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: stable/9/sys/vm/vm_object.c 284100 2015-06-06 20:37:40Z jhb $");
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/vnode.h>
82 #include <sys/vmmeter.h>
83 #include <sys/sx.h>
84
85 #include <vm/vm.h>
86 #include <vm/vm_param.h>
87 #include <vm/pmap.h>
88 #include <vm/vm_map.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_page.h>
91 #include <vm/vm_pageout.h>
92 #include <vm/vm_pager.h>
93 #include <vm/swap_pager.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_reserv.h>
97 #include <vm/uma.h>
98
99 static int old_msync;
100 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
101 "Use old (insecure) msync behavior");
102
103 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
104 int pagerflags, int flags, boolean_t *clearobjflags,
105 boolean_t *eio);
106 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
107 boolean_t *clearobjflags);
108 static void vm_object_qcollapse(vm_object_t object);
109 static void vm_object_vndeallocate(vm_object_t object);
110
111 /*
112 * Virtual memory objects maintain the actual data
113 * associated with allocated virtual memory. A given
114 * page of memory exists within exactly one object.
115 *
116 * An object is only deallocated when all "references"
117 * are given up. Only one "reference" to a given
118 * region of an object should be writeable.
119 *
120 * Associated with each object is a list of all resident
121 * memory pages belonging to that object; this list is
122 * maintained by the "vm_page" module, and locked by the object's
123 * lock.
124 *
125 * Each object also records a "pager" routine which is
126 * used to retrieve (and store) pages to the proper backing
127 * storage. In addition, objects may be backed by other
128 * objects from which they were virtual-copied.
129 *
130 * The only items within the object structure which are
131 * modified after time of creation are:
132 * reference count locked by object's lock
133 * pager routine locked by object's lock
134 *
135 */
136
137 struct object_q vm_object_list;
138 struct mtx vm_object_list_mtx; /* lock for object list and count */
139
140 struct vm_object kernel_object_store;
141 struct vm_object kmem_object_store;
142
143 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD, 0,
144 "VM object stats");
145
146 static long object_collapses;
147 SYSCTL_LONG(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
148 &object_collapses, 0, "VM object collapses");
149
150 static long object_bypasses;
151 SYSCTL_LONG(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
152 &object_bypasses, 0, "VM object bypasses");
153
154 static uma_zone_t obj_zone;
155
156 static int vm_object_zinit(void *mem, int size, int flags);
157
158 #ifdef INVARIANTS
159 static void vm_object_zdtor(void *mem, int size, void *arg);
160
161 static void
vm_object_zdtor(void * mem,int size,void * arg)162 vm_object_zdtor(void *mem, int size, void *arg)
163 {
164 vm_object_t object;
165
166 object = (vm_object_t)mem;
167 KASSERT(object->ref_count == 0,
168 ("object %p ref_count = %d", object, object->ref_count));
169 KASSERT(TAILQ_EMPTY(&object->memq),
170 ("object %p has resident pages",
171 object));
172 #if VM_NRESERVLEVEL > 0
173 KASSERT(LIST_EMPTY(&object->rvq),
174 ("object %p has reservations",
175 object));
176 #endif
177 KASSERT(object->cache == NULL,
178 ("object %p has cached pages",
179 object));
180 KASSERT(object->paging_in_progress == 0,
181 ("object %p paging_in_progress = %d",
182 object, object->paging_in_progress));
183 KASSERT(object->resident_page_count == 0,
184 ("object %p resident_page_count = %d",
185 object, object->resident_page_count));
186 KASSERT(object->shadow_count == 0,
187 ("object %p shadow_count = %d",
188 object, object->shadow_count));
189 KASSERT(object->type == OBJT_DEAD,
190 ("object %p has non-dead type %d",
191 object, object->type));
192 }
193 #endif
194
195 static int
vm_object_zinit(void * mem,int size,int flags)196 vm_object_zinit(void *mem, int size, int flags)
197 {
198 vm_object_t object;
199
200 object = (vm_object_t)mem;
201 bzero(&object->mtx, sizeof(object->mtx));
202 VM_OBJECT_LOCK_INIT(object, "standard object");
203
204 /* These are true for any object that has been freed */
205 object->type = OBJT_DEAD;
206 object->ref_count = 0;
207 object->paging_in_progress = 0;
208 object->resident_page_count = 0;
209 object->shadow_count = 0;
210
211 mtx_lock(&vm_object_list_mtx);
212 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
213 mtx_unlock(&vm_object_list_mtx);
214 return (0);
215 }
216
217 void
_vm_object_allocate(objtype_t type,vm_pindex_t size,vm_object_t object)218 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
219 {
220
221 TAILQ_INIT(&object->memq);
222 LIST_INIT(&object->shadow_head);
223
224 object->root = NULL;
225 object->type = type;
226 object->size = size;
227 object->generation = 1;
228 object->ref_count = 1;
229 object->memattr = VM_MEMATTR_DEFAULT;
230 object->flags = 0;
231 object->cred = NULL;
232 object->charge = 0;
233 if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
234 object->flags = OBJ_ONEMAPPING;
235 object->pg_color = 0;
236 object->handle = NULL;
237 object->backing_object = NULL;
238 object->backing_object_offset = (vm_ooffset_t) 0;
239 #if VM_NRESERVLEVEL > 0
240 LIST_INIT(&object->rvq);
241 #endif
242 object->cache = NULL;
243 }
244
245 /*
246 * vm_object_init:
247 *
248 * Initialize the VM objects module.
249 */
250 void
vm_object_init(void)251 vm_object_init(void)
252 {
253 TAILQ_INIT(&vm_object_list);
254 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
255
256 VM_OBJECT_LOCK_INIT(kernel_object, "kernel object");
257 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
258 kernel_object);
259 #if VM_NRESERVLEVEL > 0
260 kernel_object->flags |= OBJ_COLORED;
261 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
262 #endif
263
264 VM_OBJECT_LOCK_INIT(kmem_object, "kmem object");
265 _vm_object_allocate(OBJT_PHYS, OFF_TO_IDX(VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS),
266 kmem_object);
267 #if VM_NRESERVLEVEL > 0
268 kmem_object->flags |= OBJ_COLORED;
269 kmem_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
270 #endif
271
272 /*
273 * The lock portion of struct vm_object must be type stable due
274 * to vm_pageout_fallback_object_lock locking a vm object
275 * without holding any references to it.
276 */
277 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
278 #ifdef INVARIANTS
279 vm_object_zdtor,
280 #else
281 NULL,
282 #endif
283 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM|UMA_ZONE_NOFREE);
284 }
285
286 void
vm_object_clear_flag(vm_object_t object,u_short bits)287 vm_object_clear_flag(vm_object_t object, u_short bits)
288 {
289
290 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
291 object->flags &= ~bits;
292 }
293
294 /*
295 * Sets the default memory attribute for the specified object. Pages
296 * that are allocated to this object are by default assigned this memory
297 * attribute.
298 *
299 * Presently, this function must be called before any pages are allocated
300 * to the object. In the future, this requirement may be relaxed for
301 * "default" and "swap" objects.
302 */
303 int
vm_object_set_memattr(vm_object_t object,vm_memattr_t memattr)304 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
305 {
306
307 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
308 switch (object->type) {
309 case OBJT_DEFAULT:
310 case OBJT_DEVICE:
311 case OBJT_PHYS:
312 case OBJT_SG:
313 case OBJT_SWAP:
314 case OBJT_VNODE:
315 if (!TAILQ_EMPTY(&object->memq))
316 return (KERN_FAILURE);
317 break;
318 case OBJT_DEAD:
319 return (KERN_INVALID_ARGUMENT);
320 }
321 object->memattr = memattr;
322 return (KERN_SUCCESS);
323 }
324
325 void
vm_object_pip_add(vm_object_t object,short i)326 vm_object_pip_add(vm_object_t object, short i)
327 {
328
329 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
330 object->paging_in_progress += i;
331 }
332
333 void
vm_object_pip_subtract(vm_object_t object,short i)334 vm_object_pip_subtract(vm_object_t object, short i)
335 {
336
337 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
338 object->paging_in_progress -= i;
339 }
340
341 void
vm_object_pip_wakeup(vm_object_t object)342 vm_object_pip_wakeup(vm_object_t object)
343 {
344
345 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
346 object->paging_in_progress--;
347 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
348 vm_object_clear_flag(object, OBJ_PIPWNT);
349 wakeup(object);
350 }
351 }
352
353 void
vm_object_pip_wakeupn(vm_object_t object,short i)354 vm_object_pip_wakeupn(vm_object_t object, short i)
355 {
356
357 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
358 if (i)
359 object->paging_in_progress -= i;
360 if ((object->flags & OBJ_PIPWNT) && object->paging_in_progress == 0) {
361 vm_object_clear_flag(object, OBJ_PIPWNT);
362 wakeup(object);
363 }
364 }
365
366 void
vm_object_pip_wait(vm_object_t object,char * waitid)367 vm_object_pip_wait(vm_object_t object, char *waitid)
368 {
369
370 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
371 while (object->paging_in_progress) {
372 object->flags |= OBJ_PIPWNT;
373 msleep(object, VM_OBJECT_MTX(object), PVM, waitid, 0);
374 }
375 }
376
377 /*
378 * vm_object_allocate:
379 *
380 * Returns a new object with the given size.
