm_object_collapse(prev_object);
2101
2102 /*
2103 * Can't coalesce if: . more than one reference . paged out . shadows
2104 * another object . has a copy elsewhere (any of which mean that the
2105 * pages not mapped to prev_entry may be in use anyway)
2106 */
2107 if (prev_object->backing_object != NULL) {
2108 VM_OBJECT_WUNLOCK(prev_object);
2109 return (FALSE);
2110 }
2111
2112 prev_size >>= PAGE_SHIFT;
2113 next_size >>= PAGE_SHIFT;
2114 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2115
2116 if (prev_object->ref_count > 1 &&
2117 prev_object->size != next_pindex &&
2118 (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2119 VM_OBJECT_WUNLOCK(prev_object);
2120 return (FALSE);
2121 }
2122
2123 /*
2124 * Account for the charge.
2125 */
2126 if (prev_object->cred != NULL) {
2127
2128 /*
2129 * If prev_object was charged, then this mapping,
2130 * although not charged now, may become writable
2131 * later. Non-NULL cred in the object would prevent
2132 * swap reservation during enabling of the write
2133 * access, so reserve swap now. Failed reservation
2134 * cause allocation of the separate object for the map
2135 * entry, and swap reservation for this entry is
2136 * managed in appropriate time.
2137 */
2138 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2139 prev_object->cred)) {
2140 VM_OBJECT_WUNLOCK(prev_object);
2141 return (FALSE);
2142 }
2143 prev_object->charge += ptoa(next_size);
2144 }
2145
2146 /*
2147 * Remove any pages that may still be in the object from a previous
2148 * deallocation.
2149 */
2150 if (next_pindex < prev_object->size) {
2151 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2152 next_size, 0);
2153 if (prev_object->type == OBJT_SWAP)
2154 swap_pager_freespace(prev_object,
2155 next_pindex, next_size);
2156 #if 0
2157 if (prev_object->cred != NULL) {
2158 KASSERT(prev_object->charge >=
2159 ptoa(prev_object->size - next_pindex),
2160 ("object %p overcharged 1 %jx %jx", prev_object,
2161 (uintmax_t)next_pindex, (uintmax_t)next_size));
2162 prev_object->charge -= ptoa(prev_object->size -
2163 next_pindex);
2164 }
2165 #endif
2166 }
2167
2168 /*
2169 * Extend the object if necessary.
2170 */
2171 if (next_pindex + next_size > prev_object->size)
2172 prev_object->size = next_pindex + next_size;
2173
2174 VM_OBJECT_WUNLOCK(prev_object);
2175 return (TRUE);
2176 }
2177
2178 void
2179 vm_object_set_writeable_dirty(vm_object_t object)
2180 {
2181
2182 VM_OBJECT_ASSERT_WLOCKED(object);
2183 if (object->type != OBJT_VNODE) {
2184 if ((object->flags & OBJ_TMPFS_NODE) != 0) {
2185 KASSERT(object->type == OBJT_SWAP, ("non-swap tmpfs"));
2186 vm_object_set_flag(object, OBJ_TMPFS_DIRTY);
2187 }
2188 return;
2189 }
2190 object->generation++;
2191 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2192 return;
2193 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2194 }
2195
2196 /*
2197 * vm_object_unwire:
2198 *
2199 * For each page offset within the specified range of the given object,
2200 * find the highest-level page in the shadow chain and unwire it. A page
2201 * must exist at every page offset, and the highest-level page must be
2202 * wired.
2203 */
2204 void
2205 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2206 uint8_t queue)
2207 {
2208 vm_object_t tobject, t1object;
2209 vm_page_t m, tm;
2210 vm_pindex_t end_pindex, pindex, tpindex;
2211 int depth, locked_depth;
2212
2213 KASSERT((offset & PAGE_MASK) == 0,
2214 ("vm_object_unwire: offset is not page aligned"));
2215 KASSERT((length & PAGE_MASK) == 0,
2216 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2217 /* The wired count of a fictitious page never changes. */
2218 if ((object->flags & OBJ_FICTITIOUS) != 0)
2219 return;
2220 pindex = OFF_TO_IDX(offset);
2221 end_pindex = pindex + atop(length);
2222 again:
2223 locked_depth = 1;
2224 VM_OBJECT_RLOCK(object);
2225 m = vm_page_find_least(object, pindex);
2226 while (pindex < end_pindex) {
2227 if (m == NULL || pindex < m->pindex) {
2228 /*
2229 * The first object in the shadow chain doesn't
2230 * contain a page at the current index. Therefore,
2231 * the page must exist in a backing object.
