1 /*-
2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU)
3 *
4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
6 *
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
21 *
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 *
34 *
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 "opt_vm.h"
66
67 #include <sys/systm.h>
68 #include <sys/blockcount.h>
69 #include <sys/cpuset.h>
70 #include <sys/jail.h>
71 #include <sys/limits.h>
72 #include <sys/lock.h>
73 #include <sys/mman.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/mutex.h>
77 #include <sys/pctrie.h>
78 #include <sys/proc.h>
79 #include <sys/refcount.h>
80 #include <sys/sx.h>
81 #include <sys/sysctl.h>
82 #include <sys/resourcevar.h>
83 #include <sys/refcount.h>
84 #include <sys/rwlock.h>
85 #include <sys/user.h>
86 #include <sys/vnode.h>
87 #include <sys/vmmeter.h>
88
89 #include <vm/vm.h>
90 #include <vm/vm_param.h>
91 #include <vm/pmap.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_page.h>
95 #include <vm/vm_pageout.h>
96 #include <vm/vm_pager.h>
97 #include <vm/vm_phys.h>
98 #include <vm/vm_pagequeue.h>
99 #include <vm/swap_pager.h>
100 #include <vm/vm_kern.h>
101 #include <vm/vm_extern.h>
102 #include <vm/vm_radix.h>
103 #include <vm/vm_reserv.h>
104 #include <vm/uma.h>
105
106 static int old_msync;
107 SYSCTL_INT(_vm, OID_AUTO, old_msync, CTLFLAG_RW, &old_msync, 0,
108 "Use old (insecure) msync behavior");
109
110 static int vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
111 int pagerflags, int flags, boolean_t *allclean,
112 boolean_t *eio);
113 static boolean_t vm_object_page_remove_write(vm_page_t p, int flags,
114 boolean_t *allclean);
115 static void vm_object_backing_remove(vm_object_t object);
116
117 /*
118 * Virtual memory objects maintain the actual data
119 * associated with allocated virtual memory. A given
120 * page of memory exists within exactly one object.
121 *
122 * An object is only deallocated when all "references"
123 * are given up. Only one "reference" to a given
124 * region of an object should be writeable.
125 *
126 * Associated with each object is a list of all resident
127 * memory pages belonging to that object; this list is
128 * maintained by the "vm_page" module, and locked by the object's
129 * lock.
130 *
131 * Each object also records a "pager" routine which is
132 * used to retrieve (and store) pages to the proper backing
133 * storage. In addition, objects may be backed by other
134 * objects from which they were virtual-copied.
135 *
136 * The only items within the object structure which are
137 * modified after time of creation are:
138 * reference count locked by object's lock
139 * pager routine locked by object's lock
140 *
141 */
142
143 struct object_q vm_object_list;
144 struct mtx vm_object_list_mtx; /* lock for object list and count */
145
146 struct vm_object kernel_object_store;
147
148 static SYSCTL_NODE(_vm_stats, OID_AUTO, object, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
149 "VM object stats");
150
151 static COUNTER_U64_DEFINE_EARLY(object_collapses);
152 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapses, CTLFLAG_RD,
153 &object_collapses,
154 "VM object collapses");
155
156 static COUNTER_U64_DEFINE_EARLY(object_bypasses);
157 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, bypasses, CTLFLAG_RD,
158 &object_bypasses,
159 "VM object bypasses");
160
161 static COUNTER_U64_DEFINE_EARLY(object_collapse_waits);
162 SYSCTL_COUNTER_U64(_vm_stats_object, OID_AUTO, collapse_waits, CTLFLAG_RD,
163 &object_collapse_waits,
164 "Number of sleeps for collapse");
165
166 static uma_zone_t obj_zone;
167
168 static int vm_object_zinit(void *mem, int size, int flags);
169
170 #ifdef INVARIANTS
171 static void vm_object_zdtor(void *mem, int size, void *arg);
172
173 static void
vm_object_zdtor(void * mem,int size,void * arg)174 vm_object_zdtor(void *mem, int size, void *arg)
175 {
176 vm_object_t object;
177
178 object = (vm_object_t)mem;
179 KASSERT(object->ref_count == 0,
180 ("object %p ref_count = %d", object, object->ref_count));
181 KASSERT(TAILQ_EMPTY(&object->memq),
182 ("object %p has resident pages in its memq", object));
183 KASSERT(vm_radix_is_empty(&object->rtree),
184 ("object %p has resident pages in its trie", object));
185 #if VM_NRESERVLEVEL > 0
186 KASSERT(LIST_EMPTY(&object->rvq),
187 ("object %p has reservations",
188 object));
189 #endif
190 KASSERT(!vm_object_busied(object),
191 ("object %p busy = %d", object, blockcount_read(&object->busy)));
192 KASSERT(object->resident_page_count == 0,
193 ("object %p resident_page_count = %d",
194 object, object->resident_page_count));
195 KASSERT(atomic_load_int(&object->shadow_count) == 0,
196 ("object %p shadow_count = %d",
197 object, atomic_load_int(&object->shadow_count)));
198 KASSERT(object->type == OBJT_DEAD,
199 ("object %p has non-dead type %d",
200 object, object->type));
201 KASSERT(object->charge == 0 && object->cred == NULL,
202 ("object %p has non-zero charge %ju (%p)",
203 object, (uintmax_t)object->charge, object->cred));
204 }
205 #endif
206
207 static int
vm_object_zinit(void * mem,int size,int flags)208 vm_object_zinit(void *mem, int size, int flags)
209 {
210 vm_object_t object;
211
212 object = (vm_object_t)mem;
213 rw_init_flags(&object->lock, "vmobject", RW_DUPOK | RW_NEW);
214
215 /* These are true for any object that has been freed */
216 object->type = OBJT_DEAD;
217 vm_radix_init(&object->rtree);
218 refcount_init(&object->ref_count, 0);
219 blockcount_init(&object->paging_in_progress);
220 blockcount_init(&object->busy);
221 object->resident_page_count = 0;
222 atomic_store_int(&object->shadow_count, 0);
223 object->flags = OBJ_DEAD;
224
225 mtx_lock(&vm_object_list_mtx);
226 TAILQ_INSERT_TAIL(&vm_object_list, object, object_list);
227 mtx_unlock(&vm_object_list_mtx);
228 return (0);
229 }
230
231 static void
_vm_object_allocate(objtype_t type,vm_pindex_t size,u_short flags,vm_object_t object,void * handle)232 _vm_object_allocate(objtype_t type, vm_pindex_t size, u_short flags,
233 vm_object_t object, void *handle)
234 {
235
236 TAILQ_INIT(&object->memq);
237 LIST_INIT(&object->shadow_head);
238
239 object->type = type;
240 object->flags = flags;
241 if ((flags & OBJ_SWAP) != 0) {
242 pctrie_init(&object->un_pager.swp.swp_blks);
243 object->un_pager.swp.writemappings = 0;
244 }
245
246 /*
247 * Ensure that swap_pager_swapoff() iteration over object_list
248 * sees up to date type and pctrie head if it observed
249 * non-dead object.
250 */
251 atomic_thread_fence_rel();
252
253 object->pg_color = 0;
254 object->size = size;
255 object->domain.dr_policy = NULL;
256 object->generation = 1;
257 object->cleangeneration = 1;
258 refcount_init(&object->ref_count, 1);
259 object->memattr = VM_MEMATTR_DEFAULT;
260 object->cred = NULL;
261 object->charge = 0;
262 object->handle = handle;
263 object->backing_object = NULL;
264 object->backing_object_offset = (vm_ooffset_t) 0;
265 #if VM_NRESERVLEVEL > 0
266 LIST_INIT(&object->rvq);
267 #endif
268 umtx_shm_object_init(object);
269 }
270
271 /*
272 * vm_object_init:
273 *
274 * Initialize the VM objects module.
275 */
276 void
vm_object_init(void)277 vm_object_init(void)
278 {
279 TAILQ_INIT(&vm_object_list);
280 mtx_init(&vm_object_list_mtx, "vm object_list", NULL, MTX_DEF);
281
282 rw_init(&kernel_object->lock, "kernel vm object");
283 vm_radix_init(&kernel_object->rtree);
284 _vm_object_allocate(OBJT_PHYS, atop(VM_MAX_KERNEL_ADDRESS -
285 VM_MIN_KERNEL_ADDRESS), OBJ_UNMANAGED, kernel_object, NULL);
286 #if VM_NRESERVLEVEL > 0
287 kernel_object->flags |= OBJ_COLORED;
288 kernel_object->pg_color = (u_short)atop(VM_MIN_KERNEL_ADDRESS);
289 #endif
290 kernel_object->un_pager.phys.ops = &default_phys_pg_ops;
291
292 /*
293 * The lock portion of struct vm_object must be type stable due
294 * to vm_pageout_fallback_object_lock locking a vm object
295 * without holding any references to it.
296 *
297 * paging_in_progress is valid always. Lockless references to
298 * the objects may acquire pip and then check OBJ_DEAD.
299 */
300 obj_zone = uma_zcreate("VM OBJECT", sizeof (struct vm_object), NULL,
301 #ifdef INVARIANTS
302 vm_object_zdtor,
303 #else
304 NULL,
305 #endif
306 vm_object_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
307
308 vm_radix_zinit();
309 }
310
311 void
vm_object_clear_flag(vm_object_t object,u_short bits)312 vm_object_clear_flag(vm_object_t object, u_short bits)
313 {
314
315 VM_OBJECT_ASSERT_WLOCKED(object);
316 object->flags &= ~bits;
317 }
318
319 /*
320 * Sets the default memory attribute for the specified object. Pages
321 * that are allocated to this object are by default assigned this memory
322 * attribute.
323 *
324 * Presently, this function must be called before any pages are allocated
325 * to the object. In the future, this requirement may be relaxed for
326 * "default" and "swap" objects.
327 */
328 int
vm_object_set_memattr(vm_object_t object,vm_memattr_t memattr)329 vm_object_set_memattr(vm_object_t object, vm_memattr_t memattr)
330 {
331
332 VM_OBJECT_ASSERT_WLOCKED(object);
333
334 if (object->type == OBJT_DEAD)
335 return (KERN_INVALID_ARGUMENT);
336 if (!TAILQ_EMPTY(&object->memq))
337 return (KERN_FAILURE);
338
339 object->memattr = memattr;
340 return (KERN_SUCCESS);
341 }
342
343 void
vm_object_pip_add(vm_object_t object,short i)344 vm_object_pip_add(vm_object_t object, short i)
345 {
346
347 if (i > 0)
348 blockcount_acquire(&object->paging_in_progress, i);
349 }
350
351 void
vm_object_pip_wakeup(vm_object_t object)352 vm_object_pip_wakeup(vm_object_t object)
353 {
354
355 vm_object_pip_wakeupn(object, 1);
356 }
357
358 void
vm_object_pip_wakeupn(vm_object_t object,short i)359 vm_object_pip_wakeupn(vm_object_t object, short i)
360 {
361
362 if (i > 0)
363 blockcount_release(&object->paging_in_progress, i);
364 }
365
366 /*
367 * Atomically drop the object lock and wait for pip to drain. This protects
368 * from sleep/wakeup races due to identity changes. The lock is not re-acquired
369 * on return.
370 */
371 static void
vm_object_pip_sleep(vm_object_t object,const char * waitid)372 vm_object_pip_sleep(vm_object_t object, const char *waitid)
373 {
374
375 (void)blockcount_sleep(&object->paging_in_progress, &object->lock,
376 waitid, PVM | PDROP);
377 }
378
379 void
vm_object_pip_wait(vm_object_t object,const char * waitid)380 vm_object_pip_wait(vm_object_t object, const char *waitid)
381 {
382
383 VM_OBJECT_ASSERT_WLOCKED(object);
384
385 blockcount_wait(&object->paging_in_progress, &object->lock, waitid,
386 PVM);
387 }
388
389 void
vm_object_pip_wait_unlocked(vm_object_t object,const char * waitid)390 vm_object_pip_wait_unlocked(vm_object_t object, const char *waitid)
391 {
392
393 VM_OBJECT_ASSERT_UNLOCKED(object);
394
395 blockcount_wait(&object->paging_in_progress, NULL, waitid, PVM);
396 }
397
398 /*
399 * vm_object_allocate:
400 *
401 * Returns a new object with the given size.