381 */
382 vm_object_t
vm_object_allocate(objtype_t type,vm_pindex_t size)383 vm_object_allocate(objtype_t type, vm_pindex_t size)
384 {
385 vm_object_t object;
386
387 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
388 _vm_object_allocate(type, size, object);
389 return (object);
390 }
391
392
393 /*
394 * vm_object_reference:
395 *
396 * Gets another reference to the given object. Note: OBJ_DEAD
397 * objects can be referenced during final cleaning.
398 */
399 void
vm_object_reference(vm_object_t object)400 vm_object_reference(vm_object_t object)
401 {
402 if (object == NULL)
403 return;
404 VM_OBJECT_LOCK(object);
405 vm_object_reference_locked(object);
406 VM_OBJECT_UNLOCK(object);
407 }
408
409 /*
410 * vm_object_reference_locked:
411 *
412 * Gets another reference to the given object.
413 *
414 * The object must be locked.
415 */
416 void
vm_object_reference_locked(vm_object_t object)417 vm_object_reference_locked(vm_object_t object)
418 {
419 struct vnode *vp;
420
421 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
422 object->ref_count++;
423 if (object->type == OBJT_VNODE) {
424 vp = object->handle;
425 vref(vp);
426 }
427 }
428
429 /*
430 * Handle deallocating an object of type OBJT_VNODE.
431 */
432 static void
vm_object_vndeallocate(vm_object_t object)433 vm_object_vndeallocate(vm_object_t object)
434 {
435 struct vnode *vp = (struct vnode *) object->handle;
436
437 VFS_ASSERT_GIANT(vp->v_mount);
438 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
439 KASSERT(object->type == OBJT_VNODE,
440 ("vm_object_vndeallocate: not a vnode object"));
441 KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
442 #ifdef INVARIANTS
443 if (object->ref_count == 0) {
444 vprint("vm_object_vndeallocate", vp);
445 panic("vm_object_vndeallocate: bad object reference count");
446 }
447 #endif
448
449 if (object->ref_count > 1) {
450 object->ref_count--;
451 VM_OBJECT_UNLOCK(object);
452 /* vrele may need the vnode lock. */
453 vrele(vp);
454 } else {
455 vhold(vp);
456 VM_OBJECT_UNLOCK(object);
457 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
458 vdrop(vp);
459 VM_OBJECT_LOCK(object);
460 object->ref_count--;
461 if (object->type == OBJT_DEAD) {
462 VM_OBJECT_UNLOCK(object);
463 VOP_UNLOCK(vp, 0);
464 } else {
465 if (object->ref_count == 0)
466 VOP_UNSET_TEXT(vp);
467 VM_OBJECT_UNLOCK(object);
468 vput(vp);
469 }
470 }
471 }
472
473 /*
474 * vm_object_deallocate:
475 *
476 * Release a reference to the specified object,
477 * gained either through a vm_object_allocate
478 * or a vm_object_reference call. When all references
479 * are gone, storage associated with this object
480 * may be relinquished.
481 *
482 * No object may be locked.
483 */
484 void
vm_object_deallocate(vm_object_t object)485 vm_object_deallocate(vm_object_t object)
486 {
487 vm_object_t temp;
488
489 while (object != NULL) {
490 int vfslocked;
491
492 vfslocked = 0;
493 restart:
494 VM_OBJECT_LOCK(object);
495 if (object->type == OBJT_VNODE) {
496 struct vnode *vp = (struct vnode *) object->handle;
497
498 /*
499 * Conditionally acquire Giant for a vnode-backed
500 * object. We have to be careful since the type of
501 * a vnode object can change while the object is
502 * unlocked.
503 */
504 if (VFS_NEEDSGIANT(vp->v_mount) && !vfslocked) {
505 vfslocked = 1;
506 if (!mtx_trylock(&Giant)) {
507 VM_OBJECT_UNLOCK(object);
508 mtx_lock(&Giant);
509 goto restart;
510 }
511 }
512 vm_object_vndeallocate(object);
513 VFS_UNLOCK_GIANT(vfslocked);
514 return;
515 } else
516 /*
517 * This is to handle the case that the object
518 * changed type while we dropped its lock to
519 * obtain Giant.
520 */
521 VFS_UNLOCK_GIANT(vfslocked);
522
523 KASSERT(object->ref_count != 0,
524 ("vm_object_deallocate: object deallocated too many times: %d", object->type));
525
526 /*
527 * If the reference count goes to 0 we start calling
528 * vm_object_terminate() on the object chain.
529 * A ref count of 1 may be a special case depending on the
530 * shadow count being 0 or 1.
531 */
532 object->ref_count--;
533 if (object->ref_count > 1) {
534 VM_OBJECT_UNLOCK(object);
535 return;
536 } else if (object->ref_count == 1) {
537 if (object->shadow_count == 0 &&
538 object->handle == NULL &&
539 (object->type == OBJT_DEFAULT ||
540 object->type == OBJT_SWAP)) {
541 vm_object_set_flag(object, OBJ_ONEMAPPING);
542 } else if ((object->shadow_count == 1) &&
543 (object->handle == NULL) &&
544 (object->type == OBJT_DEFAULT ||
545 object->type == OBJT_SWAP)) {
546 vm_object_t robject;
547
548 robject = LIST_FIRST(&object->shadow_head);
549 KASSERT(robject != NULL,
550 ("vm_object_deallocate: ref_count: %d, shadow_count: %d",
551 object->ref_count,
552 object->shadow_count));
553 if (!VM_OBJECT_TRYLOCK(robject)) {
554 /*
555 * Avoid a potential deadlock.
556 */
557 object->ref_count++;
558 VM_OBJECT_UNLOCK(object);
559 /*
560 * More likely than not the thread
561 * holding robject's lock has lower
562 * priority than the current thread.
563 * Let the lower priority thread run.
564 */
565 pause("vmo_de", 1);
566 continue;
567 }
568 /*
569 * Collapse object into its shadow unless its
570 * shadow is dead. In that case, object will
571 * be deallocated by the thread that is
572 * deallocating its shadow.
573 */
574 if ((robject->flags & OBJ_DEAD) == 0 &&
575 (robject->handle == NULL) &&
576 (robject->type == OBJT_DEFAULT ||
577 robject->type == OBJT_SWAP)) {
578
579 robject->ref_count++;
580 retry:
581 if (robject->paging_in_progress) {
582 VM_OBJECT_UNLOCK(object);
583 vm_object_pip_wait(robject,
584 "objde1");
585 temp = robject->backing_object;
586 if (object == temp) {
587 VM_OBJECT_LOCK(object);
588 goto retry;
589 }
590 } else if (object->paging_in_progress) {
591 VM_OBJECT_UNLOCK(robject);
592 object->flags |= OBJ_PIPWNT;
593 msleep(object,
594 VM_OBJECT_MTX(object),
595 PDROP | PVM, "objde2", 0);
596 VM_OBJECT_LOCK(robject);
597 temp = robject->backing_object;
598 if (object == temp) {
599 VM_OBJECT_LOCK(object);
600 goto retry;
601 }
602 } else
603 VM_OBJECT_UNLOCK(object);
604
605 if (robject->ref_count == 1) {
606 robject->ref_count--;
607 object = robject;
608 goto doterm;
609 }
610 object = robject;
611 vm_object_collapse(object);
612 VM_OBJECT_UNLOCK(object);
613 continue;
614 }
615 VM_OBJECT_UNLOCK(robject);
616 }
617 VM_OBJECT_UNLOCK(object);
618 return;
619 }
620 doterm:
621 temp = object->backing_object;
622 if (temp != NULL) {
623 VM_OBJECT_LOCK(temp);
624 LIST_REMOVE(object, shadow_list);
625 temp->shadow_count--;
626 VM_OBJECT_UNLOCK(temp);
627 object->backing_object = NULL;
628 }
629 /*
630 * Don't double-terminate, we could be in a termination
631 * recursion due to the terminate having to sync data
632 * to disk.
633 */
634 if ((object->flags & OBJ_DEAD) == 0)
635 vm_object_terminate(object);
636 else
637 VM_OBJECT_UNLOCK(object);
638 object = temp;
639 }
640 }
641
642 /*
643 * vm_object_destroy removes the object from the global object list
644 * and frees the space for the object.
645 */
646 void
vm_object_destroy(vm_object_t object)647 vm_object_destroy(vm_object_t object)
648 {
649
650 /*
651 * Release the allocation charge.
652 */
653 if (object->cred != NULL) {
654 swap_release_by_cred(object->charge, object->cred);
655 object->charge = 0;
656 crfree(object->cred);
657 object->cred = NULL;
658 }
659
660 /*
661 * Free the space for the object.
662 */
663 uma_zfree(obj_zone, object);
664 }
665
666 /*
667 * vm_object_terminate actually destroys the specified object, freeing
668 * up all previously used resources.
669 *
670 * The object must be locked.
671 * This routine may block.
672 */
673 void
vm_object_terminate(vm_object_t object)674 vm_object_terminate(vm_object_t object)
675 {
676 vm_page_t p, p_next;
677
678 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
679
680 /*
681 * Make sure no one uses us.
682 */
683 vm_object_set_flag(object, OBJ_DEAD);
684
685 /*
686 * wait for the pageout daemon to be done with the object
687 */
688 vm_object_pip_wait(object, "objtrm");
689
690 KASSERT(!object->paging_in_progress,
691 ("vm_object_terminate: pageout in progress"));
692
693 /*
694 * Clean and free the pages, as appropriate. All references to the
695 * object are gone, so we don't need to lock it.