2232 */
2233 tobject = object;
2234 tpindex = pindex;
2235 depth = 0;
2236 do {
2237 tpindex +=
2238 OFF_TO_IDX(tobject->backing_object_offset);
2239 tobject = tobject->backing_object;
2240 KASSERT(tobject != NULL,
2241 ("vm_object_unwire: missing page"));
2242 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2243 goto next_page;
2244 depth++;
2245 if (depth == locked_depth) {
2246 locked_depth++;
2247 VM_OBJECT_RLOCK(tobject);
2248 }
2249 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2250 NULL);
2251 } else {
2252 tm = m;
2253 m = TAILQ_NEXT(m, listq);
2254 }
2255 vm_page_lock(tm);
2256 if (vm_page_xbusied(tm)) {
2257 for (tobject = object; locked_depth >= 1;
2258 locked_depth--) {
2259 t1object = tobject->backing_object;
2260 VM_OBJECT_RUNLOCK(tobject);
2261 tobject = t1object;
2262 }
2263 vm_page_busy_sleep(tm, "unwbo", true);
2264 goto again;
2265 }
2266 vm_page_unwire(tm, queue);
2267 vm_page_unlock(tm);
2268 next_page:
2269 pindex++;
2270 }
2271 /* Release the accumulated object locks. */
2272 for (tobject = object; locked_depth >= 1; locked_depth--) {
2273 t1object = tobject->backing_object;
2274 VM_OBJECT_RUNLOCK(tobject);
2275 tobject = t1object;
2276 }
2277 }
2278
2279 struct vnode *
2280 vm_object_vnode(vm_object_t object)
2281 {
2282
2283 VM_OBJECT_ASSERT_LOCKED(object);
2284 if (object->type == OBJT_VNODE)
2285 return (object->handle);
2286 if (object->type == OBJT_SWAP && (object->flags & OBJ_TMPFS) != 0)
2287 return (object->un_pager.swp.swp_tmpfs);
2288 return (NULL);
2289 }
2290
2291 static int
2292 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2293 {
2294 struct kinfo_vmobject *kvo;
2295 char *fullpath, *freepath;
2296 struct vnode *vp;
2297 struct vattr va;
2298 vm_object_t obj;
2299 vm_page_t m;
2300 int count, error;
2301
2302 if (req->oldptr == NULL) {
2303 /*
2304 * If an old buffer has not been provided, generate an
2305 * estimate of the space needed for a subsequent call.
2306 */
2307 mtx_lock(&vm_object_list_mtx);
2308 count = 0;
2309 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2310 if (obj->type == OBJT_DEAD)
2311 continue;
2312 count++;
2313 }
2314 mtx_unlock(&vm_object_list_mtx);
2315 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2316 count * 11 / 10));
2317 }
2318
2319 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO);
2320 error = 0;
2321
2322 /*
2323 * VM objects are type stable and are never removed from the
2324 * list once added. This allows us to safely read obj->object_list
2325 * after reacquiring the VM object lock.
2326 */
2327 mtx_lock(&vm_object_list_mtx);
2328 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2329 if (obj->type == OBJT_DEAD)
2330 continue;
2331 VM_OBJECT_RLOCK(obj);
2332 if (obj->type == OBJT_DEAD) {
2333 VM_OBJECT_RUNLOCK(obj);
2334 continue;
2335 }
2336 mtx_unlock(&vm_object_list_mtx);
2337 kvo->kvo_size = ptoa(obj->size);
2338 kvo->kvo_resident = obj->resident_page_count;
2339 kvo->kvo_ref_count = obj->ref_count;
2340 kvo->kvo_shadow_count = obj->shadow_count;
2341 kvo->kvo_memattr = obj->memattr;
2342 kvo->kvo_active = 0;
2343 kvo->kvo_inactive = 0;
2344 TAILQ_FOREACH(m, &obj->memq, listq) {
2345 /*
2346 * A page may belong to the object but be
2347 * dequeued and set to PQ_NONE while the
2348 * object lock is not held. This makes the
2349 * reads of m->queue below racy, and we do not
2350 * count pages set to PQ_NONE. However, this
2351 * sysctl is only meant to give an
2352 * approximation of the system anyway.
