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