696 */
697 if (object->type == OBJT_VNODE) {
698 struct vnode *vp = (struct vnode *)object->handle;
699
700 /*
701 * Clean pages and flush buffers.
702 */
703 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
704 VM_OBJECT_UNLOCK(object);
705
706 vinvalbuf(vp, V_SAVE, 0, 0);
707
708 VM_OBJECT_LOCK(object);
709 }
710
711 KASSERT(object->ref_count == 0,
712 ("vm_object_terminate: object with references, ref_count=%d",
713 object->ref_count));
714
715 /*
716 * Free any remaining pageable pages. This also removes them from the
717 * paging queues. However, don't free wired pages, just remove them
718 * from the object. Rather than incrementally removing each page from
719 * the object, the page and object are reset to any empty state.
720 */
721 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
722 KASSERT(!p->busy && (p->oflags & VPO_BUSY) == 0,
723 ("vm_object_terminate: freeing busy page %p", p));
724 vm_page_lock(p);
725 /*
726 * Optimize the page's removal from the object by resetting
727 * its "object" field. Specifically, if the page is not
728 * wired, then the effect of this assignment is that
729 * vm_page_free()'s call to vm_page_remove() will return
730 * immediately without modifying the page or the object.
731 */
732 p->object = NULL;
733 if (p->wire_count == 0) {
734 vm_page_free(p);
735 PCPU_INC(cnt.v_pfree);
736 }
737 vm_page_unlock(p);
738 }
739 /*
740 * If the object contained any pages, then reset it to an empty state.
741 * None of the object's fields, including "resident_page_count", were
742 * modified by the preceding loop.
743 */
744 if (object->resident_page_count != 0) {
745 object->root = NULL;
746 TAILQ_INIT(&object->memq);
747 object->resident_page_count = 0;
748 if (object->type == OBJT_VNODE)
749 vdrop(object->handle);
750 }
751
752 #if VM_NRESERVLEVEL > 0
753 if (__predict_false(!LIST_EMPTY(&object->rvq)))
754 vm_reserv_break_all(object);
755 #endif
756 if (__predict_false(object->cache != NULL))
757 vm_page_cache_free(object, 0, 0);
758
759 KASSERT(object->cred == NULL || object->type == OBJT_DEFAULT ||
760 object->type == OBJT_SWAP,
761 ("%s: non-swap obj %p has cred", __func__, object));
762
763 /*
764 * Let the pager know object is dead.
765 */
766 vm_pager_deallocate(object);
767 VM_OBJECT_UNLOCK(object);
768
769 vm_object_destroy(object);
770 }
771
772 /*
773 * Make the page read-only so that we can clear the object flags. However, if
774 * this is a nosync mmap then the object is likely to stay dirty so do not
775 * mess with the page and do not clear the object flags. Returns TRUE if the
776 * page should be flushed, and FALSE otherwise.
777 */
778 static boolean_t
vm_object_page_remove_write(vm_page_t p,int flags,boolean_t * clearobjflags)779 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *clearobjflags)
780 {
781
782 /*
783 * If we have been asked to skip nosync pages and this is a
784 * nosync page, skip it. Note that the object flags were not
785 * cleared in this case so we do not have to set them.
786 */
787 if ((flags & OBJPC_NOSYNC) != 0 && (p->oflags & VPO_NOSYNC) != 0) {
788 *clearobjflags = FALSE;
789 return (FALSE);
790 } else {
791 pmap_remove_write(p);
792 return (p->dirty != 0);
793 }
794 }
795
796 /*
797 * vm_object_page_clean
798 *
799 * Clean all dirty pages in the specified range of object. Leaves page
800 * on whatever queue it is currently on. If NOSYNC is set then do not
801 * write out pages with VPO_NOSYNC set (originally comes from MAP_NOSYNC),
802 * leaving the object dirty.
803 *
804 * When stuffing pages asynchronously, allow clustering. XXX we need a
805 * synchronous clustering mode implementation.
806 *
807 * Odd semantics: if start == end, we clean everything.
808 *
809 * The object must be locked.
810 *
811 * Returns FALSE if some page from the range was not written, as
812 * reported by the pager, and TRUE otherwise.
813 */
814 boolean_t
vm_object_page_clean(vm_object_t object,vm_ooffset_t start,vm_ooffset_t end,int flags)815 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
816 int flags)
817 {
818 vm_page_t np, p;
819 vm_pindex_t pi, tend, tstart;
820 int curgeneration, n, pagerflags;
821 boolean_t clearobjflags, eio, res;
822
823 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
824 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
825 KASSERT(object->type == OBJT_VNODE, ("Not a vnode object"));
826 if ((object->flags & OBJ_MIGHTBEDIRTY) == 0 ||
827 object->resident_page_count == 0)
828 return (TRUE);
829
830 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
831 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
832 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
833
834 tstart = OFF_TO_IDX(start);
835 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
836 clearobjflags = tstart == 0 && tend >= object->size;
837 res = TRUE;
838
839 rescan:
840 curgeneration = object->generation;
841
842 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
843 pi = p->pindex;
844 if (pi >= tend)
845 break;
846 np = TAILQ_NEXT(p, listq);
847 if (p->valid == 0)
848 continue;
849 if (vm_page_sleep_if_busy(p, TRUE, "vpcwai")) {
850 if (object->generation != curgeneration) {
851 if ((flags & OBJPC_SYNC) != 0)
852 goto rescan;
853 else
854 clearobjflags = FALSE;
855 }
856 np = vm_page_find_least(object, pi);
857 continue;
858 }
859 if (!vm_object_page_remove_write(p, flags, &clearobjflags))
860 continue;
861
862 n = vm_object_page_collect_flush(object, p, pagerflags,
863 flags, &clearobjflags, &eio);
864 if (eio) {
865 res = FALSE;
866 clearobjflags = FALSE;
867 }
868 if (object->generation != curgeneration) {
869 if ((flags & OBJPC_SYNC) != 0)
870 goto rescan;
871 else
872 clearobjflags = FALSE;
873 }
874
875 /*
876 * If the VOP_PUTPAGES() did a truncated write, so
877 * that even the first page of the run is not fully
878 * written, vm_pageout_flush() returns 0 as the run
879 * length. Since the condition that caused truncated
880 * write may be permanent, e.g. exhausted free space,
881 * accepting n == 0 would cause an infinite loop.
882 *
883 * Forwarding the iterator leaves the unwritten page
884 * behind, but there is not much we can do there if
885 * filesystem refuses to write it.
886 */
887 if (n == 0) {
888 n = 1;
889 clearobjflags = FALSE;
890 }
891 np = vm_page_find_least(object, pi + n);
892 }
893 #if 0
894 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
895 #endif
896
897 if (clearobjflags)
898 vm_object_clear_flag(object, OBJ_MIGHTBEDIRTY);
899 return (res);
900 }
901
902 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)903 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
904 int flags, boolean_t *clearobjflags, boolean_t *eio)
905 {
906 vm_page_t ma[vm_pageout_page_count], p_first, tp;
907 int count, i, mreq, runlen;
908
909 mtx_assert(&vm_page_queue_mtx, MA_NOTOWNED);
910 vm_page_lock_assert(p, MA_NOTOWNED);
911 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
912
913 count = 1;
914 mreq = 0;
915
916 for (tp = p; count < vm_pageout_page_count; count++) {
917 tp = vm_page_next(tp);
918 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
919 break;
920 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
921 break;
922 }
923
924 for (p_first = p; count < vm_pageout_page_count; count++) {
925 tp = vm_page_prev(p_first);
926 if (tp == NULL || tp->busy != 0 || (tp->oflags & VPO_BUSY) != 0)
927 break;
928 if (!vm_object_page_remove_write(tp, flags, clearobjflags))
929 break;
930 p_first = tp;
931 mreq++;
932 }
933
934 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
935 ma[i] = tp;
936
937 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
938 return (runlen);
939 }
940
941 /*
942 * Note that there is absolutely no sense in writing out
943 * anonymous objects, so we track down the vnode object
944 * to write out.
945 * We invalidate (remove) all pages from the address space
946 * for semantic correctness.
947 *
948 * If the backing object is a device object with unmanaged pages, then any
949 * mappings to the specified range of pages must be removed before this
950 * function is called.
951 *
952 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
953 * may start out with a NULL object.
954 */
955 boolean_t
vm_object_sync(vm_object_t object,vm_ooffset_t offset,vm_size_t size,boolean_t syncio,boolean_t invalidate)956 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
957 boolean_t syncio, boolean_t invalidate)
958 {
959 vm_object_t backing_object;
960 struct vnode *vp;
961 struct mount *mp;
962 int error, flags, fsync_after;
963 boolean_t res;
964
965 if (object == NULL)
966 return (TRUE);
967 res = TRUE;
968 error = 0;
969 VM_OBJECT_LOCK(object);
970 while ((backing_object = object->backing_object) != NULL) {
971 VM_OBJECT_LOCK(backing_object);
972 offset += object->backing_object_offset;
973 VM_OBJECT_UNLOCK(object);
974 object = backing_object;
975 if (object->size < OFF_TO_IDX(offset + size))
976 size = IDX_TO_OFF(object->size) - offset;
977 }
978 /*
979 * Flush pages if writing is allowed, invalidate them
980 * if invalidation requested. Pages undergoing I/O
981 * will be ignored by vm_object_page_remove().
982 *
983 * We cannot lock the vnode and then wait for paging
984 * to complete without deadlocking against vm_fault.