2353 */
2354 if (m->queue == PQ_ACTIVE)
2355 kvo->kvo_active++;
2356 else if (m->queue == PQ_INACTIVE)
2357 kvo->kvo_inactive++;
2358 }
2359
2360 kvo->kvo_vn_fileid = 0;
2361 kvo->kvo_vn_fsid = 0;
2362 kvo->kvo_vn_fsid_freebsd11 = 0;
2363 freepath = NULL;
2364 fullpath = "";
2365 vp = NULL;
2366 switch (obj->type) {
2367 case OBJT_DEFAULT:
2368 kvo->kvo_type = KVME_TYPE_DEFAULT;
2369 break;
2370 case OBJT_VNODE:
2371 kvo->kvo_type = KVME_TYPE_VNODE;
2372 vp = obj->handle;
2373 vref(vp);
2374 break;
2375 case OBJT_SWAP:
2376 kvo->kvo_type = KVME_TYPE_SWAP;
2377 break;
2378 case OBJT_DEVICE:
2379 kvo->kvo_type = KVME_TYPE_DEVICE;
2380 break;
2381 case OBJT_PHYS:
2382 kvo->kvo_type = KVME_TYPE_PHYS;
2383 break;
2384 case OBJT_DEAD:
2385 kvo->kvo_type = KVME_TYPE_DEAD;
2386 break;
2387 case OBJT_SG:
2388 kvo->kvo_type = KVME_TYPE_SG;
2389 break;
2390 case OBJT_MGTDEVICE:
2391 kvo->kvo_type = KVME_TYPE_MGTDEVICE;
2392 break;
2393 default:
2394 kvo->kvo_type = KVME_TYPE_UNKNOWN;
2395 break;
2396 }
2397 VM_OBJECT_RUNLOCK(obj);
2398 if (vp != NULL) {
2399 vn_fullpath(curthread, vp, &fullpath, &freepath);
2400 vn_lock(vp, LK_SHARED | LK_RETRY);
2401 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2402 kvo->kvo_vn_fileid = va.va_fileid;
2403 kvo->kvo_vn_fsid = va.va_fsid;
2404 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2405 /* truncate */
2406 }
2407 vput(vp);
2408 }
2409
2410 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2411 if (freepath != NULL)
2412 free(freepath, M_TEMP);
2413
2414 /* Pack record size down */
2415 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2416 + strlen(kvo->kvo_path) + 1;
2417 kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2418 sizeof(uint64_t));
2419 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2420 mtx_lock(&vm_object_list_mtx);
2421 if (error)
2422 break;
2423 }
2424 mtx_unlock(&vm_object_list_mtx);
2425 free(kvo, M_TEMP);
2426 return (error);
2427 }
2428 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2429 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2430 "List of VM objects");
2431
2432 #include "opt_ddb.h"
2433 #ifdef DDB
2434 #include <sys/kernel.h>
2435
2436 #include <sys/cons.h>
2437
2438 #include <ddb/ddb.h>
2439
2440 static int
2441 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2442 {
2443 vm_map_t tmpm;
2444 vm_map_entry_t tmpe;
2445 vm_object_t obj;
2446 int entcount;
2447
2448 if (map == 0)
2449 return 0;
2450
2451 if (entry == 0) {
2452 tmpe = map->header.next;
2453 entcount = map->nentries;
2454 while (entcount-- && (tmpe != &map->header)) {
2455 if (_vm_object_in_map(map, object, tmpe)) {
2456 return 1;
2457 }
2458 tmpe = tmpe->next;
2459 }
2460 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2461 tmpm = entry->object.sub_map;
2462 tmpe = tmpm->header.next;
2463 entcount = tmpm->nentries;
2464 while (entcount-- && tmpe != &tmpm->header) {
2465 if (_vm_object_in_map(tmpm, object, tmpe)) {
2466 return 1;
2467 }
2468 tmpe = tmpe->next;
2469 }
2470 } else if ((obj = entry->object.vm_object) != NULL) {
2471 for (; obj; obj = obj->backing_object)
2472 if (obj == object) {
2473 return 1;
2474 }
2475 }
2476 return 0;
2477 }
2478
2479 static int
2480 vm_object_in_map(vm_object_t object)
2481 {
2482 struct proc *p;
2483
2484 /* sx_slock(&allproc_lock); */
2485 FOREACH_PROC_IN_SYSTEM(p) {
2486 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2487 continue;
2488 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2489 /* sx_sunlock(&allproc_lock); */
2490 return 1;
2491 }
2492 }
2493 /* sx_sunlock(&allproc_lock); */
2494 if (_vm_object_in_map(kernel_map, object, 0))
2495 return 1;
2496 return 0;
2497 }
2498
2499 DB_SHOW_COMMAND(vmochk, vm_object_check)
2500 {
2501 vm_object_t object;
2502
2503 /*
2504 * make sure that internal objs are in a map somewhere
2505 * and none have zero ref counts.
