kernel - Remove object->agg_pv_list_count
[dragonfly.git] / sys / vm / vm_object.c
1 /*
2  * Copyright (c) 1991, 1993, 2013
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  * 3. 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  * $FreeBSD: src/sys/vm/vm_object.c,v 1.171.2.8 2003/05/26 19:17:56 alc Exp $
61  */
62
63 /*
64  *      Virtual memory object module.
65  */
66
67 #include <sys/param.h>
68 #include <sys/systm.h>
69 #include <sys/proc.h>           /* for curproc, pageproc */
70 #include <sys/thread.h>
71 #include <sys/vnode.h>
72 #include <sys/vmmeter.h>
73 #include <sys/mman.h>
74 #include <sys/mount.h>
75 #include <sys/kernel.h>
76 #include <sys/sysctl.h>
77 #include <sys/refcount.h>
78
79 #include <vm/vm.h>
80 #include <vm/vm_param.h>
81 #include <vm/pmap.h>
82 #include <vm/vm_map.h>
83 #include <vm/vm_object.h>
84 #include <vm/vm_page.h>
85 #include <vm/vm_pageout.h>
86 #include <vm/vm_pager.h>
87 #include <vm/swap_pager.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_extern.h>
90 #include <vm/vm_zone.h>
91
92 #include <vm/vm_page2.h>
93
94 #include <machine/specialreg.h>
95
96 #define EASY_SCAN_FACTOR        8
97
98 static void     vm_object_qcollapse(vm_object_t object,
99                                     vm_object_t backing_object);
100 static void     vm_object_page_collect_flush(vm_object_t object, vm_page_t p,
101                                              int pagerflags);
102 static void     vm_object_lock_init(vm_object_t);
103
104
105 /*
106  *      Virtual memory objects maintain the actual data
107  *      associated with allocated virtual memory.  A given
108  *      page of memory exists within exactly one object.
109  *
110  *      An object is only deallocated when all "references"
111  *      are given up.  Only one "reference" to a given
112  *      region of an object should be writeable.
113  *
114  *      Associated with each object is a list of all resident
115  *      memory pages belonging to that object; this list is
116  *      maintained by the "vm_page" module, and locked by the object's
117  *      lock.
118  *
119  *      Each object also records a "pager" routine which is
120  *      used to retrieve (and store) pages to the proper backing
121  *      storage.  In addition, objects may be backed by other
122  *      objects from which they were virtual-copied.
123  *
124  *      The only items within the object structure which are
125  *      modified after time of creation are:
126  *              reference count         locked by object's lock
127  *              pager routine           locked by object's lock
128  *
129  */
130
131 struct vm_object kernel_object;
132
133 static long object_collapses;
134 static long object_bypasses;
135
136 struct vm_object_hash vm_object_hash[VMOBJ_HSIZE];
137
138 MALLOC_DEFINE(M_VM_OBJECT, "vm_object", "vm_object structures");
139
140 #if defined(DEBUG_LOCKS)
141
142 #define vm_object_vndeallocate(obj, vpp)        \
143                 debugvm_object_vndeallocate(obj, vpp, __FILE__, __LINE__)
144
145 /*
146  * Debug helper to track hold/drop/ref/deallocate calls.
147  */
148 static void
149 debugvm_object_add(vm_object_t obj, char *file, int line, int addrem)
150 {
151         int i;
152
153         i = atomic_fetchadd_int(&obj->debug_index, 1);
154         i = i & (VMOBJ_DEBUG_ARRAY_SIZE - 1);
155         ksnprintf(obj->debug_hold_thrs[i],
156                   sizeof(obj->debug_hold_thrs[i]),
157                   "%c%d:(%d):%s",
158                   (addrem == -1 ? '-' : (addrem == 1 ? '+' : '=')),
159                   (curthread->td_proc ? curthread->td_proc->p_pid : -1),
160                   obj->ref_count,
161                   curthread->td_comm);
162         obj->debug_hold_file[i] = file;
163         obj->debug_hold_line[i] = line;
164 #if 0
165         /* Uncomment for debugging obj refs/derefs in reproducable cases */
166         if (strcmp(curthread->td_comm, "sshd") == 0) {
167                 kprintf("%d %p refs=%d ar=%d file: %s/%d\n",
168                         (curthread->td_proc ? curthread->td_proc->p_pid : -1),
169                         obj, obj->ref_count, addrem, file, line);
170         }
171 #endif
172 }
173
174 #endif
175
176 /*
177  * Misc low level routines
178  */
179 static void
180 vm_object_lock_init(vm_object_t obj)
181 {
182 #if defined(DEBUG_LOCKS)
183         int i;
184
185         obj->debug_index = 0;
186         for (i = 0; i < VMOBJ_DEBUG_ARRAY_SIZE; i++) {
187                 obj->debug_hold_thrs[i][0] = 0;
188                 obj->debug_hold_file[i] = NULL;
189                 obj->debug_hold_line[i] = 0;
190         }
191 #endif
192 }
193
194 void
195 vm_object_lock_swap(void)
196 {
197         lwkt_token_swap();
198 }
199
200 void
201 vm_object_lock(vm_object_t obj)
202 {
203         lwkt_gettoken(&obj->token);
204 }
205
206 /*
207  * Returns TRUE on sucesss
208  */
209 static int
210 vm_object_lock_try(vm_object_t obj)
211 {
212         return(lwkt_trytoken(&obj->token));
213 }
214
215 void
216 vm_object_lock_shared(vm_object_t obj)
217 {
218         lwkt_gettoken_shared(&obj->token);
219 }
220
221 void
222 vm_object_unlock(vm_object_t obj)
223 {
224         lwkt_reltoken(&obj->token);
225 }
226
227 void
228 vm_object_upgrade(vm_object_t obj)
229 {
230         lwkt_reltoken(&obj->token);
231         lwkt_gettoken(&obj->token);
232 }
233
234 void
235 vm_object_downgrade(vm_object_t obj)
236 {
237         lwkt_reltoken(&obj->token);
238         lwkt_gettoken_shared(&obj->token);
239 }
240
241 static __inline void
242 vm_object_assert_held(vm_object_t obj)
243 {
244         ASSERT_LWKT_TOKEN_HELD(&obj->token);
245 }
246
247 static __inline int
248 vm_quickcolor(void)
249 {
250         globaldata_t gd = mycpu;
251         int pg_color;
252
253         pg_color = (int)(intptr_t)gd->gd_curthread >> 10;
254         pg_color += gd->gd_quick_color;
255         gd->gd_quick_color += PQ_PRIME2;
256
257         return pg_color;
258 }
259
260 void
261 VMOBJDEBUG(vm_object_hold)(vm_object_t obj VMOBJDBARGS)
262 {
263         KKASSERT(obj != NULL);
264
265         /*
266          * Object must be held (object allocation is stable due to callers
267          * context, typically already holding the token on a parent object)
268          * prior to potentially blocking on the lock, otherwise the object
269          * can get ripped away from us.
270          */
271         refcount_acquire(&obj->hold_count);
272         vm_object_lock(obj);
273
274 #if defined(DEBUG_LOCKS)
275         debugvm_object_add(obj, file, line, 1);
276 #endif
277 }
278
279 int
280 VMOBJDEBUG(vm_object_hold_try)(vm_object_t obj VMOBJDBARGS)
281 {
282         KKASSERT(obj != NULL);
283
284         /*
285          * Object must be held (object allocation is stable due to callers
286          * context, typically already holding the token on a parent object)
287          * prior to potentially blocking on the lock, otherwise the object
288          * can get ripped away from us.
289          */
290         refcount_acquire(&obj->hold_count);
291         if (vm_object_lock_try(obj) == 0) {
292                 if (refcount_release(&obj->hold_count)) {
293                         if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD))
294                                 kfree(obj, M_VM_OBJECT);
295                 }
296                 return(0);
297         }
298
299 #if defined(DEBUG_LOCKS)
300         debugvm_object_add(obj, file, line, 1);
301 #endif
302         return(1);
303 }
304
305 void
306 VMOBJDEBUG(vm_object_hold_shared)(vm_object_t obj VMOBJDBARGS)
307 {
308         KKASSERT(obj != NULL);
309
310         /*
311          * Object must be held (object allocation is stable due to callers
312          * context, typically already holding the token on a parent object)
313          * prior to potentially blocking on the lock, otherwise the object
314          * can get ripped away from us.
315          */
316         refcount_acquire(&obj->hold_count);
317         vm_object_lock_shared(obj);
318
319 #if defined(DEBUG_LOCKS)
320         debugvm_object_add(obj, file, line, 1);
321 #endif
322 }
323
324 /*
325  * Drop the token and hold_count on the object.
326  *
327  * WARNING! Token might be shared.
328  */
329 void
330 VMOBJDEBUG(vm_object_drop)(vm_object_t obj VMOBJDBARGS)
331 {
332         if (obj == NULL)
333                 return;
334
335         /*
336          * No new holders should be possible once we drop hold_count 1->0 as
337          * there is no longer any way to reference the object.
338          */
339         KKASSERT(obj->hold_count > 0);
340         if (refcount_release(&obj->hold_count)) {
341 #if defined(DEBUG_LOCKS)
342                 debugvm_object_add(obj, file, line, -1);
343 #endif
344
345                 if (obj->ref_count == 0 && (obj->flags & OBJ_DEAD)) {
346                         vm_object_unlock(obj);
347                         kfree(obj, M_VM_OBJECT);
348                 } else {
349                         vm_object_unlock(obj);
350                 }
351         } else {
352 #if defined(DEBUG_LOCKS)
353                 debugvm_object_add(obj, file, line, -1);
354 #endif
355                 vm_object_unlock(obj);
356         }
357 }
358
359 /*
360  * Initialize a freshly allocated object, returning a held object.
361  *
362  * Used only by vm_object_allocate(), zinitna() and vm_object_init().
363  *
364  * No requirements.
365  */
366 void
367 _vm_object_allocate(objtype_t type, vm_pindex_t size, vm_object_t object)
368 {
369         struct vm_object_hash *hash;
370
371         RB_INIT(&object->rb_memq);
372         LIST_INIT(&object->shadow_head);
373         lwkt_token_init(&object->token, "vmobj");
374
375         object->type = type;
376         object->size = size;
377         object->ref_count = 1;
378         object->memattr = VM_MEMATTR_DEFAULT;
379         object->hold_count = 0;
380         object->flags = 0;
381         if ((object->type == OBJT_DEFAULT) || (object->type == OBJT_SWAP))
382                 vm_object_set_flag(object, OBJ_ONEMAPPING);
383         object->paging_in_progress = 0;
384         object->resident_page_count = 0;
385         object->shadow_count = 0;
386         /* cpu localization twist */
387         object->pg_color = vm_quickcolor();
388         object->handle = NULL;
389         object->backing_object = NULL;
390         object->backing_object_offset = (vm_ooffset_t)0;
391
392         object->generation++;
393         object->swblock_count = 0;
394         RB_INIT(&object->swblock_root);
395         vm_object_lock_init(object);
396         pmap_object_init(object);
397
398         vm_object_hold(object);
399
400         hash = VMOBJ_HASH(object);
401         lwkt_gettoken(&hash->token);
402         TAILQ_INSERT_TAIL(&hash->list, object, object_list);
403         lwkt_reltoken(&hash->token);
404 }
405
406 /*
407  * Initialize a VM object.
408  */
409 void
410 vm_object_init(vm_object_t object, vm_pindex_t size)
411 {
412         _vm_object_allocate(OBJT_DEFAULT, size, object);
413         vm_object_drop(object);
414 }
415
416 /*
417  * Initialize the VM objects module.
418  *
419  * Called from the low level boot code only.  Note that this occurs before
420  * kmalloc is initialized so we cannot allocate any VM objects.
421  */
422 void
423 vm_object_init1(void)
424 {
425         int i;
426
427         for (i = 0; i < VMOBJ_HSIZE; ++i) {
428                 TAILQ_INIT(&vm_object_hash[i].list);
429                 lwkt_token_init(&vm_object_hash[i].token, "vmobjlst");
430         }
431
432         _vm_object_allocate(OBJT_DEFAULT, OFF_TO_IDX(KvaEnd),
433                             &kernel_object);
434         vm_object_drop(&kernel_object);
435 }
436
437 void
438 vm_object_init2(void)
439 {
440         kmalloc_set_unlimited(M_VM_OBJECT);
441 }
442
443 /*
444  * Allocate and return a new object of the specified type and size.
445  *
446  * No requirements.
447  */
448 vm_object_t
449 vm_object_allocate(objtype_t type, vm_pindex_t size)
450 {
451         vm_object_t obj;
452
453         obj = kmalloc(sizeof(*obj), M_VM_OBJECT, M_INTWAIT|M_ZERO);
454         _vm_object_allocate(type, size, obj);
455         vm_object_drop(obj);
456
457         return (obj);
458 }
459
460 /*
461  * This version returns a held object, allowing further atomic initialization
462  * of the object.
463  */
464 vm_object_t
465 vm_object_allocate_hold(objtype_t type, vm_pindex_t size)
466 {
467         vm_object_t obj;
468
469         obj = kmalloc(sizeof(*obj), M_VM_OBJECT, M_INTWAIT|M_ZERO);
470         _vm_object_allocate(type, size, obj);
471
472         return (obj);
473 }
474
475 /*
476  * Add an additional reference to a vm_object.  The object must already be
477  * held.  The original non-lock version is no longer supported.  The object
478  * must NOT be chain locked by anyone at the time the reference is added.
479  *
480  * Referencing a chain-locked object can blow up the fairly sensitive
481  * ref_count and shadow_count tests in the deallocator.  Most callers
482  * will call vm_object_chain_wait() prior to calling
483  * vm_object_reference_locked() to avoid the case.
484  *
485  * The object must be held, but may be held shared if desired (hence why
486  * we use an atomic op).