985 * Instead we simply call vm_object_page_remove() and
986 * allow it to block internally on a page-by-page
987 * basis when it encounters pages undergoing async
988 * I/O.
989 */
990 if (object->type == OBJT_VNODE &&
991 (object->flags & OBJ_MIGHTBEDIRTY) != 0) {
992 int vfslocked;
993 vp = object->handle;
994 VM_OBJECT_UNLOCK(object);
995 (void) vn_start_write(vp, &mp, V_WAIT);
996 vfslocked = VFS_LOCK_GIANT(vp->v_mount);
997 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
998 if (syncio && !invalidate && offset == 0 &&
999 OFF_TO_IDX(size) == object->size) {
1000 /*
1001 * If syncing the whole mapping of the file,
1002 * it is faster to schedule all the writes in
1003 * async mode, also allowing the clustering,
1004 * and then wait for i/o to complete.
1005 */
1006 flags = 0;
1007 fsync_after = TRUE;
1008 } else {
1009 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1010 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1011 fsync_after = FALSE;
1012 }
1013 VM_OBJECT_LOCK(object);
1014 res = vm_object_page_clean(object, offset, offset + size,
1015 flags);
1016 VM_OBJECT_UNLOCK(object);
1017 if (fsync_after)
1018 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1019 VOP_UNLOCK(vp, 0);
1020 VFS_UNLOCK_GIANT(vfslocked);
1021 vn_finished_write(mp);
1022 if (error != 0)
1023 res = FALSE;
1024 VM_OBJECT_LOCK(object);
1025 }
1026 if ((object->type == OBJT_VNODE ||
1027 object->type == OBJT_DEVICE) && invalidate) {
1028 if (object->type == OBJT_DEVICE)
1029 /*
1030 * The option OBJPR_NOTMAPPED must be passed here
1031 * because vm_object_page_remove() cannot remove
1032 * unmanaged mappings.
1033 */
1034 flags = OBJPR_NOTMAPPED;
1035 else if (old_msync)
1036 flags = OBJPR_NOTWIRED;
1037 else
1038 flags = OBJPR_CLEANONLY | OBJPR_NOTWIRED;
1039 vm_object_page_remove(object, OFF_TO_IDX(offset),
1040 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1041 }
1042 VM_OBJECT_UNLOCK(object);
1043 return (res);
1044 }
1045
1046 /*
1047 * vm_object_madvise:
1048 *
1049 * Implements the madvise function at the object/page level.
1050 *
1051 * MADV_WILLNEED (any object)
1052 *
1053 * Activate the specified pages if they are resident.
1054 *
1055 * MADV_DONTNEED (any object)
1056 *
1057 * Deactivate the specified pages if they are resident.
1058 *
1059 * MADV_FREE (OBJT_DEFAULT/OBJT_SWAP objects,
1060 * OBJ_ONEMAPPING only)
1061 *
1062 * Deactivate and clean the specified pages if they are
1063 * resident. This permits the process to reuse the pages
1064 * without faulting or the kernel to reclaim the pages
1065 * without I/O.
1066 */
1067 void
vm_object_madvise(vm_object_t object,vm_pindex_t pindex,vm_pindex_t end,int advise)1068 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1069 int advise)
1070 {
1071 vm_pindex_t tpindex;
1072 vm_object_t backing_object, tobject;
1073 vm_page_t m;
1074
1075 if (object == NULL)
1076 return;
1077 VM_OBJECT_LOCK(object);
1078 /*
1079 * Locate and adjust resident pages
1080 */
1081 for (; pindex < end; pindex += 1) {
1082 relookup:
1083 tobject = object;
1084 tpindex = pindex;
1085 shadowlookup:
1086 /*
1087 * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1088 * and those pages must be OBJ_ONEMAPPING.
1089 */
1090 if (advise == MADV_FREE) {
1091 if ((tobject->type != OBJT_DEFAULT &&
1092 tobject->type != OBJT_SWAP) ||
1093 (tobject->flags & OBJ_ONEMAPPING) == 0) {
1094 goto unlock_tobject;
1095 }
1096 } else if (tobject->type == OBJT_PHYS)
1097 goto unlock_tobject;
1098 m = vm_page_lookup(tobject, tpindex);
1099 if (m == NULL && advise == MADV_WILLNEED) {
1100 /*
1101 * If the page is cached, reactivate it.
1102 */
1103 m = vm_page_alloc(tobject, tpindex, VM_ALLOC_IFCACHED |
1104 VM_ALLOC_NOBUSY);
1105 }
1106 if (m == NULL) {
1107 /*
1108 * There may be swap even if there is no backing page
1109 */
1110 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1111 swap_pager_freespace(tobject, tpindex, 1);
1112 /*
1113 * next object
1114 */
1115 backing_object = tobject->backing_object;
1116 if (backing_object == NULL)
1117 goto unlock_tobject;
1118 VM_OBJECT_LOCK(backing_object);
1119 tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1120 if (tobject != object)
1121 VM_OBJECT_UNLOCK(tobject);
1122 tobject = backing_object;
1123 goto shadowlookup;
1124 } else if (m->valid != VM_PAGE_BITS_ALL)
1125 goto unlock_tobject;
1126 /*
1127 * If the page is not in a normal state, skip it.
1128 */
1129 vm_page_lock(m);
1130 if (m->hold_count != 0 || m->wire_count != 0) {
1131 vm_page_unlock(m);
1132 goto unlock_tobject;
1133 }
1134 KASSERT((m->flags & PG_FICTITIOUS) == 0,
1135 ("vm_object_madvise: page %p is fictitious", m));
1136 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1137 ("vm_object_madvise: page %p is not managed", m));
1138 if ((m->oflags & VPO_BUSY) || m->busy) {
1139 if (advise == MADV_WILLNEED) {
1140 /*
1141 * Reference the page before unlocking and
1142 * sleeping so that the page daemon is less
1143 * likely to reclaim it.
1144 */
1145 vm_page_aflag_set(m, PGA_REFERENCED);
1146 }
1147 vm_page_unlock(m);
1148 if (object != tobject)
1149 VM_OBJECT_UNLOCK(object);
1150 m->oflags |= VPO_WANTED;
1151 msleep(m, VM_OBJECT_MTX(tobject), PDROP | PVM, "madvpo",
1152 0);
1153 VM_OBJECT_LOCK(object);
1154 goto relookup;
1155 }
1156 if (advise == MADV_WILLNEED) {
1157 vm_page_activate(m);
1158 } else if (advise == MADV_DONTNEED) {
1159 vm_page_dontneed(m);
1160 } else if (advise == MADV_FREE) {
1161 /*
1162 * Mark the page clean. This will allow the page
1163 * to be freed up by the system. However, such pages
1164 * are often reused quickly by malloc()/free()
1165 * so we do not do anything that would cause
1166 * a page fault if we can help it.
1167 *
1168 * Specifically, we do not try to actually free
1169 * the page now nor do we try to put it in the
1170 * cache (which would cause a page fault on reuse).
1171 *
1172 * But we do make the page is freeable as we
1173 * can without actually taking the step of unmapping
1174 * it.
1175 */
1176 pmap_clear_modify(m);
1177 m->dirty = 0;
1178 m->act_count = 0;
1179 vm_page_dontneed(m);
1180 }
1181 vm_page_unlock(m);
1182 if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1183 swap_pager_freespace(tobject, tpindex, 1);
1184 unlock_tobject:
1185 if (tobject != object)
1186 VM_OBJECT_UNLOCK(tobject);
1187 }
1188 VM_OBJECT_UNLOCK(object);
1189 }
1190
1191 /*
1192 * vm_object_shadow:
1193 *
1194 * Create a new object which is backed by the
1195 * specified existing object range. The source
1196 * object reference is deallocated.
1197 *
1198 * The new object and offset into that object
1199 * are returned in the source parameters.
1200 */
1201 void
vm_object_shadow(vm_object_t * object,vm_ooffset_t * offset,vm_size_t length)1202 vm_object_shadow(
1203 vm_object_t *object, /* IN/OUT */
1204 vm_ooffset_t *offset, /* IN/OUT */
1205 vm_size_t length)
1206 {
1207 vm_object_t source;
1208 vm_object_t result;
1209
1210 source = *object;
1211
1212 /*
1213 * Don't create the new object if the old object isn't shared.
1214 */
1215 if (source != NULL) {
1216 VM_OBJECT_LOCK(source);
1217 if (source->ref_count == 1 &&
1218 source->handle == NULL &&
1219 (source->type == OBJT_DEFAULT ||
1220 source->type == OBJT_SWAP)) {
1221 VM_OBJECT_UNLOCK(source);
1222 return;
1223 }
1224 VM_OBJECT_UNLOCK(source);
1225 }
1226
1227 /*
1228 * Allocate a new object with the given length.
1229 */
1230 result = vm_object_allocate(OBJT_DEFAULT, atop(length));
1231
1232 /*
1233 * The new object shadows the source object, adding a reference to it.
1234 * Our caller changes his reference to point to the new object,
1235 * removing a reference to the source object. Net result: no change
1236 * of reference count.
1237 *
1238 * Try to optimize the result object's page color when shadowing
1239 * in order to maintain page coloring consistency in the combined
1240 * shadowed object.
1241 */
1242 result->backing_object = source;
1243 /*
1244 * Store the offset into the source object, and fix up the offset into
1245 * the new object.