2506 */
2507 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2508 if (object->handle == NULL &&
2509 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2510 if (object->ref_count == 0) {
2511 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2512 (long)object->size);
2513 }
2514 if (!vm_object_in_map(object)) {
2515 db_printf(
2516 "vmochk: internal obj is not in a map: "
2517 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2518 object->ref_count, (u_long)object->size,
2519 (u_long)object->size,
2520 (void *)object->backing_object);
2521 }
2522 }
2523 }
2524 }
2525
2526 /*
2527 * vm_object_print: [ debug ]
2528 */
2529 DB_SHOW_COMMAND(object, vm_object_print_static)
2530 {
2531 /* XXX convert args. */
2532 vm_object_t object = (vm_object_t)addr;
2533 boolean_t full = have_addr;
2534
2535 vm_page_t p;
2536
2537 /* XXX count is an (unused) arg. Avoid shadowing it. */
2538 #define count was_count
2539
2540 int count;
2541
2542 if (object == NULL)
2543 return;
2544
2545 db_iprintf(
2546 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2547 object, (int)object->type, (uintmax_t)object->size,
2548 object->resident_page_count, object->ref_count, object->flags,
2549 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2550 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2551 object->shadow_count,
2552 object->backing_object ? object->backing_object->ref_count : 0,
2553 object->backing_object, (uintmax_t)object->backing_object_offset);
2554
2555 if (!full)
2556 return;
2557
2558 db_indent += 2;
2559 count = 0;
2560 TAILQ_FOREACH(p, &object->memq, listq) {
2561 if (count == 0)
2562 db_iprintf("memory:=");
2563 else if (count == 6) {
2564 db_printf("\n");
2565 db_iprintf(" ...");
2566 count = 0;
2567 } else
2568 db_printf(",");
2569 count++;
2570
2571 db_printf("(off=0x%jx,page=0x%jx)",
2572 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2573 }
2574 if (count != 0)
2575 db_printf("\n");
2576 db_indent -= 2;
2577 }
2578
2579 /* XXX. */
2580 #undef count
2581
2582 /* XXX need this non-static entry for calling from vm_map_print. */
2583 void
2584 vm_object_print(
2585 /* db_expr_t */ long addr,
2586 boolean_t have_addr,
2587 /* db_expr_t */ long count,
2588 char *modif)
2589 {
2590 vm_object_print_static(addr, have_addr, count, modif);
2591 }
2592
2593 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2594 {
2595 vm_object_t object;
2596 vm_pindex_t fidx;
2597 vm_paddr_t pa;
2598 vm_page_t m, prev_m;
2599 int rcount, nl, c;
2600
2601 nl = 0;
2602 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2603 db_printf("new object: %p\n", (void *)object);
2604 if (nl > 18) {
2605 c = cngetc();
2606 if (c != ' ')
2607 return;
2608 nl = 0;
2609 }
2610 nl++;
2611 rcount = 0;
2612 fidx = 0;
2613 pa = -1;
2614 TAILQ_FOREACH(m, &object->memq, listq) {
2615 if (m->pindex > 128)
2616 break;
2617 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2618 prev_m->pindex + 1 != m->pindex) {
2619 if (rcount) {
2620 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2621 (long)fidx, rcount, (long)pa);
2622 if (nl > 18) {
2623 c = cngetc();
2624 if (c != ' ')
2625 return;
2626 nl = 0;
2627 }
2628 nl++;
2629 rcount = 0;
2630 }
2631 }
2632 if (rcount &&
2633 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2634 ++rcount;
2635 continue;
2636 }
2637 if (rcount) {
2638 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2639 (long)fidx, rcount, (long)pa);
2640 if (nl > 18) {
2641 c = cngetc();
2642 if (c != ' ')
2643 return;
2644 nl = 0;
2645 }
2646 nl++;
2647 }
2648 fidx = m->pindex;
2649 pa = VM_PAGE_TO_PHYS(m);
2650 rcount = 1;
2651 }
2652 if (rcount) {
2653 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2654 (long)fidx, rcount, (long)pa);
2655 if (nl > 18) {
2656 c = cngetc();
2657 if (c != ' ')
2658 return;
2659 nl = 0;
2660 }
2661 nl++;
2662 }
2663 }
2664 }
2665 #endif /* DDB */
Cache object: af7fe769bd26ab36cc9849bddb27e35c
|