402 */
403 vm_object_t
vm_object_allocate(objtype_t type,vm_pindex_t size)404 vm_object_allocate(objtype_t type, vm_pindex_t size)
405 {
406 vm_object_t object;
407 u_short flags;
408
409 switch (type) {
410 case OBJT_DEAD:
411 panic("vm_object_allocate: can't create OBJT_DEAD");
412 case OBJT_SWAP:
413 flags = OBJ_COLORED | OBJ_SWAP;
414 break;
415 case OBJT_DEVICE:
416 case OBJT_SG:
417 flags = OBJ_FICTITIOUS | OBJ_UNMANAGED;
418 break;
419 case OBJT_MGTDEVICE:
420 flags = OBJ_FICTITIOUS;
421 break;
422 case OBJT_PHYS:
423 flags = OBJ_UNMANAGED;
424 break;
425 case OBJT_VNODE:
426 flags = 0;
427 break;
428 default:
429 panic("vm_object_allocate: type %d is undefined or dynamic",
430 type);
431 }
432 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
433 _vm_object_allocate(type, size, flags, object, NULL);
434
435 return (object);
436 }
437
438 vm_object_t
vm_object_allocate_dyn(objtype_t dyntype,vm_pindex_t size,u_short flags)439 vm_object_allocate_dyn(objtype_t dyntype, vm_pindex_t size, u_short flags)
440 {
441 vm_object_t object;
442
443 MPASS(dyntype >= OBJT_FIRST_DYN /* && dyntype < nitems(pagertab) */);
444 object = (vm_object_t)uma_zalloc(obj_zone, M_WAITOK);
445 _vm_object_allocate(dyntype, size, flags, object, NULL);
446
447 return (object);
448 }
449
450 /*
451 * vm_object_allocate_anon:
452 *
453 * Returns a new default object of the given size and marked as
454 * anonymous memory for special split/collapse handling. Color
455 * to be initialized by the caller.
456 */
457 vm_object_t
vm_object_allocate_anon(vm_pindex_t size,vm_object_t backing_object,struct ucred * cred,vm_size_t charge)458 vm_object_allocate_anon(vm_pindex_t size, vm_object_t backing_object,
459 struct ucred *cred, vm_size_t charge)
460 {
461 vm_object_t handle, object;
462
463 if (backing_object == NULL)
464 handle = NULL;
465 else if ((backing_object->flags & OBJ_ANON) != 0)
466 handle = backing_object->handle;
467 else
468 handle = backing_object;
469 object = uma_zalloc(obj_zone, M_WAITOK);
470 _vm_object_allocate(OBJT_SWAP, size,
471 OBJ_ANON | OBJ_ONEMAPPING | OBJ_SWAP, object, handle);
472 object->cred = cred;
473 object->charge = cred != NULL ? charge : 0;
474 return (object);
475 }
476
477 static void
vm_object_reference_vnode(vm_object_t object)478 vm_object_reference_vnode(vm_object_t object)
479 {
480 u_int old;
481
482 /*
483 * vnode objects need the lock for the first reference
484 * to serialize with vnode_object_deallocate().
485 */
486 if (!refcount_acquire_if_gt(&object->ref_count, 0)) {
487 VM_OBJECT_RLOCK(object);
488 old = refcount_acquire(&object->ref_count);
489 if (object->type == OBJT_VNODE && old == 0)
490 vref(object->handle);
491 VM_OBJECT_RUNLOCK(object);
492 }
493 }
494
495 /*
496 * vm_object_reference:
497 *
498 * Acquires a reference to the given object.
499 */
500 void
vm_object_reference(vm_object_t object)501 vm_object_reference(vm_object_t object)
502 {
503
504 if (object == NULL)
505 return;
506
507 if (object->type == OBJT_VNODE)
508 vm_object_reference_vnode(object);
509 else
510 refcount_acquire(&object->ref_count);
511 KASSERT((object->flags & OBJ_DEAD) == 0,
512 ("vm_object_reference: Referenced dead object."));
513 }
514
515 /*
516 * vm_object_reference_locked:
517 *
518 * Gets another reference to the given object.
519 *
520 * The object must be locked.
521 */
522 void
vm_object_reference_locked(vm_object_t object)523 vm_object_reference_locked(vm_object_t object)
524 {
525 u_int old;
526
527 VM_OBJECT_ASSERT_LOCKED(object);
528 old = refcount_acquire(&object->ref_count);
529 if (object->type == OBJT_VNODE && old == 0)
530 vref(object->handle);
531 KASSERT((object->flags & OBJ_DEAD) == 0,
532 ("vm_object_reference: Referenced dead object."));
533 }
534
535 /*
536 * Handle deallocating an object of type OBJT_VNODE.
537 */
538 static void
vm_object_deallocate_vnode(vm_object_t object)539 vm_object_deallocate_vnode(vm_object_t object)
540 {
541 struct vnode *vp = (struct vnode *) object->handle;
542 bool last;
543
544 KASSERT(object->type == OBJT_VNODE,
545 ("vm_object_deallocate_vnode: not a vnode object"));
546 KASSERT(vp != NULL, ("vm_object_deallocate_vnode: missing vp"));
547
548 /* Object lock to protect handle lookup. */
549 last = refcount_release(&object->ref_count);
550 VM_OBJECT_RUNLOCK(object);
551
552 if (!last)
553 return;
554
555 if (!umtx_shm_vnobj_persistent)
556 umtx_shm_object_terminated(object);
557
558 /* vrele may need the vnode lock. */
559 vrele(vp);
560 }
561
562 /*
563 * We dropped a reference on an object and discovered that it had a
564 * single remaining shadow. This is a sibling of the reference we
565 * dropped. Attempt to collapse the sibling and backing object.
566 */
567 static vm_object_t
vm_object_deallocate_anon(vm_object_t backing_object)568 vm_object_deallocate_anon(vm_object_t backing_object)
569 {
570 vm_object_t object;
571
572 /* Fetch the final shadow. */
573 object = LIST_FIRST(&backing_object->shadow_head);
574 KASSERT(object != NULL &&
575 atomic_load_int(&backing_object->shadow_count) == 1,
576 ("vm_object_anon_deallocate: ref_count: %d, shadow_count: %d",
577 backing_object->ref_count,
578 atomic_load_int(&backing_object->shadow_count)));
579 KASSERT((object->flags & OBJ_ANON) != 0,
580 ("invalid shadow object %p", object));
581
582 if (!VM_OBJECT_TRYWLOCK(object)) {
583 /*
584 * Prevent object from disappearing since we do not have a
585 * reference.
586 */
587 vm_object_pip_add(object, 1);
588 VM_OBJECT_WUNLOCK(backing_object);
589 VM_OBJECT_WLOCK(object);
590 vm_object_pip_wakeup(object);
591 } else
592 VM_OBJECT_WUNLOCK(backing_object);
593
594 /*
595 * Check for a collapse/terminate race with the last reference holder.
596 */
597 if ((object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) != 0 ||
598 !refcount_acquire_if_not_zero(&object->ref_count)) {
599 VM_OBJECT_WUNLOCK(object);
600 return (NULL);
601 }
602 backing_object = object->backing_object;
603 if (backing_object != NULL && (backing_object->flags & OBJ_ANON) != 0)
604 vm_object_collapse(object);
605 VM_OBJECT_WUNLOCK(object);
606
607 return (object);
608 }
609
610 /*
611 * vm_object_deallocate:
612 *
613 * Release a reference to the specified object,
614 * gained either through a vm_object_allocate
615 * or a vm_object_reference call. When all references
616 * are gone, storage associated with this object
617 * may be relinquished.
618 *
619 * No object may be locked.
620 */
621 void
vm_object_deallocate(vm_object_t object)622 vm_object_deallocate(vm_object_t object)
623 {
624 vm_object_t temp;
625 bool released;
626
627 while (object != NULL) {
628 /*
629 * If the reference count goes to 0 we start calling
630 * vm_object_terminate() on the object chain. A ref count
631 * of 1 may be a special case depending on the shadow count
632 * being 0 or 1. These cases require a write lock on the
633 * object.
634 */
635 if ((object->flags & OBJ_ANON) == 0)
636 released = refcount_release_if_gt(&object->ref_count, 1);
637 else
638 released = refcount_release_if_gt(&object->ref_count, 2);
639 if (released)
640 return;
641
642 if (object->type == OBJT_VNODE) {
643 VM_OBJECT_RLOCK(object);
644 if (object->type == OBJT_VNODE) {
645 vm_object_deallocate_vnode(object);
646 return;
647 }
648 VM_OBJECT_RUNLOCK(object);
649 }
650
651 VM_OBJECT_WLOCK(object);
652 KASSERT(object->ref_count > 0,
653 ("vm_object_deallocate: object deallocated too many times: %d",
654 object->type));
655
656 /*
657 * If this is not the final reference to an anonymous
658 * object we may need to collapse the shadow chain.
659 */
660 if (!refcount_release(&object->ref_count)) {
661 if (object->ref_count > 1 ||
662 atomic_load_int(&object->shadow_count) == 0) {
663 if ((object->flags & OBJ_ANON) != 0 &&
664 object->ref_count == 1)
665 vm_object_set_flag(object,
666 OBJ_ONEMAPPING);
667 VM_OBJECT_WUNLOCK(object);
668 return;
669 }
670
671 /* Handle collapsing last ref on anonymous objects. */
672 object = vm_object_deallocate_anon(object);
673 continue;
674 }
675
676 /*
677 * Handle the final reference to an object. We restart
678 * the loop with the backing object to avoid recursion.
679 */
680 umtx_shm_object_terminated(object);
681 temp = object->backing_object;
682 if (temp != NULL) {
683 KASSERT(object->type == OBJT_SWAP,
684 ("shadowed tmpfs v_object 2 %p", object));
685 vm_object_backing_remove(object);
686 }
687
688 KASSERT((object->flags & OBJ_DEAD) == 0,
689 ("vm_object_deallocate: Terminating dead object."));
690 vm_object_set_flag(object, OBJ_DEAD);
691 vm_object_terminate(object);
692 object = temp;
693 }
694 }
695
696 void
vm_object_destroy(vm_object_t object)697 vm_object_destroy(vm_object_t object)
698 {
699 uma_zfree(obj_zone, object);
700 }
701
702 static void
vm_object_sub_shadow(vm_object_t object)703 vm_object_sub_shadow(vm_object_t object)
704 {
705 KASSERT(object->shadow_count >= 1,
706 ("object %p sub_shadow count zero", object));
707 atomic_subtract_int(&object->shadow_count, 1);
708 }
709
710 static void
vm_object_backing_remove_locked(vm_object_t object)711 vm_object_backing_remove_locked(vm_object_t object)
712 {
713 vm_object_t backing_object;
714
715 backing_object = object->backing_object;
716 VM_OBJECT_ASSERT_WLOCKED(object);
717 VM_OBJECT_ASSERT_WLOCKED(backing_object);
718
719 KASSERT((object->flags & OBJ_COLLAPSING) == 0,
720 ("vm_object_backing_remove: Removing collapsing object."));
721
722 vm_object_sub_shadow(backing_object);
723 if ((object->flags & OBJ_SHADOWLIST) != 0) {
724 LIST_REMOVE(object, shadow_list);
725 vm_object_clear_flag(object, OBJ_SHADOWLIST);
726 }
727 object->backing_object = NULL;
728 }
729
730 static void
vm_object_backing_remove(vm_object_t object)731 vm_object_backing_remove(vm_object_t object)
732 {
733 vm_object_t backing_object;
734
735 VM_OBJECT_ASSERT_WLOCKED(object);
736
737 backing_object = object->backing_object;
738 if ((object->flags & OBJ_SHADOWLIST) != 0) {
739 VM_OBJECT_WLOCK(backing_object);
740 vm_object_backing_remove_locked(object);
741 VM_OBJECT_WUNLOCK(backing_object);
742 } else {
743 object->backing_object = NULL;
744 vm_object_sub_shadow(backing_object);
745 }
746 }
747
748 static void
vm_object_backing_insert_locked(vm_object_t object,vm_object_t backing_object)749 vm_object_backing_insert_locked(vm_object_t object, vm_object_t backing_object)
750 {
751
752 VM_OBJECT_ASSERT_WLOCKED(object);
753
754 atomic_add_int(&backing_object->shadow_count, 1);
755 if ((backing_object->flags & OBJ_ANON) != 0) {
756 VM_OBJECT_ASSERT_WLOCKED(backing_object);
757 LIST_INSERT_HEAD(&backing_object->shadow_head, object,
758 shadow_list);
759 vm_object_set_flag(object, OBJ_SHADOWLIST);
760 }
761 object->backing_object = backing_object;
762 }
763
764 static void
vm_object_backing_insert(vm_object_t object,vm_object_t backing_object)765 vm_object_backing_insert(vm_object_t object, vm_object_t backing_object)
766 {
767
768 VM_OBJECT_ASSERT_WLOCKED(object);
769
770 if ((backing_object->flags & OBJ_ANON) != 0) {
771 VM_OBJECT_WLOCK(backing_object);
772 vm_object_backing_insert_locked(object, backing_object);
773 VM_OBJECT_WUNLOCK(backing_object);
774 } else {
775 object->backing_object = backing_object;
776 atomic_add_int(&backing_object->shadow_count, 1);
777 }
778 }
779
780 /*
781 * Insert an object into a backing_object's shadow list with an additional
782 * reference to the backing_object added.