1246 */
1247 result->backing_object_offset = *offset;
1248 if (source != NULL) {
1249 VM_OBJECT_LOCK(source);
1250 LIST_INSERT_HEAD(&source->shadow_head, result, shadow_list);
1251 source->shadow_count++;
1252 #if VM_NRESERVLEVEL > 0
1253 result->flags |= source->flags & OBJ_COLORED;
1254 result->pg_color = (source->pg_color + OFF_TO_IDX(*offset)) &
1255 ((1 << (VM_NFREEORDER - 1)) - 1);
1256 #endif
1257 VM_OBJECT_UNLOCK(source);
1258 }
1259
1260
1261 /*
1262 * Return the new things
1263 */
1264 *offset = 0;
1265 *object = result;
1266 }
1267
1268 /*
1269 * vm_object_split:
1270 *
1271 * Split the pages in a map entry into a new object. This affords
1272 * easier removal of unused pages, and keeps object inheritance from
1273 * being a negative impact on memory usage.
1274 */
1275 void
vm_object_split(vm_map_entry_t entry)1276 vm_object_split(vm_map_entry_t entry)
1277 {
1278 vm_page_t m, m_next;
1279 vm_object_t orig_object, new_object, source;
1280 vm_pindex_t idx, offidxstart;
1281 vm_size_t size;
1282
1283 orig_object = entry->object.vm_object;
1284 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP)
1285 return;
1286 if (orig_object->ref_count <= 1)
1287 return;
1288 VM_OBJECT_UNLOCK(orig_object);
1289
1290 offidxstart = OFF_TO_IDX(entry->offset);
1291 size = atop(entry->end - entry->start);
1292
1293 /*
1294 * If swap_pager_copy() is later called, it will convert new_object
1295 * into a swap object.
1296 */
1297 new_object = vm_object_allocate(OBJT_DEFAULT, size);
1298
1299 /*
1300 * At this point, the new object is still private, so the order in
1301 * which the original and new objects are locked does not matter.
1302 */
1303 VM_OBJECT_LOCK(new_object);
1304 VM_OBJECT_LOCK(orig_object);
1305 source = orig_object->backing_object;
1306 if (source != NULL) {
1307 VM_OBJECT_LOCK(source);
1308 if ((source->flags & OBJ_DEAD) != 0) {
1309 VM_OBJECT_UNLOCK(source);
1310 VM_OBJECT_UNLOCK(orig_object);
1311 VM_OBJECT_UNLOCK(new_object);
1312 vm_object_deallocate(new_object);
1313 VM_OBJECT_LOCK(orig_object);
1314 return;
1315 }
1316 LIST_INSERT_HEAD(&source->shadow_head,
1317 new_object, shadow_list);
1318 source->shadow_count++;
1319 vm_object_reference_locked(source); /* for new_object */
1320 vm_object_clear_flag(source, OBJ_ONEMAPPING);
1321 VM_OBJECT_UNLOCK(source);
1322 new_object->backing_object_offset =
1323 orig_object->backing_object_offset + entry->offset;
1324 new_object->backing_object = source;
1325 }
1326 if (orig_object->cred != NULL) {
1327 new_object->cred = orig_object->cred;
1328 crhold(orig_object->cred);
1329 new_object->charge = ptoa(size);
1330 KASSERT(orig_object->charge >= ptoa(size),
1331 ("orig_object->charge < 0"));
1332 orig_object->charge -= ptoa(size);
1333 }
1334 retry:
1335 m = vm_page_find_least(orig_object, offidxstart);
1336 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1337 m = m_next) {
1338 m_next = TAILQ_NEXT(m, listq);
1339
1340 /*
1341 * We must wait for pending I/O to complete before we can
1342 * rename the page.
1343 *
1344 * We do not have to VM_PROT_NONE the page as mappings should
1345 * not be changed by this operation.
1346 */
1347 if ((m->oflags & VPO_BUSY) || m->busy) {
1348 VM_OBJECT_UNLOCK(new_object);
1349 m->oflags |= VPO_WANTED;
1350 msleep(m, VM_OBJECT_MTX(orig_object), PVM, "spltwt", 0);
1351 VM_OBJECT_LOCK(new_object);
1352 goto retry;
1353 }
1354 vm_page_lock(m);
1355 vm_page_rename(m, new_object, idx);
1356 vm_page_unlock(m);
1357 /* page automatically made dirty by rename and cache handled */
1358 vm_page_busy(m);
1359 }
1360 if (orig_object->type == OBJT_SWAP) {
1361 /*
1362 * swap_pager_copy() can sleep, in which case the orig_object's
1363 * and new_object's locks are released and reacquired.
1364 */
1365 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1366
1367 /*
1368 * Transfer any cached pages from orig_object to new_object.
1369 */
1370 if (__predict_false(orig_object->cache != NULL))
1371 vm_page_cache_transfer(orig_object, offidxstart,
1372 new_object);
1373 }
1374 VM_OBJECT_UNLOCK(orig_object);
1375 TAILQ_FOREACH(m, &new_object->memq, listq)
1376 vm_page_wakeup(m);
1377 VM_OBJECT_UNLOCK(new_object);
1378 entry->object.vm_object = new_object;
1379 entry->offset = 0LL;
1380 vm_object_deallocate(orig_object);
1381 VM_OBJECT_LOCK(new_object);
1382 }
1383
1384 #define OBSC_TEST_ALL_SHADOWED 0x0001
1385 #define OBSC_COLLAPSE_NOWAIT 0x0002
1386 #define OBSC_COLLAPSE_WAIT 0x0004
1387
1388 static int
vm_object_backing_scan(vm_object_t object,int op)1389 vm_object_backing_scan(vm_object_t object, int op)
1390 {
1391 int r = 1;
1392 vm_page_t p;
1393 vm_object_t backing_object;
1394 vm_pindex_t backing_offset_index;
1395
1396 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1397 VM_OBJECT_LOCK_ASSERT(object->backing_object, MA_OWNED);
1398
1399 backing_object = object->backing_object;
1400 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1401
1402 /*
1403 * Initial conditions
1404 */
1405 if (op & OBSC_TEST_ALL_SHADOWED) {
1406 /*
1407 * We do not want to have to test for the existence of cache
1408 * or swap pages in the backing object. XXX but with the
1409 * new swapper this would be pretty easy to do.
1410 *
1411 * XXX what about anonymous MAP_SHARED memory that hasn't
1412 * been ZFOD faulted yet? If we do not test for this, the
1413 * shadow test may succeed! XXX
1414 */
1415 if (backing_object->type != OBJT_DEFAULT) {
1416 return (0);
1417 }
1418 }
1419 if (op & OBSC_COLLAPSE_WAIT) {
1420 vm_object_set_flag(backing_object, OBJ_DEAD);
1421 }
1422
1423 /*
1424 * Our scan
1425 */
1426 p = TAILQ_FIRST(&backing_object->memq);
1427 while (p) {
1428 vm_page_t next = TAILQ_NEXT(p, listq);
1429 vm_pindex_t new_pindex = p->pindex - backing_offset_index;
1430
1431 if (op & OBSC_TEST_ALL_SHADOWED) {
1432 vm_page_t pp;
1433
1434 /*
1435 * Ignore pages outside the parent object's range
1436 * and outside the parent object's mapping of the
1437 * backing object.
1438 *
1439 * note that we do not busy the backing object's
1440 * page.
1441 */
1442 if (
1443 p->pindex < backing_offset_index ||
1444 new_pindex >= object->size
1445 ) {
1446 p = next;
1447 continue;
1448 }
1449
1450 /*
1451 * See if the parent has the page or if the parent's
1452 * object pager has the page. If the parent has the
1453 * page but the page is not valid, the parent's
1454 * object pager must have the page.
1455 *
1456 * If this fails, the parent does not completely shadow
1457 * the object and we might as well give up now.
1458 */
1459
1460 pp = vm_page_lookup(object, new_pindex);
1461 if (
1462 (pp == NULL || pp->valid == 0) &&
1463 !vm_pager_has_page(object, new_pindex, NULL, NULL)
1464 ) {
1465 r = 0;
1466 break;
1467 }
1468 }
1469
1470 /*
1471 * Check for busy page
1472 */
1473 if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
1474 vm_page_t pp;
1475
1476 if (op & OBSC_COLLAPSE_NOWAIT) {
1477 if ((p->oflags & VPO_BUSY) ||
1478 !p->valid ||
1479 p->busy) {
1480 p = next;
1481 continue;
1482 }
1483 } else if (op & OBSC_COLLAPSE_WAIT) {
1484 if ((p->oflags & VPO_BUSY) || p->busy) {
1485 VM_OBJECT_UNLOCK(object);
1486 p->oflags |= VPO_WANTED;
1487 msleep(p, VM_OBJECT_MTX(backing_object),
1488 PDROP | PVM, "vmocol", 0);
1489 VM_OBJECT_LOCK(object);
1490 VM_OBJECT_LOCK(backing_object);
1491 /*
1492 * If we slept, anything could have
1493 * happened. Since the object is
1494 * marked dead, the backing offset
1495 * should not have changed so we
1496 * just restart our scan.
1497 */
1498 p = TAILQ_FIRST(&backing_object->memq);
1499 continue;
1500 }
1501 }
1502
1503 KASSERT(
1504 p->object == backing_object,
1505 ("vm_object_backing_scan: object mismatch")
1506 );
1507
1508 /*
1509 * Destroy any associated swap
1510 */
1511 if (backing_object->type == OBJT_SWAP) {
1512 swap_pager_freespace(
1513 backing_object,
1514 p->pindex,
1515 1
1516 );
1517 }
1518
1519 if (
1520 p->pindex < backing_offset_index ||
1521 new_pindex >= object->size
1522 ) {
1523 /*
1524 * Page is out of the parent object's range, we
1525 * can simply destroy it.