487  */
488 void
489 VMOBJDEBUG(vm_object_reference_locked)(vm_object_t object VMOBJDBARGS)
490 {
491         KKASSERT(object != NULL);
492         ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
493         KKASSERT((object->chainlk & (CHAINLK_EXCL | CHAINLK_MASK)) == 0);
494         atomic_add_int(&object->ref_count, 1);
495         if (object->type == OBJT_VNODE) {
496                 vref(object->handle);
497                 /* XXX what if the vnode is being destroyed? */
498         }
499 #if defined(DEBUG_LOCKS)
500         debugvm_object_add(object, file, line, 1);
501 #endif
502 }
503
504 /*
505  * This version is only allowed for vnode objects.
506  */
507 void
508 VMOBJDEBUG(vm_object_reference_quick)(vm_object_t object VMOBJDBARGS)
509 {
510         KKASSERT(object->type == OBJT_VNODE);
511         atomic_add_int(&object->ref_count, 1);
512         vref(object->handle);
513 #if defined(DEBUG_LOCKS)
514         debugvm_object_add(object, file, line, 1);
515 #endif
516 }
517
518 /*
519  * Object OBJ_CHAINLOCK lock handling.
520  *
521  * The caller can chain-lock backing objects recursively and then
522  * use vm_object_chain_release_all() to undo the whole chain.
523  *
524  * Chain locks are used to prevent collapses and are only applicable
525  * to OBJT_DEFAULT and OBJT_SWAP objects.  Chain locking operations
526  * on other object types are ignored.  This is also important because
527  * it allows e.g. the vnode underlying a memory mapping to take concurrent
528  * faults.
529  *
530  * The object must usually be held on entry, though intermediate
531  * objects need not be held on release.  The object must be held exclusively,
532  * NOT shared.  Note that the prefault path checks the shared state and
533  * avoids using the chain functions.
534  */
535 void
536 vm_object_chain_wait(vm_object_t object, int shared)
537 {
538         ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
539         for (;;) {
540                 uint32_t chainlk = object->chainlk;
541
542                 cpu_ccfence();
543                 if (shared) {
544                         if (chainlk & (CHAINLK_EXCL | CHAINLK_EXCLREQ)) {
545                                 tsleep_interlock(object, 0);
546                                 if (atomic_cmpset_int(&object->chainlk,
547                                                       chainlk,
548                                                       chainlk | CHAINLK_WAIT)) {
549                                         tsleep(object, PINTERLOCKED,
550                                                "objchns", 0);
551                                 }
552                                 /* retry */
553                         } else {
554                                 break;
555                         }
556                         /* retry */
557                 } else {
558                         if (chainlk & (CHAINLK_MASK | CHAINLK_EXCL)) {
559                                 tsleep_interlock(object, 0);
560                                 if (atomic_cmpset_int(&object->chainlk,
561                                                       chainlk,
562                                                       chainlk | CHAINLK_WAIT))
563                                 {
564                                         tsleep(object, PINTERLOCKED,
565                                                "objchnx", 0);
566                                 }
567                                 /* retry */
568                         } else {
569                                 if (atomic_cmpset_int(&object->chainlk,
570                                                       chainlk,
571                                                       chainlk & ~CHAINLK_WAIT))
572                                 {
573                                         if (chainlk & CHAINLK_WAIT)
574                                                 wakeup(object);
575                                         break;
576                                 }
577                                 /* retry */
578                         }
579                 }
580                 /* retry */
581         }
582 }
583
584 void
585 vm_object_chain_acquire(vm_object_t object, int shared)
586 {
587         if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP)
588                 return;
589         if (vm_shared_fault == 0)
590                 shared = 0;
591
592         for (;;) {
593                 uint32_t chainlk = object->chainlk;
594
595                 cpu_ccfence();
596                 if (shared) {
597                         if (chainlk & (CHAINLK_EXCL | CHAINLK_EXCLREQ)) {
598                                 tsleep_interlock(object, 0);
599                                 if (atomic_cmpset_int(&object->chainlk,
600                                                       chainlk,
601                                                       chainlk | CHAINLK_WAIT)) {
602                                         tsleep(object, PINTERLOCKED,
603                                                "objchns", 0);
604                                 }
605                                 /* retry */
606                         } else if (atomic_cmpset_int(&object->chainlk,
607                                               chainlk, chainlk + 1)) {
608                                 break;
609                         }
610                         /* retry */
611                 } else {
612                         if (chainlk & (CHAINLK_MASK | CHAINLK_EXCL)) {
613                                 tsleep_interlock(object, 0);
614                                 if (atomic_cmpset_int(&object->chainlk,
615                                                       chainlk,
616                                                       chainlk |
617                                                        CHAINLK_WAIT |
618                                                        CHAINLK_EXCLREQ)) {
619                                         tsleep(object, PINTERLOCKED,
620                                                "objchnx", 0);
621                                 }
622                                 /* retry */
623                         } else {
624                                 if (atomic_cmpset_int(&object->chainlk,
625                                                       chainlk,
626                                                       (chainlk | CHAINLK_EXCL) &
627                                                       ~(CHAINLK_EXCLREQ |
628                                                         CHAINLK_WAIT))) {
629                                         if (chainlk & CHAINLK_WAIT)
630                                                 wakeup(object);
631                                         break;
632                                 }
633                                 /* retry */
634                         }
635                 }
636                 /* retry */
637         }
638 }
639
640 void
641 vm_object_chain_release(vm_object_t object)
642 {
643         /*ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));*/
644         if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP)
645                 return;
646         KKASSERT(object->chainlk & (CHAINLK_MASK | CHAINLK_EXCL));
647         for (;;) {
648                 uint32_t chainlk = object->chainlk;
649
650                 cpu_ccfence();
651                 if (chainlk & CHAINLK_MASK) {
652                         if ((chainlk & CHAINLK_MASK) == 1 &&
653                             atomic_cmpset_int(&object->chainlk,
654                                               chainlk,
655                                               (chainlk - 1) & ~CHAINLK_WAIT)) {
656                                 if (chainlk & CHAINLK_WAIT)
657                                         wakeup(object);
658                                 break;
659                         }
660                         if ((chainlk & CHAINLK_MASK) > 1 &&
661                             atomic_cmpset_int(&object->chainlk,
662                                               chainlk, chainlk - 1)) {
663                                 break;
664                         }
665                         /* retry */
666                 } else {
667                         KKASSERT(chainlk & CHAINLK_EXCL);
668                         if (atomic_cmpset_int(&object->chainlk,
669                                               chainlk,
670                                               chainlk & ~(CHAINLK_EXCL |
671                                                           CHAINLK_WAIT))) {
672                                 if (chainlk & CHAINLK_WAIT)
673                                         wakeup(object);
674                                 break;
675                         }
676                 }
677         }
678 }
679
680 /*
681  * Release the chain from first_object through and including stopobj.
682  * The caller is typically holding the first and last object locked
683  * (shared or exclusive) to prevent destruction races.
684  *
685  * We release stopobj first as an optimization as this object is most
686  * likely to be shared across multiple processes.
687  */
688 void
689 vm_object_chain_release_all(vm_object_t first_object, vm_object_t stopobj)
690 {
691         vm_object_t backing_object;
692         vm_object_t object;
693
694         vm_object_chain_release(stopobj);
695         object = first_object;
696
697         while (object != stopobj) {
698                 KKASSERT(object);
699                 backing_object = object->backing_object;
700                 vm_object_chain_release(object);
701                 object = backing_object;
702         }
703 }
704
705 /*
706  * Dereference an object and its underlying vnode.  The object may be
707  * held shared.  On return the object will remain held.
708  *
709  * This function may return a vnode in *vpp which the caller must release
710  * after the caller drops its own lock.  If vpp is NULL, we assume that
711  * the caller was holding an exclusive lock on the object and we vrele()
712  * the vp ourselves.
713  */
714 static void
715 VMOBJDEBUG(vm_object_vndeallocate)(vm_object_t object, struct vnode **vpp
716                                    VMOBJDBARGS)
717 {
718         struct vnode *vp = (struct vnode *) object->handle;
719
720         KASSERT(object->type == OBJT_VNODE,
721             ("vm_object_vndeallocate: not a vnode object"));
722         KASSERT(vp != NULL, ("vm_object_vndeallocate: missing vp"));
723         ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
724 #ifdef INVARIANTS
725         if (object->ref_count == 0) {
726                 vprint("vm_object_vndeallocate", vp);
727                 panic("vm_object_vndeallocate: bad object reference count");
728         }
729 #endif
730         for (;;) {
731                 int count = object->ref_count;
732                 cpu_ccfence();
733                 if (count == 1) {
734                         vm_object_upgrade(object);
735                         if (atomic_cmpset_int(&object->ref_count, count, 0)) {
736                                 vclrflags(vp, VTEXT);
737                                 break;
738                         }
739                 } else {
740                         if (atomic_cmpset_int(&object->ref_count,
741                                               count, count - 1)) {
742                                 break;
743                         }
744                 }
745                 /* retry */
746         }
747 #if defined(DEBUG_LOCKS)
748         debugvm_object_add(object, file, line, -1);
749 #endif
750
751         /*
752          * vrele or return the vp to vrele.  We can only safely vrele(vp)
753          * if the object was locked exclusively.  But there are two races
754          * here.
755          *
756          * We had to upgrade the object above to safely clear VTEXT
757          * but the alternative path where the shared lock is retained
758          * can STILL race to 0 in other paths and cause our own vrele()
759          * to terminate the vnode.  We can't allow that if the VM object
760          * is still locked shared.
761          */
762         if (vpp)
763                 *vpp = vp;
764         else
765                 vrele(vp);
766 }
767
768 /*
769  * Release a reference to the specified object, gained either through a
770  * vm_object_allocate or a vm_object_reference call.  When all references
771  * are gone, storage associated with this object may be relinquished.
772  *
773  * The caller does not have to hold the object locked but must have control
774  * over the reference in question in order to guarantee that the object
775  * does not get ripped out from under us.
776  *
777  * XXX Currently all deallocations require an exclusive lock.
778  */
779 void
780 VMOBJDEBUG(vm_object_deallocate)(vm_object_t object VMOBJDBARGS)
781 {
782         struct vnode *vp;
783         int count;
784
785         if (object == NULL)
786                 return;
787
788         for (;;) {
789                 count = object->ref_count;
790                 cpu_ccfence();
791
792                 /*
793                  * If decrementing the count enters into special handling
794                  * territory (0, 1, or 2) we have to do it the hard way.
795                  * Fortunate though, objects with only a few refs like this
796                  * are not likely to be heavily contended anyway.
797                  *
798                  * For vnode objects we only care about 1->0 transitions.
799                  */
800                 if (count <= 3 || (object->type == OBJT_VNODE && count <= 1)) {
801 #if defined(DEBUG_LOCKS)
802                         debugvm_object_add(object, file, line, 0);
803 #endif
804                         vm_object_hold(object);
805                         vm_object_deallocate_locked(object);
806                         vm_object_drop(object);
807                         break;
808                 }
809
810                 /*
811                  * Try to decrement ref_count without acquiring a hold on
812                  * the object.  This is particularly important for the exec*()
813                  * and exit*() code paths because the program binary may
814                  * have a great deal of sharing and an exclusive lock will
815                  * crowbar performance in those circumstances.
816                  */
817                 if (object->type == OBJT_VNODE) {
818                         vp = (struct vnode *)object->handle;
819                         if (atomic_cmpset_int(&object->ref_count,
820                                               count, count - 1)) {
821 #if defined(DEBUG_LOCKS)
822                                 debugvm_object_add(object, file, line, -1);
823 #endif
824
825                                 vrele(vp);
826                                 break;
827                         }
828                         /* retry */
829                 } else {
830                         if (atomic_cmpset_int(&object->ref_count,
831                                               count, count - 1)) {
832 #if defined(DEBUG_LOCKS)
833                                 debugvm_object_add(object, file, line, -1);
834 #endif
835                                 break;
836                         }
837                         /* retry */
838                 }
839                 /* retry */
840         }
841 }
842
843 void
844 VMOBJDEBUG(vm_object_deallocate_locked)(vm_object_t object VMOBJDBARGS)
845 {
846         struct vm_object_dealloc_list *dlist = NULL;
847         struct vm_object_dealloc_list *dtmp;
848         vm_object_t temp;
849         int must_drop = 0;
850
851         /*
852          * We may chain deallocate object, but additional objects may
853          * collect on the dlist which also have to be deallocated.  We
854          * must avoid a recursion, vm_object chains can get deep.
855          */
856
857 again:
858         while (object != NULL) {
859                 /*
860                  * vnode case, caller either locked the object exclusively
861                  * or this is a recursion with must_drop != 0 and the vnode
862                  * object will be locked shared.
863                  *
864                  * If locked shared we have to drop the object before we can
865                  * call vrele() or risk a shared/exclusive livelock.
866                  */
867                 if (object->type == OBJT_VNODE) {
868                         ASSERT_LWKT_TOKEN_HELD(&object->token);
869                         if (must_drop) {
870                                 struct vnode *tmp_vp;
871
872                                 vm_object_vndeallocate(object, &tmp_vp);
873                                 vm_object_drop(object);
874                                 must_drop = 0;
875                                 object = NULL;
876                                 vrele(tmp_vp);
877                         } else {
878                                 vm_object_vndeallocate(object, NULL);
879                         }
880                         break;
881                 }
882                 ASSERT_LWKT_TOKEN_HELD_EXCL(&object->token);
883
884                 /*
885                  * Normal case (object is locked exclusively)
886                  */
887                 if (object->ref_count == 0) {
888                         panic("vm_object_deallocate: object deallocated "
889                               "too many times: %d", object->type);
890                 }
891                 if (object->ref_count > 2) {
892                         atomic_add_int(&object->ref_count, -1);
893 #if defined(DEBUG_LOCKS)
894                         debugvm_object_add(object, file, line, -1);
895 #endif
896                         break;
897                 }
898
899                 /*
900                  * Here on ref_count of one or two, which are special cases for
901                  * objects.