783 */
784 static void
vm_object_backing_insert_ref(vm_object_t object,vm_object_t backing_object)785 vm_object_backing_insert_ref(vm_object_t object, vm_object_t backing_object)
786 {
787
788 VM_OBJECT_ASSERT_WLOCKED(object);
789
790 if ((backing_object->flags & OBJ_ANON) != 0) {
791 VM_OBJECT_WLOCK(backing_object);
792 KASSERT((backing_object->flags & OBJ_DEAD) == 0,
793 ("shadowing dead anonymous object"));
794 vm_object_reference_locked(backing_object);
795 vm_object_backing_insert_locked(object, backing_object);
796 vm_object_clear_flag(backing_object, OBJ_ONEMAPPING);
797 VM_OBJECT_WUNLOCK(backing_object);
798 } else {
799 vm_object_reference(backing_object);
800 atomic_add_int(&backing_object->shadow_count, 1);
801 object->backing_object = backing_object;
802 }
803 }
804
805 /*
806 * Transfer a backing reference from backing_object to object.
807 */
808 static void
vm_object_backing_transfer(vm_object_t object,vm_object_t backing_object)809 vm_object_backing_transfer(vm_object_t object, vm_object_t backing_object)
810 {
811 vm_object_t new_backing_object;
812
813 /*
814 * Note that the reference to backing_object->backing_object
815 * moves from within backing_object to within object.
816 */
817 vm_object_backing_remove_locked(object);
818 new_backing_object = backing_object->backing_object;
819 if (new_backing_object == NULL)
820 return;
821 if ((new_backing_object->flags & OBJ_ANON) != 0) {
822 VM_OBJECT_WLOCK(new_backing_object);
823 vm_object_backing_remove_locked(backing_object);
824 vm_object_backing_insert_locked(object, new_backing_object);
825 VM_OBJECT_WUNLOCK(new_backing_object);
826 } else {
827 /*
828 * shadow_count for new_backing_object is left
829 * unchanged, its reference provided by backing_object
830 * is replaced by object.
831 */
832 object->backing_object = new_backing_object;
833 backing_object->backing_object = NULL;
834 }
835 }
836
837 /*
838 * Wait for a concurrent collapse to settle.
839 */
840 static void
vm_object_collapse_wait(vm_object_t object)841 vm_object_collapse_wait(vm_object_t object)
842 {
843
844 VM_OBJECT_ASSERT_WLOCKED(object);
845
846 while ((object->flags & OBJ_COLLAPSING) != 0) {
847 vm_object_pip_wait(object, "vmcolwait");
848 counter_u64_add(object_collapse_waits, 1);
849 }
850 }
851
852 /*
853 * Waits for a backing object to clear a pending collapse and returns
854 * it locked if it is an ANON object.
855 */
856 static vm_object_t
vm_object_backing_collapse_wait(vm_object_t object)857 vm_object_backing_collapse_wait(vm_object_t object)
858 {
859 vm_object_t backing_object;
860
861 VM_OBJECT_ASSERT_WLOCKED(object);
862
863 for (;;) {
864 backing_object = object->backing_object;
865 if (backing_object == NULL ||
866 (backing_object->flags & OBJ_ANON) == 0)
867 return (NULL);
868 VM_OBJECT_WLOCK(backing_object);
869 if ((backing_object->flags & (OBJ_DEAD | OBJ_COLLAPSING)) == 0)
870 break;
871 VM_OBJECT_WUNLOCK(object);
872 vm_object_pip_sleep(backing_object, "vmbckwait");
873 counter_u64_add(object_collapse_waits, 1);
874 VM_OBJECT_WLOCK(object);
875 }
876 return (backing_object);
877 }
878
879 /*
880 * vm_object_terminate_pages removes any remaining pageable pages
881 * from the object and resets the object to an empty state.
882 */
883 static void
vm_object_terminate_pages(vm_object_t object)884 vm_object_terminate_pages(vm_object_t object)
885 {
886 vm_page_t p, p_next;
887
888 VM_OBJECT_ASSERT_WLOCKED(object);
889
890 /*
891 * Free any remaining pageable pages. This also removes them from the
892 * paging queues. However, don't free wired pages, just remove them
893 * from the object. Rather than incrementally removing each page from
894 * the object, the page and object are reset to any empty state.
895 */
896 TAILQ_FOREACH_SAFE(p, &object->memq, listq, p_next) {
897 vm_page_assert_unbusied(p);
898 KASSERT(p->object == object &&
899 (p->ref_count & VPRC_OBJREF) != 0,
900 ("vm_object_terminate_pages: page %p is inconsistent", p));
901
902 p->object = NULL;
903 if (vm_page_drop(p, VPRC_OBJREF) == VPRC_OBJREF) {
904 VM_CNT_INC(v_pfree);
905 vm_page_free(p);
906 }
907 }
908
909 /*
910 * If the object contained any pages, then reset it to an empty state.
911 * None of the object's fields, including "resident_page_count", were
912 * modified by the preceding loop.
913 */
914 if (object->resident_page_count != 0) {
915 vm_radix_reclaim_allnodes(&object->rtree);
916 TAILQ_INIT(&object->memq);
917 object->resident_page_count = 0;
918 if (object->type == OBJT_VNODE)
919 vdrop(object->handle);
920 }
921 }
922
923 /*
924 * vm_object_terminate actually destroys the specified object, freeing
925 * up all previously used resources.
926 *
927 * The object must be locked.
928 * This routine may block.
929 */
930 void
vm_object_terminate(vm_object_t object)931 vm_object_terminate(vm_object_t object)
932 {
933
934 VM_OBJECT_ASSERT_WLOCKED(object);
935 KASSERT((object->flags & OBJ_DEAD) != 0,
936 ("terminating non-dead obj %p", object));
937 KASSERT((object->flags & OBJ_COLLAPSING) == 0,
938 ("terminating collapsing obj %p", object));
939 KASSERT(object->backing_object == NULL,
940 ("terminating shadow obj %p", object));
941
942 /*
943 * Wait for the pageout daemon and other current users to be
944 * done with the object. Note that new paging_in_progress
945 * users can come after this wait, but they must check
946 * OBJ_DEAD flag set (without unlocking the object), and avoid
947 * the object being terminated.
948 */
949 vm_object_pip_wait(object, "objtrm");
950
951 KASSERT(object->ref_count == 0,
952 ("vm_object_terminate: object with references, ref_count=%d",
953 object->ref_count));
954
955 if ((object->flags & OBJ_PG_DTOR) == 0)
956 vm_object_terminate_pages(object);
957
958 #if VM_NRESERVLEVEL > 0
959 if (__predict_false(!LIST_EMPTY(&object->rvq)))
960 vm_reserv_break_all(object);
961 #endif
962
963 KASSERT(object->cred == NULL || (object->flags & OBJ_SWAP) != 0,
964 ("%s: non-swap obj %p has cred", __func__, object));
965
966 /*
967 * Let the pager know object is dead.
968 */
969 vm_pager_deallocate(object);
970 VM_OBJECT_WUNLOCK(object);
971
972 vm_object_destroy(object);
973 }
974
975 /*
976 * Make the page read-only so that we can clear the object flags. However, if
977 * this is a nosync mmap then the object is likely to stay dirty so do not
978 * mess with the page and do not clear the object flags. Returns TRUE if the
979 * page should be flushed, and FALSE otherwise.
980 */
981 static boolean_t
vm_object_page_remove_write(vm_page_t p,int flags,boolean_t * allclean)982 vm_object_page_remove_write(vm_page_t p, int flags, boolean_t *allclean)
983 {
984
985 vm_page_assert_busied(p);
986
987 /*
988 * If we have been asked to skip nosync pages and this is a
989 * nosync page, skip it. Note that the object flags were not
990 * cleared in this case so we do not have to set them.
991 */
992 if ((flags & OBJPC_NOSYNC) != 0 && (p->a.flags & PGA_NOSYNC) != 0) {
993 *allclean = FALSE;
994 return (FALSE);
995 } else {
996 pmap_remove_write(p);
997 return (p->dirty != 0);
998 }
999 }
1000
1001 /*
1002 * vm_object_page_clean
1003 *
1004 * Clean all dirty pages in the specified range of object. Leaves page
1005 * on whatever queue it is currently on. If NOSYNC is set then do not
1006 * write out pages with PGA_NOSYNC set (originally comes from MAP_NOSYNC),
1007 * leaving the object dirty.
1008 *
1009 * For swap objects backing tmpfs regular files, do not flush anything,
1010 * but remove write protection on the mapped pages to update mtime through
1011 * mmaped writes.
1012 *
1013 * When stuffing pages asynchronously, allow clustering. XXX we need a
1014 * synchronous clustering mode implementation.
1015 *
1016 * Odd semantics: if start == end, we clean everything.
1017 *
1018 * The object must be locked.
1019 *
1020 * Returns FALSE if some page from the range was not written, as
1021 * reported by the pager, and TRUE otherwise.
1022 */
1023 boolean_t
vm_object_page_clean(vm_object_t object,vm_ooffset_t start,vm_ooffset_t end,int flags)1024 vm_object_page_clean(vm_object_t object, vm_ooffset_t start, vm_ooffset_t end,
1025 int flags)
1026 {
1027 vm_page_t np, p;
1028 vm_pindex_t pi, tend, tstart;
1029 int curgeneration, n, pagerflags;
1030 boolean_t eio, res, allclean;
1031
1032 VM_OBJECT_ASSERT_WLOCKED(object);
1033
1034 if (!vm_object_mightbedirty(object) || object->resident_page_count == 0)
1035 return (TRUE);
1036
1037 pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) != 0 ?
1038 VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1039 pagerflags |= (flags & OBJPC_INVAL) != 0 ? VM_PAGER_PUT_INVAL : 0;
1040
1041 tstart = OFF_TO_IDX(start);
1042 tend = (end == 0) ? object->size : OFF_TO_IDX(end + PAGE_MASK);
1043 allclean = tstart == 0 && tend >= object->size;
1044 res = TRUE;
1045
1046 rescan:
1047 curgeneration = object->generation;
1048
1049 for (p = vm_page_find_least(object, tstart); p != NULL; p = np) {
1050 pi = p->pindex;
1051 if (pi >= tend)
1052 break;
1053 np = TAILQ_NEXT(p, listq);
1054 if (vm_page_none_valid(p))
1055 continue;
1056 if (vm_page_busy_acquire(p, VM_ALLOC_WAITFAIL) == 0) {
1057 if (object->generation != curgeneration &&
1058 (flags & OBJPC_SYNC) != 0)
1059 goto rescan;
1060 np = vm_page_find_least(object, pi);
1061 continue;
1062 }
1063 if (!vm_object_page_remove_write(p, flags, &allclean)) {
1064 vm_page_xunbusy(p);
1065 continue;
1066 }
1067 if (object->type == OBJT_VNODE) {
1068 n = vm_object_page_collect_flush(object, p, pagerflags,
1069 flags, &allclean, &eio);
1070 if (eio) {
1071 res = FALSE;
1072 allclean = FALSE;
1073 }
1074 if (object->generation != curgeneration &&
1075 (flags & OBJPC_SYNC) != 0)
1076 goto rescan;
1077
1078 /*
1079 * If the VOP_PUTPAGES() did a truncated write, so
1080 * that even the first page of the run is not fully
1081 * written, vm_pageout_flush() returns 0 as the run
1082 * length. Since the condition that caused truncated
1083 * write may be permanent, e.g. exhausted free space,
1084 * accepting n == 0 would cause an infinite loop.
1085 *
1086 * Forwarding the iterator leaves the unwritten page
1087 * behind, but there is not much we can do there if
1088 * filesystem refuses to write it.
1089 */
1090 if (n == 0) {
1091 n = 1;
1092 allclean = FALSE;
1093 }
1094 } else {
1095 n = 1;
1096 vm_page_xunbusy(p);
1097 }
1098 np = vm_page_find_least(object, pi + n);
1099 }
1100 #if 0
1101 VOP_FSYNC(vp, (pagerflags & VM_PAGER_PUT_SYNC) ? MNT_WAIT : 0);
1102 #endif
1103
1104 /*
1105 * Leave updating cleangeneration for tmpfs objects to tmpfs
1106 * scan. It needs to update mtime, which happens for other
1107 * filesystems during page writeouts.
1108 */
1109 if (allclean && object->type == OBJT_VNODE)
1110 object->cleangeneration = curgeneration;
1111 return (res);
1112 }
1113
1114 static int
vm_object_page_collect_flush(vm_object_t object,vm_page_t p,int pagerflags,int flags,boolean_t * allclean,boolean_t * eio)1115 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags,
1116 int flags, boolean_t *allclean, boolean_t *eio)
1117 {
1118 vm_page_t ma[vm_pageout_page_count], p_first, tp;
1119 int count, i, mreq, runlen;
1120
1121 vm_page_lock_assert(p, MA_NOTOWNED);
1122 vm_page_assert_xbusied(p);
1123 VM_OBJECT_ASSERT_WLOCKED(object);
1124
1125 count = 1;
1126 mreq = 0;
1127
1128 for (tp = p; count < vm_pageout_page_count; count++) {
1129 tp = vm_page_next(tp);
1130 if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1131 break;
1132 if (!vm_object_page_remove_write(tp, flags, allclean)) {
1133 vm_page_xunbusy(tp);
1134 break;
1135 }
1136 }
1137
1138 for (p_first = p; count < vm_pageout_page_count; count++) {
1139 tp = vm_page_prev(p_first);
1140 if (tp == NULL || vm_page_tryxbusy(tp) == 0)
1141 break;
1142 if (!vm_object_page_remove_write(tp, flags, allclean)) {
1143 vm_page_xunbusy(tp);
1144 break;
1145 }
1146 p_first = tp;
1147 mreq++;
1148 }
1149
1150 for (tp = p_first, i = 0; i < count; tp = TAILQ_NEXT(tp, listq), i++)
1151 ma[i] = tp;
1152
1153 vm_pageout_flush(ma, count, pagerflags, mreq, &runlen, eio);
1154 return (runlen);
1155 }
1156
1157 /*
1158 * Note that there is absolutely no sense in writing out
1159 * anonymous objects, so we track down the vnode object
1160 * to write out.