1526 */
1527 vm_page_lock(p);
1528 KASSERT(!pmap_page_is_mapped(p),
1529 ("freeing mapped page %p", p));
1530 if (p->wire_count == 0)
1531 vm_page_free(p);
1532 else
1533 vm_page_remove(p);
1534 vm_page_unlock(p);
1535 p = next;
1536 continue;
1537 }
1538
1539 pp = vm_page_lookup(object, new_pindex);
1540 if (
1541 (op & OBSC_COLLAPSE_NOWAIT) != 0 &&
1542 (pp != NULL && pp->valid == 0)
1543 ) {
1544 /*
1545 * The page in the parent is not (yet) valid.
1546 * We don't know anything about the state of
1547 * the original page. It might be mapped,
1548 * so we must avoid the next if here.
1549 *
1550 * This is due to a race in vm_fault() where
1551 * we must unbusy the original (backing_obj)
1552 * page before we can (re)lock the parent.
1553 * Hence we can get here.
1554 */
1555 p = next;
1556 continue;
1557 }
1558 if (
1559 pp != NULL ||
1560 vm_pager_has_page(object, new_pindex, NULL, NULL)
1561 ) {
1562 /*
1563 * page already exists in parent OR swap exists
1564 * for this location in the parent. Destroy
1565 * the original page from the backing object.
1566 *
1567 * Leave the parent's page alone
1568 */
1569 vm_page_lock(p);
1570 KASSERT(!pmap_page_is_mapped(p),
1571 ("freeing mapped page %p", p));
1572 if (p->wire_count == 0)
1573 vm_page_free(p);
1574 else
1575 vm_page_remove(p);
1576 vm_page_unlock(p);
1577 p = next;
1578 continue;
1579 }
1580
1581 #if VM_NRESERVLEVEL > 0
1582 /*
1583 * Rename the reservation.
1584 */
1585 vm_reserv_rename(p, object, backing_object,
1586 backing_offset_index);
1587 #endif
1588
1589 /*
1590 * Page does not exist in parent, rename the
1591 * page from the backing object to the main object.
1592 *
1593 * If the page was mapped to a process, it can remain
1594 * mapped through the rename.
1595 */
1596 vm_page_lock(p);
1597 vm_page_rename(p, object, new_pindex);
1598 vm_page_unlock(p);
1599 /* page automatically made dirty by rename */
1600 }
1601 p = next;
1602 }
1603 return (r);
1604 }
1605
1606
1607 /*
1608 * this version of collapse allows the operation to occur earlier and
1609 * when paging_in_progress is true for an object... This is not a complete
1610 * operation, but should plug 99.9% of the rest of the leaks.
1611 */
1612 static void
vm_object_qcollapse(vm_object_t object)1613 vm_object_qcollapse(vm_object_t object)
1614 {
1615 vm_object_t backing_object = object->backing_object;
1616
1617 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1618 VM_OBJECT_LOCK_ASSERT(backing_object, MA_OWNED);
1619
1620 if (backing_object->ref_count != 1)
1621 return;
1622
1623 vm_object_backing_scan(object, OBSC_COLLAPSE_NOWAIT);
1624 }
1625
1626 /*
1627 * vm_object_collapse:
1628 *
1629 * Collapse an object with the object backing it.
1630 * Pages in the backing object are moved into the
1631 * parent, and the backing object is deallocated.
1632 */
1633 void
vm_object_collapse(vm_object_t object)1634 vm_object_collapse(vm_object_t object)
1635 {
1636 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1637
1638 while (TRUE) {
1639 vm_object_t backing_object;
1640
1641 /*
1642 * Verify that the conditions are right for collapse:
1643 *
1644 * The object exists and the backing object exists.
1645 */
1646 if ((backing_object = object->backing_object) == NULL)
1647 break;
1648
1649 /*
1650 * we check the backing object first, because it is most likely
1651 * not collapsable.
1652 */
1653 VM_OBJECT_LOCK(backing_object);
1654 if (backing_object->handle != NULL ||
1655 (backing_object->type != OBJT_DEFAULT &&
1656 backing_object->type != OBJT_SWAP) ||
1657 (backing_object->flags & OBJ_DEAD) ||
1658 object->handle != NULL ||
1659 (object->type != OBJT_DEFAULT &&
1660 object->type != OBJT_SWAP) ||
1661 (object->flags & OBJ_DEAD)) {
1662 VM_OBJECT_UNLOCK(backing_object);
1663 break;
1664 }
1665
1666 if (
1667 object->paging_in_progress != 0 ||
1668 backing_object->paging_in_progress != 0
1669 ) {
1670 vm_object_qcollapse(object);
1671 VM_OBJECT_UNLOCK(backing_object);
1672 break;
1673 }
1674 /*
1675 * We know that we can either collapse the backing object (if
1676 * the parent is the only reference to it) or (perhaps) have
1677 * the parent bypass the object if the parent happens to shadow
1678 * all the resident pages in the entire backing object.
1679 *
1680 * This is ignoring pager-backed pages such as swap pages.
1681 * vm_object_backing_scan fails the shadowing test in this
1682 * case.
1683 */
1684 if (backing_object->ref_count == 1) {
1685 /*
1686 * If there is exactly one reference to the backing
1687 * object, we can collapse it into the parent.
1688 */
1689 vm_object_backing_scan(object, OBSC_COLLAPSE_WAIT);
1690
1691 #if VM_NRESERVLEVEL > 0
1692 /*
1693 * Break any reservations from backing_object.
1694 */
1695 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1696 vm_reserv_break_all(backing_object);
1697 #endif
1698
1699 /*
1700 * Move the pager from backing_object to object.
1701 */
1702 if (backing_object->type == OBJT_SWAP) {
1703 /*
1704 * swap_pager_copy() can sleep, in which case
1705 * the backing_object's and object's locks are
1706 * released and reacquired.
1707 */
1708 swap_pager_copy(
1709 backing_object,
1710 object,
1711 OFF_TO_IDX(object->backing_object_offset), TRUE);
1712
1713 /*
1714 * Free any cached pages from backing_object.
1715 */
1716 if (__predict_false(backing_object->cache != NULL))
1717 vm_page_cache_free(backing_object, 0, 0);
1718 }
1719 /*
1720 * Object now shadows whatever backing_object did.
1721 * Note that the reference to
1722 * backing_object->backing_object moves from within
1723 * backing_object to within object.
1724 */
1725 LIST_REMOVE(object, shadow_list);
1726 backing_object->shadow_count--;
1727 if (backing_object->backing_object) {
1728 VM_OBJECT_LOCK(backing_object->backing_object);
1729 LIST_REMOVE(backing_object, shadow_list);
1730 LIST_INSERT_HEAD(
1731 &backing_object->backing_object->shadow_head,
1732 object, shadow_list);
1733 /*
1734 * The shadow_count has not changed.
1735 */
1736 VM_OBJECT_UNLOCK(backing_object->backing_object);
1737 }
1738 object->backing_object = backing_object->backing_object;
1739 object->backing_object_offset +=
1740 backing_object->backing_object_offset;
1741
1742 /*
1743 * Discard backing_object.
1744 *
1745 * Since the backing object has no pages, no pager left,
1746 * and no object references within it, all that is
1747 * necessary is to dispose of it.
1748 */
1749 KASSERT(backing_object->ref_count == 1, (
1750 "backing_object %p was somehow re-referenced during collapse!",
1751 backing_object));
1752 backing_object->type = OBJT_DEAD;
1753 backing_object->ref_count = 0;
1754 VM_OBJECT_UNLOCK(backing_object);
1755 vm_object_destroy(backing_object);
1756
1757 object_collapses++;
1758 } else {
1759 vm_object_t new_backing_object;
1760
1761 /*
1762 * If we do not entirely shadow the backing object,
1763 * there is nothing we can do so we give up.
1764 */
1765 if (object->resident_page_count != object->size &&
1766 vm_object_backing_scan(object,
1767 OBSC_TEST_ALL_SHADOWED) == 0) {
1768 VM_OBJECT_UNLOCK(backing_object);
1769 break;
1770 }
1771
1772 /*
1773 * Make the parent shadow the next object in the
1774 * chain. Deallocating backing_object will not remove
1775 * it, since its reference count is at least 2.
1776 */
1777 LIST_REMOVE(object, shadow_list);
1778 backing_object->shadow_count--;
1779
1780 new_backing_object = backing_object->backing_object;
1781 if ((object->backing_object = new_backing_object) != NULL) {
1782 VM_OBJECT_LOCK(new_backing_object);
1783 LIST_INSERT_HEAD(
1784 &new_backing_object->shadow_head,
1785 object,
1786 shadow_list
1787 );
1788 new_backing_object->shadow_count++;
1789 vm_object_reference_locked(new_backing_object);
1790 VM_OBJECT_UNLOCK(new_backing_object);
1791 object->backing_object_offset +=
1792 backing_object->backing_object_offset;
1793 }
1794
1795 /*
1796 * Drop the reference count on backing_object. Since
1797 * its ref_count was at least 2, it will not vanish.
1798 */
1799 backing_object->ref_count--;
1800 VM_OBJECT_UNLOCK(backing_object);
1801 object_bypasses++;
1802 }
1803
1804 /*
1805 * Try again with this object's new backing object.