902                  *
903                  * Nominal ref_count > 1 case if the second ref is not from
904                  * a shadow.
905                  *
906                  * (ONEMAPPING only applies to DEFAULT AND SWAP objects)
907                  */
908                 if (object->ref_count == 2 && object->shadow_count == 0) {
909                         if (object->type == OBJT_DEFAULT ||
910                             object->type == OBJT_SWAP) {
911                                 vm_object_set_flag(object, OBJ_ONEMAPPING);
912                         }
913                         atomic_add_int(&object->ref_count, -1);
914 #if defined(DEBUG_LOCKS)
915                         debugvm_object_add(object, file, line, -1);
916 #endif
917                         break;
918                 }
919
920                 /*
921                  * If the second ref is from a shadow we chain along it
922                  * upwards if object's handle is exhausted.
923                  *
924                  * We have to decrement object->ref_count before potentially
925                  * collapsing the first shadow object or the collapse code
926                  * will not be able to handle the degenerate case to remove
927                  * object.  However, if we do it too early the object can
928                  * get ripped out from under us.
929                  */
930                 if (object->ref_count == 2 && object->shadow_count == 1 &&
931                     object->handle == NULL && (object->type == OBJT_DEFAULT ||
932                                                object->type == OBJT_SWAP)) {
933                         temp = LIST_FIRST(&object->shadow_head);
934                         KKASSERT(temp != NULL);
935                         vm_object_hold(temp);
936
937                         /*
938                          * Wait for any paging to complete so the collapse
939                          * doesn't (or isn't likely to) qcollapse.  pip
940                          * waiting must occur before we acquire the
941                          * chainlock.
942                          */
943                         while (
944                                 temp->paging_in_progress ||
945                                 object->paging_in_progress
946                         ) {
947                                 vm_object_pip_wait(temp, "objde1");
948                                 vm_object_pip_wait(object, "objde2");
949                         }
950
951                         /*
952                          * If the parent is locked we have to give up, as
953                          * otherwise we would be acquiring locks in the
954                          * wrong order and potentially deadlock.
955                          */
956                         if (temp->chainlk & (CHAINLK_EXCL | CHAINLK_MASK)) {
957                                 vm_object_drop(temp);
958                                 goto skip;
959                         }
960                         vm_object_chain_acquire(temp, 0);
961
962                         /*
963                          * Recheck/retry after the hold and the paging
964                          * wait, both of which can block us.
965                          */
966                         if (object->ref_count != 2 ||
967                             object->shadow_count != 1 ||
968                             object->handle ||
969                             LIST_FIRST(&object->shadow_head) != temp ||
970                             (object->type != OBJT_DEFAULT &&
971                              object->type != OBJT_SWAP)) {
972                                 vm_object_chain_release(temp);
973                                 vm_object_drop(temp);
974                                 continue;
975                         }
976
977                         /*
978                          * We can safely drop object's ref_count now.
979                          */
980                         KKASSERT(object->ref_count == 2);
981                         atomic_add_int(&object->ref_count, -1);
982 #if defined(DEBUG_LOCKS)
983                         debugvm_object_add(object, file, line, -1);
984 #endif
985
986                         /*
987                          * If our single parent is not collapseable just
988                          * decrement ref_count (2->1) and stop.
989                          */
990                         if (temp->handle || (temp->type != OBJT_DEFAULT &&
991                                              temp->type != OBJT_SWAP)) {
992                                 vm_object_chain_release(temp);
993                                 vm_object_drop(temp);
994                                 break;
995                         }
996
997                         /*
998                          * At this point we have already dropped object's
999                          * ref_count so it is possible for a race to
1000                          * deallocate obj out from under us.  Any collapse
1001                          * will re-check the situation.  We must not block
1002                          * until we are able to collapse.
1003                          *
1004                          * Bump temp's ref_count to avoid an unwanted
1005                          * degenerate recursion (can't call
1006                          * vm_object_reference_locked() because it asserts
1007                          * that CHAINLOCK is not set).
1008                          */
1009                         atomic_add_int(&temp->ref_count, 1);
1010                         KKASSERT(temp->ref_count > 1);
1011
1012                         /*
1013                          * Collapse temp, then deallocate the extra ref
1014                          * formally.
1015                          */
1016                         vm_object_collapse(temp, &dlist);
1017                         vm_object_chain_release(temp);
1018                         if (must_drop) {
1019                                 vm_object_lock_swap();
1020                                 vm_object_drop(object);
1021                         }
1022                         object = temp;
1023                         must_drop = 1;
1024                         continue;
1025                 }
1026
1027                 /*
1028                  * Drop the ref and handle termination on the 1->0 transition.
1029                  * We may have blocked above so we have to recheck.
1030                  */
1031 skip:
1032                 KKASSERT(object->ref_count != 0);
1033                 if (object->ref_count >= 2) {
1034                         atomic_add_int(&object->ref_count, -1);
1035 #if defined(DEBUG_LOCKS)
1036                         debugvm_object_add(object, file, line, -1);
1037 #endif
1038                         break;
1039                 }
1040                 KKASSERT(object->ref_count == 1);
1041
1042                 /*
1043                  * 1->0 transition.  Chain through the backing_object.
1044                  * Maintain the ref until we've located the backing object,
1045                  * then re-check.
1046                  */
1047                 while ((temp = object->backing_object) != NULL) {
1048                         if (temp->type == OBJT_VNODE)
1049                                 vm_object_hold_shared(temp);
1050                         else
1051                                 vm_object_hold(temp);
1052                         if (temp == object->backing_object)
1053                                 break;
1054                         vm_object_drop(temp);
1055                 }
1056
1057                 /*
1058                  * 1->0 transition verified, retry if ref_count is no longer
1059                  * 1.  Otherwise disconnect the backing_object (temp) and
1060                  * clean up.
1061                  */
1062                 if (object->ref_count != 1) {
1063                         vm_object_drop(temp);
1064                         continue;
1065                 }
1066
1067                 /*
1068                  * It shouldn't be possible for the object to be chain locked
1069                  * if we're removing the last ref on it.
1070                  *
1071                  * Removing object from temp's shadow list requires dropping
1072                  * temp, which we will do on loop.
1073                  *
1074                  * NOTE! vnodes do not use the shadow list, but still have
1075                  *       the backing_object reference.
1076                  */
1077                 KKASSERT((object->chainlk & (CHAINLK_EXCL|CHAINLK_MASK)) == 0);
1078
1079                 if (temp) {
1080                         if (object->flags & OBJ_ONSHADOW) {
1081                                 LIST_REMOVE(object, shadow_list);
1082                                 temp->shadow_count--;
1083                                 temp->generation++;
1084                                 vm_object_clear_flag(object, OBJ_ONSHADOW);
1085                         }
1086                         object->backing_object = NULL;
1087                 }
1088
1089                 atomic_add_int(&object->ref_count, -1);
1090                 if ((object->flags & OBJ_DEAD) == 0)
1091                         vm_object_terminate(object);
1092                 if (must_drop && temp)
1093                         vm_object_lock_swap();
1094                 if (must_drop)
1095                         vm_object_drop(object);
1096                 object = temp;
1097                 must_drop = 1;
1098         }
1099
1100         if (must_drop && object)
1101                 vm_object_drop(object);
1102
1103         /*
1104          * Additional tail recursion on dlist.  Avoid a recursion.  Objects
1105          * on the dlist have a hold count but are not locked.
1106          */
1107         if ((dtmp = dlist) != NULL) {
1108                 dlist = dtmp->next;
1109                 object = dtmp->object;
1110                 kfree(dtmp, M_TEMP);
1111
1112                 vm_object_lock(object); /* already held, add lock */
1113                 must_drop = 1;          /* and we're responsible for it */
1114                 goto again;
1115         }
1116 }
1117
1118 /*
1119  * Destroy the specified object, freeing up related resources.
1120  *
1121  * The object must have zero references.
1122  *
1123  * The object must held.  The caller is responsible for dropping the object
1124  * after terminate returns.  Terminate does NOT drop the object.
1125  */
1126 static int vm_object_terminate_callback(vm_page_t p, void *data);
1127
1128 void
1129 vm_object_terminate(vm_object_t object)
1130 {
1131         struct rb_vm_page_scan_info info;
1132         struct vm_object_hash *hash;
1133
1134         /*
1135          * Make sure no one uses us.  Once we set OBJ_DEAD we should be
1136          * able to safely block.
1137          */
1138         ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1139         KKASSERT((object->flags & OBJ_DEAD) == 0);
1140         vm_object_set_flag(object, OBJ_DEAD);
1141
1142         /*
1143          * Wait for the pageout daemon to be done with the object
1144          */
1145         vm_object_pip_wait(object, "objtrm1");
1146
1147         KASSERT(!object->paging_in_progress,
1148                 ("vm_object_terminate: pageout in progress"));
1149
1150         /*
1151          * Clean and free the pages, as appropriate. All references to the
1152          * object are gone, so we don't need to lock it.
1153          */
1154         if (object->type == OBJT_VNODE) {
1155                 struct vnode *vp;
1156
1157                 /*
1158                  * Clean pages and flush buffers.
1159                  *
1160                  * NOTE!  TMPFS buffer flushes do not typically flush the
1161                  *        actual page to swap as this would be highly
1162                  *        inefficient, and normal filesystems usually wrap
1163                  *        page flushes with buffer cache buffers.
1164                  *
1165                  *        To deal with this we have to call vinvalbuf() both
1166                  *        before and after the vm_object_page_clean().
1167                  */
1168                 vp = (struct vnode *) object->handle;
1169                 vinvalbuf(vp, V_SAVE, 0, 0);
1170                 vm_object_page_clean(object, 0, 0, OBJPC_SYNC);
1171                 vinvalbuf(vp, V_SAVE, 0, 0);
1172         }
1173
1174         /*
1175          * Wait for any I/O to complete, after which there had better not
1176          * be any references left on the object.
1177          */
1178         vm_object_pip_wait(object, "objtrm2");
1179
1180         if (object->ref_count != 0) {
1181                 panic("vm_object_terminate: object with references, "
1182                       "ref_count=%d", object->ref_count);
1183         }
1184
1185         /*
1186          * Cleanup any shared pmaps associated with this object.
1187          */
1188         pmap_object_free(object);
1189
1190         /*
1191          * Now free any remaining pages. For internal objects, this also
1192          * removes them from paging queues. Don't free wired pages, just
1193          * remove them from the object. 
1194          */
1195         info.count = 0;
1196         info.object = object;
1197         vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1198                                 vm_object_terminate_callback, &info);
1199
1200         /*
1201          * Let the pager know object is dead.
1202          */
1203         vm_pager_deallocate(object);
1204
1205         /*
1206          * Wait for the object hold count to hit 1, clean out pages as
1207          * we go.  vmobj_token interlocks any race conditions that might
1208          * pick the object up from the vm_object_list after we have cleared
1209          * rb_memq.
1210          */
1211         for (;;) {
1212                 if (RB_ROOT(&object->rb_memq) == NULL)
1213                         break;
1214                 kprintf("vm_object_terminate: Warning, object %p "
1215                         "still has %ld pages\n",
1216                         object, object->resident_page_count);
1217                 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1218                                         vm_object_terminate_callback, &info);
1219         }
1220
1221         /*
1222          * There had better not be any pages left
1223          */
1224         KKASSERT(object->resident_page_count == 0);
1225
1226         /*
1227          * Remove the object from the global object list.
1228          */
1229         hash = VMOBJ_HASH(object);
1230         lwkt_gettoken(&hash->token);
1231         TAILQ_REMOVE(&hash->list, object, object_list);
1232         lwkt_reltoken(&hash->token);
1233
1234         if (object->ref_count != 0) {
1235                 panic("vm_object_terminate2: object with references, "
1236                       "ref_count=%d", object->ref_count);
1237         }
1238
1239         /*
1240          * NOTE: The object hold_count is at least 1, so we cannot kfree()
1241          *       the object here.  See vm_object_drop().
1242          */
1243 }
1244
1245 /*
1246  * The caller must hold the object.
1247  */
1248 static int
1249 vm_object_terminate_callback(vm_page_t p, void *data)
1250 {
1251         struct rb_vm_page_scan_info *info = data;
1252         vm_object_t object;
1253
1254         if ((++info->count & 63) == 0)
1255                 lwkt_user_yield();
1256         object = p->object;
1257         if (object != info->object) {
1258                 kprintf("vm_object_terminate_callback: obj/pg race %p/%p\n",
1259                         info->object, p);
1260                 return(0);
1261         }
1262         vm_page_busy_wait(p, TRUE, "vmpgtrm");
1263         if (object != p->object) {
1264                 kprintf("vm_object_terminate: Warning: Encountered "
1265                         "busied page %p on queue %d\n", p, p->queue);
1266                 vm_page_wakeup(p);
1267         } else if (p->wire_count == 0) {
1268                 /*
1269                  * NOTE: p->dirty and PG_NEED_COMMIT are ignored.
1270                  */
1271                 vm_page_free(p);
1272                 mycpu->gd_cnt.v_pfree++;
1273         } else {
1274                 if (p->queue != PQ_NONE)
1275                         kprintf("vm_object_terminate: Warning: Encountered "
1276                                 "wired page %p on queue %d\n", p, p->queue);
1277                 vm_page_remove(p);
1278                 vm_page_wakeup(p);
1279         }
1280         return(0);
1281 }
1282
1283 /*
1284  * Clean all dirty pages in the specified range of object.  Leaves page
1285  * on whatever queue it is currently on.   If NOSYNC is set then do not
1286  * write out pages with PG_NOSYNC set (originally comes from MAP_NOSYNC),
1287  * leaving the object dirty.
1288  *
1289  * When stuffing pages asynchronously, allow clustering.  XXX we need a
1290  * synchronous clustering mode implementation.