1161 * We invalidate (remove) all pages from the address space
1162 * for semantic correctness.
1163 *
1164 * If the backing object is a device object with unmanaged pages, then any
1165 * mappings to the specified range of pages must be removed before this
1166 * function is called.
1167 *
1168 * Note: certain anonymous maps, such as MAP_NOSYNC maps,
1169 * may start out with a NULL object.
1170 */
1171 boolean_t
vm_object_sync(vm_object_t object,vm_ooffset_t offset,vm_size_t size,boolean_t syncio,boolean_t invalidate)1172 vm_object_sync(vm_object_t object, vm_ooffset_t offset, vm_size_t size,
1173 boolean_t syncio, boolean_t invalidate)
1174 {
1175 vm_object_t backing_object;
1176 struct vnode *vp;
1177 struct mount *mp;
1178 int error, flags, fsync_after;
1179 boolean_t res;
1180
1181 if (object == NULL)
1182 return (TRUE);
1183 res = TRUE;
1184 error = 0;
1185 VM_OBJECT_WLOCK(object);
1186 while ((backing_object = object->backing_object) != NULL) {
1187 VM_OBJECT_WLOCK(backing_object);
1188 offset += object->backing_object_offset;
1189 VM_OBJECT_WUNLOCK(object);
1190 object = backing_object;
1191 if (object->size < OFF_TO_IDX(offset + size))
1192 size = IDX_TO_OFF(object->size) - offset;
1193 }
1194 /*
1195 * Flush pages if writing is allowed, invalidate them
1196 * if invalidation requested. Pages undergoing I/O
1197 * will be ignored by vm_object_page_remove().
1198 *
1199 * We cannot lock the vnode and then wait for paging
1200 * to complete without deadlocking against vm_fault.
1201 * Instead we simply call vm_object_page_remove() and
1202 * allow it to block internally on a page-by-page
1203 * basis when it encounters pages undergoing async
1204 * I/O.
1205 */
1206 if (object->type == OBJT_VNODE &&
1207 vm_object_mightbedirty(object) != 0 &&
1208 ((vp = object->handle)->v_vflag & VV_NOSYNC) == 0) {
1209 VM_OBJECT_WUNLOCK(object);
1210 (void)vn_start_write(vp, &mp, V_WAIT);
1211 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1212 if (syncio && !invalidate && offset == 0 &&
1213 atop(size) == object->size) {
1214 /*
1215 * If syncing the whole mapping of the file,
1216 * it is faster to schedule all the writes in
1217 * async mode, also allowing the clustering,
1218 * and then wait for i/o to complete.
1219 */
1220 flags = 0;
1221 fsync_after = TRUE;
1222 } else {
1223 flags = (syncio || invalidate) ? OBJPC_SYNC : 0;
1224 flags |= invalidate ? (OBJPC_SYNC | OBJPC_INVAL) : 0;
1225 fsync_after = FALSE;
1226 }
1227 VM_OBJECT_WLOCK(object);
1228 res = vm_object_page_clean(object, offset, offset + size,
1229 flags);
1230 VM_OBJECT_WUNLOCK(object);
1231 if (fsync_after) {
1232 for (;;) {
1233 error = VOP_FSYNC(vp, MNT_WAIT, curthread);
1234 if (error != ERELOOKUP)
1235 break;
1236
1237 /*
1238 * Allow SU/bufdaemon to handle more
1239 * dependencies in the meantime.
1240 */
1241 VOP_UNLOCK(vp);
1242 vn_finished_write(mp);
1243
1244 (void)vn_start_write(vp, &mp, V_WAIT);
1245 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1246 }
1247 }
1248 VOP_UNLOCK(vp);
1249 vn_finished_write(mp);
1250 if (error != 0)
1251 res = FALSE;
1252 VM_OBJECT_WLOCK(object);
1253 }
1254 if ((object->type == OBJT_VNODE ||
1255 object->type == OBJT_DEVICE) && invalidate) {
1256 if (object->type == OBJT_DEVICE)
1257 /*
1258 * The option OBJPR_NOTMAPPED must be passed here
1259 * because vm_object_page_remove() cannot remove
1260 * unmanaged mappings.
1261 */
1262 flags = OBJPR_NOTMAPPED;
1263 else if (old_msync)
1264 flags = 0;
1265 else
1266 flags = OBJPR_CLEANONLY;
1267 vm_object_page_remove(object, OFF_TO_IDX(offset),
1268 OFF_TO_IDX(offset + size + PAGE_MASK), flags);
1269 }
1270 VM_OBJECT_WUNLOCK(object);
1271 return (res);
1272 }
1273
1274 /*
1275 * Determine whether the given advice can be applied to the object. Advice is
1276 * not applied to unmanaged pages since they never belong to page queues, and
1277 * since MADV_FREE is destructive, it can apply only to anonymous pages that
1278 * have been mapped at most once.
1279 */
1280 static bool
vm_object_advice_applies(vm_object_t object,int advice)1281 vm_object_advice_applies(vm_object_t object, int advice)
1282 {
1283
1284 if ((object->flags & OBJ_UNMANAGED) != 0)
1285 return (false);
1286 if (advice != MADV_FREE)
1287 return (true);
1288 return ((object->flags & (OBJ_ONEMAPPING | OBJ_ANON)) ==
1289 (OBJ_ONEMAPPING | OBJ_ANON));
1290 }
1291
1292 static void
vm_object_madvise_freespace(vm_object_t object,int advice,vm_pindex_t pindex,vm_size_t size)1293 vm_object_madvise_freespace(vm_object_t object, int advice, vm_pindex_t pindex,
1294 vm_size_t size)
1295 {
1296
1297 if (advice == MADV_FREE)
1298 vm_pager_freespace(object, pindex, size);
1299 }
1300
1301 /*
1302 * vm_object_madvise:
1303 *
1304 * Implements the madvise function at the object/page level.
1305 *
1306 * MADV_WILLNEED (any object)
1307 *
1308 * Activate the specified pages if they are resident.
1309 *
1310 * MADV_DONTNEED (any object)
1311 *
1312 * Deactivate the specified pages if they are resident.
1313 *
1314 * MADV_FREE (OBJT_SWAP objects, OBJ_ONEMAPPING only)
1315 *
1316 * Deactivate and clean the specified pages if they are
1317 * resident. This permits the process to reuse the pages
1318 * without faulting or the kernel to reclaim the pages
1319 * without I/O.
1320 */
1321 void
vm_object_madvise(vm_object_t object,vm_pindex_t pindex,vm_pindex_t end,int advice)1322 vm_object_madvise(vm_object_t object, vm_pindex_t pindex, vm_pindex_t end,
1323 int advice)
1324 {
1325 vm_pindex_t tpindex;
1326 vm_object_t backing_object, tobject;
1327 vm_page_t m, tm;
1328
1329 if (object == NULL)
1330 return;
1331
1332 relookup:
1333 VM_OBJECT_WLOCK(object);
1334 if (!vm_object_advice_applies(object, advice)) {
1335 VM_OBJECT_WUNLOCK(object);
1336 return;
1337 }
1338 for (m = vm_page_find_least(object, pindex); pindex < end; pindex++) {
1339 tobject = object;
1340
1341 /*
1342 * If the next page isn't resident in the top-level object, we
1343 * need to search the shadow chain. When applying MADV_FREE, we
1344 * take care to release any swap space used to store
1345 * non-resident pages.
1346 */
1347 if (m == NULL || pindex < m->pindex) {
1348 /*
1349 * Optimize a common case: if the top-level object has
1350 * no backing object, we can skip over the non-resident
1351 * range in constant time.
1352 */
1353 if (object->backing_object == NULL) {
1354 tpindex = (m != NULL && m->pindex < end) ?
1355 m->pindex : end;
1356 vm_object_madvise_freespace(object, advice,
1357 pindex, tpindex - pindex);
1358 if ((pindex = tpindex) == end)
1359 break;
1360 goto next_page;
1361 }
1362
1363 tpindex = pindex;
1364 do {
1365 vm_object_madvise_freespace(tobject, advice,
1366 tpindex, 1);
1367 /*
1368 * Prepare to search the next object in the
1369 * chain.
1370 */
1371 backing_object = tobject->backing_object;
1372 if (backing_object == NULL)
1373 goto next_pindex;
1374 VM_OBJECT_WLOCK(backing_object);
1375 tpindex +=
1376 OFF_TO_IDX(tobject->backing_object_offset);
1377 if (tobject != object)
1378 VM_OBJECT_WUNLOCK(tobject);
1379 tobject = backing_object;
1380 if (!vm_object_advice_applies(tobject, advice))
1381 goto next_pindex;
1382 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
1383 NULL);
1384 } else {
1385 next_page:
1386 tm = m;
1387 m = TAILQ_NEXT(m, listq);
1388 }
1389
1390 /*
1391 * If the page is not in a normal state, skip it. The page
1392 * can not be invalidated while the object lock is held.
1393 */
1394 if (!vm_page_all_valid(tm) || vm_page_wired(tm))
1395 goto next_pindex;
1396 KASSERT((tm->flags & PG_FICTITIOUS) == 0,
1397 ("vm_object_madvise: page %p is fictitious", tm));
1398 KASSERT((tm->oflags & VPO_UNMANAGED) == 0,
1399 ("vm_object_madvise: page %p is not managed", tm));
1400 if (vm_page_tryxbusy(tm) == 0) {
1401 if (object != tobject)
1402 VM_OBJECT_WUNLOCK(object);
1403 if (advice == MADV_WILLNEED) {
1404 /*
1405 * Reference the page before unlocking and
1406 * sleeping so that the page daemon is less
1407 * likely to reclaim it.
1408 */
1409 vm_page_aflag_set(tm, PGA_REFERENCED);
1410 }
1411 if (!vm_page_busy_sleep(tm, "madvpo", 0))
1412 VM_OBJECT_WUNLOCK(tobject);
1413 goto relookup;
1414 }
1415 vm_page_advise(tm, advice);
1416 vm_page_xunbusy(tm);
1417 vm_object_madvise_freespace(tobject, advice, tm->pindex, 1);
1418 next_pindex:
1419 if (tobject != object)
1420 VM_OBJECT_WUNLOCK(tobject);
1421 }
1422 VM_OBJECT_WUNLOCK(object);
1423 }
1424
1425 /*
1426 * vm_object_shadow:
1427 *
1428 * Create a new object which is backed by the
1429 * specified existing object range. The source
1430 * object reference is deallocated.
1431 *
1432 * The new object and offset into that object
1433 * are returned in the source parameters.
1434 */
1435 void
vm_object_shadow(vm_object_t * object,vm_ooffset_t * offset,vm_size_t length,struct ucred * cred,bool shared)1436 vm_object_shadow(vm_object_t *object, vm_ooffset_t *offset, vm_size_t length,
1437 struct ucred *cred, bool shared)
1438 {
1439 vm_object_t source;
1440 vm_object_t result;
1441
1442 source = *object;
1443
1444 /*
1445 * Don't create the new object if the old object isn't shared.
1446 *
1447 * If we hold the only reference we can guarantee that it won't
1448 * increase while we have the map locked. Otherwise the race is
1449 * harmless and we will end up with an extra shadow object that
1450 * will be collapsed later.
1451 */
1452 if (source != NULL && source->ref_count == 1 &&
1453 (source->flags & OBJ_ANON) != 0)
1454 return;
1455
1456 /*
1457 * Allocate a new object with the given length.
1458 */
1459 result = vm_object_allocate_anon(atop(length), source, cred, length);
1460
1461 /*
1462 * Store the offset into the source object, and fix up the offset into
1463 * the new object.
1464 */
1465 result->backing_object_offset = *offset;
1466
1467 if (shared || source != NULL) {
1468 VM_OBJECT_WLOCK(result);
1469
1470 /*
1471 * The new object shadows the source object, adding a
1472 * reference to it. Our caller changes his reference
1473 * to point to the new object, removing a reference to
1474 * the source object. Net result: no change of
1475 * reference count, unless the caller needs to add one
1476 * more reference due to forking a shared map entry.
1477 */
1478 if (shared) {
1479 vm_object_reference_locked(result);
1480 vm_object_clear_flag(result, OBJ_ONEMAPPING);
1481 }
1482
1483 /*
1484 * Try to optimize the result object's page color when
1485 * shadowing in order to maintain page coloring
1486 * consistency in the combined shadowed object.