1806 */
1807 }
1808 }
1809
1810 /*
1811 * vm_object_page_remove:
1812 *
1813 * For the given object, either frees or invalidates each of the
1814 * specified pages. In general, a page is freed. However, if a page is
1815 * wired for any reason other than the existence of a managed, wired
1816 * mapping, then it may be invalidated but not removed from the object.
1817 * Pages are specified by the given range ["start", "end") and the option
1818 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
1819 * extends from "start" to the end of the object. If the option
1820 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
1821 * specified range are affected. If the option OBJPR_NOTMAPPED is
1822 * specified, then the pages within the specified range must have no
1823 * mappings. Otherwise, if this option is not specified, any mappings to
1824 * the specified pages are removed before the pages are freed or
1825 * invalidated.
1826 *
1827 * In general, this operation should only be performed on objects that
1828 * contain managed pages. There are, however, two exceptions. First, it
1829 * is performed on the kernel and kmem objects by vm_map_entry_delete().
1830 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
1831 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
1832 * not be specified and the option OBJPR_NOTMAPPED must be specified.
1833 *
1834 * The object must be locked.
1835 */
1836 void
vm_object_page_remove(vm_object_t object,vm_pindex_t start,vm_pindex_t end,int options)1837 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1838 int options)
1839 {
1840 vm_page_t p, next;
1841 int wirings;
1842
1843 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1844 KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_PHYS) ||
1845 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
1846 ("vm_object_page_remove: illegal options for object %p", object));
1847 if (object->resident_page_count == 0)
1848 goto skipmemq;
1849 vm_object_pip_add(object, 1);
1850 again:
1851 p = vm_page_find_least(object, start);
1852
1853 /*
1854 * Here, the variable "p" is either (1) the page with the least pindex
1855 * greater than or equal to the parameter "start" or (2) NULL.
1856 */
1857 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1858 next = TAILQ_NEXT(p, listq);
1859
1860 /*
1861 * If the page is wired for any reason besides the existence
1862 * of managed, wired mappings, then it cannot be freed. For
1863 * example, fictitious pages, which represent device memory,
1864 * are inherently wired and cannot be freed. They can,
1865 * however, be invalidated if the option OBJPR_CLEANONLY is
1866 * not specified.
1867 */
1868 vm_page_lock(p);
1869 if ((wirings = p->wire_count) != 0 &&
1870 (wirings = pmap_page_wired_mappings(p)) != p->wire_count) {
1871 if ((options & (OBJPR_NOTWIRED | OBJPR_NOTMAPPED)) ==
1872 0) {
1873 pmap_remove_all(p);
1874 /* Account for removal of wired mappings. */
1875 if (wirings != 0)
1876 p->wire_count -= wirings;
1877 }
1878 if ((options & OBJPR_CLEANONLY) == 0) {
1879 p->valid = 0;
1880 vm_page_undirty(p);
1881 }
1882 goto next;
1883 }
1884 if (vm_page_sleep_if_busy(p, TRUE, "vmopar"))
1885 goto again;
1886 KASSERT((p->flags & PG_FICTITIOUS) == 0,
1887 ("vm_object_page_remove: page %p is fictitious", p));
1888 if ((options & OBJPR_CLEANONLY) != 0 && p->valid != 0) {
1889 if ((options & OBJPR_NOTMAPPED) == 0)
1890 pmap_remove_write(p);
1891 if (p->dirty)
1892 goto next;
1893 }
1894 if ((options & OBJPR_NOTMAPPED) == 0) {
1895 if ((options & OBJPR_NOTWIRED) != 0 && wirings != 0)
1896 goto next;
1897 pmap_remove_all(p);
1898 /* Account for removal of wired mappings. */
1899 if (wirings != 0) {
1900 KASSERT(p->wire_count == wirings,
1901 ("inconsistent wire count %d %d %p",
1902 p->wire_count, wirings, p));
1903 p->wire_count = 0;
1904 atomic_subtract_int(&cnt.v_wire_count, 1);
1905 }
1906 }
1907 vm_page_free(p);
1908 next:
1909 vm_page_unlock(p);
1910 }
1911 vm_object_pip_wakeup(object);
1912 skipmemq:
1913 if (__predict_false(object->cache != NULL))
1914 vm_page_cache_free(object, start, end);
1915 }
1916
1917 /*
1918 * vm_object_page_cache:
1919 *
1920 * For the given object, attempt to move the specified clean
1921 * pages to the cache queue. If a page is wired for any reason,
1922 * then it will not be changed. Pages are specified by the given
1923 * range ["start", "end"). As a special case, if "end" is zero,
1924 * then the range extends from "start" to the end of the object.
1925 * Any mappings to the specified pages are removed before the
1926 * pages are moved to the cache queue.
1927 *
1928 * This operation should only be performed on objects that
1929 * contain managed pages.
1930 *
1931 * The object must be locked.
1932 */
1933 void
vm_object_page_cache(vm_object_t object,vm_pindex_t start,vm_pindex_t end)1934 vm_object_page_cache(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_LOCK_ASSERT(object, MA_OWNED);
1940 KASSERT((object->type != OBJT_DEVICE && object->type != OBJT_SG &&
1941 object->type != OBJT_PHYS),
1942 ("vm_object_page_cache: illegal object %p", object));
1943 if (object->resident_page_count == 0)
1944 return;
1945 p = vm_page_find_least(object, start);
1946
1947 /*
1948 * Here, the variable "p" is either (1) the page with the least pindex
1949 * greater than or equal to the parameter "start" or (2) NULL.
1950 */
1951 mtx = NULL;
1952 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
1953 next = TAILQ_NEXT(p, listq);
1954
1955 /*
1956 * Avoid releasing and reacquiring the same page lock.
1957 */
1958 new_mtx = vm_page_lockptr(p);
1959 if (mtx != new_mtx) {
1960 if (mtx != NULL)
1961 mtx_unlock(mtx);
1962 mtx = new_mtx;
1963 mtx_lock(mtx);
1964 }
1965 vm_page_try_to_cache(p);
1966 }
1967 if (mtx != NULL)
1968 mtx_unlock(mtx);
1969 }
1970
1971 /*
1972 * Populate the specified range of the object with valid pages. Returns
1973 * TRUE if the range is successfully populated and FALSE otherwise.
1974 *
1975 * Note: This function should be optimized to pass a larger array of
1976 * pages to vm_pager_get_pages() before it is applied to a non-
1977 * OBJT_DEVICE object.
1978 *
1979 * The object must be locked.
1980 */
1981 boolean_t
vm_object_populate(vm_object_t object,vm_pindex_t start,vm_pindex_t end)1982 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1983 {
1984 vm_page_t m, ma[1];
1985 vm_pindex_t pindex;
1986 int rv;
1987
1988 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1989 for (pindex = start; pindex < end; pindex++) {
1990 m = vm_page_grab(object, pindex, VM_ALLOC_NORMAL |
1991 VM_ALLOC_RETRY);
1992 if (m->valid != VM_PAGE_BITS_ALL) {
1993 ma[0] = m;
1994 rv = vm_pager_get_pages(object, ma, 1, 0);
1995 m = vm_page_lookup(object, pindex);
1996 if (m == NULL)
1997 break;
1998 if (rv != VM_PAGER_OK) {
1999 vm_page_lock(m);
2000 vm_page_free(m);
2001 vm_page_unlock(m);
2002 break;
2003 }
2004 }
2005 /*
2006 * Keep "m" busy because a subsequent iteration may unlock
2007 * the object.
2008 */
2009 }
2010 if (pindex > start) {
2011 m = vm_page_lookup(object, start);
2012 while (m != NULL && m->pindex < pindex) {
2013 vm_page_wakeup(m);
2014 m = TAILQ_NEXT(m, listq);
2015 }
2016 }
2017 return (pindex == end);
2018 }
2019
2020 /*
2021 * Routine: vm_object_coalesce
2022 * Function: Coalesces two objects backing up adjoining
2023 * regions of memory into a single object.
2024 *
2025 * returns TRUE if objects were combined.
2026 *
2027 * NOTE: Only works at the moment if the second object is NULL -
2028 * if it's not, which object do we lock first?
2029 *
2030 * Parameters:
2031 * prev_object First object to coalesce
2032 * prev_offset Offset into prev_object
2033 * prev_size Size of reference to prev_object
2034 * next_size Size of reference to the second object
2035 * reserved Indicator that extension region has
2036 * swap accounted for
2037 *
2038 * Conditions:
2039 * The object must *not* be locked.
2040 */
2041 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)2042 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2043 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2044 {
2045 vm_pindex_t next_pindex;
2046
2047 if (prev_object == NULL)
2048 return (TRUE);
2049 VM_OBJECT_LOCK(prev_object);
2050 if (prev_object->type != OBJT_DEFAULT &&
2051 prev_object->type != OBJT_SWAP) {
2052 VM_OBJECT_UNLOCK(prev_object);
2053 return (FALSE);
2054 }
2055
2056 /*
2057 * Try to collapse the object first
2058 */
2059 vm_object_collapse(prev_object);
2060
2061 /*
2062 * Can't coalesce if: . more than one reference . paged out . shadows
2063 * another object . has a copy elsewhere (any of which mean that the
2064 * pages not mapped to prev_entry may be in use anyway)
2065 */
2066 if (prev_object->backing_object != NULL) {
2067 VM_OBJECT_UNLOCK(prev_object);
2068 return (FALSE);
2069 }
2070
2071 prev_size >>= PAGE_SHIFT;
2072 next_size >>= PAGE_SHIFT;
2073 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2074
2075 if ((prev_object->ref_count > 1) &&
2076 (prev_object->size != next_pindex)) {
2077 VM_OBJECT_UNLOCK(prev_object);
2078 return (FALSE);
2079 }
2080
2081 /*
2082 * Account for the charge.