1291  *
1292  * Odd semantics: if start == end, we clean everything.
1293  *
1294  * The object must be locked? XXX
1295  */
1296 static int vm_object_page_clean_pass1(struct vm_page *p, void *data);
1297 static int vm_object_page_clean_pass2(struct vm_page *p, void *data);
1298
1299 void
1300 vm_object_page_clean(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
1301                      int flags)
1302 {
1303         struct rb_vm_page_scan_info info;
1304         struct vnode *vp;
1305         int wholescan;
1306         int pagerflags;
1307         int generation;
1308
1309         vm_object_hold(object);
1310         if (object->type != OBJT_VNODE ||
1311             (object->flags & OBJ_MIGHTBEDIRTY) == 0) {
1312                 vm_object_drop(object);
1313                 return;
1314         }
1315
1316         pagerflags = (flags & (OBJPC_SYNC | OBJPC_INVAL)) ? 
1317                         VM_PAGER_PUT_SYNC : VM_PAGER_CLUSTER_OK;
1318         pagerflags |= (flags & OBJPC_INVAL) ? VM_PAGER_PUT_INVAL : 0;
1319
1320         vp = object->handle;
1321
1322         /*
1323          * Interlock other major object operations.  This allows us to 
1324          * temporarily clear OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY.
1325          */
1326         vm_object_set_flag(object, OBJ_CLEANING);
1327
1328         /*
1329          * Handle 'entire object' case
1330          */
1331         info.start_pindex = start;
1332         if (end == 0) {
1333                 info.end_pindex = object->size - 1;
1334         } else {
1335                 info.end_pindex = end - 1;
1336         }
1337         wholescan = (start == 0 && info.end_pindex == object->size - 1);
1338         info.limit = flags;
1339         info.pagerflags = pagerflags;
1340         info.object = object;
1341
1342         /*
1343          * If cleaning the entire object do a pass to mark the pages read-only.
1344          * If everything worked out ok, clear OBJ_WRITEABLE and
1345          * OBJ_MIGHTBEDIRTY.
1346          */
1347         if (wholescan) {
1348                 info.error = 0;
1349                 info.count = 0;
1350                 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1351                                         vm_object_page_clean_pass1, &info);
1352                 if (info.error == 0) {
1353                         vm_object_clear_flag(object,
1354                                              OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
1355                         if (object->type == OBJT_VNODE &&
1356                             (vp = (struct vnode *)object->handle) != NULL) {
1357                                 /*
1358                                  * Use new-style interface to clear VISDIRTY
1359                                  * because the vnode is not necessarily removed
1360                                  * from the syncer list(s) as often as it was
1361                                  * under the old interface, which can leave
1362                                  * the vnode on the syncer list after reclaim.
1363                                  */
1364                                 vclrobjdirty(vp);
1365                         }
1366                 }
1367         }
1368
1369         /*
1370          * Do a pass to clean all the dirty pages we find.
1371          */
1372         do {
1373                 info.error = 0;
1374                 info.count = 0;
1375                 generation = object->generation;
1376                 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1377                                         vm_object_page_clean_pass2, &info);
1378         } while (info.error || generation != object->generation);
1379
1380         vm_object_clear_flag(object, OBJ_CLEANING);
1381         vm_object_drop(object);
1382 }
1383
1384 /*
1385  * The caller must hold the object.
1386  */
1387 static 
1388 int
1389 vm_object_page_clean_pass1(struct vm_page *p, void *data)
1390 {
1391         struct rb_vm_page_scan_info *info = data;
1392
1393         if ((++info->count & 63) == 0)
1394                 lwkt_user_yield();
1395         if (p->object != info->object ||
1396             p->pindex < info->start_pindex ||
1397             p->pindex > info->end_pindex) {
1398                 kprintf("vm_object_page_clean_pass1: obj/pg race %p/%p\n",
1399                         info->object, p);
1400                 return(0);
1401         }
1402         vm_page_flag_set(p, PG_CLEANCHK);
1403         if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1404                 info->error = 1;
1405         } else if (vm_page_busy_try(p, FALSE) == 0) {
1406                 if (p->object == info->object)
1407                         vm_page_protect(p, VM_PROT_READ);
1408                 vm_page_wakeup(p);
1409         } else {
1410                 info->error = 1;
1411         }
1412         return(0);
1413 }
1414
1415 /*
1416  * The caller must hold the object
1417  */
1418 static 
1419 int
1420 vm_object_page_clean_pass2(struct vm_page *p, void *data)
1421 {
1422         struct rb_vm_page_scan_info *info = data;
1423         int generation;
1424
1425         if (p->object != info->object ||
1426             p->pindex < info->start_pindex ||
1427             p->pindex > info->end_pindex) {
1428                 kprintf("vm_object_page_clean_pass2: obj/pg race %p/%p\n",
1429                         info->object, p);
1430                 return(0);
1431         }
1432
1433         /*
1434          * Do not mess with pages that were inserted after we started
1435          * the cleaning pass.
1436          */
1437         if ((p->flags & PG_CLEANCHK) == 0)
1438                 goto done;
1439
1440         generation = info->object->generation;
1441         vm_page_busy_wait(p, TRUE, "vpcwai");
1442
1443         if (p->object != info->object ||
1444             p->pindex < info->start_pindex ||
1445             p->pindex > info->end_pindex ||
1446             info->object->generation != generation) {
1447                 info->error = 1;
1448                 vm_page_wakeup(p);
1449                 goto done;
1450         }
1451
1452         /*
1453          * Before wasting time traversing the pmaps, check for trivial
1454          * cases where the page cannot be dirty.
1455          */
1456         if (p->valid == 0 || (p->queue - p->pc) == PQ_CACHE) {
1457                 KKASSERT((p->dirty & p->valid) == 0 &&
1458                          (p->flags & PG_NEED_COMMIT) == 0);
1459                 vm_page_wakeup(p);
1460                 goto done;
1461         }
1462
1463         /*
1464          * Check whether the page is dirty or not.  The page has been set
1465          * to be read-only so the check will not race a user dirtying the
1466          * page.
1467          */
1468         vm_page_test_dirty(p);
1469         if ((p->dirty & p->valid) == 0 && (p->flags & PG_NEED_COMMIT) == 0) {
1470                 vm_page_flag_clear(p, PG_CLEANCHK);
1471                 vm_page_wakeup(p);
1472                 goto done;
1473         }
1474
1475         /*
1476          * If we have been asked to skip nosync pages and this is a
1477          * nosync page, skip it.  Note that the object flags were
1478          * not cleared in this case (because pass1 will have returned an
1479          * error), so we do not have to set them.
1480          */
1481         if ((info->limit & OBJPC_NOSYNC) && (p->flags & PG_NOSYNC)) {
1482                 vm_page_flag_clear(p, PG_CLEANCHK);
1483                 vm_page_wakeup(p);
1484                 goto done;
1485         }
1486
1487         /*
1488          * Flush as many pages as we can.  PG_CLEANCHK will be cleared on
1489          * the pages that get successfully flushed.  Set info->error if
1490          * we raced an object modification.
1491          */
1492         vm_object_page_collect_flush(info->object, p, info->pagerflags);
1493         /* vm_wait_nominal(); this can deadlock the system in syncer/pageout */
1494 done:
1495         if ((++info->count & 63) == 0)
1496                 lwkt_user_yield();
1497
1498         return(0);
1499 }
1500
1501 /*
1502  * Collect the specified page and nearby pages and flush them out.
1503  * The number of pages flushed is returned.  The passed page is busied
1504  * by the caller and we are responsible for its disposition.
1505  *
1506  * The caller must hold the object.
1507  */
1508 static void
1509 vm_object_page_collect_flush(vm_object_t object, vm_page_t p, int pagerflags)
1510 {
1511         int error;
1512         int is;
1513         int ib;
1514         int i;
1515         int page_base;
1516         vm_pindex_t pi;
1517         vm_page_t ma[BLIST_MAX_ALLOC];
1518
1519         ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1520
1521         pi = p->pindex;
1522         page_base = pi % BLIST_MAX_ALLOC;
1523         ma[page_base] = p;
1524         ib = page_base - 1;
1525         is = page_base + 1;
1526
1527         while (ib >= 0) {
1528                 vm_page_t tp;
1529
1530                 tp = vm_page_lookup_busy_try(object, pi - page_base + ib,
1531                                              TRUE, &error);
1532                 if (error)
1533                         break;
1534                 if (tp == NULL)
1535                         break;
1536                 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1537                     (tp->flags & PG_CLEANCHK) == 0) {
1538                         vm_page_wakeup(tp);
1539                         break;
1540                 }
1541                 if ((tp->queue - tp->pc) == PQ_CACHE) {
1542                         vm_page_flag_clear(tp, PG_CLEANCHK);
1543                         vm_page_wakeup(tp);
1544                         break;
1545                 }
1546                 vm_page_test_dirty(tp);
1547                 if ((tp->dirty & tp->valid) == 0 &&
1548                     (tp->flags & PG_NEED_COMMIT) == 0) {
1549                         vm_page_flag_clear(tp, PG_CLEANCHK);
1550                         vm_page_wakeup(tp);
1551                         break;
1552                 }
1553                 ma[ib] = tp;
1554                 --ib;
1555         }
1556         ++ib;   /* fixup */
1557
1558         while (is < BLIST_MAX_ALLOC &&
1559                pi - page_base + is < object->size) {
1560                 vm_page_t tp;
1561
1562                 tp = vm_page_lookup_busy_try(object, pi - page_base + is,
1563                                              TRUE, &error);
1564                 if (error)
1565                         break;
1566                 if (tp == NULL)
1567                         break;
1568                 if ((pagerflags & VM_PAGER_IGNORE_CLEANCHK) == 0 &&
1569                     (tp->flags & PG_CLEANCHK) == 0) {
1570                         vm_page_wakeup(tp);
1571                         break;
1572                 }
1573                 if ((tp->queue - tp->pc) == PQ_CACHE) {
1574                         vm_page_flag_clear(tp, PG_CLEANCHK);
1575                         vm_page_wakeup(tp);
1576                         break;
1577                 }
1578                 vm_page_test_dirty(tp);
1579                 if ((tp->dirty & tp->valid) == 0 &&
1580                     (tp->flags & PG_NEED_COMMIT) == 0) {
1581                         vm_page_flag_clear(tp, PG_CLEANCHK);
1582                         vm_page_wakeup(tp);
1583                         break;
1584                 }
1585                 ma[is] = tp;
1586                 ++is;
1587         }
1588
1589         /*
1590          * All pages in the ma[] array are busied now
1591          */
1592         for (i = ib; i < is; ++i) {
1593                 vm_page_flag_clear(ma[i], PG_CLEANCHK);
1594                 vm_page_hold(ma[i]);    /* XXX need this any more? */
1595         }
1596         vm_pageout_flush(&ma[ib], is - ib, pagerflags);
1597         for (i = ib; i < is; ++i)       /* XXX need this any more? */
1598                 vm_page_unhold(ma[i]);
1599 }
1600
1601 /*
1602  * Same as vm_object_pmap_copy, except range checking really
1603  * works, and is meant for small sections of an object.
1604  *
1605  * This code protects resident pages by making them read-only
1606  * and is typically called on a fork or split when a page
1607  * is converted to copy-on-write.  
1608  *
1609  * NOTE: If the page is already at VM_PROT_NONE, calling
1610  * vm_page_protect will have no effect.
1611  */
1612 void
1613 vm_object_pmap_copy_1(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1614 {
1615         vm_pindex_t idx;
1616         vm_page_t p;
1617
1618         if (object == NULL || (object->flags & OBJ_WRITEABLE) == 0)
1619                 return;
1620
1621         vm_object_hold(object);
1622         for (idx = start; idx < end; idx++) {
1623                 p = vm_page_lookup(object, idx);
1624                 if (p == NULL)
1625                         continue;
1626                 vm_page_protect(p, VM_PROT_READ);
1627         }
1628         vm_object_drop(object);
1629 }
1630
1631 /*
1632  * Removes all physical pages in the specified object range from all
1633  * physical maps.
1634  *
1635  * The object must *not* be locked.
1636  */
1637
1638 static int vm_object_pmap_remove_callback(vm_page_t p, void *data);
1639
1640 void
1641 vm_object_pmap_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end)
1642 {
1643         struct rb_vm_page_scan_info info;
1644
1645         if (object == NULL)
1646                 return;
1647         info.start_pindex = start;
1648         info.end_pindex = end - 1;
1649         info.count = 0;
1650         info.object = object;
1651
1652         vm_object_hold(object);
1653         vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
1654                                 vm_object_pmap_remove_callback, &info);
1655         if (start == 0 && end == object->size)
1656                 vm_object_clear_flag(object, OBJ_WRITEABLE);
1657         vm_object_drop(object);
1658 }
1659
1660 /*
1661  * The caller must hold the object
1662  */
1663 static int
1664 vm_object_pmap_remove_callback(vm_page_t p, void *data)
1665 {
1666         struct rb_vm_page_scan_info *info = data;
1667
1668         if ((++info->count & 63) == 0)
1669                 lwkt_user_yield();
1670
1671         if (info->object != p->object ||
1672             p->pindex < info->start_pindex ||
1673             p->pindex > info->end_pindex) {
1674                 kprintf("vm_object_pmap_remove_callback: obj/pg race %p/%p\n",
1675                         info->object, p);
1676                 return(0);
1677         }
1678
1679         vm_page_protect(p, VM_PROT_NONE);
1680
1681         return(0);
1682 }
1683
1684 /*
1685  * Implements the madvise function at the object/page level.
1686  *
1687  * MADV_WILLNEED        (any object)
1688  *
1689  *      Activate the specified pages if they are resident.
1690  *
1691  * MADV_DONTNEED        (any object)
1692  *
1693  *      Deactivate the specified pages if they are resident.