1487 */
1488 if (source != NULL) {
1489 vm_object_backing_insert(result, source);
1490 result->domain = source->domain;
1491 #if VM_NRESERVLEVEL > 0
1492 vm_object_set_flag(result,
1493 (source->flags & OBJ_COLORED));
1494 result->pg_color = (source->pg_color +
1495 OFF_TO_IDX(*offset)) & ((1 << (VM_NFREEORDER -
1496 1)) - 1);
1497 #endif
1498 }
1499 VM_OBJECT_WUNLOCK(result);
1500 }
1501
1502 /*
1503 * Return the new things
1504 */
1505 *offset = 0;
1506 *object = result;
1507 }
1508
1509 /*
1510 * vm_object_split:
1511 *
1512 * Split the pages in a map entry into a new object. This affords
1513 * easier removal of unused pages, and keeps object inheritance from
1514 * being a negative impact on memory usage.
1515 */
1516 void
vm_object_split(vm_map_entry_t entry)1517 vm_object_split(vm_map_entry_t entry)
1518 {
1519 vm_page_t m, m_next;
1520 vm_object_t orig_object, new_object, backing_object;
1521 vm_pindex_t idx, offidxstart;
1522 vm_size_t size;
1523
1524 orig_object = entry->object.vm_object;
1525 KASSERT((orig_object->flags & OBJ_ONEMAPPING) != 0,
1526 ("vm_object_split: Splitting object with multiple mappings."));
1527 if ((orig_object->flags & OBJ_ANON) == 0)
1528 return;
1529 if (orig_object->ref_count <= 1)
1530 return;
1531 VM_OBJECT_WUNLOCK(orig_object);
1532
1533 offidxstart = OFF_TO_IDX(entry->offset);
1534 size = atop(entry->end - entry->start);
1535
1536 new_object = vm_object_allocate_anon(size, orig_object,
1537 orig_object->cred, ptoa(size));
1538
1539 /*
1540 * We must wait for the orig_object to complete any in-progress
1541 * collapse so that the swap blocks are stable below. The
1542 * additional reference on backing_object by new object will
1543 * prevent further collapse operations until split completes.
1544 */
1545 VM_OBJECT_WLOCK(orig_object);
1546 vm_object_collapse_wait(orig_object);
1547
1548 /*
1549 * At this point, the new object is still private, so the order in
1550 * which the original and new objects are locked does not matter.
1551 */
1552 VM_OBJECT_WLOCK(new_object);
1553 new_object->domain = orig_object->domain;
1554 backing_object = orig_object->backing_object;
1555 if (backing_object != NULL) {
1556 vm_object_backing_insert_ref(new_object, backing_object);
1557 new_object->backing_object_offset =
1558 orig_object->backing_object_offset + entry->offset;
1559 }
1560 if (orig_object->cred != NULL) {
1561 crhold(orig_object->cred);
1562 KASSERT(orig_object->charge >= ptoa(size),
1563 ("orig_object->charge < 0"));
1564 orig_object->charge -= ptoa(size);
1565 }
1566
1567 /*
1568 * Mark the split operation so that swap_pager_getpages() knows
1569 * that the object is in transition.
1570 */
1571 vm_object_set_flag(orig_object, OBJ_SPLIT);
1572 #ifdef INVARIANTS
1573 idx = 0;
1574 #endif
1575 retry:
1576 m = vm_page_find_least(orig_object, offidxstart);
1577 KASSERT(m == NULL || idx <= m->pindex - offidxstart,
1578 ("%s: object %p was repopulated", __func__, orig_object));
1579 for (; m != NULL && (idx = m->pindex - offidxstart) < size;
1580 m = m_next) {
1581 m_next = TAILQ_NEXT(m, listq);
1582
1583 /*
1584 * We must wait for pending I/O to complete before we can
1585 * rename the page.
1586 *
1587 * We do not have to VM_PROT_NONE the page as mappings should
1588 * not be changed by this operation.
1589 */
1590 if (vm_page_tryxbusy(m) == 0) {
1591 VM_OBJECT_WUNLOCK(new_object);
1592 if (vm_page_busy_sleep(m, "spltwt", 0))
1593 VM_OBJECT_WLOCK(orig_object);
1594 VM_OBJECT_WLOCK(new_object);
1595 goto retry;
1596 }
1597
1598 /*
1599 * The page was left invalid. Likely placed there by
1600 * an incomplete fault. Just remove and ignore.
1601 */
1602 if (vm_page_none_valid(m)) {
1603 if (vm_page_remove(m))
1604 vm_page_free(m);
1605 continue;
1606 }
1607
1608 /* vm_page_rename() will dirty the page. */
1609 if (vm_page_rename(m, new_object, idx)) {
1610 vm_page_xunbusy(m);
1611 VM_OBJECT_WUNLOCK(new_object);
1612 VM_OBJECT_WUNLOCK(orig_object);
1613 vm_radix_wait();
1614 VM_OBJECT_WLOCK(orig_object);
1615 VM_OBJECT_WLOCK(new_object);
1616 goto retry;
1617 }
1618
1619 #if VM_NRESERVLEVEL > 0
1620 /*
1621 * If some of the reservation's allocated pages remain with
1622 * the original object, then transferring the reservation to
1623 * the new object is neither particularly beneficial nor
1624 * particularly harmful as compared to leaving the reservation
1625 * with the original object. If, however, all of the
1626 * reservation's allocated pages are transferred to the new
1627 * object, then transferring the reservation is typically
1628 * beneficial. Determining which of these two cases applies
1629 * would be more costly than unconditionally renaming the
1630 * reservation.
1631 */
1632 vm_reserv_rename(m, new_object, orig_object, offidxstart);
1633 #endif
1634 }
1635
1636 /*
1637 * swap_pager_copy() can sleep, in which case the orig_object's
1638 * and new_object's locks are released and reacquired.
1639 */
1640 swap_pager_copy(orig_object, new_object, offidxstart, 0);
1641
1642 TAILQ_FOREACH(m, &new_object->memq, listq)
1643 vm_page_xunbusy(m);
1644
1645 vm_object_clear_flag(orig_object, OBJ_SPLIT);
1646 VM_OBJECT_WUNLOCK(orig_object);
1647 VM_OBJECT_WUNLOCK(new_object);
1648 entry->object.vm_object = new_object;
1649 entry->offset = 0LL;
1650 vm_object_deallocate(orig_object);
1651 VM_OBJECT_WLOCK(new_object);
1652 }
1653
1654 static vm_page_t
vm_object_collapse_scan_wait(vm_object_t object,vm_page_t p)1655 vm_object_collapse_scan_wait(vm_object_t object, vm_page_t p)
1656 {
1657 vm_object_t backing_object;
1658
1659 VM_OBJECT_ASSERT_WLOCKED(object);
1660 backing_object = object->backing_object;
1661 VM_OBJECT_ASSERT_WLOCKED(backing_object);
1662
1663 KASSERT(p == NULL || p->object == object || p->object == backing_object,
1664 ("invalid ownership %p %p %p", p, object, backing_object));
1665 /* The page is only NULL when rename fails. */
1666 if (p == NULL) {
1667 VM_OBJECT_WUNLOCK(object);
1668 VM_OBJECT_WUNLOCK(backing_object);
1669 vm_radix_wait();
1670 VM_OBJECT_WLOCK(object);
1671 } else if (p->object == object) {
1672 VM_OBJECT_WUNLOCK(backing_object);
1673 if (vm_page_busy_sleep(p, "vmocol", 0))
1674 VM_OBJECT_WLOCK(object);
1675 } else {
1676 VM_OBJECT_WUNLOCK(object);
1677 if (!vm_page_busy_sleep(p, "vmocol", 0))
1678 VM_OBJECT_WUNLOCK(backing_object);
1679 VM_OBJECT_WLOCK(object);
1680 }
1681 VM_OBJECT_WLOCK(backing_object);
1682 return (TAILQ_FIRST(&backing_object->memq));
1683 }
1684
1685 static bool
vm_object_scan_all_shadowed(vm_object_t object)1686 vm_object_scan_all_shadowed(vm_object_t object)
1687 {
1688 vm_object_t backing_object;
1689 vm_page_t p, pp;
1690 vm_pindex_t backing_offset_index, new_pindex, pi, ps;
1691
1692 VM_OBJECT_ASSERT_WLOCKED(object);
1693 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1694
1695 backing_object = object->backing_object;
1696
1697 if ((backing_object->flags & OBJ_ANON) == 0)
1698 return (false);
1699
1700 pi = backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1701 p = vm_page_find_least(backing_object, pi);
1702 ps = swap_pager_find_least(backing_object, pi);
1703
1704 /*
1705 * Only check pages inside the parent object's range and
1706 * inside the parent object's mapping of the backing object.
1707 */
1708 for (;; pi++) {
1709 if (p != NULL && p->pindex < pi)
1710 p = TAILQ_NEXT(p, listq);
1711 if (ps < pi)
1712 ps = swap_pager_find_least(backing_object, pi);
1713 if (p == NULL && ps >= backing_object->size)
1714 break;
1715 else if (p == NULL)
1716 pi = ps;
1717 else
1718 pi = MIN(p->pindex, ps);
1719
1720 new_pindex = pi - backing_offset_index;
1721 if (new_pindex >= object->size)
1722 break;
1723
1724 if (p != NULL) {
1725 /*
1726 * If the backing object page is busy a
1727 * grandparent or older page may still be
1728 * undergoing CoW. It is not safe to collapse
1729 * the backing object until it is quiesced.
1730 */
1731 if (vm_page_tryxbusy(p) == 0)
1732 return (false);
1733
1734 /*
1735 * We raced with the fault handler that left
1736 * newly allocated invalid page on the object
1737 * queue and retried.
1738 */
1739 if (!vm_page_all_valid(p))
1740 goto unbusy_ret;
1741 }
1742
1743 /*
1744 * See if the parent has the page or if the parent's object
1745 * pager has the page. If the parent has the page but the page
1746 * is not valid, the parent's object pager must have the page.
1747 *
1748 * If this fails, the parent does not completely shadow the
1749 * object and we might as well give up now.
1750 */
1751 pp = vm_page_lookup(object, new_pindex);
1752
1753 /*
1754 * The valid check here is stable due to object lock
1755 * being required to clear valid and initiate paging.
1756 * Busy of p disallows fault handler to validate pp.
1757 */
1758 if ((pp == NULL || vm_page_none_valid(pp)) &&
1759 !vm_pager_has_page(object, new_pindex, NULL, NULL))
1760 goto unbusy_ret;
1761 if (p != NULL)
1762 vm_page_xunbusy(p);
1763 }
1764 return (true);
1765
1766 unbusy_ret:
1767 if (p != NULL)
1768 vm_page_xunbusy(p);
1769 return (false);
1770 }
1771
1772 static void
vm_object_collapse_scan(vm_object_t object)1773 vm_object_collapse_scan(vm_object_t object)
1774 {
1775 vm_object_t backing_object;
1776 vm_page_t next, p, pp;
1777 vm_pindex_t backing_offset_index, new_pindex;
1778
1779 VM_OBJECT_ASSERT_WLOCKED(object);
1780 VM_OBJECT_ASSERT_WLOCKED(object->backing_object);
1781
1782 backing_object = object->backing_object;
1783 backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1784
1785 /*
1786 * Our scan
1787 */
1788 for (p = TAILQ_FIRST(&backing_object->memq); p != NULL; p = next) {
1789 next = TAILQ_NEXT(p, listq);
1790 new_pindex = p->pindex - backing_offset_index;
1791
1792 /*
1793 * Check for busy page
1794 */
1795 if (vm_page_tryxbusy(p) == 0) {
1796 next = vm_object_collapse_scan_wait(object, p);
1797 continue;
1798 }
1799
1800 KASSERT(object->backing_object == backing_object,
1801 ("vm_object_collapse_scan: backing object mismatch %p != %p",
1802 object->backing_object, backing_object));
1803 KASSERT(p->object == backing_object,
1804 ("vm_object_collapse_scan: object mismatch %p != %p",
1805 p->object, backing_object));
1806
1807 if (p->pindex < backing_offset_index ||
1808 new_pindex >= object->size) {
1809 vm_pager_freespace(backing_object, p->pindex, 1);
1810
1811 KASSERT(!pmap_page_is_mapped(p),
1812 ("freeing mapped page %p", p));
1813 if (vm_page_remove(p))
1814 vm_page_free(p);
1815 continue;
1816 }
1817
1818 if (!vm_page_all_valid(p)) {
1819 KASSERT(!pmap_page_is_mapped(p),
1820 ("freeing mapped page %p", p));
1821 if (vm_page_remove(p))
1822 vm_page_free(p);
1823 continue;
1824 }
1825
1826 pp = vm_page_lookup(object, new_pindex);
1827 if (pp != NULL && vm_page_tryxbusy(pp) == 0) {
1828 vm_page_xunbusy(p);
1829 /*
1830 * The page in the parent is busy and possibly not
1831 * (yet) valid. Until its state is finalized by the
1832 * busy bit owner, we can't tell whether it shadows the
1833 * original page.
