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