1694  *
1695  * MADV_FREE    (OBJT_DEFAULT/OBJT_SWAP objects, OBJ_ONEMAPPING only)
1696  *
1697  *      Deactivate and clean the specified pages if they are
1698  *      resident.  This permits the process to reuse the pages
1699  *      without faulting or the kernel to reclaim the pages
1700  *      without I/O.
1701  *
1702  * No requirements.
1703  */
1704 void
1705 vm_object_madvise(vm_object_t object, vm_pindex_t pindex,
1706                   vm_pindex_t count, int advise)
1707 {
1708         vm_pindex_t end, tpindex;
1709         vm_object_t tobject;
1710         vm_object_t xobj;
1711         vm_page_t m;
1712         int error;
1713
1714         if (object == NULL)
1715                 return;
1716
1717         end = pindex + count;
1718
1719         vm_object_hold(object);
1720         tobject = object;
1721
1722         /*
1723          * Locate and adjust resident pages
1724          */
1725         for (; pindex < end; pindex += 1) {
1726 relookup:
1727                 if (tobject != object)
1728                         vm_object_drop(tobject);
1729                 tobject = object;
1730                 tpindex = pindex;
1731 shadowlookup:
1732                 /*
1733                  * MADV_FREE only operates on OBJT_DEFAULT or OBJT_SWAP pages
1734                  * and those pages must be OBJ_ONEMAPPING.
1735                  */
1736                 if (advise == MADV_FREE) {
1737                         if ((tobject->type != OBJT_DEFAULT &&
1738                              tobject->type != OBJT_SWAP) ||
1739                             (tobject->flags & OBJ_ONEMAPPING) == 0) {
1740                                 continue;
1741                         }
1742                 }
1743
1744                 m = vm_page_lookup_busy_try(tobject, tpindex, TRUE, &error);
1745
1746                 if (error) {
1747                         vm_page_sleep_busy(m, TRUE, "madvpo");
1748                         goto relookup;
1749                 }
1750                 if (m == NULL) {
1751                         /*
1752                          * There may be swap even if there is no backing page
1753                          */
1754                         if (advise == MADV_FREE && tobject->type == OBJT_SWAP)
1755                                 swap_pager_freespace(tobject, tpindex, 1);
1756
1757                         /*
1758                          * next object
1759                          */
1760                         while ((xobj = tobject->backing_object) != NULL) {
1761                                 KKASSERT(xobj != object);
1762                                 vm_object_hold(xobj);
1763                                 if (xobj == tobject->backing_object)
1764                                         break;
1765                                 vm_object_drop(xobj);
1766                         }
1767                         if (xobj == NULL)
1768                                 continue;
1769                         tpindex += OFF_TO_IDX(tobject->backing_object_offset);
1770                         if (tobject != object) {
1771                                 vm_object_lock_swap();
1772                                 vm_object_drop(tobject);
1773                         }
1774                         tobject = xobj;
1775                         goto shadowlookup;
1776                 }
1777
1778                 /*
1779                  * If the page is not in a normal active state, we skip it.
1780                  * If the page is not managed there are no page queues to
1781                  * mess with.  Things can break if we mess with pages in
1782                  * any of the below states.
1783                  */
1784                 if (m->wire_count ||
1785                     (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
1786                     m->valid != VM_PAGE_BITS_ALL
1787                 ) {
1788                         vm_page_wakeup(m);
1789                         continue;
1790                 }
1791
1792                 /*
1793                  * Theoretically once a page is known not to be busy, an
1794                  * interrupt cannot come along and rip it out from under us.
1795                  */
1796
1797                 if (advise == MADV_WILLNEED) {
1798                         vm_page_activate(m);
1799                 } else if (advise == MADV_DONTNEED) {
1800                         vm_page_dontneed(m);
1801                 } else if (advise == MADV_FREE) {
1802                         /*
1803                          * Mark the page clean.  This will allow the page
1804                          * to be freed up by the system.  However, such pages
1805                          * are often reused quickly by malloc()/free()
1806                          * so we do not do anything that would cause
1807                          * a page fault if we can help it.
1808                          *
1809                          * Specifically, we do not try to actually free
1810                          * the page now nor do we try to put it in the
1811                          * cache (which would cause a page fault on reuse).
1812                          *
1813                          * But we do make the page is freeable as we
1814                          * can without actually taking the step of unmapping
1815                          * it.
1816                          */
1817                         pmap_clear_modify(m);
1818                         m->dirty = 0;
1819                         m->act_count = 0;
1820                         vm_page_dontneed(m);
1821                         if (tobject->type == OBJT_SWAP)
1822                                 swap_pager_freespace(tobject, tpindex, 1);
1823                 }
1824                 vm_page_wakeup(m);
1825         }       
1826         if (tobject != object)
1827                 vm_object_drop(tobject);
1828         vm_object_drop(object);
1829 }
1830
1831 /*
1832  * Create a new object which is backed by the specified existing object
1833  * range.  Replace the pointer and offset that was pointing at the existing
1834  * object with the pointer/offset for the new object.
1835  *
1836  * If addref is non-zero the returned object is given an additional reference.
1837  * This mechanic exists to avoid the situation where refs might be 1 and
1838  * race against a collapse when the caller intends to bump it.  So the
1839  * caller cannot add the ref after the fact.  Used when the caller is
1840  * duplicating a vm_map_entry.
1841  *
1842  * No other requirements.
1843  */
1844 void
1845 vm_object_shadow(vm_object_t *objectp, vm_ooffset_t *offset, vm_size_t length,
1846                  int addref)
1847 {
1848         vm_object_t source;
1849         vm_object_t result;
1850         int useshadowlist;
1851
1852         source = *objectp;
1853
1854         /*
1855          * Don't create the new object if the old object isn't shared.
1856          * We have to chain wait before adding the reference to avoid
1857          * racing a collapse or deallocation.
1858          *
1859          * Clear OBJ_ONEMAPPING flag when shadowing.
1860          *
1861          * The caller owns a ref on source via *objectp which we are going
1862          * to replace.  This ref is inherited by the backing_object assignment.
1863          * from nobject and does not need to be incremented here.
1864          *
1865          * However, we add a temporary extra reference to the original source
1866          * prior to holding nobject in case we block, to avoid races where
1867          * someone else might believe that the source can be collapsed.
1868          */
1869         useshadowlist = 0;
1870         if (source) {
1871                 if (source->type != OBJT_VNODE) {
1872                         useshadowlist = 1;
1873                         vm_object_hold(source);
1874                         vm_object_chain_wait(source, 0);
1875                         if (source->ref_count == 1 &&
1876                             source->handle == NULL &&
1877                             (source->type == OBJT_DEFAULT ||
1878                              source->type == OBJT_SWAP)) {
1879                                 if (addref) {
1880                                         vm_object_reference_locked(source);
1881                                         vm_object_clear_flag(source,
1882                                                              OBJ_ONEMAPPING);
1883                                 }
1884                                 vm_object_drop(source);
1885                                 return;
1886                         }
1887                         vm_object_reference_locked(source);
1888                         vm_object_clear_flag(source, OBJ_ONEMAPPING);
1889                 } else {
1890                         vm_object_reference_quick(source);
1891                         vm_object_clear_flag(source, OBJ_ONEMAPPING);
1892                 }
1893         }
1894
1895         /*
1896          * Allocate a new object with the given length.  The new object
1897          * is returned referenced but we may have to add another one.
1898          * If we are adding a second reference we must clear OBJ_ONEMAPPING.
1899          * (typically because the caller is about to clone a vm_map_entry).
1900          *
1901          * The source object currently has an extra reference to prevent
1902          * collapses into it while we mess with its shadow list, which
1903          * we will remove later in this routine.
1904          *
1905          * The target object may require a second reference if asked for one
1906          * by the caller.
1907          */
1908         result = vm_object_allocate(OBJT_DEFAULT, length);
1909         if (result == NULL)
1910                 panic("vm_object_shadow: no object for shadowing");
1911         vm_object_hold(result);
1912         if (addref) {
1913                 vm_object_reference_locked(result);
1914                 vm_object_clear_flag(result, OBJ_ONEMAPPING);
1915         }
1916
1917         /*
1918          * The new object shadows the source object.  Chain wait before
1919          * adjusting shadow_count or the shadow list to avoid races.
1920          *
1921          * Try to optimize the result object's page color when shadowing
1922          * in order to maintain page coloring consistency in the combined 
1923          * shadowed object.
1924          *
1925          * The backing_object reference to source requires adding a ref to
1926          * source.  We simply inherit the ref from the original *objectp
1927          * (which we are replacing) so no additional refs need to be added.
1928          * (we must still clean up the extra ref we had to prevent collapse
1929          * races).
1930          *
1931          * SHADOWING IS NOT APPLICABLE TO OBJT_VNODE OBJECTS
1932          */
1933         KKASSERT(result->backing_object == NULL);
1934         result->backing_object = source;
1935         if (source) {
1936                 if (useshadowlist) {
1937                         vm_object_chain_wait(source, 0);
1938                         LIST_INSERT_HEAD(&source->shadow_head,
1939                                          result, shadow_list);
1940                         source->shadow_count++;
1941                         source->generation++;
1942                         vm_object_set_flag(result, OBJ_ONSHADOW);
1943                 }
1944                 /* cpu localization twist */
1945                 result->pg_color = vm_quickcolor();
1946         }
1947
1948         /*
1949          * Adjust the return storage.  Drop the ref on source before
1950          * returning.
1951          */
1952         result->backing_object_offset = *offset;
1953         vm_object_drop(result);
1954         *offset = 0;
1955         if (source) {
1956                 if (useshadowlist) {
1957                         vm_object_deallocate_locked(source);
1958                         vm_object_drop(source);
1959                 } else {
1960                         vm_object_deallocate(source);
1961                 }
1962         }
1963
1964         /*
1965          * Return the new things
1966          */
1967         *objectp = result;
1968 }
1969
1970 #define OBSC_TEST_ALL_SHADOWED  0x0001
1971 #define OBSC_COLLAPSE_NOWAIT    0x0002
1972 #define OBSC_COLLAPSE_WAIT      0x0004
1973
1974 static int vm_object_backing_scan_callback(vm_page_t p, void *data);
1975
1976 /*
1977  * The caller must hold the object.
1978  */
1979 static __inline int
1980 vm_object_backing_scan(vm_object_t object, vm_object_t backing_object, int op)
1981 {
1982         struct rb_vm_page_scan_info info;
1983         struct vm_object_hash *hash;
1984
1985         vm_object_assert_held(object);
1986         vm_object_assert_held(backing_object);
1987
1988         KKASSERT(backing_object == object->backing_object);
1989         info.backing_offset_index = OFF_TO_IDX(object->backing_object_offset);
1990
1991         /*
1992          * Initial conditions
1993          */
1994         if (op & OBSC_TEST_ALL_SHADOWED) {
1995                 /*
1996                  * We do not want to have to test for the existence of
1997                  * swap pages in the backing object.  XXX but with the
1998                  * new swapper this would be pretty easy to do.
1999                  *
2000                  * XXX what about anonymous MAP_SHARED memory that hasn't
2001                  * been ZFOD faulted yet?  If we do not test for this, the
2002                  * shadow test may succeed! XXX
2003                  */
2004                 if (backing_object->type != OBJT_DEFAULT)
2005                         return(0);
2006         }
2007         if (op & OBSC_COLLAPSE_WAIT) {
2008                 KKASSERT((backing_object->flags & OBJ_DEAD) == 0);
2009                 vm_object_set_flag(backing_object, OBJ_DEAD);
2010
2011                 hash = VMOBJ_HASH(backing_object);
2012                 lwkt_gettoken(&hash->token);
2013                 TAILQ_REMOVE(&hash->list, backing_object, object_list);
2014                 lwkt_reltoken(&hash->token);
2015         }
2016
2017         /*
2018          * Our scan.   We have to retry if a negative error code is returned,
2019          * otherwise 0 or 1 will be returned in info.error.  0 Indicates that
2020          * the scan had to be stopped because the parent does not completely
2021          * shadow the child.
2022          */
2023         info.object = object;
2024         info.backing_object = backing_object;
2025         info.limit = op;
2026         info.count = 0;
2027         do {
2028                 info.error = 1;
2029                 vm_page_rb_tree_RB_SCAN(&backing_object->rb_memq, NULL,
2030                                         vm_object_backing_scan_callback,
2031                                         &info);
2032         } while (info.error < 0);
2033
2034         return(info.error);
2035 }
2036
2037 /*
2038  * The caller must hold the object.
2039  */
2040 static int
2041 vm_object_backing_scan_callback(vm_page_t p, void *data)
2042 {
2043         struct rb_vm_page_scan_info *info = data;
2044         vm_object_t backing_object;
2045         vm_object_t object;
2046         vm_pindex_t pindex;
2047         vm_pindex_t new_pindex;
2048         vm_pindex_t backing_offset_index;
2049         int op;
2050
2051         pindex = p->pindex;
2052         new_pindex = pindex - info->backing_offset_index;
2053         op = info->limit;
2054         object = info->object;
2055         backing_object = info->backing_object;
2056         backing_offset_index = info->backing_offset_index;
2057
2058         if (op & OBSC_TEST_ALL_SHADOWED) {
2059                 vm_page_t pp;
2060
2061                 /*
2062                  * Ignore pages outside the parent object's range
2063                  * and outside the parent object's mapping of the 
2064                  * backing object.
2065                  *
2066                  * note that we do not busy the backing object's
2067                  * page.
2068                  */
2069                 if (pindex < backing_offset_index ||
2070                     new_pindex >= object->size
2071                 ) {
2072                         return(0);
2073                 }
2074
2075                 /*
2076                  * See if the parent has the page or if the parent's
2077                  * object pager has the page.  If the parent has the
2078                  * page but the page is not valid, the parent's
2079                  * object pager must have the page.