1834 */
1835 next = vm_object_collapse_scan_wait(object, pp);
1836 continue;
1837 }
1838
1839 if (pp != NULL && vm_page_none_valid(pp)) {
1840 /*
1841 * The page was invalid in the parent. Likely placed
1842 * there by an incomplete fault. Just remove and
1843 * ignore. p can replace it.
1844 */
1845 if (vm_page_remove(pp))
1846 vm_page_free(pp);
1847 pp = NULL;
1848 }
1849
1850 if (pp != NULL || vm_pager_has_page(object, new_pindex, NULL,
1851 NULL)) {
1852 /*
1853 * The page already exists in the parent OR swap exists
1854 * for this location in the parent. Leave the parent's
1855 * page alone. Destroy the original page from the
1856 * backing object.
1857 */
1858 vm_pager_freespace(backing_object, p->pindex, 1);
1859 KASSERT(!pmap_page_is_mapped(p),
1860 ("freeing mapped page %p", p));
1861 if (vm_page_remove(p))
1862 vm_page_free(p);
1863 if (pp != NULL)
1864 vm_page_xunbusy(pp);
1865 continue;
1866 }
1867
1868 /*
1869 * Page does not exist in parent, rename the page from the
1870 * backing object to the main object.
1871 *
1872 * If the page was mapped to a process, it can remain mapped
1873 * through the rename. vm_page_rename() will dirty the page.
1874 */
1875 if (vm_page_rename(p, object, new_pindex)) {
1876 vm_page_xunbusy(p);
1877 next = vm_object_collapse_scan_wait(object, NULL);
1878 continue;
1879 }
1880
1881 /* Use the old pindex to free the right page. */
1882 vm_pager_freespace(backing_object, new_pindex +
1883 backing_offset_index, 1);
1884
1885 #if VM_NRESERVLEVEL > 0
1886 /*
1887 * Rename the reservation.
1888 */
1889 vm_reserv_rename(p, object, backing_object,
1890 backing_offset_index);
1891 #endif
1892 vm_page_xunbusy(p);
1893 }
1894 return;
1895 }
1896
1897 /*
1898 * vm_object_collapse:
1899 *
1900 * Collapse an object with the object backing it.
1901 * Pages in the backing object are moved into the
1902 * parent, and the backing object is deallocated.
1903 */
1904 void
vm_object_collapse(vm_object_t object)1905 vm_object_collapse(vm_object_t object)
1906 {
1907 vm_object_t backing_object, new_backing_object;
1908
1909 VM_OBJECT_ASSERT_WLOCKED(object);
1910
1911 while (TRUE) {
1912 KASSERT((object->flags & (OBJ_DEAD | OBJ_ANON)) == OBJ_ANON,
1913 ("collapsing invalid object"));
1914
1915 /*
1916 * Wait for the backing_object to finish any pending
1917 * collapse so that the caller sees the shortest possible
1918 * shadow chain.
1919 */
1920 backing_object = vm_object_backing_collapse_wait(object);
1921 if (backing_object == NULL)
1922 return;
1923
1924 KASSERT(object->ref_count > 0 &&
1925 object->ref_count > atomic_load_int(&object->shadow_count),
1926 ("collapse with invalid ref %d or shadow %d count.",
1927 object->ref_count, atomic_load_int(&object->shadow_count)));
1928 KASSERT((backing_object->flags &
1929 (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1930 ("vm_object_collapse: Backing object already collapsing."));
1931 KASSERT((object->flags & (OBJ_COLLAPSING | OBJ_DEAD)) == 0,
1932 ("vm_object_collapse: object is already collapsing."));
1933
1934 /*
1935 * We know that we can either collapse the backing object if
1936 * the parent is the only reference to it, or (perhaps) have
1937 * the parent bypass the object if the parent happens to shadow
1938 * all the resident pages in the entire backing object.
1939 */
1940 if (backing_object->ref_count == 1) {
1941 KASSERT(atomic_load_int(&backing_object->shadow_count)
1942 == 1,
1943 ("vm_object_collapse: shadow_count: %d",
1944 atomic_load_int(&backing_object->shadow_count)));
1945 vm_object_pip_add(object, 1);
1946 vm_object_set_flag(object, OBJ_COLLAPSING);
1947 vm_object_pip_add(backing_object, 1);
1948 vm_object_set_flag(backing_object, OBJ_DEAD);
1949
1950 /*
1951 * If there is exactly one reference to the backing
1952 * object, we can collapse it into the parent.
1953 */
1954 vm_object_collapse_scan(object);
1955
1956 #if VM_NRESERVLEVEL > 0
1957 /*
1958 * Break any reservations from backing_object.
1959 */
1960 if (__predict_false(!LIST_EMPTY(&backing_object->rvq)))
1961 vm_reserv_break_all(backing_object);
1962 #endif
1963
1964 /*
1965 * Move the pager from backing_object to object.
1966 *
1967 * swap_pager_copy() can sleep, in which case the
1968 * backing_object's and object's locks are released and
1969 * reacquired.
1970 */
1971 swap_pager_copy(backing_object, object,
1972 OFF_TO_IDX(object->backing_object_offset), TRUE);
1973
1974 /*
1975 * Object now shadows whatever backing_object did.
1976 */
1977 vm_object_clear_flag(object, OBJ_COLLAPSING);
1978 vm_object_backing_transfer(object, backing_object);
1979 object->backing_object_offset +=
1980 backing_object->backing_object_offset;
1981 VM_OBJECT_WUNLOCK(object);
1982 vm_object_pip_wakeup(object);
1983
1984 /*
1985 * Discard backing_object.
1986 *
1987 * Since the backing object has no pages, no pager left,
1988 * and no object references within it, all that is
1989 * necessary is to dispose of it.
1990 */
1991 KASSERT(backing_object->ref_count == 1, (
1992 "backing_object %p was somehow re-referenced during collapse!",
1993 backing_object));
1994 vm_object_pip_wakeup(backing_object);
1995 (void)refcount_release(&backing_object->ref_count);
1996 umtx_shm_object_terminated(backing_object);
1997 vm_object_terminate(backing_object);
1998 counter_u64_add(object_collapses, 1);
1999 VM_OBJECT_WLOCK(object);
2000 } else {
2001 /*
2002 * If we do not entirely shadow the backing object,
2003 * there is nothing we can do so we give up.
2004 *
2005 * The object lock and backing_object lock must not
2006 * be dropped during this sequence.
2007 */
2008 if (!vm_object_scan_all_shadowed(object)) {
2009 VM_OBJECT_WUNLOCK(backing_object);
2010 break;
2011 }
2012
2013 /*
2014 * Make the parent shadow the next object in the
2015 * chain. Deallocating backing_object will not remove
2016 * it, since its reference count is at least 2.
2017 */
2018 vm_object_backing_remove_locked(object);
2019 new_backing_object = backing_object->backing_object;
2020 if (new_backing_object != NULL) {
2021 vm_object_backing_insert_ref(object,
2022 new_backing_object);
2023 object->backing_object_offset +=
2024 backing_object->backing_object_offset;
2025 }
2026
2027 /*
2028 * Drop the reference count on backing_object. Since
2029 * its ref_count was at least 2, it will not vanish.
2030 */
2031 (void)refcount_release(&backing_object->ref_count);
2032 KASSERT(backing_object->ref_count >= 1, (
2033 "backing_object %p was somehow dereferenced during collapse!",
2034 backing_object));
2035 VM_OBJECT_WUNLOCK(backing_object);
2036 counter_u64_add(object_bypasses, 1);
2037 }
2038
2039 /*
2040 * Try again with this object's new backing object.
2041 */
2042 }
2043 }
2044
2045 /*
2046 * vm_object_page_remove:
2047 *
2048 * For the given object, either frees or invalidates each of the
2049 * specified pages. In general, a page is freed. However, if a page is
2050 * wired for any reason other than the existence of a managed, wired
2051 * mapping, then it may be invalidated but not removed from the object.
2052 * Pages are specified by the given range ["start", "end") and the option
2053 * OBJPR_CLEANONLY. As a special case, if "end" is zero, then the range
2054 * extends from "start" to the end of the object. If the option
2055 * OBJPR_CLEANONLY is specified, then only the non-dirty pages within the
2056 * specified range are affected. If the option OBJPR_NOTMAPPED is
2057 * specified, then the pages within the specified range must have no
2058 * mappings. Otherwise, if this option is not specified, any mappings to
2059 * the specified pages are removed before the pages are freed or
2060 * invalidated.
2061 *
2062 * In general, this operation should only be performed on objects that
2063 * contain managed pages. There are, however, two exceptions. First, it
2064 * is performed on the kernel and kmem objects by vm_map_entry_delete().
2065 * Second, it is used by msync(..., MS_INVALIDATE) to invalidate device-
2066 * backed pages. In both of these cases, the option OBJPR_CLEANONLY must
2067 * not be specified and the option OBJPR_NOTMAPPED must be specified.
2068 *
2069 * The object must be locked.
2070 */
2071 void
vm_object_page_remove(vm_object_t object,vm_pindex_t start,vm_pindex_t end,int options)2072 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2073 int options)
2074 {
2075 vm_page_t p, next;
2076
2077 VM_OBJECT_ASSERT_WLOCKED(object);
2078 KASSERT((object->flags & OBJ_UNMANAGED) == 0 ||
2079 (options & (OBJPR_CLEANONLY | OBJPR_NOTMAPPED)) == OBJPR_NOTMAPPED,
2080 ("vm_object_page_remove: illegal options for object %p", object));
2081 if (object->resident_page_count == 0)
2082 return;
2083 vm_object_pip_add(object, 1);
2084 again:
2085 p = vm_page_find_least(object, start);
2086
2087 /*
2088 * Here, the variable "p" is either (1) the page with the least pindex
2089 * greater than or equal to the parameter "start" or (2) NULL.
2090 */
2091 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2092 next = TAILQ_NEXT(p, listq);
2093
2094 /*
2095 * Skip invalid pages if asked to do so. Try to avoid acquiring
2096 * the busy lock, as some consumers rely on this to avoid
2097 * deadlocks.
2098 *
2099 * A thread may concurrently transition the page from invalid to
2100 * valid using only the busy lock, so the result of this check
2101 * is immediately stale. It is up to consumers to handle this,
2102 * for instance by ensuring that all invalid->valid transitions
2103 * happen with a mutex held, as may be possible for a
2104 * filesystem.
2105 */
2106 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p))
2107 continue;
2108
2109 /*
2110 * If the page is wired for any reason besides the existence
2111 * of managed, wired mappings, then it cannot be freed. For
2112 * example, fictitious pages, which represent device memory,
2113 * are inherently wired and cannot be freed. They can,
2114 * however, be invalidated if the option OBJPR_CLEANONLY is
2115 * not specified.
2116 */
2117 if (vm_page_tryxbusy(p) == 0) {
2118 if (vm_page_busy_sleep(p, "vmopar", 0))
2119 VM_OBJECT_WLOCK(object);
2120 goto again;
2121 }
2122 if ((options & OBJPR_VALIDONLY) != 0 && vm_page_none_valid(p)) {
2123 vm_page_xunbusy(p);
2124 continue;
2125 }
2126 if (vm_page_wired(p)) {
2127 wired:
2128 if ((options & OBJPR_NOTMAPPED) == 0 &&
2129 object->ref_count != 0)
2130 pmap_remove_all(p);
2131 if ((options & OBJPR_CLEANONLY) == 0) {
2132 vm_page_invalid(p);
2133 vm_page_undirty(p);
2134 }
2135 vm_page_xunbusy(p);
2136 continue;
2137 }
2138 KASSERT((p->flags & PG_FICTITIOUS) == 0,
2139 ("vm_object_page_remove: page %p is fictitious", p));
2140 if ((options & OBJPR_CLEANONLY) != 0 &&
2141 !vm_page_none_valid(p)) {
2142 if ((options & OBJPR_NOTMAPPED) == 0 &&
2143 object->ref_count != 0 &&
2144 !vm_page_try_remove_write(p))
2145 goto wired;
2146 if (p->dirty != 0) {
2147 vm_page_xunbusy(p);
2148 continue;
2149 }
2150 }
2151 if ((options & OBJPR_NOTMAPPED) == 0 &&
2152 object->ref_count != 0 && !vm_page_try_remove_all(p))
2153 goto wired;
2154 vm_page_free(p);
2155 }
2156 vm_object_pip_wakeup(object);
2157
2158 vm_pager_freespace(object, start, (end == 0 ? object->size : end) -
2159 start);
2160 }
2161
2162 /*
2163 * vm_object_page_noreuse:
2164 *
2165 * For the given object, attempt to move the specified pages to
2166 * the head of the inactive queue. This bypasses regular LRU
2167 * operation and allows the pages to be reused quickly under memory
2168 * pressure. If a page is wired for any reason, then it will not
2169 * be queued. Pages are specified by the range ["start", "end").
2170 * As a special case, if "end" is zero, then the range extends from
2171 * "start" to the end of the object.
2172 *
2173 * This operation should only be performed on objects that
2174 * contain non-fictitious, managed pages.