2083 */
2084 if (prev_object->cred != NULL) {
2085
2086 /*
2087 * If prev_object was charged, then this mapping,
2088 * althought not charged now, may become writable
2089 * later. Non-NULL cred in the object would prevent
2090 * swap reservation during enabling of the write
2091 * access, so reserve swap now. Failed reservation
2092 * cause allocation of the separate object for the map
2093 * entry, and swap reservation for this entry is
2094 * managed in appropriate time.
2095 */
2096 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2097 prev_object->cred)) {
2098 return (FALSE);
2099 }
2100 prev_object->charge += ptoa(next_size);
2101 }
2102
2103 /*
2104 * Remove any pages that may still be in the object from a previous
2105 * deallocation.
2106 */
2107 if (next_pindex < prev_object->size) {
2108 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2109 next_size, 0);
2110 if (prev_object->type == OBJT_SWAP)
2111 swap_pager_freespace(prev_object,
2112 next_pindex, next_size);
2113 #if 0
2114 if (prev_object->cred != NULL) {
2115 KASSERT(prev_object->charge >=
2116 ptoa(prev_object->size - next_pindex),
2117 ("object %p overcharged 1 %jx %jx", prev_object,
2118 (uintmax_t)next_pindex, (uintmax_t)next_size));
2119 prev_object->charge -= ptoa(prev_object->size -
2120 next_pindex);
2121 }
2122 #endif
2123 }
2124
2125 /*
2126 * Extend the object if necessary.
2127 */
2128 if (next_pindex + next_size > prev_object->size)
2129 prev_object->size = next_pindex + next_size;
2130
2131 VM_OBJECT_UNLOCK(prev_object);
2132 return (TRUE);
2133 }
2134
2135 void
vm_object_set_writeable_dirty(vm_object_t object)2136 vm_object_set_writeable_dirty(vm_object_t object)
2137 {
2138
2139 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2140 if (object->type != OBJT_VNODE)
2141 return;
2142 object->generation++;
2143 if ((object->flags & OBJ_MIGHTBEDIRTY) != 0)
2144 return;
2145 vm_object_set_flag(object, OBJ_MIGHTBEDIRTY);
2146 }
2147
2148 #include "opt_ddb.h"
2149 #ifdef DDB
2150 #include <sys/kernel.h>
2151
2152 #include <sys/cons.h>
2153
2154 #include <ddb/ddb.h>
2155
2156 static int
_vm_object_in_map(vm_map_t map,vm_object_t object,vm_map_entry_t entry)2157 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2158 {
2159 vm_map_t tmpm;
2160 vm_map_entry_t tmpe;
2161 vm_object_t obj;
2162 int entcount;
2163
2164 if (map == 0)
2165 return 0;
2166
2167 if (entry == 0) {
2168 tmpe = map->header.next;
2169 entcount = map->nentries;
2170 while (entcount-- && (tmpe != &map->header)) {
2171 if (_vm_object_in_map(map, object, tmpe)) {
2172 return 1;
2173 }
2174 tmpe = tmpe->next;
2175 }
2176 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2177 tmpm = entry->object.sub_map;
2178 tmpe = tmpm->header.next;
2179 entcount = tmpm->nentries;
2180 while (entcount-- && tmpe != &tmpm->header) {
2181 if (_vm_object_in_map(tmpm, object, tmpe)) {
2182 return 1;
2183 }
2184 tmpe = tmpe->next;
2185 }
2186 } else if ((obj = entry->object.vm_object) != NULL) {
2187 for (; obj; obj = obj->backing_object)
2188 if (obj == object) {
2189 return 1;
2190 }
2191 }
2192 return 0;
2193 }
2194
2195 static int
vm_object_in_map(vm_object_t object)2196 vm_object_in_map(vm_object_t object)
2197 {
2198 struct proc *p;
2199
2200 /* sx_slock(&allproc_lock); */
2201 FOREACH_PROC_IN_SYSTEM(p) {
2202 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2203 continue;
2204 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2205 /* sx_sunlock(&allproc_lock); */
2206 return 1;
2207 }
2208 }
2209 /* sx_sunlock(&allproc_lock); */
2210 if (_vm_object_in_map(kernel_map, object, 0))
2211 return 1;
2212 if (_vm_object_in_map(kmem_map, object, 0))
2213 return 1;
2214 if (_vm_object_in_map(pager_map, object, 0))
2215 return 1;
2216 if (_vm_object_in_map(buffer_map, object, 0))
2217 return 1;
2218 return 0;
2219 }
2220
DB_SHOW_COMMAND(vmochk,vm_object_check)2221 DB_SHOW_COMMAND(vmochk, vm_object_check)
2222 {
2223 vm_object_t object;
2224
2225 /*
2226 * make sure that internal objs are in a map somewhere
2227 * and none have zero ref counts.
2228 */
2229 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2230 if (object->handle == NULL &&
2231 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) {
2232 if (object->ref_count == 0) {
2233 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2234 (long)object->size);
2235 }
2236 if (!vm_object_in_map(object)) {
2237 db_printf(
2238 "vmochk: internal obj is not in a map: "
2239 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2240 object->ref_count, (u_long)object->size,
2241 (u_long)object->size,
2242 (void *)object->backing_object);
2243 }
2244 }
2245 }
2246 }
2247
2248 /*
2249 * vm_object_print: [ debug ]
2250 */
DB_SHOW_COMMAND(object,vm_object_print_static)2251 DB_SHOW_COMMAND(object, vm_object_print_static)
2252 {
2253 /* XXX convert args. */
2254 vm_object_t object = (vm_object_t)addr;
2255 boolean_t full = have_addr;
2256
2257 vm_page_t p;
2258
2259 /* XXX count is an (unused) arg. Avoid shadowing it. */
2260 #define count was_count
2261
2262 int count;
2263
2264 if (object == NULL)
2265 return;
2266
2267 db_iprintf(
2268 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2269 object, (int)object->type, (uintmax_t)object->size,
2270 object->resident_page_count, object->ref_count, object->flags,
2271 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2272 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2273 object->shadow_count,
2274 object->backing_object ? object->backing_object->ref_count : 0,
2275 object->backing_object, (uintmax_t)object->backing_object_offset);
2276
2277 if (!full)
2278 return;
2279
2280 db_indent += 2;
2281 count = 0;
2282 TAILQ_FOREACH(p, &object->memq, listq) {
2283 if (count == 0)
2284 db_iprintf("memory:=");
2285 else if (count == 6) {
2286 db_printf("\n");
2287 db_iprintf(" ...");
2288 count = 0;
2289 } else
2290 db_printf(",");
2291 count++;
2292
2293 db_printf("(off=0x%jx,page=0x%jx)",
2294 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2295 }
2296 if (count != 0)
2297 db_printf("\n");
2298 db_indent -= 2;
2299 }
2300
2301 /* XXX. */
2302 #undef count
2303
2304 /* XXX need this non-static entry for calling from vm_map_print. */
2305 void
vm_object_print(long addr,boolean_t have_addr,long count,char * modif)2306 vm_object_print(
2307 /* db_expr_t */ long addr,
2308 boolean_t have_addr,
2309 /* db_expr_t */ long count,
2310 char *modif)
2311 {
2312 vm_object_print_static(addr, have_addr, count, modif);
2313 }
2314
DB_SHOW_COMMAND(vmopag,vm_object_print_pages)2315 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
2316 {
2317 vm_object_t object;
2318 vm_pindex_t fidx;
2319 vm_paddr_t pa;
2320 vm_page_t m, prev_m;
2321 int rcount, nl, c;
2322
2323 nl = 0;
2324 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2325 db_printf("new object: %p\n", (void *)object);
2326 if (nl > 18) {
2327 c = cngetc();
2328 if (c != ' ')
2329 return;
2330 nl = 0;
2331 }
2332 nl++;
2333 rcount = 0;
2334 fidx = 0;
2335 pa = -1;
2336 TAILQ_FOREACH(m, &object->memq, listq) {
2337 if (m->pindex > 128)
2338 break;
2339 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2340 prev_m->pindex + 1 != m->pindex) {
2341 if (rcount) {
2342 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2343 (long)fidx, rcount, (long)pa);
2344 if (nl > 18) {
2345 c = cngetc();
2346 if (c != ' ')
2347 return;
2348 nl = 0;
2349 }
2350 nl++;
2351 rcount = 0;
2352 }
2353 }
2354 if (rcount &&
2355 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2356 ++rcount;
2357 continue;
2358 }
2359 if (rcount) {
2360 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2361 (long)fidx, rcount, (long)pa);
2362 if (nl > 18) {
2363 c = cngetc();
2364 if (c != ' ')
2365 return;
2366 nl = 0;
2367 }
2368 nl++;
2369 }
2370 fidx = m->pindex;
2371 pa = VM_PAGE_TO_PHYS(m);
2372 rcount = 1;
2373 }
2374 if (rcount) {
2375 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2376 (long)fidx, rcount, (long)pa);
2377 if (nl > 18) {
2378 c = cngetc();
2379 if (c != ' ')
2380 return;
2381 nl = 0;
2382 }
2383 nl++;
2384 }
2385 }
2386 }
2387 #endif /* DDB */
2388