2080                  *
2081                  * If this fails, the parent does not completely shadow
2082                  * the object and we might as well give up now.
2083                  */
2084                 pp = vm_page_lookup(object, new_pindex);
2085                 if ((pp == NULL || pp->valid == 0) &&
2086                     !vm_pager_has_page(object, new_pindex)
2087                 ) {
2088                         info->error = 0;        /* problemo */
2089                         return(-1);             /* stop the scan */
2090                 }
2091         }
2092
2093         /*
2094          * Check for busy page.  Note that we may have lost (p) when we
2095          * possibly blocked above.
2096          */
2097         if (op & (OBSC_COLLAPSE_WAIT | OBSC_COLLAPSE_NOWAIT)) {
2098                 vm_page_t pp;
2099
2100                 if (vm_page_busy_try(p, TRUE)) {
2101                         if (op & OBSC_COLLAPSE_NOWAIT) {
2102                                 return(0);
2103                         } else {
2104                                 /*
2105                                  * If we slept, anything could have
2106                                  * happened.   Ask that the scan be restarted.
2107                                  *
2108                                  * Since the object is marked dead, the
2109                                  * backing offset should not have changed.  
2110                                  */
2111                                 vm_page_sleep_busy(p, TRUE, "vmocol");
2112                                 info->error = -1;
2113                                 return(-1);
2114                         }
2115                 }
2116
2117                 /*
2118                  * If (p) is no longer valid restart the scan.
2119                  */
2120                 if (p->object != backing_object || p->pindex != pindex) {
2121                         kprintf("vm_object_backing_scan: Warning: page "
2122                                 "%p ripped out from under us\n", p);
2123                         vm_page_wakeup(p);
2124                         info->error = -1;
2125                         return(-1);
2126                 }
2127
2128                 if (op & OBSC_COLLAPSE_NOWAIT) {
2129                         if (p->valid == 0 ||
2130                             p->wire_count ||
2131                             (p->flags & PG_NEED_COMMIT)) {
2132                                 vm_page_wakeup(p);
2133                                 return(0);
2134                         }
2135                 } else {
2136                         /* XXX what if p->valid == 0 , hold_count, etc? */
2137                 }
2138
2139                 KASSERT(
2140                     p->object == backing_object,
2141                     ("vm_object_qcollapse(): object mismatch")
2142                 );
2143
2144                 /*
2145                  * Destroy any associated swap
2146                  */
2147                 if (backing_object->type == OBJT_SWAP)
2148                         swap_pager_freespace(backing_object, p->pindex, 1);
2149
2150                 if (
2151                     p->pindex < backing_offset_index ||
2152                     new_pindex >= object->size
2153                 ) {
2154                         /*
2155                          * Page is out of the parent object's range, we 
2156                          * can simply destroy it. 
2157                          */
2158                         vm_page_protect(p, VM_PROT_NONE);
2159                         vm_page_free(p);
2160                         return(0);
2161                 }
2162
2163                 pp = vm_page_lookup(object, new_pindex);
2164                 if (pp != NULL || vm_pager_has_page(object, new_pindex)) {
2165                         /*
2166                          * page already exists in parent OR swap exists
2167                          * for this location in the parent.  Destroy 
2168                          * the original page from the backing object.
2169                          *
2170                          * Leave the parent's page alone
2171                          */
2172                         vm_page_protect(p, VM_PROT_NONE);
2173                         vm_page_free(p);
2174                         return(0);
2175                 }
2176
2177                 /*
2178                  * Page does not exist in parent, rename the
2179                  * page from the backing object to the main object. 
2180                  *
2181                  * If the page was mapped to a process, it can remain 
2182                  * mapped through the rename.
2183                  */
2184                 if ((p->queue - p->pc) == PQ_CACHE)
2185                         vm_page_deactivate(p);
2186
2187                 vm_page_rename(p, object, new_pindex);
2188                 vm_page_wakeup(p);
2189                 /* page automatically made dirty by rename */
2190         }
2191         return(0);
2192 }
2193
2194 /*
2195  * This version of collapse allows the operation to occur earlier and
2196  * when paging_in_progress is true for an object...  This is not a complete
2197  * operation, but should plug 99.9% of the rest of the leaks.
2198  *
2199  * The caller must hold the object and backing_object and both must be
2200  * chainlocked.
2201  *
2202  * (only called from vm_object_collapse)
2203  */
2204 static void
2205 vm_object_qcollapse(vm_object_t object, vm_object_t backing_object)
2206 {
2207         if (backing_object->ref_count == 1) {
2208                 atomic_add_int(&backing_object->ref_count, 2);
2209 #if defined(DEBUG_LOCKS)
2210                 debugvm_object_add(backing_object, "qcollapse", 1, 2);
2211 #endif
2212                 vm_object_backing_scan(object, backing_object,
2213                                        OBSC_COLLAPSE_NOWAIT);
2214                 atomic_add_int(&backing_object->ref_count, -2);
2215 #if defined(DEBUG_LOCKS)
2216                 debugvm_object_add(backing_object, "qcollapse", 2, -2);
2217 #endif
2218         }
2219 }
2220
2221 /*
2222  * Collapse an object with the object backing it.  Pages in the backing
2223  * object are moved into the parent, and the backing object is deallocated.
2224  * Any conflict is resolved in favor of the parent's existing pages.
2225  *
2226  * object must be held and chain-locked on call.
2227  *
2228  * The caller must have an extra ref on object to prevent a race from
2229  * destroying it during the collapse.
2230  */
2231 void
2232 vm_object_collapse(vm_object_t object, struct vm_object_dealloc_list **dlistp)
2233 {
2234         struct vm_object_dealloc_list *dlist = NULL;
2235         vm_object_t backing_object;
2236
2237         /*
2238          * Only one thread is attempting a collapse at any given moment.
2239          * There are few restrictions for (object) that callers of this
2240          * function check so reentrancy is likely.
2241          */
2242         KKASSERT(object != NULL);
2243         vm_object_assert_held(object);
2244         KKASSERT(object->chainlk & (CHAINLK_MASK | CHAINLK_EXCL));
2245
2246         for (;;) {
2247                 vm_object_t bbobj;
2248                 int dodealloc;
2249
2250                 /*
2251                  * We can only collapse a DEFAULT/SWAP object with a
2252                  * DEFAULT/SWAP object.
2253                  */
2254                 if (object->type != OBJT_DEFAULT && object->type != OBJT_SWAP) {
2255                         backing_object = NULL;
2256                         break;
2257                 }
2258
2259                 backing_object = object->backing_object;
2260                 if (backing_object == NULL)
2261                         break;
2262                 if (backing_object->type != OBJT_DEFAULT &&
2263                     backing_object->type != OBJT_SWAP) {
2264                         backing_object = NULL;
2265                         break;
2266                 }
2267
2268                 /*
2269                  * Hold the backing_object and check for races
2270                  */
2271                 vm_object_hold(backing_object);
2272                 if (backing_object != object->backing_object ||
2273                     (backing_object->type != OBJT_DEFAULT &&
2274                      backing_object->type != OBJT_SWAP)) {
2275                         vm_object_drop(backing_object);
2276                         continue;
2277                 }
2278
2279                 /*
2280                  * Chain-lock the backing object too because if we
2281                  * successfully merge its pages into the top object we
2282                  * will collapse backing_object->backing_object as the
2283                  * new backing_object.  Re-check that it is still our
2284                  * backing object.
2285                  */
2286                 vm_object_chain_acquire(backing_object, 0);
2287                 if (backing_object != object->backing_object) {
2288                         vm_object_chain_release(backing_object);
2289                         vm_object_drop(backing_object);
2290                         continue;
2291                 }
2292
2293                 /*
2294                  * we check the backing object first, because it is most likely
2295                  * not collapsable.
2296                  */
2297                 if (backing_object->handle != NULL ||
2298                     (backing_object->type != OBJT_DEFAULT &&
2299                      backing_object->type != OBJT_SWAP) ||
2300                     (backing_object->flags & OBJ_DEAD) ||
2301                     object->handle != NULL ||
2302                     (object->type != OBJT_DEFAULT &&
2303                      object->type != OBJT_SWAP) ||
2304                     (object->flags & OBJ_DEAD)) {
2305                         break;
2306                 }
2307
2308                 /*
2309                  * If paging is in progress we can't do a normal collapse.
2310                  */
2311                 if (
2312                     object->paging_in_progress != 0 ||
2313                     backing_object->paging_in_progress != 0
2314                 ) {
2315                         vm_object_qcollapse(object, backing_object);
2316                         break;
2317                 }
2318
2319                 /*
2320                  * We know that we can either collapse the backing object (if
2321                  * the parent is the only reference to it) or (perhaps) have
2322                  * the parent bypass the object if the parent happens to shadow
2323                  * all the resident pages in the entire backing object.
2324                  *
2325                  * This is ignoring pager-backed pages such as swap pages.
2326                  * vm_object_backing_scan fails the shadowing test in this
2327                  * case.
2328                  */
2329                 if (backing_object->ref_count == 1) {
2330                         /*
2331                          * If there is exactly one reference to the backing
2332                          * object, we can collapse it into the parent.  
2333                          */
2334                         KKASSERT(object->backing_object == backing_object);
2335                         vm_object_backing_scan(object, backing_object,
2336                                                OBSC_COLLAPSE_WAIT);
2337
2338                         /*
2339                          * Move the pager from backing_object to object.
2340                          */
2341                         if (backing_object->type == OBJT_SWAP) {
2342                                 vm_object_pip_add(backing_object, 1);
2343
2344                                 /*
2345                                  * scrap the paging_offset junk and do a 
2346                                  * discrete copy.  This also removes major 
2347                                  * assumptions about how the swap-pager 
2348                                  * works from where it doesn't belong.  The
2349                                  * new swapper is able to optimize the
2350                                  * destroy-source case.
2351                                  */
2352                                 vm_object_pip_add(object, 1);
2353                                 swap_pager_copy(backing_object, object,
2354                                     OFF_TO_IDX(object->backing_object_offset),
2355                                     TRUE);
2356                                 vm_object_pip_wakeup(object);
2357                                 vm_object_pip_wakeup(backing_object);
2358                         }
2359
2360                         /*
2361                          * Object now shadows whatever backing_object did.
2362                          * Remove object from backing_object's shadow_list.
2363                          *
2364                          * Removing object from backing_objects shadow list
2365                          * requires releasing object, which we will do below.
2366                          */
2367                         KKASSERT(object->backing_object == backing_object);
2368                         if (object->flags & OBJ_ONSHADOW) {
2369                                 LIST_REMOVE(object, shadow_list);
2370                                 backing_object->shadow_count--;
2371                                 backing_object->generation++;
2372                                 vm_object_clear_flag(object, OBJ_ONSHADOW);
2373                         }
2374
2375                         /*
2376                          * backing_object->backing_object moves from within
2377                          * backing_object to within object.
2378                          *
2379                          * OBJT_VNODE bbobj's should have empty shadow lists.
2380                          */
2381                         while ((bbobj = backing_object->backing_object) != NULL) {
2382                                 if (bbobj->type == OBJT_VNODE)
2383                                         vm_object_hold_shared(bbobj);
2384                                 else
2385                                         vm_object_hold(bbobj);
2386                                 if (bbobj == backing_object->backing_object)
2387                                         break;
2388                                 vm_object_drop(bbobj);
2389                         }
2390
2391                         /*
2392                          * We are removing backing_object from bbobj's
2393                          * shadow list and adding object to bbobj's shadow
2394                          * list, so the ref_count on bbobj is unchanged.
2395                          */
2396                         if (bbobj) {
2397                                 if (backing_object->flags & OBJ_ONSHADOW) {
2398                                         /* not locked exclusively if vnode */
2399                                         KKASSERT(bbobj->type != OBJT_VNODE);
2400                                         LIST_REMOVE(backing_object,
2401                                                     shadow_list);
2402                                         bbobj->shadow_count--;
2403                                         bbobj->generation++;
2404                                         vm_object_clear_flag(backing_object,
2405                                                              OBJ_ONSHADOW);
2406                                 }
2407                                 backing_object->backing_object = NULL;
2408                         }
2409                         object->backing_object = bbobj;
2410                         if (bbobj) {
2411                                 if (bbobj->type != OBJT_VNODE) {
2412                                         LIST_INSERT_HEAD(&bbobj->shadow_head,
2413                                                          object, shadow_list);
2414                                         bbobj->shadow_count++;
2415                                         bbobj->generation++;
2416                                         vm_object_set_flag(object,
2417                                                            OBJ_ONSHADOW);
2418                                 }
2419                         }
2420
2421                         object->backing_object_offset +=
2422                                 backing_object->backing_object_offset;
2423
2424                         vm_object_drop(bbobj);
2425
2426                         /*
2427                          * Discard the old backing_object.  Nothing should be
2428                          * able to ref it, other than a vm_map_split(),
2429                          * and vm_map_split() will stall on our chain lock.
2430                          * And we control the parent so it shouldn't be
2431                          * possible for it to go away either.
2432                          *
2433                          * Since the backing object has no pages, no pager
2434                          * left, and no object references within it, all
2435                          * that is necessary is to dispose of it.
2436                          */
2437                         KASSERT(backing_object->ref_count == 1,
2438                                 ("backing_object %p was somehow "
2439                                  "re-referenced during collapse!",
2440                                  backing_object));
2441                         KASSERT(RB_EMPTY(&backing_object->rb_memq),
2442                                 ("backing_object %p somehow has left "
2443                                  "over pages during collapse!",
2444                                  backing_object));
2445
2446                         /*
2447                          * The object can be destroyed.
2448                          *
2449                          * XXX just fall through and dodealloc instead
2450                          *     of forcing destruction?