2175 *
2176 * The object must be locked.
2177 */
2178 void
vm_object_page_noreuse(vm_object_t object,vm_pindex_t start,vm_pindex_t end)2179 vm_object_page_noreuse(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2180 {
2181 vm_page_t p, next;
2182
2183 VM_OBJECT_ASSERT_LOCKED(object);
2184 KASSERT((object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0,
2185 ("vm_object_page_noreuse: illegal object %p", object));
2186 if (object->resident_page_count == 0)
2187 return;
2188 p = vm_page_find_least(object, start);
2189
2190 /*
2191 * Here, the variable "p" is either (1) the page with the least pindex
2192 * greater than or equal to the parameter "start" or (2) NULL.
2193 */
2194 for (; p != NULL && (p->pindex < end || end == 0); p = next) {
2195 next = TAILQ_NEXT(p, listq);
2196 vm_page_deactivate_noreuse(p);
2197 }
2198 }
2199
2200 /*
2201 * Populate the specified range of the object with valid pages. Returns
2202 * TRUE if the range is successfully populated and FALSE otherwise.
2203 *
2204 * Note: This function should be optimized to pass a larger array of
2205 * pages to vm_pager_get_pages() before it is applied to a non-
2206 * OBJT_DEVICE object.
2207 *
2208 * The object must be locked.
2209 */
2210 boolean_t
vm_object_populate(vm_object_t object,vm_pindex_t start,vm_pindex_t end)2211 vm_object_populate(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
2212 {
2213 vm_page_t m;
2214 vm_pindex_t pindex;
2215 int rv;
2216
2217 VM_OBJECT_ASSERT_WLOCKED(object);
2218 for (pindex = start; pindex < end; pindex++) {
2219 rv = vm_page_grab_valid(&m, object, pindex, VM_ALLOC_NORMAL);
2220 if (rv != VM_PAGER_OK)
2221 break;
2222
2223 /*
2224 * Keep "m" busy because a subsequent iteration may unlock
2225 * the object.
2226 */
2227 }
2228 if (pindex > start) {
2229 m = vm_page_lookup(object, start);
2230 while (m != NULL && m->pindex < pindex) {
2231 vm_page_xunbusy(m);
2232 m = TAILQ_NEXT(m, listq);
2233 }
2234 }
2235 return (pindex == end);
2236 }
2237
2238 /*
2239 * Routine: vm_object_coalesce
2240 * Function: Coalesces two objects backing up adjoining
2241 * regions of memory into a single object.
2242 *
2243 * returns TRUE if objects were combined.
2244 *
2245 * NOTE: Only works at the moment if the second object is NULL -
2246 * if it's not, which object do we lock first?
2247 *
2248 * Parameters:
2249 * prev_object First object to coalesce
2250 * prev_offset Offset into prev_object
2251 * prev_size Size of reference to prev_object
2252 * next_size Size of reference to the second object
2253 * reserved Indicator that extension region has
2254 * swap accounted for
2255 *
2256 * Conditions:
2257 * The object must *not* be locked.
2258 */
2259 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)2260 vm_object_coalesce(vm_object_t prev_object, vm_ooffset_t prev_offset,
2261 vm_size_t prev_size, vm_size_t next_size, boolean_t reserved)
2262 {
2263 vm_pindex_t next_pindex;
2264
2265 if (prev_object == NULL)
2266 return (TRUE);
2267 if ((prev_object->flags & OBJ_ANON) == 0)
2268 return (FALSE);
2269
2270 VM_OBJECT_WLOCK(prev_object);
2271 /*
2272 * Try to collapse the object first.
2273 */
2274 vm_object_collapse(prev_object);
2275
2276 /*
2277 * Can't coalesce if: . more than one reference . paged out . shadows
2278 * another object . has a copy elsewhere (any of which mean that the
2279 * pages not mapped to prev_entry may be in use anyway)
2280 */
2281 if (prev_object->backing_object != NULL) {
2282 VM_OBJECT_WUNLOCK(prev_object);
2283 return (FALSE);
2284 }
2285
2286 prev_size >>= PAGE_SHIFT;
2287 next_size >>= PAGE_SHIFT;
2288 next_pindex = OFF_TO_IDX(prev_offset) + prev_size;
2289
2290 if (prev_object->ref_count > 1 &&
2291 prev_object->size != next_pindex &&
2292 (prev_object->flags & OBJ_ONEMAPPING) == 0) {
2293 VM_OBJECT_WUNLOCK(prev_object);
2294 return (FALSE);
2295 }
2296
2297 /*
2298 * Account for the charge.
2299 */
2300 if (prev_object->cred != NULL) {
2301 /*
2302 * If prev_object was charged, then this mapping,
2303 * although not charged now, may become writable
2304 * later. Non-NULL cred in the object would prevent
2305 * swap reservation during enabling of the write
2306 * access, so reserve swap now. Failed reservation
2307 * cause allocation of the separate object for the map
2308 * entry, and swap reservation for this entry is
2309 * managed in appropriate time.
2310 */
2311 if (!reserved && !swap_reserve_by_cred(ptoa(next_size),
2312 prev_object->cred)) {
2313 VM_OBJECT_WUNLOCK(prev_object);
2314 return (FALSE);
2315 }
2316 prev_object->charge += ptoa(next_size);
2317 }
2318
2319 /*
2320 * Remove any pages that may still be in the object from a previous
2321 * deallocation.
2322 */
2323 if (next_pindex < prev_object->size) {
2324 vm_object_page_remove(prev_object, next_pindex, next_pindex +
2325 next_size, 0);
2326 #if 0
2327 if (prev_object->cred != NULL) {
2328 KASSERT(prev_object->charge >=
2329 ptoa(prev_object->size - next_pindex),
2330 ("object %p overcharged 1 %jx %jx", prev_object,
2331 (uintmax_t)next_pindex, (uintmax_t)next_size));
2332 prev_object->charge -= ptoa(prev_object->size -
2333 next_pindex);
2334 }
2335 #endif
2336 }
2337
2338 /*
2339 * Extend the object if necessary.
2340 */
2341 if (next_pindex + next_size > prev_object->size)
2342 prev_object->size = next_pindex + next_size;
2343
2344 VM_OBJECT_WUNLOCK(prev_object);
2345 return (TRUE);
2346 }
2347
2348 void
vm_object_set_writeable_dirty_(vm_object_t object)2349 vm_object_set_writeable_dirty_(vm_object_t object)
2350 {
2351 atomic_add_int(&object->generation, 1);
2352 }
2353
2354 bool
vm_object_mightbedirty_(vm_object_t object)2355 vm_object_mightbedirty_(vm_object_t object)
2356 {
2357 return (object->generation != object->cleangeneration);
2358 }
2359
2360 /*
2361 * vm_object_unwire:
2362 *
2363 * For each page offset within the specified range of the given object,
2364 * find the highest-level page in the shadow chain and unwire it. A page
2365 * must exist at every page offset, and the highest-level page must be
2366 * wired.
2367 */
2368 void
vm_object_unwire(vm_object_t object,vm_ooffset_t offset,vm_size_t length,uint8_t queue)2369 vm_object_unwire(vm_object_t object, vm_ooffset_t offset, vm_size_t length,
2370 uint8_t queue)
2371 {
2372 vm_object_t tobject, t1object;
2373 vm_page_t m, tm;
2374 vm_pindex_t end_pindex, pindex, tpindex;
2375 int depth, locked_depth;
2376
2377 KASSERT((offset & PAGE_MASK) == 0,
2378 ("vm_object_unwire: offset is not page aligned"));
2379 KASSERT((length & PAGE_MASK) == 0,
2380 ("vm_object_unwire: length is not a multiple of PAGE_SIZE"));
2381 /* The wired count of a fictitious page never changes. */
2382 if ((object->flags & OBJ_FICTITIOUS) != 0)
2383 return;
2384 pindex = OFF_TO_IDX(offset);
2385 end_pindex = pindex + atop(length);
2386 again:
2387 locked_depth = 1;
2388 VM_OBJECT_RLOCK(object);
2389 m = vm_page_find_least(object, pindex);
2390 while (pindex < end_pindex) {
2391 if (m == NULL || pindex < m->pindex) {
2392 /*
2393 * The first object in the shadow chain doesn't
2394 * contain a page at the current index. Therefore,
2395 * the page must exist in a backing object.
2396 */
2397 tobject = object;
2398 tpindex = pindex;
2399 depth = 0;
2400 do {
2401 tpindex +=
2402 OFF_TO_IDX(tobject->backing_object_offset);
2403 tobject = tobject->backing_object;
2404 KASSERT(tobject != NULL,
2405 ("vm_object_unwire: missing page"));
2406 if ((tobject->flags & OBJ_FICTITIOUS) != 0)
2407 goto next_page;
2408 depth++;
2409 if (depth == locked_depth) {
2410 locked_depth++;
2411 VM_OBJECT_RLOCK(tobject);
2412 }
2413 } while ((tm = vm_page_lookup(tobject, tpindex)) ==
2414 NULL);
2415 } else {
2416 tm = m;
2417 m = TAILQ_NEXT(m, listq);
2418 }
2419 if (vm_page_trysbusy(tm) == 0) {
2420 for (tobject = object; locked_depth >= 1;
2421 locked_depth--) {
2422 t1object = tobject->backing_object;
2423 if (tm->object != tobject)
2424 VM_OBJECT_RUNLOCK(tobject);
2425 tobject = t1object;
2426 }
2427 tobject = tm->object;
2428 if (!vm_page_busy_sleep(tm, "unwbo",
2429 VM_ALLOC_IGN_SBUSY))
2430 VM_OBJECT_RUNLOCK(tobject);
2431 goto again;
2432 }
2433 vm_page_unwire(tm, queue);
2434 vm_page_sunbusy(tm);
2435 next_page:
2436 pindex++;
2437 }
2438 /* Release the accumulated object locks. */
2439 for (tobject = object; locked_depth >= 1; locked_depth--) {
2440 t1object = tobject->backing_object;
2441 VM_OBJECT_RUNLOCK(tobject);
2442 tobject = t1object;
2443 }
2444 }
2445
2446 /*
2447 * Return the vnode for the given object, or NULL if none exists.
2448 * For tmpfs objects, the function may return NULL if there is
2449 * no vnode allocated at the time of the call.
2450 */
2451 struct vnode *
vm_object_vnode(vm_object_t object)2452 vm_object_vnode(vm_object_t object)
2453 {
2454 struct vnode *vp;
2455
2456 VM_OBJECT_ASSERT_LOCKED(object);
2457 vm_pager_getvp(object, &vp, NULL);
2458 return (vp);
2459 }
2460
2461 /*
2462 * Busy the vm object. This prevents new pages belonging to the object from
2463 * becoming busy. Existing pages persist as busy. Callers are responsible
2464 * for checking page state before proceeding.
2465 */
2466 void
vm_object_busy(vm_object_t obj)2467 vm_object_busy(vm_object_t obj)
2468 {
2469
2470 VM_OBJECT_ASSERT_LOCKED(obj);
2471
2472 blockcount_acquire(&obj->busy, 1);
2473 /* The fence is required to order loads of page busy. */
2474 atomic_thread_fence_acq_rel();
2475 }
2476
2477 void
vm_object_unbusy(vm_object_t obj)2478 vm_object_unbusy(vm_object_t obj)
2479 {
2480
2481 blockcount_release(&obj->busy, 1);
2482 }
2483
2484 void
vm_object_busy_wait(vm_object_t obj,const char * wmesg)2485 vm_object_busy_wait(vm_object_t obj, const char *wmesg)
2486 {
2487
2488 VM_OBJECT_ASSERT_UNLOCKED(obj);
2489
2490 (void)blockcount_sleep(&obj->busy, NULL, wmesg, PVM);
2491 }
2492
2493 /*
2494 * This function aims to determine if the object is mapped,
2495 * specifically, if it is referenced by a vm_map_entry. Because
2496 * objects occasionally acquire transient references that do not
2497 * represent a mapping, the method used here is inexact. However, it
2498 * has very low overhead and is good enough for the advisory
2499 * vm.vmtotal sysctl.
2500 */
2501 bool
vm_object_is_active(vm_object_t obj)2502 vm_object_is_active(vm_object_t obj)
2503 {
2504
2505 return (obj->ref_count > atomic_load_int(&obj->shadow_count));
2506 }
2507
2508 static int
vm_object_list_handler(struct sysctl_req * req,bool swap_only)2509 vm_object_list_handler(struct sysctl_req *req, bool swap_only)
2510 {
2511 struct kinfo_vmobject *kvo;
2512 char *fullpath, *freepath;
2513 struct vnode *vp;
2514 struct vattr va;
2515 vm_object_t obj;
2516 vm_page_t m;
2517 u_long sp;
2518 int count, error;
2519 bool want_path;
2520
2521 if (req->oldptr == NULL) {
2522 /*
2523 * If an old buffer has not been provided, generate an
2524 * estimate of the space needed for a subsequent call.
2525 */
2526 mtx_lock(&vm_object_list_mtx);
2527 count = 0;
2528 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2529 if (obj->type == OBJT_DEAD)
2530 continue;
2531 count++;
2532 }
2533 mtx_unlock(&vm_object_list_mtx);
2534 return (SYSCTL_OUT(req, NULL, sizeof(struct kinfo_vmobject) *
2535 count * 11 / 10));
2536 }
2537
2538 want_path = !(swap_only || jailed(curthread->td_ucred));
2539 kvo = malloc(sizeof(*kvo), M_TEMP, M_WAITOK | M_ZERO);
2540 error = 0;
2541
2542 /*
2543 * VM objects are type stable and are never removed from the
2544 * list once added. This allows us to safely read obj->object_list
2545 * after reacquiring the VM object lock.
2546 */
2547 mtx_lock(&vm_object_list_mtx);
2548 TAILQ_FOREACH(obj, &vm_object_list, object_list) {
2549 if (obj->type == OBJT_DEAD ||
2550 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0))
2551 continue;
2552 VM_OBJECT_RLOCK(obj);
2553 if (obj->type == OBJT_DEAD ||
2554 (swap_only && (obj->flags & (OBJ_ANON | OBJ_SWAP)) == 0)) {
2555 VM_OBJECT_RUNLOCK(obj);
2556 continue;
2557 }
2558 mtx_unlock(&vm_object_list_mtx);
2559 kvo->kvo_size = ptoa(obj->size);
2560 kvo->kvo_resident = obj->resident_page_count;
2561 kvo->kvo_ref_count = obj->ref_count;
2562 kvo->kvo_shadow_count = atomic_load_int(&obj->shadow_count);
2563 kvo->kvo_memattr = obj->memattr;
2564 kvo->kvo_active = 0;
2565 kvo->kvo_inactive = 0;
2566 if (!swap_only) {
2567 TAILQ_FOREACH(m, &obj->memq, listq) {
2568 /*
2569 * A page may belong to the object but be
2570 * dequeued and set to PQ_NONE while the
2571 * object lock is not held. This makes the
2572 * reads of m->queue below racy, and we do not
2573 * count pages set to PQ_NONE. However, this
2574 * sysctl is only meant to give an
2575 * approximation of the system anyway.
2576 */
2577 if (m->a.queue == PQ_ACTIVE)
2578 kvo->kvo_active++;
2579 else if (m->a.queue == PQ_INACTIVE)
2580 kvo->kvo_inactive++;
2581 }
2582 }
2583
2584 kvo->kvo_vn_fileid = 0;
2585 kvo->kvo_vn_fsid = 0;
2586 kvo->kvo_vn_fsid_freebsd11 = 0;
2587 freepath = NULL;
2588 fullpath = "";
2589 vp = NULL;
2590 kvo->kvo_type = vm_object_kvme_type(obj, want_path ? &vp :
2591 NULL);
2592 if (vp != NULL) {
2593 vref(vp);
2594 } else if ((obj->flags & OBJ_ANON) != 0) {
2595 MPASS(kvo->kvo_type == KVME_TYPE_SWAP);
2596 kvo->kvo_me = (uintptr_t)obj;
2597 /* tmpfs objs are reported as vnodes */
2598 kvo->kvo_backing_obj = (uintptr_t)obj->backing_object;
2599 sp = swap_pager_swapped_pages(obj);
2600 kvo->kvo_swapped = sp > UINT32_MAX ? UINT32_MAX : sp;
2601 }
2602 VM_OBJECT_RUNLOCK(obj);
2603 if (vp != NULL) {
2604 vn_fullpath(vp, &fullpath, &freepath);
2605 vn_lock(vp, LK_SHARED | LK_RETRY);
2606 if (VOP_GETATTR(vp, &va, curthread->td_ucred) == 0) {
2607 kvo->kvo_vn_fileid = va.va_fileid;
2608 kvo->kvo_vn_fsid = va.va_fsid;
2609 kvo->kvo_vn_fsid_freebsd11 = va.va_fsid;
2610 /* truncate */
2611 }
2612 vput(vp);
2613 }
2614
2615 strlcpy(kvo->kvo_path, fullpath, sizeof(kvo->kvo_path));
2616 free(freepath, M_TEMP);
2617
2618 /* Pack record size down */
2619 kvo->kvo_structsize = offsetof(struct kinfo_vmobject, kvo_path)
2620 + strlen(kvo->kvo_path) + 1;
2621 kvo->kvo_structsize = roundup(kvo->kvo_structsize,
2622 sizeof(uint64_t));
2623 error = SYSCTL_OUT(req, kvo, kvo->kvo_structsize);
2624 maybe_yield();
2625 mtx_lock(&vm_object_list_mtx);
2626 if (error)
2627 break;
2628 }
2629 mtx_unlock(&vm_object_list_mtx);
2630 free(kvo, M_TEMP);
2631 return (error);
2632 }
2633
2634 static int
sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)2635 sysctl_vm_object_list(SYSCTL_HANDLER_ARGS)
2636 {
2637 return (vm_object_list_handler(req, false));
2638 }
2639
2640 SYSCTL_PROC(_vm, OID_AUTO, objects, CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP |
2641 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_object_list, "S,kinfo_vmobject",
2642 "List of VM objects");
2643
2644 static int
sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)2645 sysctl_vm_object_list_swap(SYSCTL_HANDLER_ARGS)
2646 {
2647 return (vm_object_list_handler(req, true));
2648 }
2649
2650 /*
2651 * This sysctl returns list of the anonymous or swap objects. Intent
2652 * is to provide stripped optimized list useful to analyze swap use.
2653 * Since technically non-swap (default) objects participate in the
2654 * shadow chains, and are converted to swap type as needed by swap
2655 * pager, we must report them.
2656 */
2657 SYSCTL_PROC(_vm, OID_AUTO, swap_objects,
2658 CTLTYPE_STRUCT | CTLFLAG_RW | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0,
2659 sysctl_vm_object_list_swap, "S,kinfo_vmobject",
2660 "List of swap VM objects");
2661
2662 #include "opt_ddb.h"
2663 #ifdef DDB
2664 #include <sys/kernel.h>
2665
2666 #include <sys/cons.h>
2667
2668 #include <ddb/ddb.h>
2669
2670 static int
_vm_object_in_map(vm_map_t map,vm_object_t object,vm_map_entry_t entry)2671 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2672 {
2673 vm_map_t tmpm;
2674 vm_map_entry_t tmpe;
2675 vm_object_t obj;
2676
2677 if (map == 0)
2678 return 0;
2679
2680 if (entry == 0) {
2681 VM_MAP_ENTRY_FOREACH(tmpe, map) {
2682 if (_vm_object_in_map(map, object, tmpe)) {
2683 return 1;
2684 }
2685 }
2686 } else if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) {
2687 tmpm = entry->object.sub_map;
2688 VM_MAP_ENTRY_FOREACH(tmpe, tmpm) {
2689 if (_vm_object_in_map(tmpm, object, tmpe)) {
2690 return 1;
2691 }
2692 }
2693 } else if ((obj = entry->object.vm_object) != NULL) {
2694 for (; obj; obj = obj->backing_object)
2695 if (obj == object) {
2696 return 1;
2697 }
2698 }
2699 return 0;
2700 }
2701
2702 static int
vm_object_in_map(vm_object_t object)2703 vm_object_in_map(vm_object_t object)
2704 {
2705 struct proc *p;
2706
2707 /* sx_slock(&allproc_lock); */
2708 FOREACH_PROC_IN_SYSTEM(p) {
2709 if (!p->p_vmspace /* || (p->p_flag & (P_SYSTEM|P_WEXIT)) */)
2710 continue;
2711 if (_vm_object_in_map(&p->p_vmspace->vm_map, object, 0)) {
2712 /* sx_sunlock(&allproc_lock); */
2713 return 1;
2714 }
2715 }
2716 /* sx_sunlock(&allproc_lock); */
2717 if (_vm_object_in_map(kernel_map, object, 0))
2718 return 1;
2719 return 0;
2720 }
2721
DB_SHOW_COMMAND_FLAGS(vmochk,vm_object_check,DB_CMD_MEMSAFE)2722 DB_SHOW_COMMAND_FLAGS(vmochk, vm_object_check, DB_CMD_MEMSAFE)
2723 {
2724 vm_object_t object;
2725
2726 /*
2727 * make sure that internal objs are in a map somewhere
2728 * and none have zero ref counts.
2729 */
2730 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2731 if ((object->flags & OBJ_ANON) != 0) {
2732 if (object->ref_count == 0) {
2733 db_printf("vmochk: internal obj has zero ref count: %ld\n",
2734 (long)object->size);
2735 }
2736 if (!vm_object_in_map(object)) {
2737 db_printf(
2738 "vmochk: internal obj is not in a map: "
2739 "ref: %d, size: %lu: 0x%lx, backing_object: %p\n",
2740 object->ref_count, (u_long)object->size,
2741 (u_long)object->size,
2742 (void *)object->backing_object);
2743 }
2744 }
2745 if (db_pager_quit)
2746 return;
2747 }
2748 }
2749
2750 /*
2751 * vm_object_print: [ debug ]
2752 */
DB_SHOW_COMMAND(object,vm_object_print_static)2753 DB_SHOW_COMMAND(object, vm_object_print_static)
2754 {
2755 /* XXX convert args. */
2756 vm_object_t object = (vm_object_t)addr;
2757 boolean_t full = have_addr;
2758
2759 vm_page_t p;
2760
2761 /* XXX count is an (unused) arg. Avoid shadowing it. */
2762 #define count was_count
2763
2764 int count;
2765
2766 if (object == NULL)
2767 return;
2768
2769 db_iprintf(
2770 "Object %p: type=%d, size=0x%jx, res=%d, ref=%d, flags=0x%x ruid %d charge %jx\n",
2771 object, (int)object->type, (uintmax_t)object->size,
2772 object->resident_page_count, object->ref_count, object->flags,
2773 object->cred ? object->cred->cr_ruid : -1, (uintmax_t)object->charge);
2774 db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%jx\n",
2775 atomic_load_int(&object->shadow_count),
2776 object->backing_object ? object->backing_object->ref_count : 0,
2777 object->backing_object, (uintmax_t)object->backing_object_offset);
2778
2779 if (!full)
2780 return;
2781
2782 db_indent += 2;
2783 count = 0;
2784 TAILQ_FOREACH(p, &object->memq, listq) {
2785 if (count == 0)
2786 db_iprintf("memory:=");
2787 else if (count == 6) {
2788 db_printf("\n");
2789 db_iprintf(" ...");
2790 count = 0;
2791 } else
2792 db_printf(",");
2793 count++;
2794
2795 db_printf("(off=0x%jx,page=0x%jx)",
2796 (uintmax_t)p->pindex, (uintmax_t)VM_PAGE_TO_PHYS(p));
2797
2798 if (db_pager_quit)
2799 break;
2800 }
2801 if (count != 0)
2802 db_printf("\n");
2803 db_indent -= 2;
2804 }
2805
2806 /* XXX. */
2807 #undef count
2808
2809 /* XXX need this non-static entry for calling from vm_map_print. */
2810 void
vm_object_print(long addr,boolean_t have_addr,long count,char * modif)2811 vm_object_print(
2812 /* db_expr_t */ long addr,
2813 boolean_t have_addr,
2814 /* db_expr_t */ long count,
2815 char *modif)
2816 {
2817 vm_object_print_static(addr, have_addr, count, modif);
2818 }
2819
DB_SHOW_COMMAND_FLAGS(vmopag,vm_object_print_pages,DB_CMD_MEMSAFE)2820 DB_SHOW_COMMAND_FLAGS(vmopag, vm_object_print_pages, DB_CMD_MEMSAFE)
2821 {
2822 vm_object_t object;
2823 vm_pindex_t fidx;
2824 vm_paddr_t pa;
2825 vm_page_t m, prev_m;
2826 int rcount;
2827
2828 TAILQ_FOREACH(object, &vm_object_list, object_list) {
2829 db_printf("new object: %p\n", (void *)object);
2830 if (db_pager_quit)
2831 return;
2832
2833 rcount = 0;
2834 fidx = 0;
2835 pa = -1;
2836 TAILQ_FOREACH(m, &object->memq, listq) {
2837 if ((prev_m = TAILQ_PREV(m, pglist, listq)) != NULL &&
2838 prev_m->pindex + 1 != m->pindex) {
2839 if (rcount) {
2840 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2841 (long)fidx, rcount, (long)pa);
2842 if (db_pager_quit)
2843 return;
2844 rcount = 0;
2845 }
2846 }
2847 if (rcount &&
2848 (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
2849 ++rcount;
2850 continue;
2851 }
2852 if (rcount) {
2853 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2854 (long)fidx, rcount, (long)pa);
2855 if (db_pager_quit)
2856 return;
2857 }
2858 fidx = m->pindex;
2859 pa = VM_PAGE_TO_PHYS(m);
2860 rcount = 1;
2861 }
2862 if (rcount) {
2863 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
2864 (long)fidx, rcount, (long)pa);
2865 if (db_pager_quit)
2866 return;
2867 }
2868 }
2869 }
2870 #endif /* DDB */
2871