2451                          */
2452                         atomic_add_int(&backing_object->ref_count, -1);
2453 #if defined(DEBUG_LOCKS)
2454                         debugvm_object_add(backing_object, "collapse", 1, -1);
2455 #endif
2456                         if ((backing_object->flags & OBJ_DEAD) == 0)
2457                                 vm_object_terminate(backing_object);
2458                         object_collapses++;
2459                         dodealloc = 0;
2460                 } else {
2461                         /*
2462                          * If we do not entirely shadow the backing object,
2463                          * there is nothing we can do so we give up.
2464                          */
2465                         if (vm_object_backing_scan(object, backing_object,
2466                                                 OBSC_TEST_ALL_SHADOWED) == 0) {
2467                                 break;
2468                         }
2469
2470                         /*
2471                          * bbobj is backing_object->backing_object.  Since
2472                          * object completely shadows backing_object we can
2473                          * bypass it and become backed by bbobj instead.
2474                          *
2475                          * The shadow list for vnode backing objects is not
2476                          * used and a shared hold is allowed.
2477                          */
2478                         while ((bbobj = backing_object->backing_object) != NULL) {
2479                                 if (bbobj->type == OBJT_VNODE)
2480                                         vm_object_hold_shared(bbobj);
2481                                 else
2482                                         vm_object_hold(bbobj);
2483                                 if (bbobj == backing_object->backing_object)
2484                                         break;
2485                                 vm_object_drop(bbobj);
2486                         }
2487
2488                         /*
2489                          * Make object shadow bbobj instead of backing_object.
2490                          * Remove object from backing_object's shadow list.
2491                          *
2492                          * Deallocating backing_object will not remove
2493                          * it, since its reference count is at least 2.
2494                          *
2495                          * Removing object from backing_object's shadow
2496                          * list requires releasing a ref, which we do
2497                          * below by setting dodealloc to 1.
2498                          */
2499                         KKASSERT(object->backing_object == backing_object);
2500                         if (object->flags & OBJ_ONSHADOW) {
2501                                 LIST_REMOVE(object, shadow_list);
2502                                 backing_object->shadow_count--;
2503                                 backing_object->generation++;
2504                                 vm_object_clear_flag(object, OBJ_ONSHADOW);
2505                         }
2506
2507                         /*
2508                          * Add a ref to bbobj, bbobj now shadows object.
2509                          *
2510                          * NOTE: backing_object->backing_object still points
2511                          *       to bbobj.  That relationship remains intact
2512                          *       because backing_object has > 1 ref, so
2513                          *       someone else is pointing to it (hence why
2514                          *       we can't collapse it into object and can
2515                          *       only handle the all-shadowed bypass case).
2516                          */
2517                         if (bbobj) {
2518                                 if (bbobj->type != OBJT_VNODE) {
2519                                         vm_object_chain_wait(bbobj, 0);
2520                                         vm_object_reference_locked(bbobj);
2521                                         LIST_INSERT_HEAD(&bbobj->shadow_head,
2522                                                          object, shadow_list);
2523                                         bbobj->shadow_count++;
2524                                         bbobj->generation++;
2525                                         vm_object_set_flag(object,
2526                                                            OBJ_ONSHADOW);
2527                                 } else {
2528                                         vm_object_reference_quick(bbobj);
2529                                 }
2530                                 object->backing_object_offset +=
2531                                         backing_object->backing_object_offset;
2532                                 object->backing_object = bbobj;
2533                                 vm_object_drop(bbobj);
2534                         } else {
2535                                 object->backing_object = NULL;
2536                         }
2537
2538                         /*
2539                          * Drop the reference count on backing_object.  To
2540                          * handle ref_count races properly we can't assume
2541                          * that the ref_count is still at least 2 so we
2542                          * have to actually call vm_object_deallocate()
2543                          * (after clearing the chainlock).
2544                          */
2545                         object_bypasses++;
2546                         dodealloc = 1;
2547                 }
2548
2549                 /*
2550                  * Ok, we want to loop on the new object->bbobj association,
2551                  * possibly collapsing it further.  However if dodealloc is
2552                  * non-zero we have to deallocate the backing_object which
2553                  * itself can potentially undergo a collapse, creating a
2554                  * recursion depth issue with the LWKT token subsystem.
2555                  *
2556                  * In the case where we must deallocate the backing_object
2557                  * it is possible now that the backing_object has a single
2558                  * shadow count on some other object (not represented here
2559                  * as yet), since it no longer shadows us.  Thus when we
2560                  * call vm_object_deallocate() it may attempt to collapse
2561                  * itself into its remaining parent.
2562                  */
2563                 if (dodealloc) {
2564                         struct vm_object_dealloc_list *dtmp;
2565
2566                         vm_object_chain_release(backing_object);
2567                         vm_object_unlock(backing_object);
2568                         /* backing_object remains held */
2569
2570                         /*
2571                          * Auto-deallocation list for caller convenience.
2572                          */
2573                         if (dlistp == NULL)
2574                                 dlistp = &dlist;
2575
2576                         dtmp = kmalloc(sizeof(*dtmp), M_TEMP, M_WAITOK);
2577                         dtmp->object = backing_object;
2578                         dtmp->next = *dlistp;
2579                         *dlistp = dtmp;
2580                 } else {
2581                         vm_object_chain_release(backing_object);
2582                         vm_object_drop(backing_object);
2583                 }
2584                 /* backing_object = NULL; not needed */
2585                 /* loop */
2586         }
2587
2588         /*
2589          * Clean up any left over backing_object
2590          */
2591         if (backing_object) {
2592                 vm_object_chain_release(backing_object);
2593                 vm_object_drop(backing_object);
2594         }
2595
2596         /*
2597          * Clean up any auto-deallocation list.  This is a convenience
2598          * for top-level callers so they don't have to pass &dlist.
2599          * Do not clean up any caller-passed dlistp, the caller will
2600          * do that.
2601          */
2602         if (dlist)
2603                 vm_object_deallocate_list(&dlist);
2604
2605 }
2606
2607 /*
2608  * vm_object_collapse() may collect additional objects in need of
2609  * deallocation.  This routine deallocates these objects.  The
2610  * deallocation itself can trigger additional collapses (which the
2611  * deallocate function takes care of).  This procedure is used to
2612  * reduce procedural recursion since these vm_object shadow chains
2613  * can become quite long.
2614  */
2615 void
2616 vm_object_deallocate_list(struct vm_object_dealloc_list **dlistp)
2617 {
2618         struct vm_object_dealloc_list *dlist;
2619
2620         while ((dlist = *dlistp) != NULL) {
2621                 *dlistp = dlist->next;
2622                 vm_object_lock(dlist->object);
2623                 vm_object_deallocate_locked(dlist->object);
2624                 vm_object_drop(dlist->object);
2625                 kfree(dlist, M_TEMP);
2626         }
2627 }
2628
2629 /*
2630  * Removes all physical pages in the specified object range from the
2631  * object's list of pages.
2632  *
2633  * No requirements.
2634  */
2635 static int vm_object_page_remove_callback(vm_page_t p, void *data);
2636
2637 void
2638 vm_object_page_remove(vm_object_t object, vm_pindex_t start, vm_pindex_t end,
2639                       boolean_t clean_only)
2640 {
2641         struct rb_vm_page_scan_info info;
2642         int all;
2643
2644         /*
2645          * Degenerate cases and assertions
2646          */
2647         vm_object_hold(object);
2648         if (object == NULL ||
2649             (object->resident_page_count == 0 && object->swblock_count == 0)) {
2650                 vm_object_drop(object);
2651                 return;
2652         }
2653         KASSERT(object->type != OBJT_PHYS, 
2654                 ("attempt to remove pages from a physical object"));
2655
2656         /*
2657          * Indicate that paging is occuring on the object
2658          */
2659         vm_object_pip_add(object, 1);
2660
2661         /*
2662          * Figure out the actual removal range and whether we are removing
2663          * the entire contents of the object or not.  If removing the entire
2664          * contents, be sure to get all pages, even those that might be 
2665          * beyond the end of the object.
2666          */
2667         info.object = object;
2668         info.start_pindex = start;
2669         if (end == 0)
2670                 info.end_pindex = (vm_pindex_t)-1;
2671         else
2672                 info.end_pindex = end - 1;
2673         info.limit = clean_only;
2674         info.count = 0;
2675         all = (start == 0 && info.end_pindex >= object->size - 1);
2676
2677         /*
2678          * Loop until we are sure we have gotten them all.
2679          */
2680         do {
2681                 info.error = 0;
2682                 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2683                                         vm_object_page_remove_callback, &info);
2684         } while (info.error);
2685
2686         /*
2687          * Remove any related swap if throwing away pages, or for
2688          * non-swap objects (the swap is a clean copy in that case).
2689          */
2690         if (object->type != OBJT_SWAP || clean_only == FALSE) {
2691                 if (all)
2692                         swap_pager_freespace_all(object);
2693                 else
2694                         swap_pager_freespace(object, info.start_pindex,
2695                              info.end_pindex - info.start_pindex + 1);
2696         }
2697
2698         /*
2699          * Cleanup
2700          */
2701         vm_object_pip_wakeup(object);
2702         vm_object_drop(object);
2703 }
2704
2705 /*
2706  * The caller must hold the object.
2707  *
2708  * NOTE: User yields are allowed when removing more than one page, but not
2709  *       allowed if only removing one page (the path for single page removals
2710  *       might hold a spinlock).
2711  */
2712 static int
2713 vm_object_page_remove_callback(vm_page_t p, void *data)
2714 {
2715         struct rb_vm_page_scan_info *info = data;
2716
2717         if ((++info->count & 63) == 0)
2718                 lwkt_user_yield();
2719
2720         if (info->object != p->object ||
2721             p->pindex < info->start_pindex ||
2722             p->pindex > info->end_pindex) {
2723                 kprintf("vm_object_page_remove_callbackA: obj/pg race %p/%p\n",
2724                         info->object, p);
2725                 return(0);
2726         }
2727         if (vm_page_busy_try(p, TRUE)) {
2728                 vm_page_sleep_busy(p, TRUE, "vmopar");
2729                 info->error = 1;
2730                 return(0);
2731         }
2732         if (info->object != p->object) {
2733                 /* this should never happen */
2734                 kprintf("vm_object_page_remove_callbackB: obj/pg race %p/%p\n",
2735                         info->object, p);
2736                 vm_page_wakeup(p);
2737                 return(0);
2738         }
2739
2740         /*
2741          * Wired pages cannot be destroyed, but they can be invalidated
2742          * and we do so if clean_only (limit) is not set.
2743          *
2744          * WARNING!  The page may be wired due to being part of a buffer
2745          *           cache buffer, and the buffer might be marked B_CACHE.
2746          *           This is fine as part of a truncation but VFSs must be
2747          *           sure to fix the buffer up when re-extending the file.
2748          *
2749          * NOTE!     PG_NEED_COMMIT is ignored.
2750          */
2751         if (p->wire_count != 0) {
2752                 vm_page_protect(p, VM_PROT_NONE);
2753                 if (info->limit == 0)
2754                         p->valid = 0;
2755                 vm_page_wakeup(p);
2756                 return(0);
2757         }
2758
2759         /*
2760          * limit is our clean_only flag.  If set and the page is dirty or
2761          * requires a commit, do not free it.  If set and the page is being
2762          * held by someone, do not free it.
2763          */
2764         if (info->limit && p->valid) {
2765                 vm_page_test_dirty(p);
2766                 if ((p->valid & p->dirty) || (p->flags & PG_NEED_COMMIT)) {
2767                         vm_page_wakeup(p);
2768                         return(0);
2769                 }
2770         }
2771
2772         /*
2773          * Destroy the page
2774          */
2775         vm_page_protect(p, VM_PROT_NONE);
2776         vm_page_free(p);
2777
2778         return(0);
2779 }
2780
2781 /*
2782  * Coalesces two objects backing up adjoining regions of memory into a
2783  * single object.
2784  *
2785  * returns TRUE if objects were combined.
2786  *
2787  * NOTE: Only works at the moment if the second object is NULL -
2788  *       if it's not, which object do we lock first?
2789  *
2790  * Parameters:
2791  *      prev_object     First object to coalesce
2792  *      prev_offset     Offset into prev_object
2793  *      next_object     Second object into coalesce
2794  *      next_offset     Offset into next_object
2795  *
2796  *      prev_size       Size of reference to prev_object
2797  *      next_size       Size of reference to next_object
2798  *
2799  * The caller does not need to hold (prev_object) but must have a stable
2800  * pointer to it (typically by holding the vm_map locked).
2801  */
2802 boolean_t
2803 vm_object_coalesce(vm_object_t prev_object, vm_pindex_t prev_pindex,
2804                    vm_size_t prev_size, vm_size_t next_size)
2805 {
2806         vm_pindex_t next_pindex;
2807
2808         if (prev_object == NULL)
2809                 return (TRUE);
2810
2811         vm_object_hold(prev_object);
2812
2813         if (prev_object->type != OBJT_DEFAULT &&
2814             prev_object->type != OBJT_SWAP) {
2815                 vm_object_drop(prev_object);
2816                 return (FALSE);
2817         }
2818
2819         /*
2820          * Try to collapse the object first
2821          */
2822         vm_object_chain_acquire(prev_object, 0);
2823         vm_object_collapse(prev_object, NULL);
2824
2825         /*
2826          * Can't coalesce if: . more than one reference . paged out . shadows
2827          * another object . has a copy elsewhere (any of which mean that the
2828          * pages not mapped to prev_entry may be in use anyway)
2829          */
2830
2831         if (prev_object->backing_object != NULL) {
2832                 vm_object_chain_release(prev_object);
2833                 vm_object_drop(prev_object);
2834                 return (FALSE);
2835         }
2836
2837         prev_size >>= PAGE_SHIFT;
2838         next_size >>= PAGE_SHIFT;
2839         next_pindex = prev_pindex + prev_size;
2840
2841         if ((prev_object->ref_count > 1) &&
2842             (prev_object->size != next_pindex)) {
2843                 vm_object_chain_release(prev_object);
2844                 vm_object_drop(prev_object);
2845                 return (FALSE);
2846         }
2847
2848         /*
2849          * Remove any pages that may still be in the object from a previous
2850          * deallocation.
2851          */
2852         if (next_pindex < prev_object->size) {
2853                 vm_object_page_remove(prev_object,
2854                                       next_pindex,
2855                                       next_pindex + next_size, FALSE);
2856                 if (prev_object->type == OBJT_SWAP)
2857                         swap_pager_freespace(prev_object,
2858                                              next_pindex, next_size);
2859         }
2860
2861         /*
2862          * Extend the object if necessary.
2863          */
2864         if (next_pindex + next_size > prev_object->size)
2865                 prev_object->size = next_pindex + next_size;
2866
2867         vm_object_chain_release(prev_object);
2868         vm_object_drop(prev_object);
2869         return (TRUE);
2870 }
2871
2872 /*
2873  * Make the object writable and flag is being possibly dirty.
2874  *
2875  * The object might not be held (or might be held but held shared),
2876  * the related vnode is probably not held either.  Object and vnode are
2877  * stable by virtue of the vm_page busied by the caller preventing
2878  * destruction.
2879  *
2880  * If the related mount is flagged MNTK_THR_SYNC we need to call
2881  * vsetobjdirty().  Filesystems using this option usually shortcut
2882  * synchronization by only scanning the syncer list.
2883  */
2884 void
2885 vm_object_set_writeable_dirty(vm_object_t object)
2886 {
2887         struct vnode *vp;
2888
2889         /*vm_object_assert_held(object);*/
2890         /*
2891          * Avoid contention in vm fault path by checking the state before
2892          * issuing an atomic op on it.
2893          */
2894         if ((object->flags & (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) !=
2895             (OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY)) {
2896                 vm_object_set_flag(object, OBJ_WRITEABLE|OBJ_MIGHTBEDIRTY);
2897         }
2898         if (object->type == OBJT_VNODE &&
2899             (vp = (struct vnode *)object->handle) != NULL) {
2900                 if ((vp->v_flag & VOBJDIRTY) == 0) {
2901                         if (vp->v_mount &&
2902                             (vp->v_mount->mnt_kern_flag & MNTK_THR_SYNC)) {
2903                                 /*
2904                                  * New style THR_SYNC places vnodes on the
2905                                  * syncer list more deterministically.
2906                                  */
2907                                 vsetobjdirty(vp);
2908                         } else {
2909                                 /*
2910                                  * Old style scan would not necessarily place
2911                                  * a vnode on the syncer list when possibly
2912                                  * modified via mmap.
2913                                  */
2914                                 vsetflags(vp, VOBJDIRTY);
2915                         }
2916                 }
2917         }
2918 }
2919
2920 #include "opt_ddb.h"
2921 #ifdef DDB
2922 #include <sys/kernel.h>
2923
2924 #include <sys/cons.h>
2925
2926 #include <ddb/ddb.h>
2927
2928 static int      _vm_object_in_map (vm_map_t map, vm_object_t object,
2929                                        vm_map_entry_t entry);
2930 static int      vm_object_in_map (vm_object_t object);
2931
2932 /*
2933  * The caller must hold the object.
2934  */
2935 static int
2936 _vm_object_in_map(vm_map_t map, vm_object_t object, vm_map_entry_t entry)
2937 {
2938         vm_map_t tmpm;
2939         vm_map_entry_t tmpe;
2940         vm_object_t obj, nobj;
2941         int entcount;
2942
2943         if (map == 0)
2944                 return 0;
2945         if (entry == 0) {
2946                 tmpe = map->header.next;
2947                 entcount = map->nentries;
2948                 while (entcount-- && (tmpe != &map->header)) {
2949                         if( _vm_object_in_map(map, object, tmpe)) {
2950                                 return 1;
2951                         }
2952                         tmpe = tmpe->next;
2953                 }
2954                 return (0);
2955         }
2956         switch(entry->maptype) {
2957         case VM_MAPTYPE_SUBMAP:
2958                 tmpm = entry->object.sub_map;
2959                 tmpe = tmpm->header.next;
2960                 entcount = tmpm->nentries;
2961                 while (entcount-- && tmpe != &tmpm->header) {
2962                         if( _vm_object_in_map(tmpm, object, tmpe)) {
2963                                 return 1;
2964                         }
2965                         tmpe = tmpe->next;
2966                 }
2967                 break;
2968         case VM_MAPTYPE_NORMAL:
2969         case VM_MAPTYPE_VPAGETABLE:
2970                 obj = entry->object.vm_object;
2971                 while (obj) {
2972                         if (obj == object) {
2973                                 if (obj != entry->object.vm_object)
2974                                         vm_object_drop(obj);
2975                                 return 1;
2976                         }
2977                         while ((nobj = obj->backing_object) != NULL) {
2978                                 vm_object_hold(nobj);
2979                                 if (nobj == obj->backing_object)
2980                                         break;
2981                                 vm_object_drop(nobj);
2982                         }
2983                         if (obj != entry->object.vm_object) {
2984                                 if (nobj)
2985                                         vm_object_lock_swap();
2986                                 vm_object_drop(obj);
2987                         }
2988                         obj = nobj;
2989                 }
2990                 break;
2991         default:
2992                 break;
2993         }
2994         return 0;
2995 }
2996
2997 static int vm_object_in_map_callback(struct proc *p, void *data);
2998
2999 struct vm_object_in_map_info {
3000         vm_object_t object;
3001         int rv;
3002 };
3003
3004 /*
3005  * Debugging only
3006  */
3007 static int
3008 vm_object_in_map(vm_object_t object)
3009 {
3010         struct vm_object_in_map_info info;
3011
3012         info.rv = 0;
3013         info.object = object;
3014
3015         allproc_scan(vm_object_in_map_callback, &info);
3016         if (info.rv)
3017                 return 1;
3018         if( _vm_object_in_map(&kernel_map, object, 0))
3019                 return 1;
3020         if( _vm_object_in_map(&pager_map, object, 0))
3021                 return 1;
3022         if( _vm_object_in_map(&buffer_map, object, 0))
3023                 return 1;
3024         return 0;
3025 }
3026
3027 /*
3028  * Debugging only
3029  */
3030 static int
3031 vm_object_in_map_callback(struct proc *p, void *data)
3032 {
3033         struct vm_object_in_map_info *info = data;
3034
3035         if (p->p_vmspace) {
3036                 if (_vm_object_in_map(&p->p_vmspace->vm_map, info->object, 0)) {
3037                         info->rv = 1;
3038                         return -1;
3039                 }
3040         }
3041         return (0);
3042 }
3043
3044 DB_SHOW_COMMAND(vmochk, vm_object_check)
3045 {
3046         struct vm_object_hash *hash;
3047         vm_object_t object;
3048         int n;
3049
3050         /*
3051          * make sure that internal objs are in a map somewhere
3052          * and none have zero ref counts.
3053          */
3054         for (n = 0; n < VMOBJ_HSIZE; ++n) {
3055                 hash = &vm_object_hash[n];
3056                 for (object = TAILQ_FIRST(&hash->list);
3057                                 object != NULL;
3058                                 object = TAILQ_NEXT(object, object_list)) {
3059                         if (object->type == OBJT_MARKER)
3060                                 continue;
3061                         if (object->handle != NULL ||
3062                             (object->type != OBJT_DEFAULT &&
3063                              object->type != OBJT_SWAP)) {
3064                                 continue;
3065                         }
3066                         if (object->ref_count == 0) {
3067                                 db_printf("vmochk: internal obj has "
3068                                           "zero ref count: %ld\n",
3069                                           (long)object->size);
3070                         }
3071                         if (vm_object_in_map(object))
3072                                 continue;
3073                         db_printf("vmochk: internal obj is not in a map: "
3074                                   "ref: %d, size: %lu: 0x%lx, "
3075                                   "backing_object: %p\n",
3076                                   object->ref_count, (u_long)object->size,
3077                                   (u_long)object->size,
3078                                   (void *)object->backing_object);
3079                 }
3080         }
3081 }
3082
3083 /*
3084  * Debugging only
3085  */
3086 DB_SHOW_COMMAND(object, vm_object_print_static)
3087 {
3088         /* XXX convert args. */
3089         vm_object_t object = (vm_object_t)addr;
3090         boolean_t full = have_addr;
3091
3092         vm_page_t p;
3093
3094         /* XXX count is an (unused) arg.  Avoid shadowing it. */
3095 #define count   was_count
3096
3097         int count;
3098
3099         if (object == NULL)
3100                 return;
3101
3102         db_iprintf(
3103             "Object %p: type=%d, size=0x%lx, res=%ld, ref=%d, flags=0x%x\n",
3104             object, (int)object->type, (u_long)object->size,
3105             object->resident_page_count, object->ref_count, object->flags);
3106         /*
3107          * XXX no %qd in kernel.  Truncate object->backing_object_offset.
3108          */
3109         db_iprintf(" sref=%d, backing_object(%d)=(%p)+0x%lx\n",
3110             object->shadow_count, 
3111             object->backing_object ? object->backing_object->ref_count : 0,
3112             object->backing_object, (long)object->backing_object_offset);
3113
3114         if (!full)
3115                 return;
3116
3117         db_indent += 2;
3118         count = 0;
3119         RB_FOREACH(p, vm_page_rb_tree, &object->rb_memq) {
3120                 if (count == 0)
3121                         db_iprintf("memory:=");
3122                 else if (count == 6) {
3123                         db_printf("\n");
3124                         db_iprintf(" ...");
3125                         count = 0;
3126                 } else
3127                         db_printf(",");
3128                 count++;
3129
3130                 db_printf("(off=0x%lx,page=0x%lx)",
3131                     (u_long) p->pindex, (u_long) VM_PAGE_TO_PHYS(p));
3132         }
3133         if (count != 0)
3134                 db_printf("\n");
3135         db_indent -= 2;
3136 }
3137
3138 /* XXX. */
3139 #undef count
3140
3141 /*
3142  * XXX need this non-static entry for calling from vm_map_print.
3143  *
3144  * Debugging only
3145  */
3146 void
3147 vm_object_print(/* db_expr_t */ long addr,
3148                 boolean_t have_addr,
3149                 /* db_expr_t */ long count,
3150                 char *modif)
3151 {
3152         vm_object_print_static(addr, have_addr, count, modif);
3153 }
3154
3155 /*
3156  * Debugging only
3157  */
3158 DB_SHOW_COMMAND(vmopag, vm_object_print_pages)
3159 {
3160         struct vm_object_hash *hash;
3161         vm_object_t object;
3162         int nl = 0;
3163         int c;
3164         int n;
3165
3166         for (n = 0; n < VMOBJ_HSIZE; ++n) {
3167                 hash = &vm_object_hash[n];
3168                 for (object = TAILQ_FIRST(&hash->list);
3169                                 object != NULL;
3170                                 object = TAILQ_NEXT(object, object_list)) {
3171                         vm_pindex_t idx, fidx;
3172                         vm_pindex_t osize;
3173                         vm_paddr_t pa = -1, padiff;
3174                         int rcount;
3175                         vm_page_t m;
3176
3177                         if (object->type == OBJT_MARKER)
3178                                 continue;
3179                         db_printf("new object: %p\n", (void *)object);
3180                         if ( nl > 18) {
3181                                 c = cngetc();
3182                                 if (c != ' ')
3183                                         return;
3184                                 nl = 0;
3185                         }
3186                         nl++;
3187                         rcount = 0;
3188                         fidx = 0;
3189                         osize = object->size;
3190                         if (osize > 128)
3191                                 osize = 128;
3192                         for (idx = 0; idx < osize; idx++) {
3193                                 m = vm_page_lookup(object, idx);
3194                                 if (m == NULL) {
3195                                         if (rcount) {
3196                                                 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
3197                                                         (long)fidx, rcount, (long)pa);
3198                                                 if ( nl > 18) {
3199                                                         c = cngetc();
3200                                                         if (c != ' ')
3201                                                                 return;
3202                                                         nl = 0;
3203                                                 }
3204                                                 nl++;
3205                                                 rcount = 0;
3206                                         }
3207                                         continue;
3208                                 }
3209
3210                                 if (rcount &&
3211                                         (VM_PAGE_TO_PHYS(m) == pa + rcount * PAGE_SIZE)) {
3212                                         ++rcount;
3213                                         continue;
3214                                 }
3215                                 if (rcount) {
3216                                         padiff = pa + rcount * PAGE_SIZE - VM_PAGE_TO_PHYS(m);
3217                                         padiff >>= PAGE_SHIFT;
3218                                         padiff &= PQ_L2_MASK;
3219                                         if (padiff == 0) {
3220                                                 pa = VM_PAGE_TO_PHYS(m) - rcount * PAGE_SIZE;
3221                                                 ++rcount;
3222                                                 continue;
3223                                         }
3224                                         db_printf(" index(%ld)run(%d)pa(0x%lx)",
3225                                                 (long)fidx, rcount, (long)pa);
3226                                         db_printf("pd(%ld)\n", (long)padiff);
3227                                         if ( nl > 18) {
3228                                                 c = cngetc();
3229                                                 if (c != ' ')
3230                                                         return;
3231                                                 nl = 0;
3232                                         }
3233                                         nl++;
3234                                 }
3235                                 fidx = idx;
3236                                 pa = VM_PAGE_TO_PHYS(m);
3237                                 rcount = 1;
3238                         }
3239                         if (rcount) {
3240                                 db_printf(" index(%ld)run(%d)pa(0x%lx)\n",
3241                                         (long)fidx, rcount, (long)pa);
3242                                 if ( nl > 18) {
3243                                         c = cngetc();
3244                                         if (c != ' ')
3245                                                 return;
3246                                         nl = 0;
3247                                 }
3248                                 nl++;
3249                         }
3250                 }
3251         }
3252 }
3253 #endif /* DDB */